From acd8b8ea842da106ec13405744cce1f6c0670597 Mon Sep 17 00:00:00 2001
From: liuhy <liuhy6@sugon.com>
Date: Thu, 24 Aug 2023 11:01:43 +0000
Subject: [PATCH] =?UTF-8?q?=E6=8F=90=E4=BA=A4FT=E5=92=8CCK=E4=BA=A4?=
 =?UTF-8?q?=E5=8F=89=E7=BC=96=E8=AF=91=E4=BB=A3=E7=A0=81?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 3rdparty/composable_kernel/.clang-format      |   90 +
 3rdparty/composable_kernel/.clang-tidy        |    3 +
 3rdparty/composable_kernel/.gitignore         |   49 +
 3rdparty/composable_kernel/CITATION.cff       |   67 +
 3rdparty/composable_kernel/CMakeLists.txt     |  311 +++
 3rdparty/composable_kernel/CONTRIBUTORS.md    |   31 +
 3rdparty/composable_kernel/Config.cmake.in    |   11 +
 3rdparty/composable_kernel/Dockerfile         |  110 +
 3rdparty/composable_kernel/Jenkinsfile        |  705 ++++++
 3rdparty/composable_kernel/LICENSE            |   28 +
 3rdparty/composable_kernel/README.md          |   96 +
 .../client_example/01_gemm/CMakeLists.txt     |    2 +
 .../client_example/01_gemm/gemm.cpp           |  227 ++
 .../02_gemm_add_add_fastgelu/CMakeLists.txt   |   13 +
 .../gemm_add_add_fastgelu.cpp                 |  241 ++
 .../gemm_add_fastgelu.cpp                     |  233 ++
 .../gemm_fastgelu.cpp                         |  225 ++
 .../03_gemm_layernorm/CMakeLists.txt          |    2 +
 .../gemm_add_add_layernorm.cpp                |  274 +++
 .../04_contraction/CMakeLists.txt             |    6 +
 .../04_contraction/contraction_bilinear.cpp   |  236 ++
 .../04_contraction/contraction_scale.cpp      |  222 ++
 .../05_layernorm/CMakeLists.txt               |    2 +
 .../05_layernorm/layernorm2d.cpp              |  163 ++
 .../client_example/06_softmax/CMakeLists.txt  |    2 +
 .../client_example/06_softmax/softmax4d.cpp   |  150 ++
 .../07_grouped_conv2d_fwd/CMakeLists.txt      |    2 +
 .../grouped_conv2d_fwd.cpp                    |  255 +++
 .../08_fused_attention/CMakeLists.txt         |    2 +
 .../08_fused_attention/fused_attention.cpp    |  213 ++
 .../09_quantization/CMakeLists.txt            |   11 +
 ..._fwd_bias_relu_perchannel_quantization.cpp |  205 ++
 ...2d_fwd_bias_relu_perlayer_quantization.cpp |  198 ++
 .../conv2d_fwd_perchannel_quantization.cpp    |  198 ++
 .../conv2d_fwd_perlayer_quantization.cpp      |  192 ++
 .../10_grouped_conv2d_bwd_data/CMakeLists.txt |    2 +
 .../grouped_conv2d_bwd_data.cpp               |  226 ++
 .../11_grouped_conv_bwd_weight/CMakeLists.txt |    2 +
 .../grouped_conv2d_bwd_weight.cpp             |  190 ++
 .../CMakeLists.txt                            |    2 +
 .../elementwise_layernorm2d.cpp               |  175 ++
 .../13_batchnorm/CMakeLists.txt               |    4 +
 .../13_batchnorm/batchnorm_bwd_nhwc.cpp       |  201 ++
 .../13_batchnorm/batchnorm_fwd_nhwc.cpp       |  197 ++
 .../14_instance_id/CMakeLists.txt             |    2 +
 .../batchnorm_fwd_instance_id.cpp             |  206 ++
 .../client_example/CMakeLists.txt             |   15 +
 .../client_example/README.md                  |   18 +
 .../composable_kernel/cmake/Analyzers.cmake   |   34 +
 .../composable_kernel/cmake/ClangTidy.cmake   |  162 ++
 .../composable_kernel/cmake/CppCheck.cmake    |  130 ++
 .../composable_kernel/cmake/DoxygenDoc.cmake  |  355 +++
 .../cmake/EnableCompilerWarnings.cmake        |  110 +
 .../composable_kernel/cmake/TargetFlags.cmake |   50 +
 .../composable_kernel/cmake/googletest.cmake  |   50 +
 3rdparty/composable_kernel/compile.sh         |   13 +
 .../composable_kernel/dev-requirements.txt    |    3 +
 .../doc/image/ck_component.png                |  Bin 0 -> 565049 bytes
 .../composable_kernel/doc/image/ck_layer.png  |  Bin 0 -> 549343 bytes
 .../doc/markdown/dockerhub.md                 |   93 +
 .../example/01_gemm/CMakeLists.txt            |   37 +
 .../example/01_gemm/README.md                 |   23 +
 .../example/01_gemm/common.hpp                |   89 +
 .../example/01_gemm/gemm_dl_fp16.cpp          |   37 +
 .../example/01_gemm/gemm_dl_fp32.cpp          |   37 +
 .../example/01_gemm/gemm_dl_int4.cpp          |   45 +
 .../example/01_gemm/gemm_dl_int8.cpp          |   37 +
 .../example/01_gemm/gemm_xdl_bf16.cpp         |   38 +
 .../example/01_gemm/gemm_xdl_fp16.cpp         |   50 +
 .../example/01_gemm/gemm_xdl_fp64.cpp         |   46 +
 .../example/01_gemm/gemm_xdl_int4.cpp         |   46 +
 .../example/01_gemm/gemm_xdl_int8.cpp         |   38 +
 .../01_gemm/gemm_xdl_skip_b_lds_fp16.cpp      |  249 +++
 .../example/01_gemm/run_gemm_example.inc      |  151 ++
 .../example/02_gemm_bilinear/CMakeLists.txt   |    1 +
 .../example/02_gemm_bilinear/README.md        |   28 +
 .../gemm_bilinear_xdl_fp16.cpp                |  306 +++
 .../example/03_gemm_bias_relu/CMakeLists.txt  |    1 +
 .../example/03_gemm_bias_relu/README.md       |   10 +
 .../gemm_bias_relu_xdl_fp16.cpp               |  283 +++
 .../04_gemm_add_add_fastgelu/CMakeLists.txt   |   17 +
 .../04_gemm_add_add_fastgelu/README.md        |   23 +
 .../04_gemm_add_add_fastgelu/common.hpp       |  106 +
 .../gemm_add_add_fastgelu_xdl_bf16.cpp        |   47 +
 .../gemm_add_add_fastgelu_xdl_fp16.cpp        |   47 +
 .../gemm_add_add_fastgelu_xdl_fp32.cpp        |   47 +
 .../gemm_add_add_fastgelu_xdl_int4.cpp        |   59 +
 .../gemm_add_add_fastgelu_xdl_int8.cpp        |   47 +
 .../run_gemm_add_add_fastgelu_example.inc     |  166 ++
 .../example/09_convnd_fwd/CMakeLists.txt      |   11 +
 .../example/09_convnd_fwd/README.md           |   32 +
 .../example/09_convnd_fwd/compile.sh          |   10 +
 .../09_convnd_fwd/convnd_fwd_common.hpp       |  172 ++
 .../09_convnd_fwd/convnd_fwd_dl_common.hpp    |  196 ++
 .../09_convnd_fwd/convnd_fwd_dl_fp16.cpp      |   40 +
 .../09_convnd_fwd/convnd_fwd_dl_fp32.cpp      |   40 +
 .../09_convnd_fwd/convnd_fwd_dl_int8.cpp      |   40 +
 .../09_convnd_fwd/convnd_fwd_xdl_bf16.cpp     |   79 +
 .../09_convnd_fwd/convnd_fwd_xdl_fp16.cpp     |   79 +
 .../09_convnd_fwd/convnd_fwd_xdl_fp32.cpp     |   79 +
 .../09_convnd_fwd/convnd_fwd_xdl_fp64.cpp     |   79 +
 .../09_convnd_fwd/convnd_fwd_xdl_int8.cpp     |   79 +
 .../run_convnd_fwd_dl_example.inc             |   98 +
 .../09_convnd_fwd/run_convnd_fwd_example.inc  |   97 +
 .../CMakeLists.txt                            |   16 +
 .../common.hpp                                |  159 ++
 .../convnd_fwd_max_xdl_bf16.cpp               |   18 +
 .../convnd_fwd_max_xdl_fp16.cpp               |   18 +
 .../convnd_fwd_max_xdl_fp32.cpp               |   18 +
 .../convnd_fwd_max_xdl_int4.cpp               |   26 +
 .../convnd_fwd_max_xdl_int8.cpp               |   18 +
 .../run_convnd_fwd_max_example.inc            |  307 +++
 .../example/12_reduce/CMakeLists.txt          |    3 +
 .../example/12_reduce/README.md               |   62 +
 .../example/12_reduce/reduce_blockwise.cpp    |  299 +++
 .../12_reduce/reduce_blockwise_impl.hpp       |  338 +++
 .../12_reduce/reduce_blockwise_two_call.cpp   |  301 +++
 .../12_reduce/reduce_example_common.hpp       |   49 +
 .../reduce_multiblock_atomic_add.cpp          |  216 ++
 .../reduce_multiblock_atomic_add_impl.hpp     |  233 ++
 .../example/13_pool2d_fwd/CMakeLists.txt      |    3 +
 .../example/13_pool2d_fwd/README.md           |   41 +
 .../13_pool2d_fwd/pool2d_fwd_common.hpp       |  283 +++
 .../example/13_pool2d_fwd/pool2d_fwd_fp16.cpp |  117 +
 .../example/13_pool2d_fwd/pool2d_fwd_fp32.cpp |  117 +
 .../14_gemm_quantization/CMakeLists.txt       |    2 +
 .../gemm_xdl_bias_relu_quantization_int8.cpp  |  235 ++
 .../gemm_xdl_quantization_int8.cpp            |  207 ++
 .../example/15_grouped_gemm/CMakeLists.txt    |   17 +
 .../example/15_grouped_gemm/README.md         |   25 +
 .../grouped_gemm_xdl_bfp16.cpp                |   62 +
 .../15_grouped_gemm/grouped_gemm_xdl_fp16.cpp |   62 +
 .../15_grouped_gemm/grouped_gemm_xdl_fp32.cpp |   62 +
 .../15_grouped_gemm/grouped_gemm_xdl_int4.cpp |  102 +
 .../15_grouped_gemm/grouped_gemm_xdl_int8.cpp |   59 +
 .../run_grouped_gemm_example.inc              |  265 +++
 .../CMakeLists.txt                            |   40 +
 .../gemm_add_add_mean_meansquare_xdl_fp16.cpp |  276 +++
 .../gemm_add_addsquare_xdl_int8.cpp           |  364 +++
 .../gemm_max_xdl_bf16.cpp                     |  167 ++
 .../gemm_max_xdl_fp16.cpp                     |  167 ++
 .../gemm_max_xdl_fp32.cpp                     |  166 ++
 .../gemm_max_xdl_int4.cpp                     |  172 ++
 .../gemm_max_xdl_int8.cpp                     |  166 ++
 .../gemm_mean_meansquare_xdl_bf16.cpp         |  174 ++
 .../gemm_mean_meansquare_xdl_fp16.cpp         |  174 ++
 .../gemm_mean_meansquare_xdl_fp32.cpp         |  174 ++
 .../gemm_reduce_xdl_common.hpp                |  491 +++++
 .../example/17_convnd_bwd_data/CMakeLists.txt |    5 +
 .../example/17_convnd_bwd_data/README.md      |   47 +
 .../convnd_bwd_data_common.hpp                |  152 ++
 .../convnd_bwd_data_dl_fp16.cpp               |  180 ++
 .../convnd_bwd_data_xdl_fp16.cpp              |  207 ++
 .../18_batched_gemm_reduce/CMakeLists.txt     |    2 +
 .../batched_gemm_reduce_xdl_fp16.cpp          |  311 +++
 .../19_binary_elementwise/CMakeLists.txt      |    4 +
 .../broadcast_add_2d_amn_bn.cpp               |  136 ++
 .../broadcast_add_3d_am_bmnk.cpp              |  120 +
 .../elementwise_add_1d.cpp                    |  111 +
 .../elementwise_add_4d.cpp                    |  120 +
 .../20_grouped_conv_bwd_weight/CMakeLists.txt |    8 +
 .../20_grouped_conv_bwd_weight/common.hpp     |  138 ++
 .../grouped_conv_bwd_weight_xdl_bf16.cpp      |   18 +
 .../grouped_conv_bwd_weight_xdl_fp16.cpp      |   17 +
 .../run_grouped_conv_bwd_weight_example.inc   |  206 ++
 .../example/21_gemm_layernorm/CMakeLists.txt  |    3 +
 .../gemm_bias_relu_add_layernorm_xdl_fp16.cpp |  406 ++++
 .../gemm_layernorm_xdl_fp16.cpp               |  375 ++++
 .../gemm_xdl_layernorm_single_kernel_fp16.cpp |  287 +++
 .../example/22_cgemm/CMakeLists.txt           |   17 +
 .../example/22_cgemm/cgemm_xdl_bf16.cpp       |  132 ++
 .../example/22_cgemm/cgemm_xdl_common.hpp     |  254 +++
 .../example/22_cgemm/cgemm_xdl_fp16.cpp       |  131 ++
 .../example/22_cgemm/cgemm_xdl_fp32.cpp       |  132 ++
 .../example/22_cgemm/cgemm_xdl_int4.cpp       |  140 ++
 .../example/22_cgemm/cgemm_xdl_int8.cpp       |  132 ++
 .../example/23_softmax/CMakeLists.txt         |    1 +
 .../example/23_softmax/README.md              |   18 +
 .../example/23_softmax/softmax_blockwise.cpp  |  264 +++
 .../example/24_batched_gemm/CMakeLists.txt    |   17 +
 .../batched_gemm_xdl_bfp16.cpp                |   59 +
 .../24_batched_gemm/batched_gemm_xdl_fp16.cpp |   59 +
 .../24_batched_gemm/batched_gemm_xdl_fp32.cpp |   58 +
 .../24_batched_gemm/batched_gemm_xdl_int4.cpp |   99 +
 .../24_batched_gemm/batched_gemm_xdl_int8.cpp |   56 +
 .../run_batched_gemm_example.inc              |  240 ++
 .../25_gemm_bias_e_permute/CMakeLists.txt     |    2 +
 .../gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp |  397 ++++
 .../gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp |  398 ++++
 .../example/26_contraction/CMakeLists.txt     |    2 +
 .../example/26_contraction/README.md          |   20 +
 .../contraction_bilinear_xdl_fp32.cpp         |  427 ++++
 .../contraction_scale_xdl_fp32.cpp            |  409 ++++
 .../example/27_layernorm/CMakeLists.txt       |    1 +
 .../27_layernorm/layernorm_blockwise.cpp      |  139 ++
 .../CMakeLists.txt                            |    1 +
 .../grouped_gemm_bias_e_permute_xdl_fp16.cpp  |  466 ++++
 .../CMakeLists.txt                            |    1 +
 .../batched_gemm_bias_e_permute_xdl_fp16.cpp  |  397 ++++
 .../CMakeLists.txt                            |   22 +
 .../30_grouped_conv_fwd_multiple_d/README.md  |   30 +
 .../30_grouped_conv_fwd_multiple_d/common.hpp |  355 +++
 ...rouped_conv_fwd_bias_relu_add_xdl_bf16.cpp |   26 +
 ...rouped_conv_fwd_bias_relu_add_xdl_fp16.cpp |   26 +
 ...rouped_conv_fwd_bias_relu_add_xdl_fp32.cpp |   26 +
 ...rouped_conv_fwd_bias_relu_add_xdl_int4.cpp |   31 +
 ...rouped_conv_fwd_bias_relu_add_xdl_int8.cpp |   26 +
 .../grouped_conv_fwd_xdl_fp16.cpp             |   24 +
 ...grouped_conv_fwd_bias_relu_add_example.inc |  288 +++
 .../run_grouped_conv_fwd_example.inc          |  223 ++
 .../31_batched_gemm_gemm/CMakeLists.txt       |    8 +
 .../batched_gemm_gemm_xdl_bf16.cpp            |  136 ++
 .../batched_gemm_gemm_xdl_fp16.cpp            |  136 ++
 .../batched_gemm_gemm_xdl_fp32.cpp            |  135 ++
 .../batched_gemm_gemm_xdl_int4.cpp            |  146 ++
 .../batched_gemm_gemm_xdl_int8.cpp            |  133 ++
 .../run_batched_gemm_gemm_example.inc         |  277 +++
 .../CMakeLists.txt                            |   16 +
 ...le_scale_softmax_gemm_permute_xdl_fp16.cpp |  160 ++
 ...mm_scale_softmax_gemm_permute_xdl_bf16.cpp |  159 ++
 ...mm_scale_softmax_gemm_permute_xdl_fp16.cpp |  160 ++
 ...tched_gemm_scale_softmax_gemm_xdl_bf16.cpp |  143 ++
 ...tched_gemm_scale_softmax_gemm_xdl_fp16.cpp |  144 ++
 ...le_scale_softmax_gemm_permute_xdl_fp16.cpp |  159 ++
 ...mm_scale_softmax_gemm_permute_xdl_fp16.cpp |  160 ++
 .../run_batched_gemm_scale_softmax_gemm.inc   |  261 +++
 ...atched_gemm_scale_softmax_gemm_permute.inc |  278 +++
 ...rouped_gemm_scale_softmax_gemm_permute.inc |  319 +++
 .../example/33_multiple_reduce/CMakeLists.txt |    2 +
 .../example/33_multiple_reduce/README.md      |   37 +
 .../33_multiple_reduce/dual_reduce_common.hpp |  314 +++
 .../dual_reduce_multiblock.cpp                |   98 +
 .../dual_reduce_threadwise.cpp                |   93 +
 .../example/34_batchnorm/CMakeLists.txt       |    3 +
 .../example/34_batchnorm/README.md            |   81 +
 .../34_batchnorm/batchnorm_backward_nhwc.cpp  |  506 +++++
 .../example/34_batchnorm/batchnorm_common.hpp |   68 +
 .../batchnorm_forward_inferring_nhwc.cpp      |  366 +++
 .../batchnorm_forward_training_nhwc.cpp       |  591 +++++
 .../34_batchnorm/batchnorm_infer_impl.hpp     |  131 ++
 .../example/35_splitK_gemm/CMakeLists.txt     |   17 +
 .../run_splitK_gemm_example.inc               |  217 ++
 .../35_splitK_gemm/splitK_gemm_xdl_bfp16.cpp  |   58 +
 .../35_splitK_gemm/splitK_gemm_xdl_fp16.cpp   |   58 +
 .../35_splitK_gemm/splitK_gemm_xdl_fp32.cpp   |   58 +
 .../35_splitK_gemm/splitK_gemm_xdl_int4.cpp   |   92 +
 .../35_splitK_gemm/splitK_gemm_xdl_int8.cpp   |   55 +
 .../36_sparse_embedding/CMakeLists.txt        |    1 +
 .../sparse_embedding3_forward_layernorm.cpp   |  219 ++
 .../CMakeLists.txt                            |    1 +
 ...ed_gemm_add_add_relu_gemm_add_xdl_fp16.cpp |  519 +++++
 .../CMakeLists.txt                            |    7 +
 .../common.hpp                                |  103 +
 .../grouped_conv_bwd_data_bias_relu_fp16.cpp  |   33 +
 .../grouped_conv_bwd_data_fp16.cpp            |   33 +
 ...rouped_conv_bwd_data_bias_relu_example.inc |  224 ++
 .../run_grouped_conv_bwd_data_example.inc     |  190 ++
 .../example/39_permute/CMakeLists.txt         |    9 +
 .../example/39_permute/common.hpp             |  456 ++++
 .../example/39_permute/permute_1xHxW_fp16.cpp |   20 +
 .../example/39_permute/permute_HxWx4_fp16.cpp |   22 +
 .../example/39_permute/permute_NxHxW_fp16.cpp |   20 +
 .../39_permute/run_permute_bundle_example.inc |   78 +
 .../run_permute_element_example.inc           |   65 +
 .../41_grouped_conv_conv_fwd/CMakeLists.txt   |    8 +
 .../grouped_conv_conv_fwd_xdl_bf16.cpp        |  109 +
 .../grouped_conv_conv_fwd_xdl_fp16.cpp        |  109 +
 .../grouped_conv_conv_fwd_xdl_fp32.cpp        |  109 +
 .../grouped_conv_conv_fwd_xdl_int4.cpp        |  122 +
 .../grouped_conv_conv_fwd_xdl_int8.cpp        |  109 +
 .../run_grouped_conv_conv_fwd_example.inc     |  379 ++++
 .../example/42_groupnorm/CMakeLists.txt       |    1 +
 .../42_groupnorm/groupnorm_sigmoid_fp16.cpp   |  174 ++
 .../CMakeLists.txt                            |    2 +
 .../splitk_gemm_bias_e_permute_xdl_fp16.cpp   |  407 ++++
 .../splitk_gemm_bias_e_permute_xdl_fp32.cpp   |  407 ++++
 .../44_conv2d_fwd_quantization/CMakeLists.txt |    3 +
 ...bias_relu_perchannel_quantization_int8.cpp |  342 +++
 ...l_bias_relu_perlayer_quantization_int8.cpp |  318 +++
 ...v2d_fwd_xdl_perlayer_quantization_int8.cpp |  279 +++
 .../44_elementwise_permute/CMakeLists.txt     |    2 +
 .../elementwise_permute_4D_fp16.cpp           |  116 +
 .../elementwise_permute_4D_fp16_2d.cpp        |  130 ++
 .../CMakeLists.txt                            |    1 +
 .../elementwise_layernorm_blockwise.cpp       |  195 ++
 .../composable_kernel/example/CMakeLists.txt  |   34 +
 3rdparty/composable_kernel/include/ck/ck.hpp  |  201 ++
 .../include/ck/host_utility/device_prop.hpp   |   54 +
 .../ck/host_utility/hip_check_error.hpp       |   17 +
 .../include/ck/host_utility/io.hpp            |   41 +
 .../include/ck/host_utility/kernel_launch.hpp |   74 +
 ...volution_into_gemm_v4r1_nhwc_kyxc_nhwk.hpp |  275 +++
 ...lution_into_gemm_v4r1r2_nhwc_kyxc_nhwk.hpp |  355 +++
 ...into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw.hpp |  150 ++
 ...lution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp |  132 ++
 ...into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk.hpp |  150 ++
 ...lution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp |  135 ++
 ...lution_into_gemm_v4r4r5_nhwc_kyxc_nhwk.hpp |  147 ++
 ...n3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp |  153 ++
 ...volution_into_gemm_v4r4_nchw_kcyx_nkhw.hpp |  260 +++
 ...volution_into_gemm_v4r4_nhwc_kyxc_nhwk.hpp |  179 ++
 ...lution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp |  132 ++
 ...lution_into_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp |  132 ++
 ...lution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp |  134 ++
 ...volution_into_gemm_v6r1_nchw_kcyx_nkhw.hpp |  135 ++
 .../include/ck/stream_config.hpp              |   14 +
 .../include/ck/tensor/static_tensor.hpp       |  273 +++
 .../tensor_description/cluster_descriptor.hpp |   34 +
 .../multi_index_transform.hpp                 | 1954 +++++++++++++++++
 .../multi_index_transform_helper.hpp          |  130 ++
 .../ck/tensor_description/tensor_adaptor.hpp  |  482 ++++
 .../tensor_description/tensor_descriptor.hpp  |  615 ++++++
 .../tensor_descriptor_helper.hpp              |  165 ++
 .../tensor_space_filling_curve.hpp            |  162 ++
 .../gpu/block/blockwise_gemm_dl_v2r3.hpp      |  412 ++++
 .../gpu/block/blockwise_gemm_dlops_v2r2.hpp   |  397 ++++
 .../gpu/block/blockwise_gemm_dlops_v3.hpp     |  178 ++
 .../gpu/block/blockwise_gemm_xdlops.hpp       | 1000 +++++++++
 .../blockwise_gemm_xdlops_skip_b_lds.hpp      |  321 +++
 .../gpu/block/blockwise_softmax.hpp           |  115 +
 .../blockwise_tensor_slice_transfer_v5r1.hpp  |  156 ++
 .../gpu/block/blockwise_welford.hpp           |  108 +
 .../block/reduction_functions_blockwise.hpp   |  244 ++
 ...hread_group_tensor_slice_transfer_v4r1.hpp |  173 ++
 ...hread_group_tensor_slice_transfer_v6r1.hpp |  134 ++
 ...hread_group_tensor_slice_transfer_v6r2.hpp |  158 ++
 ...hread_group_tensor_slice_transfer_v6r3.hpp |  183 ++
 .../thread_group_tensor_slice_transfer_v7.hpp |  172 ++
 ...nvolution_backward_data_specialization.hpp |   30 +
 ...olution_backward_weight_specialization.hpp |   33 +
 .../convolution_forward_specialization.hpp    |   34 +
 .../gpu/device/device_base.hpp                |   74 +
 .../device_batched_contraction_multiple_d.hpp |   64 +
 .../gpu/device/device_batched_gemm.hpp        |   68 +
 .../device/device_batched_gemm_e_permute.hpp  |   50 +
 .../gpu/device/device_batched_gemm_gemm.hpp   |   59 +
 .../device/device_batched_gemm_multi_d.hpp    |   58 +
 ...atched_gemm_multiple_d_gemm_multiple_d.hpp |   72 +
 .../device_batched_gemm_softmax_gemm.hpp      |   60 +
 ...vice_batched_gemm_softmax_gemm_permute.hpp |   70 +
 .../gpu/device/device_batchnorm_backward.hpp  |   77 +
 .../gpu/device/device_batchnorm_forward.hpp   |   72 +
 .../gpu/device/device_batchnorm_infer.hpp     |   69 +
 .../gpu/device/device_cgemm.hpp               |   51 +
 .../device/device_contraction_multiple_d.hpp  |   63 +
 .../gpu/device/device_conv_bwd_data.hpp       |   49 +
 .../gpu/device/device_conv_fwd.hpp            |   49 +
 .../device_conv_fwd_bias_activation.hpp       |   53 +
 .../device_conv_fwd_bias_activation_add.hpp   |   53 +
 .../gpu/device/device_elementwise_2d.hpp      |  341 +++
 .../gpu/device/device_elementwise_base.hpp    |   45 +
 .../device_elementwise_normalization.hpp      |   68 +
 .../gpu/device/device_gemm.hpp                |   42 +
 .../gpu/device/device_gemm_bias_e_permute.hpp |   51 +
 .../gpu/device/device_gemm_multiple_d.hpp     |   58 +
 .../device_gemm_multiple_d_multiple_r.hpp     |   97 +
 .../gpu/device/device_gemm_reduce.hpp         |   46 +
 .../gpu/device/device_gemm_splitk.hpp         |   64 +
 .../device_grouped_contraction_multiple_d.hpp |   72 +
 ...evice_grouped_conv_bwd_data_multiple_d.hpp |   67 +
 .../device/device_grouped_conv_bwd_weight.hpp |   51 +
 .../gpu/device/device_grouped_conv_fwd.hpp    |   55 +
 ..._conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp |  959 ++++++++
 ...ice_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp |  837 +++++++
 .../device_grouped_conv_fwd_multiple_d.hpp    |   65 +
 .../gpu/device/device_grouped_gemm.hpp        |   51 +
 ...vice_grouped_gemm_softmax_gemm_permute.hpp |   75 +
 ...gemm_softmax_gemm_permute_xdl_cshuffle.hpp |  881 ++++++++
 .../gpu/device/device_multiple_reduce.hpp     |   58 +
 .../gpu/device/device_normalization.hpp       |   62 +
 .../gpu/device/device_permute.hpp             |   37 +
 .../gpu/device/device_pool2d_fwd.hpp          |   40 +
 .../gpu/device/device_reduce.hpp              |   51 +
 .../gpu/device/device_softmax.hpp             |   69 +
 .../device_splitk_contraction_multiple_d.hpp  |   65 +
 ...tk_contraction_multiple_d_xdl_cshuffle.hpp | 1147 ++++++++++
 .../gpu/device/gemm_specialization.hpp        |   58 +
 ...ed_contraction_multiple_d_xdl_cshuffle.hpp | 1040 +++++++++
 .../device_batched_gemm_e_permute_xdl.hpp     |  683 ++++++
 .../device_batched_gemm_gemm_xdl_cshuffle.hpp |  747 +++++++
 .../impl/device_batched_gemm_multi_d_xdl.hpp  |  716 ++++++
 ...ultiple_d_gemm_multiple_d_xdl_cshuffle.hpp |  951 ++++++++
 ...evice_batched_gemm_reduce_xdl_cshuffle.hpp | 1001 +++++++++
 ...gemm_softmax_gemm_permute_xdl_cshuffle.hpp |  859 ++++++++
 ...batched_gemm_softmax_gemm_xdl_cshuffle.hpp |  771 +++++++
 .../device/impl/device_batched_gemm_xdl.hpp   |  668 ++++++
 .../impl/device_batchnorm_backward_impl.hpp   |  874 ++++++++
 .../impl/device_batchnorm_forward_impl.hpp    |  718 ++++++
 .../impl/device_cgemm_4gemm_xdl_cshuffle.hpp  |  948 ++++++++
 ...ce_contraction_multiple_d_xdl_cshuffle.hpp |  779 +++++++
 ...rd_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp |  787 +++++++
 ...ice_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp |  835 +++++++
 ...fle_bias_activation_add_nhwc_kyxc_nhwk.hpp |  968 ++++++++
 ...shuffle_bias_activation_nhwc_kyxc_nhwk.hpp |  925 ++++++++
 ...onv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp |  893 ++++++++
 .../device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp  |  733 +++++++
 ...ice_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp |  268 +++
 ...evice_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp |  642 ++++++
 .../device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp | 1583 +++++++++++++
 ...device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp | 1568 +++++++++++++
 .../gpu/device/impl/device_elementwise.hpp    |  304 +++
 .../device_elementwise_normalization_impl.hpp |  592 +++++
 ...vice_gemm_bias_add_reduce_xdl_cshuffle.hpp |  875 ++++++++
 .../impl/device_gemm_bias_e_permute_xdl.hpp   |  572 +++++
 .../gpu/device/impl/device_gemm_dl.hpp        |  595 +++++
 ...emm_multiple_d_multiple_r_xdl_cshuffle.hpp |  682 ++++++
 .../device_gemm_multiple_d_xdl_cshuffle.hpp   |  698 ++++++
 .../impl/device_gemm_reduce_xdl_cshuffle.hpp  |  835 +++++++
 .../gpu/device/impl/device_gemm_xdl.hpp       |  570 +++++
 .../device/impl/device_gemm_xdl_cshuffle.hpp  |  700 ++++++
 .../device_gemm_xdl_layernorm_cshuffle.hpp    |  773 +++++++
 .../impl/device_gemm_xdl_skip_b_lds.hpp       |  523 +++++
 .../impl/device_gemm_xdl_splitk_c_shuffle.hpp |  650 ++++++
 ...ed_contraction_multiple_d_xdl_cshuffle.hpp |  907 ++++++++
 ...nv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp | 1015 +++++++++
 ...bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp | 1244 +++++++++++
 ...grouped_conv_fwd_multiple_d_multiple_r.hpp |   77 +
 ...fwd_multiple_d_multiple_r_xdl_cshuffle.hpp | 1105 ++++++++++
 ...ouped_conv_fwd_multiple_d_xdl_cshuffle.hpp |  952 ++++++++
 .../device/impl/device_grouped_gemm_xdl.hpp   |  677 ++++++
 .../device_multiple_reduce_multiblock.hpp     |  595 +++++
 .../device_multiple_reduce_threadwise.hpp     |  422 ++++
 .../device/impl/device_normalization_impl.hpp |  476 ++++
 .../gpu/device/impl/device_permute_impl.hpp   |  282 +++
 .../impl/device_pool2d_fwd_nhwc_nhwc.hpp      |  327 +++
 .../gpu/device/impl/device_reduce_common.hpp  |  142 ++
 .../device/impl/device_reduce_multiblock.hpp  |  543 +++++
 .../device/impl/device_reduce_threadwise.hpp  |  382 ++++
 .../gpu/device/impl/device_softmax_impl.hpp   |  423 ++++
 ...ce_sparse_embedding3_forward_layernorm.hpp |  210 ++
 .../gpu/device/masking_specialization.hpp     |   82 +
 .../gpu/device/matrix_padder.hpp              |  382 ++++
 .../gpu/device/reduction_operator_mapping.hpp |  186 ++
 .../gpu/device/tensor_layout.hpp              |  417 ++++
 .../gpu/device/tensor_specialization.hpp      |   28 +
 .../gpu/device/welford_helper.hpp             |   89 +
 .../element/binary_element_wise_operation.hpp |  286 +++
 .../gpu/element/element_wise_operation.hpp    |  304 +++
 .../gpu/element/quantization_operation.hpp    |  124 ++
 .../element/unary_element_wise_operation.hpp  |  268 +++
 ...e_second_half_batchnorm_backward_final.hpp |  498 +++++
 ...gridwise_multiblock_welford_first_half.hpp |  261 +++
 ...rd_second_half_batchnorm_forward_final.hpp |  571 +++++
 ...cond_half_multiblock_reduce_first_half.hpp |  556 +++++
 .../gpu/grid/block_to_ctile_map.hpp           |  546 +++++
 ...dwise_2d_multiple_reduction_multiblock.hpp |  321 +++
 ...dwise_2d_multiple_reduction_threadwise.hpp |  264 +++
 .../grid/gridwise_2d_reduction_multiblock.hpp |  613 ++++++
 .../grid/gridwise_2d_reduction_threadwise.hpp |  474 ++++
 ...wise_batched_gemm_gemm_xdl_cshuffle_v1.hpp |  931 ++++++++
 ...iple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp | 1268 +++++++++++
 ...ched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp | 1131 ++++++++++
 ...e_batchnorm_backward_blockwise_welford.hpp |  554 +++++
 ...se_batchnorm_forward_blockwise_welford.hpp |  483 ++++
 .../grid/gridwise_contraction_dlops_v1r2.hpp  |  662 ++++++
 .../gpu/grid/gridwise_elementwise_1d.hpp      |  195 ++
 .../gpu/grid/gridwise_elementwise_2d.hpp      |  230 ++
 ...elementwise_layernorm_welford_variance.hpp |  500 +++++
 ...e_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp |  997 +++++++++
 .../gpu/grid/gridwise_gemm_dl_multiple_d.hpp  |  685 ++++++
 .../gpu/grid/gridwise_gemm_dl_v1r3.hpp        |  577 +++++
 .../gpu/grid/gridwise_gemm_dlops_v1r2.hpp     |  608 +++++
 .../gpu/grid/gridwise_gemm_dlops_v2.hpp       |  461 ++++
 .../gpu/grid/gridwise_gemm_dlops_v3.hpp       | 1597 ++++++++++++++
 ...emm_multiple_d_multiple_r_xdl_cshuffle.hpp |  944 ++++++++
 .../gridwise_gemm_multiple_d_xdl_cshuffle.hpp |  753 +++++++
 .../grid/gridwise_gemm_pipeline_selector.hpp  |   43 +
 .../gpu/grid/gridwise_gemm_pipeline_v1.hpp    |  369 ++++
 .../gpu/grid/gridwise_gemm_pipeline_v2.hpp    |  128 ++
 .../gridwise_gemm_reduce_xdl_cshuffle_v1.hpp  |  879 ++++++++
 ...e_gemm_split_k_multiple_d_xdl_cshuffle.hpp | 1263 +++++++++++
 .../grid/gridwise_gemm_xdl_cshuffle_v1.hpp    |  653 ++++++
 ...ridwise_gemm_xdl_layernorm_cshuffle_v1.hpp | 1068 +++++++++
 .../grid/gridwise_gemm_xdlops_bwd_weight.hpp  |  983 +++++++++
 .../gridwise_gemm_xdlops_skip_b_lds_v1.hpp    |  678 ++++++
 .../gpu/grid/gridwise_gemm_xdlops_v2r3.hpp    |  557 +++++
 .../gpu/grid/gridwise_gemm_xdlops_v2r4.hpp    |  616 ++++++
 .../gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp  |  721 ++++++
 .../gpu/grid/gridwise_gemm_xdlops_v3r1.hpp    |  723 ++++++
 .../gpu/grid/gridwise_gemm_xdlops_v3r2.hpp    |  762 +++++++
 .../gpu/grid/gridwise_gemm_xdlops_v3r3.hpp    |  801 +++++++
 .../gridwise_normalization_naive_variance.hpp |  360 +++
 ...ridwise_normalization_welford_variance.hpp |  384 ++++
 .../gpu/grid/gridwise_permute.hpp             |  339 +++
 .../gpu/grid/gridwise_set_buffer_value.hpp    |   57 +
 .../gridwise_set_multiple_buffer_value.hpp    |   86 +
 .../gpu/grid/gridwise_softmax.hpp             |  407 ++++
 ...se_sparse_embedding3_forward_layernorm.hpp |  344 +++
 .../thread/reduction_functions_threadwise.hpp |   99 +
 .../gpu/thread/threadwise_contraction_dl.hpp  |  224 ++
 .../gpu/thread/threadwise_gemm_dlops_v3.hpp   |  168 ++
 .../thread/threadwise_tensor_slice_set.hpp    |   60 +
 .../threadwise_tensor_slice_transfer.hpp      | 1301 +++++++++++
 .../threadwise_tensor_slice_transfer_v3r1.hpp |  794 +++++++
 .../threadwise_tensor_slice_transfer_v3r3.hpp |  886 ++++++++
 .../threadwise_tensor_slice_transfer_v4r1.hpp |  175 ++
 .../threadwise_tensor_slice_transfer_v5r1.hpp |  614 ++++++
 .../threadwise_tensor_slice_transfer_v6r1.hpp |  213 ++
 .../threadwise_tensor_slice_transfer_v6r2.hpp |  260 +++
 .../threadwise_tensor_slice_transfer_v6r3.hpp |  310 +++
 .../threadwise_tensor_slice_transfer_v7.hpp   |  298 +++
 .../gpu/thread/threadwise_welford.hpp         |  137 ++
 .../tensor_operation/gpu/warp/xdlops_gemm.hpp |  851 +++++++
 .../transform_contraction_to_gemm.hpp         |  288 +++
 .../transform_conv_bwd_data_to_gemm_v1.hpp    |  583 +++++
 .../transform_conv_fwd_to_gemm.hpp            |  880 ++++++++
 .../include/ck/utility/amd_address_space.hpp  |   45 +
 .../ck/utility/amd_buffer_addressing.hpp      | 1177 ++++++++++
 .../include/ck/utility/amd_inline_asm.hpp     |  359 +++
 .../include/ck/utility/amd_llvm_intrinsic.hpp |   14 +
 .../include/ck/utility/amd_wmma.hpp           |  102 +
 .../include/ck/utility/amd_xdlops.hpp         |  320 +++
 .../include/ck/utility/array.hpp              |   66 +
 .../include/ck/utility/array_multi_index.hpp  |   80 +
 .../ck/utility/c_style_pointer_cast.hpp       |   25 +
 .../include/ck/utility/common_header.hpp      |   51 +
 .../ck/utility/container_element_picker.hpp   |  158 ++
 .../include/ck/utility/container_helper.hpp   |  393 ++++
 .../include/ck/utility/data_type.hpp          | 1059 +++++++++
 .../include/ck/utility/debug.hpp              |   85 +
 .../include/ck/utility/dynamic_buffer.hpp     |  398 ++++
 .../include/ck/utility/enable_if.hpp          |   14 +
 .../include/ck/utility/functional.hpp         |  131 ++
 .../include/ck/utility/functional2.hpp        |   49 +
 .../include/ck/utility/functional3.hpp        |  144 ++
 .../include/ck/utility/functional4.hpp        |   65 +
 .../utility/generic_memory_space_atomic.hpp   |  123 ++
 .../include/ck/utility/get_id.hpp             |   28 +
 .../include/ck/utility/ignore.hpp             |   22 +
 .../include/ck/utility/inner_product.hpp      |  211 ++
 .../include/ck/utility/integral_constant.hpp  |   51 +
 .../ck/utility/is_known_at_compile_time.hpp   |   56 +
 .../include/ck/utility/magic_division.hpp     |  160 ++
 .../include/ck/utility/math.hpp               |  240 ++
 .../include/ck/utility/math_v2.hpp            |  164 ++
 .../include/ck/utility/multi_index.hpp        |   12 +
 .../include/ck/utility/number.hpp             |   18 +
 .../include/ck/utility/print.hpp              |   25 +
 .../include/ck/utility/reduction_common.hpp   |   40 +
 .../include/ck/utility/reduction_enums.hpp    |   41 +
 .../reduction_functions_accumulate.hpp        |  115 +
 .../include/ck/utility/reduction_operator.hpp |  292 +++
 .../include/ck/utility/sequence.hpp           |  899 ++++++++
 .../include/ck/utility/sequence_helper.hpp    |   37 +
 .../include/ck/utility/span.hpp               |   67 +
 .../include/ck/utility/static_buffer.hpp      |  195 ++
 .../ck/utility/statically_indexed_array.hpp   |  105 +
 .../statically_indexed_array_multi_index.hpp  |  162 ++
 .../include/ck/utility/synchronization.hpp    |   32 +
 .../include/ck/utility/thread_group.hpp       |   22 +
 .../include/ck/utility/transpose_vectors.hpp  |  165 ++
 .../include/ck/utility/tuple.hpp              |  221 ++
 .../include/ck/utility/tuple_helper.hpp       |   81 +
 .../include/ck/utility/type.hpp               |   60 +
 .../composable_kernel/library/CMakeLists.txt  |    2 +
 .../cpu/reference_batched_gemm.hpp            |  138 ++
 .../cpu/reference_batchnorm_backward.hpp      |  412 ++++
 .../cpu/reference_batchnorm_forward.hpp       |  368 ++++
 .../cpu/reference_batchnorm_infer.hpp         |  300 +++
 .../cpu/reference_cgemm.hpp                   |  184 ++
 .../cpu/reference_conv_bwd_data.hpp           |  378 ++++
 .../cpu/reference_conv_bwd_weight.hpp         |  327 +++
 .../cpu/reference_conv_fwd.hpp                |  339 +++
 .../reference_conv_fwd_bias_activation.hpp    |  192 ++
 ...reference_conv_fwd_bias_activation_add.hpp |  200 ++
 .../cpu/reference_gemm.hpp                    |  137 ++
 .../cpu/reference_gemm_bias_2d.hpp            |  136 ++
 .../cpu/reference_gemm_bias_activation.hpp    |  140 ++
 .../reference_gemm_bias_activation_add.hpp    |  148 ++
 .../cpu/reference_gemm_layernorm.hpp          |  236 ++
 .../cpu/reference_groupnorm.hpp               |  191 ++
 .../cpu/reference_layernorm.hpp               |  171 ++
 .../cpu/reference_softmax.hpp                 |  175 ++
 ...ce_sparse_embedding3_forward_layernorm.hpp |  205 ++
 .../gpu/naive_conv_fwd.hpp                    |  125 ++
 .../add_device_operation_instance.hpp         |   35 +
 .../device_operation_instance_factory.hpp     |  114 +
 .../gpu/batched_gemm.hpp                      |  259 +++
 .../gpu/batched_gemm_add_relu_gemm_add.hpp    |  139 ++
 .../gpu/batched_gemm_gemm.hpp                 |  113 +
 .../gpu/batched_gemm_softmax_gemm.hpp         |  121 +
 .../gpu/batched_gemm_softmax_gemm_permute.hpp |  185 ++
 .../gpu/batchnorm_backward.hpp                |  124 ++
 .../gpu/batchnorm_forward.hpp                 |  117 +
 .../gpu/contraction_bilinear.hpp              |  128 ++
 .../gpu/contraction_scale.hpp                 |  127 ++
 .../gpu/convolution_backward_data.hpp         |  309 +++
 .../gpu/convolution_forward.hpp               |  128 ++
 .../gpu/device_elementwise_instance.hpp       |   54 +
 .../device_gemm_mean_squaremean_instance.hpp  |   84 +
 .../gpu/elementwise_normalization.hpp         |   79 +
 .../tensor_operation_instance/gpu/gemm.hpp    |  385 ++++
 .../gpu/gemm_add_add_fastgelu.hpp             |  155 ++
 .../gpu/gemm_add_fastgelu.hpp                 |  145 ++
 .../gpu/gemm_bilinear.hpp                     |  148 ++
 .../gpu/gemm_fastgelu.hpp                     |  138 ++
 .../gpu/gemm_splitk.hpp                       |  147 ++
 .../gpu/grouped_convolution_backward_data.hpp |   90 +
 .../grouped_convolution_backward_weight.hpp   |  235 ++
 .../gpu/grouped_convolution_forward.hpp       |  396 ++++
 .../gpu/grouped_gemm.hpp                      |  137 ++
 .../gpu/normalization.hpp                     |  109 +
 ...n_bias_forward_perchannel_quantization.hpp |  114 +
 ...ion_bias_forward_perlayer_quantization.hpp |  114 +
 ...lution_forward_perchannel_quantization.hpp |  113 +
 ...volution_forward_perlayer_quantization.hpp |  110 +
 .../gpu/reduce/device_reduce_instance.hpp     |   79 +
 .../device_reduce_instance_blockwise.hpp      |  123 ++
 ...uce_instance_blockwise_b16_f32_b16_add.hpp |   27 +
 ...ce_instance_blockwise_b16_f32_b16_amax.hpp |   31 +
 ...uce_instance_blockwise_b16_f32_b16_avg.hpp |   27 +
 ...uce_instance_blockwise_b16_f32_b16_max.hpp |   31 +
 ...uce_instance_blockwise_b16_f32_b16_min.hpp |   31 +
 ...e_instance_blockwise_b16_f32_b16_norm2.hpp |   27 +
 ...ce_instance_blockwise_f16_f16_f16_amax.hpp |   31 +
 ...uce_instance_blockwise_f16_f16_f16_max.hpp |   31 +
 ...uce_instance_blockwise_f16_f16_f16_min.hpp |   31 +
 ...uce_instance_blockwise_f16_f32_f16_add.hpp |   27 +
 ...uce_instance_blockwise_f16_f32_f16_avg.hpp |   27 +
 ...e_instance_blockwise_f16_f32_f16_norm2.hpp |   27 +
 ...uce_instance_blockwise_f32_f32_f32_add.hpp |   27 +
 ...ce_instance_blockwise_f32_f32_f32_amax.hpp |   31 +
 ...uce_instance_blockwise_f32_f32_f32_avg.hpp |   27 +
 ...uce_instance_blockwise_f32_f32_f32_max.hpp |   31 +
 ...uce_instance_blockwise_f32_f32_f32_min.hpp |   31 +
 ...e_instance_blockwise_f32_f32_f32_norm2.hpp |   27 +
 ...uce_instance_blockwise_f32_f64_f32_add.hpp |   27 +
 ...uce_instance_blockwise_f32_f64_f32_avg.hpp |   27 +
 ...e_instance_blockwise_f32_f64_f32_norm2.hpp |   27 +
 ...uce_instance_blockwise_f64_f64_f64_add.hpp |   27 +
 ...ce_instance_blockwise_f64_f64_f64_amax.hpp |   31 +
 ...uce_instance_blockwise_f64_f64_f64_avg.hpp |   27 +
 ...uce_instance_blockwise_f64_f64_f64_max.hpp |   31 +
 ...uce_instance_blockwise_f64_f64_f64_min.hpp |   31 +
 ...e_instance_blockwise_f64_f64_f64_norm2.hpp |   27 +
 ...educe_instance_blockwise_i8_i32_i8_add.hpp |   27 +
 ...educe_instance_blockwise_i8_i32_i8_avg.hpp |   27 +
 ...educe_instance_blockwise_i8_i8_i8_amax.hpp |   31 +
 ...reduce_instance_blockwise_i8_i8_i8_max.hpp |   31 +
 ...reduce_instance_blockwise_i8_i8_i8_min.hpp |   31 +
 .../device_reduce_instance_impl_common.hpp    |   53 +
 ..._reduce_instance_multiblock_atomic_add.hpp |  123 ++
 ..._multiblock_atomic_add_b16_f32_f32_add.hpp |   27 +
 ..._multiblock_atomic_add_b16_f32_f32_avg.hpp |   27 +
 ..._multiblock_atomic_add_f16_f32_f32_add.hpp |   27 +
 ..._multiblock_atomic_add_f16_f32_f32_avg.hpp |   27 +
 ..._multiblock_atomic_add_f32_f32_f32_add.hpp |   27 +
 ..._multiblock_atomic_add_f32_f32_f32_avg.hpp |   27 +
 ..._multiblock_atomic_add_f32_f64_f32_add.hpp |   28 +
 ..._multiblock_atomic_add_f32_f64_f32_avg.hpp |   28 +
 ..._multiblock_atomic_add_f64_f64_f64_add.hpp |   27 +
 ..._multiblock_atomic_add_f64_f64_f64_avg.hpp |   27 +
 .../device_reduce_instance_threadwise.hpp     |  100 +
 ...ce_instance_threadwise_b16_f32_b16_add.hpp |   27 +
 ...e_instance_threadwise_b16_f32_b16_amax.hpp |   31 +
 ...ce_instance_threadwise_b16_f32_b16_avg.hpp |   27 +
 ...ce_instance_threadwise_b16_f32_b16_max.hpp |   31 +
 ...ce_instance_threadwise_b16_f32_b16_min.hpp |   31 +
 ..._instance_threadwise_b16_f32_b16_norm2.hpp |   27 +
 ...e_instance_threadwise_f16_f16_f16_amax.hpp |   31 +
 ...ce_instance_threadwise_f16_f16_f16_max.hpp |   31 +
 ...ce_instance_threadwise_f16_f16_f16_min.hpp |   31 +
 ...ce_instance_threadwise_f16_f32_f16_add.hpp |   27 +
 ...ce_instance_threadwise_f16_f32_f16_avg.hpp |   27 +
 ..._instance_threadwise_f16_f32_f16_norm2.hpp |   27 +
 ...ce_instance_threadwise_f32_f32_f32_add.hpp |   27 +
 ...e_instance_threadwise_f32_f32_f32_amax.hpp |   31 +
 ...ce_instance_threadwise_f32_f32_f32_avg.hpp |   27 +
 ...ce_instance_threadwise_f32_f32_f32_max.hpp |   31 +
 ...ce_instance_threadwise_f32_f32_f32_min.hpp |   31 +
 ..._instance_threadwise_f32_f32_f32_norm2.hpp |   27 +
 ...ce_instance_threadwise_f32_f64_f32_add.hpp |   27 +
 ...ce_instance_threadwise_f32_f64_f32_avg.hpp |   27 +
 ..._instance_threadwise_f32_f64_f32_norm2.hpp |   27 +
 ...ce_instance_threadwise_f64_f64_f64_add.hpp |   27 +
 ...e_instance_threadwise_f64_f64_f64_amax.hpp |   31 +
 ...ce_instance_threadwise_f64_f64_f64_avg.hpp |   27 +
 ...ce_instance_threadwise_f64_f64_f64_max.hpp |   31 +
 ...ce_instance_threadwise_f64_f64_f64_min.hpp |   31 +
 ..._instance_threadwise_f64_f64_f64_norm2.hpp |   27 +
 ...duce_instance_threadwise_i8_i32_i8_add.hpp |   27 +
 ...duce_instance_threadwise_i8_i32_i8_avg.hpp |   27 +
 ...duce_instance_threadwise_i8_i8_i8_amax.hpp |   31 +
 ...educe_instance_threadwise_i8_i8_i8_max.hpp |   31 +
 ...educe_instance_threadwise_i8_i8_i8_min.hpp |   31 +
 .../tensor_operation_instance/gpu/softmax.hpp |   77 +
 .../device_softmax_f16_f16_instance.hpp       |   22 +
 ...softmax_f16_f16_instance_rank3_reduce1.hpp |   22 +
 ...softmax_f16_f16_instance_rank3_reduce2.hpp |   22 +
 ...softmax_f16_f16_instance_rank3_reduce3.hpp |   22 +
 ...softmax_f16_f16_instance_rank4_reduce1.hpp |   22 +
 ...softmax_f16_f16_instance_rank4_reduce2.hpp |   22 +
 ...softmax_f16_f16_instance_rank4_reduce3.hpp |   22 +
 ...softmax_f16_f16_instance_rank4_reduce4.hpp |   22 +
 .../device_softmax_f16_f16_instance_type.hpp  |   39 +
 .../device_softmax_f32_f32_instance.hpp       |   22 +
 ...softmax_f32_f32_instance_rank3_reduce1.hpp |   22 +
 ...softmax_f32_f32_instance_rank3_reduce2.hpp |   22 +
 ...softmax_f32_f32_instance_rank3_reduce3.hpp |   22 +
 ...softmax_f32_f32_instance_rank4_reduce1.hpp |   22 +
 ...softmax_f32_f32_instance_rank4_reduce2.hpp |   22 +
 ...softmax_f32_f32_instance_rank4_reduce3.hpp |   22 +
 ...softmax_f32_f32_instance_rank4_reduce4.hpp |   22 +
 .../device_softmax_f32_f32_instance_type.hpp  |   38 +
 .../softmax/device_softmax_i8_i8_instance.hpp |   22 +
 ...e_softmax_i8_i8_instance_rank3_reduce1.hpp |   22 +
 ...e_softmax_i8_i8_instance_rank3_reduce2.hpp |   22 +
 ...e_softmax_i8_i8_instance_rank3_reduce3.hpp |   22 +
 ...e_softmax_i8_i8_instance_rank4_reduce1.hpp |   22 +
 ...e_softmax_i8_i8_instance_rank4_reduce2.hpp |   22 +
 ...e_softmax_i8_i8_instance_rank4_reduce3.hpp |   22 +
 ...e_softmax_i8_i8_instance_rank4_reduce4.hpp |   22 +
 .../device_softmax_i8_i8_instance_type.hpp    |   40 +
 .../gpu/softmax/device_softmax_instance.hpp   |    8 +
 .../include/ck/library/utility/algorithm.hpp  |   43 +
 .../include/ck/library/utility/check_err.hpp  |  218 ++
 .../ck/library/utility/conv_common.hpp        |   77 +
 ...volution_host_tensor_descriptor_helper.hpp |  355 +++
 .../library/utility/convolution_parameter.hpp |  101 +
 .../ck/library/utility/device_memory.hpp      |   43 +
 .../include/ck/library/utility/fill.hpp       |  118 +
 .../ck/library/utility/host_common_util.hpp   |  136 ++
 .../include/ck/library/utility/host_conv.hpp  |  152 ++
 .../include/ck/library/utility/host_gemm.hpp  |   47 +
 .../ck/library/utility/host_reduction.hpp     |  374 ++++
 .../ck/library/utility/host_tensor.hpp        |  472 ++++
 .../library/utility/host_tensor_generator.hpp |  190 ++
 .../include/ck/library/utility/iterator.hpp   |   22 +
 .../include/ck/library/utility/literals.hpp   |   20 +
 .../include/ck/library/utility/numeric.hpp    |   16 +
 .../ck/library/utility/op_instance_engine.hpp |  249 +++
 .../include/ck/library/utility/ranges.hpp     |   60 +
 .../gpu/CMakeLists.txt                        |   69 +
 .../gpu/batched_gemm/CMakeLists.txt           |   18 +
 ...dl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp |   59 +
 ...dl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp |   58 +
 ...dl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp |   62 +
 ...dl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp |   63 +
 ...m_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp |   83 +
 ...m_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp |   83 +
 ...m_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp |  110 +
 ...m_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp |   98 +
 ...m_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp |   58 +
 ...m_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp |   58 +
 ...m_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp |   58 +
 ...m_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp |   63 +
 ...dl_int8_int8_int8_gkm_gkn_gmn_instance.cpp |   79 +
 ...dl_int8_int8_int8_gkm_gnk_gmn_instance.cpp |   79 +
 ...dl_int8_int8_int8_gmk_gkn_gmn_instance.cpp |   79 +
 ...dl_int8_int8_int8_gmk_gnk_gmn_instance.cpp |   71 +
 .../CMakeLists.txt                            |    4 +
 ...6_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |   80 +
 ...6_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp |   81 +
 .../gpu/batched_gemm_gemm/CMakeLists.txt      |    4 +
 ...6_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |   80 +
 ...6_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp |   80 +
 .../gpu/batched_gemm_reduce/CMakeLists.txt    |    7 +
 ...6_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp |   80 +
 ...6_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp |   80 +
 ...6_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp |   80 +
 ...6_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp |   77 +
 .../batched_gemm_softmax_gemm/CMakeLists.txt  |    4 +
 ...6_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |  131 ++
 .../CMakeLists.txt                            |    5 +
 ...f16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp |  133 ++
 ...6_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |  133 ++
 .../gpu/batchnorm/CMakeLists.txt              |   10 +
 ...evice_batchnorm_backward_bf16_instance.cpp |  146 ++
 ...device_batchnorm_backward_f16_instance.cpp |  147 ++
 ...device_batchnorm_backward_f32_instance.cpp |  145 ++
 ...device_batchnorm_backward_f64_instance.cpp |  145 ++
 ...device_batchnorm_forward_bf16_instance.cpp |  147 ++
 .../device_batchnorm_forward_f16_instance.cpp |  147 ++
 .../device_batchnorm_forward_f32_instance.cpp |  145 ++
 .../device_batchnorm_forward_f64_instance.cpp |  145 ++
 .../gpu/contraction_bilinear/CMakeLists.txt   |    7 +
 ..._shuffle_f32_f32_f32_f32_kknn_instance.cpp |   79 +
 ..._shuffle_f32_f32_f32_f32_knnn_instance.cpp |   82 +
 ..._shuffle_f32_f32_f32_f32_mknn_instance.cpp |   82 +
 ..._shuffle_f32_f32_f32_f32_mnnn_instance.cpp |   82 +
 .../gpu/contraction_scale/CMakeLists.txt      |    7 +
 ...xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp |   78 +
 ...xdl_c_shuffle_f32_f32_f32_knn_instance.cpp |   81 +
 ...xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp |   81 +
 ...xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp |   81 +
 .../gpu/conv1d_bwd_data/CMakeLists.txt        |    6 +
 ...bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp |  102 +
 ..._bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp |   95 +
 ..._bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp |   94 +
 ...bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp |   99 +
 .../gpu/conv2d_bwd_data/CMakeLists.txt        |   10 +
 ...wd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp |   83 +
 ...wd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp |   83 +
 ...d_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp |   83 +
 ..._data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp |  157 ++
 ...d_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp |  158 ++
 ...d_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp |  155 ++
 ..._data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp |  153 ++
 .../gpu/conv2d_fwd/CMakeLists.txt             |    7 +
 ..._c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp |  155 ++
 ...d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp |  128 ++
 ...2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp |  120 +
 ...2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp |  119 +
 ...d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp |  125 ++
 .../gpu/conv2d_fwd_bias_relu/CMakeLists.txt   |    3 +
 ..._bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp |  154 ++
 .../conv2d_fwd_bias_relu_add/CMakeLists.txt   |    4 +
 ...s_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp |  154 ++
 .../gpu/conv3d_bwd_data/CMakeLists.txt        |    6 +
 ...ta_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp |  102 +
 ...ata_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp |  102 +
 ...ata_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp |  101 +
 ...ta_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp |   99 +
 .../gpu/elementwise/CMakeLists.txt            |    3 +
 .../elementwise/device_normalize_instance.cpp |   49 +
 .../elementwise_normalization/CMakeLists.txt  |    3 +
 ...elementwise_normalization_f16_instance.cpp |   54 +
 .../gpu/gemm/CMakeLists.txt                   |   43 +
 ..._gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp |   52 +
 ..._gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp |   52 +
 ..._gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp |   52 +
 ..._gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp |   53 +
 ..._gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp |   52 +
 ..._gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp |   53 +
 ..._gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp |   53 +
 ..._gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp |   53 +
 ...ice_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp |   49 +
 ...ice_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp |   49 +
 ...ice_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp |   49 +
 ...ice_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp |   49 +
 ..._2_stage_f16_f16_f16_mk_nk_mn_instance.cpp |  100 +
 ...uffle_bf16_bf16_bf16_km_kn_mn_instance.cpp |   68 +
 ...uffle_bf16_bf16_bf16_km_nk_mn_instance.cpp |   68 +
 ...uffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp |   68 +
 ...uffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp |   65 +
 ..._shuffle_f16_f16_f16_km_kn_mn_instance.cpp |  111 +
 ..._shuffle_f16_f16_f16_km_nk_mn_instance.cpp |  111 +
 ..._shuffle_f16_f16_f16_mk_kn_mn_instance.cpp |  111 +
 ..._shuffle_f16_f16_f16_mk_nk_mn_instance.cpp |  102 +
 ..._shuffle_f32_f32_f32_km_kn_mn_instance.cpp |   67 +
 ..._shuffle_f32_f32_f32_km_nk_mn_instance.cpp |   67 +
 ..._shuffle_f32_f32_f32_mk_kn_mn_instance.cpp |   67 +
 ..._shuffle_f32_f32_f32_mk_nk_mn_instance.cpp |   64 +
 ...l_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp |   68 +
 ...l_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp |   68 +
 ...l_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp |   68 +
 ...l_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp |   65 +
 ...gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp |  110 +
 ...gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp |  110 +
 ...gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp |  137 ++
 ...gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp |  130 ++
 ...gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp |   60 +
 ...gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp |   60 +
 ...gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp |   60 +
 ...gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp |   65 +
 ...gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp |   56 +
 ...gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp |   56 +
 ...gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp |   56 +
 ...gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp |   61 +
 .../gpu/gemm_add_add_fastgelu/CMakeLists.txt  |    6 +
 ...16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp |  153 ++
 ...16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp |  153 ++
 ...16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp |  153 ++
 ...16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp |  144 ++
 .../gpu/gemm_add_fastgelu/CMakeLists.txt      |    6 +
 ...e_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp |  136 ++
 ...e_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp |  136 ++
 ...e_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp |  136 ++
 ...e_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp |  127 ++
 .../gpu/gemm_bias_add_reduce/CMakeLists.txt   |    6 +
 ..._f16_f16_f16_f32_f32_km_kn_mn_instance.cpp |   83 +
 ..._f16_f16_f16_f32_f32_km_nk_mn_instance.cpp |   82 +
 ..._f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp |   82 +
 ..._f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp |   79 +
 .../gpu/gemm_bilinear/CMakeLists.txt          |    6 +
 ...e_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp |  105 +
 ...e_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp |  105 +
 ...e_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp |  105 +
 ...e_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp |  142 ++
 .../gpu/gemm_fastgelu/CMakeLists.txt          |    6 +
 ..._shuffle_f16_f16_f16_km_kn_mn_instance.cpp |  135 ++
 ..._shuffle_f16_f16_f16_km_nk_mn_instance.cpp |  135 ++
 ..._shuffle_f16_f16_f16_mk_kn_mn_instance.cpp |  135 ++
 ..._shuffle_f16_f16_f16_mk_nk_mn_instance.cpp |  126 ++
 .../gpu/gemm_reduce/CMakeLists.txt            |    6 +
 ..._f16_f16_f16_f32_f32_km_kn_mn_instance.cpp |   80 +
 ..._f16_f16_f16_f32_f32_km_nk_mn_instance.cpp |   80 +
 ..._f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp |   80 +
 ..._f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp |   77 +
 .../gpu/gemm_splitk/CMakeLists.txt            |   10 +
 ...l_splitk_f16_f16_f16_km_kn_mn_instance.cpp |   61 +
 ...l_splitk_f16_f16_f16_km_nk_mn_instance.cpp |   61 +
 ...l_splitk_f16_f16_f16_mk_kn_mn_instance.cpp |   61 +
 ...l_splitk_f16_f16_f16_mk_nk_mn_instance.cpp |   66 +
 ...l_splitk_f32_f32_f32_km_kn_mn_instance.cpp |   61 +
 ...l_splitk_f32_f32_f32_km_nk_mn_instance.cpp |   61 +
 ...l_splitk_f32_f32_f32_mk_kn_mn_instance.cpp |   61 +
 ...l_splitk_f32_f32_f32_mk_nk_mn_instance.cpp |   66 +
 .../grouped_conv1d_bwd_weight/CMakeLists.txt  |    5 +
 ...eight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp |  106 +
 ...weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp |  104 +
 ...weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp |  103 +
 .../gpu/grouped_conv1d_fwd/CMakeLists.txt     |    6 +
 ...d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp |  129 ++
 ...1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp |  129 ++
 ...1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp |  128 ++
 ...d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp |  125 ++
 .../grouped_conv2d_bwd_data/CMakeLists.txt    |    3 +
 ...ata_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   99 +
 .../grouped_conv2d_bwd_weight/CMakeLists.txt  |    6 +
 ...ht_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp |  106 +
 ...ght_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |  105 +
 ...ght_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |  104 +
 .../gpu/grouped_conv2d_fwd/CMakeLists.txt     |   13 +
 ..._fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |  103 +
 ..._fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |  107 +
 ...fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp |  104 +
 ...wd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp |  156 ++
 ...fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |  156 ++
 ...fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |  128 ++
 ...wd_xdl_gnhwc_gkyxc_gnhwk_int8_instance.cpp |  125 ++
 ...fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp |  156 ++
 .../grouped_conv3d_bwd_weight/CMakeLists.txt  |    5 +
 ...xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp |  106 +
 ..._xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |  106 +
 ..._xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp |  105 +
 .../gpu/grouped_conv3d_fwd/CMakeLists.txt     |    6 +
 ...xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp |  129 ++
 ..._xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |  129 ++
 ..._xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp |  128 ++
 ...xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp |  125 ++
 .../gpu/grouped_gemm/CMakeLists.txt           |    6 +
 ...gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp |   94 +
 ...gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp |   94 +
 ...gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp |   94 +
 ...gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp |   91 +
 .../gpu/normalization/CMakeLists.txt          |    4 +
 .../device_normalization_f16_instance.cpp     |   65 +
 .../device_normalization_f32_instance.cpp     |   60 +
 .../gpu/quantization/CMakeLists.txt           |    6 +
 ..._perchannel_quantization_int8_instance.cpp |   74 +
 ...as_perlayer_quantization_int8_instance.cpp |   68 +
 .../device_conv2d_xdl_int8_instance.hpp       |  111 +
 ..._perchannel_quantization_int8_instance.cpp |   62 +
 ...dl_perlayer_quantization_int8_instance.cpp |   62 +
 .../gpu/reduce/CMakeLists.txt                 |   76 +
 ...uce_instance_blockwise_b16_f32_b16_add.cpp |   24 +
 ...ce_instance_blockwise_b16_f32_b16_amax.cpp |   28 +
 ...uce_instance_blockwise_b16_f32_b16_avg.cpp |   24 +
 ...uce_instance_blockwise_b16_f32_b16_max.cpp |   28 +
 ...uce_instance_blockwise_b16_f32_b16_min.cpp |   28 +
 ...e_instance_blockwise_b16_f32_b16_norm2.cpp |   24 +
 ...ce_instance_blockwise_f16_f16_f16_amax.cpp |   28 +
 ...uce_instance_blockwise_f16_f16_f16_max.cpp |   28 +
 ...uce_instance_blockwise_f16_f16_f16_min.cpp |   28 +
 ...uce_instance_blockwise_f16_f32_f16_add.cpp |   24 +
 ...uce_instance_blockwise_f16_f32_f16_avg.cpp |   24 +
 ...e_instance_blockwise_f16_f32_f16_norm2.cpp |   24 +
 ...uce_instance_blockwise_f32_f32_f32_add.cpp |   24 +
 ...ce_instance_blockwise_f32_f32_f32_amax.cpp |   28 +
 ...uce_instance_blockwise_f32_f32_f32_avg.cpp |   24 +
 ...uce_instance_blockwise_f32_f32_f32_max.cpp |   28 +
 ...uce_instance_blockwise_f32_f32_f32_min.cpp |   28 +
 ...e_instance_blockwise_f32_f32_f32_norm2.cpp |   25 +
 ...uce_instance_blockwise_f32_f64_f32_add.cpp |   23 +
 ...uce_instance_blockwise_f32_f64_f32_avg.cpp |   23 +
 ...e_instance_blockwise_f32_f64_f32_norm2.cpp |   23 +
 ...uce_instance_blockwise_f64_f64_f64_add.cpp |   24 +
 ...ce_instance_blockwise_f64_f64_f64_amax.cpp |   28 +
 ...uce_instance_blockwise_f64_f64_f64_avg.cpp |   24 +
 ...uce_instance_blockwise_f64_f64_f64_max.cpp |   28 +
 ...uce_instance_blockwise_f64_f64_f64_min.cpp |   28 +
 ...e_instance_blockwise_f64_f64_f64_norm2.cpp |   24 +
 ...educe_instance_blockwise_i8_i32_i8_add.cpp |   24 +
 ...educe_instance_blockwise_i8_i32_i8_avg.cpp |   24 +
 ...educe_instance_blockwise_i8_i8_i8_amax.cpp |   28 +
 ...reduce_instance_blockwise_i8_i8_i8_max.cpp |   28 +
 ...reduce_instance_blockwise_i8_i8_i8_min.cpp |   28 +
 ..._multiblock_atomic_add_b16_f32_f32_add.cpp |   23 +
 ..._multiblock_atomic_add_b16_f32_f32_avg.cpp |   23 +
 ..._multiblock_atomic_add_f16_f32_f32_add.cpp |   24 +
 ..._multiblock_atomic_add_f16_f32_f32_avg.cpp |   24 +
 ..._multiblock_atomic_add_f32_f32_f32_add.cpp |   23 +
 ..._multiblock_atomic_add_f32_f32_f32_avg.cpp |   23 +
 ..._multiblock_atomic_add_f32_f64_f32_add.cpp |   23 +
 ..._multiblock_atomic_add_f32_f64_f32_avg.cpp |   23 +
 ..._multiblock_atomic_add_f64_f64_f64_add.cpp |   24 +
 ..._multiblock_atomic_add_f64_f64_f64_avg.cpp |   24 +
 ...ce_instance_threadwise_b16_f32_b16_add.cpp |   24 +
 ...e_instance_threadwise_b16_f32_b16_amax.cpp |   28 +
 ...ce_instance_threadwise_b16_f32_b16_avg.cpp |   24 +
 ...ce_instance_threadwise_b16_f32_b16_max.cpp |   28 +
 ...ce_instance_threadwise_b16_f32_b16_min.cpp |   28 +
 ..._instance_threadwise_b16_f32_b16_norm2.cpp |   24 +
 ...e_instance_threadwise_f16_f16_f16_amax.cpp |   28 +
 ...ce_instance_threadwise_f16_f16_f16_max.cpp |   28 +
 ...ce_instance_threadwise_f16_f16_f16_min.cpp |   28 +
 ...ce_instance_threadwise_f16_f32_f16_add.cpp |   23 +
 ...ce_instance_threadwise_f16_f32_f16_avg.cpp |   23 +
 ..._instance_threadwise_f16_f32_f16_norm2.cpp |   23 +
 ...ce_instance_threadwise_f32_f32_f32_add.cpp |   24 +
 ...e_instance_threadwise_f32_f32_f32_amax.cpp |   28 +
 ...ce_instance_threadwise_f32_f32_f32_avg.cpp |   24 +
 ...ce_instance_threadwise_f32_f32_f32_max.cpp |   28 +
 ...ce_instance_threadwise_f32_f32_f32_min.cpp |   28 +
 ..._instance_threadwise_f32_f32_f32_norm2.cpp |   24 +
 ...ce_instance_threadwise_f32_f64_f32_add.cpp |   24 +
 ...ce_instance_threadwise_f32_f64_f32_avg.cpp |   24 +
 ..._instance_threadwise_f32_f64_f32_norm2.cpp |   24 +
 ...ce_instance_threadwise_f64_f64_f64_add.cpp |   23 +
 ...e_instance_threadwise_f64_f64_f64_amax.cpp |   27 +
 ...ce_instance_threadwise_f64_f64_f64_avg.cpp |   23 +
 ...ce_instance_threadwise_f64_f64_f64_max.cpp |   27 +
 ...ce_instance_threadwise_f64_f64_f64_min.cpp |   27 +
 ..._instance_threadwise_f64_f64_f64_norm2.cpp |   23 +
 ...duce_instance_threadwise_i8_i32_i8_add.cpp |   25 +
 ...duce_instance_threadwise_i8_i32_i8_avg.cpp |   24 +
 ...duce_instance_threadwise_i8_i8_i8_amax.cpp |   28 +
 ...educe_instance_threadwise_i8_i8_i8_max.cpp |   28 +
 ...educe_instance_threadwise_i8_i8_i8_min.cpp |   28 +
 .../gpu/softmax/CMakeLists.txt                |   26 +
 .../device_softmax_f16_f16_instance.cpp       |   40 +
 ...softmax_f16_f16_instance_rank3_reduce1.cpp |   27 +
 ...softmax_f16_f16_instance_rank3_reduce2.cpp |   27 +
 ...softmax_f16_f16_instance_rank3_reduce3.cpp |   27 +
 ...softmax_f16_f16_instance_rank4_reduce1.cpp |   27 +
 ...softmax_f16_f16_instance_rank4_reduce2.cpp |   27 +
 ...softmax_f16_f16_instance_rank4_reduce3.cpp |   27 +
 ...softmax_f16_f16_instance_rank4_reduce4.cpp |   27 +
 .../device_softmax_f32_f32_instance.cpp       |   40 +
 ...softmax_f32_f32_instance_rank3_reduce1.cpp |   27 +
 ...softmax_f32_f32_instance_rank3_reduce2.cpp |   27 +
 ...softmax_f32_f32_instance_rank3_reduce3.cpp |   27 +
 ...softmax_f32_f32_instance_rank4_reduce1.cpp |   27 +
 ...softmax_f32_f32_instance_rank4_reduce2.cpp |   27 +
 ...softmax_f32_f32_instance_rank4_reduce3.cpp |   27 +
 ...softmax_f32_f32_instance_rank4_reduce4.cpp |   27 +
 .../softmax/device_softmax_i8_i8_instance.cpp |   40 +
 ...e_softmax_i8_i8_instance_rank3_reduce1.cpp |   27 +
 ...e_softmax_i8_i8_instance_rank3_reduce2.cpp |   27 +
 ...e_softmax_i8_i8_instance_rank3_reduce3.cpp |   27 +
 ...e_softmax_i8_i8_instance_rank4_reduce1.cpp |   27 +
 ...e_softmax_i8_i8_instance_rank4_reduce2.cpp |   27 +
 ...e_softmax_i8_i8_instance_rank4_reduce3.cpp |   27 +
 ...e_softmax_i8_i8_instance_rank4_reduce4.cpp |   27 +
 .../library/src/utility/CMakeLists.txt        |   28 +
 .../src/utility/convolution_parameter.cpp     |  171 ++
 .../library/src/utility/device_memory.cpp     |   29 +
 .../library/src/utility/host_tensor.cpp       |   59 +
 .../composable_kernel/profiler/CMakeLists.txt |    5 +
 3rdparty/composable_kernel/profiler/README.md |   48 +
 .../include/profiler/data_type_enum.hpp       |   20 +
 .../profiler/data_type_enum_helper.hpp        |   77 +
 ...le_batched_gemm_add_relu_gemm_add_impl.hpp |  360 +++
 .../profile_batched_gemm_gemm_impl.hpp        |  319 +++
 .../profiler/profile_batched_gemm_impl.hpp    |  233 ++
 .../profile_batched_gemm_reduce_impl.hpp      |  362 +++
 ...profile_batched_gemm_softmax_gemm_impl.hpp |  347 +++
 ...batched_gemm_softmax_gemm_permute_impl.hpp |  367 ++++
 .../profile_batchnorm_backward_impl.hpp       |  390 ++++
 .../profile_batchnorm_forward_impl.hpp        |  412 ++++
 .../profiler/profile_conv_bwd_data_impl.hpp   |  248 +++
 .../profile_conv_fwd_bias_relu_add_impl.hpp   |  278 +++
 .../profile_conv_fwd_bias_relu_impl.hpp       |  266 +++
 .../profiler/profile_conv_fwd_impl.hpp        |  221 ++
 .../profiler/profile_convnd_bwd_data_impl.hpp |  486 ++++
 .../profile_convnd_bwd_weight_impl.hpp        |  474 ++++
 .../profile_elementwise_layernorm_impl.hpp    |  266 +++
 .../profile_gemm_add_add_fastgelu_impl.hpp    |  242 ++
 .../profile_gemm_add_fastgelu_impl.hpp        |  232 ++
 .../profile_gemm_bias_add_reduce_impl.hpp     |  384 ++++
 .../profiler/profile_gemm_bilinear_impl.hpp   |  234 ++
 .../profiler/profile_gemm_fastgelu_impl.hpp   |  222 ++
 .../include/profiler/profile_gemm_impl.hpp    |  254 +++
 .../profiler/profile_gemm_reduce_impl.hpp     |  353 +++
 .../profiler/profile_gemm_splitk_impl.hpp     |  257 +++
 .../profile_grouped_conv_bwd_weight_impl.hpp  |  252 +++
 .../profile_grouped_conv_fwd_impl.hpp         |  253 +++
 .../profiler/profile_grouped_gemm_impl.hpp    |  291 +++
 .../profiler/profile_groupnorm_impl.hpp       |  208 ++
 .../profiler/profile_layernorm_impl.hpp       |  227 ++
 .../include/profiler/profile_reduce_impl.hpp  |  520 +++++
 .../include/profiler/profile_softmax_impl.hpp |  219 ++
 .../profiler/src/CMakeLists.txt               |   67 +
 .../profiler/src/profile_batched_gemm.cpp     |  203 ++
 ...profile_batched_gemm_add_relu_gemm_add.cpp |  214 ++
 .../src/profile_batched_gemm_gemm.cpp         |  187 ++
 .../src/profile_batched_gemm_reduce.cpp       |  159 ++
 .../profiler/src/profile_batchnorm_bwd.cpp    |  207 ++
 .../profiler/src/profile_batchnorm_fwd.cpp    |  219 ++
 .../profiler/src/profile_conv_bwd_data.cpp    |  190 ++
 .../profiler/src/profile_conv_fwd.cpp         |  192 ++
 .../src/profile_conv_fwd_bias_relu.cpp        |  122 +
 .../src/profile_conv_fwd_bias_relu_add.cpp    |  122 +
 .../profiler/src/profile_gemm.cpp             |  192 ++
 .../src/profile_gemm_add_add_fastgelu.cpp     |  158 ++
 .../src/profile_gemm_add_fastgelu.cpp         |  146 ++
 .../src/profile_gemm_bias_add_reduce.cpp      |  167 ++
 .../profiler/src/profile_gemm_bilinear.cpp    |  152 ++
 .../profiler/src/profile_gemm_fastgelu.cpp    |  137 ++
 .../profiler/src/profile_gemm_reduce.cpp      |  154 ++
 .../profiler/src/profile_gemm_splitk.cpp      |  154 ++
 .../src/profile_grouped_conv_bwd_weight.cpp   |  182 ++
 .../profiler/src/profile_grouped_conv_fwd.cpp |  260 +++
 .../profiler/src/profile_grouped_gemm.cpp     |  169 ++
 .../profiler/src/profile_groupnorm.cpp        |  112 +
 .../profiler/src/profile_layernorm.cpp        |  101 +
 .../profiler/src/profile_reduce.cpp           |  434 ++++
 .../profiler/src/profile_softmax.cpp          |  169 ++
 .../profiler/src/profiler.cpp                 |   30 +
 .../src/profiler_operation_registry.hpp       |   79 +
 3rdparty/composable_kernel/rbuild.ini         |    8 +
 3rdparty/composable_kernel/requirements.txt   |    1 +
 .../script/clang-format-overwrite.sh          |    2 +
 .../composable_kernel/script/cmake-ck-dev.sh  |   19 +
 .../script/cmake-ck-release.sh                |   19 +
 .../composable_kernel/script/count_vgpr.sh    |   20 +
 .../composable_kernel/script/hipclang_opt.sh  |   25 +
 .../script/parse_perf_data.py                 |  290 +++
 .../script/process_perf_data.py               |  308 +++
 .../script/process_perf_data.sh               |   15 +
 .../script/process_qa_data.sh                 |   23 +
 .../script/profile_batched_gemm.sh            |   44 +
 .../script/profile_conv_bwd_data.sh           |   38 +
 .../script/profile_conv_fwd.sh                |   38 +
 .../composable_kernel/script/profile_gemm.sh  |   58 +
 .../script/profile_gemm_bilinear.sh           |   41 +
 .../script/profile_grouped_gemm.sh            |   18 +
 .../script/profile_onnx_gemm.sh               |   31 +
 .../script/profile_reduce_no_index.sh         |   78 +
 .../script/profile_reduce_with_index.sh       |   70 +
 .../script/profile_resnet50.sh                |   69 +
 .../script/profile_splitK_gemm.sh             |   41 +
 .../script/run_full_performance_tests.sh      |  149 ++
 .../script/run_performance_tests.sh           |   60 +
 .../script/test_convnd_fwd.sh                 |  110 +
 .../script/test_reduce_no_index.sh            |   63 +
 .../script/test_reduce_with_index.sh          |   63 +
 .../composable_kernel/test/CMakeLists.txt     |   60 +
 .../test/batched_gemm/CMakeLists.txt          |   15 +
 .../test/batched_gemm/batched_gemm_bf16.cpp   |   44 +
 .../test/batched_gemm/batched_gemm_fp16.cpp   |   44 +
 .../test/batched_gemm/batched_gemm_fp32.cpp   |   44 +
 .../test/batched_gemm/batched_gemm_int8.cpp   |   44 +
 .../test/batched_gemm_gemm/CMakeLists.txt     |    5 +
 .../test_batched_gemm_gemm_fp16.cpp           |  149 ++
 .../test_batched_gemm_gemm_util.hpp           |  189 ++
 .../test/batched_gemm_reduce/CMakeLists.txt   |    3 +
 .../batched_gemm_reduce_fp16.cpp              |   67 +
 .../batched_gemm_softmax_gemm/CMakeLists.txt  |    5 +
 .../test_batched_gemm_softmax_gemm_fp16.cpp   |  176 ++
 .../test_batched_gemm_softmax_gemm_util.hpp   |  197 ++
 .../CMakeLists.txt                            |    8 +
 ...batched_gemm_softmax_gemm_permute_bf16.cpp |  182 ++
 ...batched_gemm_softmax_gemm_permute_fp16.cpp |  182 ++
 ...batched_gemm_softmax_gemm_permute_util.hpp |  368 ++++
 .../test/batchnorm/CMakeLists.txt             |    4 +
 .../test/batchnorm/batchnorm_bwd_rank_4.cpp   |   92 +
 .../test/batchnorm/batchnorm_fwd_rank_4.cpp   |  109 +
 .../test/block_to_ctile_map/CMakeLists.txt    |    1 +
 .../test_block_to_ctile_map.cpp               |  322 +++
 .../test/conv_util/CMakeLists.txt             |    2 +
 .../test/conv_util/conv_util.cpp              |  156 ++
 .../test/convnd_bwd_data/CMakeLists.txt       |    2 +
 .../test/convnd_bwd_data/convnd_bwd_data.cpp  |   93 +
 .../test/convnd_fwd/CMakeLists.txt            |    2 +
 .../test/convnd_fwd/convnd_fwd.cpp            |   92 +
 .../test/data_type/CMakeLists.txt             |    4 +
 .../composable_kernel/test/data_type/int4.cpp |  211 ++
 .../elementwise_normalization/CMakeLists.txt  |    7 +
 .../test_elementwise_layernorm_fp16.cpp       |   47 +
 .../test/gemm/CMakeLists.txt                  |   25 +
 .../composable_kernel/test/gemm/gemm_bf16.cpp |   34 +
 .../composable_kernel/test/gemm/gemm_fp16.cpp |   34 +
 .../composable_kernel/test/gemm/gemm_fp32.cpp |   34 +
 .../composable_kernel/test/gemm/gemm_fp64.cpp |   34 +
 .../composable_kernel/test/gemm/gemm_int8.cpp |   34 +
 .../test/gemm/gemm_standalone_xdl_fp16.cpp    |  162 ++
 .../composable_kernel/test/gemm/gemm_util.hpp |  267 +++
 .../gemm/instance/gemm_f16_nn_instance.cpp    |   86 +
 .../gemm/instance/gemm_f16_nn_instance.hpp    |   41 +
 .../gemm/instance/gemm_f16_nt_instance.cpp    |   86 +
 .../gemm/instance/gemm_f16_nt_instance.hpp    |   41 +
 .../gemm/instance/gemm_f16_tn_instance.cpp    |   86 +
 .../gemm/instance/gemm_f16_tn_instance.hpp    |   41 +
 .../gemm/instance/gemm_f16_tt_instance.cpp    |   86 +
 .../gemm/instance/gemm_f16_tt_instance.hpp    |   41 +
 .../test/gemm/run_gemm_test.inc               |   41 +
 .../test/gemm_reduce/CMakeLists.txt           |    3 +
 .../test/gemm_reduce/gemm_reduce_fp16.cpp     |   49 +
 .../test/gemm_split_k/CMakeLists.txt          |    3 +
 .../test/gemm_split_k/gemm_split_k.cpp        |  261 +++
 .../grouped_convnd_bwd_weight/CMakeLists.txt  |    2 +
 .../grouped_convnd_bwd_weight.cpp             |   91 +
 .../test/grouped_convnd_fwd/CMakeLists.txt    |    3 +
 .../grouped_convnd_fwd/grouped_convnd_fwd.cpp |  270 +++
 .../test/grouped_gemm/CMakeLists.txt          |    3 +
 .../test/grouped_gemm/grouped_gemm_fp16.cpp   |   65 +
 .../test/magic_number_division/CMakeLists.txt |    2 +
 .../magic_number_division.cpp                 |  150 ++
 .../test/normalization/CMakeLists.txt         |   17 +
 .../normalization/test_groupnorm_fp16.cpp     |   51 +
 .../normalization/test_groupnorm_fp32.cpp     |   49 +
 .../normalization/test_layernorm2d_fp16.cpp   |   45 +
 .../normalization/test_layernorm2d_fp32.cpp   |   45 +
 .../test/reduce/CMakeLists.txt                |    7 +
 .../test/reduce/reduce_no_index.cpp           |  248 +++
 .../test/reduce/reduce_with_index.cpp         |  248 +++
 .../test/reference_conv_fwd/CMakeLists.txt    |    2 +
 .../reference_conv_fwd/reference_conv_fwd.cpp |  392 ++++
 .../test/softmax/CMakeLists.txt               |   11 +
 .../test/softmax/test_softmax_interface.cpp   |   86 +
 .../test/softmax/test_softmax_rank3.cpp       |   34 +
 .../test/softmax/test_softmax_rank4.cpp       |   34 +
 .../test/softmax/test_softmax_ut_cases.inc    |   60 +
 .../test/softmax/test_softmax_util.hpp        |  142 ++
 .../test/space_filling_curve/CMakeLists.txt   |    1 +
 .../space_filling_curve.cpp                   |  195 ++
 .../test/wmma_op/CMakeLists.txt               |    2 +
 .../test/wmma_op/wmma_op.cpp                  |   67 +
 .../test/wmma_op/wmma_op_util.hpp             |  369 ++++
 CMakeLists.txt                                |   23 +-
 .../include/ck_extensions/ck_utils.h          |   37 +
 .../epilogue/epilogue_quant_helper.h          |   48 +
 .../include/ck_extensions/ft_gemm_configs.h   |   58 +
 .../gemm/device/DeviceGemmDequant.hpp         |  137 ++
 .../gemm/kernel/default_fpA_intB_traits.h     |  125 ++
 .../interleaved_numeric_conversion.h          |  120 +
 src/fastertransformer/kernels/CMakeLists.txt  |    3 +-
 .../kernels/activation_fp8_kernels.cu         |    1 +
 .../kernels/activation_kernels.cu             |    1 +
 .../kernels/add_residual_kernels.h            |    1 +
 .../kernels/ban_bad_words.cu                  |    1 +
 .../kernels/beam_search_penalty_kernels.h     |    1 +
 .../kernels/beam_search_topk_kernels.cu       |    1 +
 .../kernels/bert_preprocess_kernels.h         |    1 +
 .../kernels/ck_kernels/CMakeLists.txt         |   47 +
 .../kernels/ck_kernels/cmake/Analyzers.cmake  |   34 +
 .../kernels/ck_kernels/cmake/ClangTidy.cmake  |  162 ++
 .../kernels/ck_kernels/cmake/CppCheck.cmake   |  130 ++
 .../kernels/ck_kernels/cmake/DoxygenDoc.cmake |  355 +++
 .../cmake/EnableCompilerWarnings.cmake        |  110 +
 .../ck_kernels/cmake/TargetFlags.cmake        |   50 +
 .../kernels/ck_kernels/cmake/googletest.cmake |   50 +
 .../ck_kernels/cutlass_preprocessors.cc       |  719 ++++++
 .../ck_kernels/cutlass_preprocessors.h        |   70 +
 .../ck_kernels/fpA_intB_gemm/fpA_intB_gemm.h  |  131 ++
 .../fpA_intB_gemm_bf16_uint4.cpp              |   23 +
 .../fpA_intB_gemm_bf16_uint8.cpp              |   23 +
 .../fpA_intB_gemm_dummy_stubs.cpp             |   22 +
 .../fpA_intB_gemm/fpA_intB_gemm_fp16_int4.cpp |   21 +
 .../fpA_intB_gemm/fpA_intB_gemm_fp16_int8.cpp |   21 +
 .../fpA_intB_gemm/fpA_intB_gemm_template.h    |  120 +
 .../kernels/ck_kernels/int8_gemm/int8_gemm.h  |   49 +
 .../ck_kernels/int8_gemm/int8_gemm_bf16.cpp   |   23 +
 .../ck_kernels/int8_gemm/int8_gemm_fp16.cpp   |   21 +
 .../ck_kernels/int8_gemm/int8_gemm_fp32.cpp   |   21 +
 .../ck_kernels/int8_gemm/int8_gemm_int32.cpp  |   21 +
 .../ck_kernels/int8_gemm/int8_gemm_template.h |   71 +
 .../kernels/ck_kernels/kernels.h              |   19 +
 .../ck_kernels/moe_gemm/moe_gemm_kernels.h    |   75 +
 .../moe_gemm/moe_gemm_kernels_bf16_bf16.cpp   |   23 +
 .../moe_gemm/moe_gemm_kernels_bf16_uint4.cpp  |   23 +
 .../moe_gemm/moe_gemm_kernels_bf16_uint8.cpp  |   23 +
 .../moe_gemm/moe_gemm_kernels_fp16_fp16.cpp   |   21 +
 .../moe_gemm/moe_gemm_kernels_fp16_uint4.cpp  |   21 +
 .../moe_gemm/moe_gemm_kernels_fp16_uint8.cpp  |   21 +
 .../moe_gemm/moe_gemm_kernels_fp32_fp32.cpp   |   21 +
 .../moe_gemm/moe_gemm_kernels_template.h      |   56 +
 .../kernels/ck_kernels/utils.h                |   30 +
 .../kernels/custom_ar_kernels.h               |    1 +
 .../kernels/cutlass_kernels/CMakeLists.txt    |   11 +-
 .../cutlass_kernels/cutlass_heuristic.h       |    1 +
 .../cutlass_kernels/cutlass_preprocessors.h   |    1 +
 .../fpA_intB_gemm/fpA_intB_gemm_template.h    |    1 +
 .../int8_gemm/int8_gemm_template.h            |    1 +
 .../moe_gemm/moe_gemm_kernels_template.h      |    1 +
 .../kernels/decoding_kernels.cu               |    1 +
 .../kernels/dequantize_kernels.h              |    1 +
 .../kernels/disentangled_attention_kernels.cu |    1 +
 .../kernels/gen_relative_pos_bias.cu          |    1 +
 .../kernels/image_merge_kernels.h             |    1 +
 .../kernels/image_shift_partition_kernels.h   |    1 +
 .../kernels/layernorm_fp8_kernels.cu          |    1 +
 .../kernels/layernorm_int8_kernels.h          |    1 +
 .../kernels/layernorm_kernels.h               |    1 +
 .../kernels/longformer_kernels.cu             |    1 +
 src/fastertransformer/kernels/moe_kernels.cu  |    1 +
 src/fastertransformer/kernels/moe_kernels.h   |    3 +-
 .../kernels/normalize_kernels.h               |    1 +
 .../online_softmax_beamsearch_kernels.cu      |    1 +
 .../kernels/sampling_penalty_kernels.h        |    1 +
 .../kernels/sampling_topp_kernels.cu          |    1 +
 .../kernels/softmax_int8_kernels.cu           |    1 +
 .../kernels/stop_criteria_kernels.cu          |    1 +
 .../kernels/unfused_attention_fp8_kernels.cu  |    1 +
 .../kernels/unfused_attention_kernels.cu      |    1 +
 src/fastertransformer/kernels/vit_kernels.cu  |    1 +
 .../kernels/xlnet_preprocess_kernels.h        |    1 +
 src/fastertransformer/layers/CMakeLists.txt   |    2 +-
 src/fastertransformer/layers/FfnLayer.cc      |  229 +-
 src/fastertransformer/layers/FfnLayer.h       |    8 +-
 .../layers/attention_layers/CMakeLists.txt    |    4 +-
 .../DecoderSelfAttentionLayer.cc              |  197 +-
 .../DecoderSelfAttentionLayer.h               |    6 +-
 .../GptContextAttentionLayer.cc               |  229 +-
 .../GptContextAttentionLayer.h                |    6 +-
 .../utils/activation_types.h                  |    3 +-
 src/fastertransformer/utils/cuda_utils.h      |   48 +-
 1309 files changed, 200659 insertions(+), 367 deletions(-)
 create mode 100644 3rdparty/composable_kernel/.clang-format
 create mode 100644 3rdparty/composable_kernel/.clang-tidy
 create mode 100644 3rdparty/composable_kernel/.gitignore
 create mode 100644 3rdparty/composable_kernel/CITATION.cff
 create mode 100644 3rdparty/composable_kernel/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/CONTRIBUTORS.md
 create mode 100644 3rdparty/composable_kernel/Config.cmake.in
 create mode 100644 3rdparty/composable_kernel/Dockerfile
 create mode 100644 3rdparty/composable_kernel/Jenkinsfile
 create mode 100644 3rdparty/composable_kernel/LICENSE
 create mode 100644 3rdparty/composable_kernel/README.md
 create mode 100644 3rdparty/composable_kernel/client_example/01_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/01_gemm/gemm.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/03_gemm_layernorm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/03_gemm_layernorm/gemm_add_add_layernorm.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/04_contraction/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/04_contraction/contraction_bilinear.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/04_contraction/contraction_scale.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/05_layernorm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/05_layernorm/layernorm2d.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/06_softmax/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/06_softmax/softmax4d.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/07_grouped_conv2d_fwd/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/07_grouped_conv2d_fwd/grouped_conv2d_fwd.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/08_fused_attention/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/08_fused_attention/fused_attention.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/09_quantization/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/10_grouped_conv2d_bwd_data/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/10_grouped_conv2d_bwd_data/grouped_conv2d_bwd_data.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/12_elementwise_normalization/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/13_batchnorm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/14_instance_id/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp
 create mode 100644 3rdparty/composable_kernel/client_example/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/client_example/README.md
 create mode 100644 3rdparty/composable_kernel/cmake/Analyzers.cmake
 create mode 100644 3rdparty/composable_kernel/cmake/ClangTidy.cmake
 create mode 100644 3rdparty/composable_kernel/cmake/CppCheck.cmake
 create mode 100644 3rdparty/composable_kernel/cmake/DoxygenDoc.cmake
 create mode 100644 3rdparty/composable_kernel/cmake/EnableCompilerWarnings.cmake
 create mode 100644 3rdparty/composable_kernel/cmake/TargetFlags.cmake
 create mode 100644 3rdparty/composable_kernel/cmake/googletest.cmake
 create mode 100644 3rdparty/composable_kernel/compile.sh
 create mode 100644 3rdparty/composable_kernel/dev-requirements.txt
 create mode 100644 3rdparty/composable_kernel/doc/image/ck_component.png
 create mode 100644 3rdparty/composable_kernel/doc/image/ck_layer.png
 create mode 100644 3rdparty/composable_kernel/doc/markdown/dockerhub.md
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/README.md
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/common.hpp
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/gemm_dl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/gemm_dl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/gemm_dl_int4.cpp
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/gemm_dl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/gemm_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/gemm_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/gemm_xdl_fp64.cpp
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/gemm_xdl_int4.cpp
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/gemm_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/01_gemm/run_gemm_example.inc
 create mode 100644 3rdparty/composable_kernel/example/02_gemm_bilinear/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/02_gemm_bilinear/README.md
 create mode 100644 3rdparty/composable_kernel/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/03_gemm_bias_relu/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/03_gemm_bias_relu/README.md
 create mode 100644 3rdparty/composable_kernel/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/README.md
 create mode 100644 3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/common.hpp
 create mode 100644 3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp
 create mode 100644 3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/README.md
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/compile.sh
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_common.hpp
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_common.hpp
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
 create mode 100644 3rdparty/composable_kernel/example/09_convnd_fwd/run_convnd_fwd_example.inc
 create mode 100644 3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
 create mode 100644 3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp
 create mode 100644 3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
 create mode 100644 3rdparty/composable_kernel/example/12_reduce/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/12_reduce/README.md
 create mode 100644 3rdparty/composable_kernel/example/12_reduce/reduce_blockwise.cpp
 create mode 100644 3rdparty/composable_kernel/example/12_reduce/reduce_blockwise_impl.hpp
 create mode 100644 3rdparty/composable_kernel/example/12_reduce/reduce_blockwise_two_call.cpp
 create mode 100644 3rdparty/composable_kernel/example/12_reduce/reduce_example_common.hpp
 create mode 100644 3rdparty/composable_kernel/example/12_reduce/reduce_multiblock_atomic_add.cpp
 create mode 100644 3rdparty/composable_kernel/example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
 create mode 100644 3rdparty/composable_kernel/example/13_pool2d_fwd/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/13_pool2d_fwd/README.md
 create mode 100644 3rdparty/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_common.hpp
 create mode 100644 3rdparty/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/14_gemm_quantization/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/15_grouped_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/15_grouped_gemm/README.md
 create mode 100644 3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_bfp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp
 create mode 100644 3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/15_grouped_gemm/run_grouped_gemm_example.inc
 create mode 100644 3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp
 create mode 100644 3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
 create mode 100644 3rdparty/composable_kernel/example/17_convnd_bwd_data/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/17_convnd_bwd_data/README.md
 create mode 100644 3rdparty/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
 create mode 100644 3rdparty/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/18_batched_gemm_reduce/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/19_binary_elementwise/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
 create mode 100644 3rdparty/composable_kernel/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
 create mode 100644 3rdparty/composable_kernel/example/19_binary_elementwise/elementwise_add_1d.cpp
 create mode 100644 3rdparty/composable_kernel/example/19_binary_elementwise/elementwise_add_4d.cpp
 create mode 100644 3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/common.hpp
 create mode 100644 3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
 create mode 100644 3rdparty/composable_kernel/example/21_gemm_layernorm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/21_gemm_layernorm/gemm_xdl_layernorm_single_kernel_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/22_cgemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_common.hpp
 create mode 100644 3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_int4.cpp
 create mode 100644 3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/23_softmax/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/23_softmax/README.md
 create mode 100644 3rdparty/composable_kernel/example/23_softmax/softmax_blockwise.cpp
 create mode 100644 3rdparty/composable_kernel/example/24_batched_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_bfp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_int4.cpp
 create mode 100644 3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/24_batched_gemm/run_batched_gemm_example.inc
 create mode 100644 3rdparty/composable_kernel/example/25_gemm_bias_e_permute/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/26_contraction/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/26_contraction/README.md
 create mode 100644 3rdparty/composable_kernel/example/26_contraction/contraction_bilinear_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/26_contraction/contraction_scale_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/27_layernorm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/27_layernorm/layernorm_blockwise.cpp
 create mode 100644 3rdparty/composable_kernel/example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/29_batched_gemm_bias_e_permute/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/README.md
 create mode 100644 3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/common.hpp
 create mode 100644 3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp
 create mode 100644 3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc
 create mode 100644 3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc
 create mode 100644 3rdparty/composable_kernel/example/31_batched_gemm_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp
 create mode 100644 3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
 create mode 100644 3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc
 create mode 100644 3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
 create mode 100644 3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
 create mode 100644 3rdparty/composable_kernel/example/33_multiple_reduce/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/33_multiple_reduce/README.md
 create mode 100644 3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_common.hpp
 create mode 100644 3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_multiblock.cpp
 create mode 100644 3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_threadwise.cpp
 create mode 100644 3rdparty/composable_kernel/example/34_batchnorm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/34_batchnorm/README.md
 create mode 100644 3rdparty/composable_kernel/example/34_batchnorm/batchnorm_backward_nhwc.cpp
 create mode 100644 3rdparty/composable_kernel/example/34_batchnorm/batchnorm_common.hpp
 create mode 100644 3rdparty/composable_kernel/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp
 create mode 100644 3rdparty/composable_kernel/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp
 create mode 100644 3rdparty/composable_kernel/example/34_batchnorm/batchnorm_infer_impl.hpp
 create mode 100644 3rdparty/composable_kernel/example/35_splitK_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/35_splitK_gemm/run_splitK_gemm_example.inc
 create mode 100644 3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_bfp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp
 create mode 100644 3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/36_sparse_embedding/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
 create mode 100644 3rdparty/composable_kernel/example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/common.hpp
 create mode 100644 3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc
 create mode 100644 3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc
 create mode 100644 3rdparty/composable_kernel/example/39_permute/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/39_permute/common.hpp
 create mode 100644 3rdparty/composable_kernel/example/39_permute/permute_1xHxW_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/39_permute/permute_HxWx4_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/39_permute/permute_NxHxW_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/39_permute/run_permute_bundle_example.inc
 create mode 100644 3rdparty/composable_kernel/example/39_permute/run_permute_element_example.inc
 create mode 100644 3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp
 create mode 100644 3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc
 create mode 100644 3rdparty/composable_kernel/example/42_groupnorm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/42_groupnorm/groupnorm_sigmoid_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp
 create mode 100644 3rdparty/composable_kernel/example/44_elementwise_permute/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
 create mode 100644 3rdparty/composable_kernel/example/45_elementwise_normalization/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp
 create mode 100644 3rdparty/composable_kernel/example/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/include/ck/ck.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/host_utility/device_prop.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/host_utility/hip_check_error.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/host_utility/io.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/host_utility/kernel_launch.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_data_convolution_into_gemm_v4r1_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v6r1_nchw_kcyx_nkhw.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/stream_config.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor/static_tensor.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_description/cluster_descriptor.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_description/multi_index_transform.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_description/multi_index_transform_helper.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_description/tensor_adaptor.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_description/tensor_descriptor.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_description/tensor_descriptor_helper.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_description/tensor_space_filling_curve.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_welford.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_base.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_cgemm.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_2d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_base.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_normalization.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_permute.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_reduce.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_softmax.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_e_permute_xdl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_sparse_embedding3_forward_layernorm.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/masking_specialization.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/matrix_padder.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/tensor_layout.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/tensor_specialization.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/welford_helper.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/quantization_operation.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_contraction_dlops_v1r2.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dlops_v1r2.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dlops_v2.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dlops_v3.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_normalization_naive_variance.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_sparse_embedding3_forward_layernorm.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r3.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/amd_address_space.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/amd_buffer_addressing.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/amd_inline_asm.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/amd_llvm_intrinsic.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/amd_wmma.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/amd_xdlops.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/array.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/array_multi_index.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/c_style_pointer_cast.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/common_header.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/container_element_picker.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/container_helper.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/data_type.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/debug.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/dynamic_buffer.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/enable_if.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/functional.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/functional2.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/functional3.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/functional4.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/generic_memory_space_atomic.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/get_id.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/ignore.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/inner_product.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/integral_constant.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/is_known_at_compile_time.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/magic_division.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/math.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/math_v2.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/multi_index.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/number.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/print.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/reduction_common.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/reduction_enums.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/reduction_functions_accumulate.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/reduction_operator.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/sequence.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/sequence_helper.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/span.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/static_buffer.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/statically_indexed_array.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/statically_indexed_array_multi_index.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/synchronization.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/thread_group.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/transpose_vectors.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/tuple.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/tuple_helper.hpp
 create mode 100644 3rdparty/composable_kernel/include/ck/utility/type.hpp
 create mode 100644 3rdparty/composable_kernel/library/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_2d.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/normalization.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/algorithm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/check_err.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/conv_common.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/convolution_parameter.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/device_memory.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/fill.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/host_common_util.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/host_conv.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/host_gemm.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/host_reduction.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/host_tensor.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/host_tensor_generator.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/iterator.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/literals.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/numeric.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/op_instance_engine.hpp
 create mode 100644 3rdparty/composable_kernel/library/include/ck/library/utility/ranges.hpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_normalization_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_normalization_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_int8_instance.hpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/utility/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/library/src/utility/convolution_parameter.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/utility/device_memory.cpp
 create mode 100644 3rdparty/composable_kernel/library/src/utility/host_tensor.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/profiler/README.md
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/data_type_enum.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/data_type_enum_helper.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_batchnorm_backward_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_batchnorm_forward_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_conv_bwd_data_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_conv_fwd_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_convnd_bwd_data_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_convnd_bwd_weight_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_bilinear_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_fastgelu_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_reduce_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_splitk_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_grouped_gemm_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_groupnorm_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_layernorm_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_reduce_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/include/profiler/profile_softmax_impl.hpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_batched_gemm.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_batched_gemm_gemm.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_batched_gemm_reduce.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_batchnorm_bwd.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_batchnorm_fwd.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_conv_bwd_data.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_conv_fwd.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_conv_fwd_bias_relu.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_conv_fwd_bias_relu_add.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_gemm.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_gemm_add_add_fastgelu.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_gemm_add_fastgelu.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_gemm_bias_add_reduce.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_gemm_bilinear.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_gemm_fastgelu.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_gemm_reduce.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_gemm_splitk.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_grouped_conv_bwd_weight.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_grouped_conv_fwd.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_grouped_gemm.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_groupnorm.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_layernorm.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_reduce.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profile_softmax.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profiler.cpp
 create mode 100644 3rdparty/composable_kernel/profiler/src/profiler_operation_registry.hpp
 create mode 100644 3rdparty/composable_kernel/rbuild.ini
 create mode 100644 3rdparty/composable_kernel/requirements.txt
 create mode 100644 3rdparty/composable_kernel/script/clang-format-overwrite.sh
 create mode 100644 3rdparty/composable_kernel/script/cmake-ck-dev.sh
 create mode 100644 3rdparty/composable_kernel/script/cmake-ck-release.sh
 create mode 100644 3rdparty/composable_kernel/script/count_vgpr.sh
 create mode 100644 3rdparty/composable_kernel/script/hipclang_opt.sh
 create mode 100644 3rdparty/composable_kernel/script/parse_perf_data.py
 create mode 100644 3rdparty/composable_kernel/script/process_perf_data.py
 create mode 100644 3rdparty/composable_kernel/script/process_perf_data.sh
 create mode 100644 3rdparty/composable_kernel/script/process_qa_data.sh
 create mode 100644 3rdparty/composable_kernel/script/profile_batched_gemm.sh
 create mode 100644 3rdparty/composable_kernel/script/profile_conv_bwd_data.sh
 create mode 100644 3rdparty/composable_kernel/script/profile_conv_fwd.sh
 create mode 100644 3rdparty/composable_kernel/script/profile_gemm.sh
 create mode 100644 3rdparty/composable_kernel/script/profile_gemm_bilinear.sh
 create mode 100644 3rdparty/composable_kernel/script/profile_grouped_gemm.sh
 create mode 100644 3rdparty/composable_kernel/script/profile_onnx_gemm.sh
 create mode 100644 3rdparty/composable_kernel/script/profile_reduce_no_index.sh
 create mode 100644 3rdparty/composable_kernel/script/profile_reduce_with_index.sh
 create mode 100644 3rdparty/composable_kernel/script/profile_resnet50.sh
 create mode 100644 3rdparty/composable_kernel/script/profile_splitK_gemm.sh
 create mode 100644 3rdparty/composable_kernel/script/run_full_performance_tests.sh
 create mode 100644 3rdparty/composable_kernel/script/run_performance_tests.sh
 create mode 100644 3rdparty/composable_kernel/script/test_convnd_fwd.sh
 create mode 100644 3rdparty/composable_kernel/script/test_reduce_no_index.sh
 create mode 100644 3rdparty/composable_kernel/script/test_reduce_with_index.sh
 create mode 100644 3rdparty/composable_kernel/test/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm/batched_gemm_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm/batched_gemm_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm/batched_gemm_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm/batched_gemm_int8.cpp
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm_reduce/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm_softmax_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp
 create mode 100644 3rdparty/composable_kernel/test/batchnorm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/batchnorm/batchnorm_bwd_rank_4.cpp
 create mode 100644 3rdparty/composable_kernel/test/batchnorm/batchnorm_fwd_rank_4.cpp
 create mode 100644 3rdparty/composable_kernel/test/block_to_ctile_map/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/block_to_ctile_map/test_block_to_ctile_map.cpp
 create mode 100644 3rdparty/composable_kernel/test/conv_util/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/conv_util/conv_util.cpp
 create mode 100644 3rdparty/composable_kernel/test/convnd_bwd_data/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/convnd_bwd_data/convnd_bwd_data.cpp
 create mode 100644 3rdparty/composable_kernel/test/convnd_fwd/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/convnd_fwd/convnd_fwd.cpp
 create mode 100644 3rdparty/composable_kernel/test/data_type/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/data_type/int4.cpp
 create mode 100644 3rdparty/composable_kernel/test/elementwise_normalization/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/gemm/gemm_bf16.cpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/gemm_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/gemm_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/gemm_fp64.cpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/gemm_int8.cpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/gemm_standalone_xdl_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/gemm_util.hpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nn_instance.hpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nt_instance.cpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nt_instance.hpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tn_instance.cpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tn_instance.hpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tt_instance.cpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tt_instance.hpp
 create mode 100644 3rdparty/composable_kernel/test/gemm/run_gemm_test.inc
 create mode 100644 3rdparty/composable_kernel/test/gemm_reduce/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/gemm_reduce/gemm_reduce_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/test/gemm_split_k/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/gemm_split_k/gemm_split_k.cpp
 create mode 100644 3rdparty/composable_kernel/test/grouped_convnd_bwd_weight/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/grouped_convnd_bwd_weight/grouped_convnd_bwd_weight.cpp
 create mode 100644 3rdparty/composable_kernel/test/grouped_convnd_fwd/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/grouped_convnd_fwd/grouped_convnd_fwd.cpp
 create mode 100644 3rdparty/composable_kernel/test/grouped_gemm/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/grouped_gemm/grouped_gemm_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/test/magic_number_division/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/magic_number_division/magic_number_division.cpp
 create mode 100644 3rdparty/composable_kernel/test/normalization/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/normalization/test_groupnorm_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/test/normalization/test_groupnorm_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/test/normalization/test_layernorm2d_fp16.cpp
 create mode 100644 3rdparty/composable_kernel/test/normalization/test_layernorm2d_fp32.cpp
 create mode 100644 3rdparty/composable_kernel/test/reduce/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/reduce/reduce_no_index.cpp
 create mode 100644 3rdparty/composable_kernel/test/reduce/reduce_with_index.cpp
 create mode 100644 3rdparty/composable_kernel/test/reference_conv_fwd/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/reference_conv_fwd/reference_conv_fwd.cpp
 create mode 100644 3rdparty/composable_kernel/test/softmax/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/softmax/test_softmax_interface.cpp
 create mode 100644 3rdparty/composable_kernel/test/softmax/test_softmax_rank3.cpp
 create mode 100644 3rdparty/composable_kernel/test/softmax/test_softmax_rank4.cpp
 create mode 100644 3rdparty/composable_kernel/test/softmax/test_softmax_ut_cases.inc
 create mode 100644 3rdparty/composable_kernel/test/softmax/test_softmax_util.hpp
 create mode 100644 3rdparty/composable_kernel/test/space_filling_curve/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/space_filling_curve/space_filling_curve.cpp
 create mode 100644 3rdparty/composable_kernel/test/wmma_op/CMakeLists.txt
 create mode 100644 3rdparty/composable_kernel/test/wmma_op/wmma_op.cpp
 create mode 100644 3rdparty/composable_kernel/test/wmma_op/wmma_op_util.hpp
 create mode 100644 src/fastertransformer/ck_extensions/include/ck_extensions/ck_utils.h
 create mode 100644 src/fastertransformer/ck_extensions/include/ck_extensions/epilogue/epilogue_quant_helper.h
 create mode 100644 src/fastertransformer/ck_extensions/include/ck_extensions/ft_gemm_configs.h
 create mode 100644 src/fastertransformer/ck_extensions/include/ck_extensions/gemm/device/DeviceGemmDequant.hpp
 create mode 100644 src/fastertransformer/ck_extensions/include/ck_extensions/gemm/kernel/default_fpA_intB_traits.h
 create mode 100644 src/fastertransformer/ck_extensions/include/ck_extensions/interleaved_numeric_conversion.h
 create mode 100644 src/fastertransformer/kernels/ck_kernels/CMakeLists.txt
 create mode 100644 src/fastertransformer/kernels/ck_kernels/cmake/Analyzers.cmake
 create mode 100644 src/fastertransformer/kernels/ck_kernels/cmake/ClangTidy.cmake
 create mode 100644 src/fastertransformer/kernels/ck_kernels/cmake/CppCheck.cmake
 create mode 100644 src/fastertransformer/kernels/ck_kernels/cmake/DoxygenDoc.cmake
 create mode 100644 src/fastertransformer/kernels/ck_kernels/cmake/EnableCompilerWarnings.cmake
 create mode 100644 src/fastertransformer/kernels/ck_kernels/cmake/TargetFlags.cmake
 create mode 100644 src/fastertransformer/kernels/ck_kernels/cmake/googletest.cmake
 create mode 100644 src/fastertransformer/kernels/ck_kernels/cutlass_preprocessors.cc
 create mode 100644 src/fastertransformer/kernels/ck_kernels/cutlass_preprocessors.h
 create mode 100644 src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm.h
 create mode 100644 src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_bf16_uint4.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_bf16_uint8.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_dummy_stubs.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_fp16_int4.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_fp16_int8.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h
 create mode 100644 src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm.h
 create mode 100644 src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_bf16.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_fp16.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_fp32.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_int32.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_template.h
 create mode 100644 src/fastertransformer/kernels/ck_kernels/kernels.h
 create mode 100644 src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels.h
 create mode 100644 src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_bf16_bf16.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_bf16_uint4.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_bf16_uint8.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp16_fp16.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp16_uint4.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp16_uint8.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp32_fp32.cpp
 create mode 100644 src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_template.h
 create mode 100644 src/fastertransformer/kernels/ck_kernels/utils.h

diff --git a/3rdparty/composable_kernel/.clang-format b/3rdparty/composable_kernel/.clang-format
new file mode 100644
index 0000000..22f2674
--- /dev/null
+++ b/3rdparty/composable_kernel/.clang-format
@@ -0,0 +1,90 @@
+---
+Language:        Cpp
+AccessModifierOffset: 0
+AlignAfterOpenBracket: Align
+AlignConsecutiveAssignments: true
+AlignConsecutiveDeclarations: false
+AlignEscapedNewlinesLeft: true
+AlignOperands:   true
+AlignTrailingComments: true
+AllowAllParametersOfDeclarationOnNextLine: true
+AllowShortBlocksOnASingleLine: true
+AllowShortCaseLabelsOnASingleLine: true
+AllowShortFunctionsOnASingleLine: All
+AllowShortIfStatementsOnASingleLine: false
+AllowShortLoopsOnASingleLine: false
+AlwaysBreakAfterDefinitionReturnType: None
+AlwaysBreakAfterReturnType: None
+AlwaysBreakBeforeMultilineStrings: false
+AlwaysBreakTemplateDeclarations: true
+BinPackArguments: false
+BinPackParameters: false
+BraceWrapping:   
+  AfterClass:      true
+  AfterControlStatement: true
+  AfterEnum:       true
+  AfterFunction:   true
+  AfterNamespace:  false
+  AfterObjCDeclaration: true
+  AfterStruct:     true
+  AfterUnion:      true
+  BeforeCatch:     true
+  BeforeElse:      true
+  IndentBraces:    false
+BreakBeforeBinaryOperators: None
+BreakBeforeBraces: Custom
+BreakBeforeTernaryOperators: true
+BreakConstructorInitializersBeforeComma: false
+ColumnLimit:     100
+CommentPragmas:  '^ IWYU pragma:'
+ConstructorInitializerAllOnOneLineOrOnePerLine: true
+ConstructorInitializerIndentWidth: 4
+ContinuationIndentWidth: 4
+Cpp11BracedListStyle: true
+DerivePointerAlignment: false
+DisableFormat:   false
+ExperimentalAutoDetectBinPacking: false
+ForEachMacros:   [ foreach, Q_FOREACH, BOOST_FOREACH ]
+IncludeCategories: 
+  - Regex:           '^"(llvm|llvm-c|clang|clang-c)/'
+    Priority:        2
+  - Regex:           '^(<|"(gtest|isl|json)/)'
+    Priority:        3
+  - Regex:           '.*'
+    Priority:        1
+IndentCaseLabels: false
+IndentWidth:     4
+IndentWrappedFunctionNames: false
+KeepEmptyLinesAtTheStartOfBlocks: true
+MacroBlockBegin: ''
+MacroBlockEnd:   ''
+MaxEmptyLinesToKeep: 1
+NamespaceIndentation: None
+ObjCBlockIndentWidth: 2
+ObjCSpaceAfterProperty: false
+ObjCSpaceBeforeProtocolList: true
+PenaltyBreakBeforeFirstCallParameter: 19
+PenaltyBreakComment: 300
+PenaltyBreakFirstLessLess: 120
+PenaltyBreakString: 1000
+PenaltyExcessCharacter: 1000000
+PenaltyReturnTypeOnItsOwnLine: 60
+PointerAlignment: Left
+ReflowComments:  true
+SortIncludes:    false
+SpaceAfterCStyleCast: false
+# SpaceAfterTemplateKeyword: true
+SpaceBeforeAssignmentOperators: true
+SpaceBeforeParens: Never
+SpaceInEmptyParentheses: false
+SpacesBeforeTrailingComments: 1
+SpacesInAngles:  false
+SpacesInContainerLiterals: true
+SpacesInCStyleCastParentheses: false
+SpacesInParentheses: false
+SpacesInSquareBrackets: false
+Standard:        Cpp11
+TabWidth:        8
+UseTab:          Never
+...
+
diff --git a/3rdparty/composable_kernel/.clang-tidy b/3rdparty/composable_kernel/.clang-tidy
new file mode 100644
index 0000000..5c2b781
--- /dev/null
+++ b/3rdparty/composable_kernel/.clang-tidy
@@ -0,0 +1,3 @@
+CheckOptions:
+  - key: bugprone-reserved-identifier.AllowedIdentifiers
+    value: '__HIP_PLATFORM_HCC__;__HIP_ROCclr__'
diff --git a/3rdparty/composable_kernel/.gitignore b/3rdparty/composable_kernel/.gitignore
new file mode 100644
index 0000000..71059ec
--- /dev/null
+++ b/3rdparty/composable_kernel/.gitignore
@@ -0,0 +1,49 @@
+# Compiled Object files
+*.slo
+*.lo
+*.o
+*.obj
+
+# Precompiled Headers
+*.gch
+*.pch
+*.ipch
+
+# Compiled Dynamic libraries
+*.so
+*.dylib
+*.dll
+
+# Fortran module files
+*.mod
+
+# Compiled Static libraries
+*.lai
+*.la
+*.a
+*.lib
+
+# Executables
+*.exe
+*.out
+*.app
+
+# vim tags
+tags
+.tags
+.*.swp
+
+# Editors
+.vscode
+
+# build-in-source directory
+build*
+
+# emacs temporary/backup files
+.\#*
+\#*\#
+*~
+
+# GDB temporary files
+.gdb_history
+install.dir*
diff --git a/3rdparty/composable_kernel/CITATION.cff b/3rdparty/composable_kernel/CITATION.cff
new file mode 100644
index 0000000..d35fe9e
--- /dev/null
+++ b/3rdparty/composable_kernel/CITATION.cff
@@ -0,0 +1,67 @@
+cff-version: 1.2.0
+title: Composable Kernel
+message: If you use this software, please cite using the following metadata.
+type: software
+authors:
+  - given-names: Chao
+    family-names: Liu
+    email: chao.liu2@amd.com
+    affiliation: AMD
+  - given-names: Jing
+    family-names: Zhang
+    email: jing.zhang3@amd.com
+    affiliation: AMD
+  - given-names: Letao
+    family-names: Qin
+    email: letao.qin@amd.com
+    affiliation: AMD
+  - given-names: Qianfeng
+    family-names: Zhang
+    email: qianfeng.zhang@amd.com
+    affiliation: AMD
+  - given-names: Liang
+    family-names: Huang
+    email: carlus.huang@amd.com
+    affiliation: AMD
+  - given-names: Shaojie
+    family-names: Wang
+    email: shaojie.wang@amd.com
+    affiliation: AMD
+  - given-names: Anthony
+    family-names: Chang
+    email: antc@amd.com
+    affiliation: AMD
+  - given-names: Chunyu
+    family-names: Lai
+    email: chunyu.lai@amd.com
+    affiliation: AMD
+  - given-names: Illia
+    family-names: Silin
+    email: illia.silin@amd.com
+    affiliation: AMD
+  - given-names: Adam
+    family-names: Osewski
+    email: adam.osewski@amd.com
+    affiliation: AMD
+  - given-names: Poyen
+    family-names: Chen
+    email: poyen.chen@amd.com
+    affiliation: AMD
+  - given-names: Rosty
+    family-names: Geyyer
+    email: rosty.geyyer@amd.com
+    affiliation: AMD
+  - given-names: Hanwen
+    family-names: Chen
+  - given-names: Tejash
+    family-names: Shah
+  - given-names: Xiaoyan
+    family-names: Zhou
+  - given-names: Jianfeng
+    family-names: Yan
+repository-code: 'https://github.com/ROCmSoftwarePlatform/composable_kernel'
+abstract: Composable Kernel (CK) library aims to provide a programming model for writing performance critical kernels for Machine Learning workloads across multiple architectures including GPUs, CPUs, etc, through general purpose kernel progarmming languages, like HIP C++.
+keywords:
+  - 'CK, Composable Kernel, Tensor Coordinate Transformation'
+license: MIT
+license-url: https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/7fc3ed761aa35709d87c8fbbe41dd368648b3541/LICENSE
diff --git a/3rdparty/composable_kernel/CMakeLists.txt b/3rdparty/composable_kernel/CMakeLists.txt
new file mode 100644
index 0000000..53e5889
--- /dev/null
+++ b/3rdparty/composable_kernel/CMakeLists.txt
@@ -0,0 +1,311 @@
+cmake_minimum_required(VERSION 3.14)
+
+# Check support for CUDA/HIP in Cmake
+project(composable_kernel)
+
+list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
+
+enable_testing()
+set(ROCM_PATH $ENV{ROCM_PATH})
+set(ROCM_SYMLINK_LIBS OFF)
+find_package(ROCM REQUIRED PATHS ${ROCM_PATH})
+
+include(ROCMInstallTargets)
+include(ROCMPackageConfigHelpers)
+include(ROCMSetupVersion)
+include(ROCMInstallSymlinks)
+include(ROCMCreatePackage)
+include(CheckCXXCompilerFlag)
+
+rocm_setup_version(VERSION 0.2.0)
+include(TargetFlags)
+list(APPEND CMAKE_PREFIX_PATH ${CMAKE_INSTALL_PREFIX} ${CMAKE_INSTALL_PREFIX}/llvm ${CMAKE_INSTALL_PREFIX}/hip ${ROCM_PATH} ${ROCM_PATH}/llvm ${ROCM_PATH}/hip)
+
+option(USE_BITINT_EXTENSION_INT4, "Whether to enable clang's BitInt extension to provide int4 data type." OFF)
+
+if(USE_BITINT_EXTENSION_INT4)
+    add_compile_definitions(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    add_compile_options(-Wno-bit-int-extension)
+    message("CK compiled with USE_BITINT_EXTENSION_INT4 set to ${USE_BITINT_EXTENSION_INT4}")
+endif()
+
+## Threads
+set(THREADS_PREFER_PTHREAD_FLAG ON)
+find_package(Threads REQUIRED)
+link_libraries(Threads::Threads)
+
+## C++
+enable_language(CXX)
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CXX_EXTENSIONS OFF)
+message("CMAKE_CXX_COMPILER_ID: ${CMAKE_CXX_COMPILER_ID}")
+
+## OpenMP
+if(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+	# workaround issue hipcc in rocm3.5 cannot find openmp
+	set(OpenMP_CXX "${CMAKE_CXX_COMPILER}")
+	set(OpenMP_CXX_FLAGS "-fopenmp=libomp -Wno-unused-command-line-argument")
+	set(OpenMP_CXX_LIB_NAMES "libomp" "libgomp" "libiomp5")
+	set(OpenMP_libomp_LIBRARY ${OpenMP_CXX_LIB_NAMES})
+	set(OpenMP_libgomp_LIBRARY ${OpenMP_CXX_LIB_NAMES})
+	set(OpenMP_libiomp5_LIBRARY ${OpenMP_CXX_LIB_NAMES})
+else()
+	find_package(OpenMP REQUIRED)
+endif()
+
+message("OpenMP_CXX_LIB_NAMES: ${OpenMP_CXX_LIB_NAMES}")
+message("OpenMP_gomp_LIBRARY: ${OpenMP_gomp_LIBRARY}")
+message("OpenMP_pthread_LIBRARY: ${OpenMP_pthread_LIBRARY}")
+message("OpenMP_CXX_FLAGS: ${OpenMP_CXX_FLAGS}")
+
+link_libraries(${OpenMP_gomp_LIBRARY})
+link_libraries(${OpenMP_pthread_LIBRARY})
+
+## HIP
+find_package(HIP REQUIRED)
+# Override HIP version in config.h, if necessary.
+# The variables set by find_package() can't be overwritten,
+# therefore let's use intermediate variables.
+set(CK_HIP_VERSION_MAJOR "${HIP_VERSION_MAJOR}")
+set(CK_HIP_VERSION_MINOR "${HIP_VERSION_MINOR}")
+set(CK_HIP_VERSION_PATCH "${HIP_VERSION_PATCH}")
+if( DEFINED CK_OVERRIDE_HIP_VERSION_MAJOR )
+    set(CK_HIP_VERSION_MAJOR "${CK_OVERRIDE_HIP_VERSION_MAJOR}")
+    message(STATUS "CK_HIP_VERSION_MAJOR overriden with ${CK_OVERRIDE_HIP_VERSION_MAJOR}")
+endif()
+if( DEFINED CK_OVERRIDE_HIP_VERSION_MINOR )
+    set(CK_HIP_VERSION_MINOR "${CK_OVERRIDE_HIP_VERSION_MINOR}")
+    message(STATUS "CK_HIP_VERSION_MINOR overriden with ${CK_OVERRIDE_HIP_VERSION_MINOR}")
+endif()
+if( DEFINED CK_OVERRIDE_HIP_VERSION_PATCH )
+    set(CK_HIP_VERSION_PATCH "${CK_OVERRIDE_HIP_VERSION_PATCH}")
+    message(STATUS "CK_HIP_VERSION_PATCH overriden with ${CK_OVERRIDE_HIP_VERSION_PATCH}")
+endif()
+message(STATUS "Build with HIP ${HIP_VERSION}")
+link_libraries(hip::device)
+add_compile_definitions(__HIP_PLATFORM_HCC__=1)
+
+## tidy
+include(EnableCompilerWarnings)
+set(CK_TIDY_ERRORS ERRORS * -readability-inconsistent-declaration-parameter-name)
+if(CMAKE_CXX_COMPILER MATCHES ".*hcc" OR CMAKE_CXX_COMPILER MATCHES ".*clang\\+\\+")
+    set(CK_TIDY_CHECKS -modernize-use-override -readability-non-const-parameter)
+# Enable tidy on hip
+elseif(CK_BACKEND STREQUAL "HIP" OR CK_BACKEND STREQUAL "HIPNOGPU")
+    set(CK_TIDY_ERRORS ALL)
+endif()
+
+
+include(ClangTidy)
+enable_clang_tidy(
+    CHECKS
+        *
+        -abseil-*
+        -android-cloexec-fopen
+        # Yea we shouldn't be using rand()
+        -cert-msc30-c
+        -bugprone-exception-escape
+        -bugprone-macro-parentheses
+        -cert-env33-c
+        -cert-msc32-c
+        -cert-msc50-cpp
+        -cert-msc51-cpp
+        -cert-dcl37-c
+        -cert-dcl51-cpp
+        -clang-analyzer-alpha.core.CastToStruct
+        -clang-analyzer-optin.performance.Padding
+        -clang-diagnostic-deprecated-declarations
+        -clang-diagnostic-extern-c-compat
+        -clang-diagnostic-unused-command-line-argument
+        -cppcoreguidelines-avoid-c-arrays
+        -cppcoreguidelines-avoid-magic-numbers
+        -cppcoreguidelines-explicit-virtual-functions
+        -cppcoreguidelines-init-variables
+        -cppcoreguidelines-macro-usage
+        -cppcoreguidelines-non-private-member-variables-in-classes
+        -cppcoreguidelines-pro-bounds-array-to-pointer-decay
+        -cppcoreguidelines-pro-bounds-constant-array-index
+        -cppcoreguidelines-pro-bounds-pointer-arithmetic
+        -cppcoreguidelines-pro-type-member-init
+        -cppcoreguidelines-pro-type-reinterpret-cast
+        -cppcoreguidelines-pro-type-union-access
+        -cppcoreguidelines-pro-type-vararg
+        -cppcoreguidelines-special-member-functions
+        -fuchsia-*
+        -google-explicit-constructor
+        -google-readability-braces-around-statements
+        -google-readability-todo
+        -google-runtime-int
+        -google-runtime-references
+        -hicpp-vararg
+        -hicpp-braces-around-statements
+        -hicpp-explicit-conversions
+        -hicpp-named-parameter
+        -hicpp-no-array-decay
+        # We really shouldn't use bitwise operators with signed integers, but
+        # opencl leaves us no choice
+        -hicpp-avoid-c-arrays
+        -hicpp-signed-bitwise
+        -hicpp-special-member-functions
+        -hicpp-uppercase-literal-suffix
+        -hicpp-use-auto
+        -hicpp-use-equals-default
+        -hicpp-use-override
+        -llvm-header-guard
+        -llvm-include-order
+        #-llvmlibc-*
+        -llvmlibc-restrict-system-libc-headers
+        -llvmlibc-callee-namespace
+        -llvmlibc-implementation-in-namespace
+        -llvm-else-after-return
+        -llvm-qualified-auto
+        -misc-misplaced-const
+        -misc-non-private-member-variables-in-classes
+        -misc-no-recursion
+        -modernize-avoid-bind
+        -modernize-avoid-c-arrays
+        -modernize-pass-by-value
+        -modernize-use-auto
+        -modernize-use-default-member-init
+        -modernize-use-equals-default
+        -modernize-use-trailing-return-type
+        -modernize-use-transparent-functors
+        -performance-unnecessary-value-param
+        -readability-braces-around-statements
+        -readability-else-after-return
+        # we are not ready to use it, but very useful
+        -readability-function-cognitive-complexity
+        -readability-isolate-declaration
+        -readability-magic-numbers
+        -readability-named-parameter
+        -readability-uppercase-literal-suffix
+        -readability-convert-member-functions-to-static
+        -readability-qualified-auto
+        -readability-redundant-string-init
+        # too many narrowing conversions in our code
+        -bugprone-narrowing-conversions
+        -cppcoreguidelines-narrowing-conversions
+        -altera-struct-pack-align
+        -cppcoreguidelines-prefer-member-initializer
+        ${CK_TIDY_CHECKS}
+        ${CK_TIDY_ERRORS}
+    HEADER_FILTER
+        "\.hpp$"
+    EXTRA_ARGS
+        -DCK_USE_CLANG_TIDY
+)
+
+include(CppCheck)
+enable_cppcheck(
+    CHECKS
+        warning
+        style
+        performance
+        portability
+    SUPPRESS
+        ConfigurationNotChecked
+        constStatement
+        duplicateCondition
+        noExplicitConstructor
+        passedByValue
+        preprocessorErrorDirective
+        shadowVariable
+        unusedFunction
+        unusedPrivateFunction
+        unusedStructMember
+        unmatchedSuppression
+    FORCE
+    SOURCES
+        library/src
+    INCLUDE
+        ${CMAKE_CURRENT_SOURCE_DIR}/include
+        ${CMAKE_CURRENT_BINARY_DIR}/include
+        ${CMAKE_CURRENT_SOURCE_DIR}/library/include
+    DEFINE
+        CPPCHECK=1
+        __linux__=1
+)
+
+set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib)
+set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib)
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/bin)
+
+include_directories(BEFORE
+    ${PROJECT_SOURCE_DIR}/include
+    ${PROJECT_SOURCE_DIR}/library/include
+    ${HIP_INCLUDE_DIRS}
+)
+
+
+SET(BUILD_DEV ON CACHE BOOL "BUILD_DEV")
+if(BUILD_DEV)
+    add_compile_options(-Werror)
+    add_compile_options(-Weverything)
+endif()
+message("CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}")
+
+add_custom_target(check COMMAND ${CMAKE_CTEST_COMMAND} --output-on-failure -C ${CMAKE_CFG_INTDIR})
+
+file(GLOB_RECURSE INSTANCE_FILES "${PROJECT_SOURCE_DIR}/*/device_*_instance.cpp")
+file(GLOB dir_list RELATIVE ${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu ${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu/*)
+set(CK_DEVICE_INSTANCES)
+FOREACH(subdir_path ${dir_list})
+    IF(IS_DIRECTORY "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu/${subdir_path}")
+       list(APPEND CK_DEVICE_INSTANCES device_${subdir_path}_instance)
+    ENDIF()
+ENDFOREACH()
+add_custom_target(instances DEPENDS utility;${CK_DEVICE_INSTANCES}  SOURCES ${INSTANCE_FILES})
+
+rocm_package_setup_component(tests
+        LIBRARY_NAME composablekernel
+        PACKAGE_NAME tests # Prevent -static suffix on package name
+)
+
+rocm_package_setup_component(examples
+        LIBRARY_NAME composablekernel
+        PACKAGE_NAME examples
+)
+
+rocm_package_setup_component(profiler
+        LIBRARY_NAME composablekernel
+        PACKAGE_NAME ckProfiler
+)
+
+add_subdirectory(library)
+add_subdirectory(example)
+add_subdirectory(test)
+add_subdirectory(profiler)
+
+#Create an interface target for the include only files and call it "composablekernels"
+include(CMakePackageConfigHelpers)
+
+set(version 1.0.0)
+write_basic_package_version_file(
+    "${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfigVersion.cmake"
+    VERSION "${version}"
+    COMPATIBILITY AnyNewerVersion
+)
+
+configure_package_config_file(${CMAKE_CURRENT_SOURCE_DIR}/Config.cmake.in
+        "${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfig.cmake"
+        INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+        NO_CHECK_REQUIRED_COMPONENTS_MACRO
+)
+
+rocm_install(FILES
+    "${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfig.cmake"
+    "${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfigVersion.cmake"
+    DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+)
+
+set(CPACK_RESOURCE_FILE_LICENSE "${CMAKE_CURRENT_SOURCE_DIR}/LICENSE")
+set(CPACK_RPM_PACKAGE_LICENSE "MIT")
+
+rocm_create_package(
+    NAME composablekernel
+    DESCRIPTION "High Performance Composable Kernel for AMD GPUs"
+    MAINTAINER "MIOpen Kernels Dev Team <dl.MIOpen@amd.com>"
+    LDCONFIG
+    HEADER_ONLY
+)
diff --git a/3rdparty/composable_kernel/CONTRIBUTORS.md b/3rdparty/composable_kernel/CONTRIBUTORS.md
new file mode 100644
index 0000000..8ccfe99
--- /dev/null
+++ b/3rdparty/composable_kernel/CONTRIBUTORS.md
@@ -0,0 +1,31 @@
+# Composable Kernel Developers and Contributors
+
+This is the list of developers and contributors to Composable Kernel library
+
+
+## Developers
+[Chao Liu](https://github.com/asroy), [Jing Zhang](https://github.com/zjing14), 2018-2022
+
+[Letao Qin](https://github.com/ltqin), [Qianfeng Zhang](https://github.com/qianfengz), [Liang Huang](https://github.com/carlushuang), [Shaojie Wang](https://github.com/shaojiewang), 2019-2022
+
+[Anthony Chang](https://github.com/rosenrodt), [Chunyu Lai](https://github.com/rocking5566), [Illia Silin](https://github.com/illsilin), [Adam Osewski](https://github.com/aosewski), [Poyen Chen](https://github.com/poyenc), [Rosty Geyyer](https://github.com/geyyer), 2022
+
+Hanwen Chang, 2019-2021,
+
+Tejash Shah, 2019-2020
+
+Xiaoyan Zhou, 2020
+
+[Jianfeng Yan](https://github.com/j4yan), 2021-2022
+
+
+## Product Manager
+[Jun Liu](https://github.com/junliume)
+
+
+## Contributors
+[Dan Yao](https://github.com/danyao12), [Guangzhao Lu](https://github.com/guangzlu), [Raman Jana](https://github.com/ramjana), [Jehandad Khan](https://github.com/JehandadKhan), [Wen-Heng (Jack) Chung](https://github.com/whchung)
+
+
+## Acknowledgement
+CK team works closely with Meta [AITemplate](https://github.com/facebookincubator/AITemplate) team ([Bing Xu](https://github.com/antinucleon), [Hao Lu](https://github.com/hlu1), [Ying Zhang](https://github.com/ipiszy), etc). Most of the lucrative graph optimization opportunities in ML models were identified by AITemplate team, and we also co-designed many high performance fused kernels for AMD GPUs. Without this collaboration, CK would not reach its current potential.
diff --git a/3rdparty/composable_kernel/Config.cmake.in b/3rdparty/composable_kernel/Config.cmake.in
new file mode 100644
index 0000000..02978cd
--- /dev/null
+++ b/3rdparty/composable_kernel/Config.cmake.in
@@ -0,0 +1,11 @@
+@PACKAGE_INIT@
+
+set(_composable_kernel_supported_components device_operations utility)
+
+foreach(_comp ${composable_kernel_FIND_COMPONENTS})
+	if(NOT _comp IN_LIST _composable_kernel_supported_components)
+		set(composable_kernel_FOUND False)
+		set(composable_kernel_NOT_FOUND_MESSAGE "Unsupported component: ${_comp}")
+	endif()
+	include("${CMAKE_CURRENT_LIST_DIR}/composable_kernel${_comp}Targets.cmake")
+endforeach()
diff --git a/3rdparty/composable_kernel/Dockerfile b/3rdparty/composable_kernel/Dockerfile
new file mode 100644
index 0000000..d024f96
--- /dev/null
+++ b/3rdparty/composable_kernel/Dockerfile
@@ -0,0 +1,110 @@
+FROM ubuntu:20.04
+
+ARG ROCMVERSION=5.3
+ARG compiler_version="release"
+ARG compiler_commit=""
+
+RUN set -xe
+
+ARG DEB_ROCM_REPO=http://repo.radeon.com/rocm/apt/.apt_$ROCMVERSION/
+# Add rocm repository
+RUN apt-get update
+RUN apt-get install -y wget gnupg
+RUN wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -
+RUN sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.list.d/rocm.list"
+RUN wget --no-check-certificate -qO - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | apt-key add -
+RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
+
+# Install dependencies
+RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
+    apt-utils \
+    build-essential \
+    ccache \
+    cmake-data \
+    cmake \
+    curl \
+    git \
+    hip-rocclr \
+    jq \
+    libelf-dev \
+    libncurses5-dev \
+    libnuma-dev \
+    libpthread-stubs0-dev \
+    llvm-amdgpu \
+    pkg-config \
+    python \
+    python3 \
+    python-dev \
+    python3-dev \
+    python3-pip \
+    software-properties-common \
+    rocm-dev \
+    rocm-device-libs \
+    rocm-cmake \
+    vim \
+    zlib1g-dev \
+    openssh-server \
+    clang-format-10 \
+    kmod && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Setup ubsan environment to printstacktrace
+RUN ln -s /usr/bin/llvm-symbolizer-3.8 /usr/local/bin/llvm-symbolizer
+ENV UBSAN_OPTIONS=print_stacktrace=1
+
+# Install an init system
+RUN wget https://github.com/Yelp/dumb-init/releases/download/v1.2.0/dumb-init_1.2.0_amd64.deb
+RUN dpkg -i dumb-init_*.deb && rm dumb-init_*.deb
+
+ARG PREFIX=/opt/rocm
+# Install packages for processing the performance results
+RUN pip3 install --upgrade pip
+RUN pip3 install sqlalchemy
+RUN pip3 install pymysql
+RUN pip3 install pandas
+RUN pip3 install setuptools-rust
+RUN pip3 install sshtunnel
+# Setup ubsan environment to printstacktrace
+ENV UBSAN_OPTIONS=print_stacktrace=1
+
+ENV LC_ALL=C.UTF-8
+ENV LANG=C.UTF-8
+RUN groupadd -f render
+
+# Install the new rocm-cmake version
+RUN git clone -b master https://github.com/RadeonOpenCompute/rocm-cmake.git  && \
+  cd rocm-cmake && mkdir build && cd build && \
+  cmake  .. && cmake --build . && cmake --build . --target install
+
+WORKDIR /
+
+ENV compiler_version=$compiler_version
+ENV compiler_commit=$compiler_commit
+RUN sh -c "echo compiler version = '$compiler_version'"
+RUN sh -c "echo compiler commit = '$compiler_commit'"
+
+RUN --mount=type=ssh if [ "$compiler_version" = "amd-stg-open" ]; then \
+        sed -i '/$HIP_CLANG_TARGET = chomp($HIP_CLANG_TARGET);/c\    chomp($HIP_CLANG_TARGET);' /opt/rocm/hip/bin/hipcc.pl && \
+        sed -i '/$HIP_CLANG_TARGET = chomp($HIP_CLANG_TARGET);/c\    chomp($HIP_CLANG_TARGET);' /opt/rocm/bin/hipcc.pl; \
+    fi
+
+RUN --mount=type=ssh if [ "$compiler_version" != "release" ] && [ "$compiler_commit" = "" ]; then \
+        git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \
+        cd llvm-project && mkdir build && cd build && \
+        cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \
+        make -j 8 ; \
+    else echo "using the release compiler"; \
+    fi
+
+RUN --mount=type=ssh if [ "$compiler_version" != "release" ] && [ "$compiler_commit" != "" ]; then \
+        git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \
+        cd llvm-project && git checkout "$compiler_commit" && echo "checking out commit $compiler_commit" && mkdir build && cd build && \
+        cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \
+        make -j 8 ; \
+    else echo "using the release compiler"; \
+    fi
+
+
+#ENV HIP_CLANG_PATH='/llvm-project/build/bin'
+#RUN sh -c "echo HIP_CLANG_PATH = '$HIP_CLANG_PATH'"
diff --git a/3rdparty/composable_kernel/Jenkinsfile b/3rdparty/composable_kernel/Jenkinsfile
new file mode 100644
index 0000000..7b2e57c
--- /dev/null
+++ b/3rdparty/composable_kernel/Jenkinsfile
@@ -0,0 +1,705 @@
+def rocmnode(name) {
+    return 'rocmtest && miopen && ' + name
+}
+
+def show_node_info() {
+    sh """
+        echo "NODE_NAME = \$NODE_NAME"
+        lsb_release -sd
+        uname -r
+        ls /opt/ -la
+    """
+}
+
+def runShell(String command){
+    def responseCode = sh returnStatus: true, script: "${command} > tmp.txt"
+    def output = readFile(file: "tmp.txt")
+    echo "tmp.txt contents: $output"
+    return (output != "")
+}
+
+def getDockerImageName(){
+    def img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}"
+    return img
+}
+
+def check_host() {
+    if ("${env.CK_CCACHE}" != "null"){
+        def CCACHE_SERVER="${env.CK_CCACHE.split(':')[0]}"
+        echo "ccache server: ${CCACHE_SERVER}"
+        sh '''ping -c 1 -p 6379 "${CCACHE_SERVER}" | echo $? > tmp.txt'''
+        def output = readFile(file: "tmp.txt")
+        echo "tmp.txt contents: \$output"
+        return (output != "0")
+    }
+    else{
+        return 1
+    }
+}
+
+def build_compiler(){
+    def compiler
+    if (params.BUILD_COMPILER == "hipcc"){
+        compiler = '/opt/rocm/bin/hipcc'
+    }
+    else{
+        if (params.COMPILER_VERSION == "release"){
+            compiler = "/opt/rocm/llvm/bin/clang++"
+        }
+        else{
+            compiler = "/llvm-project/build/bin/clang++"
+        }        
+    }
+    return compiler
+}
+
+def getDockerImage(Map conf=[:]){
+    env.DOCKER_BUILDKIT=1
+    def prefixpath = conf.get("prefixpath", "/opt/rocm") // prefix:/opt/rocm
+    def no_cache = conf.get("no_cache", false)
+    def dockerArgs = "--build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
+    echo "ccache server: ${env.CK_CCACHE}"
+    if(env.CK_CCACHE)
+    {
+        if(check_host())
+        {
+            echo "FOUND CCACHE SERVER: ${env.CK_CCACHE}"
+        }
+        else 
+        {
+            echo "CCACHE SERVER: ${env.CK_CCACHE} NOT FOUND, got ${check_host} response"
+        }
+        dockerArgs = dockerArgs + " --build-arg CCACHE_SECONDARY_STORAGE='redis://${env.CK_CCACHE}' --build-arg COMPILER_LAUNCHER='ccache' "
+        env.CCACHE_DIR = """/tmp/ccache_store"""
+        env.CCACHE_SECONDARY_STORAGE="""redis://${env.CK_CCACHE}"""
+    }
+    if(no_cache)
+    {
+        dockerArgs = dockerArgs + " --no-cache "
+    }
+    echo "Docker Args: ${dockerArgs}"
+    def image = getDockerImageName()
+    //Check if image exists 
+    def retimage
+    try 
+    {
+        echo "Pulling down image: ${image}"
+        retimage = docker.image("${image}")
+        retimage.pull()
+    }
+    catch(Exception ex)
+    {
+        error "Unable to locate image: ${image}"
+    }
+    return [retimage, image]
+}
+
+def buildDocker(install_prefix){
+    show_node_info()
+    env.DOCKER_BUILDKIT=1
+    checkout scm
+    def image_name = getDockerImageName()
+    echo "Building Docker for ${image_name}"
+    def dockerArgs = "--build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${install_prefix} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
+    echo "ccache server: ${env.CK_CCACHE}"
+    if(env.CK_CCACHE)
+    {
+        if(check_host())
+        {
+            echo "FOUND CCACHE SERVER: ${env.CK_CCACHE}"
+        }
+        else 
+        {
+            echo "CCACHE SERVER: ${env.CK_CCACHE} NOT FOUND, got ${check_host} response"
+        }
+        dockerArgs = dockerArgs + " --build-arg CCACHE_SECONDARY_STORAGE='redis://${env.CK_CCACHE}' --build-arg COMPILER_LAUNCHER='ccache' "
+        env.CCACHE_DIR = """/tmp/ccache_store"""
+        env.CCACHE_SECONDARY_STORAGE="""redis://${env.CK_CCACHE}"""
+    }
+
+    echo "Build Args: ${dockerArgs}"
+    try{
+        if(params.BUILD_DOCKER){
+            //force building the new docker if that parameter is true
+            echo "Building image: ${image_name}"
+            retimage = docker.build("${image_name}", dockerArgs + ' .')
+            retimage.push()
+        }
+        else{
+            echo "Checking for image: ${image_name}"
+            sh "docker manifest inspect --insecure ${image_name}"
+            echo "Image: ${image_name} found!! Skipping building image"
+        }
+    }
+    catch(Exception ex){
+        echo "Unable to locate image: ${image_name}. Building image now"
+        retimage = docker.build("${image_name}", dockerArgs + ' .')
+        retimage.push()
+    }
+}
+
+def cmake_build(Map conf=[:]){
+
+    def compiler = build_compiler()
+    def config_targets = conf.get("config_targets","check")
+    def debug_flags = "-g -fno-omit-frame-pointer -fsanitize=undefined -fno-sanitize-recover=undefined " + conf.get("extradebugflags", "")
+    def build_envs = "CTEST_PARALLEL_LEVEL=4 " + conf.get("build_env","")
+    def prefixpath = conf.get("prefixpath","/opt/rocm")
+    def setup_args = conf.get("setup_args","")
+
+    if (prefixpath != "/usr/local"){
+        setup_args = setup_args + " -DCMAKE_PREFIX_PATH=${prefixpath} "
+    }
+
+    def build_type_debug = (conf.get("build_type",'release') == 'debug')
+
+    //cmake_env can overwrite default CXX variables.
+    def cmake_envs = "CXX=${compiler} CXXFLAGS='-Werror' " + conf.get("cmake_ex_env","")
+
+    def package_build = (conf.get("package_build","") == "true")
+
+    if (package_build == true) {
+        config_targets = "package"
+    }
+
+    if(conf.get("build_install","") == "true")
+    {
+        config_targets = 'install ' + config_targets
+        setup_args = ' -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install' + setup_args
+    } else{
+        setup_args = ' -DBUILD_DEV=On' + setup_args
+    }
+
+    if(build_type_debug){
+        setup_args = " -DCMAKE_BUILD_TYPE=debug -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}'" + setup_args
+    }else{
+        setup_args = " -DCMAKE_BUILD_TYPE=release" + setup_args
+    }
+    if(env.CK_CCACHE)
+    {
+        setup_args = " -DCMAKE_CXX_COMPILER_LAUNCHER='ccache' -DCMAKE_C_COMPILER_LAUNCHER='ccache' " + setup_args
+    }
+    echo "ccache server: ${env.CK_CCACHE}"
+
+    def pre_setup_cmd = """
+            echo \$HSA_ENABLE_SDMA
+            ulimit -c unlimited
+            rm -rf build
+            mkdir build
+            rm -rf install
+            mkdir install
+            cd build
+        """
+    def setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args}   .. ")
+    // reduce parallelism when compiling, clang uses too much memory
+    def build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make  -j\$(( \$(nproc) / 2 )) ${config_targets}")
+    def execute_cmd = conf.get("execute_cmd", "")
+
+    def cmd = conf.get("cmd", """
+            ${pre_setup_cmd}
+            ${setup_cmd}
+            ${build_cmd}
+            ${execute_cmd}
+        """)
+
+    echo cmd
+    sh cmd
+
+    // Only archive from master or develop
+    if (package_build == true && (env.BRANCH_NAME == "develop" || env.BRANCH_NAME == "master")) {
+        archiveArtifacts artifacts: "build/*.deb", allowEmptyArchive: true, fingerprint: true
+    }
+}
+
+def buildHipClangJob(Map conf=[:]){
+        show_node_info()
+
+        env.HSA_ENABLE_SDMA=0
+        checkout scm
+
+        def image = getDockerImageName() 
+        def prefixpath = conf.get("prefixpath", "/opt/rocm")
+
+        // Jenkins is complaining about the render group 
+        def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+        if (conf.get("enforce_xnack_on", false)) {
+            dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
+        }
+        def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
+        if (params.COMPILER_VERSION != "release"){
+            dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
+        }
+
+        def variant = env.STAGE_NAME
+
+        def retimage
+        (retimage, image) = getDockerImage(conf)
+
+        gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
+            withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
+                timeout(time: 5, unit: 'HOURS')
+                {
+                    cmake_build(conf)
+                }
+            }
+        }
+        return retimage
+}
+
+def reboot(){
+    build job: 'reboot-slaves', propagate: false , parameters: [string(name: 'server', value: "${env.NODE_NAME}"),]
+}
+
+def buildHipClangJobAndReboot(Map conf=[:]){
+    try{
+        buildHipClangJob(conf)
+    }
+    catch(e){
+        echo "throwing error exception for the stage"
+        echo 'Exception occurred: ' + e.toString()
+        throw e
+    }
+    finally{
+        if (!conf.get("no_reboot", false)) {
+            reboot()
+        }
+    }
+}
+
+def runCKProfiler(Map conf=[:]){
+        show_node_info()
+
+        env.HSA_ENABLE_SDMA=0
+        checkout scm
+
+        def image = getDockerImageName()
+        def prefixpath = conf.get("prefixpath", "/opt/rocm")
+
+        // Jenkins is complaining about the render group 
+        def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+        if (conf.get("enforce_xnack_on", false)) {
+            dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
+        }
+        def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
+        if (params.COMPILER_VERSION != "release"){
+            dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
+        }
+
+        def variant = env.STAGE_NAME
+        def retimage
+
+        gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
+            try {
+                (retimage, image) = getDockerImage(conf)
+                withDockerContainer(image: image, args: dockerOpts) {
+                    timeout(time: 5, unit: 'MINUTES'){
+                        sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
+                        if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
+                            throw new Exception ("GPU not found")
+                        }
+                        else{
+                            echo "GPU is OK"
+                        }
+                    }
+                }
+            }
+            catch (org.jenkinsci.plugins.workflow.steps.FlowInterruptedException e){
+                echo "The job was cancelled or aborted"
+                throw e
+            }
+            catch(Exception ex) {
+                retimage = docker.build("${image}", dockerArgs + " --no-cache .")
+                withDockerContainer(image: image, args: dockerOpts) {
+                    timeout(time: 5, unit: 'MINUTES'){
+                        sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
+                        if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
+                            throw new Exception ("GPU not found")
+                        }
+                        else{
+                            echo "GPU is OK"
+                        }
+                    }
+                }
+            }
+
+            withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
+                timeout(time: 24, unit: 'HOURS')
+                {
+                    //cmake_build(conf)
+                    //instead of building, just unstash the ckProfiler and install it
+                    sh """
+                        rm -rf build
+                        mkdir build
+                    """
+                    dir("build"){
+                        unstash 'ckProfiler.tar.gz'
+                        sh 'tar -xvf ckProfiler.tar.gz'
+                    }
+
+					dir("script"){
+                        if (params.RUN_FULL_QA){
+                            sh "./run_full_performance_tests.sh 1 QA_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME}"
+                            archiveArtifacts "perf_gemm.log"
+                            archiveArtifacts "perf_resnet50_N256.log"
+                            archiveArtifacts "perf_resnet50_N4.log"
+                            archiveArtifacts "perf_batched_gemm.log"
+                            archiveArtifacts "perf_grouped_gemm.log"
+                            archiveArtifacts "perf_conv_fwd.log"
+                            archiveArtifacts "perf_conv_bwd_data.log"
+                            archiveArtifacts "perf_gemm_bilinear.log"
+                            archiveArtifacts "perf_reduction.log"
+                            archiveArtifacts "perf_splitK_gemm_verify.log"
+                            archiveArtifacts "perf_splitK_gemm.log"
+                            archiveArtifacts "perf_onnx_gemm.log"
+                           // stash perf files to master
+                            stash name: "perf_gemm.log"
+                            stash name: "perf_resnet50_N256.log"
+                            stash name: "perf_resnet50_N4.log"
+                            stash name: "perf_batched_gemm.log"
+                            stash name: "perf_grouped_gemm.log"
+                            stash name: "perf_conv_fwd.log"
+                            stash name: "perf_conv_bwd_data.log"
+                            stash name: "perf_gemm_bilinear.log"
+                            stash name: "perf_reduction.log"
+                            stash name: "perf_splitK_gemm.log"
+                            stash name: "perf_onnx_gemm.log"
+                            //we will process results on the master node
+                        }
+                        else{
+                            sh "./run_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME}"
+                            archiveArtifacts "perf_gemm.log"
+                            archiveArtifacts "perf_resnet50_N256.log"
+                            archiveArtifacts "perf_resnet50_N4.log"
+                            // stash perf files to master
+                            stash name: "perf_gemm.log"
+                            stash name: "perf_resnet50_N256.log"
+                            stash name: "perf_resnet50_N4.log"
+                            //we will process the results on the master node
+                        }
+					}
+                }
+            }
+        }
+        return retimage
+}
+
+def runPerfTest(Map conf=[:]){
+    try{
+        runCKProfiler(conf)
+    }
+    catch(e){
+        echo "throwing error exception in performance tests"
+        echo 'Exception occurred: ' + e.toString()
+        throw e
+    }
+    finally{
+        if (!conf.get("no_reboot", false)) {
+            reboot()
+        }
+    }
+}
+
+def Build_CK(Map conf=[:]){
+        show_node_info()
+
+        env.HSA_ENABLE_SDMA=0
+        checkout scm
+
+        def image = getDockerImageName() 
+        def prefixpath = conf.get("prefixpath", "/opt/rocm")
+
+        // Jenkins is complaining about the render group 
+        def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+        if (conf.get("enforce_xnack_on", false)) {
+            dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
+        }
+        def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
+        if (params.COMPILER_VERSION != "release"){
+            dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
+        }
+
+        def variant = env.STAGE_NAME
+        def retimage
+
+        gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
+            try {
+                (retimage, image) = getDockerImage(conf)
+                withDockerContainer(image: image, args: dockerOpts) {
+                    timeout(time: 5, unit: 'MINUTES'){
+                        sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
+                        if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
+                            throw new Exception ("GPU not found")
+                        }
+                        else{
+                            echo "GPU is OK"
+                        }
+                    }
+                }
+            }
+            catch (org.jenkinsci.plugins.workflow.steps.FlowInterruptedException e){
+                echo "The job was cancelled or aborted"
+                throw e
+            }
+            catch(Exception ex) {
+                retimage = docker.build("${image}", dockerArgs + " --no-cache .")
+                withDockerContainer(image: image, args: dockerOpts) {
+                    timeout(time: 5, unit: 'MINUTES'){
+                        sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo |tee clinfo.log'
+                        if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
+                            throw new Exception ("GPU not found")
+                        }
+                        else{
+                            echo "GPU is OK"
+                        }
+                    }
+                }
+            }
+            withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
+                timeout(time: 24, unit: 'HOURS')
+                {
+                    cmake_build(conf)
+                    dir("build"){
+                        //run tests and examples
+                        sh 'make -j check'
+                        //we only need the ckProfiler to run the performance tests, so we pack and stash it
+                        sh 'tar -zcvf ckProfiler.tar.gz bin/ckProfiler'
+                        stash "ckProfiler.tar.gz"
+                    }
+                }
+            }
+        }
+        return retimage
+}
+
+def Build_CK_and_Reboot(Map conf=[:]){
+    try{
+        Build_CK(conf)
+    }
+    catch(e){
+        echo "throwing error exception while building CK"
+        echo 'Exception occurred: ' + e.toString()
+        throw e
+    }
+    finally{
+        if (!conf.get("no_reboot", false)) {
+            reboot()
+        }
+    }
+}
+
+def process_results(Map conf=[:]){
+    env.HSA_ENABLE_SDMA=0
+    checkout scm
+    def image = getDockerImageName() 
+    def prefixpath = "/opt/rocm"
+
+    // Jenkins is complaining about the render group 
+    def dockerOpts="--cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+    if (conf.get("enforce_xnack_on", false)) {
+        dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
+    }
+
+    def variant = env.STAGE_NAME
+    def retimage
+
+    gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
+        try {
+            (retimage, image) = getDockerImage(conf)
+        }
+        catch (org.jenkinsci.plugins.workflow.steps.FlowInterruptedException e){
+            echo "The job was cancelled or aborted"
+            throw e
+        }
+    }
+
+    withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
+        timeout(time: 1, unit: 'HOURS'){
+            try{
+                dir("script"){
+                    if (params.RUN_FULL_QA){
+                        // unstash perf files to master
+                        unstash "perf_gemm.log"
+                        unstash "perf_resnet50_N256.log"
+                        unstash "perf_resnet50_N4.log"
+                        unstash "perf_batched_gemm.log"
+                        unstash "perf_grouped_gemm.log"
+                        unstash "perf_conv_fwd.log"
+                        unstash "perf_conv_bwd_data.log"
+                        unstash "perf_gemm_bilinear.log"
+                        unstash "perf_reduction.log"
+                        unstash "perf_splitK_gemm.log"
+                        unstash "perf_onnx_gemm.log"
+                        sh "./process_qa_data.sh"
+                    }
+                    else{
+                        // unstash perf files to master
+                        unstash "perf_gemm.log"
+                        unstash "perf_resnet50_N256.log"
+                        unstash "perf_resnet50_N4.log"
+                        sh "./process_perf_data.sh"
+                    }
+                }
+            }
+            catch(e){
+                echo "throwing error exception while processing performance test results"
+                echo 'Exception occurred: ' + e.toString()
+                throw e
+            }
+        }
+    }
+}
+
+//launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
+CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;COMPILER_VERSION=release
+                                              0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open''' : ""
+
+pipeline {
+    agent none
+    triggers {
+        parameterizedCron(CRON_SETTINGS)
+    }
+    options {
+        parallelsAlwaysFailFast()
+    }
+    parameters {
+        booleanParam(
+            name: "BUILD_DOCKER",
+            defaultValue: false,
+            description: "Force building docker image (default: false), set to true if docker image needs to be updated.")
+        string(
+            name: 'ROCMVERSION', 
+            defaultValue: '5.3', 
+            description: 'Specify which ROCM version to use: 5.2.3, or 5.3 (default), etc.')
+        string(
+            name: 'COMPILER_VERSION', 
+            defaultValue: 'release', 
+            description: 'Specify which version of compiler to use: ck-9110, release (default), or amd-stg-open.')
+        string(
+            name: 'COMPILER_COMMIT', 
+            defaultValue: '', 
+            description: 'Specify which commit of compiler branch to use: leave empty to use the latest commit (default), or use 8a82e4eb7ba28521ba9a9424a0315a8a16590424 commit of amd-stg-open branch.')
+        string(
+            name: 'BUILD_COMPILER', 
+            defaultValue: 'hipcc', 
+            description: 'Specify whether to build CK with hipcc (default) or with clang.')
+        booleanParam(
+            name: "RUN_FULL_QA",
+            defaultValue: false,
+            description: "Select whether to run small set of performance tests (default) or full QA")
+    }
+    environment{
+        dbuser = "${dbuser}"
+        dbpassword = "${dbpassword}"
+        dbsship = "${dbsship}"
+        dbsshport = "${dbsshport}"
+        dbsshuser = "${dbsshuser}"
+        dbsshpassword = "${dbsshpassword}"
+        status_wrapper_creds = "${status_wrapper_creds}"
+        gerrit_cred="${gerrit_cred}"
+        DOCKER_BUILDKIT = "1"
+    }
+    stages{
+        stage("Build Docker"){
+            //when {
+            //    beforeAgent true
+            //    expression { params.BUILD_DOCKER.toBoolean() }
+            //}
+            parallel{
+                stage('Docker /opt/rocm'){
+                    agent{ label rocmnode("nogpu") }
+                    steps{
+                        buildDocker('/opt/rocm')
+                    }
+                }
+            }
+        }
+        stage("Static checks") {
+            parallel{
+                stage('Clang Format') {
+                    agent{ label rocmnode("nogpu") }
+                    environment{
+                        execute_cmd = "find .. -not -path \'*.git*\' -iname \'*.h\' \
+                                -o -not -path \'*.git*\' -iname \'*.hpp\' \
+                                -o -not -path \'*.git*\' -iname \'*.cpp\' \
+                                -o -iname \'*.h.in\' \
+                                -o -iname \'*.hpp.in\' \
+                                -o -iname \'*.cpp.in\' \
+                                -o -iname \'*.cl\' \
+                                | grep -v 'build/' \
+                                | xargs -n 1 -P 1 -I{} -t sh -c \'clang-format-10 -style=file {} | diff - {}\'"
+                    }
+                    steps{
+                        buildHipClangJobAndReboot(setup_cmd: "", build_cmd: "", execute_cmd: execute_cmd, no_reboot:true)
+                    }
+                }
+            }
+        }
+    
+		stage("Build CK and run Tests")
+        {
+            parallel
+            {
+                stage("Build CK and run Tests")
+                {
+                    agent{ label rocmnode("gfx908 || gfx90a") }
+                    environment{
+                        setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS="-O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" """ : """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS="-O3 " """ }"
+                        execute_args = "${params.COMPILER_VERSION == "ck-9110" ? """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS="-O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ : """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908,gfx90a" -DCMAKE_CXX_FLAGS="-O3" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ }"
+                    }
+                    steps{
+                        Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
+                    }
+                }
+            }
+        }
+
+        stage("Performance Tests")
+        {
+            parallel
+            {
+                stage("Run ckProfiler: gfx908 or gfx90a")
+                {
+                    when {
+                        beforeAgent true
+                        expression { !params.RUN_FULL_QA.toBoolean() }
+                    }
+                    options { retry(2) }
+                    agent{ label rocmnode("gfx908 || gfx90a")}
+                    environment{
+                        setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS=" -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -DBUILD_DEV=On """ : """ -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS=" -O3 " -DBUILD_DEV=On """}"
+                   }
+                    steps{
+                        runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
+                    }
+                }
+                stage("Run ckProfiler: gfx90a")
+                {
+                    when {
+                        beforeAgent true
+                        expression { params.RUN_FULL_QA.toBoolean() }
+                    }
+                    options { retry(2) }
+                    agent{ label rocmnode("gfx90a")}
+                    environment{
+                        setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DGPU_TARGETS="gfx90a" -DCMAKE_CXX_FLAGS=" -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -DBUILD_DEV=On """ : """ -DGPU_TARGETS="gfx90a" -DCMAKE_CXX_FLAGS=" -O3 " -DBUILD_DEV=On """}"
+                    }
+                    steps{
+                        runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
+                    }
+                }
+            }
+        }
+        stage("Process Performance Test Results")
+        {
+            parallel
+            {
+                stage("Process results"){
+                    agent { label 'mici' }
+                    steps{
+                        process_results()
+                    }
+                }
+            }
+        }
+    }
+}
diff --git a/3rdparty/composable_kernel/LICENSE b/3rdparty/composable_kernel/LICENSE
new file mode 100644
index 0000000..2fe9a84
--- /dev/null
+++ b/3rdparty/composable_kernel/LICENSE
@@ -0,0 +1,28 @@
+Copyright (c) 2018-    , Advanced Micro Devices, Inc. (Chao Liu, Jing Zhang)
+Copyright (c) 2019-    , Advanced Micro Devices, Inc. (Letao Qin, Qianfeng Zhang, Liang Huang, Shaojie Wang)
+Copyright (c) 2022-    , Advanced Micro Devices, Inc. (Anthony Chang, Chunyu Lai, Illia Silin, Adam Osewski, Poyen Chen, Jehandad Khan)
+Copyright (c) 2019-2021, Advanced Micro Devices, Inc. (Hanwen Chang)
+Copyright (c) 2019-2020, Advanced Micro Devices, Inc. (Tejash Shah)
+Copyright (c) 2020     , Advanced Micro Devices, Inc. (Xiaoyan Zhou)
+Copyright (c) 2021-2022, Advanced Micro Devices, Inc. (Jianfeng Yan)
+
+SPDX-License-Identifier: MIT
+Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/3rdparty/composable_kernel/README.md b/3rdparty/composable_kernel/README.md
new file mode 100644
index 0000000..129996e
--- /dev/null
+++ b/3rdparty/composable_kernel/README.md
@@ -0,0 +1,96 @@
+# Composable Kernel
+
+## Methodology
+Composable Kernel (CK) library aims to provide a programming model for writing performance critical kernels for machine learning workloads across multiple architectures including GPUs, CPUs, etc, through general purpose kernel languages, like HIP C++.
+
+CK utilizes two concepts to achieve performance portability and code maintainability:
+* A tile-based programming model
+* Algorithm complexity reduction for complex ML operators, using innovative technique we call "Tensor Coordinate Transformation".
+
+![ALT](/doc/image/ck_component.png "CK Components")
+
+## Code Structure
+Current CK library are structured into 4 layers:
+* "Templated Tile Operators" layer
+* "Templated Kernel and Invoker" layer
+* "Instantiated Kernel and Invoker" layer
+* "Client API" layer
+
+![ALT](/doc/image/ck_layer.png "CK Layers")
+
+## Contributors
+The list of developers and contributors is here: [Contributors](/CONTRIBUTORS.md)
+
+## Citation
+If you use CK, please use following citations:
+* CK paper will be freely available on arXiv soon: [Realizing Tensor Operators Using Coordinate Transformations and Tile Based Programming](???)
+* [CITATION.cff](/CITATION.cff)
+
+## License
+CK is released under the MIT license. [License File](/LICENSE)
+
+
+# Build CK
+
+## Build docker image
+```bash
+DOCKER_BUILDKIT=1 docker build -t ck:latest -f Dockerfile .
+```
+
+## Launch docker
+```bash
+docker run                                     \
+-it                                            \
+--privileged                                   \
+--group-add sudo                               \
+-w /root/workspace                             \
+-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace  \
+ck:latest                                      \
+/bin/bash
+```
+
+## Build CK
+```bash
+mkdir build && cd build
+
+# Need to specify target ID, example below is for gfx908 and gfx90a
+cmake                                                                                             \
+-D CMAKE_PREFIX_PATH=/opt/dtk-23.04                                                                    \
+-D CMAKE_CXX_COMPILER=/opt/dtk-23.04/bin/hipcc                                                         \
+-D CMAKE_CXX_FLAGS="-O3"                                                                          \
+-D CMAKE_BUILD_TYPE=Release                                                                       \
+-D GPU_TARGETS="gfx906;gfx926"                                                                    \
+..
+```
+
+### Build examples and tests
+```bash
+ make -j examples tests
+ make test
+```
+
+Instructions for running each individual examples are under [example](/example)
+
+
+## Build ckProfiler
+```bash
+ make -j ckProfiler
+```
+Instructions for running ckProfiler are under [profiler](/profiler)
+
+## Install CK
+```bash
+make install
+```
+
+## Using CK as pre-built kernel library
+Instructions for using CK as a pre-built kernel library are under [client_example](/client_example)
+
+## Caveat
+### Kernel Timing and Verification
+CK's own kernel timer will warn up kernel once, and then run it multiple times
+to get average kernel time. For some kernels that use atomic add, this will cause
+output buffer to be accumulated multiple times, causing verification failure.
+To work around it, do not use CK's own timer and do verification at the same time.
+CK's own timer and verification in each example and ckProfiler can be enabled or
+disabled from command line.
diff --git a/3rdparty/composable_kernel/client_example/01_gemm/CMakeLists.txt b/3rdparty/composable_kernel/client_example/01_gemm/CMakeLists.txt
new file mode 100644
index 0000000..9e74119
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/01_gemm/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_gemm gemm.cpp)
+target_link_libraries(client_gemm PRIVATE composable_kernel::device_operations)
diff --git a/3rdparty/composable_kernel/client_example/01_gemm/gemm.cpp b/3rdparty/composable_kernel/client_example/01_gemm/gemm.cpp
new file mode 100644
index 0000000..668c1bf
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/01_gemm/gemm.cpp
@@ -0,0 +1,227 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+using ADataType = F32;
+using BDataType = F32;
+using CDataType = F32;
+
+using ALayout = Row;
+using BLayout = Row;
+using CLayout = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 7)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideC = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1 to 6: M, N, K, StrideA, StrideB, StrideC\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * f_matrix_space_size(M, N, StrideC, CLayout{}));
+    ADataType *in_data=new ADataType[K*N];
+    for(int i=0;i<K*N;i++){
+        in_data[i]=(ADataType)1;
+    }
+    hipMemcpy(a_device_buf.GetDeviceBuffer(),in_data,sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}),hipMemcpyHostToDevice);
+    hipMemcpy(b_device_buf.GetDeviceBuffer(),in_data,sizeof(BDataType) * f_matrix_space_size(K, N, StrideA, BLayout{}),hipMemcpyHostToDevice);
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGemm<ALayout,
+                                                 BLayout,
+                                                 CLayout,
+                                                 ADataType,
+                                                 BDataType,
+                                                 CDataType,
+                                                 ck::tensor_operation::element_wise::PassThrough,
+                                                 ck::tensor_operation::element_wise::PassThrough,
+                                                 ck::tensor_operation::element_wise::PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        StrideC,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+            std::cout << op_name << "support this problem" << std::endl;
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+           // std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        StrideC,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+            hipMemcpy(in_data,c_device_buf.GetDeviceBuffer(),sizeof(CDataType) * f_matrix_space_size(M, N, StrideA, CLayout{}),hipMemcpyDeviceToHost);
+            for(int i=0;i<100;i++){
+                std::cout<<in_data[i*M+1]<<std::endl;
+            }
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt b/3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
new file mode 100644
index 0000000..b7b724c
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
@@ -0,0 +1,13 @@
+add_custom_target(client_gemm_fastgelu_examples)
+
+add_executable(client_gemm_add_add_fastgelu gemm_add_add_fastgelu.cpp)
+target_link_libraries(client_gemm_add_add_fastgelu PRIVATE composable_kernel::device_operations)
+
+add_executable(client_gemm_add_fastgelu gemm_add_fastgelu.cpp)
+target_link_libraries(client_gemm_add_fastgelu PRIVATE composable_kernel::device_operations)
+
+add_executable(client_gemm_fastgelu gemm_fastgelu.cpp)
+target_link_libraries(client_gemm_fastgelu PRIVATE composable_kernel::device_operations)
+
+add_dependencies(client_gemm_fastgelu_examples client_gemm_add_add_fastgelu client_gemm_add_fastgelu
+                 client_gemm_fastgelu)
diff --git a/3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp b/3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp
new file mode 100644
index 0000000..f88e72b
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp
@@ -0,0 +1,241 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using D1DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using ELayout  = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 9)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA  = std::stoi(argv[4]);
+        StrideB  = std::stoi(argv[5]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideD1 = std::stoi(argv[7]);
+        StrideE  = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 8: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem d1_m_n_device_buf(sizeof(D1DataType) *
+                                      f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddAddFastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp b/3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp
new file mode 100644
index 0000000..512555f
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp
@@ -0,0 +1,233 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddFastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using ELayout  = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 8)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA  = std::stoi(argv[4]);
+        StrideB  = std::stoi(argv[5]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideE  = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD0, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddFastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp b/3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp
new file mode 100644
index 0000000..7237231
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp
@@ -0,0 +1,225 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = FastGelu;
+
+using ADataType = F16;
+using BDataType = F16;
+using EDataType = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideE = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 7)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideE = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 6: M, N, K, StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::FastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        {},
+                                                        StrideE,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        {},
+                                                        StrideE,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/client_example/03_gemm_layernorm/CMakeLists.txt b/3rdparty/composable_kernel/client_example/03_gemm_layernorm/CMakeLists.txt
new file mode 100644
index 0000000..3742e70
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/03_gemm_layernorm/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_gemm_add_add_reduce_normalize gemm_add_add_layernorm.cpp)
+target_link_libraries(client_gemm_add_add_reduce_normalize PRIVATE composable_kernel::device_operations)
diff --git a/3rdparty/composable_kernel/client_example/03_gemm_layernorm/gemm_add_add_layernorm.cpp b/3rdparty/composable_kernel/client_example/03_gemm_layernorm/gemm_add_add_layernorm.cpp
new file mode 100644
index 0000000..6c25940
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/03_gemm_layernorm/gemm_add_add_layernorm.cpp
@@ -0,0 +1,274 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType            = F16;
+using BDataType            = F16;
+using BiasDataType         = F32;
+using CDataType            = F16;
+using D0DataType           = F16;
+using ReduceDataType       = F32;
+using GammaDataType        = F16;
+using BetaDataType         = F16;
+using LayerNormOutDataType = F16;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+template <typename gemm_reduce_op_ptr>
+bool RunDeviceGemmMeanSquareMean(gemm_reduce_op_ptr& p_op,
+                                 const void* p_a,
+                                 const void* p_b,
+                                 const void* p_bias,
+                                 const void* p_d0,
+                                 void* p_c,
+                                 void* p_mean,
+                                 void* p_square_mean,
+                                 int M,
+                                 int N,
+                                 int K,
+                                 int StrideA,
+                                 int StrideB,
+                                 int StrideC,
+                                 int StrideD0,
+                                 bool time_kernel)
+{
+    using PassThrough          = ck::tensor_operation::element_wise::PassThrough;
+    using UnaryDivElementOp    = ck::tensor_operation::element_wise::UnaryDivide;
+    using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
+
+    auto passOp   = PassThrough{};
+    auto squareOp = UnarySquareElementOp{};
+    auto divOp    = UnaryDivElementOp{N};
+
+    auto argument_ptr =
+        p_op->MakeArgumentPointer(p_a,
+                                  p_b,
+                                  p_bias,
+                                  {p_d0},
+                                  p_c,
+                                  {p_mean, p_square_mean},
+                                  M,
+                                  N,
+                                  K,
+                                  StrideA,
+                                  StrideB,
+                                  StrideC,
+                                  {StrideD0},
+                                  {&passOp, &passOp, &passOp}, // functor for a, b, c
+                                  {&passOp},                   // functor for d0
+                                  {&passOp, &squareOp},        // functor for inputs of reduction
+                                  {&divOp, &divOp});           // functor for outputs of reduction
+
+    if(p_op->IsSupportedArgument(argument_ptr.get()))
+    {
+        auto invoker_ptr = p_op->MakeInvokerPointer();
+
+        // If we evaluate running time of gemm_reduce. The output may wrong.
+        // Because we need to initialize the reduction tensor before runing the kernel.
+        // However we run kernel many times for time_kernel = trie without reinitialize the out
+        // of reduction tensor.
+        float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+            std::cout << "Gemm + reduce Perf: " << std::setw(10) << ave_time << " ms" << std::endl;
+
+        return true;
+    }
+
+    return false;
+}
+
+template <typename normalize_op_ptr>
+bool RunDeviceNormalize2D(normalize_op_ptr& p_op,
+                          const void* p_x,
+                          const void* p_mean,
+                          const void* p_square_mean,
+                          const void* p_gamma,
+                          const void* p_beta,
+                          void* p_y,
+                          int M,
+                          int N,
+                          int StrideX,
+                          bool time_kernel)
+{
+    std::array<const void*, 5> input = {p_x, p_mean, p_square_mean, p_gamma, p_beta};
+    std::array<void*, 1> output      = {p_y};
+    auto normalize_functor           = ck::tensor_operation::element_wise::Normalize{};
+
+    std::array<ck::index_t, 2> xyLengths = {M, N};
+    std::array<ck::index_t, 2> xyStrides = {StrideX, 1};
+
+    auto argument_ptr = p_op->MakeArgumentPointer(xyLengths,
+                                                  {xyStrides, {1, 0}, {1, 0}, {0, 1}, {0, 1}},
+                                                  {xyStrides},
+                                                  input,
+                                                  output,
+                                                  ck::tensor_operation::element_wise::Normalize{});
+
+    if(p_op->IsSupportedArgument(argument_ptr.get()))
+    {
+        auto invoker_ptr = p_op->MakeInvokerPointer();
+        float ave_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+            std::cout << "Normalize Perf: " << std::setw(10) << ave_time << " ms" << std::endl;
+
+        return true;
+    }
+
+    return false;
+}
+
+int main()
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = 1024;
+    ck::index_t StrideB  = 1024;
+    ck::index_t StrideC  = 1024;
+    ck::index_t StrideD0 = 1024;
+
+    const auto gemm_reduce_ptrs =
+        ck::tensor_operation::device::instance::get_device_gemm_add_add_mean_squaremean_instances<
+            ADataType,
+            BDataType,
+            CDataType,
+            ALayout,
+            BLayout,
+            CLayout>();
+
+    const auto normalize_ptrs =
+        ck::tensor_operation::device::instance::get_device_normalize_from_mean_meansquare_instances<
+            CDataType,
+            ReduceDataType,
+            ReduceDataType,
+            GammaDataType,
+            BetaDataType,
+            LayerNormOutDataType>();
+
+    std::cout << "found " << gemm_reduce_ptrs.size()
+              << " gemm_reduceMean_reduceSquareMean instances" << std::endl;
+
+    std::cout << "found " << normalize_ptrs.size() << " normalize instances" << std::endl;
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem bias_device_buf(sizeof(BiasDataType) * N);
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * f_matrix_space_size(M, N, StrideC, CLayout{}));
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) *
+                                  f_matrix_space_size(M, N, StrideD0, CLayout{}));
+    SimpleDeviceMem reduceMean_device_buf(sizeof(ReduceDataType) * M);
+    SimpleDeviceMem reduceMeanSquare_device_buf(sizeof(ReduceDataType) * M);
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) * M * N);
+
+    bool b_time_kernel           = true;
+    bool b_only_run_first_kernel = true;
+
+    // layernorm => (1) + (2)
+    // (1). c = gemm(a, b), reduce_mean(c), reduce_square_mean(c)
+    // (2). normalize(c, mean, square_mean, gamma, beta)
+    for(auto& gemm_reduce_ptr : gemm_reduce_ptrs)
+    {
+        // run first available kernel
+        if(RunDeviceGemmMeanSquareMean(gemm_reduce_ptr,
+                                       a_device_buf.GetDeviceBuffer(),
+                                       b_device_buf.GetDeviceBuffer(),
+                                       bias_device_buf.GetDeviceBuffer(),
+                                       d0_device_buf.GetDeviceBuffer(),
+                                       c_device_buf.GetDeviceBuffer(),
+                                       reduceMean_device_buf.GetDeviceBuffer(),
+                                       reduceMeanSquare_device_buf.GetDeviceBuffer(),
+                                       M,
+                                       N,
+                                       K,
+                                       StrideA,
+                                       StrideB,
+                                       StrideC,
+                                       StrideD0,
+                                       b_time_kernel))
+        {
+            if(b_only_run_first_kernel)
+                break;
+        }
+        else
+        {
+            std::cout << gemm_reduce_ptr->GetTypeString() << " does not support this problem"
+                      << std::endl;
+        }
+    }
+
+    for(auto& normalize_ptr : normalize_ptrs)
+    {
+        if(RunDeviceNormalize2D(normalize_ptr,
+                                c_device_buf.GetDeviceBuffer(),
+                                reduceMean_device_buf.GetDeviceBuffer(),
+                                reduceMeanSquare_device_buf.GetDeviceBuffer(),
+                                gamma_device_buf.GetDeviceBuffer(),
+                                beta_device_buf.GetDeviceBuffer(),
+                                layerNorm_device_buf.GetDeviceBuffer(),
+                                M,
+                                N,
+                                StrideC,
+                                b_time_kernel))
+        {
+            if(b_only_run_first_kernel)
+                break;
+        }
+        else
+        {
+            std::cout << normalize_ptr->GetTypeString() << " does not support this problem"
+                      << std::endl;
+        }
+    }
+}
diff --git a/3rdparty/composable_kernel/client_example/04_contraction/CMakeLists.txt b/3rdparty/composable_kernel/client_example/04_contraction/CMakeLists.txt
new file mode 100644
index 0000000..4bc6780
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/04_contraction/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_executable(client_contraction_scale contraction_scale.cpp)
+target_link_libraries(client_contraction_scale PRIVATE composable_kernel::device_operations)
+
+add_executable(client_contraction_bilinear contraction_bilinear.cpp)
+target_link_libraries(client_contraction_bilinear PRIVATE composable_kernel::device_operations)
+
diff --git a/3rdparty/composable_kernel/client_example/04_contraction/contraction_bilinear.cpp b/3rdparty/composable_kernel/client_example/04_contraction/contraction_bilinear.cpp
new file mode 100644
index 0000000..91dead4
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/04_contraction/contraction_bilinear.cpp
@@ -0,0 +1,236 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Bilinear;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F32;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F32;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float alpha = 1.f;
+    float beta  = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 25)
+    {
+        const ck::index_t M0 = std::stoi(argv[1]);
+        const ck::index_t M1 = std::stoi(argv[2]);
+
+        const ck::index_t N0 = std::stoi(argv[3]);
+        const ck::index_t N1 = std::stoi(argv[4]);
+
+        const ck::index_t K0 = std::stoi(argv[5]);
+        const ck::index_t K1 = std::stoi(argv[6]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
+
+        d_ms_ns_lengths = {M0, M1, N0, N1};
+        d_ms_ns_strides = {
+            std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21]), std::stoi(argv[22])};
+
+        alpha = std::stof(argv[23]);
+        beta  = std::stof(argv[24]);
+    }
+    else
+    {
+        printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
+        printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg15 to 18: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
+        printf("arg19 to 22: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg23 to 24: alpha, beta\n");
+        exit(0);
+    }
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
+    SimpleDeviceMem d_device_buf(sizeof(DDataType) *
+                                 f_tensor_space_size(d_ms_ns_lengths, d_ms_ns_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<DDataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Bilinear>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{alpha, beta};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                        b_device_buf.GetDeviceBuffer(),
+                                        std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                        e_device_buf.GetDeviceBuffer(),
+                                        a_ms_ks_lengths,
+                                        a_ms_ks_strides,
+                                        b_ns_ks_lengths,
+                                        b_ns_ks_strides,
+                                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                        e_ms_ns_lengths,
+                                        e_ms_ns_strides,
+                                        a_element_op,
+                                        b_element_op,
+                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop      = std::size_t(2) * M * N * K;
+            std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                    sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/client_example/04_contraction/contraction_scale.cpp b/3rdparty/composable_kernel/client_example/04_contraction/contraction_scale.cpp
new file mode 100644
index 0000000..4e08ee1
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/04_contraction/contraction_scale.cpp
@@ -0,0 +1,222 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_scale.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Scale;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float scale = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 20)
+    {
+        const ck::index_t M0 = std::stoi(argv[1]);
+        const ck::index_t M1 = std::stoi(argv[2]);
+
+        const ck::index_t N0 = std::stoi(argv[3]);
+        const ck::index_t N1 = std::stoi(argv[4]);
+
+        const ck::index_t K0 = std::stoi(argv[5]);
+        const ck::index_t K1 = std::stoi(argv[6]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
+
+        scale = std::stof(argv[19]);
+    }
+    else
+    {
+        printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
+        printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg15 to 18: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg19: scale\n");
+        exit(0);
+    }
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Scale>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{scale};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        std::array<const void*, 0>{},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        a_ms_ks_lengths,
+                                                        a_ms_ks_strides,
+                                                        b_ns_ks_lengths,
+                                                        b_ns_ks_strides,
+                                                        std::array<std::vector<ck::index_t>, 0>{},
+                                                        std::array<std::vector<ck::index_t>, 0>{},
+                                                        e_ms_ns_lengths,
+                                                        e_ms_ns_strides,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/client_example/05_layernorm/CMakeLists.txt b/3rdparty/composable_kernel/client_example/05_layernorm/CMakeLists.txt
new file mode 100644
index 0000000..b582b48
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/05_layernorm/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_layernorm2d layernorm2d.cpp)
+target_link_libraries(client_layernorm2d PRIVATE composable_kernel::device_operations)
diff --git a/3rdparty/composable_kernel/client_example/05_layernorm/layernorm2d.cpp b/3rdparty/composable_kernel/client_example/05_layernorm/layernorm2d.cpp
new file mode 100644
index 0000000..adb4117
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/05_layernorm/layernorm2d.cpp
@@ -0,0 +1,163 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/normalization.hpp"
+
+using XDataType     = ck::half_t;
+using GammaDataType = ck::half_t;
+using BetaDataType  = ck::half_t;
+using YDataType     = ck::half_t;
+using AccDataType   = float;
+using PassThrough   = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    ck::index_t M      = 1024;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = 1024;
+
+    auto xy_size = (M - 1) * Stride + N;
+
+    SimpleDeviceMem x_device_buf(sizeof(XDataType) * xy_size);
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
+                                                                       GammaDataType,
+                                                                       BetaDataType,
+                                                                       AccDataType,
+                                                                       YDataType,
+                                                                       PassThrough,
+                                                                       Rank,
+                                                                       NumReduceDim>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({M, N},      // lengths
+                                                        {Stride, 1}, // xStrides
+                                                        {0, 1},      // gammaStrides
+                                                        {0, 1},      // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        x_device_buf.GetDeviceBuffer(),
+                                                        gamma_device_buf.GetDeviceBuffer(),
+                                                        beta_device_buf.GetDeviceBuffer(),
+                                                        y_device_buf.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        nullptr,
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte = sizeof(XDataType) * M * N + sizeof(GammaDataType) * N +
+                                   sizeof(BetaDataType) * N + sizeof(YDataType) * M * N;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({M, N},      // lengths
+                                                        {Stride, 1}, // xStrides
+                                                        {1},         // gammaStrides
+                                                        {1},         // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        x_device_buf.GetDeviceBuffer(),
+                                                        gamma_device_buf.GetDeviceBuffer(),
+                                                        beta_device_buf.GetDeviceBuffer(),
+                                                        y_device_buf.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        nullptr,
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/client_example/06_softmax/CMakeLists.txt b/3rdparty/composable_kernel/client_example/06_softmax/CMakeLists.txt
new file mode 100644
index 0000000..b38a0fd
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/06_softmax/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_softmax4d softmax4d.cpp)
+target_link_libraries(client_softmax4d PRIVATE composable_kernel::device_operations)
diff --git a/3rdparty/composable_kernel/client_example/06_softmax/softmax4d.cpp b/3rdparty/composable_kernel/client_example/06_softmax/softmax4d.cpp
new file mode 100644
index 0000000..7745ddf
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/06_softmax/softmax4d.cpp
@@ -0,0 +1,150 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/softmax.hpp"
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 2;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::vector<ck::index_t> in_lengths{2, 8, 128, 1024};
+    std::vector<ck::index_t> in_strides{8 * 128 * 1024, 128 * 1024, 1024, 1};
+    std::vector<ck::index_t> reduce_dims{2, 3};
+
+    ck::index_t num_elements =
+        std::accumulate(in_lengths.begin(), in_lengths.end(), 1, std::multiplies<ck::index_t>());
+
+    AccDataType alpha{2.0f};
+    AccDataType beta{2.0f};
+
+    SimpleDeviceMem in(sizeof(InDataType) * num_elements);
+    SimpleDeviceMem out(sizeof(OutDataType) * num_elements);
+
+    using DeviceOp = ck::tensor_operation::device::
+        DeviceSoftmax<InDataType, AccDataType, OutDataType, PassThrough, PassThrough, Rank>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        if(op_ptr->GetRank() != Rank || op_ptr->GetNumReduceDim() != NumReduceDim)
+        {
+            continue;
+        }
+
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in_lengths,
+                                                        in_strides,
+                                                        reduce_dims,
+                                                        &alpha,
+                                                        &beta,
+                                                        in.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        PassThrough{},
+                                                        PassThrough{});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes = num_elements * sizeof(InDataType) +
+                                    (beta == 0.0f ? 1 : 2) * num_elements * sizeof(OutDataType);
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in_lengths,
+                                                        in_strides,
+                                                        reduce_dims,
+                                                        &alpha,
+                                                        &beta,
+                                                        in.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        PassThrough{},
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/client_example/07_grouped_conv2d_fwd/CMakeLists.txt b/3rdparty/composable_kernel/client_example/07_grouped_conv2d_fwd/CMakeLists.txt
new file mode 100644
index 0000000..ddc8316
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/07_grouped_conv2d_fwd/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_grouped_conv2d_fwd grouped_conv2d_fwd.cpp)
+target_link_libraries(client_grouped_conv2d_fwd PRIVATE composable_kernel::device_operations)
diff --git a/3rdparty/composable_kernel/client_example/07_grouped_conv2d_fwd/grouped_conv2d_fwd.cpp b/3rdparty/composable_kernel/client_example/07_grouped_conv2d_fwd/grouped_conv2d_fwd.cpp
new file mode 100644
index 0000000..c119c78
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/07_grouped_conv2d_fwd/grouped_conv2d_fwd.cpp
@@ -0,0 +1,255 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+// using InDataType  = ck::half_t;
+// using WeiDataType = ck::half_t;
+// using OutDataType = ck::half_t;
+using InDataType  = float;
+using WeiDataType = float;
+using OutDataType = float;
+
+using InLayout    = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout   = ck::tensor_layout::convolution::GKYXC;
+using OutLayout   = ck::tensor_layout::convolution::GNHWK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Ho            = 26;
+static constexpr ck::index_t Wo            = 26;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+void printArray(std::array<ck::index_t, NumDimSpatial + 3> array)
+{
+    for(int i=0;i<NumDimSpatial + 3;i++){
+        std::cout<<array[i]<<", ";
+    }
+    printf("\n");
+}
+int main()
+{
+    std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Hi, Wi, C};
+    std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, Y, X, C};
+    std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Ho, Wo, K};
+    std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 0, 1};
+
+    std::partial_sum(rbegin(in_lengths),
+                     std::prev(rend(in_lengths)),
+                     std::next(rbegin(in_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(wei_lengths),
+                     std::prev(rend(wei_lengths)),
+                     std::next(rbegin(wei_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(out_lengths),
+                     std::prev(rend(out_lengths)),
+                     std::next(rbegin(out_strides)),
+                     std::multiplies<>{});
+    
+    printArray(in_lengths);
+    printArray(in_strides);
+    
+    // transpose GNHWC/GKYXC/GNHWK to GNCHW/GKCYX/GNCHW
+    std::rotate(
+        rbegin(in_lengths), std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 3));
+    std::rotate(
+        rbegin(in_strides), std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 3));
+    std::rotate(
+        rbegin(wei_lengths), std::next(rbegin(wei_lengths)), std::next(rbegin(wei_lengths), 3));
+    std::rotate(
+        rbegin(wei_strides), std::next(rbegin(wei_strides)), std::next(rbegin(wei_strides), 3));
+    std::rotate(
+        rbegin(out_lengths), std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 3));
+    std::rotate(
+        rbegin(out_strides), std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 3));
+    
+    printArray(in_lengths);
+    printArray(in_strides);
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{0, 0};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{0, 0};
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
+
+    InDataType *in_data=new InDataType[G * N * Hi * Wi * C];
+    for(int i=0;i<G * N * Hi * Wi * C;i++){
+        in_data[i]=(InDataType)1;
+    }
+    hipMemcpy(in.GetDeviceBuffer(),in_data,sizeof(InDataType) * G * N * Hi * Wi * C,hipMemcpyHostToDevice);
+    hipMemcpy(wei.GetDeviceBuffer(),in_data,sizeof(WeiDataType) * G * K * Y * X * C,hipMemcpyHostToDevice);
+    
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                                 InLayout,
+                                                                                 WeiLayout,
+                                                                                 ck::Tuple<>,
+                                                                                 OutLayout,
+                                                                                 InDataType,
+                                                                                 WeiDataType,
+                                                                                 ck::Tuple<>,
+                                                                                 OutDataType,
+                                                                                 PassThrough,
+                                                                                 PassThrough,
+                                                                                 PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::cout << op_name << std::endl;
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = std::size_t(2) * G * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = sizeof(InDataType) * G * N * Hi * Wi * C +
+                                    sizeof(WeiDataType) * G * K * Y * X * C +
+                                    sizeof(OutDataType) * G * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            //std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      //<< gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            //std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+            hipMemcpy(in_data,out.GetDeviceBuffer(),sizeof(OutDataType) * G * N * Ho * Wo * K,hipMemcpyDeviceToHost);
+            for(int i=0;i<10;i++){
+                std::cout<<in_data[i*Wo * K]<<std::endl;
+            }
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+}
diff --git a/3rdparty/composable_kernel/client_example/08_fused_attention/CMakeLists.txt b/3rdparty/composable_kernel/client_example/08_fused_attention/CMakeLists.txt
new file mode 100644
index 0000000..5cdea72
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/08_fused_attention/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_fused_attention fused_attention.cpp)
+target_link_libraries(client_fused_attention PRIVATE composable_kernel::device_operations)
diff --git a/3rdparty/composable_kernel/client_example/08_fused_attention/fused_attention.cpp b/3rdparty/composable_kernel/client_example/08_fused_attention/fused_attention.cpp
new file mode 100644
index 0000000..fe927da
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/08_fused_attention/fused_attention.cpp
@@ -0,0 +1,213 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using AElementOp    = ck::tensor_operation::element_wise::PassThrough;
+using B0ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp    = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr static auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+using ADataType   = ck::half_t;
+using B0DataType  = ck::half_t;
+using B1DataType  = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    int G0 = 48;
+    int G1 = 16;
+    int M  = 1024;
+    int N  = 1024;
+    int K  = 64;
+    int O  = 64;
+
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * G0 * G1 * M * K);
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) * G0 * G1 * N * K);
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * G0 * G1 * O * N);
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * G0 * G1 * M * O);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          ADataType,
+                                                                          B0DataType,
+                                                                          B1DataType,
+                                                                          CDataType,
+                                                                          ck::Tuple<>,
+                                                                          ck::Tuple<>,
+                                                                          AElementOp,
+                                                                          B0ElementOp,
+                                                                          Acc0ElementOp,
+                                                                          B1ElementOp,
+                                                                          CElementOp,
+                                                                          MaskingSpec>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b0_device_buf.GetDeviceBuffer(),
+                                                        b1_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        {}, // p_acc0_biases
+                                                        {}, // p_acc1_biases
+                                                        a_gs_ms_ks_lengths,
+                                                        a_gs_ms_ks_strides,
+                                                        b0_gs_ns_ks_lengths,
+                                                        b0_gs_ns_ks_strides,
+                                                        b1_gs_os_ns_lengths,
+                                                        b1_gs_os_ns_strides,
+                                                        c_gs_ms_os_lengths,
+                                                        c_gs_ms_os_strides,
+                                                        {}, // acc0_biases_gs_ms_ns_lengths
+                                                        {}, // acc0_biases_gs_ms_ns_strides
+                                                        {}, // acc1_biases_gs_ms_os_lengths
+                                                        {}, // acc1_biases_gs_ms_os_strides
+                                                        AElementOp{},
+                                                        B0ElementOp{},
+                                                        Acc0ElementOp{1 / sqrtf(K)},
+                                                        B1ElementOp{},
+                                                        CElementOp{});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * G0 * G1;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                                    G0 * G1;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best instance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b0_device_buf.GetDeviceBuffer(),
+                                                        b1_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        {}, // p_acc0_biases
+                                                        {}, // p_acc1_biases
+                                                        a_gs_ms_ks_lengths,
+                                                        a_gs_ms_ks_strides,
+                                                        b0_gs_ns_ks_lengths,
+                                                        b0_gs_ns_ks_strides,
+                                                        b1_gs_os_ns_lengths,
+                                                        b1_gs_os_ns_strides,
+                                                        c_gs_ms_os_lengths,
+                                                        c_gs_ms_os_strides,
+                                                        {}, // acc0_biases_gs_ms_ns_lengths
+                                                        {}, // acc0_biases_gs_ms_ns_strides
+                                                        {}, // acc1_biases_gs_ms_os_lengths
+                                                        {}, // acc1_biases_gs_ms_os_strides
+                                                        AElementOp{},
+                                                        B0ElementOp{},
+                                                        Acc0ElementOp{1 / sqrtf(K)},
+                                                        B1ElementOp{},
+                                                        CElementOp{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/client_example/09_quantization/CMakeLists.txt b/3rdparty/composable_kernel/client_example/09_quantization/CMakeLists.txt
new file mode 100644
index 0000000..7dc9b86
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/09_quantization/CMakeLists.txt
@@ -0,0 +1,11 @@
+add_executable(client_conv2d_fwd_bias_relu_perchannel_quantization conv2d_fwd_bias_relu_perchannel_quantization.cpp)
+target_link_libraries(client_conv2d_fwd_bias_relu_perchannel_quantization PRIVATE composable_kernel::device_operations)
+
+add_executable(client_conv2d_fwd_bias_relu_perlayer_quantization conv2d_fwd_bias_relu_perlayer_quantization.cpp)
+target_link_libraries(client_conv2d_fwd_bias_relu_perlayer_quantization PRIVATE composable_kernel::device_operations)
+
+add_executable(client_conv2d_fwd_perchannel_quantization conv2d_fwd_perchannel_quantization.cpp)
+target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composable_kernel::device_operations)
+
+add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp)
+target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_operations)
diff --git a/3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp b/3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
new file mode 100644
index 0000000..bcb0cef
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
@@ -0,0 +1,205 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using BiasDataType         = int32_t;
+using RequantScaleDataType = float;
+using OutDataType          = int8_t;
+
+using InLayout           = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout          = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout         = ck::tensor_layout::convolution::G_K;
+using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+using OutLayout          = ck::tensor_layout::convolution::GNHWK;
+using PassThrough        = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp       = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 4;
+static constexpr ck::index_t K             = 64;
+static constexpr ck::index_t C             = 32;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 71;
+static constexpr ck::index_t Wi            = 71;
+static constexpr ck::index_t Ho            = 36;
+static constexpr ck::index_t Wo            = 36;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C};
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+        NumDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout, RequantScaleLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<BiasDataType, RequantScaleDataType>,
+        OutDataType,
+        PassThrough,
+        PassThrough,
+        OutElementOp>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                        wei.GetDeviceBuffer(),
+                                        {bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths, requant_scale_lengths},
+                                        {bias_strides, requant_scale_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{ActivationOp{}});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
+                                    G * sizeof(WeiDataType) * K * Y * X * C +
+                                    G * sizeof(OutDataType) * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                        wei.GetDeviceBuffer(),
+                                        {bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths, requant_scale_lengths},
+                                        {bias_strides, requant_scale_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{ActivationOp{}});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp b/3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
new file mode 100644
index 0000000..26c7aa1
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
@@ -0,0 +1,198 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType   = int8_t;
+using WeiDataType  = int8_t;
+using BiasDataType = int32_t;
+using OutDataType  = int8_t;
+
+using InLayout     = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout    = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout   = ck::tensor_layout::convolution::G_K;
+using OutLayout    = ck::tensor_layout::convolution::GNHWK;
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 4;
+static constexpr ck::index_t K             = 64;
+static constexpr ck::index_t C             = 32;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 71;
+static constexpr ck::index_t Wi            = 71;
+static constexpr ck::index_t Ho            = 36;
+static constexpr ck::index_t Wo            = 36;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C};
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                    InLayout,
+                                                                    WeiLayout,
+                                                                    ck::Tuple<BiasLayout>,
+                                                                    OutLayout,
+                                                                    InDataType,
+                                                                    WeiDataType,
+                                                                    ck::Tuple<BiasDataType>,
+                                                                    OutDataType,
+                                                                    PassThrough,
+                                                                    PassThrough,
+                                                                    OutElementOp>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {bias.GetDeviceBuffer()},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {bias_lengths},
+                                                        {bias_strides},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{0.5f, ActivationOp{}});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
+                                    G * sizeof(WeiDataType) * K * Y * X * C +
+                                    G * sizeof(OutDataType) * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {bias.GetDeviceBuffer()},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {bias_lengths},
+                                                        {bias_strides},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{0.5f, ActivationOp{}});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp b/3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
new file mode 100644
index 0000000..475b2f0
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
@@ -0,0 +1,198 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using RequantScaleDataType = float;
+using OutDataType          = int8_t;
+
+using InLayout           = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout          = ck::tensor_layout::convolution::GKYXC;
+using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+using OutLayout          = ck::tensor_layout::convolution::GNHWK;
+using PassThrough        = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp       = PassThrough;
+using OutElementOp       = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 4;
+static constexpr ck::index_t K             = 64;
+static constexpr ck::index_t C             = 32;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 71;
+static constexpr ck::index_t Wi            = 71;
+static constexpr ck::index_t Ho            = 36;
+static constexpr ck::index_t Wo            = 36;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C};
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                    InLayout,
+                                                                    WeiLayout,
+                                                                    ck::Tuple<RequantScaleLayout>,
+                                                                    OutLayout,
+                                                                    InDataType,
+                                                                    WeiDataType,
+                                                                    ck::Tuple<RequantScaleDataType>,
+                                                                    OutDataType,
+                                                                    PassThrough,
+                                                                    PassThrough,
+                                                                    OutElementOp>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {requant_scale.GetDeviceBuffer()},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {requant_scale_lengths},
+                                                        {requant_scale_strides},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{ActivationOp{}});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
+                                    G * sizeof(WeiDataType) * K * Y * X * C +
+                                    G * sizeof(OutDataType) * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{ActivationOp{}});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp b/3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
new file mode 100644
index 0000000..da7b7e6
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
@@ -0,0 +1,192 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = int8_t;
+using WeiDataType = int8_t;
+using OutDataType = int8_t;
+
+using InLayout     = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout    = ck::tensor_layout::convolution::GKYXC;
+using OutLayout    = ck::tensor_layout::convolution::GNHWK;
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 4;
+static constexpr ck::index_t K             = 64;
+static constexpr ck::index_t C             = 32;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 71;
+static constexpr ck::index_t Wi            = 71;
+static constexpr ck::index_t Ho            = 36;
+static constexpr ck::index_t Wo            = 36;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C};
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                                 InLayout,
+                                                                                 WeiLayout,
+                                                                                 ck::Tuple<>,
+                                                                                 OutLayout,
+                                                                                 InDataType,
+                                                                                 WeiDataType,
+                                                                                 ck::Tuple<>,
+                                                                                 OutDataType,
+                                                                                 PassThrough,
+                                                                                 PassThrough,
+                                                                                 OutElementOp>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{0.5f, ActivationOp{}});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
+                                    G * sizeof(WeiDataType) * K * Y * X * C +
+                                    G * sizeof(OutDataType) * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{0.5f, ActivationOp{}});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/client_example/10_grouped_conv2d_bwd_data/CMakeLists.txt b/3rdparty/composable_kernel/client_example/10_grouped_conv2d_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000..e564f31
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/10_grouped_conv2d_bwd_data/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_grouped_conv2d_bwd_data grouped_conv2d_bwd_data.cpp)
+target_link_libraries(client_grouped_conv2d_bwd_data PRIVATE composable_kernel::device_operations)
diff --git a/3rdparty/composable_kernel/client_example/10_grouped_conv2d_bwd_data/grouped_conv2d_bwd_data.cpp b/3rdparty/composable_kernel/client_example/10_grouped_conv2d_bwd_data/grouped_conv2d_bwd_data.cpp
new file mode 100644
index 0000000..55c7898
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/10_grouped_conv2d_bwd_data/grouped_conv2d_bwd_data.cpp
@@ -0,0 +1,226 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout    = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout   = ck::tensor_layout::convolution::GKYXC;
+using OutLayout   = ck::tensor_layout::convolution::GNHWK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Hi, Wi, C};
+    std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, Y, X, C};
+    std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Ho, Wo, K};
+    std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 0, 1};
+
+    std::partial_sum(rbegin(in_lengths),
+                     std::prev(rend(in_lengths)),
+                     std::next(rbegin(in_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(wei_lengths),
+                     std::prev(rend(wei_lengths)),
+                     std::next(rbegin(wei_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(out_lengths),
+                     std::prev(rend(out_lengths)),
+                     std::next(rbegin(out_strides)),
+                     std::multiplies<>{});
+
+    // transpose GNHWC/GKYXC/GNHWK to GNCHW/GKCYX/GNCHW
+    std::rotate(
+        rbegin(in_lengths), std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 3));
+    std::rotate(
+        rbegin(in_strides), std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 3));
+    std::rotate(
+        rbegin(wei_lengths), std::next(rbegin(wei_lengths)), std::next(rbegin(wei_lengths), 3));
+    std::rotate(
+        rbegin(wei_strides), std::next(rbegin(wei_strides)), std::next(rbegin(wei_strides), 3));
+    std::rotate(
+        rbegin(out_lengths), std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 3));
+    std::rotate(
+        rbegin(out_strides), std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 3));
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
+                                                                                     OutLayout,
+                                                                                     WeiLayout,
+                                                                                     ck::Tuple<>,
+                                                                                     InLayout,
+                                                                                     OutDataType,
+                                                                                     WeiDataType,
+                                                                                     ck::Tuple<>,
+                                                                                     InDataType,
+                                                                                     PassThrough,
+                                                                                     PassThrough,
+                                                                                     PassThrough>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        in.GetDeviceBuffer(),
+                                                        out_lengths,
+                                                        out_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        in_lengths,
+                                                        in_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = std::size_t(2) * G * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = sizeof(InDataType) * G * N * Hi * Wi * C +
+                                    sizeof(WeiDataType) * G * K * Y * X * C +
+                                    sizeof(OutDataType) * G * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        in.GetDeviceBuffer(),
+                                                        out_lengths,
+                                                        out_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        in_lengths,
+                                                        in_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+}
diff --git a/3rdparty/composable_kernel/client_example/11_grouped_conv_bwd_weight/CMakeLists.txt b/3rdparty/composable_kernel/client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
new file mode 100644
index 0000000..3e3f667
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_grouped_conv2d_bwd_weight grouped_conv2d_bwd_weight.cpp)
+target_link_libraries(client_grouped_conv2d_bwd_weight PRIVATE composable_kernel::device_operations)
diff --git a/3rdparty/composable_kernel/client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight.cpp b/3rdparty/composable_kernel/client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight.cpp
new file mode 100644
index 0000000..1ecc856
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight.cpp
@@ -0,0 +1,190 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout    = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout   = ck::tensor_layout::convolution::GKYXC;
+using OutLayout   = ck::tensor_layout::convolution::GNHWK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    std::array<ck::index_t, NumDimSpatial> input_spatial_lengths{Hi, Wi};
+    std::array<ck::index_t, NumDimSpatial> filter_spatial_lengths{Y, X};
+    std::array<ck::index_t, NumDimSpatial> output_spatial_lengths{Ho, Wo};
+
+    std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1};
+    std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+    ck::index_t split_k = 2;
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NumDimSpatial,
+                                                                              InLayout,
+                                                                              WeiLayout,
+                                                                              OutLayout,
+                                                                              InDataType,
+                                                                              WeiDataType,
+                                                                              OutDataType,
+                                                                              PassThrough,
+                                                                              PassThrough,
+                                                                              PassThrough>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        G,
+                                                        N,
+                                                        K,
+                                                        C,
+                                                        input_spatial_lengths,
+                                                        filter_spatial_lengths,
+                                                        output_spatial_lengths,
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        split_k);
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = std::size_t(2) * G * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = sizeof(InDataType) * G * N * Hi * Wi * C +
+                                    sizeof(WeiDataType) * G * K * Y * X * C +
+                                    sizeof(OutDataType) * G * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        G,
+                                                        N,
+                                                        K,
+                                                        C,
+                                                        input_spatial_lengths,
+                                                        filter_spatial_lengths,
+                                                        output_spatial_lengths,
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        split_k);
+        auto invoker_ptr  = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+}
diff --git a/3rdparty/composable_kernel/client_example/12_elementwise_normalization/CMakeLists.txt b/3rdparty/composable_kernel/client_example/12_elementwise_normalization/CMakeLists.txt
new file mode 100644
index 0000000..1ba0e12
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/12_elementwise_normalization/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_elementwise_layernorm2d elementwise_layernorm2d.cpp)
+target_link_libraries(client_elementwise_layernorm2d PRIVATE composable_kernel::device_operations)
diff --git a/3rdparty/composable_kernel/client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp b/3rdparty/composable_kernel/client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp
new file mode 100644
index 0000000..8cf46ed
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp
@@ -0,0 +1,175 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp"
+
+using ADataType             = ck::half_t; // Input 1
+using BDataType             = ck::half_t; // Input 2
+using XDataType             = ck::half_t;
+using GammaDataType         = ck::half_t;
+using BetaDataType          = ck::half_t;
+using YDataType             = ck::half_t;
+using AccDataType           = float;
+using XElementwiseOperation = ck::tensor_operation::element_wise::Add;
+using YElementwiseOperation = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    bool time_kernel = true;
+
+    ck::index_t M      = 48 * 256;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = N;
+
+    auto mn_size = (M - 1) * Stride + N;
+
+    SimpleDeviceMem a_dev_buf(sizeof(ADataType) * mn_size);
+    SimpleDeviceMem b_dev_buf(sizeof(BDataType) * mn_size);
+    SimpleDeviceMem gamma_dev_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_dev_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem y_dev_buf(sizeof(YDataType) * mn_size);
+
+    std::array<const void*, 2> ab_input               = {a_dev_buf.GetDeviceBuffer(),
+                                           b_dev_buf.GetDeviceBuffer()};
+    std::vector<ck::index_t> abStride                 = {Stride, 1};
+    std::array<std::vector<ck::index_t>, 2> abStrides = {abStride, abStride};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwiseNormalization<
+        ck::Tuple<ADataType, BDataType>,
+        GammaDataType,
+        BetaDataType,
+        AccDataType,
+        YDataType,
+        XElementwiseOperation,
+        YElementwiseOperation,
+        Rank,
+        NumReduceDim>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({M, N}, // lengths
+                                                        abStrides,
+                                                        {0, 1},      // gammaStrides
+                                                        {0, 1},      // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        ab_input,
+                                                        gamma_dev_buf.GetDeviceBuffer(),
+                                                        beta_dev_buf.GetDeviceBuffer(),
+                                                        y_dev_buf.GetDeviceBuffer(),
+                                                        XElementwiseOperation{},
+                                                        YElementwiseOperation{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte = sizeof(ADataType) * M * N + sizeof(BDataType) * M * N +
+                                   sizeof(GammaDataType) * N + sizeof(BetaDataType) * N +
+                                   sizeof(YDataType) * M * N;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({M, N}, // lengths
+                                                        abStrides,
+                                                        {1},         // gammaStrides
+                                                        {1},         // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        ab_input,
+                                                        gamma_dev_buf.GetDeviceBuffer(),
+                                                        beta_dev_buf.GetDeviceBuffer(),
+                                                        y_dev_buf.GetDeviceBuffer(),
+                                                        XElementwiseOperation{},
+                                                        YElementwiseOperation{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/client_example/13_batchnorm/CMakeLists.txt b/3rdparty/composable_kernel/client_example/13_batchnorm/CMakeLists.txt
new file mode 100644
index 0000000..5466967
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/13_batchnorm/CMakeLists.txt
@@ -0,0 +1,4 @@
+add_executable(client_batchnorm_fwd_nhwc batchnorm_fwd_nhwc.cpp)
+add_executable(client_batchnorm_bwd_nhwc batchnorm_bwd_nhwc.cpp)
+target_link_libraries(client_batchnorm_fwd_nhwc PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_batchnorm_bwd_nhwc PRIVATE composable_kernel::device_operations)
diff --git a/3rdparty/composable_kernel/client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp b/3rdparty/composable_kernel/client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp
new file mode 100644
index 0000000..8ef2198
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp
@@ -0,0 +1,201 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp"
+
+using XDataType           = ck::half_t;
+using DxDataType          = float;
+using DyDataType          = float;
+using AccDataType         = float;
+using ScaleDataType       = ck::half_t;
+using DscaleDbiasDataType = float;
+using MeanVarDataType     = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+const double epsilon = std::numeric_limits<float>::epsilon();
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem dy(sizeof(DyDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem dx(sizeof(DxDataType) * numXYElement);
+    SimpleDeviceMem dscale(sizeof(DscaleDbiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem dbias(sizeof(DscaleDbiasDataType) * numScaleBiasMeanVarElement);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
+                                                                      DxDataType,
+                                                                      DyDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      DscaleDbiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThrough,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        dy.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        dx.GetDeviceBuffer(),
+                                                        dscale.GetDeviceBuffer(),
+                                                        dbias.GetDeviceBuffer());
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                numXYElement * (sizeof(XDataType) + sizeof(DyDataType) + sizeof(DxDataType)) +
+                numScaleBiasMeanVarElement *
+                    (sizeof(ScaleDataType) + sizeof(DscaleDbiasDataType) * 2 +
+                     sizeof(MeanVarDataType) * 2);
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        // run the best intance
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        dy.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        dx.GetDeviceBuffer(),
+                                                        dscale.GetDeviceBuffer(),
+                                                        dbias.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp b/3rdparty/composable_kernel/client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp
new file mode 100644
index 0000000..322667a
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp
@@ -0,0 +1,197 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp"
+
+using XDataType       = float;
+using YDataType       = float;
+using AccDataType     = float;
+using ScaleDataType   = AccDataType;
+using BiasDataType    = AccDataType;
+using MeanVarDataType = AccDataType;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+const double epsilon       = std::numeric_limits<float>::epsilon();
+const double averageFactor = 0.1;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem y(sizeof(YDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                                      YDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      BiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThrough,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        bias.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        y.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        averageFactor,
+                                                        nullptr,
+                                                        nullptr);
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                numXYElement * (sizeof(XDataType) + sizeof(YDataType)) +
+                numScaleBiasMeanVarElement * (sizeof(ScaleDataType) + sizeof(BiasDataType) +
+                                              sizeof(MeanVarDataType) + sizeof(MeanVarDataType));
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        // run the best intance
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        bias.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        y.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        averageFactor,
+                                                        nullptr,
+                                                        nullptr);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/client_example/14_instance_id/CMakeLists.txt b/3rdparty/composable_kernel/client_example/14_instance_id/CMakeLists.txt
new file mode 100644
index 0000000..87b2a9a
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/14_instance_id/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_batchnorm_fwd_instance_id batchnorm_fwd_instance_id.cpp)
+target_link_libraries(client_batchnorm_fwd_instance_id PRIVATE composable_kernel::device_operations)
diff --git a/3rdparty/composable_kernel/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp b/3rdparty/composable_kernel/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp
new file mode 100644
index 0000000..9cfeee1
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp
@@ -0,0 +1,206 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp"
+
+using XDataType       = float;
+using YDataType       = float;
+using AccDataType     = float;
+using ScaleDataType   = AccDataType;
+using BiasDataType    = AccDataType;
+using MeanVarDataType = AccDataType;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+const double epsilon       = std::numeric_limits<float>::epsilon();
+const double averageFactor = 0.1;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+// In the actual application, the instance index and name are usually from the perf db
+static int instance_index = -1;
+static std::string instance_name;
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem y(sizeof(YDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                                      YDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      BiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThrough,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    bool found          = false;
+    int best_op_index   = -1;
+    float best_ave_time = std::numeric_limits<float>::max();
+
+    // profile device operation instances and save the best performant instance index and instance
+    // name
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        bias.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        y.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        averageFactor,
+                                                        nullptr,
+                                                        nullptr);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            if(ave_time < best_ave_time)
+            {
+                found         = true;
+                best_op_index = i;
+                best_ave_time = ave_time;
+            }
+        }
+    }
+
+    if(found)
+    {
+        instance_index = best_op_index;
+        instance_name  = op_ptrs[instance_index]->GetTypeIdHashCode();
+    };
+
+    // simulate the execution of the operation when the instance index and name are available
+    const auto op_ptrs_2 = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    if(instance_index >= 0 && instance_index < op_ptrs_2.size())
+    {
+        auto& op_ptr = op_ptrs_2[instance_index];
+
+        if(op_ptr->GetTypeIdHashCode() == instance_name)
+        {
+
+            auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                            xyStrides,
+                                                            xyStrides,
+                                                            reduceDims,
+                                                            scaleBiasMeanVarLengths,
+                                                            scaleBiasMeanVarStrides,
+                                                            scaleBiasMeanVarStrides,
+                                                            scaleBiasMeanVarStrides,
+                                                            x.GetDeviceBuffer(),
+                                                            scale.GetDeviceBuffer(),
+                                                            bias.GetDeviceBuffer(),
+                                                            epsilon,
+                                                            PassThrough{},
+                                                            y.GetDeviceBuffer(),
+                                                            mean.GetDeviceBuffer(),
+                                                            invVariance.GetDeviceBuffer(),
+                                                            averageFactor,
+                                                            nullptr,
+                                                            nullptr);
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+                SimpleDeviceMem workspace(workspace_sz);
+
+                op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+                float exec_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+                size_t num_bytes = numXYElement * (sizeof(XDataType) + sizeof(YDataType)) +
+                                   numScaleBiasMeanVarElement *
+                                       (sizeof(ScaleDataType) + sizeof(BiasDataType) +
+                                        sizeof(MeanVarDataType) + sizeof(MeanVarDataType));
+
+                float gb_per_sec = num_bytes / 1.E6 / exec_time;
+
+                std::cout << "Kernel execution time: " << std::setw(10) << exec_time
+                          << " ms,  effective data transfer bandwidth: " << gb_per_sec << " GB/s"
+                          << std::endl;
+            }
+        };
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/client_example/CMakeLists.txt b/3rdparty/composable_kernel/client_example/CMakeLists.txt
new file mode 100644
index 0000000..a5993d0
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/CMakeLists.txt
@@ -0,0 +1,15 @@
+cmake_minimum_required(VERSION 3.15)
+project(ck_app)
+add_compile_options(-std=c++17)
+
+find_package(composable_kernel 1.0.0 COMPONENTS device_operations)
+find_package(hip REQUIRED PATHS $ENV{ROCM_PATH})
+message(STATUS "Build with HIP ${hip_VERSION}")
+
+# add all example subdir
+file(GLOB dir_list LIST_DIRECTORIES true *)
+FOREACH(subdir ${dir_list})
+    IF(IS_DIRECTORY "${subdir}" AND (NOT "${subdir}" MATCHES "build"))
+        add_subdirectory(${subdir})
+    ENDIF()
+ENDFOREACH()
diff --git a/3rdparty/composable_kernel/client_example/README.md b/3rdparty/composable_kernel/client_example/README.md
new file mode 100644
index 0000000..5c84f1d
--- /dev/null
+++ b/3rdparty/composable_kernel/client_example/README.md
@@ -0,0 +1,18 @@
+##
+Client application links to CK library, and therefore CK library needs to be installed before building client applications.
+
+
+## Build
+```bash
+mkdir -p client_example/build
+cd client_example/build
+```
+
+```bash
+cmake -D CMAKE_CXX_COMPILER=${ROCM_PATH}/bin/hipcc  -D CMAKE_PREFIX_PATH="${ROCM_PATH};${PATH_TO_CK_INSTALL_DIRECTORY}" ..
+```
+
+### Build client example
+```bash
+ make -j 
+```
diff --git a/3rdparty/composable_kernel/cmake/Analyzers.cmake b/3rdparty/composable_kernel/cmake/Analyzers.cmake
new file mode 100644
index 0000000..1bf1a52
--- /dev/null
+++ b/3rdparty/composable_kernel/cmake/Analyzers.cmake
@@ -0,0 +1,34 @@
+################################################################################
+# 
+# MIT License
+# 
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+# 
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+# 
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+# 
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+# 
+################################################################################
+
+if(NOT TARGET analyze)
+    add_custom_target(analyze)
+endif()
+
+function(mark_as_analyzer)
+    add_dependencies(analyze ${ARGN})
+endfunction()
+
diff --git a/3rdparty/composable_kernel/cmake/ClangTidy.cmake b/3rdparty/composable_kernel/cmake/ClangTidy.cmake
new file mode 100644
index 0000000..654d4b6
--- /dev/null
+++ b/3rdparty/composable_kernel/cmake/ClangTidy.cmake
@@ -0,0 +1,162 @@
+################################################################################
+#
+# MIT License
+#
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+#
+################################################################################
+include(CMakeParseArguments)
+include(Analyzers)
+
+get_filename_component(CLANG_TIDY_EXE_HINT "${CMAKE_CXX_COMPILER}" PATH)
+set(ROCM_PATH $ENV{ROCM_PATH})
+find_program(CLANG_TIDY_EXE
+    NAMES
+        clang-tidy
+        clang-tidy-5.0
+        clang-tidy-4.0
+        clang-tidy-3.9
+        clang-tidy-3.8
+        clang-tidy-3.7
+        clang-tidy-3.6
+        clang-tidy-3.5
+    HINTS
+        ${CLANG_TIDY_EXE_HINT}
+    PATH_SUFFIXES
+        compiler/bin
+    PATHS
+        ${ROCM_PATH}/llvm/bin
+        ${ROCM_PATH}/hcc
+        /usr/local/opt/llvm/bin
+)
+
+function(find_clang_tidy_version VAR)
+    execute_process(COMMAND ${CLANG_TIDY_EXE} -version OUTPUT_VARIABLE VERSION_OUTPUT)
+    separate_arguments(VERSION_OUTPUT_LIST UNIX_COMMAND "${VERSION_OUTPUT}")
+    list(FIND VERSION_OUTPUT_LIST "version" VERSION_INDEX)
+    if(VERSION_INDEX GREATER 0)
+        math(EXPR VERSION_INDEX "${VERSION_INDEX} + 1")
+        list(GET VERSION_OUTPUT_LIST ${VERSION_INDEX} VERSION)
+        set(${VAR} ${VERSION} PARENT_SCOPE)
+    else()
+        set(${VAR} "0.0" PARENT_SCOPE)
+    endif()
+
+endfunction()
+
+if( NOT CLANG_TIDY_EXE )
+    message( STATUS "Clang tidy not found" )
+    set(CLANG_TIDY_VERSION "0.0")
+else()
+    find_clang_tidy_version(CLANG_TIDY_VERSION)
+    message( STATUS "Clang tidy found: ${CLANG_TIDY_VERSION}")
+endif()
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+set(CLANG_TIDY_FIXIT_DIR ${CMAKE_BINARY_DIR}/fixits)
+file(MAKE_DIRECTORY ${CLANG_TIDY_FIXIT_DIR})
+set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES ${CLANG_TIDY_FIXIT_DIR})
+
+macro(enable_clang_tidy)
+    set(options ANALYZE_TEMPORARY_DTORS ALL)
+    set(oneValueArgs HEADER_FILTER)
+    set(multiValueArgs CHECKS ERRORS EXTRA_ARGS)
+
+    cmake_parse_arguments(PARSE "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    string(REPLACE ";" "," CLANG_TIDY_CHECKS "${PARSE_CHECKS}")
+    string(REPLACE ";" "," CLANG_TIDY_ERRORS "${PARSE_ERRORS}")
+    set(CLANG_TIDY_EXTRA_ARGS)
+    foreach(ARG ${PARSE_EXTRA_ARGS})
+        list(APPEND CLANG_TIDY_EXTRA_ARGS "-extra-arg=${ARG}")
+    endforeach()
+
+    set(CLANG_TIDY_ALL)
+    if(PARSE_ALL)
+        set(CLANG_TIDY_ALL ALL)
+    endif()
+
+    message(STATUS "Clang tidy checks: ${CLANG_TIDY_CHECKS}")
+
+    if (${PARSE_ANALYZE_TEMPORARY_DTORS})
+        set(CLANG_TIDY_ANALYZE_TEMPORARY_DTORS "-analyze-temporary-dtors")
+    endif()
+
+    if (${CLANG_TIDY_VERSION} VERSION_LESS "3.9.0")
+        set(CLANG_TIDY_ERRORS_ARG "")
+    else()
+        set(CLANG_TIDY_ERRORS_ARG "-warnings-as-errors='${CLANG_TIDY_ERRORS}'")
+    endif()
+
+    if (${CLANG_TIDY_VERSION} VERSION_LESS "3.9.0")
+        set(CLANG_TIDY_QUIET_ARG "")
+    else()
+        set(CLANG_TIDY_QUIET_ARG "-quiet")
+    endif()
+
+    if(PARSE_HEADER_FILTER)
+        string(REPLACE "$" "$$" CLANG_TIDY_HEADER_FILTER "${PARSE_HEADER_FILTER}")
+    else()
+        set(CLANG_TIDY_HEADER_FILTER ".*")
+    endif()
+
+    set(CLANG_TIDY_COMMAND
+        ${CLANG_TIDY_EXE}
+        ${CLANG_TIDY_QUIET_ARG}
+        -p ${CMAKE_BINARY_DIR}
+        -checks='${CLANG_TIDY_CHECKS}'
+        ${CLANG_TIDY_ERRORS_ARG}
+        ${CLANG_TIDY_EXTRA_ARGS}
+        ${CLANG_TIDY_ANALYZE_TEMPORARY_DTORS}
+        -header-filter='${CLANG_TIDY_HEADER_FILTER}'
+    )
+    add_custom_target(tidy ${CLANG_TIDY_ALL})
+    mark_as_analyzer(tidy)
+    add_custom_target(tidy-base)
+    add_custom_target(tidy-make-fixit-dir COMMAND ${CMAKE_COMMAND} -E make_directory ${CLANG_TIDY_FIXIT_DIR})
+    add_custom_target(tidy-rm-fixit-dir COMMAND ${CMAKE_COMMAND} -E remove_directory ${CLANG_TIDY_FIXIT_DIR})
+    add_dependencies(tidy-make-fixit-dir tidy-rm-fixit-dir)
+    add_dependencies(tidy-base tidy-make-fixit-dir)
+endmacro()
+
+function(clang_tidy_check TARGET)
+    get_target_property(SOURCES ${TARGET} SOURCES)
+    # TODO: Use generator expressions instead
+    # COMMAND ${CLANG_TIDY_COMMAND} $<TARGET_PROPERTY:${TARGET},SOURCES>
+    # COMMAND ${CLANG_TIDY_COMMAND} $<JOIN:$<TARGET_PROPERTY:${TARGET},SOURCES>, >
+    foreach(SOURCE ${SOURCES})
+        if((NOT "${SOURCE}" MATCHES "(h|hpp|hxx)$") AND (NOT "${SOURCE}" MATCHES "TARGET_OBJECTS"))
+            string(MAKE_C_IDENTIFIER "${SOURCE}" tidy_file)
+            set(tidy_target tidy-target-${TARGET}-${tidy_file})
+            add_custom_target(${tidy_target}
+                # for some targets clang-tidy not able to get information from .clang-tidy
+                DEPENDS ${SOURCE}
+                COMMAND ${CLANG_TIDY_COMMAND} "-config=\{CheckOptions: \[\{key: bugprone-reserved-identifier.AllowedIdentifiers,value: __HIP_PLATFORM_HCC__\; __HIP_ROCclr__\}\]\}" ${SOURCE} "-export-fixes=${CLANG_TIDY_FIXIT_DIR}/${TARGET}-${tidy_file}.yaml"
+                WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+                COMMENT "clang-tidy: Running clang-tidy on target ${SOURCE}..."
+            )
+            add_dependencies(${tidy_target} ${TARGET})
+            add_dependencies(${tidy_target} tidy-base)
+            add_dependencies(tidy ${tidy_target})
+        endif()
+    endforeach()
+endfunction()
+
diff --git a/3rdparty/composable_kernel/cmake/CppCheck.cmake b/3rdparty/composable_kernel/cmake/CppCheck.cmake
new file mode 100644
index 0000000..63e1441
--- /dev/null
+++ b/3rdparty/composable_kernel/cmake/CppCheck.cmake
@@ -0,0 +1,130 @@
+################################################################################
+# 
+# MIT License
+# 
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+# 
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+# 
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+# 
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+# 
+################################################################################
+
+include(CMakeParseArguments)
+include(ProcessorCount)
+include(Analyzers)
+
+find_program(CPPCHECK_EXE 
+    NAMES 
+        cppcheck
+    PATHS
+        $ENV{ROCM_PATH}/bin
+)
+
+ProcessorCount(CPPCHECK_JOBS)
+
+set(CPPCHECK_BUILD_DIR ${CMAKE_BINARY_DIR}/cppcheck-build)
+file(MAKE_DIRECTORY ${CPPCHECK_BUILD_DIR})
+set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES ${CPPCHECK_BUILD_DIR})
+
+macro(enable_cppcheck)
+    set(options FORCE)
+    set(oneValueArgs)
+    set(multiValueArgs CHECKS SUPPRESS DEFINE UNDEFINE INCLUDE SOURCES)
+
+    cmake_parse_arguments(PARSE "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    string(REPLACE ";" "," CPPCHECK_CHECKS "${PARSE_CHECKS}")
+    string(REPLACE ";" "\n" CPPCHECK_SUPPRESS "${PARSE_SUPPRESS};*:/usr/*")
+    file(WRITE ${CMAKE_BINARY_DIR}/cppcheck-supressions "${CPPCHECK_SUPPRESS}")
+    set(CPPCHECK_DEFINES)
+    foreach(DEF ${PARSE_DEFINE})
+        set(CPPCHECK_DEFINES "${CPPCHECK_DEFINES} -D${DEF}")
+    endforeach()
+
+    set(CPPCHECK_UNDEFINES)
+    foreach(DEF ${PARSE_UNDEFINE})
+        set(CPPCHECK_UNDEFINES "${CPPCHECK_UNDEFINES} -U${DEF}")
+    endforeach()
+
+    set(CPPCHECK_INCLUDES)
+    foreach(INC ${PARSE_INCLUDE})
+        set(CPPCHECK_INCLUDES "${CPPCHECK_INCLUDES} -I${INC}")
+    endforeach()
+
+    # set(CPPCHECK_FORCE)
+    set(CPPCHECK_FORCE "--project=${CMAKE_BINARY_DIR}/compile_commands.json")
+    if(PARSE_FORCE)
+        set(CPPCHECK_FORCE --force)
+    endif()
+
+    set(SOURCES)
+    set(GLOBS)
+    foreach(SOURCE ${PARSE_SOURCES})
+        get_filename_component(ABS_SOURCE ${SOURCE} ABSOLUTE)
+        if(EXISTS ${ABS_SOURCE})
+            if(IS_DIRECTORY ${ABS_SOURCE})
+                set(GLOBS "${GLOBS} ${ABS_SOURCE}/*.cpp ${ABS_SOURCE}/*.hpp ${ABS_SOURCE}/*.cxx ${ABS_SOURCE}/*.c ${ABS_SOURCE}/*.h")
+            else()
+                set(SOURCES "${SOURCES} ${ABS_SOURCE}")
+            endif()
+        else()
+            set(GLOBS "${GLOBS} ${ABS_SOURCE}")
+        endif()
+    endforeach()
+
+    file(WRITE ${CMAKE_BINARY_DIR}/cppcheck.cmake "
+        file(GLOB_RECURSE GSRCS ${GLOBS})
+        set(CPPCHECK_COMMAND
+            ${CPPCHECK_EXE}
+            -q
+            # -v
+            # --report-progress
+            ${CPPCHECK_FORCE}
+            --cppcheck-build-dir=${CPPCHECK_BUILD_DIR}
+            --platform=native
+            --template=gcc
+            --error-exitcode=1
+            -j ${CPPCHECK_JOBS}
+            ${CPPCHECK_DEFINES}
+            ${CPPCHECK_UNDEFINES}
+            ${CPPCHECK_INCLUDES}
+            --enable=${CPPCHECK_CHECKS}
+            --inline-suppr
+            --suppressions-list=${CMAKE_BINARY_DIR}/cppcheck-supressions
+            ${SOURCES} \${GSRCS}
+        )
+        string(REPLACE \";\" \" \" CPPCHECK_SHOW_COMMAND \"\${CPPCHECK_COMMAND}\")
+        message(\"\${CPPCHECK_SHOW_COMMAND}\")
+        execute_process(
+            COMMAND \${CPPCHECK_COMMAND}
+            WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+            RESULT_VARIABLE RESULT
+        )
+        if(NOT RESULT EQUAL 0)
+            message(FATAL_ERROR \"Cppcheck failed\")
+        endif()
+")
+
+    add_custom_target(cppcheck
+        COMMAND ${CMAKE_COMMAND} -P ${CMAKE_BINARY_DIR}/cppcheck.cmake
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        COMMENT "cppcheck: Running cppcheck..."
+    )
+    mark_as_analyzer(cppcheck)
+endmacro()
+
+
diff --git a/3rdparty/composable_kernel/cmake/DoxygenDoc.cmake b/3rdparty/composable_kernel/cmake/DoxygenDoc.cmake
new file mode 100644
index 0000000..2e3669f
--- /dev/null
+++ b/3rdparty/composable_kernel/cmake/DoxygenDoc.cmake
@@ -0,0 +1,355 @@
+################################################################################
+# 
+# MIT License
+# 
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+# 
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+# 
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+# 
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+# 
+################################################################################
+include(CMakeParseArguments)
+include(MainDoc)
+
+find_program(DOXYGEN_EXECUTABLE NAMES doxygen
+    PATH_SUFFIXES bin
+    DOC "Doxygen documentation generator"
+)
+mark_as_advanced(DOXYGEN_EXECUTABLE)
+
+find_path(DOT_EXECUTABLE NAMES dot
+    PATH_SUFFIXES bin
+    DOC "Graphviz"
+)
+mark_as_advanced(DOT_EXECUTABLE)
+
+set(DOXYGEN_ARGS
+ABBREVIATE_BRIEF
+ALIASES
+ALLEXTERNALS
+ALLOW_UNICODE_NAMES
+ALPHABETICAL_INDEX
+ALWAYS_DETAILED_SEC
+AUTOLINK_SUPPORT
+BINARY_TOC
+BRIEF_MEMBER_DESC
+BUILTIN_STL_SUPPORT
+CALLER_GRAPH
+CALL_GRAPH
+CASE_SENSE_NAMES
+CHM_FILE
+CHM_INDEX_ENCODING
+CITE_BIB_FILES
+CLANG_ASSISTED_PARSING
+CLANG_OPTIONS
+CLASS_DIAGRAMS
+CLASS_GRAPH
+COLLABORATION_GRAPH
+COLS_IN_ALPHA_INDEX
+COMPACT_LATEX
+COMPACT_RTF
+CPP_CLI_SUPPORT
+CREATE_SUBDIRS
+DIAFILE_DIRS
+DIA_PATH
+DIRECTORY_GRAPH
+DISABLE_INDEX
+DISTRIBUTE_GROUP_DOC
+DOCBOOK_OUTPUT
+DOCBOOK_PROGRAMLISTING
+DOCSET_BUNDLE_ID
+DOCSET_FEEDNAME
+DOCSET_PUBLISHER_ID
+DOCSET_PUBLISHER_NAME
+DOTFILE_DIRS
+DOT_CLEANUP
+DOT_FONTNAME
+DOT_FONTPATH
+DOT_FONTSIZE
+DOT_GRAPH_MAX_NODES
+DOT_IMAGE_FORMAT
+DOT_MULTI_TARGETS
+DOT_NUM_THREADS
+# DOT_PATH
+DOT_TRANSPARENT
+DOXYFILE_ENCODING
+ECLIPSE_DOC_ID
+ENABLED_SECTIONS
+ENABLE_PREPROCESSING
+ENUM_VALUES_PER_LINE
+EXAMPLE_PATH
+EXAMPLE_PATTERNS
+EXAMPLE_RECURSIVE
+EXCLUDE
+EXCLUDE_PATTERNS
+EXCLUDE_SYMBOLS
+EXCLUDE_SYMLINKS
+EXPAND_AS_DEFINED
+EXPAND_ONLY_PREDEF
+EXTENSION_MAPPING
+EXTERNAL_GROUPS
+EXTERNAL_PAGES
+EXTERNAL_SEARCH
+EXTERNAL_SEARCH_ID
+EXTRACT_ALL
+EXTRACT_ANON_NSPACES
+EXTRACT_LOCAL_CLASSES
+EXTRACT_LOCAL_METHODS
+EXTRACT_PACKAGE
+EXTRACT_PRIVATE
+EXTRACT_STATIC
+EXTRA_PACKAGES
+EXTRA_SEARCH_MAPPINGS
+EXT_LINKS_IN_WINDOW
+FILE_PATTERNS
+FILE_VERSION_FILTER
+FILTER_PATTERNS
+FILTER_SOURCE_FILES
+FILTER_SOURCE_PATTERNS
+FORCE_LOCAL_INCLUDES
+FORMULA_FONTSIZE
+FORMULA_TRANSPARENT
+FULL_PATH_NAMES
+GENERATE_AUTOGEN_DEF
+GENERATE_BUGLIST
+GENERATE_CHI
+GENERATE_DEPRECATEDLIST
+GENERATE_DOCBOOK
+GENERATE_DOCSET
+GENERATE_ECLIPSEHELP
+GENERATE_HTML
+GENERATE_HTMLHELP
+GENERATE_LATEX
+GENERATE_LEGEND
+GENERATE_MAN
+GENERATE_PERLMOD
+GENERATE_QHP
+GENERATE_RTF
+GENERATE_TAGFILE
+GENERATE_TESTLIST
+GENERATE_TODOLIST
+GENERATE_TREEVIEW
+GENERATE_XML
+GRAPHICAL_HIERARCHY
+GROUP_GRAPHS
+GROUP_NESTED_COMPOUNDS
+# HAVE_DOT
+HHC_LOCATION
+HIDE_COMPOUND_REFERENCE
+HIDE_FRIEND_COMPOUNDS
+HIDE_IN_BODY_DOCS
+HIDE_SCOPE_NAMES
+HIDE_UNDOC_CLASSES
+HIDE_UNDOC_MEMBERS
+HIDE_UNDOC_RELATIONS
+HTML_COLORSTYLE_GAMMA
+HTML_COLORSTYLE_HUE
+HTML_COLORSTYLE_SAT
+HTML_DYNAMIC_SECTIONS
+HTML_EXTRA_FILES
+HTML_EXTRA_STYLESHEET
+HTML_FILE_EXTENSION
+HTML_FOOTER
+HTML_HEADER
+HTML_INDEX_NUM_ENTRIES
+HTML_OUTPUT
+HTML_STYLESHEET
+HTML_TIMESTAMP
+IDL_PROPERTY_SUPPORT
+IGNORE_PREFIX
+IMAGE_PATH
+INCLUDED_BY_GRAPH
+INCLUDE_FILE_PATTERNS
+INCLUDE_GRAPH
+INCLUDE_PATH
+INHERIT_DOCS
+INLINE_GROUPED_CLASSES
+INLINE_INFO
+INLINE_INHERITED_MEMB
+INLINE_SIMPLE_STRUCTS
+INLINE_SOURCES
+INPUT
+INPUT_ENCODING
+INPUT_FILTER
+INTERACTIVE_SVG
+INTERNAL_DOCS
+JAVADOC_AUTOBRIEF
+LATEX_BATCHMODE
+LATEX_BIB_STYLE
+LATEX_CMD_NAME
+LATEX_EXTRA_FILES
+LATEX_EXTRA_STYLESHEET
+LATEX_FOOTER
+LATEX_HEADER
+LATEX_HIDE_INDICES
+LATEX_OUTPUT
+LATEX_SOURCE_CODE
+LATEX_TIMESTAMP
+LAYOUT_FILE
+LOOKUP_CACHE_SIZE
+MACRO_EXPANSION
+MAKEINDEX_CMD_NAME
+MAN_EXTENSION
+MAN_LINKS
+MAN_OUTPUT
+MAN_SUBDIR
+MARKDOWN_SUPPORT
+MATHJAX_CODEFILE
+MATHJAX_EXTENSIONS
+MATHJAX_FORMAT
+MATHJAX_RELPATH
+MAX_DOT_GRAPH_DEPTH
+MAX_INITIALIZER_LINES
+MSCFILE_DIRS
+MSCGEN_PATH
+MULTILINE_CPP_IS_BRIEF
+OPTIMIZE_FOR_FORTRAN
+OPTIMIZE_OUTPUT_FOR_C
+OPTIMIZE_OUTPUT_JAVA
+OPTIMIZE_OUTPUT_VHDL
+OUTPUT_DIRECTORY
+OUTPUT_LANGUAGE
+PAPER_TYPE
+PDF_HYPERLINKS
+PERLMOD_LATEX
+PERLMOD_MAKEVAR_PREFIX
+PERLMOD_PRETTY
+PERL_PATH
+PLANTUML_CFG_FILE
+PLANTUML_INCLUDE_PATH
+PLANTUML_JAR_PATH
+PREDEFINED
+PROJECT_BRIEF
+PROJECT_LOGO
+PROJECT_NAME
+PROJECT_NUMBER
+QCH_FILE
+QHG_LOCATION
+QHP_CUST_FILTER_ATTRS
+QHP_CUST_FILTER_NAME
+QHP_NAMESPACE
+QHP_SECT_FILTER_ATTRS
+QHP_VIRTUAL_FOLDER
+QT_AUTOBRIEF
+QUIET
+RECURSIVE
+REFERENCED_BY_RELATION
+REFERENCES_LINK_SOURCE
+REFERENCES_RELATION
+REPEAT_BRIEF
+RTF_EXTENSIONS_FILE
+RTF_HYPERLINKS
+RTF_OUTPUT
+RTF_SOURCE_CODE
+RTF_STYLESHEET_FILE
+SEARCHDATA_FILE
+SEARCHENGINE
+SEARCHENGINE_URL
+SEARCH_INCLUDES
+SEPARATE_MEMBER_PAGES
+SERVER_BASED_SEARCH
+SHORT_NAMES
+SHOW_FILES
+SHOW_GROUPED_MEMB_INC
+SHOW_INCLUDE_FILES
+SHOW_NAMESPACES
+SHOW_USED_FILES
+SIP_SUPPORT
+SKIP_FUNCTION_MACROS
+SORT_BRIEF_DOCS
+SORT_BY_SCOPE_NAME
+SORT_GROUP_NAMES
+SORT_MEMBERS_CTORS_1ST
+SORT_MEMBER_DOCS
+SOURCE_BROWSER
+SOURCE_TOOLTIPS
+STRICT_PROTO_MATCHING
+STRIP_CODE_COMMENTS
+STRIP_FROM_INC_PATH
+STRIP_FROM_PATH
+SUBGROUPING
+TAB_SIZE
+TAGFILES
+TCL_SUBST
+TEMPLATE_RELATIONS
+TOC_EXPAND
+TOC_INCLUDE_HEADINGS
+TREEVIEW_WIDTH
+TYPEDEF_HIDES_STRUCT
+UML_LIMIT_NUM_FIELDS
+UML_LOOK
+USE_HTAGS
+USE_MATHJAX
+USE_MDFILE_AS_MAINPAGE
+USE_PDFLATEX
+VERBATIM_HEADERS
+WARNINGS
+WARN_AS_ERROR
+WARN_FORMAT
+WARN_IF_DOC_ERROR
+WARN_IF_UNDOCUMENTED
+WARN_LOGFILE
+WARN_NO_PARAMDOC
+XML_OUTPUT
+XML_PROGRAMLISTING
+)
+
+set(DOXYGEN_CONFIG_FILE "${CMAKE_CURRENT_BINARY_DIR}/doxygen/doxygen.conf" CACHE PATH "Path to generated doxygen configuration file")
+
+function(add_doxygen_doc)
+    set(options)
+    set(oneValueArgs)
+    set(multiValueArgs DEPENDS ${DOXYGEN_ARGS})
+
+    cmake_parse_arguments(PARSE "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+
+    file(WRITE ${DOXYGEN_CONFIG_FILE} "# Auto-generated doxygen configuration file\n")
+
+    foreach(ARG ${DOXYGEN_ARGS})
+        if(PARSE_${ARG})
+            string(REPLACE ";" " " ARG_VALUE ${PARSE_${ARG}})
+            file(APPEND ${DOXYGEN_CONFIG_FILE} "\n${ARG} = ${ARG_VALUE}\n")
+        endif()
+    endforeach()
+
+    if(PARSE_OUTPUT_DIRECTORY)
+        if(NOT EXISTS ${PARSE_OUTPUT_DIRECTORY})
+            file(MAKE_DIRECTORY ${PARSE_OUTPUT_DIRECTORY})
+        endif()
+    endif()
+
+    if(DOT_EXECUTABLE)
+        file(APPEND ${DOXYGEN_CONFIG_FILE} "\nDOT_PATH = \"${DOT_EXECUTABLE}\"\n")
+        file(APPEND ${DOXYGEN_CONFIG_FILE} "\nHAVE_DOT = YES\n")
+    else()
+        file(APPEND ${DOXYGEN_CONFIG_FILE} "\nHAVE_DOT = NO\n")
+    endif()
+
+    add_custom_target(doxygen
+        ${DOXYGEN_EXECUTABLE} ${DOXYGEN_CONFIG_FILE}
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        COMMENT "Building documentation with doxygen"
+    )
+    if(PARSE_OUTPUT_DIRECTORY)
+        clean_doc_output(${PARSE_OUTPUT_DIRECTORY})
+    endif()
+    mark_as_doc(doxygen)
+    if(PARSE_DEPENDS)
+        add_dependencies(doxygen ${PARSE_DEPENDS})
+    endif()
+endfunction()
diff --git a/3rdparty/composable_kernel/cmake/EnableCompilerWarnings.cmake b/3rdparty/composable_kernel/cmake/EnableCompilerWarnings.cmake
new file mode 100644
index 0000000..78133af
--- /dev/null
+++ b/3rdparty/composable_kernel/cmake/EnableCompilerWarnings.cmake
@@ -0,0 +1,110 @@
+################################################################################
+#
+# MIT License
+#
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+#
+################################################################################
+# - Enable warning all for gcc/clang or use /W4 for visual studio
+
+## Strict warning level
+if (MSVC)
+    # Use the highest warning level for visual studio.
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /w")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /w")
+    # set(CMAKE_CXX_WARNING_LEVEL 4)
+    # if (CMAKE_CXX_FLAGS MATCHES "/W[0-4]")
+    #     string(REGEX REPLACE "/W[0-4]" "/W4" CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
+    # else ()
+    #     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /W4")
+    # endif ()
+
+    # set(CMAKE_C_WARNING_LEVEL 4)
+    # if (CMAKE_C_FLAGS MATCHES "/W[0-4]")
+    #     string(REGEX REPLACE "/W[0-4]" "/W4" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
+    # else ()
+    #     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /W4")
+    # endif ()
+
+else()
+    foreach(COMPILER C CXX)
+        set(CMAKE_COMPILER_WARNINGS)
+        # use -Wall for gcc and clang
+        list(APPEND CMAKE_COMPILER_WARNINGS
+            -Wall
+            -Wextra
+            -Wcomment
+            -Wendif-labels
+            -Wformat
+            -Winit-self
+            -Wreturn-type
+            -Wsequence-point
+            # Shadow is broken on gcc when using lambdas
+            # -Wshadow
+            -Wswitch
+            -Wtrigraphs
+            -Wundef
+            -Wuninitialized
+            -Wunreachable-code
+            -Wunused
+
+            -Wsign-compare
+            -Wno-extra-semi-stmt
+        )
+        if (CMAKE_${COMPILER}_COMPILER_ID MATCHES "Clang")
+            list(APPEND CMAKE_COMPILER_WARNINGS
+                -Weverything
+                -Wno-c++98-compat
+                -Wno-c++98-compat-pedantic
+                -Wno-conversion
+                -Wno-double-promotion
+                -Wno-exit-time-destructors
+                -Wno-extra-semi
+                -Wno-float-conversion
+                -Wno-gnu-anonymous-struct
+                -Wno-gnu-zero-variadic-macro-arguments
+                -Wno-missing-prototypes
+                -Wno-nested-anon-types
+                -Wno-padded
+                -Wno-return-std-move-in-c++11
+                -Wno-shorten-64-to-32
+                -Wno-sign-conversion
+                -Wno-unknown-warning-option
+                -Wno-unused-command-line-argument
+                -Wno-weak-vtables
+                -Wno-covered-switch-default
+            )
+        else()
+            if (CMAKE_${COMPILER}_COMPILER_ID MATCHES "GNU" AND ${COMPILER} MATCHES "CXX")
+                # cmake 3.5.2 does not support >=.
+                if(NOT CMAKE_CXX_COMPILER_VERSION VERSION_LESS "6.1")
+                    list(APPEND CMAKE_COMPILER_WARNINGS
+                        -Wno-ignored-attributes)
+                endif()
+            endif()
+            list(APPEND CMAKE_COMPILER_WARNINGS
+                -Wno-missing-field-initializers
+                -Wno-deprecated-declarations
+            )
+        endif()
+        add_definitions(${CMAKE_COMPILER_WARNINGS})
+    endforeach()
+endif ()
diff --git a/3rdparty/composable_kernel/cmake/TargetFlags.cmake b/3rdparty/composable_kernel/cmake/TargetFlags.cmake
new file mode 100644
index 0000000..4f83fb5
--- /dev/null
+++ b/3rdparty/composable_kernel/cmake/TargetFlags.cmake
@@ -0,0 +1,50 @@
+
+function(get_target_property2 VAR TARGET PROPERTY)
+    get_target_property(_pflags ${TARGET} ${PROPERTY})
+    if(_pflags)
+        set(${VAR} ${_pflags} PARENT_SCOPE)
+    else()
+        set(${VAR} "" PARENT_SCOPE)
+    endif()
+endfunction()
+
+
+macro(append_flags FLAGS TARGET PROPERTY PREFIX)
+    get_target_property2(_pflags ${TARGET} ${PROPERTY})
+    foreach(FLAG ${_pflags})
+        if(TARGET ${FLAG})
+            target_flags(_pflags2 ${FLAG})
+            string(APPEND ${FLAGS} " ${_pflags2}")
+        else()
+            string(APPEND ${FLAGS} " ${PREFIX}${FLAG}")
+        endif()
+    endforeach()
+endmacro()
+
+macro(append_link_flags FLAGS TARGET PROPERTY)
+    get_target_property2(_pflags ${TARGET} ${PROPERTY})
+    foreach(FLAG ${_pflags})
+        if(TARGET ${FLAG})
+            target_flags(_pflags2 ${FLAG})
+            string(APPEND ${FLAGS} " ${_pflags2}")
+        elseif(FLAG MATCHES "^-.*")
+            string(APPEND ${FLAGS} " ${FLAG}")
+        elseif(EXISTS ${FLAG})
+            string(APPEND ${FLAGS} " ${FLAG}")
+        else()
+            string(APPEND ${FLAGS} " -l${FLAG}")
+        endif()
+    endforeach()
+endmacro()
+
+function(target_flags FLAGS TARGET)
+    set(_flags)
+    append_flags(_flags ${TARGET} "INTERFACE_COMPILE_OPTIONS" "")
+    append_flags(_flags ${TARGET} "INTERFACE_COMPILE_DEFINITIONS" "-D")
+    append_flags(_flags ${TARGET} "INTERFACE_INCLUDE_DIRECTORIES" "-isystem ")
+    append_flags(_flags ${TARGET} "INTERFACE_LINK_DIRECTORIES" "-L ")
+    append_flags(_flags ${TARGET} "INTERFACE_LINK_OPTIONS" "")
+    append_link_flags(_flags ${TARGET} "INTERFACE_LINK_LIBRARIES" "")
+    # message("_flags: ${_flags}")
+    set(${FLAGS} ${_flags} PARENT_SCOPE)
+endfunction()
diff --git a/3rdparty/composable_kernel/cmake/googletest.cmake b/3rdparty/composable_kernel/cmake/googletest.cmake
new file mode 100644
index 0000000..3da4877
--- /dev/null
+++ b/3rdparty/composable_kernel/cmake/googletest.cmake
@@ -0,0 +1,50 @@
+include(FetchContent)
+
+set(GOOGLETEST_DIR "" CACHE STRING "Location of local GoogleTest repo to build against")
+
+if(GOOGLETEST_DIR)
+  set(FETCHCONTENT_SOURCE_DIR_GOOGLETEST ${GOOGLETEST_DIR} CACHE STRING "GoogleTest source directory override")
+endif()
+
+message(STATUS "Fetching GoogleTest")
+
+list(APPEND GTEST_CMAKE_CXX_FLAGS
+     -Wno-undef
+     -Wno-reserved-identifier
+     -Wno-global-constructors
+     -Wno-missing-noreturn
+     -Wno-disabled-macro-expansion
+     -Wno-used-but-marked-unused
+     -Wno-switch-enum
+     -Wno-zero-as-null-pointer-constant
+     -Wno-unused-member-function
+     -Wno-comma
+     -Wno-old-style-cast
+     -Wno-deprecated
+)
+message(STATUS "Suppressing googltest warnings with flags: ${GTEST_CMAKE_CXX_FLAGS}")
+
+FetchContent_Declare(
+  googletest
+  GIT_REPOSITORY http://10.0.50.24/Mirrors/googletest.git
+# GIT_REPOSITORY /work/home/zhangshao/installer/googletest
+  GIT_TAG        b85864c64758dec007208e56af933fc3f52044ee
+)
+
+# Will be necessary for windows build
+# set(gtest_force_shared_crt ON CACHE BOOL "" FORCE)
+FetchContent_GetProperties(googletest)
+if(NOT googletest_POPULATED)
+  FetchContent_Populate(googletest)
+  add_subdirectory(${googletest_SOURCE_DIR} ${googletest_BINARY_DIR} EXCLUDE_FROM_ALL)
+endif()
+
+target_compile_options(gtest PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+target_compile_options(gtest_main PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+target_compile_options(gmock PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+target_compile_options(gmock_main PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+
+set_target_properties(gtest PROPERTIES POSITION_INDEPENDENT_CODE ON)
+set_target_properties(gtest_main PROPERTIES POSITION_INDEPENDENT_CODE ON)
+set_target_properties(gmock PROPERTIES POSITION_INDEPENDENT_CODE ON)
+set_target_properties(gmock_main PROPERTIES POSITION_INDEPENDENT_CODE ON)
diff --git a/3rdparty/composable_kernel/compile.sh b/3rdparty/composable_kernel/compile.sh
new file mode 100644
index 0000000..92521ea
--- /dev/null
+++ b/3rdparty/composable_kernel/compile.sh
@@ -0,0 +1,13 @@
+mkdir build -p
+cd build
+
+cmake                                                                                             \
+-D CMAKE_PREFIX_PATH=${ROCM_PATH}                                                              \
+-D CMAKE_CXX_COMPILER=${ROCM_PATH}/bin/hipcc                                                    \
+-D CMAKE_CXX_FLAGS="-O3"                                                                          \
+-D CMAKE_BUILD_TYPE=Release                                                                       \
+-D GPU_TARGETS="gfx906;gfx926"                                                                           \
+-D CMAKE_INSTALL_PREFIX=~/composable_kernel/install_ck  \
+..
+
+cd -
diff --git a/3rdparty/composable_kernel/dev-requirements.txt b/3rdparty/composable_kernel/dev-requirements.txt
new file mode 100644
index 0000000..9e7b9f0
--- /dev/null
+++ b/3rdparty/composable_kernel/dev-requirements.txt
@@ -0,0 +1,3 @@
+ROCmSoftwarePlatform/rocm-recipes
+RadeonOpenCompute/rocm-cmake@04f694df2a8dc9d7e35fa4dee4ba5fa407ec04f8 --build
+danmar/cppcheck@2.9
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/doc/image/ck_component.png b/3rdparty/composable_kernel/doc/image/ck_component.png
new file mode 100644
index 0000000000000000000000000000000000000000..db892331d77273208f861eb68f1df46d0f2667f4
GIT binary patch
literal 565049
zcmeFZdHnQbc`y8cw1A=rE-fk`RTMYOGD#)~!X%S@GBcUUB$Kp?Stt8slFUpJm8w<T
zYD-<JmRhKYdqr`pia<rJAfRHaRZ&D#upR{!MFHj90est@_w;k{{r5f3@AsQz?j-km
z-QVkbEqPu!9di$T@M9i);J|^03jA<(;K1Qm95`_3GY|WHQ1b0ZoV5gpgUT649eCfh
zkNd%a1E;;SVwP1}`(ff8I1#1qJ3A49f^}J)hz?Ig5GL`Qs#s^>6qKiHH}pf#z3&<V
zA;>de=ot{ILdc2e04Bf(N}No<uzBD8*2>H7uLwIC0u`L3hnc(Hm*8})f)5x1#V7^N
zzy-W_-)W8ao&rh+d7gN>XG>vqBEn*zViX6(CrP|2k4{8raGr#*2M(O)#Ov-6x3Kmn
zhih;<0-cQB+aetl>lEC?!cgyI=wu9rk&|J#2X6k4p>=n|C!?TjZuwRb{-@@;QG1oO
zx^KiSR0o{2>lbL6&G>A|oSEJ~(w$e7A(*CnC+)Y5cMn%h?%h}Fdb`j8J?TnwFf-+S
z#h_U*c?_6-H?&rTm4eIR{WI+By}7#Nu@D_Ijwx5=Gc^9b3Ni}}>fXkj_)KS==Iabh
zHJI0~F2c-B2bX({0A`jQ8l)Jy(;l>N`V(R*+T6O`VhNToyn(EEhTf&tlIYJvsw5^u
zoZgU@w%H8zz@7Q?27<$=CsuiX9f{_&X7m#gYQ1974LAns%pETjQeN~G-^olx9;p*;
zL<dN2?Umhx^5=sMl}Pv=OerNw%EA&_lv=RrIv$GrP>MuBa+5`fW=sKkYJjJ-4bAo(
zeCQ<w86*un+oU+HfQvnQZ{!(V3?!53(`8Ai%pBAj&ecU4Pn!r2s1Xfq+|>l)KzTZX
z%Vj0vWs6|ot5MC^k+5ZLtk2DB&nK`gsI$i9<qjON2Fy$VZd<O^!fJvYCJmFW&!cgb
z$UTa78zqf5F3n`401=iovrcr)#Ei0}mk|<<kfjgq9!@aLkP9ekl(EK4(tIzBV_HM>
zxfr`skIbgQObzB*oXm6yu@<0Zxz*NX!euqj4s=Oq62YV0lmV3qO1dXms-KUW#@ufD
zsDe8a1@Y)rstb8oZiLAsq6;QtI9!oxg&6O!px)Dv#Ozwz?He0=RVZ@BEyOaexmf9&
z%f$Ctr~s>%v|cx-Q))VySqx?6h%ttmVm{vKGAuQsD>jXmr==ImW}lFP2|m>q!}TmO
z?4E*TMlfGd<USNC#b(f!&NO1YVtAenc4Gt7%U6s=Fg;zx{o3aoEu?{zvk4^1P-AR`
zQk591zmjqu>CNSl*asI!TEO<VS%6!qGwiR)Tv&qAkfPaWI$l;y(@jaNhVfkH{291}
znG753>535zq44H9RAXggZdb`_l?og|;982#_V9KwvYaqeu#FX}u|f8Cx}o3!7^{<Q
zH;7Qcvbc6j8tf}-5n^1T?tw&59s(Dw9u7HZs>6{ga0=_|fv(79wbqsWreZ@`wGF%8
zYuuV-RT4Zi<cpPTFZ*Ekg1YvqEZ9oiL(~2$i1AXaXa<GjV&v?>Wkqh6iVRNXGTw~P
z$PxBrK$D^;T5J>#RDtf3u{Oz!9)MnU&QIj1RB{+RpiV|(kssvIS}*s4F~=vuOdg0~
zsDuIPn{q-&b*t*5VC~Gvy(CrywyV3jtY~u>BWYH;@ECk`t9CEa{mI(c08ve%^}aSr
zD);Gqny+FtAM!Yv_i>BN*rrX@om;MGC|hm0gd9O@iO&nv+se}*<h%hdH!CuRk`|1d
zT{*KMhLpoqn+uJ=rMs|z^G!CQ3RPt^cTdCcZYK)+0b;Xcw>Mb>#zlFQPhyx)mc_cq
zK^#Qu(+TJ%h3_h4ImY^In*s&G7ISPzxhhG}0g<}2PwlsD;X)D3wn@<rwC)M;*XR35
z#%}Zy@_Ni3tE;}St2voMS$jm9<Jl0-kqJIgS2}@&8b2gPLb{wS=Y(wGvq_dIlW7~9
z&C)E>Eyho0ob3?oh)a?bC>IKE#-g>zCg4;p!fLQ6dY**#Y?Rue1V-AmIue;VK3wn3
z41t+)3saaqGs7j`3};ZKlDd&3ZXwU{O}>DHeJ>fcqmXxME*hu?ZtV~oRJ{t<%U*>N
z)^rK`SZnU^ATNAI$L3)|sbbVmTS?i}J<FVibS%>TrZwp~Y!Hk)7o|K|CnKOsSQHV$
zxvp({{Z*tU;EJ^yclvPOtTv&hD5|Syo*sLX2o#o1pIl8+++B5F44*4{8`~nGL4~=+
zZVyLx2b-kVVdFS0nL#Z=$YSdWes9)dnvMri?F2q*^e8#rN`(;-Os9jyp-dB_O&F;x
z>&qU~#}#s~IWd(hGBn#4*lxCFk{vSjQ>w$^nrQ$%pbBX?#`mI)61j->#&|;|=2*pD
zDhU=r7I;z9gornHKnr?b6P9v8M(Juj9>#dwg6F+u6^(0Avz&4b@ULHE@l<BQPOkC<
z9ZBFh@QdRXT{%%IcvieD<Y)rL+XUR=6ygT!HuffB8nJJVGgbmJNM(o$XQEWPInCAH
z$dj@$A{!#O(xKeQg*KRg=DPv03!|wt#u|hlGbu^7ldKm`D#efO!kJ7Q6zB}LH-X6=
zUPi{UiG5_aoU$po%UIM^xKfV>SwuD-G2r*+l&=kT1E7)`m}>XrY6>_5IR2J4{5Kq8
zyX-AnXS#Jp%h4)Jfa;*Z)N~}|zi^1eY5;|~pFe1=ZC7QFnIhI;k`9L^zo?gl-0Dc(
z>iB>i2M7wZUTGxKI`={b?WDqwQ*oP165-4msL1KnVim=+QI=8hWa3S{wN(My#hQGC
zd6FK?6?_HZy=gpMn()9Im>sAm++3aSZDpqQwrM6?SS%{EN8~=FjQBb1LM9zYio+Th
zw`y>6U*{6jGP7Y2C7Iesc|M62{-m6@y*Su{6=8!$>;y?&OzK8jL@l1h**<hfy)`<3
zd_P?BbxBz<T{lgRtp(I<0;!K14S;4gBxkyqOxM!RQ!6om(FPzl8;KN0jD?OQ`cv?M
z)aYzLthyBmMt%(%696Y<X|*?T!SVnN%NF^-+fG@tq6a-u&4A8_bxqj1N^KfxKW^zQ
zm4zl<c3_x{@;cX*pcjYUAj^PipRU7LmrWQ7LbhFQo1P7Oy}l`U_yUQ<g#(*Hq$CQ@
zx>l=-!!Y9+6De#g7quDXjRDv#28holbD;65M)xF#Q!y#WW1@*0yIrU4gd*JmXGlt{
zM>+#G>J}cx>)4zDtry5tr!9mTIv);Zd9h8)p~7MmA$gK&;XBhPtr%Jd@T&0=0$vJo
zsYYOnr}88+R=6_|;GMe4TNp_~B9>LKSE-pPcQXUXjpOjMWlPWYk;6*H9%V-TK21;T
zSuo6YLv#qXQsAJYc(~7p)D8$On+dBZ-5Uyq@5!b$CKxXyt_ZD$-2$LqqIB$kq+0|;
zqp8~Beg=tSvJoAEH05dPtexD?sdUYy{Mz3-IqFKI&D=z%Fo{m)G@xwZY9V6FZqSSf
zrk`C#s;k1?+RSG1dY;vmFUFqfCxnCOi$Y(QDJhoDJjn;TU@rwhAn8ml)<it=LoPPV
z-5f2Au+?VLyqr)(ic<o=_QJRVReEx?A0>vk!r7P(Y`ll@ZZo&y5KPv{))YpP8d5C8
zeH$C`3ex<!-N^37?butwWGc--q-<4lCxTVV8bd>R4@oODsTC($n5#;Fxh0;J12yla
zMalTp4oTHs2V-kq*<#B%H=5M7&1|*ZX4N8DAF<F<M<Gj0YG!YvrM567cw0mSDkDBx
zkL|jzruy7MoMM_ul47}-u*n5dSHskxk|GKdW#-L>`j&HjcP081OiiSH(Duu)0L@C%
zVbR=ck(oh;4B$;mPM9WM6-ZzSrb437NLZNlW{;H7*1!|Yma1UlX0yWHZ`o097m|dC
zl0XiK7>tRf-2QyE<kRJ%<V+p+a4ApM&3?%*NUSW?$%KiQ6TOhBq$;pcSq-bnsMIz)
zVBJz06$_<?%x%$+S+I3kyYz;e%JRzs!M8i5UO7>vK{3;+re%TaqNrEv*+fmb#d@}k
zhFU7|x<fGYK|U^t8IcOu7HACC<|a+B?sNxGBy5U#RzE;^XhyWSu^5yTY=F_CMM|1+
zuf@X^C#|G0v9LvQ7j#$(?Hb8lq?WAtCaTXhbF*FeLPm&C$pP-tfQoUwrXs*~SYw^W
z@I;MT52<85jS(l&5MqNuyt*XjEMbmBcTDDEdYi<g=uWJzOSq7`n8na#UZKj`HPK>+
zrkaM<!+>jt?GCII3mQ?J)wTpYhc!)}=JuXoh2Er}bCf+uIuZ1vH4jC2yKLLVJY1!d
zT;Ft^^D&Q%gbg^udS0iAouZWbCAtH%Ijdz2>!+Eihl_jwt}#Bj@`wtoCRI(mPLFPt
zlA#i{41uL?i7SA6>R`Z^ts<gx22sUPt++EpgVSoD&GRxh!<B{S3wiIlakS7-RIyzj
zbF;3Cyn)lZg7n~&^lYSDxIImoEOpLpoQBbTf)t1IjmqY$M&%k7KsSIVLKza)P-jNd
z1kednB*hw%{nR8^3fvi&y#=ycgwfcuI-%k1z6+QZ<@tpcET)lwIk32JN9cN<b`)Qf
z`9k7Yz2l4ZA_hv9TGC5GS!p9|MwslTBOBqgspPsWB7SMtCX5r2*l&koKQ(1+iPLg2
z*JG}8da&t+aJ;l~17{JTuHN4*qr6w=8$(==#vEB`V5?IOx=ll>LVEt*4X}aI*&I+6
zvJ|e2iY>WCA}44%wm#YuC9bbkQn5C3yVMd7aA;Ou?1fER1sR!a79o-cDsEzsFMwM5
zDO7W$VUhhoZ?baxk}EM)-i$I`AIytp<xbjmx80PCw^dwv+)A2eEJ&{D18HFS4gof9
ztU(T}b~?5!18e1u1|w25R|DIoSGFhia(vq3H)D7QxG|2>mBKh%RicX-TussWx-X;a
z&AlA9)y%2G7p2o-slIepepJ;HcMVPI5RY<#FiAt2uUgmV$8|m4I=yJeKuFC4S~h7%
z-jq=1xiDhLEgY8^LNv&7%etzBkfS<q9CXKM#bUZ1EgS$Btd1^6ls3v{#4br02vaA^
zE#Gf~!bym2(`k&c!UwH5Gf9;%21AI4NO)vTI#Pi6wy&1u;Qo!pV}Nc`I;e>DAa_Si
z2$C6-Ga~RHb6nC0UJz0-ocjwM9ZcK#OpFa)SW)Wy-VKvSR%A>C*+vE>np=x%OKGUG
zlREncnAxZ^Q28{{0z(Q=IES{T?Pv<U!Fu?7+icQR=igJ85bsSxc5kztP97@2!64yq
zUgat&w*s76v%b?}v*j*bVDfr43hF4D(M3O-vw%{uB;Uz4d#^+)i38Lx6+&bx4Cg4#
z?&xt;Ef<BG?Y00%*qJjJCBQl@H~aCXh(}UX%2+9gT0iwsRdXp^OUOn}r?niDom@1L
zr1{nk8;*UC?=5?Cz=#w#k?U>U(7B)oOT(Yd1Ad0~+hmz$LAK58Q86<77H@!PP{6MO
zwS*?K18+0aCNMKuPC+rKdNSC7&wR&J!*0p;og1zD?MCu{;L`J<KskE{vPR@|G+1i5
zWB@oo{tK5*lQR=5b4$kpj=%fdaT;v9WFeHh9=y!9K1Hbbdb^DPI#DgdVB=n#0qC4}
zUrmu7i|c_QEE{u4fD*!8OgDf4ai~2Sc$}GV3PZan>{d8|<7*Ykdwt)?`ruYRo<{~R
z($FNTgHaC*M=)lg5)<7i5<v-zQdtToKu9Pebi|pj!)TA)^=ym+zP@Z>Mr)BJ3{7Mi
zWbUvycx)onX^GGVED@M9+mkR$A{>NnvbI2p{V3lTUNFVDZbpN25@##|+e-!C6}&J4
zlGU*$dEpZ>Op{*5l14OzXCzBwGnt%NoeG|;LzS@o!WcGYpEPo~aWQnawaxx|$L{Af
za6H;mv;oF$E^<BG1M<&OZMXR?b?3JzREn94Rk%Nzc6w^kEi%$*lbYf=C=(#s4_rv8
z`eY=_orW3OC@?VD)}$H}c2mSrX}@IjPPs?BmYwiW<r9?9+Xl<6PLw0Gmvy3@^$B7%
z?S->m!M2^<*jr5%RT;;-P~6RDprG1%V8SS4t*0YtG+vU%Tv<io3Yi-S?8LJ{zTCE1
zGg=i>goC|Fv-GFvs#5EAw`=<hXoPKnB(cC1+jIR0b97MKGNrt%+ieLps&wG>iS1^R
zOmb^Rj6f;gi`;r!t-YeqnXVIqNuEz2bji$Q21h$8te~JC;Ainta1eJH_}pOx@QYB$
zcs+y6X<vmw*>GkJlb~8>1_QnsE?NOM^2j4Z77~5U*Xe-TfsIU2)tIRox(~R7yxsE<
z5>7^tEqZ;oh|2(XaW00kBvI{V>a{V6*Uk>=>SraMQVcAhcU#G$1aHH3+OPnEKHD=Z
zf49~A(XK~$<4%&IC2-U6?i6al8iCtvNeV=;Z3nK^2uh!s%kFroce3S$3v?J`zTj>K
z?i#5`o})-PoVC?h(VHEXaWid7Qbs=RxS~n-R+ta(Nxc(2c26$4t`VD-@waH(8ShiO
z+<<Bg)b_{AEmP>)d<lwUywABaX#?PEH|{1rocB|L<CVq=R%};E-7Hjj>+i=r8FFh3
zpaRQH2btzBi`6E!t%){=Wg%n3!D=^O_Y`^?;o7JtV){y*h}cx*AiZtKiOr~S9fquc
z3RYrMp*~pS<_HghMh0pT===^Px`7LGQuq6n?x6EFWi@$Cs|c|ZBcdcB(0B(O`<m7b
zpk~;fMRuxwQ6qC`%4;RLXa_5nk@S1PyF$_d8(7N5U>B;h?~o;xwm!4$nS&x-#tNtk
zN&wVlsmj-=-Bq8F)ShB%YvwGN)fBD)s7|AOx~AAB9Im<z<_Anx=1tWE19?D^We|iJ
zk<JI<l0yW2LaxV=A&l2^VMI(<<RqsPl`$y0+iP+;$)_<>&|y$Y)q2<f*&pbI$_H}^
zQ0xV`h4lRq=tl?V{&G7qLURlcOAdmURxPyyLKsu^6sR?p+7u9};vzCK<rPubxxODN
zm7&W)=f&7B`rCA{Q{?fe^JszJO+JEYi_b7jC2`S1Vn!~eT{w%~=Sw5sa+WpKhLkE#
zbDO82JynHcPE%P)Ol)*c<EAaHeW?JHC?r!V0<g?Nv|DH*7kZnvZzR@WFrXAHNS0>5
zT<ObYX$~pif}iY?xkULQ;4B2`W5g8oOM=)<<z`3h9kL<qxIcxkA@EZ5AS9w3UzNpT
zvTA1W+J(ZQpLj9AR?g~)>(~du13w#VMV*d4mYGqlLYweV9u{DOF3NPVDh4rMg#^Fr
z52+4P$t)ccIktf$6tac2AYi2kZRcWt00nLgxkjGWtBP4rERqkR!jpud2K;d4A}xpo
z4sW+Rbv@bSm22^ev)0labFIOum}Hws7J3|Q2K97l2}#l5<U)%H&8$Eq!}VaDDjHi0
z$f2e5G*7{Lr0Wvn9!f69^k!)KyD8e4QDMyhWlgp0dbhyi8NrtO2ncbW;u#y*8iL`@
zluh`uUWh=(1$-9hiOCHnDuXk*v#%+GCp+fO4#d3j^{KEAHqIQ0sc@jgtVyk<0r0r0
z6p7FP(xlz}jjmi41j73Icn)?eVeV;w%wQG3grSRByX~d>d9qtjw6j^etJtc|8r%2@
zKyMD+^rzA^PZEp7>c*Smz0tlffoV+ly#%Sa4om@?tX-gi)P<^MHYfli?aEi+Fy%5x
z>IKQpTP6{y41tTl2*-A!cG|!KVJAjzu_arCk~kD9S8Mc8m>LC8f55}mb5SUm&VZE_
ziP?UW^`Uq_bj2xcCvb!bH3MtbA`(ewzXCKj91Y@_>QCy~%w7oL07iW~3n35;+RrE1
zVxqcvxJ#9tX;R8?vP0pdg_$91yY^Ub`w8Ahwf$&ljR2Mt%{*d3$QS6cEuEXa2u!^*
zg7tyCX$_~h4}p1(_ksb}PB(LwT~PeA*ul=+Z;jE6nF+KwY-gpT4g^&*Qfoi$D8jbT
zV4R}}3df8UZLA6r)SB20G#G?ye5nrve7Lc9t9A?8rDvgfuhSNslYBH`I;3A3CCi$9
z9trg-L!*TQK(JGbnhL{`W@{kq#Vm%CeZ7_VHDHzUih`z#xE7X+fnFB*BtVG{<=Is3
z`+${a0l?R8rDVU}@IXUQoYyY3g0FjN)FXBhRPYnf0M*;V_SCPUH6p2K#_`~(MOrB}
zl?5TMEY}=@EP`3RETm|&?wKaec{?9ripC2XCtOy0&lWd3Es4<{x7y%t+{(~Gh5>h^
zM!BN;Lj)sMG@K6P{yetGV6ZP?U?+MkHP3ih^!r8)deIE~eFnC75k6zP;3&IOdo(Bc
zP)Zwhf6)ehF)mAw?!y7s-{vqpi#8*Cj))j#iKSgK%osTTY)o+$M9)$EK4%Ls8A7{w
zx~+Ttpo}}Q3y7g9vwd>6%69mK3Pk|atPk`#P^ayT?6Wh>p_YuGl~j>&t<_EtuF9%n
zG735mLEdiQZ4&izqL?kV4vdC|OAXw)2Cn&(6Ax{Zlcja7%E@Zu+NQk@l3qHbv#Cm~
zrzR}u!#G0+TA;^UFGjKvYBJMyM0DzEf)@TDL6vwkCPpmq_S@Vx;UsN1>gYzvB7!DK
zaKROc#uBN$sMe%{1oOZwotc!3a6+h+g+Xp+VK^&1k)PH*G9NBeTo3}g-y5U>q$Ts3
z#&g6Ry1+7XU0#uae)OTZ>F1%m@^aug(Sa*`DGTd_HTcaM8XION0Szg%3r?*1?Sfpw
z_Smr~Zd2HT+LOt(Fp>384}k7ar*H_B#klpHlu&@9C+6w7Sc(V(42sDHM`Aus=wYYE
z*5t6}T?^QAykoT$;zN$TTX&nIKadE!wK~n&@8%U33OC0Mp-O;j9m9-1gD4)gi4k8}
z;vLJ~q@e^)k>$>mb{%1Jss(PcaLB?-1oy<MVxSS@@6q|d6LeRrHa=br44uki5GrDp
z@j{^BO~DPgln#5#GAylqIaZk+Gjd8y=d9UuxJPZQ2ZeFmpanCIwbqRG`3~Fjl4(MP
z*U>L@E(pYMwOTk}qg5My?If_lKzi<I>VyS~SQ9PToinCg$fmxhQevt_Gy-B`ij8ab
zx*r000=*|@+my2r$!dLtTSbXlS@C4i_GrAFh`ZsIobaXUl-yt`3=D>BJHO0Y3H5uO
zzuYt~3eXvm50Sxq$O8I1ow5$q?x*aa>?W%v-K{x}>KsYtbnIssD*%UPSQTSj4EQav
zGds|sEIJwxX69hLvKS#w>F<?UB%Arp2Jr_J3nfnR$5fT{cUr6atD!NC_X7xnD3y*i
z1;&63?q}PCC8m7lhU0?H$&!F6!RBy`*hnGk#S9BmCb@N|{Dj0!LurS#I%x83?XQSQ
zNBUZQx<<IY&Qp$7GJTBfV_|1@Z#hlT#lB?l^emePG`s{PilWL8r45bHHuG{r1OFaN
zrT|q~VCnDV0ukfQ+~U_^0??0HI#8|@tOfKAUUic`nP=5xRw;oF)QCGtRcW9z-g>FI
z{A{oo4OeNt9YP?sX@Gi}W)m;15rBuy(w66FHmA94XOB~1kLPGVuXGDGotzcZY8voP
z3^JGc8_Cf5Sb~60&};icGj?cg%`i>m5Bjdo5tO;*){Aky8(5okO4lW(4y1gv=%*VV
zX+bD>!!aVi?uq4UF<WEQteq9Pw6PGbh?cY799mJdHyl}jxvA?^8_u)Ua=3<ALd!fS
znc4jik|MY4EJ#OM;$uXTmNDN#a@r2Lriz?IhFh`3Xddk3RRp7~<*L}gu84_|@jgHC
zX@=UTrH~c!t{QD7x`7O5W4JcABe7@Ug4Ga=v)t(+1g&>S4gwZhG9q&xgOEJcdCN@$
zR0p9!I5!mFGyusPVA_Ss&&rNXgGeIi*-UWlY1s}z;)t#Cm`qXnV56?GLF}w>Y@Is<
zELc6807)aZ8s?6K>d?~LxL~{VrBa-!Zai!{8)~@~L%LeQq(PM=uqG*j+k?*qRgi;z
z*Pb&4pX?Gia+9=QXB}(v9WzO)nlyRYT=Z-&U05l%pD5klF4e}S2pV@#3u4=zvF*qe
z%yH^!&0Fdaj*M)na0#^!n?4M-4{H$<os@c`kfhUI52U*&3Z`s&^PK3cfyr_y_1t91
zk}hAZS5<A$Rjt}BVhnql2#krakE4b<AZB{kGc`wR!v%pxaUrK@wH$LosW0{fh5{1@
zY=5>$ItJ2lvkGKX$SjBoV4}QAr>hJxIxhsU#{kdBVMEt}gL(_XY_=F#0zg+bo$-9I
z29`6`&_Od)7t3YL&RJ0@7Z$8f5?3k*nc|WDFmjfYC1C8+!M*uaeSRd*<4T$>O?lq9
zK)8B+P@glc%(UpGrJ^V+Ma68c4lzYuL7PcV>@1>2s)C!9RcP$1!m3Ffc!4sot0+1_
z^+=maagNUbA7HZq5kjEz6j%VN>Kp^p6<<tL4}-Ou+AK;Lq$dnVthkS7NX=Sf3~DB%
z)3<RZma!ah4NXgma^fl={bUyu#Bd5v*`+)lXY_uAjK-PcInmyf!`xyv)Jz#ebQ<gJ
z9OP02b}K*;;E?@^qM3<iI2kpTk$j<0;y~|2Ddv1;3|to!9Yxcr-%1l_=!s-ig~Ptx
z0B%Wbn$dD-<bb7AsvL4I2!Kc%B4JX<-Uy12xA-t%X0!D&$!W(}du0ovi}grw%B8W0
zS24tBh_6m&Jr{>qWH&7V;ih630G*fCTe=akI#ujY>}wR%8;6=PnwMab1`H^L@pR{p
zw()W|-86lCrx{$hi3mQamoThqWGwNs+>785uJc1@UezU0k(##CU}cD6IKQzfL53&r
z2H#r|iu#j%%MgXwiyaAytDs-IP2)-!zcW-$gzdJ+S#Udm3n|!h&RX$CYh3K+U`Ew5
z8XdCncA3u`)(+c>bu(bB)Kv>?cVk;hQK(uYCmGv|o-f4WT<C#pw~@)?YmGxdxN6;t
zlR4oKb6U5ywxMOF*%2`tjxD}t&PP=m+gv>NeOp*fnLVdXI$(h)MVzjt5u`0xcv{e%
z_z2Q`R_a@PG37i}*~{v92m*6k9N#(XF|T7~&n!p+200CRF<cPK9X%hhduvzBK@!5m
z$ES;G!Nr{hqN3#v0A@i7LmcJBW{D!p0`K*$e(r1gVHn!@P9$x3Z0Jn4CfZ~haP<y=
zooV?(*ZrR1Y-pmC(Ui;Dr7ePlGc{Lno|LBj?NsU^>%AxUZ7uXf74xGk?%V_*V8aeV
z9Gm;L#Sf`1EZJEa>^Mv^hyCed(m_i*v8zci2OMp{g%CFnWsuoqM@}uUvs5vF-zCWu
zCRO*`M(5T;*v|nxvcPXsxHD{=j5EU$ZN_gaft!l}lPNxF_j@^5?O+H5T&7!C?2lJa
z(=*IbzAMzuzF7Amf4HGN2Mz!gqoztEvlDe&yBn!$V0Y<efSwL{X+N;`L6S}F+F>vU
zf;6{wJAFRmDmYcBJ_qoFnhQ_?XzEI_ql}(l0~R3spl1gQkU_~UlvRu-^H2--iv`Yc
zo8<s_atFH|k`drpvK*EIWlA)ej?2w9sMC&54l-!m5=H1xTfx`DL4OW|QchlqlhI~m
z?uZz0GfkI$&1!m4TQwTJQ+EQBdi`n%Cs48ggJhkp?1%b_7*)M$ZEdh>Fz$~y7U_-a
zS;2QsDoQonEg$bnH5c<x@Rk!r&!;?(0xwdHo8@jhg*7k9M%$4|w3$&&^^}RXxIgE{
zfzAUnvZ6tlI_?3-FvpjDX2tD0xea4tB#HKN*;%Z6)7-bNIWgDO90N&wDC1Ab8t!PX
zjgO&p=2IkRH^~T27?9c1%VBbg^-6Ev_8l+T^;k9EZcxi!bYd*F&8ivEte$xMU^E-I
zz^p3#z^h6EL|0MWAp_1-IzCxj>&XNoKmr9j$h|SIx$CaUiJbs98kK0?n9ecY2AL=c
zNq9v)YzRzp1>DrUc{5u|MAXlx&SDIr9pHX>>W}uMVyZ|#l8RlYiYH8G<P;!ozlk>E
zpiw54y-pHY9uKuyr^0~$MyX+87`jF3#rl$-%)QCl2?nvBg1xo3MlBHc9GDjDOcDbQ
znOQv*o@#k-ib_tc>bcEItAU$$3KOuw&gD2arG+`~n|a-BFYsG)6D1>efMk&UaL!>C
zvI~1g<3la07{~(I06#<)0?$J?cyRx2HlA+MV(ghP3Ow7?-W$7|xg3m`M3_#964+`8
zxUF=6_lhd$GDUVaT2Vr<b*TkIE%1q^j}SF+kTj|NeACohGvl19kxLoM`XFzU*H@cp
zQnhoMGCOr+FSzY^yxDD;9V8f`4T%8mSZJjRld{vAfhg|D{>JSU{dhK|soqF-=38r@
zuhvb3%=48iTALlqt2ncv(F$_BSkgxEVyrtG-&Tf+4TIf5B?%sp$OgllF}s#xb+wo;
z>8A5A;^0qeNm^92CUAg7z$V@D%qFtru@*(n!EQpfSrUe6&imWm*zi#&tz+NQc!L=Z
zoWTkqh6Ud_sJQ`%I+R<;ob@$`4Jr`Er=|rt>iZzJpTYo>sIr;@u0`=p=XLUigT=5f
z?bE^_lzvhM3wkJSgmp2>v~VX^b2v_hLYef7`8*nEJ)mBBiUA2m`bypzdyxk+Mh(S#
zQ}vk!<VtgLJRBhs-8uBB-N+h990OxqA&CeFgaNUV>Co*#07<hw$osH+$VNhTtcyl5
zzQ)GUG@pqmNcMw-p}W{@H?y^)QBAo*X=Um)v>1>(ka*MeJT5qp-Xu?U!}Ul><{;6b
zU9;VS_JTDck8DfJXl1Zo&r)GTcaDlhx3R!)EvqA>+;2@}Y|Oaf#Pn<6#$1}L2?IlV
z0v-tphFtCGEwIPOBwCCieKE_^%oedlqYfssbqG>zX8>DSuNh9Jq-r4`pTHvZ-nw<m
zu!qGUmSmu7$Az3dT7Uq*I5(%h*QUL(;2{o#3V~wKg(~u136dTb6LjXRw84Hnrsah&
z?0i?a>_s3}HFkG8t#kx<!T<|>N#u}EYJ?oa>v>L2S5`5MQHDzQE6+>gQKGP6zgGY@
z#KFAoLKz_RjupwUg3R0`Uq#*01L@ta%!w95+hxM{ri-zOPWDh@!Gx}kJGZcAp&HAD
z14f&>n;wj_xgX##1-oE=*{NyD2&(VkaEz)fAylPo*DWAx?GnS%q^P^Cx$U^~eo+?K
zJ{bu)na{Vg)}to4%rB}vyhlg4m6TcoA}6ZOQ1u{XfnRtDBK5o}8R$yCcPUv+VcBRQ
zvjMdqZkH;&=)kVt^!f=6!b54!EUd|TS|gmr^Dxn+)>AD>xuY<$RvS@V?WJsi^?Ht^
z4p6K1=WvHHIlIzRyb*?;zEj9ugS)nxs70aCDqScbF9J7Ae-*FV!eDhpFNrXoPL`Wq
zCtNQ08!YM7ifw3|ntHwQh8Y;}vNzCGac#?jSTg&}G2vYRL<s}kWkX$qcceGjO?;#6
z%!(b!fdmG?d=m^0-9W}>3^C(%XlpPV6sej?xy11*b$b?#LXCvS8NAze28Ukf<)j#G
zouI?o0h@`7;R?V9g-IyLi7ZUu<L|~N;2sB~n;j>P{ei7|{#wu(aGPSyObQWLs#-`E
z#!V|goY?IS>Kiy_>f?x`K^#wIJ3Ozu!S2By?DyN1MXgJgLS${Pr#A2p_^9Ibb!Y*o
z<gC$Gvss793r89)LVlcUnH_cxsVOnv?c-D<nNa0o0iSMdiAh7;NWe=Y<Q|PCSU4mt
zF_3iyJU@>RW8eD2dIvYzUIZDIOtqo0Y(R;6Vd0qvx1A-ri70Wz_hNQY02fQzarjy`
zNw{8)g9ys^8ORVqK(1$RX;+D*%6&1X*RD#S`?bGGVrB|(vXcUD+xRFrRR9iBeSb+8
zR5e3_B2NY~z8nGPn6BoC6)^_Rg*-W1tw4gk3DTCAs5$Imkr2!pDM%w;)#j~d&e0i2
z2P*}^HY#Kl8v+|}L90-U0P2&Q4PGfIx=tq=3=E6T`M349##pjaRhG3u<Hj6_oiaRA
zMItcCBnana-7kaw03B8tNYZ0jp_oN1T&5BhrPv;?N0P(@a?Wu3QL~dM#bv`;SXL(R
zjEq2{`M@(bAp;{uzH#e%xgDFNEfYbwgIc<?|3ISWxKr(&z^r+kQ-&D2G?YTpFb=3v
zkXQ%y^~&Hj9v4U)4rvxMuBVue5{m{;5C;zZO6)@=NHtx_D%BS9qGfE}H2j1OKtd!r
ztu4Ny2wR0v#I2num*nNWFA5NQ?cNoj<OE?5qTYEkxh6cU)S?A|B$z?Mba}sjfN=JL
zN8oKUDmJoAj2k|kfQR>0srUhS<*3EM`%4R8%x&{{xfpX>plSh-RPja;4c4}$c}&a_
z(pPh`?q_)E`EIEWwi|41Q0QRK33@RTR!p_XqP}diqyEe#%Ys$6I)+=wau8I6yKm=n
zR5cj9dsT?-v@Cd!gh03+Ml8@|$`!#jACC!YNG3e8U@#hlZs4_L6{Rgg(5zHoN&@c^
zVVbwJ&^0qlSS<z%qZ*~^X;25Z)r?;Z0gLhMaEf4U*~<s>4dS}xxLIs8YbIc8t-wL#
zrN*iaxIeagCD()ly8sD;<b1wF+sdB^BX=;{ZfIE&c$b|Up_`&o3lFDFumOHb$}-da
zUTt(X?Q@nZH)C>L&k<ve^ar8T<Jx|&*$})<quLs=B&+iPF7_iuopW`%S&Li1l6_!p
z059WuAr4Ko-skHE83&@;rWrC`?9ybBgUF@3-x7Y`Da@TV3t(qB8rTycDr!HbY<i%W
zePzw&qimUxeX`+;Muo{11}-XL4;YBy8x3lKF&0Yg%X?L2RTvM)ogq#kIb{HEf#TYF
zG-~s_@mKq14D3Fp(>SASWUb~BaCP}PGl*t((E>)^SyLG7^v8P+>9d8aR^@y;+6&#L
zU*Sf&lcFq_{fwzrd%HlaF>Fqur8Uf@0;BYvIRq+&1sPXbn`tDV)l`x<+l5q+V++~z
zFq2P3bcW=Hqjqd{4VPBhXuK;sa8;>rNtAluPb;?cp)S>y%~~<Y@<_CDV&J+aM$u4G
zTYIhrQK|tx-tx+P9Z(Dmauam{s`!Py2`bhHu8u0GhI(eBAYc}VCR<_Ha-;~&$jmFe
zA*CGd1f!kj%D7Vq+@b|B%3`=-lrhkYy4OPVo<LU%1fM`U$eCR1lc`-~SQ1P&n4V|$
zJOx=Doy`TRnXMFFBTES+waBZ)AYEiVTazFX%&s+ubPXA#VQ4rU?DDCtTY!teEwgMq
zS=$AAAc5DgBsWCoeQ2(&77D1+Txy*noJ@D)ZVuV$u5+!k{tDU#d*F{_`=cR_Z+%s<
zvVMXAW0urQP=|+Y;$2{n3|4_8)I@822V03Xo93(wQnq>uFU{u_&lqM8?m~v(<t7BY
zUbs>#eM58}FgI`4lW{vlfM-r)R<x5_8ZvdzHL}n`J}I|)j~y~l+Naw{l6+tr%qTht
zDILy708OjBt%WqQV(?<iBD6wiV=B2Vk1TEv-kkx~)1Z(HW-#*!yzgaOf*X{Y9Krkk
zCf(9OqJS(vY1Y#^BOL44B8UT0OS`P$a4?)m%8V=4AP~LNM0Z&^Ik*8_Z$V0njx^?;
z)jU757F!&6>X$ur1eJ?Iu2RU{OvNF^hsK)ePxx7_QVwu>gJApICG=Hb1;BB~4;UXr
zV|KerM?4C=SE9jXm!`z(%N$~POe(8rKUJiW)t?TaW}$DFX0392mCAChV&*78QnMH+
zU&N*RIl%rfFhDJa6|MRy2y>CkbT`ELc^2Or4Oh&bT1Y6VfD9_YTijj~)&17(IWa0v
zvTcfNsa`_m+nojm24>t+pd~Ka3t)&N>p7tdLhKDO)0@{8vZZ-F<x4sDXknC<Q@H>y
zK@KdR_tyl(qK2#2`;iUas|noE(%<?B$Wl*f0t5nF#mn0eI4@NWasW#36>y+T)T~DG
zK8JRT049K7&Do+-^T8GeX{!+;1O2NH%fzw9^F_4;UUqO3GioMkPnBnuF~GqRu8aj8
z^HgsvPm&yC!nsp4a)Kw}c(Gom&Atw|;3?7DsVT#Fin*oM<4~uEL0U)U)CHqAres7<
zw$o^Ywb)+Wiqv>e^t!fs{Qo8e5tQHmQZ~xG<C&k<4jg#=0bw{$>*xNDPadAUUjLh0
ze_j96V?Xkfv*FKv_TT>S>Pv6EYkc(4M<4vnW2sZ#@{Kcpdd;7h7yOui27HzKY~Xpg
zVDRtC$1l9}`ezaU@ORbz(bv8BoO9lP2zA)=esS92&-%9)z2+ww^x3<9`UCQN?Q<6I
zyZANm)H~~<NB&ju$<LFQmDk>L>aBPE=7;22GWhqY-#_TEGY%a5@E098<nad{;OTAF
z!Dqec#uxnK&rdq#fj@MA1Aln%n{K@2u-~b<?#XZe>=5XRU%lWFvj^D4|1_M-kKm5~
zPhTJ4>3L6n(J|NFeC|09{#_^MMF;%;oxj6m-rt{Jo_f&L`Ll0$sroyv1$Z2~;!B4<
z;I;eh@Grml@aLWVvtK>wZNJ;ncaOj5n7?@T0~M9uHh19QXHO2GfBOaEgh%~H(F2u&
ztMl7`_hrAW0h0EDOAmk7@4gn#JNtqof9GX?(H9<Z`4Kn&Oy+;@cV3GF=;L1ayD$3-
z{tigmd28yazx!H%E%=}ZI{E((I{e=)_}@?2|J{QBJ4NvSyDfOzGY-C|{N^8DyuHJ}
z;i<g~U-p2Y=f(KpZ+p^(R~~w6R=@4_&=sf6?*5a<-3Q9I)kl5q)Q{+2dGD{Uxe2_X
z^pFPvg5QQ8|Nh-IpFo^X{SVHZU!DKuBTxT>m*3ZqQ=fatbK4ib%)Qw^?YVDw`$ztJ
zY5q?D_HpvXCzfx#>ap*+q5bZMKYhd<pF8(O2j35k#~pd_t^aW5lW$$${knI(_q_AY
z8~#4`>dVf%=O(3z4u2p$J^!&sz4P7gdHuiNDb6pr?fv<$etG1%Pkho-fALpeeK>mk
z6V?Ym_<mTs<u`|%?7ro7XTHzy79Uzo|M8`VoGrYr_sa8r`IBcp?v_LU{6W=$W6wDB
ziw8ZB!{uCc{`)>KzUn=nfA1ZaUG~BsJ=FL}@G||9Uw-z8n;&(}87JI7>~EfX`486@
z{q+;xaparDUmSMWYd`nxla4z5sPMYp=I)0adgnnmJ@EDp9DCcNjy>t3OHbTB^uyo(
z^{sFF()_#MB)-~Q^N{t``E$Sg$VVP>-!gykw8N9He_(Ld9p8GE_Pw`$%6{S(FXQfb
z@6$j0k_RF_UwqBs>Q@&6BY$jh#pxIQdWu|kb}PRIzM_%m(}&-O-DlkKglYSpq4t~h
z&8n)s>8h9g-p^nE{_}79v93giJfMX-_Ypx7z5e2>>@d3S#LI`zHGcE%OMdaG=%ROH
ze*dEAVGn-LeYM>3+(W))-0|1R*FO0*^{HPxefrkt|028g$Jamctnb`>&iCH<h1<?Q
z{X_44%+>YtPKA#9!y|vO`1F?^bo+tR&N%Xm4?6vWr#!fQ#c^+a`7I}a19Cic{Lx1~
z{D1kW=~K`D`~{KU|K{v<*M0ZT_C)>7hh$(Hzxu>S9`@U5Jj^}r8&8YB@|(}Tdvwtk
zgTo(o#1D>p!v&9iKyrTE6As3H{;$O+-hK7EfAV$gvNs;{2){WNFtYzz|0TEI`QR(B
zy#{&1;XgX+Nq_bC%12)HwX;6{FBkvhnB#6g^s|5VmILjJjywO*TMlZ^xMX+<f63_7
zGyV@h-+seS&JdWt{}<vbkA01a?B8|C2kv~vMKAckYmT}0Y3II-8r;VTUU2T^&;Q|9
zue|HA_qad1^jE)<{-nO-mv<fdid(+^le5nH{~(b^9sACAp8Up7VEe1@y5T>z-B+G=
zX5^gxv#b2SpWZ)TH=LNC@Wa2oySV15?>YG6r$)haa_*tm{@ov5`{P&M^ucE+$DMl1
z@qd2g?FXGUJo3#)zUjzU9{P$0^m7Sx+&kCre8YF1^|G(rb;Avp{^IPX|M7=D6`vV^
zroaC7zkl9un||VLcYfnN-}~GT=(peYPcJ^<=tmrP^SO`u`+s?03t1j~;iEn?o1dAU
zG<wWw%(q@xzD0k;KOJ%Nzg|P%c>e~!lzHBl|Kg!P{lp#L`oM#~`K*io^7g|H`^pU$
zf9+4Nxcst@+<<@TpoiW5pwAw3%8MR${)28g01j1r>`6~K>2=>c_m1tA554AXAG_&o
zf7%uw{r-tZ=BHl&u@`a=yY-ZlKPqODnmqH(r~c_B!`okVOYaM>eaX>Z`y~10=o=rC
zR(x{&a~^%+O?QxAKYje;)1UEC@v49L*%6OD?}TG6_<iAw^H0CCKK|D?-ub=9kr&_f
z(i^}2<flCMNnbej$%7NFf8!0(Juf)troSqxKY4O+=-(X=Klcxxd2fI2?^gZthuw1B
zJ3sog&%gJsJMQ23yDQQ24&5Gob$;Zlo__QF%I8%75OMv6&s}xqGhg?rs~$Z6*(=U@
z?$Jj*=1S?$@2X!pIOo=52haY|8PB@yi!Xcm$-k}q@=yK{5X)PSA-?6G{j{f?^oG~n
zbs6#fQ}4p<??A_#ipY0A=-?k;slNQ#AN}#_O*j7dITzi+ed_}K>W|)j=4YR9(^3B!
zif_H{#h?8w`;@0&_LSGiulfF?fAfa#U369a%=iA{&X0ckO*dY0=_^lq{C%9^a^kQ{
zKKSEjeC^`rfAA^E^UgR@xaAG?k<CGerJsD-hb}xVIPy61w$B~(9~J+2a;W%I>AT-J
z`Klv`=3BKOOuqk{x4!Ol?5US6PI&q5*bgID$ZtMf`J)Ry^{i_?c+#2Z6(4qgeeNf!
zx8FW~#gD%Dt5>c6`g0$6;gi32?K8gdm+w6I`G;Kf{>?G3g^}lc`i8$e@96D|A9dT;
z!9v{qqs<k>0KMXxb07atCk&2_j<0V0O?Bj1w>{ykC!RI@FRJF2?3Db6*(<OAhYx)7
zimzV(53l?8AId-d&a0n!)w>t(dE$vDekMNC{$B+9gmaqYqW2niJ@!49f8eN}9Q~L_
zetmY(^KW0@^N{n>{ey0M|54|k{>9_}@JN6lzlFi-6ZQH3c>eIxU%&M;5cb1&z3b>N
zKkMW3_75KZ@W&r?TF+VjjQ!J(z2i5<JKjq@?36d%`1WhhJL#w|-T0R8zw2zAyWplv
zE<{g!-$f_C?(4g|9{H}LvWq@=RQ4}#v%vem@BZcQ^$&UJhi*OX!x#PV*Z=Zi>0$mq
zez$%zvVX}TFZ=xuf8*Id`tb2LTyXmJm;LPSm)uNT^S<gke>MJ0dHeBq9{GhUn&Zzn
zRRBLfc>Yho8|`m+{q;}z(Al^C>et4s{osk8zw7oNeEEBSaM>?kcF8Zk9zX1lZa(&&
zS08ijv%db~+5Lp!1%G|{h0+_pd;GWVe%ZhDv~y3n{fNWg_@>+bas2(Kf15z>_TT=A
zU;Ag@dD|bKQ9blI^nKG@pSnH2f0?>_;idol@h`mjp6ioWO0T%_**6~b-DiKI{NZ0b
z?%$eE10hIneBvu#@v2*@hh6-gyX~+2{8!I?`A6Pz;^y%7$XkqCU-eV*ySM$=Def3w
zNRH7@$!{XrJC76}efmv*_WV=-`N}_g=N%uGzWTi{zRx)uq_H3W=2QL@%)lZ0kAL#b
z|NQ2EK7M%mg=c>0qFeN<pa0vP(V;<o*~7m3#_x&Gq2Kbc%Z@nm6|X(^y#JB?{Mpa{
z$#*_;@eA+TUG)4f0Pz0d-@NqT--7q6FTdxGPY~``tv^2Jrdw}*=HXuwj{f$o=Y05|
zKKgDtJmEiEZZAIh@zcLO^E-fxoO9ZHAM(C`ddey6<sUreea=6B;>X*I(m&2G;Q!*R
z{q@MFe((1Vy|TtWb>Wk@?>q3ge|^$H*BtSV&wuCNee??X_2)nQ<yYso{n`JEHhhTw
z_Rf3r>-BFT&;R)m4|yYf=HI^lSszV5f8i<Ft(V<%>6H(=Ieo0(U&lvXb-yCN?SzB>
z;NokqeCAi*d$IW5V{ZH->IvWY+@Y7>dg>+j?2o(qxL4ocz~f`y@gI|t-*d|OXTNKW
zzi{=<r#*d%bKg1IzVe=*ANJGo4L5xJ33u%u_Ytrq&+Hk0bip5;^1>Um@4h{K{*yob
z@zc)y&Cg#ze*BKNc`yCpF-JYgdB<yB``r(n^_EMYebWbj^poc2*KNsPzUG=EzDj)T
zFGmI>P}EDlE&pb4`yEH!^2~O7&rc57kQ?%*S8QJTTaZ5W1;@VgT^Ie;-4{RhE9hH3
zcGIJ;yypc@`X1unPHS&J=ep}=haY+59UuSbHvsOuPdMq}_v6m~uuFdPi-;`lJm)Dn
zKYi!zKloSu>c@cp80Y!dU2>mBe^C3pC%ovlwB)TX|I07E;M-S1|N5w(S@PmFkGuQ(
zueko|r$8@^{{1DdM4mUj2fs^~xBu{EIJ$QJpf`Tvr#%c8jh{T@%4@ED-Ak^go^tX@
z*v()0bN>S`eCvf1`;yxq`$hea-ukJ}eBp-6uK3xxe{)^-C!Y+jdX71L>UR6{w|@Wq
z@BFvtzV=Pm>93tV;>0ID{cTs={AFaJe^t!B^3W@<x%7n35YpA8bJwT&dtNNO_LP6T
z?2wngjBkwQ)yJHCviSCA{;hj;{+8EFUX?xiqsS?b`RkK^eb?I``-JO1{k5C^eDEIh
zj#q#Ee=_#oQB7{!7pU!s9ubbBB3-3->7A%3NEZ;0E+wJ&-V(8ZC`Eb)>C%;60}9eY
z=nx=4=%LpT2oUl<mE%3<{_cC@jr9jG7z2|1?Y-7qbI!FBNT~TEGAV|g^-Xr`b5y3n
z$A{k^Df*cuXJxS~Tc~6S9C=B?R8>>x$Gu0&16<PJUhi<LG+d+9#frmkY&A8X*o?5`
zusk07aC*V7g2G>+1dP=+(d8E_o~Jv;Z(E(Oob)$i1wSd3Pzay;NKs{}r4<z&GcJmC
zUG>yYU99F<iW@`WsES+-j*5-zhs^?4N)o2o6hF%d(QjD=p)<>7Utw=cQ!a<V`!Lle
z>+kk25bwi!(;#PBq*|n;VJm~g8(SB)-kM8mIvwUbMd>6s2w3B;bx8Wii8@ZMdR5sU
z!wB2+rKZh+I<UP!5I}GR>U<<W!3|6-pRL=@6DE8&p0TQjD8_zlnp#NvBrMYSw3LKP
z?5VtUm9a7{V}NOqGK35OrhfyQT3_jpkm{7QwaRts8?qo>ANCga4&uwiX^KxY>ppM^
zQu!ZI7Wd?plC~-nKT>?Um`}K=K3`*?!O1SDuBfJi;QuHyQJf9QYG34JW+}9^O4<Vr
zc^1s!riEPQ4Ly=Pp*C5xBG%Pel&aTBO)o@O21Z6$;(8z*dkWn*%VO(PRiD<<=h3bU
z3nkMY4L$`6Ha}AEI5^JYD0$}L;9`XayJmZuB@Yl>N_WT=L<*O3!102P892pvjo8D0
z)`RmketdPv+OZW7koH4Tr41jHtB#Muh6V>a15$$K7Ix1xn{C%$beXrSG(6Z+;(jgH
zMjqP-_?nIEQ-(0BM>e+0bNg_u)^k|vvK_I7MvrwegrjzQ$JyLaz0xAp9toRmCq>41
z7AUPsbxmWf&vYVZiG%+kd;HcFdfL^NW_Rfkd=k&sbmW4}x`~vGY&<3bgQnZ?>)AP#
zO13He0()AT@|H0WjJFs!7rds_+-stvsr9OINGnC4j;t?QS;fW0Lu>i8(}RCuLw}?x
zyX&1NUCXA+nd~Yhi}{i`#Yfrp_WPew`>tM}Oul%JG{xyy(O~SiAzLEiXLoLq2}kMa
z-9z&eu=-m-p95i4bk<lJbYq^n$~4dxdU32{8})RkGr&<l#NK9x=0jhjs;nqg^n03O
z$K-({eg7)`uvG~M!Ed8tBF>L+@9u7JE!(VQSAT3-a!@U+QB9OFe&AyxFBa#1^U=ey
zuMEfdcq>tn!<>YM3||Gsd-q;E)kQtl>*LLJg@n*rsCWe6$)EFK?QN_gU32-{^In0(
z8FJ-|KJ-4%b+I7?g(2?-mfR2;zZM+g&`pjCCRE4>8Ft6extPJ;Ur|G3(@f7CSh2|l
zc{QG0D;Y(pF&${|`N2ABiW@RnSXiFX($bQHkZ&z9K6{};Y+~GpDQ7`#(2ohk3^O0S
zRs|`L$mWvVv7Wk`m>!<~Flk-L=~;fVl=oM`gC+OqGTLhhB?pNFs@;t-Pzv&8H?eV3
zK7yXGUcJPPR}GEm`=T+kv0+&}yPtwjyFwMz@uBY(kjA!HBiHJ6`3=Qp<-{zb43tT3
z43rLgEB6kaNUcQfy$!};4EV^2G!{%4qE-I^C5S;-CTY?XfvgAQw@e7q@^N8eYV-;B
z&zwsX78OY(>;q`=9ckjZ_wJBx%6iA{E56Wt<*a4Ryvb_Qz@1i|bd{4+&CkBsU0Y>=
za4?V@|JZbEi8^iur`D`h*oK+FG=KerZ++c<i5hq5;0@L?Erha_Z1&Ywz>gw+4Z{;#
zA_ldc809gRRi8O%b?@+h#dq1~dE4y{+wN^q|7nXp9(()UOUTsN8abhtqoo}+u?l5E
zR<RXKVW}dRf^?hlkiM(bOy;FNqwcxzN;STej!58F0y#nxt;n0q&8_4P?!^2~8vxi>
z&sm^LyYf!k&%HKfE6{!Z{6z&!TRYV7=r}SN<liTW9kI@|Sd1q(4|FL)4FQk@z%0Y?
zv*-8X1c0O1)1Im$!M*Xcld`S5b6!YK%`?tsA#tna>C*M<eQYI_Zk3ymFRM;g=tl^w
zq8~=w6>b|>jEIn7=aGx2of^DAO~>TlgVJ&xt<j>Mn+w=bW$-!oN9WX)rAPsv&#!mM
zzjW)GU~YM%96IylK##o>kWXTh$JE)74E*V)yoc5V!^40EqgHh$W`C`#Etnp4+N8v+
z8IL!wAK{{#PQl8q6`8}z&Z?@dSeBOWzFPu^4eW>w4k5lXm5A?#m#3e#=r44>+ROY|
zn}lMuqlCH+3hK4CwzV(jF=snAdR-yKLOrXFijTaGl7~_|&H9rsLO2ak$LY%Ob&R!8
zGpkK&)gvunU$<L>bSYBt=5R0&yT^oB=T32OvY=YuZ|dh`f!oJTkASC7W0DR#2D5|-
zQxlg3VaI+1JUSLDh4-B?Ihvu~K0e&=U8A!HT6?tcHS%B7!BomQ!SJ;h5nGt~b-vq@
z8sA>}`s{Ps>Ao>Z4nj*fV5?J30O?PrsVUCw<02&3ufaL)6^#QNw_Y%swb966n{R6Y
z;}#l!A*v}XqNwVgvWju;73}3TUN=n5uN!LR(RIpv^1wH+JvkSmWS3@DHE(ZY7tN%e
z95)bTmoR!&>#D`UH_of{S1r~sgDfr!EfU<mNr~Tbx{_{be9>S;c<;`LNF1dg=WdJ%
zJ%#Y4-q{M?QajY0EQ%#cQ7YbdNUy<9MJx$AjB#8PZfx-wRwO>!1D^Q&;Qgco6)}-!
zq@P*-L-|LQ9R)=mT(x?Gor@!Nj*c}xONsOot`BVJW9?O$QrV~=A<hPH{DG^far)}v
z!H1_4MSzlzY!Z9!^sb{HTkKbbaFC)ww5l+0Hn&}NAH!5IEhgAOlzND56^#{bMu%;B
zWK<?4i;oj_=Y*H8XLcDJz~uZJ=n>Lj-oE_v@L<<gy3Dk?gPK&-aO?n!AlJAxY4wtB
zN|)0BwKay^1Gov5!Y0tQ%0xE=g&#}U2zFew+yn6ohTT*MVQ#(hpSFCwgyQM&c+630
ztO~&vb!2>gkLk|Eix>N|C?yOl*U(jnneXo#cEFXBtYS6X6{s6=FHa!(t<6ste-Swq
zVCuIQ)>yYRE{4}WHe~v>7fvwwrr?=Gnbnocm~q`a2v4GoAiT!GMy>NmU#NLP*u;CR
zXG<Jqexu~s^U{P-F(gN&=OMH#+RD$&F`>pQ|Mi<7Z=S|=Ppdr*bHU}5zqps0$r5=i
z@M}6t_M>(7jpKa0ag6t6LiE&xf-5QjD<kdf*k}|M=}=3)P`zZ%a;pg=sZs_7uZvsB
z7SPxXQw^lVNxRqFbJwn*A9sK1l2kU}J{OOtMON6y%uh&pCY22hNiY<<r%CGy#aB4A
zRTgI=GM6JObi2Vx6M>-&cg%CbT+F7|_md0hT6v!78n8QG{37VOC44b_kjZgW6b0?g
zHL16@6=CujGYr?zO5#h;S1$A!D~DMfqT|Gj%KhLt#nrWPU#MDuUzNx}bcN+Z^#tOe
zTVI!xy?up9h07v3(`C&iY}EcJ<cpx4UEsj&Fn^XWFq4EYBMyPKC13*Mzg8E}02y`9
zNIV+VG%&HvZ?ss~^fuo$*gnLNjE!N$U7g1nnVH$2K0Q8j=1lz!YJg?5r~UIpsNR(Q
zO*?bOH>Ne>!qsD5yts#^R9j6Vl8O6OUU@-A1d_b$M}SV;IlXgx>(o}0`Atf<1R96O
zN(IMST0?;}-OXM*R)MyYm_Bl>#S{Q>8vI0+Db^*;7Mfb3;ev8PZEwz36$FX7ZqFz6
zGzzul2`CA{9^v}COzvFXz3gI#QLNdY1@Cybyf)1k?Xi1rE$ZmdS)iOfhUo7ED)zzk
z1qnKa0m1X}+0Tatvkg=>tnx}qs$zNE1%2E%eTpI9yp=7{UGLZEO|<n`^^FPKe|rHA
zE1868j@`7NXV33eZ4M3&K3>4zV4m;&LL=y~Y_(NIM=_C%sc2l%ETV#h#k^=b6+XXZ
zHDV{*cq~giYzB4iDfGmb>S0%}^cEW%A5E7EFfI=}-plBECj98-OA+q$wPT?}zxSs7
zI2mps`n}w>T_(7r+O2ltr9J-f#&RjlN9x5tdYTQsskfe;4cKb(vRZ*gA5YT}GjVEP
zqP3&rzXb-pLA1d}SOhe_2zqY=HzBp)?V6nrhf1KVu%lTf1L8!|F`%3$>AxyHY5VKr
zo8A<<ZX-kr3a{G8SGRxWD4Pp`w9T7{_fp;@`%fGBgK+a;=*<?c^7lzgn7*MN6CiiK
zycYARI6uPQgekh-7MU<Dy^0-SvbD9b*uck{ryg~j>C2W`o7ZMW1vpX|&{Vl~45XV2
zvH%^(L;SmIq*IL)m_6Y^vCiPLxpS)>mllq#feI|>F~%|b&_<n|+W`|HgQj?_XqqR1
z55B&G)Xei4pLoaBadhxvPK-YyUvFAd=%C7=xIe?7)Y3MmL-R)F=v->DL+tA6suyoN
zcjOx5m?)=+23t?Utw)6KY|JSWx8jmxS+uaK!pglKbG3;b>uCNx+Vd9`w^|2p(B2s1
zYiw;j|D)|S+&_+#@W9Po5%k%9ydsW*-gO=pALzEPcy9}%69Q*RI2!CaHH1OZ2C|>Y
z?23VftRDK*?GDTiS>*>Rf^EA(*_4fxeT^kx+JcRhV9KXL$;MJF!qsx)=#@0T9%)R(
zmNF<ktyldm*?$Tophqw;d@3r5U3sHh^dALeZec;m#ibN2hC1Bah_fEZ@!m49k5$}V
zZcTSOL0OJ>elxO<O~1-yn;0TWR_sAWp6&8YmR`L~^nv`dXZ`On8Js`&bK5+Z9=Z6D
zY{H0A*mtB--BaHP`qkBywK&}WVtZHx?D_MLyoa#};v-e}il{HDNLgz+O>{`3O};I7
z1AjI)V7()aM;MVlCTiea%5N-0Gq`%5@9<(XvyW0*mTY8XDvR?)=c}|g#v%13@*E;7
z^AK!;7VceGbSj0id!;*RaHN+5{2iPzKcpH*P#IYDnqjX(bcW`s#!U?tWmDMi6xJJ9
zmuAmnZ#-5pEbsEMNyNSLFq|nfUo*vO#Op2&FMbNF;&f$y{LJf<u&{6k(Co~21-t6h
zg`~;K2ZqyeK_Nq^O6)5D?8Q)#sU?{Y#ouJ0o``HHN;yQ<&3j(vo;V+_5~o<!s(A1D
zM?xVDn*K0=*+l(d{5lcok#brGR_9vq0B?tOOkfmAGJadp<$zx6dt$<G<5a!}T#21;
z+7Ab(^oSDGe^hNBv;1>G<vh~};$g$Cr`Oj4nT8qPU(x0jd3+BB$SUI|0)V6vLWa3X
zDfZ4at8}5*x>zwYx4q7>ScpWzqa6{A$19DM<ns$8WhJG^<m6<aKduCQpy@{Izux})
zoPKD^VW6Ay+-NY5DTMhYzCI@*tFrg)ffmbe3&4-sI2Cb)ql-JT2!=5=BITv=x!g^g
zF!>UaZjFZ4^d-}%31wB>3>{BH@80yBx3;s3EMxEKSZW4U3bsh3R(L*0&Vpn;R`P)@
z(Oel`ep!sFO&gOOMn>c%=0ZRs%{SM^(DVpb{x)yP5?k~#^lVVhJBZ$;i}p3@g@vD@
zfi&<TN#c)%r?p3giQf(=RcuPyT@E%WUACw8*49yz6g13a9a*EZNvbhONcG{WRLYWL
zaob2}X!axUns)lchS8zRFrlCnDAulqz`v7`FnTFB*wpcVJy2G+0b)1dTI^`)`U)Of
z!K5TV_GE{&w6x?h-x;G*JzS{AOgi2mt?%9hq&9yljeEb=s40{6pHkb)mWQp1h8BpD
zZYKMs>VKQ1rb5s8^^sd*Kb_BZ&E;#~<(AM16jcu)yUfHfn-dz_o0>k%7oJ4pu=&=F
z24r5?F||-T(bXZAFp|HnlTb4FSlqnt+S2ve_=I8+qb?@~rlI(0$?kVD|Kb#Z$JHeN
z5uR>be8VsTNIW})6B1sO1UtZ4(+ELA44xiLeA39QXJiJdtmB<OFZ|N-yvl!TdDrfI
zQuv(3nq~05e}oz_WrEJQ>Ga|(@rCnmg#=I}UZ|1{l2^?~l@ui<u!0Qi1e@LeNLyZ}
zEo+$zEvN)P7TM80_%J9syEV2<OX!}C4~xz%E;fF5NofvX++BGO`;&@2t%zi72;3i0
zNWnP$#_s46X2U<}<K^2Pb&Ru?v2d}KP8|N#wrvHemQaQ$J#{;5^X?aFbK$rI9%cVh
zg?0f5w~p|(p3W~_Q8z@5YTY_lF9KahuQ}|AP_lo*V2H-6uH)l?(|Z#z!^5wu>>>rv
zZIZBSD!!GTDchwkqYdD+5us-HERcsAw8l>sZJ{<;gn99>PcrcJxhk}}qz@Etp1ShZ
zxKDkN9Y?kP{-28TrUOGef5YvnjY+}1uslWU?XPuog@<_1@mBb@Itiy<Wc-AZPE@5k
zSz^b^#L#nd8F*9|-c6jyQQ^!1UFN=E<QV3ew)#8eh|SLzHo_*~7F#o|m)0QU{~<U`
zMXcQWihM9yI{B4+@0;1ufEZl4Q(G&$wNeE^E_T>K14-zv9G8vpEU@Sq_ID9NRa#&p
z)uMCrGnBSncIKDg*ha=hZInkZ&q+{$_eXB$K#S*-?)n#oBudY~+ve9I(c5dW4VpT3
zzR&DctcWvcnG$pFLGoO-T4RBTHAZZ(<=4<}I`VCmof_^{`goXYyTw?UV_5yT5q5BT
zgICDV#Rdexe)5YH6tx=w(9&YY*t;)D0m)t3(el8+7ggh5VCVs>KnXS@jRdtc$tbHg
z4y;LaE_?=fJ4AVxlWVk*l~XR7jylCnv99S#*xuR#EBsO0TgN{0|LFURcK`*mp1I~b
zs?Zl#VGHFKV(W78xt^Mm!V5b@W9FJELDBE2`4cxLp@;T_I2eIC%?lhx)6Je3{NTYa
zNtN+(K%BA3?a>;&g_(wLnXCe$9XfycexA>J=I2<)`6}G^436@E>}N&mQ}8?&&-5c|
zhmHW6yJXn0Y8i*Sq&^a2V08C5RKj{iw!^-gJQtD{(A+;Zm#a4cC)$LAWnH+`*>MV&
ztq(ouKV&3*U#MMb?zb2_>gZ?w6a+w$_Gy{6F3o~ka0oq%)`D7e6Jx-hHmUQSlKEVJ
zwt{=oPITg_fw2ZqT{c^7$>2j>Hk-htJ)B(<O*eePoVbx^pR~brY5PyrB<*O7i0<X^
zRZbf-mB-7ol}CEi#k6Jd4t>sjf*Ul*msnHlAzuzy0X&mDD61l`(IH%8-G0+{BR+9i
z-~1V(qf^9^M#{Zh&!LLDwJpX~FDGp=R`{02g0#p0BygvE;%s9vhe?5}lm7y20oKZX
zFvxlk4zK;|In?xbI4is$eC?vt^QO+uDQ^AiZza1Y@%EjzFovjb*@JeA_sXiO_O2$5
zgsVcyRpsaAPh5arGy!|VB~g1~#~sm~7|^}*Y!tCgrzrUy>kC(}k^QS@08-NW=v{Zp
z@Af@bbL&sViYKDi=sfgIBoe0+6BFkE36-;2kL5dowZ97XG9*PXf8y%Wc3M!Q6a?=o
z!XN>qzK&L2kP$M-FwY#5vDAQmS;cs_^VH<GR@`sy51~=1Y7LZ^sNHz2=kc^^dN#1L
zMQ@hn)~%1o9peT!BM<xI?$z00a+E_}vge{5I#t41jbcsKCWS|WR}sRO$dxW5XApBQ
zE_b44$)GO;T8lIkR(10p!i$O|tYN&L<QLe4O!I-Cm*UjbZ-78(bZ<8<h0Hm`)_^Ss
zC$i9VSB%GSrEtxAxp(jC4Mw&GDl>5Jrkj@Tp+^s-8vk)y=D+<_ZggiJRsET@?~$IT
z?)9wN?<h$1D%Ewg{jz1r1xU36D(mTqEy9WP+dq@?$<YoJ9Q-zl&H5^6YxaKuaQo$Z
z)1h}pkG6DHQq7h+b6Maf8+=P*?z_7#UVFgDJdDM9+vDaU1f5&OcUB;td)${pLtSQ?
zIL&XBIu_4Mzz#Aj)}^iBYKrQy3GZ{1A3t+Snjcz0hxf7aS@NZFef|11P+9SQo*!{k
zEH=TmD;oIxnX#J}H~041i-0l}wjsrdVUjZMyn?B!>X^APS&*w8_Jhpo=J!A|>MH~y
zUS_+p@ZBCZmsqh`nQtnVb6M)T?y6$95x>A=tC2Fibwu~a_Iw+fC$^{1PxI=*uJxW@
zMMZ_nbChWTkI&0#^Po;Z3&~b^{&8T-UBta~Z-p4?-&bj;Y(|0$9_G<pWt=QSu6)XK
z-W&EGzFK)HErpx&)q#PZO}-6yFIGlJ=f07kPTYw7ifmPFzP^WJpUP`76}<(X1kP81
z6n16w3^m<D+^Ep$9ys!uuy$Zrr`1k<#gKccBldI$+-Znj^!bx_U53)YjV*RWwH4>3
z@#_Q!gzS!g{|jh_^5?+Uf#qn~@r|AOS6CV8jR*fo{Xa}cwN(^Phjf&5CRtZxaaAP^
zw-Gmrs!7xcW3Tf7n>E65?c%TV@`K+v5w_O;P_}xVg|jyANLMg=R!sS;xc`0j7)~P@
z7i<D|s#R>*m0;Vw))KMa{q)x4#_ByckWC8?b*VAjhww>9ngKT^k!Fe5+^o+}J^YS~
zo<{4&Yjr&n*A5J!wF6&sm2n)J;4&ZXcU=)RZGA=~l?_fn1O}d$sN8rBwi}q&vgC<3
zixsJ`<aCw6&b#Wh8tUpASi%#<U6Vsnh5M~xNWa*cofnIO8MWD2dsSfFOSUOSMPI&X
zK*c@e!&YS+jFj|{F!?ye-OqDYWMebXvEfRK6-G`@E)iWp!y9TbZf#(NB!tu2wmZ8y
z`m=x>z&LI#-1JOJ_3go{aMtS$cC|um$S+^|;(-cruB%$D07(8oS5N!49ndFpL$oF`
zv@9*&Aq2KA>bbU5d^8qb9#%Wr<y6L!RRq?Q>Rnhp^RIOS@G@L|O8<WB)29;x*As+&
z-qjy2NMcH&gJ_=zu2iZje?O`0Da^kWCo;-av1boneN&13!}{v+oBxJSrzCreDdw=(
zI~ORcp6y*%*5!?O-LmfyzOv}AU<!>`-B*e0D=b$5*+lpASmO_4mpf@|HfMpGt{2HP
zm^A=C`Kh)}643}T&^vNVVK4Q=M)<K6T(^0s{bDmxGb=s)w%R-&e;sq1p!(ayFt_*b
zLIr8Y%-kfj$o#9d{gmP8VYBXjR|+Pnhw>pzkl5<-xU$L|HNGQY?ucHTS7b<JA=&1w
zCY4st+gRf2A*MdnQN_$-HkCHjt!A)#|Jl~go)C7(4x%3?#NK1DASvT$SD!e&FcR`A
zWwnh^^QM%*98~G=9L&P<-E-gXhqwi2AS-n7p3yZi;19yn#{fNo0r3;5$$|!C9};s}
z-Fqc6(_4bz;<n-32ZpSz<?2mGY930&6x7zX#&5Zl+3QU%$P_`U1MKn2YYG?}QHetV
zn~!v|pOFV1I7u<r9o1n8UxV{O=MdF-nXKlx=NFgznz`ceu11WgFEV75>T%YWa>dtj
zrTNlo?qJ4{PI*L&$;eSQM7p_eU97r94Z0+AHiGD7+^;7yEuLDI&a_DOALjAnSxN;m
zrHyOK#$VRR{A-EcYeA6_qUEmqQT|gLfQi5N@wF=_C-L)z=t-`Km_}o^>oJvf7aB4T
z9;d~_8K8MzRWFJ0j+aX(C*2TlrfGQOc+#cU8<8=V({vBjLmDmLtx(y<z8++ixU^$O
z^UW<u*!SfiYPci#RW8!fbD=)}9aq`8c|a{larnB59R%fDF9G!8imyoZkJ{58lwH%{
z$W^V!ksw9!4wC9RK11cJ{N%}pK@=)!V|8@Ojmd~nyDWRjU8Z6uW@nV`ee#5KZrh{p
zI6>vy+F_&1PKj?+5bKvA;+||(r(bDV41cq}+Aq{HTCT0lOXkVstgk=xrMUKW&sx6=
zLL~=dourZ?rZ~v6z%;cn1QEB#tX@K?FsucN`!bXlojH5fpduGql5-C!d~=e!Uc0fG
zyzOXeZRhMfa~NybquhU@X|HO@k3Ech^Qlz+u^DJ?gh>p__!;rDF$fg7@;uO&Wu%7V
zM8KE2_}e&S`1r|NawSGC=nOkB)B>?)^zZvgO#C#{wEH#rJwCDG^%_fmRJMj5PwNhc
z_P7?L#>y&g8To+Hb4s*=;Bic0Y3_ni5|j4UIlM*=Zi|cL67KuHKTrW-Ov@YSyZ~9H
zZSIJY8r@*MG9jjx)0JvfT3zvdR?Yhj{e-6>-?3NCgewJO1RqsScvSnRw;mgfE+z8`
zrDJh_xO;97e&vJ^bc#IHJc<RHd7nw@{=_1Uxdh{~1pny|yBnk0%2|$@-(*tMFTbg)
zT6Y)ZQA?His()iR!qUj7Az8}ZIE+br-s0F**F@y1O<)t0Gr0BfqNK?XAYK;pdn&8B
z#)=^$a}imUL(oLkCuJR+pb0lcW0A3n>y&wf&qMBk+1{(9TDwAggl(Mb*eY}7>PxxH
z>`45AgDl)@p%f5rJMt8B>T91CntK}|Eb+dO%8r+B;h7t2OHvqDP&@I~-m7s+7trJS
zd-cEEjZ+3&_td6EIP0{`)~U39c{9ro{l1|*fpLfx-p7QKj@0Pkt(_5Ly)wMm@aTCK
z^62)|hT6r|<n#Z=c*z&$us6yDn^{?Qy>DIURC4H;FVkh&vq`-+cRkUTg^fLaRnNw)
z-_E+00p!%*L5|InK5!{1+>f;|spYC=`ILxUr5O=6NTO{Vw8$Bq5O&p75S7WLZ7Mc`
z^8tvlygP4!&1h2i3+yF1eI#3-vqo$DP+t+e*tpmj&NoV3l6YV%!F;DV^uA;7W<bA~
z#C<^L1*76HdXVZjJ>5cjbCPxL+URA$GT%x|eR+BCm&F=FWTA3W?Pxw%feJ&6sq_8d
zm~?gYLse+)xHV=@*1l5FT|4yq1O9YK_5J4amBBNn44KLVn}%|zH)<!l`g6-?4Y}OJ
zql0Xmh(F7Nvac0ij3j-(4{C^s)L1jfJ$RK8=2R4wOO!+_h*SJ)t$-?C|LPuL0_`F~
zdZ!(K(g|&ceUoX@1kT9|Ckc_0a}syqZ{zBtu;7rGx&71dr+%G({&!LP-~LMEIX*)W
z^H9am!a!t=@_gu!qs$FEz@~l!sls~lXzi)I9Ep<Nu-UXOwbRF^0sgi6+<DyHG}SR*
z>r%CobLesZht>C|<Jaz7uN>|#z+urRhx<TyLHOYH;(ta{(o8@St60{@)f`UXUe&)r
z?%}g~(KhnEXTm};uYY-+yx9Kss@ew4?OzLCyUsuSO&WSfRn><I{U0+AtoHpcf9?y$
zkxBdyec;4(d?FK^0;Ga}$BVd@voRu`^WS#M{$p63`Qi=qlj~osZMaU>&`Eu2w@Mcg
zF%Z-hcf~X!GBx@%*S|=!|JUiB&HHEKT>0BmI$gB3PZ+zMuljEAc7P)231dE4U|IFw
zj@9LKL}uWd*YCx7-rFE9q9TXme;BLpGU~U419bflZ}sZAlV$EbOLzM3W&TUXLiV{V
z3*fT&XLz}7-RmoQ(7{=4qqvoZf^yJU%lRKS;2WNv5*lylH8c73Chm#d`6sKT&)R~I
zYN-ZNEwtaF%dzrS=&zRm{M-Mpr}=-MaZ}xFK<$B#M3TQgEiE13N7jvP<SJ?STuaw5
zz5Y!5ZnJv`mG6maSIZFKIo!Z%;Nd8``{zZ9m@9diMUrE3-iJ*4c9+(sO__fXp?4X9
z*YLmJ>ObH1f4<)=_L>?)a%=k4wO`{B;Z&6cK6A6mw{0TgPt(Aa3#|BnU;E2V=uAQz
zzB93+@rzANYV>W5>AuRIa6VgpR1k<5zSYJ>TbR)&I4k&*g#D)xyoo5dhZyIl`Y$v4
zFLQf24dr;P78~&VpB)w&?f51Ln8vnMjeC20N5;pGASYjnsmB4uENmj%Kqpe^Z}oCY
zWhxt```E)FmSAr9v_1>Zz%hem&oDhr`W66f694}@3vmi)DE#Efp!F`A`8N5$+MkSh
z*z8;<>~GCxXJtLqDK_NfZ#*g`XOJvM*K4v18CF><ah+h{#hbqy_f>kpj+%KsEUKyL
z-WC<emY}@bc}S4(i)DicQ4*U|#Up;GapAl^O-D^f?LWgxxlqOL-p_c>s*gQRz+RFz
z`NAL9depunlFGmBH`ISNm1*<n1(r-2G(|YwL+J(GzL&-{!BNlW{k>*^==h&>u(CD3
z_5D3HD>HNR(xFDCJdOD{h_nOpT+J?{-graP{hlr;<L^zB_*;q^pjec8a?rx^Ow1I|
zze>ruZ^wL!(g5#%aF!Ub1=#riBOd-cBmA$QUhK7;^(^3;YbzXxy-UPbh*^wgT(^fq
zj;mmYV%{r7!wIvIk&!Iy?70Od8X(+`X`t_7d(Dacd8+S<7jbRKWe#^+>xJ>O#*gfo
z^euk;hW)k3jcGa3ru4m=fO83Ot~67f#+?R4mR>U4SDDkeEB+6e`V@dKZUFrf@Q+9r
zIlZ9pGunpFb(=^rS8+Qe@Uw-c-r0cH6i>wU{Y_&ZHV_t)wn4{jB&$6yN6cw}nW_3~
z{=`4(b|pYnO0<1r(zi_#-5y|`ZE5ATV+lpHITH4HO%PAae~xj4kV&0Bnc~1T$^UT8
z0Xh9gYX9F;0rk9$9305F)q2G6wcTH|dao#kFHPL}L2EeEhDRDAjvvt!aJFu74VNi(
z8B(p#mdY)6LfX*=^c%32Zle%c8KvJ0%S<+oBgfgluk8|yt~7Gbcs@!EVguYgU-zH;
zdEI#-`;l_-?7<7!YW5eh8R3h}2XBEd>lf~V{u$+Le{V=H{vC58|DxntH$z(5Wu$2d
z4@}$G*VMgkBkc#N*a<sogO!gt;uiZ{mIw1Z!uQ@V0Q1NXxHaTYh=2d9UY5Y*7kCfi
z8vFUfT^bpym(Ln!cc`pwu5OiKz@s2^Nypx1I_$1&FWCwiOrF0_f$sDI+3yHiV0tYC
zy2|tylL*<{7KMKXZ_gz!=(ioJRHm(?ko|Zd_5)hm;#~P2o9YE5IG)VIlpo61<mD*>
z6XRK*KYt#A|Bxs6Y=;Z2_vqoD&qBmatXMlw$o*={-|4LTu0_*@;NXxiH8%_A!&a?B
z3DkVJORazWv)lQ%MEie!nktffqHoNFEjhMoAQ{yy>;bc9g&r7=jfpy4NKfjTM`Lm_
zy`qwm&HsEM{0;C2rLJXNG&AN`EuL)uMMjmJOND-A0QXcsB}-AmKYr^arC;IrEnK+H
zxu54C2>Utk=R>|LTTPleZf5xBDv~V(4Fhy-!asruOW2xwlc}lakGl2j*NjaOMT6>b
z8|+Ad&_g=v(K%F0lQj|)5|VPGFd4Zhl+0g#6?_6L+#-L^-h@$*Ux;>=V^e(c{6hzs
zB}ST99JFD7Zff?}Q~jT8apNU0bQ4Lqk6Pv6tznfIm2i=f((QbQ$eMy6G%xxev?_1B
zo{Wr+<T%@}#Ud#9An{0l<zpS5pOY_&A2j;NDY@})FTkBQghweMRw<P4)fah{VU-Uv
zW;0S|%Ewo{xBGn4%4=K8xYHs+OqWvL!RY8{EqcOKBs`{G4|wY~;Dr=d(-z%`wY5G3
zfH;%z$NA>^lU-JhV^1^S_?mt@<z3^>&YyE-l<*Jlq$-1wYC{vml@??r4191G+&Xtw
z!(gV+FMeN+l{vX<Se138Ih*Avt;OI!{7NsA9!2^u<<HHJN6&n_%Gs#xpvFW+%{#^v
zsTE~0HTCpjc*tFZzZ0y_f4#)*$PwNqkQ6_zYc^xXV!wGxkxuxY&pEP+Wy@k>S#7C4
zeC#t4Cp2N~w`SSj12&ZC7V6@sw1n39VvFrYtBs0{Y8?1(UZZE}Y1j|R%G%uvJDg>U
zI7)h>8EwMC+<Lj{Ai<m@?s`yfH~JZ=$bZ)NoR0!QW)ACE0231x>DVo#&nqa1x-WOW
zvU6iYlJseql1V0Of)KDZbhYgxKc#|6HEixU;;=1j0+imfSZ2hZkQSd7eQJk&CpVX{
z;{&anyJjeKmQ|X^Q1-Z)Ejaj{yARv^x}2QcE8O~oL1$;DzCW=IbgTj0EikF@r7#%<
zGH6NpwfU3HGtL}vNe??0XA4{_o6UYX7Vb>!vV@AYZvZCbkJ!`Sfkbsz_9j#61t3+)
zc@6{x{&4HwRt{s3zFnj1=q5f!FU@0=gG?|}(OBrPquT;oJ0}1SrGO6*E)TyAv_{WY
z&l@A6C@}_@e2UrJu$Y?nmcg*+FB(2MtSmYBVfgBJlR%?f-Zf$=VS2k`ms)QwR?`T5
z8{U`$jIDgq8M)Gp%X*5Bm8!8Tw!KhXle2lL>-ZqK7%5M}Ow6<VJ85+Z@W14uStLdE
zV~hO8;%O)B^+^X~k~K~Xk@OwXY>)o<&ah0$+)5Eevy-D};WpQR>Y4s?v0!g;g%~#a
zq4%$Zn&*ijj3CW3XU<A5%V~ov{0K`r2y(20;b`3j;_UP;rN~!*d><sQD6c|(@%7Ry
zxI2U9>R-=o)lqAv73%X7ToSdU_+)WEeYT$XTLTMf&gjO~+FX--<8EJ@FagR#Ge6I|
zLPNnunOmS#Y3`n*hnJCZ2HMT;$IcTQrWN7*+xE7{4`F}>7(}TKW6FjLLMNy`k_Je1
zXkw$edM<3BaS4;k<h}oGzuLfCLCshiqCIoMtS<JtCi5CpwhyR-CTvU^Fx*^Rk&9MX
ztUZHdUflX?nr>yfK(iBEu^#t(WOU%6LQDHwoIr*i$S#j=pZZ_JWSinE_kQ^5&=n?p
zX+6tNNAgBrvh^f$MCF9XHGO1coJI6!XcA_kO0qT~C~%;!$Na!mD!R36sQiSU6dFE(
zyb>ON(_q}YclB8D{h3ij=~5=LXut3PX7VBHI9Ls<#XMSUkLk!`?oOD-Jr^-kNtRS=
zAx*kf)z->a`r|9CB<QOIqU1Y{*s<juM!QAnUo`N34eDX*AU+!A^N1+K4MiuX4+a!}
zi2XU2#9PO@r{AU;ffsxv|EP1E9g<)hPqsYnxQ`?1SSu#@ZEX?ho>}szTI0T&Tz=}X
z4l2mcpGP!G#ZQlia4#$cNvZ#^fv%VAb!@%r9=*Gvpirkoxx7MWfr_1oo%ULIxg==c
zZBgapiXYT7KQmmE=`*T5zuLJ2I1XI|LujMW@uo$sQ7cEY)}YpbW(M`mKu_J})5Kul
zf!0<Xa(I2cH1SZZu8!i!YrZ^sf>UHqI!azW!V_bNnzr=iZ@@+R?N0|4;Z5qjzSv>l
z4b9lCHw*&*s@XDEN6O7#9j!Nb7G}D64{m;b8(X&jO`UGv`W;p$m$<)l1Hcp0!KVNi
z@#Bd49`v%f@UGO_O;j5>i1GVdgIN6A?<q1xviV;<e9Vu;&0~L%Pc8+xO1+G=t|eE9
zwPc&><%?Pn)lx9I%K&MRe2;Lz&nq-OD4}K1AF@@XJbItlFPL|w2ZO;Z(c1(bPLZIi
zm3V}X3^z^l`H3i5tKs?y^j^9C9c}*2eo$d_)-{7x@SzSso7PpX$3`irRdF>>O~0DU
z1HWfwDFo*Y>PXe(k&xp_rCW|^QoIrQZJ{Z``V0ICE73IFM$0$sG02t1NIKMx#qdj-
zN*<+R`$3jScTNNNv0tthtasHoF~I7%e11!7o08Lq#Kni5j+lt#RYM8?RkB?8NBiv-
z*Gu4$PwDFto}c|?D4vHk<{qOXj_1q6&zhhF4qHke*@vYiChFG!q;fGB+uqTfn^l`X
zmaKpHn)Yr`810N`#S?9Cj9%CfTqVEG85|RVhjFZJjNq+bX+3hkS5~-$pWAs4#=VYM
z4_cJ7N;^34F%K-WE$o&Vz^Ycnw9Ow_s-;Bnli=U|z_{(anyWYGBmmHw19xFy?m2P`
zCTALu6iQw>8n9^$>Xpqpze^AGu_Zk)NxFDfma+5bfgd*D6fyj`ZHKOhisL6%+!U8}
zxH0iTSpr+ML8`~gbv(B9Pn%jlH#9WFwv%+cQcIlYhegDhKx;yqhWFs#$g#fm^1*-e
z9#=UUwC`}k8dSKrxVxZ4s`KqHS^^eEvu4M(@ln;s6UPZZ5yH6RB9kcN4-60NBtZay
zinnURg+azW-4VQL4z7Z-q4ZLQIX!7?8hs62Q~1F~Iy$;XalKkm@iGHRQPbJpFWsc!
zE#Jv=&j$jDR`Pqc(N})#*k!^L$e_!~|LSb^t&-kot8eAbf_uTE!;zs3G>zcehsy^g
zEd1CUz>BKrZ`4p>H`PNj&BiG&VOtdA@?3>&T5w?zfE&ZyVjx>Hj&vlklr&g1yOqrs
z0i@j3v4a$W>r0WJVy?$5c-7IU#Sv|};f+3o#6r?e?m|2@lt4xEA+>+`z~;HjDy2`6
zvsP<+77bOFWzXGkEQ&`pH!L+)MoN&EFVTM^A&oa;E*fey32!l><Ui)&=893JI6s^X
zGP=`DPcCjB%4B4$TSl{7-KQG*Q90EF7#HD{?+pt9nGETggi{X9<*((U2PL9d(@kF5
z9c*4I;7#*^et8`jC>^;~M7#RwOPR3Hczs?hg=gDgS!X8~VLOU9(tRmEFEw#oK7{7r
z@xG@5T$;%?c+bit)p7se=n=tNR?e?uhi08gf92GU?_M%yW!YwaarBiY=7}TMe&@FC
zd<C^mmGoUJEJ>CyN98^JhFX7ZC{uxwlb1JUb90j-4MFYELXN#u^pNIcGwj`)N^y7W
zh@GgMPxcAKMF%%0!Fbo4D4@@pDQ)*nL$UY&Rq$VxtjK}C8;w?42UV=*(55Z$BwBZA
z0|pfawN*{SXLn5OM<6uDngd#MOKE%&Z{MaC1%?YW`yXXr8!x9Bfbb@pb~zQJiDCRA
z&BQg5pGAttWK`Mw;{^fWc6^m}?w(>l`HZ1etf_0mad>WbVxe{|G8v@Gm)yYf2I=la
zU^20bp;1FRVeAMxYK)cOKspJB>JH=Fg{u?7QRulPDrt&4s1?9m9*Klu?&wjMv6}DS
zTB7OGQlpmycZ+yzTj^C>q-sH`b&e7r;x}YDd_5paO8TWd#L<DPSC)>R?5wBMmCl!~
z4Qtk<o8B&6O!!044)<5o^<HOJw5+OG=^qY<enH^313vuvwM~`u+<Qzn0uI+OK1aun
zl%10)2@A=}u$-KDQzIX!@B6wiy@ugUfo91<$}wz}T{&-22JOx_n1N}6f|TTOtv*xj
zOVX=h5*^B+-6Tn*%RK}X0?hMhCwbHXNaZUjDlQI~i|`_t5p^L13wy+M5KQend3k~j
z6Mlb~)zDMYs&LkD`|iH<BHq&ziX`8AlQ_)>sbko~i!K)L=0Zi!l=JT~g+GAZgoS``
zWJK_>DG*z3dk0p#(?>&PktUIrf5s(7voY1zeRaJ*;kPsH+S5U8G6LP4tL8O<8;aUh
zSGi6MykY9=j-O_{;>CAzm50|~zvm88eCC=2uF?&mk+VW5G>xnjX0qulV_20hjtrnA
z(!t$kKl^Q2DMrBSuoys9x>I1Jb6wbAY%YN+DUQCOYRi9XmDPMWzbH=xYjE6O7AYXz
ze7mJ0r3X?tx@s1fG(5j$y<TC6c%1QOe~CX9vDXHR-T@g~mHDlYX$k$cHeQOji5`3`
z=6dW#SYK(-VK6>!(+2vKkz~0MkHCXk$<aLJeFg`C!d>+>ev?8{zh6mUB`YFL7(p>s
zS?O)-KTJZMZ`KWUa~G*7-Af`aw10_eYkyFtViT@YzXto1r(=#l<s2u@+pklV0S<T5
zgnME{)9Ha7@J#s@QV{j!_#>hNT4i2BKj}GWI0lK=V>^2u?>RGI>`#FEPbL&~IW@~v
zl8Bpb3}rT1gLY3DQb<bglq?1mlCnR0E6)4Mz?o-fq087JY)?ZhJB$co5)F42sqgl@
zTVII*Nsh^;E~mvs@zWhwDsrtqy;-(V#xKnr5pku|5e4ly!oQGE*)^Sva%H*oBrXqH
zy>-vuNUn^gJLik2P;xix+Dcvto;rzUotsQ3IqZ?hSK0mJ&Clq#6`Ap^bB<nh?_qUi
zshTEIrxz%3ltejz={8B-P;$85L8)uECs$i^Lhm)Z22_%&GBO^RnVJ2ZYzw0kbBM;K
z!Qjo<?U*0%L`hjg?SyqD`%T;X;zO0AZ=aR-GTv)~%BrtuZZ(9S@*6L?q(u7olTiRz
zph3%}E;3mmrXS+GisfjWJ)$U^lE$h_ulBX2<xEN<;JZ6`?>HAk1xnlr)p|8NfLypt
z=KsVTnG+8G9AZ&M-5pJHDaW0#-_elB4I^MgnS&h~_csc%?lwNFv@hi)h&dm4C6ML-
zy*zJiQCHD9PmlIV+31q9Ia^pPGR+S>biXGhI&yB1v|_sw+PC?zj7H5{B|${5(r%m$
zRa&g<^|1PAZXO#7%&D!c1P}R@TDHleog2N3OZhAisVRs0W%#>HD}t4Dpb?hr@8H(K
znfOO4B^zT*f`OE=xw^thyXzX5AzYSyA<bELtykH4v7)z{X7{Uob2a%=8#v;+jjo6s
zYisM!F{f-R(j{}f3ZjS{slG|aO*yZT*^XUl6qb*W^`8ZgzVA`8b40(IP`2bp5QL~%
z`PtPS;$q%I8U{q^WPLtmQPOdIz5lfcR{%g43+Z~Ng$#nj*GO+c@79M`c8TSgo?|50
z<DJ~r`kT~D+&v{eYCe0xE1Lf8<L*t%XVvU|!q<k3Nv^9T{nMVrTwX?PZEZHXESaR3
z+*t;6k1PaPz5OIDCx^|vxWVnX)Uc;BBID(qfy&xlz`;D}@1sl_w1^u0tjmv%?19+6
z3Xi?LqFLDy2-)EFdHlePDeAiYVmrB_nz9Spq>y6p(UCj-h)7pdYNc*kHxx72%i`+^
zeuo#(ZK(H|PxL!L_2?hqm3tI7_7~t26efUw^6)zK4aDe151r~DxX``c3eC#K-uYqr
z6}Y$8ZRyH*vvs|Mi#1JoHz_x9Yy48rW2<MMtT&j{Q!7H+!+3`3>uk{uKvU+WER&Xq
z&d|0WY;)$*Dmv)sNvq<OSsOdPqY14Ww=%$I+hxA@DRjt}8!L&_ic|rGz7MAY!5VcA
zW}ZRS-7A-`U*jm=QCGm-kCQ>kmaIiYrEn9x|5(t>PbVJi3HgIqi(mzKr=k-7;?i=n
z)w|X<c9C>G>jT*2?uimuMzD;VHQrbQX<q)mRb4k+p&JE<dgwTgC8-l8O$UW!$3hzy
z`Ek@Z+6IX13{l&6uWDlHklB)=eIozQfwj0UBQD26C}LpkCA!gZqE<VWN3Vam?(N$P
zocvW22A)-yg$zAyHI{#k!-w<!X1WJKcW*t(2}!;!bLQ1UU2z+m|F&{Sw@Ny#2fc5l
z2V-Iz%T)ry=>1G;H;G$J4@0jWxJY^5D}DHISuhW(Ml7uUTtVZo@X~tE-N<vxK8#Mt
z(8=kuot5b8U1hasFh5_6abLq216gjkKX4O%Epcvs9vC)cP7_I*#W{KiC7VkukH=5n
z;7Q<4sH5x*m<Z~sxgjf~&Ze5i0X?>F!+LUa&As0hoz0-VI+*3bvg5@qWElSOUO2rv
zBVqF9mf1`O_`5P9%kzO_u8)bhtDY=zMZN&^Nx)?_OC!Z5F7Uu-v|?8(FE1~m{C-}f
z^ex3ZK9AluHkuUd3_V442FL`3(0otT?j5*1{%MQZ0-ur4&&}0uUz1;uZ=qGTq=a+=
z4#sf&vWFibwBmIMV0a%jpBCC-^><3-ZHzRcemK_)&xF(%ObE)8$L7X?tjT%oC)I=7
z;gFpSMoY20@*&xj8j!0<O32fJj<fM7rT5Ay2X%=dU4nTGgFV{@^QB9(NQkxSL@LZV
zkDNiId}u3d4Qp*Ux)KM=?8T9kv3ym&jV1EgTUxCh<q}b<Z1SG-4pvstLb=_vXAAHN
zm7;6>Lewg8rkQSp=OhZ(MGpQ$XfdSvQBPOgi>4+r51F>@2`=8a9#mEN=zCI6&kz1;
zpivkx7nfEU|4=I5yc^rd>)Y5NJSvGcFWOb3i-s&&C?K5>I<m~S)6)Y{l)&{w_^nS6
zUF3^gbVJvrF_S^R0(8??>+RJ`>U29P<}!{LE-|(R{Ia#>U6=6-33XmM!0j<^lHukA
znJt$=8WaKebBw$9AG+x|uN`_zh@tJn^)BTgK19g<m=G(PtX}SK`g+y^H7t{{^C?%g
zZLxeb3cTH=*gqucT{0U+^funVwV-iuo%jy=-S@&T2W>#{Z~#ERyNL9}Uib&IWxsbk
z;vRH#<VRS{4DW)qg`=S?xK8epb{G)_o7k!Gn?BX$1n*UfK5@46QvBw&d@n1~un`Yf
z#9$s?b>_$4EYVtqQhZ)bua29%?opm-y{pAOpIXC>3_J~#^=$CjSY27>1~VU-O)m|h
z8M~t?SMGwb<VQl_g_kr)^|m@Sm}{x)9Far}xK7!)jWLOo4c;oXAD|$>C448;jg8sl
z<QFxIQ+JDtE7dCf;RE&UsO&RC?OG-ntB}`;oZZqKhD29IwI%E|^<ls<6ISBUu_ecj
zl$~V{k9nkwMeG*TY)|{HATfa<H+g!F<~Yqqtdt!~`Awo4N#IO@_jv*jFP}YxcSLV^
zj0iS|rU~lYu)YDmby%9)z1uf`&eQH#K8#70Q%-?G07XtgQQC<f<S0E{ktv1k-FKB0
ziR3leO$K^s>E2_f+@ml%0O@voo3CcYtuUL)G(U{X9a?=D#wLMe!QtNJvKW^Y*0gLS
z2B^f%*7pMg6v*3Tc=}j0xC*&6)RMK(2G~*oW9yWKPk&>9w|uB?7;zfn0562WUUei>
zGt%&Z4!^rZRpb}p^En#1(_q`#7AB=&qe1c)iSzjdyAR_E%J)u-d<vzvG-0c-#co7q
z2<9Fik9TGZYQ{lUr-)JJq#+(}Zs>76E-c)>jA%EzoUOB#4r`2!F^_#$UR7|cdikCI
z0sac<6s(FUzn36r8`(8L{w5%xdkr<m4oa0|HjqV6=9l~tbnZRm?m)=9FqGC#cO*9L
zGx(XCo4SYTQM4<s?xcz3z}-6NIef3cWX0Obx*rxM&Yvbq*P=;w`2!Y;FJB_z_cvS8
zqXOd`*5Qt}en4T2DG_z3D9v1wh|Yyz1`ez}Kt}o~xJ22FngoMe7za|ZcHKx@6=QCo
zo(K~<HXScSqrBECqT4XI0+5Q-v3h09v)XZallboxKL`vh6_Iml2@-iV>gciD>Q5~^
z4A5DhR4t>`UAJ7h+RQ|r3dc9EN8@_&LZh7F4AOmtMsG4`$lc3qvyWkob8tcPW8sVB
ztGA1H!rA?M;YE7bN{557ybY-^b`-wXc1gT{%R;cA7T4yFp}(sPUJ(Uo)C>f1tjtz2
zZLd;`sh3!R%;`$D3w)}uJ8v37woX>f^bs@$Uj5nw3K46in+@H9gp2*W6K%7Q%#+Ij
z5m8o&6|4PQPaMW}ANRbuME%p>4(Oc-;I<tpU)mB`3ZZ0OG0Hc8%XT3rRHQu(yP41T
zj**ay3{oE&P8^9pEFQHo0aEkVf&9roHx{#!Q%sc;CY12y%n-}=3$Y}c;_r_|E&F()
zz<C8?Kqquu+o+c-5?FCsamH#UUfBF4kC%<{&L#gnRt_$X*vqiPE*^U^|H-Ebbyk(p
zvEoUuR<Rz^uV`z8nG?W)TFqlZm$LHn$ILb?inmYQi;F|=AZ17QkkH`Z_Yv1YCXg<h
zgHZ?Zy9;5KH*PWq1D9P$r}<~;rMe*dD|mA*qbF-y*5c9}Vq8JVVrjh1hHsnu(~ZXO
z)t<|P<PWeC+_akC<K>J3%fGvsbEESzbHSM6v$3ZiUUs%wf5I4+MElSVA`V#j){+t_
zEP0bTx7g@Xkzcdf6hz{zJh<vc9dNHTw&z^mXDK|PTzQs6qp8YYSt!50Uz5>C4)xf{
zh}{Kd2Vy?Of8t}~Q1frPIoJwV1&w?;Q3AgFs9IOFgI77Ph?M!HaMVN(#srn<K52;V
znW*mH0dzQ_C(AkJ`k{~gqYflf%F@%nE+!>PuZ<Q<hl;fcI3S*gJbFquc)=1gC>#xA
zHi1vT+w|9W?^hNXqcobw=w^*ZXBd`t!n$x{t@j`y-|Za}!B5aV4u5G;<wIz-kY9E@
z+Qwp+wj|2tKKr{4jZ`i~GO49T^nfPqDBQtIpS~Or8@ey8AH*yV<M}<_L6nDhD?JMZ
zd@grAcGZ8W9mj@j)OhaJE`Ug9Pc1fY8F=Of25451JGSTV)tz519jwjTeG$Mp>{VLT
zmYUKHE)TJ-Ei7nW5EmOI74D|(&DkmT+_XY!`4|s%I>HZz1!D~$dWVtyYwjHIlglQ5
z4wW^TKc1R<W3K5_H!|qgzsR}0p0vU!_h(9hAf&q6d6wf>qw@mw*KZznTZK_T^^VV2
zncoQXc>#hnA|~7XuR`*Wl8IV|r(TZ(Mq07VXgxP^15|d8ttS!pWJV6TUj|&;HGe&R
z1*s>k4B?OOo8QWVJU=WU?dle&k=Ati>1IP9u7~9{>o$-+a;rW84aO4KcDprUdw_RA
zn!lk^v;TaqlC?&#-|~|nljCKuq9Dz8{e`F>I-1Xp8oSW=O!lSFD-SM9)<!2gBre%6
zcXjm7O|d&AhXWfvQK6I%{GN~MV6c>3MW;$8k!Ic20M8)exmdFX5l0bt4$Uf$)*8Y?
zr|pIc_#cx_E2{1C^4IawfeyaKJGu>ofOgUnZ4*Rm{h)SNC7^TBeifz#HL1f`PGqTw
zN3}pfMXhnv<KdVBHbkV16lxGC_LEeuArXD>M%qWh%>yx)cp=pn2U-ystUD~dU#B1#
zlTZ$_J$g~StM2as_>H`A6O&s-l?Hr{8e(BV_G4h0C9L-0OV!rA;jLE!!>i+#UA(n<
z1l!4((K{4@5ydlksLy)=V4FZXv6+Atp<*JD5Tgu!xKlCyU0pRaRMWev^QIBSnv_{W
zpI{(5HsNyk;?n;^*;mI^wLfc<q5=W}8zcp!OF+82K}s5=yF0h2bW2G~cXuP*NW-Q}
zx|_|0cX{-jd+vMR&*yjFfBY<P3$C^1JM+vl&&;?sH`J#44Dp;g5=8>6c;BhosLatd
zSLC{SZ<A>z15<Si&D)9XNzIs)OZ89t>V!6fFBOwm#?JA;F(-vC&|MFX-RJ$ow>b-v
z;e=zGw>c+45t2;-+z}_GyMl+id0?j$(n<T<iG=io>(&zmzF*~pt#~_t_ZoXguu<S$
zY9tzNZ)r&!HNDIzXj1Y5?x3eZ6V<eD-`_I(oDEvEI;dpi!ZhZypsQcyx2xQecFwvt
zu6DD*Oc+KxB)~k2UVJ#>)?uJmzE`<h<ip36ra0^fQB<BlaVJXOtUKz<y-42NHA(KD
zQ2In%cElsWcOf`))0--gUom--#3PWdRjWbnC0=iXVUi^)Q53AfX(Pyfa3Q<9SKVgH
zxVXs2x#K%<HkZzCb3T=-r!(5flspZ#uz-%IGeiS*nI>ZJVFdWcy5E@-xDS@WPUNB2
zadwXK{K!2-@`wI89!wlO7C3#Cb(=tlJhQnl#r6*6<_v;CNnv9y8@T!ac?C@7e)dP7
z1I%duXE=_De}TAVo8OnTH?s@H5Igdl8CSL^DF1H1jx2h%p*I{S$cx{-|1&=D`}%{~
z5l?$zW90<`GO|*IQRfosyM5gXJS5zGb|sZJF~%wV#!F77r>yI8tYcl~hM~Hu`z`dB
zOrU*YEG4WQY49>6(CM&YE6uB6rTgtYjFPm2-z))Kc5&1)dUK=&l%*I0yb?+{CFfT2
zwMRx2sGz+3EJUI0fO^4VD|gZP=1y(wyxc)F-xj9T2o+WEucLrzMZ|X6RFr>D!B+@R
zPs1<83tc68`h`JZ{gRuKr0vd)K}X6BJU}a`&ZI+eJ0Oy2T9O>Hciv_H>>09rZ?r_$
z&Ku1Sy++3*0#PBjPArWIylG7Q+*;|vs)klgL7RP_cMt?g5*V!MHOfES_;?bGb+aXy
zYd3y~@W=Nx@(D&(a(r0o=|)vbX9$=GnLqErDt#{LsNW;^q9Dj=V^3dFLTRONSR}Ud
zV?s1G?DMr!St}=B?Uc6I1}o?Wop@b#%1UiG{wz?-ITO@9>)Bd;8I7B6Gd}*Q@#0WU
zJOsqz^=x7!oq;v3H|tXJ5D`^+Gj=v}jPq|TK*-N&rxMb_?mEX7OCm3?t5>77!^0!$
zmX-v>OI~8xmW-N(NK;li36?WOiHv7OKP^C={dWd?H|&MX;oFmA6V-0>@GQebj1^^W
zieyw%wa}GI#HUhtQ59wA#0T;@gK;dpZI*&-jZo-iMsAvXHba%^$d=_;VtkwOmoFA|
zGWOQNb-=T<Y~2V-WvFd&xX8t?qXdsbG8mSW^&ok<W5RJOiCMa?X2Q0Q?x@b*)Y;j#
zuD|)H?*yB~?(EN$Z%DU{W<H}0CE@PKq=b%#TtZxJ8I^M(c&~%L)EIIx3P5ga+$&e2
zAL@+pX4{M&(n|P(w*1}ol9)U{rLKBlO%oP#uD~A)DXjZ_&PL+FA`<JT?O$79oi-zv
zKTS@Gqws8PK{2fN=Don@!h{1x6hGcTMruvvjBzFU(t5JLwjpA}&$GtzjAVU<MWMDi
zt-}GocQwV8NCJHP^;t8LUS%4J#`_w>@0KP|-urbHa;;hDaTeYv_rW&RQ^m;Qb0#l)
zg$53Hq|ntA`AgnN624qOFK5xCp{di05L&WV8#)Scwx<026#1z=M8ZhW^YWa#L>nZ<
ztS_dVM?w^rkL}>vi#JLkzN_ki=6Ov?Ng13k?ubLI01t8cfXB^+^KF~6f?SPG(u-be
z$p5YvBkO70f*eFqo|2~;ib>4x{&66dUS3q<=B*}YM5z&VW@JTsc&MryhNJz;d0LMv
zCLW<I;f!?-FXv%yWE@4S0}oIHhnzU?%t?oFHth@BfcbpeDntu1y>8IvjmRG5&HA`M
zRN!+)CT4l}!@@$nWC_)6GI@8RR~c5xBL$h5A~;GRB5t-zxGt3|6%JJr64$S&^Tt~b
zpgYc<I1CX|Ga1@V=dndS?oJ`^=IRP(3aP}{#Y1t2$9+?fQ!NQc3d`!%XsE~c;+Z;Y
zjrbM&>Mai_s$eTY8V5XexrOXSO5Z=``kLC_SNq};VpES#u`=0ZF1;-i)_ZpX%5!^j
z2@t27u2bnVo7*m}KK(f~PBNFJMOt3N45hn!7yBIaF!eL!8A0nXj47)wrAKlH62nFv
zq|hr&tl@0cuG|?XnGyE3670p&4ZUb@V1Z8X7ozV=d?U?z))))o{;5wxy$IN?gt~Qa
z(%#6($Yk(qU(pd+1=lg@*OO;TB(2LT%z4mlqaDoED!PPMr1rs7%TTrSe)>{UQB7=~
zCG31(l*GB93*nou$?h5E@~<%QYV=+uu+eGJtk#e5U!rL4pG%~8OK?~fP<HKkaGO;V
z`S~*vq``p`IWNz_<hx#^?;_(Z^~@mGNm8v*7LIguR>^JHG1biTZfBVETH1HjMzp+Y
z4nbe5nX2Mrb&2}*v;9r(6;$uly2NF-?{+Jk$As7D?r?nXBT!10&~wYTl*Jw_;|eDR
zND-Dd@e`Sk{PCIe6?h+^BYl7aPJ;)$pV8Jw^Pb|JfwxK1Ho_({u`#|0qnM|8*Sb6^
za2T^FcZhdz0!XBW{Oz7!F|H6Xj`&$+$moI<=*;?15}5UJ!i%yKE~asm?ASryUR-7|
zlj{*Y=`SwmU0@eM<8+J)Ld#n`5@;-=2#`IJ6}xuQYV$_q2VaVKCJK%x>uA2uSp5X_
zxCg1G5v^Ly&XSRf)8QT9pC8N0Y~c7l$_T(TMK=c4J*{mkOI|*DB21W&hUgWfS^&As
z%F?{6(G+{J-lvr!Gs(igxE0-^#gYOXn94rJmh3V+gA7HB;k6nZ7*)WdNc>mIJeH#<
zzs!Crn8)QPDudtHMbL+J`^~`|d9Z@zOc`b76gMtYih&nx0WG;aX&%RBW>4&vAd^5E
zN9jWG<M}QwPv`RbW!&g^)hil&i%S&FvL)rueC&JE-@e}?_Rskvt2}D1ff-fk?M0M`
zea<RQ>8<`aQhXaiE=8+7&?QiPnR(Sz0(7%|yZV-<n&zL*17$wA+c^njhfP81Fd_Tz
zo=DDNpn_EQ4}R8YSDOaE8ei*Z9<Q`<ax7a_P;lf~u%2bM>`k7!*@QmAAbmAXV}?x1
zuE+307dr2WB^*OhaRamq2XnjhjDmT(TbtERO@2z+J@|>N6txI2#%=etzT|0EmpND{
zR#VyK58@6*Zq_`D%-#Unw8bd{s#cDOo~VM7XgiG2^#k}>>~n%g^LxsEmc1k9E0Aon
zRV^3tL%X4;XjLsMldwvDN@{f@({yisKtXrPZlh5nEggQq>-d`8T5FjOs#Hen{mB)k
z0!?OpMok@ht<jX$QslGl+vyn>&A;Y^OJRoifT+KMDu6(0;_`ra4+s7rc*+dsi}aMh
zvrAEtdp{jEH`$7eA}ZVu%;4%YdBmuL`rBBAUZRk_K4sHyB=uB0&o`VKm@Lnge6~y6
zi16p{;U2@mzl>7Be4jD9mL5;Hcw!K(>J$-R(x4*%_UAAL401K5;t`9E=M+P+0a+4}
zT8Yus<z%kPURcsJY++<6;;ECTQU(Vr6NIG@3Mj3`0H>K5YWA^PzLOd)iwE73Q0@5A
z2Ej(RDZCU}ACDXZg#9>LvJqc_|51yBV^ZT{vSz-6-oiwtyc1S`NVmO?hV%1s<5bMl
zH^=#shyfRPMoM^~onpP_cXt*m<-Q_9<ai(qq6Hs~I$IZFqHeSKTVnD`qP%^Ey>~rG
z^am;S8&wK=JLHuN!(lhc%F4>-%a@2D-@;Y0E5ByLg-RFGoVs5tgK`{H$+yTA)<)`Q
z-N}wuXxw#*E3B-J3|PoQEe;nOnY*HW*8`tR`l=-4tC#uzI6L|DmU2yB{OakuFqwzv
zCFo`~<e*Em7YCv_Q-<|&s<}0h@S;$q773T(>GrWKF7oM6Gv@8x9Lz>sN9~5`8Ziel
zU*n9pTJtTD(X!$l4yc}-bJ$muT4Y;60j8zmGW#y>%ji9I+pnsluwu2c+?5ZGn}b?5
zdDG`Ayn0rLEN<cKwP{MQ{(jnOyLn$8m-DdtSnh9Cc#@rOHq{%LKp>}cIpB{ILSC`#
zh$#>wDdeLOt4f|H&`nmcpnlN!ZBwUJp@m6<z)nKrY-`BBVyxo)T$*dK)gi$)aDnP4
zCX<$J%wAP|S@o>>!6#&!B0S{B(QnGEI;n4!{AGwz+18)KsFD!WZ8K5aNJ2tuT(p}w
z3INF-|BEXeb-;e$kpA+F08W7OYyQ-JA}dKl#iWInZ<hdnJ~f%>VP=2i@CviDMY4Ph
z=Qo~!gV!c2fBl|>aj_g&Zd(2UvBqvX@hP_y^Ta+^s{gfRt?DuF@dy5P(sq;;(P9)0
zj0)1eWvE|X#%i_Y7{Lj7itGcKk9cfJ*mPiV$m--}E_L(n{?yn{=v3ep>5C`#lN|(y
zqc;<H=7U<aYL@c~B?VrB3UPE#DzDF|GYmpBzG)vZO=fs`(RScrwS)y_y9^@Y3NLpl
z_9$&yrkGBbx4xpHOvLSUfY~jH(XY{!;QI+GDoDQ*1Gbi8gGaF9jahUCIfe`JWX!48
zOF&}o)#h?OKAxy#3DXbjhHDnE0d$dVlV=#(2c!*(UK~d`RXgOupl)N+`za0@REZEA
z2hgRF0*ztt(fQ$t`O<6`iHfqy&hZjOS0*}>Wk!@vn^to^h|4t2PLm<#m}6}J=3paQ
z1*iJq>vp8pw_KairKL0?RBp_Np@G}78n<PvrAaAV|Do|~4Sct)TU|)2K_2<}YIOmT
zUzvu30;!co<Z$rn(A4NP9|l-q$ZV8C@3E57#+g(OaW|RT!+SCu$vk?R@oa*-xTEDt
zSHcRoF&zg-*UU!ry|*!-@B+iQkdSirqO-x-^7;^E?x>`ln^*e|U%xok6_{b~LYmlc
z88qI`kUZr9d2ih}hlD*h9QhHbIAj<#h9b|;pMSI>AfZI}f)oqcZpu^~zd+BaG9;aw
zB}LEY`xZ03h^>=%zfqI@ZDu9YvBoXhp}ap4aomkkXdhQ0O&%PhQpeQVZCrd3m9ouQ
z%V;`(-ZAwk&8c{H&}+-m{BHZ*6OBQ2$b}@M#`^8#Ue?kr@KD!{mX%iu!$Y<;22r#O
zlGEBO_UcUeRiH&B(^<vG2Ja4VVVuW|KQTtLRxplHj+xUW4fuzbN3&i(B&I)TN80{q
zdAlQP$~3}MZ8}0lw^JB$mo_Gho=!OOVB!5-5+31~p~C}K?}(1Td8C_n!X`ljvw54v
z($I9md4RsW@-tArx&IB$I@iNB(RI7P3h4^U3;1J-X^i6ZM|SPEU)5(MBLu2^yB~>P
ze>Z$xc2AT#7-gzr2mTrxD<T~%-63`^0KQ-i;Xt;|FjJ)HVvi9ZE(<}UKA}g2n`Y83
zd)%*K-eX8(qu1iLn0`8VFD`hW%U}n5Nb+!MA#OR{U-AbeXHQ&=LSdQ9c=Ej?<7%O|
z?_HlXnJK+$K4$p6JyHW|MRFGgghguojrYiml;LN7=gZGHFCr$GE=eY3v5Fr%Or=rC
zK!g-t%+{(9=D^g;UKu<Mb;|bg`tnKCQ%0Dar~2wK@dBlIzyvhq#0$0?&N>RYow`0v
zIh?Hy9Z7Gb*|l0xN`G{Fk)WJF*Uhm<fvVwZvsAM3W$p$82rPgr!D%T}t9x`4aV{N+
z)@xW_hg0wPDX6l`U**Ar1`Fc?I=c3}#p`^Z`CJLVM2|sSrBJD<F-V455>-w?A-84z
z5<Bnh2CKOS*m4Q-@^wv(kJE#OGccJT<GjKXJM_BpQ(MjDNWG(sg4IoQ`kQn?m$cgA
zn6A8^$73VuB;ex~qRoTziYn)|K3dAP<bk*0_x%VQ3XFF=yW!EKmG}zt4_#fxjm{ij
zfns=U#>=p*eV3z&xLwZD1d_yYj{A(Xw%#jIi`7V3=bLri8>J(lg3UArppao<i>YP*
za<`A94q{^Ltqv6q9d_ug7S?5&Cy^W3IBYtGjRF<;5m|}--gCglWJlGLf(3yxgS~f1
zAvas*NxnvbxsnrdS=DcNdA{#;#J6OU=+!lz@28M(k2>Esfbm9kGGm1E7@jWIGl6LF
z9B4(Nne~GkHPJZ5RypoNRol2*$Jg?7!QPSt>m_=m!1t}gnB$nVA$0dE0_E;n5$&`m
z40+pe(~_Xt{$^v1D-CZ;U+}QVI%46)(88TDN}IF`WC>idDOrtiX8=AH<O6>{!z;Es
zwG7%)oE2Pwg{}TD(ID0DZO+C1E*Try`t=;w!|O)NMW9nW_wh?sqGE*KC*y+<ljus7
zUG8||ic028j_%;w!2p|VGg`-0{#U*zB;_bi;I9#{;k+LMIQb(+1<6}wN7e&~G3F}^
ze|}9}<_mGdxbSIxhvCx;a-L^+WC6Q#V?XyMvb-X3$)401y(+Gt02UlSA1l--Za3*d
z62!U)mi4~VhBTX^`s*|ZUl2~^2m=tazyBqg*U4QD=6HXB9<@B7;-MNn;fa-UX2tfP
z6k0+n>8g}<Q9^MJkWY{a5pHk_$;4So;ym|GoxZD8nA{QcrJ1r*9nHV*^g2gE*)i-n
zzqQk8kuCzE<Rkg)_)N}Wq;h$O0Cr6Qp;61t#`G(fA}@mGHE|w|0R=Mu{49yoUDZZ8
zS}YGmb+m-wn;G2B2CgSi7J8hD@`s6*JaGzLudqlj*<{35P4A;Oj!!zpZspR-)R~8Z
z%`5F53P7p&PadqBm!}}SlAWjKxiUr8;$BM8JqS9Zu8hIRNkltxKM4<XXatH1({DZ@
z*Zm}~uaGM7&N=NMy?JoN9T*=&><#bcUTW#{n`f^gjq=Y<(Tnp5nl~h!0@eBfh($Te
zG(4=;HD2w0@-lFK3b#|!&&_o7!EPPf+^PLtBZ$b5b!6z?@1{94f#G?~etMQ-p?Dr*
zz&SSgQt>Hj<Y;mEJCyVlr$<;N@4gnNF)QFUxy=j2<-5SK+fAq*`%ym3tP!74#k*QB
z(2gVR;VR(mzZ;W(Tv4RzMqT&RmR!E9^oKe-$!}nMBitg;sKy`(G)J&xD~+_buC1-n
z+WX5m7q22>v*<j!@442lN=rOb*S;a#Rg`>5cuiU5{drHV*zczN<_7K(8H%v{>1f&H
z;R)0ac1LpQbP196#JbRVbFed(<a_O@XBuf<uh()%#!9pg|F@AI3csIN61ufrt&<6y
zXWtR)4s9f8)t~|U!3m->8*BzjUq9jX({E>PM+E2+IQN&NXYDsJc;+e}$ev=NdhL2m
zdQt1SpWGoMGoWW5Q?D9=2gobvh{EOl&j0{O^FDEv0MzNSGqsZnQJY~#arbaP!eUIJ
z5LZ*aRbE`oCyoGB`}*UU6zgtrY_Nah#VuzR%7{$+Q7BFX&!8n{Y&G9dgR@EFLo^*0
zlT~uC(P!T~8hoo=Ya{IQ=;~GA$7hdD6m?R}A1*I=RjW0W2rm2t)$&z^v!xcR#Gf0h
z7Agk^gr8(Qr<BECbvNgf90#N^HP$m9Z|yg@BvDMf3up`0&M2S}&+~GKcm@JP8bR!#
zZ0j|532GtRz4MXcFNZ7NFc(-9bisJ5C2ub)@#kNG3s>~~lt0fJ3la-@D|=5D@`q-6
z5(H<S|5Qm+2hH|pwrIIng#`JIFpYe{k2uj88Ou8iYP?~?VjTXM-rJo$>UWz43LlQp
zUi3{!OAUIcV}8%v70O`!hLcogg^aM2mx%bdU7`O`mj>fGR2MP>X=1OpoR@V^7Me$J
z3+c1cA*j5bNbgSf4~Ge}gSGC1A|vr=8Orn;bFW?2O~dmwi>O5E`(5ALnc1-}Tc(`f
z{rKW@%%!ZXB1Sh2=o!bSLdnj_*Zju?&XSMJY1K?6E{rdoB9<tfG~V5&CSGzqPa5*D
z1lRP~R^KymSyf}i(ddetQ2fHvS215=%(o;WT4^1u>jZT3L)EDlJst~~t$6acsz%0a
zd2DvDXO5Vr$>%Re32bbym{&7R7AB1FkBm?quz-)2eWN0Xc`|07IH?V~4i^}1hHNMo
z2Wgm9pJnMqBGrYH$c-f3mTb~9E@}Bb_QH{%2RRm=o0IJm^P0rQ?#@5;4e`Vk--{<C
zAGExANq-+d=4mS5#%gqwuShvQ(hGe*3%?K7obfcw0{H=vc?Ee1$?hY{k{|}(Q!gR)
zBN;PuCRx*xOr9w<{a96j?}EkRaM7#9;!kJbf1wxvyI-eb1%-J^-%m8tQZY%KXXkp^
z6dsINXa$LlpPlD-tj)#mz6~Aq!`Mf>hxd(oOmr7?0l$~EYCv*{a$^U^yjjC7(QWk;
zAq|jOi={$02NJa|3*@SS$e&_{mQX{oen^hr`5WEWy_rF!Ej)AH54Y_A@7(>cd++!;
zEhEb(-#vKE`Sh5U%-Be&FDTeowLwNa>%Jw-UD!%BvkilbHHG0;woGs=1%Cw4lBok@
z?ZI2ZEGR84UQd<uCI{7V%H@)hIf%8D%KZ4VacT=;FyT1-kt$tO+Hkti)$#dAlW4p=
z%4T0lV206=0`gPGLLS&fB*n7|F4m1x7YFn_@cAXFkb;6jqJX}bxWp4Bf>4jx#=A#R
zDp=Gf(BMvLi!CbJr@E=PH5Pp|c69q2v8y-!;?iDqmY<ywJr<C=@4jMpYOZneJ4gd?
zc{nYdyo17g__;zER0rz?4=$}7uWBX{8H0_EB36?zUfXWSYNY}A?6jex1%tTTLK&T0
z)T-xd`XjzVt3EFnKp*!yjLL@dCofonlB$m}E$D_*?W)&gK_V9yrYAYVv6k`=o44G=
zhry!M7Ql$lfr$?eDw0a!5O&t;<)Onimu`Ey`zU=0liUZ-`}FvDnIY;*_ZowkbDB6J
zbhhx~(+6H4LhW`d$I3EUvM^pTi&Ktlj!vx!$ifsA<Z}7vFM}uNQ)*+g+@?#_Rfg(H
z@k`C97ccg{$e@{ZCGf0=dSJ`%sat0PBQ=x?SaZy;-4)`r;v@D2WuR`^PBk*GN)EuY
z>#wycgq|lk>@4M>)*L}1p7KbIJsvJ8OG)xl=V2}caQiLyV)46Ut1flZX;N{tg!_ir
zXVWju*WVO`Cpi>OgpEq>o>W9d9Hb*hRr)yH9>I3Vb<?(5KdIfD=K5CcCB|I_ZV|e-
zMr-=k;DTG}AJ*>g-0nNhQ24-b4^oEY$E4R7(7LcIyC>35oHy>^WfT;iL(eM&ur0nU
zt8Nm7fQ`}NgH|g82oDhue#80WC9G@LB{_(LRLIZOHJw<Vy6w{4PcrD#Bls4|Z%}hq
zoHi{OkNtew1=}%JB#SXLNGGt`?_UZYU~nak7Et6_KznpUP|k^deByDok-WUT)P#`o
zMY>DFx0R@Twk1_ar%`T~7OyPZ!eS091s)5%%fXc91+MnZuG@T@K>bw|Xh^R(?4WZ8
zJ74C4j?PXoYa`njQsKav%I*8^&Sx@vRkcO}7gF!ZrrYGL`xprYijmzxE?)Ruu-Yi9
zr=L~|9eU7}ErwMiRho>9?}Kthg@PhoPIWZOO+G(<L;pEQ%r;USKV?bvPz4|=<d3=Q
zp|AVNc@((&KO&-_NSb(OM5g};(N{+mew6pYrd{bl6*j?33hXk7<|v&jo*c8Yx3iom
zV4pjge>!zN1)E7@R#s6Ff79s0i)NVd{r*!vavp)4U3zbCbb|YkBTk`=8lCB^Ajlqx
zL0d7BR0?6Fb*Uv^H=#Dfr{bI4!Kn}#+y<+2HFL#VF&nTWc=M#HUhIg(X)CX*OAJ!$
zi;^nnqa>zEz}=r7(g0GiP7;syHS#WS?_Zm1ZCf*!sYEZ2it#sqoLFp8)2CY<q&i%f
ze=iU&5!9Cumq8en+(GF7MZ_I_(+d#_I*{cBg2nsqVSIWZG4vM)vf?3J+9}J|_3qv2
zHzNJ&*{%%_W%?~t3N3`7+t^vqE%XvMlU{bq4CGZwJW!xT$-7=%@rcjb;r^#;eA|fL
zs~)omL2H&aJ3YyX#?8H*r2~}AM&ZueQ1h6?MBvAA0dFaEMV@I`i#-Qk^!7_u){wM?
zGccb_tB&uj_sE)28;fXPO?FZDP0TGw>dJPRb4|2IeQ-ITrtAEsH^=Eac~YV^B2i9E
zR?E(6FOlo}D>AIbFQ{kgEj=!uA0JF#%s)5t;Yy)XGaX@dkL7z{XpZUCY|SLR2`7{>
z8HlemC%yY<brtXPE4@)7lT0JzfFz`_U)n^JRo@lsf~)VMqP_~v&m7_K_2d97)1=jV
zPlzGvsO0$gxR~l_YO-tIXjYe7DP<`AI-72?XY?mbgrqBs1OE7*frPdGV3VHK65WT&
zj<Rw7S@3+4AWPeH*=V`oHR;`GEtK;Z<GL3Fb=t=$7hJpGzWyohdAFbOI`B86o?V4A
z&_o8uve2{|*yQ(%P6rNOGagbzDr)A7s$6=T8tWfP195Z@dJqYUztimBv^%2T*KK^r
z&{%uv0w$uNgJSRDL4nUqZHW0G1enI!6;ho7w8<#hvVf|jz#Pm0Rkf6RrAE<-46jys
z&u=@x>Zo`c8dKjK__faFdC17wK-w=LQ+Pl%O>n0%JNp$yMk+%`bho{uO$GV*=p%9C
z?@HaJI?Z%qSe4}iFY#H<h4t-UJ~^l_e{6RKu`$LpiXFkM|GKRK!72AI&6`J*d_u%k
zP0DTwd990S*6<W0W6#sUxK~S$m8sFUb-)>=69Q7QHMQpE-%m|;>t5!v{G#4wt6Zqc
z=D#OEtqirrQpXo_j*I0TLKPpn;-$aJx5<@UwpGo{E&`I`9H+>L$rd|xGMeGT%vyM(
z)pQ%}Y3$AyVgWDea#DF)MW0J?ur+G(;#sDftWJy!H{tqPK*m2kLOY=58dP45ZP*Pm
z(Gr{tDJlLwmLf!N^orHZKdE!*MDnNQWZSKleGqx;rj$8D+GcfCBCS?*Y#P||Ajred
zM(Fvx=|bIr8vPS^-C$+Pu5PUewb0e-ngX9`0){Dalh(9}+e)EC<mMN1ORGM@<hXAV
zc54csDA`u5i8FS*Ah`@fh%e0wI^J62QDI4OZmRLl>6RERO;j{FVJl;y=)uSeIkfak
zw@6rS_N5T>ujEpq&s^ta)k;49&^N%JSfTh2jr}*_f{ie$Cy-M$lg&I}pU7^+QDZm8
z!ii4Sdx>jUPsGc<jsG{jh=@QTc$NdrxauOMYf{Uga!EoY`8Q7qklq$Z*PC7dO(avz
zSo%R_%w=~;Sj?sBmFkL>rjQHJ{z10179_pPdQ&A`*k?aZ%@CwfWnA){KcZVq*_dk{
z1Y)~8?G(vTZfP}#c6Rs!jO|$eHx*SakiV)k+s;>4G*1^)mW}{jXCD4t9r_5VO`{+*
zYV`r~=L&8O%=`!BsHY;NpF?o6Qd@YQuV+yg8=ib1@{!ke>zGG^RQmGj@;L=3d2(d<
ziQXz@bi|8`I)V?EJzjp+J=Hn&DdBYw3Gx%yLHo$TNJGSOF#8=F$9V<c0ji_%p2ayf
zQi{5e=?m8`pN$U1lll5^KT&8XjjiWp&rH=dx&0JXj=l{~ZSK;i<}k5j5t>Sdw~tX9
z>~{rSs^9RV4i4~_>9uCE%g#JWFVJm^m!c%}zOTgZ%@6J@0P@*dc2l;R0-N@v3ey@D
zEw52IZBM6O1o?-P&U`sdBfnf%rfOdoce^O5#QvOO?H8w{ys-<;cPb=XF}()WCW{;t
zwiCBNkuXHcgP{5$zrMut!19f8Hbz7Iij;b;e{WX!T@Gh_v0+K8;;_Nb>vi4vu~Svk
z8<9nU^_KmFI2q5kKV1|)F`gv^)z|JtVahrfI1oa8x(9Hb1F>LK5@HwY=k+?(7bgn`
zOO*)#;wBT=s=WsDKU?jyA41nD{{DsHsqM;R|7RCAe`^6UzLTluB=N>E>IaOZaWNlW
zy%-TuI2+afRlP)-Vy&yFkD5qJq93YZcp2Z@KQmRTE_<9n#+u@MHGf@~39ReV#csoR
zu_-#tm;6<=BvqY1-3R+3V7l~iwpU$tzE=t)apq}+bAkHT88^GdqQ}aK%bF?XLHBfv
zCKfiVfx_<&J=EuWVgJq53BAlCx4+)^4T<ogXq>4VguAue9ZG>r>C7+(tvW8oW_14k
zc+@hc2L9{Fc-yG)94;JS8*iH4#`asOY?kN=aeP>H6@{;P<S)=J+Kcgvf&>Uil;ZSR
z!R)-jkx^(~?-l7pZjCmsUs7|MX@$t*BbU$k-<%u9f#gLfMUNdQQ#lZf%#+pajFfnV
zTuZl(&dmikp6@b<p*)LtCc%e|b+RCZHq%`cgQ?YmdXK!18f9c4Sy?Bmcw}iNiE$|K
z2CGx8L{HpOvgY22hr7mHo&+&e<PrJ~^X}APmcCeX_LM9FPIIm|NAQ9s7tL94Dscsu
z9DTsv)`X<YX#O0Sukk7<en{W;<ND{w9_8Uib~z8oh(eG|wIP=Rw0B+QDpIA&b=#1?
zm9LHA=PXY7RK6fz9tk4>YQ`(c@HHA+ivo*t2ELDnOU;saBYLdQMJrzqdCkIEE;L6Q
z=&4?p9|zC>D*U_})1|kZX<)~t*OQw|Z}x${d%UH=fQ5(0Tx!yEPR#OkpdpUEGWmMe
zJ-SsH)LHTN6V%%Aner+rA)yudH!DbGa5oGq)DvKVvV!c5lMM_Oekiq)Kh@LY0b3lT
zW;HrMbY-_>K?i*3=y`*9Wly6c8KR7fxo|C#BTD4<MF3#ZX2`2uVP2z*hZ@#=>y|LC
z1~~z{eoQwsTHEeO40SfPi!x+8C91KWPkD5GsWP5j*~aGqWQcfWO8W^Dv`*O_mh6RJ
zQu$}-QYpB37UN`RfSMQVoKE+8@UHu#+J%eOL#;8m?dpc%Zinr=iKAGhurFf1i`OW_
zoZZ^I7^K-Z8FO{Oci?09mmC>7UxF7mGS~B+UdoN63exj6W@eT@vRkFJ2i)zS_?(N0
zsjv6>I_d%>O@^|G48aL?9$TD86!MoLx40syF>=jpy7}&Irvx;^`pXo$y1MbSh8u3*
z?axXC4y7^vDV@DgrJc&0A&8(+gXIYUvPx^-uhMRnD|)}e_|2Bhf_9@T_vc@k#AtfU
zYtLe}Dg|`1OJ`RH4ok^mbwanX9V2B+JCNPw?h&QE(zLLbSIk9Vx4pdf8y3xndLNMB
z;C=s2q1WXR89Btk$53Egs`v`UjW}~pozj3jVF|H0->1%Xz=bseC<;^_7F_EOM(%o8
zn{3{#+<?;TU`uzwmnYtg)MZEIinG-FDpd+=H{kfotKBU^j~(yi3E28i7-r*z)WFip
z+P;X@PwUJtr8!|EsEM%6xhgZ&Qk|cfV6)or1xNla`v#NTABw*NPG7>~g=V+Zo>Baa
zu*ED_wnD9nMt7lp49^p6uJ&guD1T8)aC~qia2Uq^C>krk_8S2n@a-D!K~-{{r3?@s
zg7X#c*7U6+M?3X@(2>3Jx>Y#Hx3isEin)cJV|DO0@Pj&sAC3?<D`~;k9}1`QGn@2V
z3B1GKG3eM#rhoTEEKoyHT-@VU2IURK^uFFa!^pR+p_71~WOu!S)EZEy3Q*;Q<NBN*
zP_zzZrlwzCyFb-wh%S8;NaORK<>NzrhH9|}%c9Rcu;b4B=ou|;y6YX8dg(-^zWP(Y
zx$6R>p@{jE)MM0Yz$C{;<ePBD6iul@Cju@{4&}p>I50(?IUO_u^hNTFZ4M!wA99^a
zh9k1011<UZ%K(1?_UFoRWg2PCFqF4(g@;rZJ>A{GL2@DZw)<}_AswA-+_X$o5z`FX
zl64~4=8nEb0m>?hk!Yi8Zng_Ko2cq9E=rbH1SVT$wRM|DhntIQu7+kmUP3NY=aonO
z_&;MxTj4=&CT`V~9iJa8RBMZ6sN#WC<7jAeM!a!v2Bb8)uxd53>wP*EPPiXNd^h*^
zi>pU3M|ONTA6c`2K7bs^YwHk@pJGCtX0eU5H>9)#?mzYAe3cEj{DQre-a9--RA(1k
zZoo1J%!Vz{@vqb*8J)x45<`+~xQ;l^jknRppSn8!P64S7=$;8+|Il5#NiVHowH0;q
zSv;5FZhQ^j`#uEo%mQ_kP|_HC4CNq*DeXz_2NUILj+P4%J710MCdV;L@yVX6+P#H<
zTlcHXkH}AxHgQ45B>eU|*%xYfB$5_-cnAApVpHWmUT+#5GUtVDZab|Stacnu!ql%K
zDS{n8bJ!U0UXfBf*{<G;<kpjwm(DrMV-Bj3VbrL3L%P6(ir;j;o4wb|U_IZYTf>p4
zc{O=&XTd)>;7dh)OW5Og_G16}78^i+{<hH%D6DIK`J83jeK&{$?;b*>^`q+&j+%(R
zceiu~pl68&J{$XyA!^YTgf*EfpX^r3R`(8$MaMRk`<(DM?RtugknY14H-aRK@h}+A
zAZf$Lvk~fG6!qU?;9Mjecs@bpjF{hkNa|~x(%^`AN`9i53YjO3dTVj_WFZl}B;%{W
zd|N94_0Oiyo(*hb9P(%1EEjdzGiVg)iPtqsDR?y)GxB>LYe3AZXpGnFSwJnB#Gfz^
z%y&)6W@?3qSyBok9T&rzv|>BJTh8D`sPmM3^gz0kF&FlCc*#Rqd0eny^;VG7FsF4j
zGcKdvBmusB>8kU$l*?G;FP0gfa|Sff`lBmlV!w1rzlZ~r1U!|3K?;vjN6OJl=ZnA?
zRkRG;5TZP0wa+#T0aiN(6{wmGf%IX%#PU`ND0I<9`0FgBVK_>Ar&<;Mqm}PVypBpI
z%PkPG)ZUi;B@~#|)2JT2G%R>c4KM}%TmzGP>SY$2nQ&yRYe9PZ^F$@Gy0a+U7q9XR
zrX+U8YdgtRBCKZ$yd>9L9j#VXzgc40hE5LyT-^P}_V(9JGJY>hhL={`O6lH`&#0))
zj8e$;#X=sNS;Cj=WnB<V<BmPhsnV6r=gofAh?Y?+BdLMu@Da+v%c>c2fy{U8D;lvP
z4jW7kIC$K4Kb~$W#Q+TnROqZ-fj8+5XML?%AP@Q4oQOq8D39h4y<}(dbwAhQc_C4O
z!$IZ6fG4G!1lEr@ddAOPyBEEq2zVi?J#~i@rHo$NC!KFYjj(+e@(9^&r(aK4#pHwx
zmzcdk2uJ-Z?fb_lgh)hIjE|K<g1o6DHhxq3OT3XxR8#>2v9zK@F8WMUg=!RFF!=U#
zkhy<q>LPW%PaanYEg{(y)d%@6j>pg2zdg6b?|iYtE{?^`cJ7l-Xd|5k+eR7Hpn>m6
z6Am4QJFt((0HYk`Pi?Bp^TWk!HmKkBwz;RXV61Jx!gKQHE}u(THwWJCuu>K{s<{<v
z=JBW%>j;Mo#NVr#$$FpNX@17+bnvOAOVLvi(_J@L*@3n}-AI2lAA!}9zTVyh0xXW&
zBFsR>@#(>AwXC$)93XumcqK;W?_35x*+(wXXbworja<zdrB;a}IFwl5a&~2q;BrxY
zeo%eb9=&=SjhH7buZWk|6YB3i=&LIXkPx9kwnhR(((HSaZ<^8<>&`xq2dG2}a!>^T
z=@;Hv&zP5nRAzQlm7(wXCF(Ej$MHQy;|oXa8Ib!u_ZQv80GEP#D<w);4*TWHIy~Hq
zBoQWkJX&_msQwO`9zPY5=>mQ2AZ8ue_?J!R&R;#fdqfJR+K_1sxgIJy#bP;SFY)mR
z>5*u{@bDDm`#MI#uD@gnI`9aA6df$43u!$t2&Md0EnL{x?!zEZe!*X)zAwGx_qCgK
zbL+SwHP%x_tDkneTQLQe=%Tj<YBZb%k$^5>0@%a1-ETzO0XkFe9@s4K&GIBCNO6iL
z?*Mc8Z1vP=Z;N<N4&Dp+{2~F2&CKk8g@hkJObF_?xvE^(>2QWFarw~Yh5x(_Z2(QR
z2y!0O9W69db4h|4ZzImh0KxC_&Rt3hNz2h}2DYKXD<(q9=_e(E**RGJ9(SKbK9)_C
z5Z3LC`ACNiooVDWHU;f!l)O{JlArhjqwb>qv4d?U?)lE6wLT3C`>sFBgMx)bA$d?l
zGdNU!2G!|8Uv*JE%hd%-t=#=&?ZqkLOz-;o#o_qepadXI08r<U6_M_bDYZgocJ}Z~
z+j&W#uK)zwpot;!SFpP)({IbgmQ)CRGXba_)XUS(``x3bb(-#9m$6@1Tb|rVIXZ?&
z>3cD9b#;}w3nU$TPZnp&m=%Frf6~}_)yg*a?WLu}&pszj(jIk;5!}RW?e)Og>Rxtg
zgNJ^CeZN0$M%bfW(%;yEEyJj<Xs&;HY}U6JPtz5uw4|~3+J(*(Yt@8EL=agx%CC0#
z?@Se`cL|=U_~A1N4s-N8Mf&ShHazm{%TL%498Lyik%^`6sg_7y6mdSs0Os{9<hAK^
zF6RV3ZioNB(ai^xKXtUf1IhpRk~<h>iK=JfUZshamRq{{fq0KCn(9WrrXRZ3>H{4g
zN#&K_U7IX`hll5|px{(P0OCp#<2ml1xPp$bJd!NU`^*M+o$*~ZzKh)=?{XM1;_t%`
z|KGEJxZgMOk9+w;MQu<bjRdA-1S%edlBpG`roOw!G>=(z*x*!%PWtl2P<i?1aI(cp
zyB~)I_5b+w$Oye7$YWue6LA%_C)+}svg&UuyRyxAJC2lbELPx%-lH5N{ws?9kDLD&
z4*cs$`4J5wD(Zh8>tD$lfG%sYnW{iR#5SjgP+)*g2CYi&*U8PbHIeA(=#7<?_c5-H
zDwLR=o4Io$`oDR#m-;XNR{n+6$oN}MW(r$TOeGMF%q-YGOP4ih^NYMMz`KNpegJ<U
zZI4`$Y{~nmL0T^G-lzTc15n;UoJLV0X}=A+z}lmjmMo6cAgvHE^qD+`yF{5KeGmQs
zga375YyvJn>SFT5hH5oG9L$@*`LMg>4ljAR?M?=N{P<CBH~F;z0;5)SpXtFyZph1k
zLHR2GR)oJ^hESU1dxLg9twv!;ERepwEOUkR^;f(`_(ldtgn&Sf`SKsY@gE=6U&Yyk
zllCdzgq<Fk)f%iUv<kH(u_{=mF}||j>j-E@cXV<BWEG%7f&@gAXK!apb#2BKY>SW#
z>lZkTwwZs^MjM()wo6S4>t<nZduEFDWDAuV-lY*UKYH~4HWs+QZsU*T{QX5=(c@-M
zY-y<C!MlK~?(DL@`N0{}PS|17(e@y7+na(i>$yrC5)uVhSE7FA2i&05)8!PSu0ILA
zDAx6Ua)C3|b?mlogR<>pg$Jj19rChGzsa%<3n|H>&aC~Nu3!q6Jv@`A!L$KhIqEF(
z^o#$6QAK&Vzudkf^SnGQaDUkQ9X`|KmsL6KmQt8>>fbAGRa;D$U+&Lw*wAyT0ea7W
zpDPB0<jY6nZm4Z)Ju_MIq;yKn*|UR{@6##fEx`j0<tN;Ksv<pnaK5qClm9lNPe~}{
zfYo%}!$B$bF)9SR?lGaFzb}5YBhHc<fx}=dm%YC(QWl}>ypgtJH&JYWxl>lXQ%c>v
zu)b`C9xgViR+-8#-U5>)<<rvAIIQUF)R2JxVlvA)-k+mSgb=O7+z(FT?C$Cgs+cMp
z3nG|4e@$=rx{kYd`GWJH9l3s+*!2Cc2j+hhs0(n^fJ*xw9!R<VlT)f3L3|Q?Qhm0~
zAtR%Ja|G_amCKg)Rk68E^ZGfIcwW~A8_ph)iK81^?~Sy9yt6F=a^HWO6!)LT2?$+}
z+CmQ3#o_Chqv2SHe^;I0kp4T&`M<0Q4gs)?4HP5=eQx0^(W8)zcYmDHnSqlkSH!bC
zMtv}h9kx#TQ10Q?P#^;ELlf2+Wq-_R3zN&0DadpvsWhxe^o&Wh!rFvZpn)ilSkMn;
zAO8MV#`&La|6fP1?}YF8)fx*iKUm7+f!C_X+HyHFOMclYT@ak6Rc)T3y_L%AmM;-b
zz+uB&_m>U*>jmNb!^eItR4<e+A*fTSEk6}H7_#0uZuks$JeHu)A@H8`7~>f68tX2f
zx}^7SC-ke*Rxa@M6z;#AP}g5xDJveOT(BLmEW}^62A~vI1pSmMtiA0b7yEL&Rrk>k
zbrmzn9qbZM-&=SY{LmuZ5>CjO3;4?9x`cVU&-hGr!}zy<bkZJ={QTk1%y$CB4f&oS
z6BmH<gjMsSuv<r?*V{(@>V1N=gX+V7k|3-ctpHy98vd_r{XadsQ+f=Y2AAm4r3Qwj
zheocZA4%NF`Po*q$*s!@va%GMoSY#+L9Yw0^dD~VhQG3O<{oVJ3;-@NyedEd5&l=a
z5-OEk`Ps7f#<%CO2~isQ{M!;b7e`qAYq<YaFoXO3Gkir~o?gg8T&$B>_nE1D#FzA1
zIReREz4}b>#vIVWb_QbfFD@?rwE0>4lq_J-hArU!Z~rXxw}i`(VRKLw;1~IG%7XgD
z?+4I)6ENXO%7IgK4G(Cze&J+HDEvN3Icd8u-2dX@W#A#ovowe(5Z;29JUpv&Nn6tr
zY>sO5+kEBsfZ>Gz=umn47BeR|SGmdcV0*U0Xl~lnwh$jUT5q(s>Hm1^CrBIz=-Edg
z-=eS_yIw!lOn2IBNpVkA@w33=)yJf1`kjGB3sUKD_lOUE{}Fop>W8vj_nXJC{joz&
z9edUGTQRZMlHfM4n8t^dSzoADG#*7Jx+I&--GT)iv|F?c@SlU0hxsr0L|Bm$BZj4N
z9k<xq&+<OD^OIR^?)ZLzSSyVG3rGF?OrG*$bWwIq)QS?)u+S>~gM)@uj_RG}-vcJ{
zueVexu`=($X3?8bVu)5Mu4^HLM;eDFN0Q`i-G1~Op<R?m;ycOvUouztYfSH;3%EV9
zOT<}>_ktf_aR2clD&`WAn(5J4YHyvt<{#JdwP@LF%YQ@^EdkFI%LoqP*CX)L@oTSe
z5`JSe?VMAl%h!@$ne0`%&CBxy=|6inIQYMf`4JLHpkSt`*(DwL>+jkQ7A9=uu@Kr7
zuOeI$%J1KF_o=;a)O*Kwo1bMR=@2MIRVl>AOBkC;UO)N&Mp(crXIk@PL_wsbyNz>y
z0M=apMYZ3Zo${hY;s6yTKJIOTKgLFk1bQ@-l|u*$hUezt8vkZ1+TTzfgEOpKQ;;CZ
z5V~>vD+=OwWgcTZfj|EAiu{q_N7C;As`AH}z<Yd5aXmJE8<mtKfi+6L{jQI((PN#%
zKafAG7lV`@C1OC0PGyrWF5mQkVDg(1Kb>Aj-2Q>b_kMR6I`}pWbXWaVgw!9h0jJX6
z^4#b61j(`J?MObXsqSdJ?cu&fnlL8MM+hK6`^Rs;rliOb&e-O{nw}9%L3`VL6@aAh
zwWT08ZSZs{TK2b3lHsMBwUUZN(&9QPRn-E5%@098eFf9?BwOC?9?aGtq7&}O0{s4q
zL$C-be}=xVjuh4^)*FKX4a=#KiE@2bHt>mq9Z(YBb>19T^E~AO2GXMC)Y&bE6cx<?
z-TrFyDp^$M1@BUmfN`v^FZFkr$xa9E>8W_)L8A}hrMl1ht4BeAY&C9yUis#!ZAdr{
zoer1my%~DjtS7KVv+IQxAGp9=>$2+Ovu6#t4-lBXmKK3-n`_LM=h*#O7Dh2!^wPmS
z&D}(+bJHpWWw>{?R(BQP?t5J5%sGtij=2+tevLBPX5ZR<0~!@Px#KKLX*<2W-xBHL
zZuH*E)R@;z(KR1VkyjnBQOzAqs{eion3tKCoD#*~g+q#eAmIIUt~5%ac1c>cVU%0d
z^%K6oPSJnSVC{$;Tks%z-B$NFU}m3Co!wH!ZFUDjneUR%XQ7diktM^nj&zA1A4&1X
zjR=ER>8kJ#CxSQLEe+={^41-W6}Q29jnNi8T`Bd!c80>TrlWYwX@YskP6<C?yjS77
zsV8mvZ-*i5;k&Xja=A71Gh@++J%;sbrQJJOf^xY4>2J~qtFBfh>(Z{H`;iY*Skyq{
zj3puh8njXKBGYrk&*I;5*Hw0P2!XJZI>}BqWqGQZgCa<HBq~@cUO*60({G>ciV`=~
zW{LPeHS_dNOvTVizqhHd*26F|G&CO@m92agFhL++FivUl!47P>Sc^M{x5DN-BCs`h
z-?82omcH5L(BggnvDSK4fjTlRGgzfprO1_0{&?mzzG2UC?YbT1C=lZVn|;8;*1iks
z*{2(L#ZP(P5m)?Dh7g&v^v0(}31f}c@M2oY?U<L*(%Dx&UVVb=vOhvGt48?kIBxL$
zMq3kQ%swhUL86N-5#yh~816NHzx~_o!u|e`&<gBjd-O6#*|H4KUDvPt37G4Vs|Rc8
zD^kRu6iL(*CudC=?zWHVD`^JApLJea)4IAWIyFgCVy_y;O=QeGQ-JzV_DvO-y}_>M
z!Iw63qaYN=pkEcOcLt4kRd&(Kje^&uo|4`Z8;?*@7LOq>bc^YQt_|jnRP^c)F4i2=
zd+dXGmk&u^99clyd2~!wO#ddJt#sb)ve2qy{XvJ7Qv2~y_-@SdcE-j#KjaXa24cUj
zUp2JO&kU1;R`F4PJ_U_lZZD?pqn5jKY{gIXjy^u2!Lx!fP)A7c27c)qp2#lJT4Ht6
zRUo3cMf!8ke+uFsghnLZkDd_YfXGaq9G{!Z3wrZ2o4IkNUpxc)cR5z^nSR@Kv;IG^
z@(oXU5k{OVW6ba}FjsLorv1wOn4e^px^)!gzwn^Y<tf}&bu*S@j4#xMyT$utd>d|M
zteC~)jg8o$+G2roqO#cKrC31$4%SJEt!^7IqBJ%4N3UYo;fwA0bkD0JaiCG$<?3WK
zO&|+!A}^I~9CvujIWM;7DF#n6p3cr;<>@lKZC}c70aRq;f@b&HS3+zTRsECACKGb0
zY`@y17n9JP-`6>WsM@Y5vOP%ESkAmIj1;GYr`JHkGae5#WlSEg$B>9o!P=4whb}Fm
zyt#sUNYE-boGL8xd|9T6K-+p>(Tc~cH{Q_@>i_nKSD(->TR!L9{9NL>*Gw&cwM!>Q
zmalrDPPA5?ojALtS{_irSMozZrlF+t?@0eV;qR2Ny-}ee(ZQ{d&U4twqMBiNBe#kM
z9@PVnK5ZRi2uN=F3hX(evd+}atL^Z8YWm&|81MziOSu;(S3{I@WXG~2slQ&!zq?IH
zg?^ssRh;NZH(^O}@r_S~_0tIu%{hG60lkrN;@Y7Ab=I)1qB7bgHZc{!+I_f3+nh36
zIuzbhm&>L+{DZ)wuH08{TdbVdzK}WIf6*+;S8y458t_F0^8-Fp-Vo9){BP6nc=j)D
zrfW87(nex<?~F6<W>YT7U5Xy}?%8ig2Ec(m_-WDKpsoEK6dqK_koJ%qjah0FIQ^zQ
zuMM-^0ixn`*@M+ps^28NH-kjCR@<TAZ*krgJz0pJx5Bn*6@iuq>O?1CeJd7RIpuEb
zI~lUNQ1mWW5`!d^#AR<PF5njWRv3l0B%l5MoOKNu#^Z*y*yNNR0-~3ZmX<!}xG%|9
z1a%hl7U9O(Z!}(C*NOD?6etVcc+;sBDNXHOYF77QEl!u}D27#JeL`&Uz-CN}?Y57I
zAbK%!=PpvJlXvvMt-Ww|lb)vnlek-%v*#b(0)`*b*bhr=o_Ft<K`{<x8TpbmyY}QK
z3ANfxKS_P3!2`j<(_Vkk32B)1_O-Fm<;*pJ1Msx8n6N5y{^;F3PtrtoVVC_hBpE{b
ze5o~uIN5D%q@)QKy1i6HkA8BMgYXR5vDhbt$2S0U->k7|)+xrbwqVN45MV@B%wTRD
zt-a@3Qv5T0fkW>cZ_6Yl!9lOq6^zmfr~iSu9d8A7Z=`cY_IR`Dcyqqa0jHk%!}FNu
z;_OB~RO7GO3I^sdANpPZF)4xVCwvPW!e8KuFQ~ol%yc|LRJ*%R28Ui)K<~b=OduP}
z;VIcBz3?@$dppSr#@ZtKNvOd!;vJCJf#FC$08JIfHvGo$6P4ohvDL{8l>~f4$vkcc
zFEa%+2g?*iJIzm+W&PiVp@tSy`MmK9!DF_q)C&|u&%1H*RqJwE5=Bu4AGLDmkJ>pI
z*FU7@q@-}M?<^5RY3O#@9VRRL(|up&$kaFre$V`h?1UG5`{?!@Z|nzG4MTz;@DCbm
z!j5I+WO6@jTI~RA^IB3Z@qCcEr6dM1pq3zHk8SpJqvrL}>t#z5RB-AujHmyu9xQIk
zkk4CvRGwetbQQ6|2~m8!F%%vcsXWhlkS)K*p|W`Evi(Vs>Pt~k(b&aI+0r+EC%h+N
z^^=7vz^HNpaa{WNI!oW@OH@B@r~I7l%O;tbjHtqMBXl9r#FDsO9wU}sRi@x3n@(3q
zwQp$q&8J9enKdf<gSn*J&6|+m2&`H#yG96|X?V-d6e%IL<RD(jx2n96V9&yLMIyft
zmdRjT#Xzes29(?7u6pe%waJaVZLG5|TD-bp|3*yCnt}1U|8?M;1@rW=rfnw%tlQ)4
zY9)Pt#p&9t2g-QOdq;d9u_e2r&!^wvX#u@p1^Af8i<JN_l)r&B%Ch6uJ()0C0(1+L
z;?KJ$3i=0Fo9Wp=qr8wKzI^x-9sh}fLar~P;^QTNGD9}|AVTb}PayPdS*xD0e!IYg
zydIyMwAkA(SvfxEfwYqgRf*F4s|E}KfYA0!lw8p2qpLw@uNRB<(4WK<EDPr^(tlKR
zQg1KpoUOfcf=v{0MysUdqRwkb^(vilAX&2pt=@BpKJJ|q_rZ8b^?&2CJC&;s)}mp;
zE>Nb9%pn&~VIv5+PU$`CZ2S4xTT`kxx|X9Xp;azB$XU#o3-+88Pq7WwC>OR2#O71t
z|HIi^heg?b?ZOKBfPzxeC9R|&U4wK8NOuS--JJu1Gz^`>(4~NM34=7!jdUa3HB5ZB
zk3P@u-QTyr{qDW_W4PylW9FKBuC=bU&ULPHiKpHtS?~2!508${+rW7s$)sX_Jxa^0
zF}af_lG5TiRH927j&fEDjw|mfqhR5zq6xI4tRF9QW92sVi8*_FLr6qq&X`}-y*P0G
zzAFd#!ZgKE)iH-N{&(~$;xu0L)6aO2drbM5W{A|fJn3}&@rJ8zo7?*K-&S?|pR7aT
zG1udxCwsT|#VR$(zkepLF7b<u@p^6PWseKssf|y&0)FQ(k6Hf3Q(X&GCYvTI-wG5M
ztVZsas;qh}-}ZSXc;|213jp<E2jI0~En^<Ncz6AHXVQkw7^jht-u~QjDz)2BFvQR(
z(!vT&O-rl#Dfx5G+6xwGKX$5P+BG?#?=fn}hlE50G`!jC7*A$@H1c9%spoM>H8Hf3
zY<CWQonl62f<&sGtsfKV*ExQ;GJq^DSpE#7TVH<ltm9bx0Zx(Ew%_KBuq$BZHGSsG
zO~8RO4+dB(Bd=m`dpyr*Mk0ap{OLl2dxD|TqaxMTt5oZW&&jWEO02AE{Kg;7Hn>+i
z)5>kHK5?-AETmj2grEW4FMy&&pjl#m%(?BgR!jQxY$mkM=fcRdh2e<QQ3%8I`TmOw
z_mBKN-sn|_QCcC-#AHt0^7*sv6k3gO66L7b@!8#s=~35}J(>IBv@h!&7U)!Vd*kkZ
znl6@oAPT77k|-A&UOSWAFwmzWYhugqnB)QZ(D=kQ?XL{sjGr!$GZU{6$AWWgLFDo$
zAQ++9S%5wH*(hPQ<`fW&piBUM6wrcWE2>L_OczO?he&m1TU|sbg!~1GSv7n25O1cy
z*YGkba?XXn)`e=5(8V`I)YJ^pA47n9>|jtO;Iby<w)Y4LPk8h?uHOWOVks>ljNdtj
z?xR<ba_!=&6$gvSS=0?$j(My=<YU#(Jkemqz#Tdj1MO;io%7mtvs3SWtEb+?@P#()
zwHQIyD3GVbk?1a%T&%2^BYIyj2}eiA<d|sA_+H(M9AI2hze!;dJ-?Fj0Tn%aHz($E
z@;M9``P_D4l;{3{fH=0RPG5X{waoeL29!tV)YsX2Ym#-b>YeD*($<A0Cj--`!>T}5
zP%QIlQ?vd>Chk5O2vQ_n`?`F7KBj2f#TeGm6|8)sCU&V++Nc4Lx=N(B+%<F5?^dnJ
z`>~4mM}SrUS%F)Kfg6W+Q&Kk9*q)dUztB+Mz%3zN6WnIaE2+iP&~u-u5y14#;&0r^
zB}|kvOFY=WEja^|qdGPM6eLLh7=**07VgJlN3O1p`LA1K)mPdP_!%Zm`711r0n}}H
z|HdxAV2wAJcwe6VEII^hKO}|vq6qj-ZQ&{?_p>AJCjb2q{KAPcolt=5&EO;AlXvYy
zauMa9g^Ua*8=AWL)vz#zMh=6Lj*%OO;6h225075xMyF41OR0?N*1Zh5a9cNx4J++_
zJDUQHHTdAn3%;m0dtx-wW0pMYW=%esN1Wuky%-KC$V!UvALj07c!V|SI!;wrq<CUl
zS|KWoWkz$Mq~o=0bQ4u>ICT2+u-#jBn&B&N;$j-TV`xNWmHdrZ@4TfV9$sjv8e7U=
zXS{FN5_hpT!v@#NG4q1|#M3B|EI84t3N3oAYIQz#%ei7x<(y)opcJ5HDaFw@3H~Y?
zUT}z5p${-4@Q3Ha>y3Kq*4MU_nG};o_H|rcYw*V%S@kYildQg<;XUQ#&(cQKMI0>o
znO-~|%q=A_N0bE0`7!uxOvJ;GRf$8;#@Lc;LS$-lSh-;n>w$TW{@jHfPfhhe7P#Ja
zrZ8+p9PuVT?Ch-#Xca!uaN;?U`(>cc!xN0pVPJ``SY<NxS);=iCKUgsE#mx<FWt;)
z5I)HVrw9uY<G1UoYHDgV+@?f#3yRpMKBeaA!yz=x^i<zRsm{BpDXEoUh#Cok`})uE
z^ndN*?xdt*JX+EG!QHyC0^Db&FAjR0IzBA*T~>#^Jb($^nq1={?ntZuG*P43<Z(k&
zrr*?^3wCN21!4MOz3QLh=xe=oIdtYV9B$4+42V|cx4_S{0gCkKqR|b9vy@H+=lb)3
zjE_bitGiIocT*Z@hL3srBY|JwZ<Naq{s3ydpF|xGcC4B_ksk#k8vX@ZOTM1Ri{M57
zW!50dn?`p3cP4~D{maI3&6heI(uB$8Ey2ux-u?#w_<b()c3=Y|=Ej?}0wb=%WgaDf
zkg<v>eW=iQ`tP%Nr^IyRIJ!W@Q!g|Nr2T~sC~N!<KO_XC@%Hk2jqkA{lQ%Watq-2v
zk_z(j*~F1L1USeyAAzZLQQd8B)1?~y#mPXI1HtR<k=BmYLnCij614(B=P_{WwJ{?4
zt^iSCR35!dgzg@uoOJr>f*~FDim<Lkk7RMQjR5)N{?DJq29PE$_QH&3UQ|LtN^v6%
z(k&6W*C5Vp0Q%9@f|fDwYOS+z(^4HP842!ajGD>FB$kHa(U9MJZ9P${0_e8pDJHR{
zy9P0IAWzmXWN-o22+M4(8>!QdHYa=UQ?0^B1&l@5H)}s#oL|pD5!+bIr@a_2vU&jD
zdg<}Krcblb*2>^p;Y40tU?0?1!cLfsKK@4)H($B60l$UA=4?!KG%*3@@JCTL&b}}Y
zdt(?A{Zk%`-VY4$%8^QmZy5CLQ26thJ^m4kfFlUTJ9RObKQaV3HQv2(@uihZT?=#A
z5f)0~mg6@s=M}K$;GY@@x}Xt5Enf(FF94PIU`Vl@R(Bhqfspk9q_LNA_*x|EY*f0`
z02cG}`OQ>d3zo92@dbBlq4(epuc;%*+0pSakX%vXT}CrBav0RK_kuVE@#^_HmqNYI
zcc11SSoQbi6$O8Z#WCp+^)VfOaop@3le;9%HoCPg+;q7k+adcfXJ9s;19`{o_J87r
zzf*8iiRG?O*t1_8%1%Q6fa?KIuowWw)59z(*9X7LQ_i%17*9|_>WcME6+e-Oyqz*B
zYjkTxk4VV{Z(+^cM*P>BIOVZ~W`5;WiNSFBI_Mpl&?DDgcaRU9_v0o~M-JD*{01QZ
zSDNmRpgI)q57V?XNB*?Ei$lUIa=?<KYI$&Z`ed(9V`pnYN1EB^g7^LAV3sO`M!v-O
zLu!FN2gA+Fmx_-@M{CR)8X5%CL2`syjmeG$R$zp5J_cX?($-I<jA=W9Re_kDZj9FN
z2T3ox9L<r^`F1Vd#sF#8#2=w0-MH#?@S`VkZ*M;lCEjoo;GcS)hu`Hy@V?pK_{WYz
zrBh|inNHBNh%U~BuT2pXpFfuB(8wX?JnsuTYDVdh+^KAfVsK1yVYA>}OcnT8<s3*D
zrz}oa{WEr)vnoJ^#zWE!gL>X2j33Kht5cL(xV}h@NbVI~kECMD&h_w~uhWef+?54J
zAXSoW{vu<ok!j>{*c&>iyq}FZ;O=S~?!PX&>oYC_tGeZJ2_*XdgMV#&lK~(4uZ!8@
z6Mv}}*!Iw!(I9xBrh=tk_Kxr5obn~Wf6^#dRZ`AN;W8Pkixx?A?X;CMXNlGZu%Ki4
zm-6`l{1*8n*zjuhMJuLsz6jkz<u&x!PoNPe7h3Ny&+pxvD{zpSN(O%$!`QZNHYjR&
zJh6s%AW%QNs-id0N8k3%e-S`-{hZm?pTINX(-o&H42x<%O+EEp9~}r{eo3y>6-kNc
zo%Of&>Yw-I{Mh`2eyCC+C9?6_%MndyXAG0cf12F?B~GeD2h2tAK}`(A5`=3ec&$$|
zeGR1884t^POTHvDWNl^2RPE|uu@YQ+au)d=b5A-LE^YdUY;k+iBvF68-~8U_Te<G}
z0}<eC{zW_u{5pUs^3Wk?!~)VewK$%~hdes}dU--9!ow5)F*CEoai#sRpylfXcBiA4
z6+bu0Z`XKbL@-k$5Zh{SRW^pTl;yA4B(0~Vk8}N!J+!a;aO^q8GTny<p<=(1QXmGV
z`v+#IYNGW#J0kcMH$Fn0r7`+b?5eihcol#mhmHLo;Q7e9pl_<_=Z}slUt;>*%=v<Q
zB;&!pc7hIw^NfmMQ;xIOPmos&hTQHXT7G9g9<dtL)%>{73#+w65eAOaqTzS>PAuiw
zC$6tyr-|H-3$?q>>6#rV#;4qI+MGsGp%;pAD^1>&0lg;J4>2+Mv2R1##Em@PIrKDS
z95@)mygBX|PjiO8YF4qc+Up#pU)W7<E32V<EXMOu$AY5IO+<Jni&VGx^bqXJ(g`v_
zZMXpP6@{ao08v?@+5%bpaj<asDlxs@Hr?)`4Fkr2_1?0NONm&G5GPfzDvhqJ{2S~V
zPO0NC((V;W{Q4F2c1y5)NS<Eb?rT@qD0~XzMpqf~$S^WKYiwdp&E7YneNRhjeCF)w
zVpr9H;0=Cu3MrP{xPM=qJ$v=1j@-D0^}Bbi4MXH$q<X%|;KSB;bWmGNjY6k44g5cx
zjn`|kPPY)HVYtn2OdU>q8JSojd22oF7YGPi*J~=HaBdd!A<VD36p53r6ZWAWB0i;*
z=LBuHb`Z>Vt5C?4Gfi$gJ-CSb7;nZ40TBK19&oDSN0fGp$ve>1aWnKoK&0R6c`Wls
zw{zV6bLmY+$3Cz{(f2B>(QwlD_D#E_Pk6cbt9A3goegb|;TWtZQ58c0HkAD~Qk&J?
z;vQG`&AVP@I+cVEE<{C9c`*yZtZNV;K=vvX;<ezSigwx<o4oI^9UW@IiG;67-f#3a
zx>A-sC8DL}@_iX|(mlE>j$Ad56W0?1^+2*7wM#Z0xv_hfX~1GQvs5KJELVn=eje;M
zuac;AeBaj6J4&ZUB~PA{+_>#czlO|I1c8na4~Gw{E`9I$269~(2syxrowx4a%@oPs
z17;cM%ea+J`*0oR!%%P=_;C+y-`re=Y!umxg(kN=fC>CHrOw5N0f>Vy$fePLh&FPM
z@yXBQ5k+1aFAby4j%r>gSx*+*!KEC%PGg{B%ryUB2S0bx2GGl*g3W&63j4xep~wbp
z+)GZZL|J8uu?ee$5`at3e6u+xa4-UDMS%+7F#l6hIt-;S)$D@E<ndz$mQD-kV5+RL
z0srWNM|D`xhgZIKEq}RLoz&o!yy+v{HWQ4&W%29RBdng_7aq2hVKQoI3bWLwa(hWq
zblj3<r#}WX?e6S-_Mrh(9we(#$G|Y12D~0X#K)-9jf&V-s!jIT!#pmA+vE#IR9@+E
zt>|AWKeqL&8b&=Kv|_IrB}XYC99~yIowjDVNXGU2gl`nRi0M803T(o<{Hks5I<&{>
zs*OMfh+5LDo{JX%wbLgJ3iJ#?WJ9|1g~NH)PK=kpEHL!*$r6*fVp<DTXdW&4xz1Uw
z)41%}N)hd=%3K*>{NrEIyF^84?8fb64hF|RHrqH@xj(YES0TM~y3UP!lMgpsYd`L2
zM^;DmPqIy04#B{LxqMk$Kw7ISty&XoU}*dC%LAdipI+2havfB!LpFC>2%LlbQ7fcO
zX1Q6=G}4YVC>BSSH_H8R$dB&ThGjJSy!o-N6y&dW$e_FeJHvXtus4lg9Nh;x)z;V7
z8M__qK!cGhxNEOLhN5VJ(U{s(204b7EY9p*kzz<#Ro|*9IhX^FJaX>k@k<3|f(w}4
zn)_(V(1OTVu@ggI4t^xD-62-piNS>V91{2FPS%mf5%s?;Cub78&erIH0BSa%==)-C
z{14M)>72~!9LZ-hmEY`9z%}6bRHvNk0kv>qE8;|{)UdH?@_fAq2PoV_s$;l+vjzeN
z|LJ7*mN%&Ke0s9`{QIU|w&cHv$DiWC89@VuZxZU<L&3Cb1B^PfA^5T9mk4Aq+eL>$
zrRj`}J}F}j3JmGOemsqj+_+0Nr=sW1E@8fTBC;ed?CC5rKD!Gr7N=S9I94qU8ZqCO
zAk-EMfsFk~Y#{wkY)gxnRZn)P3B6p%jU6BLJewb7dY2<Z&jS`YIXODVm9tRxzt$!)
zpy=!8rP_I=3O4|H{;1&{rX4U4Fhc*LtA~D)g;Yh1ER{1vW4f%25w6TTh<3USl-qvn
z=|ZqF^8ujJ%i$us4nzaOArv~%l4Ux1n~%@3@YP?#P}pqGk(;@()A8)n57@}YS+8TV
zBew@$<lQ+pTl0yeErMaI1MpTPwkG>a_}a9|Pk5wvYXe5@;2U24mf>3dk4eH;G?%3Z
zXMI~5CB7d{^hzZoZIW3PRF%H3svgn>6~Fv3)-?Gngnh^CW(uw>yI)noajg}>Ka<~x
zrF?xRwaS}eSR;N@guafa@aC=2eTyMyVbJjY{X=TLIZ_UO?~6P;X&(p;cyG2YRGY(y
z!`^PRR7g2p_yu`nqe^wlXWgp!`R~nYK%QUXS5)_1Z0bUz-(dFmi-gm5`-baYuhVFB
zmY+}T(i0%IeYT<M1{xn5``NZARpa`N=h#9;4>@X)P~u7+L|yFF8xJkn)mk#Q66@`C
zKN9*Mvz%KPbuX5ew_9A-jKhjQ>lf_QoP<iuY1+zgSPXs4(Js^9Hul;YqApd-3n8vu
zUs3{?2QU{%+@Eb0{V%;}VybBER;fPs8YwNmHO=I${~<^Y^o6uV$7Y-jnml+aT}Iqt
zC8I%pyF|x91;68ixIevmLCA-Y?FgQAr@mDRl6PNBKfTC<ujY9G|9BUqua+{8`qcX*
zUBE)z2<9Cm?7l4lpt`w0$ub-Qyw(EjCZH9)g(n$cp;CO>+U01S1JSN|(^nXukYKg{
z-L3P|Wi(=;<~^+m?3-llyAoXn!m+ykZ9Odb8H<RBioeAZt}Gb7S<9S&15R9CEj`Eu
z*^gt^P&!$P%}pcc^alM@`)g{~by#F%k<2Lx8^K~XSWtVQnr{QB2%J^G&<4_pp9~8K
z{Z6iXG;?wpZ)xH@)ARgprP<$@)c-}lwoOIV&?*{fN!cfl+f(9|!lkDG=+hC@S7A?Q
zjDxNBI68(taek?Tl83KM@DEKsk13nmhS`o?s1C{?(%k)CoyOK;sD!QXnBQ$6`JGDq
z!>4$|g@KnfpO^VU{qZBt>URT$7aYRx9P&zMi&UX1E}H|~Rux{So301*jh9o9`B{9R
zBgt)Z#NkhJ%Z(j{7)Ut?T_i%57_{7Wa~My!GH6#6v=I^6_Sjby8oVoxCH^<P=UXaP
z8${ik={F1O7dO=O9PW4u=YY@YpkEDPQyG={6}giy4vz*1o+Q_9fMH0s$r7Ep;>Dvc
zj=z#uCLXGy_T6!qPV!#Lo_x4(z7dxFQiJkyE%T4p!?61+lqQ!K`vw4-DMT&wM8+46
z7#=Tav~bDu@9yaue-9<Sr>4@;{XqIo6Hxf>LbB)HY~Mh)UwgZX0P>MdFJi_zM`IgY
zcYkO}hnRCyLOlSAK@NH-og-L;iendm_#+ZSz6<uo3_-<Im(VVC4d@^UvwD#|b<(gN
zZat70l-O5r@>;jc13BWeCg`*r$tgQ(L!h)bZle_u%-XEOBLueE(u$H(xv^e5=L1TE
zQ|u;6&2{@EeuvMqQ^E#%`~-xt(tj-VuC4e8z1{ACDiia!H8?%`3bh#i)$bS9<hHL*
zqw0+WitfCZp<ds`X#My<%(`2eJBzotGX<k=xU5=(#a?Ex7l5=@Tu@Q3TlMR#UnGd9
zKh7n2ckb(R8d3?>-}RW?)z#(QkA(D<z0U~u)2#-Nt<IL%$Widj>quj2EDgs^=*HIh
z^cXjh)5{#XI6372gPLDm8ZWr+Ys~ped)`4_Aq~w|9;K8{etI=()-^LJq*2}ly%!#S
z_7f!)0RgieXx%W)T_J>E5ip$QG;uA(ijRSQE@$1~p^tBOEYfOd-9+?7YZ0%Ls4bk)
z4rW{9MX%;!>|S<56BDJ`NGLs(jze8P(#-ovps!)*`%WA<;xI%%ZL#%vsrIpeocsDY
z1EZK$=@e*EwA1Bm^z7*`#iB@*%j!=BpeN!w#4G9gal+N|b;70))ltVa;npthbZFER
zo^Q<7jot!!1!&Ma3<R~&Hl~55_H<R^Doi4fUQ3*qKR=Z`ul9}_JoAgn$jG3Sk(N%K
zzgh5D8U@C{Wo-+e&OHEPwR-X%&)>0HTOl?(ZmOJ;06U=@iSrX_1~`PtqKN)`hw}Bd
zyFNwS)tC!>+}`Unbhw}h-G+5>8+{8DK>JoaktY82-@Ir7fNcojsRuQAY>E3_mN8l`
z^Z1_#?krZ@gxnmi$*Evh#?`*vtOpm&O5$4lq?Oash+GM|x*5r)qdM=c+ZB=9O24)U
zT84Bv*6vQIEU6X&td19wd));dXybK|HyJVet5?^C9B{JLL$*6&h`ThXVL&RUMO-O2
zUdz?=@P&f>cr{xzg}{42{xRZoc(-_(*Wsc92*9*tG}X(z`3n3<qVED*RHS~NB(7F>
zfN{();#49%4A%m`!>vG#NH#VvDDOn50z&oLy366y&$g_fi5vhe7qDU{cIUKpMgdst
zU&ysxv__#83D&zFczSB0C1YLR9)cF+Eh&dsrQc)ORK^|5EF-K_hDz=W*VcOaM}&uu
zl6F{KkXGd7CSI?~2ZFE0U9D$|1D(Z{6cn^KBLk^~?1S;WTKnON9~0SwGd8d6dXkN5
z-UgkO-2k+a1dC6}vZocrJ-sm>KjLWIzQJgMT69839n5}O5SLe7jhQ-hBRX;$0Ni#g
z<czm<9#jV-FqeB(x%!t=LYAM^^*8<U(PouFP{y<aKLI3t7Y4uXD9fmSM6<ogbe+a7
zyn!;X0a^P$X9-C-HvM*_iq4d~JX|F0fRCexvb&Wj-t`VA4`EdbyDyeY84%q}z>63|
zY(5=TQ$PU7UuNfB{r!!y)x#)+MoVk&dzln7c4j_U>-K3Nqw~N+RQxsRIi|@B=@9g=
zs<p%46!e8w6xM5rjvjt~>5F-Y=hA(8tnsa9dwz0!6?g4Foui8z@+l=gcvONC3(Xz^
zjiHwZD=RCSb+$iVPNvXug#Zl~L&Dv-cmDSH{4f6DffkexWi6Db+2LhOL@WMt8vJwM
zexbRo;?@&4DENROMI6&x$?VJfyByl>PtSBs-<FHwK(c{S?O0wYq>Je}`WgiAVw!@*
zycB<hZA_&Ua;dAl0d<^INc&AFx?J<qzS7UUSLw(@+HD0(vdMCTNbzd~1FKPCjWZHB
z{VcoUwbs`)Q{N$mX`{PysDl;-$=PDQ)PNz3A?B5Iat<!8u8&|YrJ5u-zv5_wxHSTs
z8_!v7IPSu@HRR9$Cj6@2zD2m^oN^F&`N8X9)$d>Rnl6mC6e5-0Z9((o`r@)h)A>oP
zD1_h?q$+oXE@7unu2$b3i{N+(uI3b&?TaDazulTzs5z<E$`!8DqlJBY<KV#M5AnSC
z=nwZV4~yznyDCH<Ym}}VZyT*tdBSG@`b=)4Et1zIKTK2W#`0A;TdRL+q}Avq3F3Wr
zpu*7YRU{eo4;SD{Bd1-x)TAZ%yzKKu+TjiOq8w8+qBN35LXuna_U-MKps^%Qtw#k4
za^Cm&YGU!_AK%3Y9LfJssD3w}RBfAHIpS=wpR2{=1bn%Fu($WlqxxP3{?XUNHHPNr
zr{BFUvo7Cl?zMQsT)RzdYYVmr4A$JI{LkDBOw0#qhmMq7ot)~RQz$Dt9dFS@vT`sU
z2eF-|UqmThjPIQB0(T>GDd2Vp9%D!cVB_DWm$(yl3wSarwY+42?r7(8sfmQUoxMP&
zN9XGF*<>x`NMQ}6ax<uNB(dpq16`TM)fedQjYl0wfz7l$#=l=AP*<Yu`3Pq7!@XC0
z7X69TCSdzn>@-P<8JonO2OukE@j8NRzY+1DwCh+W4A-M17>2YB;*NffHR7T^7ZZ}j
zS~Z>oN$Toji)lmzvQU#u0w=>p-v?`Q(H@UOT%yslq#fitq?JRTtw5%yAsp7(NMCsN
zn>5vlE2fCCiSU(~7IQqs6t)jQu4KKfyK$A&ufu-FFWHPl3VG4xNPmZCrYPT(>bdC4
z0yp*>EAk{(KChF@NhgJ23KAZb%&qN7r2=)nN5B4A>>UENSgRv*tLD*}O!wtzo?I%w
zld!w}Qh|Q4ZUF``Dj}~t(hx}hbSjKjQ$8p;A>j|gOv_AHFuvDm)(9)mimb6HGm`1y
zZZ)9n2s5AA#_wnkbUG%V%aQ!_%$D3xwWC09ClnD-P!-a%>cer_H<iZ1>h{kvsYi7t
zu7vuR>be}uKP}yn1SULeUNFH|&ibBLT5K&*WDV*I$ShRsK32<>=I5^44DFR1PmNZE
zXw_0^jWkKwFe;eg7wVJr!x<u1^uG;qRJtp;?)9w0@D7&h@$K|aL-}A=S9sED;kZSo
zV4po(2rQ2<H*Kf-Vz}2#FuXxXGt=Gx*8=RHyA$Nn&JVnd$@qhVr@>PDNvS^|jr_)k
z?4Tc6w^!v54Ima!9JCKohrbMOI}hCLm}?bhN1JnV8j4kfs1<q&1z%Gf{8%fMe}e+L
zLP~oVrrz#Ml>D41(J*w|D6<a*)LOmaGhV-Uou>viyAg~mXXTszsq$`Jd#KfOYdr&c
z{2cX%<kh+Fq2XBjfhY;_XD#39X=CwvvDPt90#`-_uSX_*8;<J-Q-z(<>*NkaMu-af
zQ}`9j%t9_V`gOQfg6%CH;JpZqIg-1rkWiWd`;<~)wpZ*V2)=kPyfwEOJ(}GozzL%j
zbUI@U!9APq7O)M5DgG#}FDlCbq+7X{KDq2n-Czv1+C+37wuC0IsJ?a`A)9aY!z!Dt
z?Ib<7W_bNRcozA6qd!F6deaG>Bk>JEa`_x+xa+~AhwER4mcqREhew<I*chE$#*Z32
z4qFM`_m_NbS&ik(qQ4>IL_AdFOWVDwNlo`l-Fg$aAc`_qE1~4<JLtD7JP1*TIPFW1
zeFM76j12zPFdYJ2K@Wmt4&xX}8Jt`T=z)ADfArwBAMvqbfl<<Ffp6kD!x{Mq4(;21
zLI$4(zni*<MfamNIxQ$Rvgzz+@SDDbwT(zpR<l&;#RT=!NQ#iN%KItB*Y<-Ed{<5(
zE2LZWA-kjX&4jQp@H`AHA!G4eCS-T9h>}$sH9!XrlaK=|L&O&}t+fnL7q6s)h}tX7
zZ&VnSl5zc7#FCz<fK0}?!nJiOo$nTPp9&-D-mIt({)LM~vCEzk6T=o^U=x{m-WAq5
zTRVq~tL@HFA~MiBdV7SA)L`07<FZD+E<whX?x|D-O6*muWaIQ^C;)C*J<S?)U!!3%
z9s?6WQhPFWcU6Ji{AQ`%enH)=p#uv-<~0@%`h4vEK}@UxJra^BSC_IioTYb^Tnm!l
znTiciQ^aQ#e`Z-MaI!wUKY=hrU;fn%$qFOtQ7LvZxGBnlD7z%me`uCMBVA6Rsiy0W
zHsX#Us>VHNUGRjjvTq|e+S+gBICa0i5428u!gig6P(Zn!pXN-L972h1zBm;L(O(V3
zHuge4nSN}IF>du0YQ1K5?!J9ZFm=8=9mlMkF1SBDH%r##=&(Cbx%Au79(Ch4xNR|~
zpO-?gAh;0-9-00;!|?k%In^bp_N&)F0&X2Y6P?Hw8=?7$2kdmZ>)74f(^3amW)2`n
zBp7XGSe)-BT0rsiQcutI<YzX0c2^>piDzCnlT77~WjwQ%3G`eD=)UqxIYgh)laND8
zxZubStz6gl*QJJA7veaPSh{8E`9o^$nf^g=dJkg)pVUaPQRt#^nG6#%UG#^aZ1fmP
zq#WbYmSuDGnz>)m$U~DTfg$BgE}J~BjCYPowlT6yo{H;yHryYd-u*-cc03ekrug7)
zGI4{!(P~p#@U4~)#=4_7V##zXk)FP->3pZy(tflybU_5IO%LmV(8`r6wd$4;)%dCJ
zn7f{y1vEDcr6;sp7=QH)^a8QJnpPs9Guo}=IL^ulOTbsaWs#&&NZ2VxPhkkR6c3P~
zZLjPSnCX#)qtljSx!mQ(NGLPLWF)D12JCZ}V`1W#(aehCq55ulhkmvFX(sBU(hZwu
z9v7m6f#RI!+)b{BmQm7HjZ1!4%XG}a%I&RUHxt)@#sqJEp0ephTQS^b_v#xGi(p;A
za57EI!u{-ENki#yySSBuvfi3r*oGi`g4pOL&L(#iA_sw*PEF{=))n~9<#JQ-qGYdg
z8NS|aeo9oda-7%Go5mJONU-@mLFosKV>FAyCoU}6o1mf=q>D3)Yu0t5a(tOKZo$oC
znz%}q5X<R+5={pyCJNZR&UgX%V36%Dg+j&vlHn3kISBg<y0|DhGl6p<BT>YYsQ8}i
zuY$~%n1&{17Jgo5S9Jb{+?c9RF|7wRRx`J$GrUtNBlBSC{B{ulky+Nd%zmLsx7J#V
z>U?$IAq4Qf;D<sOzkP3*`JMMH*XHNl#HEvX>P1qYD(3WL{pw=g4-R+A(Gq7|Ncd!n
zrRrs%B*59A5!#p7zZ@{<auCoCN$h3oeg#Q)sJ$vDZ~bgS8cw*QPA1#2$k<{sBM%Vx
z0`S4>T!~9&t>uP|@4w)m3ehuL&*TU8BX6!o*U0wX0?bdi`ZJge$qOb#DvHS8z~YMd
z^wyF;Yr)p~v4+;H5&^cNboF{)g2CD0cP7fU#HIqW<Ervz+5MfXaw)b31#8&_1U1N=
z>YG2EqToAgRr=Cs8k)lM63w*uxFfx)BT{AQqRIWu!dmcVqO>DLNiYlHx6iiMJS50*
zR^AJpvF@Q%VVn%q%;$#%!?6$j{_>{xSTr0xsTc7$F#UdIHDmMWdC3zO<CZu+1Clai
z&<$S6SsBkJua!m~Ock`=mooPny}k@nY!kZ80-?O~kudY~fq;o6lIV53be8a!%K+iU
zJO_cK6D>%+mH*2%7<xI7kVfB7qw&3w&sWc5!tN-t$bM3;Jj?Lhp9kB4QfMmmxbgf6
zq<(PdAn!_!n@>67bNy{T?QIWUgqB0a)I~Ph7oEjaUbi<mdVYB@-%+D5o}`+l2uu+;
z-Vkt$d!{BCdF^$4a(8nm3o_aN;JqACH~q`#QkV#Owm>@9YOGKVW2$tLjln3ywbY+!
z^?qi+T6giAbXCDNCH4J*G%-9IopKH3Pfi}Tck4Sh&q_yYY15$K3u@_Qe3y0K=!_@?
zof<V$*=~Cy2q5Wu_PziQLz@G*8QLAszw<GzpHqn5YCPH0T&v%v<lmFYr9$D`Z%PlA
zM-g3~ZEFmqnr(&)n@olbyfsVEomN4=X^9g;Iu+8Dn)?&bUhW2-(Nb*6eV-bf;{~yz
z6r7k;RN}@Ka?r9(g(e>Pjvn>9u*wzMn5z%_R(jv^&X$T189jDyDj|1Ui^MXvgHTH7
z0$_hq?K@b}*5J*C-VQvUn_lfL3EDaQsOS7zWyUPisdEqtv|*{mEv<uhc{~mbc^wu-
z?S)Ss)<|;Ci#!Xb!uTBwcT!`_UDyWd$UBsM^PZ%5tTXksX~tr+?O?O$qOeE^v4o(c
zU6sE1We&ZV>$gvCc?F0{&^U_W&9}H<jWetC9Ax@TWEth2PuN@$x0afa+S@DS);hpE
zF2zX(4T4Rk5rz9`OxcKAWdSl`gwY)A(!kgR^%8!oV-5=bvI?L6-E=1!kT*sU*3~wg
zM^mPxx-PFQ1Esl*Sr8Um^75L8OvIVpeyJ(1)5Z-}PI)hsErrVf8*UqYnwGEd0Ifze
zYb8m+W+-FH`+7in4}B9r7!rQV?q&*>cuNjGS8bx2KYa8e-=9nGizZP&*^d1K0fBHD
zi?+jhH}GS`%{J?OVE(i8H~5q9*P`A(!y@)r%v0!BzX7Vdlnj<KVq6efkh)~yxiF>W
zRMXQ|`zvBX*1&?m9$-Obxl;-?oKMXhlnwY~Ryz-W)CPrNn|8n-qzZEz8s$m521!e&
z$bHfoQ<5dD(<=GtIDdHv(xiV{O;#?#s(n3R4>~(7RWE>_eH#LKSfB1uKEv)(eB{(9
zHn$36!b!G#FRwtlTf9`!0i9}?t*S5H0i6vKyd)EQQO!|3qb?nrygQeb>GO=|nda#N
zkObuF4_ZIFBU^133XqH1YYsb(r|YP9ns_#p@@$5OJWcN9(H|kMaR$=(bcpH{CTdg`
z7zqnk06n>3s0gHzp@(fH?~%b;-Nh+98ulU~Ihte7$OPS=QlTl`Qvl*}=}l-i`R31%
zXsWsQ8B)&@0XCe-0e?1b)mNgk74x^)pZ6B6+q3R)piErWipcYtfZPCkn{{Vy1^n>S
z?vAe(7RMaFlo=B{?=sA7Tw96zBletnzT;1rH%c3$$zLp%{)2Q{RC+~2#d>SJyCjGX
zFBa?H?S}6{Bj|F>sBL>C(_C@ym?H9i0c+O+j-|O0y{FBOs<-*>3Czk7bM+2VC<~{D
zw<A_Qc7$6H{7Ld&Vb`4523MyH6~l10o<^h-@0`TFNvz2B#OXG8w)nmaK@x?MLg%1I
zz<<A=gsYcCFwGJqGhb#dzQw1H2b%4lR}oiX>YtP~5@!UE@kuiT?6v3<cG1T54Mg|J
zgmCZTY}7ZPb_&?Yh<Hf0qRbi!IRJ*x$j0Tp9tc(FVE~;z7(?!%hT6{O<I&3^$I<rQ
zWU)^XsVPrgn3-I^$1xb>ADbliIyQ07P!#1WWD=t={S9owgmtQ8u%`f}N&$gLI6lE`
z1&>D47`lR<6ccb@0*lIv&S&(W0kx2Ogp4oAN|xw@^`G_N3zY(bgO+p0B8=`X8e3xC
z2(stQ!=0NC3|=+rGu0f<KI}_iauEbHd?Hi0Dy_F!i68W9@>ujmbt_Yr{?t5V;?HNO
zeT#=SzM=I`<;H48%@!CLG#Nexn1J`5Y7|v99B4@g11MISDgH#*Oi`Ykk7lv@x5Adb
zYx>+YUi*l5=JDcnZ<|R-JQs<$=Zo*nI7H#Hw5I~Pwt`fbK{uD&Djv!e{$iRSs_W61
z{bqa$QT_84?=tHK&z1P}3Nxk%$uY%3m5lC<3)@eiNNHS)n%fp|>hbe$6goCOE`^Jv
z!%*`@2+kMh5+RxBnRlU_`?0)UJUCxzL_LeuN7IAtwx8#R%2jYaG5q{R`R2>p`nn@Q
zw-a1hJleb;uOj&e)hbXa@~KbM@~;Vv6B6^9J;;LF>ku#R@>ZWLQN_P^li+&O2UlZk
zCK5?Lt1!Ey$L1q8KpX?qpkURsaAlZ-dQlQPGW=u>mwarl#$x3duo|nctdbH;k9py{
zYoU;5RtU|#tMSJqk3`BefvD5T;53<o%3<+q!JW>~sCNG2(-@;UhCwVVyV;hb(Rsbo
zqg@~_wo6kjnNEk)+s+mSm#8a-eJKOu<-d$;&q3SI)hNqE@)|K%87yezhY_J&x)1W2
zX-<6))~s~qdlLR*AxCvbZFv2KYnz#r)?CK*)?!zlsSKw!0wPXFz=rV;2g=TJ)(YxH
zLr@Utq4&`N?+ZXs_IV1`y+3CF55@`(^XrTF?b0Mhdxeogf0nP%m;O2ZlD}b+e=!pY
zU&-apN}|*Yq^q+0>m7qdgjXWA6n*ep)RP3o<EKv%DCDK^P?m6{pU<U~y0PLJy`Pin
zN(@k>FP2wSh*u02a*2DC=gn;x<gojx*R(=|vCM{tV1i6iuEcpDgseOu-?o-De#q7R
z3QyPJQCZqUwj7mdcg5njMxT2gON#6^mF*I98)n2TeN$FX(<k1xiCsF97kv1TEO?cc
zwEEtlSti*~lf55$*446XOikE;(P&4tded~lI>@n0vudL^L{W<^_tnb#-qdX=M)}lR
zKxw}F$Mda3%k4uW9}YNI-FR7jDB;~7tbh;IG-F6uR&XrWs%PGr`jj_jm>yAc+yj50
z`}K%av;i&QXIEh#Zu*YHR>Ml2Ix*FdAy+j#n*~GA&uX^n9tD4i!|TwYRT}7Ub$4$Z
zhX_Sz{vj1;_T#4ZP?BaUkDT{X(jkOO`29U|W~+8Sq9R}8>hu#g4Fa_vJH$4}yF2IU
z$TVqp_~|+VU|6woXBqG5ReaOA+R122A5;)lt+Y~5W92X7y6E=X>3kY|C&U^<;8>4f
zc4Ht(iZaEt;Wbj0-RLV7m&sR6L+{HWBdrz=RP3QDR5a*SrYD!`;(V-pq)Y^{Y_4>O
z<`fSzK>?E2C8m2Ro=`7$yWnM#`fe*9tpfF+fq+yYgr~G-h8a~Ot}1O)=PNJ$K0bb?
z6Ilmm)>UlUgl=(FwW*Jl`fH(;>~ITJ>}%<4pG${So0vH(`(E%}ME@7C4cJ9rA7CRI
zJ#gGn)kbcUVvQo35QB)C5~f_gd6IBMvbdkOo5>raWG<5!@AI85yU@jxsV5;m{VcHy
zEWaVQh8N3MX?d1glae-{y|4A37%wL@SbZ$}51{vZr{de)u{VT;UxPT>3PJpGH~PrC
zTH+O->C^oZnW~l5VzFywA=r_OUwtxVZKg^B>oVCtyrX|cIRXBnZ8uUJApf3k{2TS0
zhnZ$9-B?OVSgQ0+t=8ur(@*vUNdxK3^deHTx(0r{lN@nhn1MJkb8b*=Qu_1QC$MV%
z<h_nFKRsR(`VR8kKD}uUD+f7T%93E;)(+pjm^qqFddE0RHzm?E4Ix<zf5uXc<{`Ty
zHIg|=K@B5%RvaiVCit-K5%tDECeR*I&^cZ-?07IBgcFx3y|^>c!f-+wn^>PLb4M<I
zYU~lYO0X#XvWupf^WZ0)5X}@GtU`JW@d^NH(J8R=tB>H5jj{bP+ABj<K+4tENKd~%
zS;4~No4UBFF66Og()>&W0mxt_n2ifLZ%D*5+Iyb;Bv7Fev2w}`^I-a(ElpdkCA}wf
z*>kPi;v3zy;(Pam-s@z&lp`Er$x44m-2bdWiq_vg%Da9wC5gl>eQ%}-$}v~hme8a6
z_`UYii-ta$zj*3iradpUkufcX^ri`bja@HmI3$r9TLq~%QtQ3gs}(I+XPQqH%Dnos
z@cR0h_1J>f&de+cI3ggemy14eUlLIfx-~2H59q`5H)fKf4LEYYroq>H3xQAEypc)u
z;J5So=gSW>xP$UIrP%zOTEI73k%zZ4ZTqM`^I09g+m5E|_861;+`+*?gWE0%A&Uw^
z7JgpGugE<=C+xB(@W%;0#mEYznQ?22^=G(|6B;dkry-HcD=n>y#513_**#Q-dd<f!
z1znB!{{*e$ofI+QIr{yA1LJNm4<KgdU`Dm27;3SYBqbev?CRX6H2mAa(&nYcpM<O4
z4d*`jZj3HBY|T|CP6zNXt@`IZNrFW#lnqSo;|E-CsDj|zosK!TpS=1KFnEKfa4f|9
zhDwD-0NaK!T@M6p?M>&19K1<kBjqx$$yX3_QOT_6WdDHk^Q2LF`N!#2DxX`O9`2>g
zgPQ#rS()Ifi={A~iZ9mJ6vR10niK*y!44S$>QAW#w=$+d3UQ6gEP$63MClMzJVW=~
z#b1fc-%r%c0dPA0n2zgn4fSeyYxVFqwdh0S&`V6sv+}|RY8vzy%t|{GNGSI06b}7D
z`cXYOvKo8M;A8-ym1=SY4zy{c&!tu;zXmudwn(2T<0@_M6otfMZOLcO-d)lt(2o8N
zdq^V~44_o{(0+gYzi3)Sm$O!p9tWs>RErN8TgjVfchigdPGFjNicprgSICE-+IQ5I
z_4(BZzJI<Huo%wj0s6vChs$Bqv)N(m*3)GlydS_M0VzAGp{42~)#Kgtr9`s9M|LAk
z&LxK#vgffC65_bjkFp=m(R$pW0eQU#qS%gfVw?Th$}72XZRJ-j@A%jasw7WM3=prY
zK*t@=Syat8YiJ$!dbvZ%q&H}FxJ<Ono)ZnHu@W8Yn;eSwBGlfA`k)5h*GtFRTwp1o
z@E3($kFYlf(t|WuR7CbN!t!8ti)0&Y$#S{D_w0s-lONEBP*&_bRjaxK^!Elnda5CO
zW};9gY@c|XuBu2T<9DD2chE!)uKA+cw($Ob@TPI}N6Vua!SSFz==!h+6hvB)0Xfwf
zD)sY8Tiw9l6zMD3x?Aa_zVSV&@@;>zyq3q$k=sBOta5Z+#E7NK@$l@-+3~j%T)uxN
zx!fdb<jAzFbjqY&E(DK<;-YQpatL92{Tb?skJnNc^f%ry<9#`N9=!YXlIT~3qH`4P
zcz9q$1#-u@L5l~RZVj#NS-Avr67awI{_!5+hhp7u`rf$bEb+`rtBGJmO+E#r1z~*S
zDoy?-Dc@+T3ndVHj_<8qjhY`F?0(V$`MN$nA!Rik_s<gZiGjXD<=?SVFP$uoh@qj@
zJfn|XDkMW5j>&}EThK^3T0Xu{!fDEl+8Tzw4hdVCA{&dIg69?z?lC#JR5!LNly<Fj
zNNSeK5Uk=-@!XP8dQ@w37`e%%9fwGimlN=3!lo3Wxj`+4hA2~Hd+ip=$3A`?HLbe6
z2Plg<T`j++f^Xr!K@EN67rjcRH_I{fa?a7*hNxxB(|v`C>P1aCTu)w<Q#V(7TpA~>
zYD+7=56YKL>eYg3!5?k&<jb|+xwkh3R*m+fx*(S^0DPC^Mkz7%#n!i$8|-nTHKA-$
zR_zipCSz9p+2Z4wdJAZKjuK?3_*y2~BOZFNkW`uORJ-9Y+PFlzGSQp;iX%5odh(QN
zgh(c0uI>`(yI!47tDdPfhg3-sk%l+Cq(25EG&`T(eDSAoLM~Cx)Olq@y!5<UCg6iK
zI&Vq$jWMZ}=-+)abON|-vGU_E{)0|?{{1yA={rD&OB4rX-AWq@{CK;l;1Q?3d_hL}
zt45453;CiwwL@6lRR_f{@r1RRQKFezKbD8I!V%T7w23GZc6EGWHXVsY3@>VUXcYQl
z&_K7^nc5xEC>0sN;=f5NO6G;L-f?P?{B<J=gbn_#%hbL(DkaC~AL}MbzuVV<L&{<K
zZMbgl?P+(}cW{9(yII5Smc%b=H+I?4$l<AIgCul%c^s;#NS>eoE!$&EUF%@uw|wnh
zGd&M$P1wcU&%96(-?(|$dRPhPz?v2*_i#G<x~Y787?|gCJa<t$dm4~N=Mo3fd7&dZ
z74A-bd9WY_s7f6lt#rsw+s}Ntt@TMe+C39}j`Jksro!M>kLvFT_94N4P-vRs)9)$8
z0r#2MG7j%6>tr0nU;Pb3&-0DZSmK>wo!q{T_3b?m*rmI~oX2wnwIOf)w-@*h7npdo
zpw6hSgyXXx8gEnVDBLA_3TYh9${M9<OaQDL(7_9J+nrjuKHch#rNb;YY7V)ROZFz5
z1mjVM3=U`3%ijw{^ZkQRTF-2c<;yjPWTI{G9l!n?g8wR8{;!_|dSC?B8fdRBhhti^
z3stYTfBe#QY;6c=j&u)$&{4on%QmRDKQ47}P_h8RD}#3t$UhZLK{?a8ERLi5+hw(L
z+3$$f`;xhaZod5${@=gnpJW?Y+VVNRpxeqFe?8c%PmG;5?5v5UXWmICJD#Dpx*U(~
z6@mH762Crb7}vjv`N5+>9Sk?N>G3LXdhYBgSjyFBRf5VA>kRV`+VuFpq5%>#|MRDT
zR|A;wH0Ry>yE{mSh0^oowG~<uW94dEPU!$MLDKhqrGJ-vrI#08Jof#@IAF|U^x#Rj
zfx0Vv3iFV~qkg_6J;cUAXL#+eJ7JLhW|~>7{UFo}|Hq(sT01e~)xQgWnA2G0c)z@)
z^yckyh#_sGxU=6ny!SkAmtk7YhGrQR00aYg4WM5aL6!%2b$g3p`!{v0C|A{knCrD<
zdb=IG#Z1z9s_+otZt(u~`&K;0(8i_$xjfqcs&Bb8RjNx0eZ+p3ab&-gfgTrHh8F0O
zUvpZOy2%E|(z1I6AS_>Dx%?e@1n&O59&eiEUrYEEhBScIQL#5t2Y8LYFP`H$f(MY?
zmp${eF2IEo`#z*%p&>+1snGfmeDk~DeU*ivmI*Lx@NhK~V*m5q2>lL)OhZc_TO`k0
zYS$+<-K|IEkb?|RM^V7h@uH29ya1@#{B16QE1_xYQYU>8@y7pf0e~`&ue>-Tt+Uv&
zcklnoeA+S=@Yr>cL_@~HWxd5OkDakkGasF%^nK-7WRT!C{0gOakNurOtP%zm?2<C@
zit5@_`E?BSrixc@Us}#y78W-ba$Ws!B$5)v#s0IR{|f^?+PuW{!8`1pE&EgxGk}{G
ziD<mvy}@Q#N8E$`3YWAepjzn9y@6rRH@OfEdQbKQ$>C|Tu=%RrOX~Nf0gDV|{ew9@
z{%<VwZ)*uqzC$DK9ngovyu4k_vg4T%LOoJhf;H;hLy%2&Kn*p02>JcVKgvD!j0b5u
z80K)wLzyv!1@?0ykKe-gAEE`Ys`&4^ZlX7r>PE#aY0t3Wn17?Q|8oQR2fO|A*TAbb
zOf~_|jzDvt$(yA=u<@1&&FMfr34oz3&X$<>{I>kIBy8Kc>bu5tEw!w$vT+Qc_xP1h
z&7FGz*hW+~wQiFO9Y=!4Ierq6E-L89pjg^yRNHybF3Tm}Jje;%iwxw(yMHM+PwNEk
zL=r?Dwp}3thVlM>cto3O88=MI7k%eC4k7Q)2$)(UK#}0(jGU)!a0NRe_P7yW-ODdf
z0>2%}`gfgy&um%rz|IiOBy*4avolBnle3~7FUH6xEyFRO%b$iRQx94uJeoYx5HCe7
z@ir}<|GG2CKww$w^>0`FUyO@8{Z|8`0H_&_;uN)@d0Izs8}jxg*6)izg4@4LY8BP{
zG7NYHDO-5Q`t_O<atii+c}Kepixv;i(7!!6wMYi|@D}c$Kc4-`BuMYHlfG=<V;a};
zh;L&0Acv#czye-EjJu7iHe7?oyR9L1<cl9y2FVvi#e)TksK*{!g&IQs#XJB1JlwA@
zuq_>n{{TU4aA~W42wq5NHb*rS$Y;QVu1>tukzsG|JPC<>8=xolXCV%EVxU41?#3K7
zDpO@Ae2T<RDPbami~mRaRD$tNxiKut3I{gbdN@6bb#B8?ua2ez-suKUnLkwV_qi2E
zaD;o-R*oq$5eNWV*&EaQWa#MQCE&q{pzl9*ga4hK=|G2vsPSc1F%d0TL?=Hre)zgI
z;VK;g`sFgwmG`zB|Ln@Bwz2a_0t%YXexr?6cyCmo<npn%!eSlUyUpOXNDzb)|FQUA
zJ~@86s}ks6zOxr_*uNM!>wt4+IDnjqwjt>*L+5Fo!M*>lixVW~66^>`v;;dwnzMsw
zXZ##lwC(f@S45w`;XK)5?bTUFbN{*l5M{worQbK;uRHb&jp6TFcGkY-VePmSjA<Qv
zUF+A;D$e@o_8*3)l!Z3b;LP6LZ$~MK!?BRS(SrWpxf$z$y32uBc!%cgA4fe;&ceOX
zbA_+5Ty1v<9E>3#-z$7B<F68>nkB2s(8bA^SGbquL+m{=znAoo06S|PGD)dZ?81|e
z#9uJD2#j#CWhEqw`9(7M7H|t0=qmwu>3F=pP(!>1n(h<-$E$Dea2mBnVct9}29t6b
zwwEM?xRR)!tPcjwd%5HQT{ghrR>rXcrO-R~+;9?uPz%jIc;7kzNu_cz&SoWwOFy#v
zq-?g8;)Y*6?*;k?3i|Ct$%q&=dB_5s<c|jx=O~-LVVLPj@KX{uZ#<f?VH@qp8u7t2
zG16)agsGGLH_5xtdWW&dpIbzYdSAe7&{9VuD}ug1!`%IV$$87q-#k_?R3v8tEq-;<
zSm8RI1qA<i9Tsr!t-tI4hFaZ-OtD6J9L@*j$;GVB|2U2Hm+wpGU;|bk4d8=vKTh7p
za{4tr*<JI$SZIO~-w3*FvZH>WA#jfsgpA==N44tV_aAWR9a=@7=zo184&*B8MJoGe
z{wT;f<ug&!;Qofy{)*jPn@djlc;4_2=)vo6F%|He$L8HpDwjILq$<~?PGS%|--6q@
zYlm9>I_9=N1=oiXa9XAOfU0Wplj$Xq`UBIYR%FCl7fB3_cxkChmZ*^P)?hTf49+LF
zJc%3(`FKW!#Pc|=MUK{GUuW@aU!D5=t`q+(lq@(8Ho9cHfB5;(;X3Wmz|5yb7d}V7
zCt~J_N_eisIdIQluj)obN0Zm*!WO_g?gGK4%;_djEKM<?IMC2Fi~-DtGc=Ky@bA+{
z!uQ`;&#k&kXNeA6F1ceoc=j>?O_|j~gtaOml1=+xYcLt2M|;;Udo!(=Fn0ZrAgqVN
zer#$vV-}XzMmS8a@~MvlJN$N$&6oJ=*4*>?&zQc9-1HRSq~M!M0ew~bB44=F3mqXB
zDEJ~0i%iUnx6bZ*9D@>=cvGRrq3#Fzf#1>h0Tn3qs?zy`Wj)v0T%8vs!Y@>>hRdu0
zAZ7zo`x4a_1*pV5$rh1+tpF4?Z1U$A1&Uy^dRhon^c#>^9G|}L67<>pta0%Q7>`n8
zzeJJ7<4=}%+*Rn805a-_n$g0_ZXU2DmpjjjOXF*mD9Y4W&k@qTLp)RTS@`0Q`QC^F
zpak7jNMNk|qH55tl~AEwY)JRrf3|9_jNi8w(C<2;;lT$`xhN0cxBIlJS>jslCIp?6
z#ah&2@Uym?zJwoVHqz3|?H?kqycX$gf7EClD{0CE*OblHNOZ9huOtg{U$+<kc&yr(
zTolG~yx#i$a`@kRs;7&nBVvGHz{m;}Xc{)HtNBFhX{o^nK{4YZ73$9h^(r>w3k%c_
zD0n5%D5P4xe4OBZfAb<hwK;_2y7>Oi2k`IxJ_qCJ^gCESDCWEvWKINHTA$Hk_WX3i
zaoAm)x1*;0bX0jC2|>vLLG)LosN>E$`M`bX2Uy)<*r-utm|=R|6-jgkzbpU<x?j3n
z9xTd@d@&_{JD~Lzv0llln7R0nDShD!9U=Kajm&de-&m5{9ih1rIqyeX4+Y)VpZw)>
zv$8p8K;ovTS+~RYNv`y;=m+Talxu&lqx0*a?(@+fc|?RP{#^Pui@$RGoqR_`m1T*{
zO+kMI*$~!Kje@YM)-D|IxBPnvY7LgJui##5QsLfj*HCN!4`pv17UkP@4GXBCASK-l
zp>%f)-Q7roNOwvj(g-3T-60{}4bmVW-7VcQ!~iqjh5qj6x$o~izV~^)W9FD+&_Cvi
z^V;X$Yp=ET?{M`bOm3`Dbh7A6&hY*s`Ff>7jYiNz_scGA;MHlWG6-a70!ZpmGTBM{
zqRC?FeEXl!sTBlB$1o2%VAEg|SGGpOm`xi@y38g+N+a8#P<hXbgAdfQU*62-qi9<w
zW|Hn=2bH@^tatyg8W|lr-)3QT8|M^g+xrwx|0Jfd?YcvIdgFGb)&?d1bv!agSFW@+
zr}cQpmY|R?^l>B}W9NrP)O74l-5^G>=8+|SO1pVZ)ieC7)I|rWJ`5|C8vOal^L@Pg
zV3?SHcKgGVRoRKO;ZO9*-+?G+aHjMPhA<h7xw+m&QWB`tq@WkOr)Qf(GE=37twcsX
zr@;&<CZtlpo|8+5`R4)d@P8h7@+6Mcd%%@zD^K=I&Xot!YKFeyYQ^9G|L>X&6CXuF
zLc$ond{4k9s3GrbWnF*wZByl_+POcGg+=&hysD(HH!fBnLt=TMYgl_ZoYX=|PYTUu
z0w%y|GlvuY@>SiOFTg3p?*cf@#E1MffvKrsil$l;VOJU0Uy~#AtZHaN;y9=kO#vs%
zSc*Vlp2Cf~C*OKd6OGR#hH#3U6Gkx|G|F$@=RNmJ5$Y@GC@{|ca6-zDlFYuVme0?2
zQKDT=`HMv+pMDCznqVY}c1p<r=)Xu<_FLL6H#~W&sxl+93x1o{>~SQev%`z(jd?o0
zJynve(8%7O!6C4$oS|QPcHBPWWDH06Zcsb@1MjVp;O&)E?Ve_CaP>>plX#vmk?s<d
zJJ<p_0gqQp@QLhJ^CEBsV89i~pb}tzD^u=8$AUYO*&fj9eYsw5<X>Mw*C=IMp|mvb
zZuk2N;;$37CKW{%jrioOk04MY9Gmjyc7Uq(s=${^_#}a=aH0PG$x2|5rk8rBL5{<l
z#M`SQv-$X4omSJ@9!lA82A|k=5n;PSPv`~RNQO}8i*`GeF)qm1hH%L>9rWT#f`}%}
zj1uubtNK+t>D8;l{R*oYa%+O4fh9KoMYTM|d(jr6>F)5CKm?4D@s*3CeNkrAj^611
z)`=9zK^xKKrzQUG{ih^>8y3+y>Ds8Ji4b9ju&SQHRQ~`|ep9IEQ!tPjk%f>K{-4zn
z^BhLlXDd&Uuwa?|cj+mX6J=cN?*_vjKYKa5NsQHWu+WU5POCHlst9+?Dc2^a`A}8s
z-|BHphIkKf^6cHNOB~uH>NCjRseX;THt41iM@F{1A53Mzpq@|0kyV{vX(m_8lUcj)
z@`A&Bqln0!%lanKDq0KrLKR-Rpq-twMXtHt!=1iLgm6ftLRW(@;!&ibH0C-LV0x{x
z9oJju;pt&Lw(h<YdF!HM08rUG9ygcono$OI<-6*vB&%DjZVTWfG8?M>c-grOh5eW(
zdn?>-uY$3{Z4(8MaTiMA0Q#cS=!5%CA-!k~gthK-?;E9h8->1~bF+CbvbaGfHt?d|
zp(82Qtd@^^P8y&yyiRp>ZKpRaWJ;~4C%#$X5_f%MPUR>3TJ2|&U+<u>1R6dkr%vk8
zGxgq{2qC`BlCcK3<Dj-1qzAce{2T(TY*v?f#-I0AWO6=gV#8y|yNugc0~@cB*sX}Z
z!mIRCGZCMh=ZbHmU+4Yq3#sjjZa3ofM|j!i7qis-6H9EL=k6(Bn<j=`PrbO}=)dvh
z{zYacA(F~?hNbeDPJ2Up#O-}T=`GC}{!QdD_~(Ksv}<7(LmSrm@pwCWinhu~ntZZ+
zPKKxQbYpZ>_!bS|NW@Gg15+W)2IKp|AEU@y{k*+_Uhsl|#{u8pnnB_v!l($w0l%?K
zXZ$N~;TpTy0ng*aM_5fueuyU=r>rHU4@EyvPIcZfX8wyOcU;$x%(fY{P;YOhAVW}R
zcds5u_v0&i`ohML!=82QQret9V0BJLDN5yZQ9)^{J20Nv0r*SOL|3-KU`w|sCm7@1
zj*zyS=VYxvu?EwuMw3si*T2Mta2Ay=)c#m$zABFQA209tZ1P&AP$UmFX9NEQ4qct<
z(+dxfAfVmV&KIVWL(?9@BvkTm#h(30-0cX`8A|8D$NQ;UD9mHI_z1{AKj@kAd^_@&
zkh!>Tkkx6l#RahDLe>lp?=r3Wun`HgZIt#Ee=d~Bs<*=*vPJ#+6J90u+|=7Ii|w3@
zy(wOdr&G_e$jk`Yu?N{!5@kXxXud-2sCuK&Cc4#X_a-Gqg@xiVS&R4IFOK`={1A6A
zq9t}sfX}`$8p&)nUfTw}je8xLP%Oj0Sa{gQnQ?!6sfXfd19UMAeZ4?dRe7-oS20F#
zkt`)N9;ZQ19Oie?DqF>;{2L|MhvUhY8G^p1GulJ(OrJ!hjHX5%d{!BJ{g0L#OF(Bp
zKZBU*4UaV7!w}1AB4&5DzE5wxqT=#=fqsfU*Sp;Qnxd6&K8UKtwiGNLbgx>hARBkX
z7hKP=k8mZw$RgSEO^~b=66%S?Vlgd##*wM*c)tl9f~>bU8~%L5{sN#3s3@uYp|~rJ
zUtXVX#>gi!Q`$^_=WeckfBsIRSc9(36_UP-45T7IHJmB_2C-F?zhKSc-Tsk1lThrY
z_<4c2Juc)~<nJwA`zpoYzHp4gt^<<`j5DO0SOCV7Mne2sbBkVFBCm7G`@x>P>{Xg%
zJ~$YArkLU%`7s<VCv@jzC{qunA;zG-*Br3c6))2}52jJH4ltr{C#32MCn(c9wwFoi
zCVA11dz%`G&umh7tdhnavDWhRMaSHP1&|?17As?aDgz|*^aiFHca%jY+dngc2q)eZ
z@5tOEU-qy>ezi6`fslVBsx96>VLwpzOS++Y7c$B&_=s8WO)U?|lG=W`AzUmN)wI8&
zvp=3sJ^oN09<S5OOBVhkK|t9)3XLp(O31UY9GG{x-qnAbi$7{)JfCM}eFx0#2}$(u
zZ#jmwaf8Fc!ro#Cd3l7^1V8q&sW50Cctb6we+}#yyvG)r^(c%;;)p6$ldob=!lQl1
z+No~qm)kD5cb>y5FX8yyTbdvdKf~NNH3#BSr}XQ%emC!ndWF{2MJWSc>lBJm-zH2J
z2f{?>S6?Zm2Y`B>OQ&Y$RArKrUKfWc-$%7skG5Yxj2avf#C=2Qj<DbZ39ophc6~I}
zsK3c241O?zz?hIp&J#9n$vd?_R=sQ63pK*K&)HxgpDfhuqHNFPb;n-5_r=kE@(E2W
zI3n}S^)b8*raU~<>5pD?9Erz(SScD#OGJ7r;CI3M^!C~+7Lmr>zPwE+mQlCLvJ|X3
z%n~E*;E%Zk$-KrkI`<mc=D5h*Spt6s_mls|Ktddw0~3yHpOc?S-)+W_BG;MS97ooo
zkNidCQUpkpVMJ`Ek6T(>sUWoCA!rTmhbdlK-Rp0F)X0S?<NH5^gLO*8`lCgOYD38B
z>KM{gbMdBi)|Q3ax{2}}&A&j`zf1Mx2t)xw4!^6mxLUFPya{FqMOeO%y$q^hBf97U
zI6d|kymw9Rtxbj!_JFZ5$2Tz7w6#2%#eNHco0!MglXOpoHoWYRJf6dal6HQ1wRo==
zdk!)QM3S+uWkDAUjb@2WWk9a_karlcr4{i5H1e>}cOOT8cu`4tbsB5<{)+tYn?D@K
z0{dC`FC}0Kv*rvibOCDzKMbJI<N``z%c+9=o-P9$?=W=X>BDqkx2NJj@_FO3J(cYM
zB<inXlw{Jbg_^r?Zs#g}L*<@$|2|PwBvCJvp^T?Dz{aK56j}CpANz|S@TZof*TP^N
z56>Oq{bwdUEgAkkaJ^l9p@8{d{A;xuR^jp+zCnL2vSVVZnce+_$p%W%+(|dJ3&g>~
zEqMtsB*E5rUQX}m0|KJoF=>s?5^lOR(F-rTBulVgDGPgVk=#AKv2lS;%@`HGHpU)i
zqTTF;5Yw+Ow6$Lq4cX1*`8db}WH1KpM4rQv)4=kJq%xg4>|F6)kz;A-Mu~RWSa-vj
zuWN&XQB2-EUgt>~Ihrc7Bmtj~2J0DqzlBGHQENm8aMl%TSD&o!`JHzFf*c4pqF<*$
zjCYJ0y-qup%ta{ldCWt`xKMA`mmL&JS=6fN7OA5=-b*_cUatrt7ji}mYVk=3`1NLW
zx8IOva9LTZvNr9N-Nn%YLvP!r^@_3Yf97Bvx{r^*Eir7>QmX?KeS2THUkGHGsbg5M
z#7vw6{;$#gu$xzEsrOJwSD|VMIayW118}Mz`O1dsPcj$i876W<?f-qbK@4)xg~9%#
zx>Q^6fTsDM6zyNXwQTpe&UPMgB~ZLZ^l9{G8T<0Yr>2K{ZnMo6EQ7vp(0jnlArG>m
zNiQf!EXV|#(7ulu;s)5h{pt#_JaQXrcZ7lNeD7=f9YJ`v4R4=4Mq@+yA@LnkP%TGQ
z@*dS(Hk2NOjKv-PsPGcMma&(@ZJ(+u8On@E96myw4<yk9?34JZ+_qwt{R9w$TnFMD
zn;7Oda&NrN@Yu9<c63nM0HTm8)+W%tF9aRz?z>aNe#MylyjF>)S0Bu21lPl!9m9fZ
zj$3@7x%RD1Y$;E%_7rdY=p|OQ3KZyY=nQcP+4ExFqN+7Ye9^0Y$h}hYC0<1je#0FE
z*7s$9Dof1p#6=wqhzu^`AC_gem~;x1l1WR&)#jrSWEn#70aEaN2dp9)Y9U!=21yX9
zTz6UF<9J%9p^G98)2Uv_rTOSnGtNLQeBBL9(v`{?7?5?vKa%+<4w7Qj&%;#9Q^4At
zxdMf7%r#Cz?rXy7pC^|LrCv?}1s$yd#D~f%TIlvl3yVx~b1f(o<r!JHA~APKxx{yR
zE3Y1AQ#A=*e&0b?AGN)g5Eg(!?H+85KVsAYRH&7{3r6Uv*p)W>b)_e6L0;U!HqMn~
zE^FJJupmccI5>&>eG31(^00f^izc5NrYGJ{PY+~l=icIsUY@Cltrmjh!e;p$Eiv6;
z<~lFUCKAoq!7U;G$(Q1d?8tFDKfR*tYU8%}s-N`oT{PylEvb{!@>*A@Y97#A8UzL!
z5NG<BM%$9(-`|wi0Lh%cc>$l&Ly+t6#|gCauIpgPtLKIu@6}6Z3rbRA%qU8NM)!n?
z#_>*hPuV~sfBH&4QBnTx!$;%|{_GST_o5e9$fFrTT^h2<d@G;XC#fq2lTN#IR{U3*
zJ;=L49$T6=l<&_}hf!=7Lx#ULNT~4x@`9J(dsgXB&3Y^toL1Abpn)j2TP%QO!r@aV
z_nzJ~=JQxrESs5BvnlUAr0y{{kNu(ws9PSNrRcR30R4PsQ>R_>1D`?Nw=Ve5PEXZU
z8{6yvlLl{B4KW-&Q(nDmS!x<AvT$B`@C+^d{moed|GS8hx7c68Wf-~d<-3DHDSCBp
z8+%BrYHm$>qdafcgRwdGB%Mujk$hWEoCabNO80Ipm_p)q%+I$OJ8Es<>Ja%>b`O?=
zNdeDUV;lzi_x@h%*Uvm5km($u7vzK=5qn)TF0Z9?GGwVm)h8uee@$v$GTAL)cQsb|
zO<a>L*5iORV>z3Jku*}4W@|72t5-B0Iv-F34p)1Pqvma$4+IL?Cpb9#c2jKB0Wc3<
zS9fi{+#bnLpCRz6t&9LQo?<Mt71xvB`fwqE33CewulMv&jVj674_?rWr=0uXzQ5=7
zKl(sRjhF4bQSc+S0K$!+lrEUR23m37Su~8~BqY9kb>+{_91h|Iu}Yj<Sm{Z4y+7Mj
z{ZKDZf4urTJPz;LXX3=hb$43#<@qi!!CL>ZFK8X_K%uCO20LHkrSI7Vol-h)zAwuQ
z6c72DvF~$LI)d$QYK*aHqwWnMqTSVi=VZsy5{d65p0Kxja?f<|bNmlR0hGSnS3HX=
zjS(vsu%TP&W$+m+_9{l-sG-f#Hyn2UvMh6?G*eoc{0D~o3ouF?_Bt0Hu3YOf1~U`)
z74zNd{(fXuR+bFc)<`<>H>GrE6Z!9QF#tMQgZy`E_WXurxjTnq!IG|O#;94t*jBnk
zqOx9((p7G5uG}-)4bWEnDFAI`BOpWg?H6#10LaDMz!abu6Ku-iq?`)BT!&w|9v%R>
zD$)(s=zW|amu(e+_A1sp?=Vj`uF3Vj*>epPaxL?gw<|~?yg;j3>&3xxEpmnsl%VH%
z!m7|LD!>$TKW%$|PK$UKRADhg23TgJ`nC$59?ZEsMti1WrJgTCO&*^0))&m--HRdj
zmNx&pL<lkW<M5?r^<8Ae-M2wvhzXWI#_oR&ZkA|tQ_`8d?9Bp%I&#$1Ibw9!Q!*d<
zi_{8g`1~=Bz$&q6+_1`LpHey1#&#AL7i(a9G4j5e#hR)gIIgZ%Vz|6LPw$=-D53WO
z08@?k>b%jURCBZ)@VH2c<20fG&BJf>Da#lf{vRZQzkojQ5;_3>l}u2Ns}UhIVs?d8
zeD$6W<#}sAlQ~{t)5;ZILjWzVPSab$cZ!*%;reaYRp<{8#FsqTnj6ov-_LrZo_D2a
z+Gx6c8Q0DcN`_kO!!&JsECMLMWaYhw@I(~q`?i<Vv&8xK_1@QDgI|z*h{R{1L*sw%
zDzVR9s=rr$dEkFv-frd2Y^RJ4H+;r#aDOM#ZANRvb)&w1wcNF%YI5y1TyDK6N`<Hj
z(9%Qq&x;dE5{!p_m*fw9Rco`yE_;p<w!;pv_VBKODw=(6dzxusb!ilHS5fUN9-~GF
zqQit!oc{}q7gB&Bo9Pr;G`Rq}Q90#9CTPvAcWHZFW?&P>s@|Xw{?kx$b#JOf?pNIi
zuZ6Qc(1KVUt7iDpKvTIIiMq{f+)r^%WI_~^cWpmahs*AFO^Hp-85T$5wXjIOZPFH9
zK3@YzKjg3Dig120S`8pKCEX*Skebj;*v$+BO<l&5Q**ky>*Ud)>lQ$q+XubAo141D
zn(n(ItgM9phW0^X_NNx=4L@bDk-(d+X3PHp`Q}^wd~T_L5~{&<z|p^i1qOqioS#QJ
zEcseM#Q|*&M*<1T-&$OU3bV}^`yESiS({P{H6aurtm7OL`%~KB2wBdYRC~$D3r78`
z=N<SA0U)jf01X`qaso_GMPWhTGfp35^yzLUkd)21KEBfq*S^POd-E4sw+h_~YSV^{
zR`1scUl^|1@tO6*7(+@7kg#iQc0NC%dCVD(`&_nIL*NcH*6IVrQ2S8;L%Id*^`>)5
zD8-R;-=!19LdgNC;)hxdy2`smTmmkTMmoznatxV4>?~8|@ag-jc9MT_0pth8;t1q(
zvTt4xM*)2aQ5I}9F#)rsx&v|Wiq}|tQKyY|IbC2*9z#56>>D6hv{hOmLcOf8nX$k2
zdNKXcGF^y2+v=S5DT!^3$JS5XD3s@&vEtGi<RyNFxJW)wLbb0%qxE%6eC2=((#g&;
zu58?^Paltm5?K-m9L~Cn_H?4t=CD|OlCXz&4^5`dri$(8zR1lk%*f|o)R>olfvL0D
zTTK<v@Vg&nNqCmkIP4lxv+3LAzG(7nb3f)pl`%)75(~{0%OBMjCeynHlz+Nt7TBUq
z*rRhCr5{V8eYl)VQ+-S(v?soGL2KJhk}^dbtg(9v<kH6x+kq5goU(G|WMD$`;%C+7
zQ1Rpl&_SpoIR5b-Z1!U|ycqr?n@T!H1oq*^Y?zFi$p*kTzw1>i)l!a4NPJIa@K6#@
z)w+b;&ABip)vHK?6F_4d5*7M@?l|&XQojUWVW_igNg&tiyg=r{Fh^!1_9g+Rw4Kqg
zI9j&YO12c*iA-*Jy<;abf^}*eSFl2_n$F~%z}ZP#SxbUn>0%{oGJJ3UeQsj^DATi5
zl67fjIR;yKUAB#5f)wbMKWHEu&S11<hr2U&C8vEx`%{>U|J89uTFUFq1bZAUhC6Q#
z&K-P~2#(uvWULGV;&^2cgAL8!445aX<CH^sY@{T3I`S!3yW~yGCk)Vv<I1ey-xdBZ
ztopi~@?<nHXlK4toVYH}yhOM8U0R3Z)sX>B%lu6uW%2xeyx{wD3gc@PT$=u-uwSC%
zxGA8n<lfh&F;<|<r2Dw0yl)m5@|0O+p!>!z=q!#B6x6|m5Q2j>9$@aCLZ)dHn)Wh5
zE0ZRD0(8P7i-p;+;bTruL<jhnvZQ#m=EZiM^~Ce^CU+Ipf$4I+nCzfWh{Og<b$hJo
zCFDUcL7VUVu2wO1hYKwule{gLm9ne#&Y$?p-$059R+m=PjNj!x4X5Spf1UL*aOf9>
zzt5ZO7WV>3OIQPwqr@ZUE%u0awv$DD+UWjZJCK}@{^I&i2JNyR{-41B^DgHjS8iWu
zv1@T%1hyVJSxBq@vHG^$B%^F60kNcmQRQxzmtLHcLWc0b(Y1k)(te%MH$Q3j`@h(;
zF$2~w>h<5z?e1cuJYI_s9lrI|nPi6hT9C|F88?s&UM99^IjwE=L_aXK4vX8Gak`8f
zWk<ODl&|F}r0*60mvp|f;bWzQZAq+JlpicN61O%JgCe&^;^f0Es<->?ouFHvQZ?Vo
z&}~Yn@Y=Gf$x@3DOB!Wz95z2!pBDd%HUF!S(GTv(x4~yS3RQjMcFzyZ@*9>aD_-7&
zTpTSCZ}fhO)NmO1LQ`omN|spP_k0+j)kxd?Wy%D4VE+XjE5tuB8AT)R%yuI^`wVcz
zR^i=&?jqp-Xq03230Bx_Gam-G9gImXs<n7l@+OCRdmMNApQt2;uOWtFh)*BrQl0kq
zoa3+U3rV~RA^wn3Yg2gce!QF;&OrSmNrF&r?TF0k?u*hiwQce36#mJ<arp0nFKd8|
zPA;tOygYr6O#0))0G$Ze`=5l>uBHuDdo7biY8$}JN)%$;kSQrPo&KG*LY<-<#XFja
z#s0NiP7}@D@%+j6*lhAp+I&EE3k+ryAJm|uT01$|<%lBYARF@2jV{jA4y->Gc|d`B
z-UA)5K1>I28fQT~?wCsRr5z9g@<?%$Pku;F!B=Ag3)O1n;t!YnPac_3_Z;wdv5e59
zWJ~yZvz&Tc3r?S(DF&5~l_)DW&^T2$kLWDq<{5F#h41=7=Bh2O{eHY<d{ZG)yHF{Q
zTw`^W9<lt^*N8!N(#zvYV>`S-GUsW5Vj{QH{$d~!TqjJ+3|b-bSpM>^JNfo3zn!tG
za4ke81h7NvtXLC|74p7(y*-3`b?YS%vwN*L!&hY!UWG5*gE?S$tgK81?%7c_tB^XM
z!lI|oU|Lk*4L7!tuw<3{%3L&Z>$X(RzgC=#d~w~4-1Y@VBIt<X>vgI5gTuQeZ@YH3
z4q4Whem7omeREKghJ8aMPa&CI?)KK^)@TQSQ&QBi8a@%d{DAlb)pADnJcFv8d$=AV
zsF*BFC>e=gSrIy<T_%A&I{(JVsLiV==j7fJ=Ux5bucwCsb?U0W(@y2JukjbwTx935
zGUfkWEY&2)YPWX9>vWVf?~(YIUt)MQ%`{z>i1U+><6ZYdNmcae*DsZe!v&t{`FA}J
zx#>R&a?fo{ohCmvv96$6r}uTIVqc?lR~nt1MgH{+{KGM{6Mg8n?aUXYPA?U4(B|r#
znyR|PXKMEoJ`c~CmKcUFZ*DV~@&<iAaFbj10i8XoASN7{@n%)Z;LT6Ltkfyx583em
z3M7rO8297(Pp+UZD8#{tP)`3|rNepPA#H+aW%@kz-Y6+=*YMYi5jCp!6Xf_dI{wUd
z(9|+3lX8<H05F81<TY-9>Wv2h{E#nxs@f8tvH49or^|Fhfhk@kDB|NrbkN4nt|8DS
zh?~swNy?x*r<2T}oypo<Y$qC-tkF(nnUB;$TA=4vGuemx3ZJ@c>h0%35PVxE`964y
zoUpzfNa-oMq`uD%yUo*yS@1GUBMhx}DLQs(R37Sdn8nPii@3bFyKza7QiZ;j%t0Sq
ze6w*lDS~k4vP$l1zcMv9ahS2)ubfuobGTPau!Uc!^hwFU>o%sC)Egp|=t5Gv(NDHq
zQ~u_5HFcGtbL6*C^p<Xnrt1MEJ#mHGU1tz70U0YIF>JOF-Q#fjN$EnJb1}Wp)xeW!
z!9x@m?-x2oW;UOyw-CHrN{%y6y!v;r=$|m6{^!2qsfRh+=D_gv&bk4f!2XxrL*UR1
z?<4om0yIdm&{tOn!02NDXe3;>TUiN?Rs~#Ve}eje>iVI<O7!n2Bg6orY?G@*(}hZ1
z+V7o&f}8~u*xA0b9+>~|eJVm9-2~o=0F+YR)lmOSy$D-I2}Ja3s-_<}siqCy9&0@P
zj3KvGh6b*C<EoO!YBWT539O=G)c#;TY#KS7UD%JZUz7+Zdz%bDlek|3j-jBjE{U2v
z2O}0sfRT^<K@2fge%y|hH4l3d6?(8{p>9ltSd6mzGr5}J_hzJv#b~x=Wex$}ZvdYx
zCWcq$_KmGnw__iN^FBBd0>a4brsXn%k>}vJEae={;E(9iMOFErQ?r_G$FU^xRNACR
zI;=B{mL-?(H9axNP&_<5zHRx>lr8OSnqa<ugX)xqU8f5Hk8=h36`E6-R~U*Endlb8
z_h2S8^vsYWV~}X%NT#lTwOqw#Fm14`aUe>oT@k<;-P?<aC1953M!l`E7qI<diNpAN
zbu%L9w2{VLyy_}$Zjqc9yG!@l`V=^F%L~n=gN2K!kI$wEM^Lm$pH)*ER2#IhUv@GZ
zNs70Wbk|enYFev?f1L9WAqU)A7{||3lh@wGIsh3R+QHD+Ii-#0LO&ZN5V$|6?>8GN
zFyg;vd)rqxf46{Bm+*Sa5$?J6;|cE6ICWL@Z`9sXK+}zEc;BM=Q3#c?8lAkgqt1D1
zkbrZRm`*K!6_#+T{q10389KOK;?rh}kAQ#flIN}izDd1Pgl(P6E}*n@-Fd>-grbu0
z0$wSTmsHmOb*tdjiCVeKzC2to1DNZv{A!KP=C<=SyTgFB?J)w-M!A0M;`tCx0^vl6
zopE=SUK`K34x)D1O+=)P&WA`RXg|Y=?Z1X2Ic1Ul!uU7RU>Fq{0Gts(A>3i#^E_|R
zMDC?lB*Nk4Hh!8QR*RvUWObFF)Xc2&R|Z&JH`ivNZ0!2?AuCxPD~14@NjCma+G54W
zPy*Eu^oF;nNq|#ELaCNBBUYox&=+}qtV_C3tCij-#Ix((*W<9^j@~mYh8)J{h*=zS
z??oA|hh0rz|M@){c0%biek}-MBtbPn8(MNXBzsXu;hzUIVE_qE)GsWv3a?FTHy@3p
zFhTk^C#!dSs2Vt&Sa+Y&S$i>CN5gXS3~^MmwLB`oYWWK0POW(aX+r173hA6LGUK5V
z;hy8>GWbQAfH8OHs;J*=t4n=xxo%{^$<7>Z_=-FUL(BI*nx=CW#7d>Zrb551;Y9pE
zN`(FlV1mJxq=WSIrKQ~~+7=_pmLC9eNa33Z%1J>_pSM|;;B(k7DBW?0^899NUBR=c
z-q1H&mI?NbmX0BxYxRREqfXw*#ZJZoHF+JM(X&|0fu%Sq)gKaUDa!W^d8Mauy*U7k
zsv0d19#wS;Ng2XI=(or2nyYXr5_gIi!O(XjXENnQHJZAa*pQCF;&(>VDrbcjqf`}Z
z)DXb$-lMYG*NkYpwS3!Fv>{`KfQ3({Zd_=yXJkEJ5CuSAv{hQM*l5Y+bv)wA2SL$D
z1)e{7jvF>kmeaLz=@>q82lhR~F_Vtz8mAf3FrpDlr45mPw9{8I_jSt4+|FqF{B1dy
z$9`GE>U*$yaJwV{16`RSM9BY?B(7Jv_pT9DW5o+{82)ZJwYMK^_MxQYE(Aj0zc_U7
z{rwtz7<nn_w{eU3hlY>Y{@%E=IJ06ulhhiV`j>G=GG(_y49=jEMo$C#3#)+(l+q)V
z()n9YVXgb50E^im(<J&2i}{3H?60(d?atIl?zY!Jh1Y>?LrCE<HbtO4i;dU5h1X@@
zMtA6?i66caoRt;nrt>5!{U1*7u#b%2Zz$#j<P<38Q+~<sVV<f3ECWv#CsVt2`=K_V
zI~ws#IinELKYc2nSVrM#i3uD;>GY8m$G05Ma8tbfnbu|x%>_Ku2EW$NoU&?(%jCfE
zVzn8z@w26Yt*46eKPi6<psoCX!Y>trDhd1Wq-H~p$3fhG6v3JV^tl+AS2*1+;Bl;j
z0{tG7dXh{iJRxVyfm5WOKYpTDX-o4V^hpUf8X_k~1C7%6Xt3uAW@lH3S=TYK3T(pp
zBoU{iNZsZ>&Qjek=x(YOR6&R*2#CS*tCGO7Y?j4cO@8-x;jApSlRAt>*4LDAq9i-H
zIki2#FtZSJ{J3Ag=5D|ZaB^4M8C{6_V-4&|7QM^wd$6<Uyw$0>uVENAGew(@@V9m`
z&YvjL?ce(G0c=URW3P?g7xBz`M1Bb3pM?$hHQ|vKEUbQu5_f`diz%MY@?{>Qp67mb
z2SdLcs(qWCeR;huUPB>7S-nh>&ph66Mmqa5gQ|cGb*=}#vzEKFw4|Oo)bUy5cYnt4
z2O~cNh8*RfbopnQe7FxPEIx2~H8=@Z7*!qeV`?><KEHARrp>GuOu=B3P}A#t2MfAI
zA>$T1*K8M&%}$l-vRImTiw!O*Qqx;~U|!<VjWaISri}u|3O`m}m#HDvIAy=i+9<J0
zt<TPkE(h*U>6Alc`9+`N6}}#~gJzmViQhfU`*_(Gs2532ML)JA&}%!>`KKOe=hW{y
z{dp#!IA-E~Z^G8*J#4?t&AgYzDDBd3H=H{s_r9-e1BZuwMoD_gR*hp-@4I~|6VIA$
z4F`}qf&-z)p%|ng5@e*p>?nLL+s+NJwrJmTSooE44f!%Sd74A}2;<r(Vs!ff>Rf}K
ztr!@gX9|W9Miy9JSy39Asb*`?9J=`2Z}ddsUqfb$*no}!qfr|P2Dz|C@>77T6U}Lb
zb_LYQ4{DgoP$nDNzjZRNR93zPRa&d7RPCDro%fvZVmFd2z?K~0Kbw`ts0dikJ_bhv
zo$5rQ%YAcshDyyQ*Ih%@3P7jVCH;ySC6h}reC<MVTn!zsbKJd)C5)J}7-Q70H5MSc
z%~&r$P50Z-P>CYuv0I0As3FEc&l%I?E6!+J>#y(mqO0csz^OnWfPlIUj{+#ERF0mH
z*!|;mo2R|drrlDN;+tsbQ{$a8zlsN_8k_O^ZCwz0UR`z0Rc~w5lmH$1&$4Ryh?r{$
zDC!aclg%*~rF7WTtSB{k-}TNJLeS2t_3BA@-7g5B7YE<=DrThBeb2XlO3&^%?U_y6
z>T1eO0VRI7rxLZ${$l19UAAFzm3h3eVz;>3MT3L~3?q!7taJ^30ThCY{!HSMWNymB
zVp|eixTEg^U-o^B_Topui&Tx8Z>nZu8qX!@w3&?p%dd<{_&o{uBp2D~2>{Vt1e@Q8
zeqkS5@r_%7u=n<2ql>iU$GiQ}_DQ-g^9}G0zgTrkscI}~7ptL>O#!_$?BREI8L$Fh
zh~jxdp6<Qh9re#t4sxJJj1lgh%wH>7VPQ~a*jn&WA3I-FF4cV{TfTlo#!>;5c<JWz
zPTIzg-jYO>OwDT?pn)a|C{=M+_^JzLZAQJ)_YcNo!^ah9Xx*f@iG)J+$}5TCvyaRD
zkO04ev7Yqpm;Z6!1Cv0krhBaV%}$yI|Cl@Yci>F$Z$SC6elY3Pp}kVh_}C6#n1MI+
z4&ffEqHRCN(MA57=(afhmlgs1g+}&iW*JQ8)NhbynZ^CzjWGX|FGz?CUrR++Rm3&e
z=_s?QR3PW{<xgTb!~(BDA#s5b1mCo;mgzP30Pda1=S)hX6vNrcyt(?Fv1a7OMKm#H
zx{>FFIoPv0kzCYE#`%M|SXh8-Ni4tnPhJ`o5?pTV;7FXRBw`DIo8wM`g|i;tpcoU(
z`FsCW?C@SsKiT)-5cUwZjTf=s^LBzgGBxiM@*Nao*@M;XWH0DtPZ?;+dK{;H@l6!_
zm`USNWLWIO%Q{Aj0vCAc%DNrvwkEzRN_mht#1Syp`Emi{4sFk;Ap1P5-4-aY<!T-|
zr__?3NiWrksOtAa@95Qv!YoRIJsuF~TrIb63W^}Jr!QXglD9jkZ1>d?D6O;^3YOsO
z$+w%T_wV7T2!{V>S55*ckSkT&_`a@o@A=j5qa?6B-FN|vf{5#65>5}$@b0xOGB*A^
z*e&A!CG4Vt)+xF8XsMBt&0r=i>0c*%K3Lw-KK^6y_yF7*iv$3gsVHAzoX(l7z&83X
z3GsVe!l75j4!E(e3a%rwy1Y9mn5I@P=GI;t<}^t9dK@oz86S6E%SWQkJo1GCUERiB
zSm`ZDb<WzKvIuDyBHvvC)O4EN3_->=oX_#-!#?@jb_BGA6beyEhw6ksez|(IBtXyp
z#R_e!iRs>>w<Z~&r4<)TtEJ2RuGgR!%6ac6Imtb15ymeqL9yiR*0<3YgSA5*_KBD~
z+q#zi1ac)Qr^xtj2j0ZBuklf6s7c*w<7MYylBx+kbY+{XxO2wgGd6T?BhmZn$i_}C
zeak)90Ykf7U)5Kj0v@oJxk-R+JXFB6v(TK$uT`Of@HqG}z${|yebRAyH=0`g2$&VP
zxA+_v-1RHVL?&6S*BpmVDkYI0{cvuc4CzCtt^{UH@BYu-YJypRijfSy&@cIN#e4~k
zY$kRUgjip-+^hx5yRN)$tISt`fw<J^X6?&c8bwTc^&&Gz25VbecxG*wZ2F<xq?$aE
zv#*Ol0NiCV1O6?D^lc|su_=Vb_O^H~GtaJdl+dS^m<;R}Ap(DtUXDI}F!m8RW1GY4
z;$)U(PNM<+i-T$TXF<|2FC_EnaY^Tod%@h+Bvw}EW2qRjubLcTVo_v#y|WqlaG9o^
zj_kne4W6K^Qauce8Y_6Zdl?lRi!%MVqw~#UfpCy+?$<mK?vrwk{RFrB?NNh(t4HJ|
zTqA8l%tlRW5Tl5TW}Eu522asF*6(qW5qP4%P<(%n?W|q#->u2{u_~k7l26>f-0u1r
zTJ!29;>9bd`oRju_{xrhzJCKu!UO#NbWif*2+H_Q?mx`Hdhm9pM29f2Xvg^ffi6m=
z|JvFTsF7`sYN9N?x%eQs8XqMb3*%{xSo<qP)H#0yBxuDM|0HMwF<%8cAf4v>K`q8_
zSVz@9Qiy%mTB#NJ@iE-AyM3Elaqkg%8n2()5=NL~es^Ebupip$a(Xj;dR{)k$V=Hc
zT2<^P+NJ)0A#`5{?A_Q0Y`&>dJ5_jzHdYdsFCMKwSQ0W*o%ssz+zvZ6d5OdQrHS)3
z^&<ePVP7>lSOJ)YOpkrlJkpzLsUZH)NkT2thDUiy>4P7D5%-K*IZ4=3i!Dy5_ky<@
z03~H-VpKifVXB_x7Ac6ek0R$~c6y$(VlsaMcbm;jnd#?5QIzp!2ZtKVDTSqYz8@ZQ
z<G6B|7WBTz7orneij#$aO{9na)!wH3&y-W70s^*#CZBtsQ^g=JBH?e!>75<0<p&jl
zwcO*-sbcL<&amSM<IzzH!7t-2<}MdIOD~RF6Llb?M=eR7lOeuK%{EpsfPJN1^CkTE
zsTn{Y3r=9rD$*#*Hp%@=DgN}@lkmWBOC^75IE`+4hNabauF64s*&pl=LyMzT=uIQv
zrav%0SZ-S`G?4C=99almhFu;DWQDJtZH`8+#&-6{QZ+xZ_Yz2Sx46TMHtmV*3(F)q
z9f~?3Ogd}TZ79B@)1;4dzqossvJG%IvmMG|^thELq1VY=M>g7eQ0opy0YM571C_^o
zRhKpvRub+}KzY<%_G!Y{QSL<~ig-;gTegXmlNN>W;Fr1V`{kDR;?Kf84rlWZL?uQ+
zXth>X=y%79b1o;R?5Kr9@eD!PfiRg&;dE;-XdSme5xwuiibF%-smY;no?HS*4nBX<
zjB?pB>9#u_Ir|>%IWB#cfydGJJjisZHuz(%#Iq$Yb2GTxTH0olsYwj?_2xk`$H|_>
zDnH>u1%*F=UiUEu*iVrxdAqB}N*gr7PH%`a!A)P_=>~9vTL0YpaX+hetz%VvPk#w`
zj{jKujuD<9CX!#WpaV&CI$zG^ta(JDc+zoRJuICY8PnrI`01fj{SRR*?8#q01xJ_S
zd<0TtN}Ga%%3s(ef2Gs^y<e}o0aEjGybmEy%qDt=i;J&Tl%uEZ^m3;>m>zO4OHKdV
zOS^}PHWH6{x-9jk%M}N%1GF-RmIOT5WV#*wVfk&gvkeFmuDDB=IeTdm_gI|?vPrMj
z@vy$@CIY&wtWarXL7!_G$v&Hf>=z-(A*;VS#VyC=y0w-xn}ns!Dzxjv5){Om&oUFF
zf1NzalVchy_$di$kk!beiS$!7iYCO509v+`jIQ6Y9i!4&r6fzLnL!m;s@XVtZpU{v
zuLSu$v|ad3#;;}0jZxbHsW=1M;+CS`<eEm2+7Fx0_}0_KfsfB71tUPaoHXN0VOdzu
zmFW^nG}K}b#|sp54VYq&#UxbfeDa}Xz&t?$&TgE9Uw%{qb;l~}c<M{Jy-^w<e)Y;a
zChv3XxW-AN+`cwIP}QHtS50e=b$hzAteLO$=`g3B#iAG5bdMz(179^6<WaaB`_}$y
zZ?ZH9px;ZDK!*~UscbKJWj?ae0WKSC*1T2Q6+V7heQ!ppd+5fjXGaDc#<lb19ln%P
z6D@|VVtzMlq<!OKE5x(jXZV-jr=&R04mXsRQ3<Pq^-I<rc+QM0?aVZ1o_VivU_I9=
zdoyQ349F0=0dT*<8UQUaB&<o;H9)p<dGg|wy^Dkk5l38Jeu-_U7P9kNFISXP__54~
z65XD&Wc<JRUDOW_i@iU%Ldd0yysIA^bQ#X7BY5Vd>qy%HtuM3RD2BK6Ka#zFbIT<w
zthUzt9?4STr}*=4Z2s?zEvyek2US0>1TbrmNOdpOd|Oi}Aq2{LQ|-)!_Y+PCY_hck
zy9*Q{aRBmr=6ajyPYE=}<uG|%pp0DT-eONIB(+CSZ1YH43xDrJA&%RZq%<-mKqrJ?
zXCptyQ5Rp@lou->F!Uv?E*O<68mGbH1wGoG_~tSA9KKwQ@7D&K*N+Iub&BhKkbmXL
zn&H*xYSSXtH-bmHVygWoyRDPiE!0#q<i0|w)fT3IaEC`v=D4|-%@6k*2>I@DhTJ@{
z7GuP!4+~IQc?D$gM&1f#jXpQeeV(GJ`QBVqpUD|l+aH$86~5%>#xWoK>i&FxMu-l`
z0LP7NZa*Kl-4&O~AtX~-poiuR0N$)JdxJD!md<e+c}hBK#Z5cRuAY$cl-}d-1kBvm
za^r4tZ+1|(Svo8#>an}|)VsI)J`p-l1~3EUuqMbs2mpr983=TvN0ae|{EE$`G<~aX
zwU~$H1=$UMJMcx9(}Br0ZZZdCJCy#ae%M~v`%=MatowAOWMe%s=<aE@mR-21CG!|=
z-9Q3cqZDWoa?ByZb*C8QKryyX&19#GOJ`q*EmZxlPUk-zuj{CeR^%&8UMp4r82jEd
z_m8MEv-xLRk2t_(jSP+Z!Th)U3vfYstbEqD_laS0AbH%Pmm<dG%jzhOWg?sy8ygGx
z|4`~3d#=I{(vdws&5-xCyJOObhUU~p7Pnp%f(i}vKBKUKJ}PuSr(}<kpf;U(Ze?dZ
z$!Hx@T&e{z!e*)+fb(r{wKZ}y@=$+CJ-!u$e}gw1doQ~PZ+o<Dw;bOuukg7*Iqztd
z`|=|bYw+Mkeh?ZWZtE~$0QFk+C^%HJOdnGZvefE=-)?kG8j4Qxq2IKtocv#0fcf(k
zf4G&wEA|LJG-r`-Wu&nJ^{mb`WkK-w*DNi9EeSsz4+UyM&rdr06DH_;^sC-J`Cba3
zwpx{2TBXhK2B()t_GEIY1xlG;J`*Z=8!*>J=r$N<bH-Saaam7)*87SFI}va7x<5am
zY!UQ{=MFo6vrs!GzycCPG;z_?tB5$T^PBia;U9?TwSDij5wa+!oY;Ql$>Mr6KSzOY
z5^{cdGWOb={tO>eH>PUUM4!Z`rFRInJ5`S_)qmj5r?}2gZ@kp>!Y{7C3StAhyBRAU
z=3yt3O8}gq^|d@9P~0-3=0K`Y46~tv(l_t=Ti&1luxXi<Qm@@CY%-dJ`<|}T0K-K?
z&E>x`%r4Z`2TfqU!nx6~m1kft>DjnB$Df5vD4}d5&K9T%JK)_&q{wK{Hh04~YmL4{
zNSit3yeF?T-#J)q)QPPyFKj{p<|TLGv=6*CrmBMLZk=a_#RxUKn@#_u5IQfJeOd*~
zpn7d|ii!h2aCtlW6}Q~5y>l3Bzxv~5rP5)=&s<M&xRSO`%8i5_Wmu0>awH8zWu9Yb
z+1uFg&VH5zrE6;8q;@bJwb9vXU-4uAdQuibab*f`2Iqcm3B#UyVQ(>O?Curd@~O9a
z7@O2H_8GlJxm9gOIjKr)>uIh}FpI+0?W!nAfl^vd!jd9u(VNoZNmP?U*!4!t+xL}~
z7$*xisI!&#uMSq4jpQ9tIqeXSpBcW^smKyd$DVm6o;>f!WZK(lMUmpagxKi(%x(x&
zuhHtgt^6@G*lgmdrWCJ}>;5t>bFThpf%={EZrIq=Qei%-YQA*>wVPUfX3kfsY`LlO
zFTWw%OU)<FSDm}+;k{s0mRM2e)vUm$kW;($)Z_H)<xi(($>%J$3HNfo-`doX3z3f1
zk-x$l2tt)f<M;QmCFM=utE@*qD)0^ZSuOJ@i!fSM1d$qef2Mp!!dr)N1fxo55^y7r
zE+5p~(q5u%7u(83BR5~(&PCgZIqkdNVBp$&<QBj$Liz~SbcMe){HzB+DVFz=^XzWx
zcEG>LXykm*Pnp_}uAx_J;FC$#uGXnxY-YTz1gnW}>p_S&3Af>B!q9gO(@&Y?9!#)a
zC#p8UuYB|Qkw~Cj?;AL(%!iAw^v_jnA$w~~)dS(;1cmdnwbKF)qMmuC6<CDbz_Sto
z=+^erukLzK0i!_U%?QoGd&#f-mHCE%cPg5&w%t;T<`CMqIoOI{^P@x)Dsyj*E#|I6
zZoI4Vm6pz<^kyth2A$E9w{rf^s#ErNVlAh0l12UYZiVl*jeb7M_yiead(-ngH}drc
zU^D5Qx&W}ErEm1_Zg<#ne|>+i-g?bmYs9%wN*q&g28HyiMmb~Y@z4^9_I%B5y}e*3
zxcbUHst>&fZ4JMe;NVKeCf{@UXre~Y-*j~ltt@=QJIBlO?I3T2i_?&*iwal#JEf57
z)yfb)3#P5{_ot6XM0SCb*c-z8RB5e%yqsv*>UD!n+3Av3sh-654VX0Nmx}mKm+yGR
zp8{Ze(iAn`znvLDFU!?8)+@{!y9Ei8A!CqzSov+`Rj^RCIrw9c9v?`g+Kn0t-MnLK
zf4N05*xK1%tjj0aaBZWQ*p|HE+P%_gy05XcieLP|Uv$cJp<zqO{BvwioB#t*ZKxKh
zNw)E9cym<u63>+Q*w@xT9xrrlJMIy=ed7NNb6fq{Gl62nC9OLIZAQ)&{Uo);eEVMb
z*)Oy{0ty6d;N73}*sIyS6R>lByScAb8Y8@dzDl1_a(&Sru|2`vNB8jhv!1fhRgi03
zSBLj`czvO0z^H_frRK||J4aAeNceX{QUNy+z#%GD^DTs(z3pS=fXnwpQ)-SITG_PM
zXtWdWu0<!seT__92?XX12I{u9!J0(VXjA(u{;=TV>u}74xp46$S44|ZvBUXVv<$fN
zBKX8r**2&jG*w_=(|TG%!6?Qy+s_ZZrEKpfdPi~Oez^E)jA|)fD<bNB#_~}=uvK=R
z2Y2=kE$L;HFTi%D_F2uxHR2g{5jle1WdVsK-S};D{jurp7xUTc4cgWN|K6tDrFuKI
z?)R(_I83Hr+64V!`u5cXSY$=!UK|>JxxFvgTt9LQCNNxS-RBG#@sPzuyd2&?T4~21
zA9dor8^Y4nG_ik{L5dQq$X4B;XDH|o1mYOuu^&CWPIq|&;<^Mx0`5lvb#{j*;r)oO
zpXOt={)%#mbNKu-JfwI3!wJgA5mob0&b(JCkOfFwVCCLM^rBYdi#oHR-cokr0<}5>
zQcoi<$kmE}m&<pa=zA};pr-<+71eRp-s&7P(kzB;9c`IcLplIsw?)=~rf$lOXgzF;
zDTV4=|Ixc2K0c@Gl0Q>PBYgRjH(ecGZy`#*LJlj{cG(;>uMpxXvl+#TOq9%)tEHEn
z7O&65V~G(<px6Ae=_Zk6AswaVTJzB6O6No%{|)cRb5C6$7hK7cC>*#X=KGRM*VN5x
zlE$jEV_g8O2}T2C+ELO_FuaW+anW6Ip=)SvCJ!R43_>tK{E60=^v^PLd<@bGHa%^U
z(RDaI^Z0R6bWP!~o~YtOu-X-n4@Rr-_1P6o&mB<(BH|>%UI5LWh^;>3NE&~n5&HcH
zAo<={KTeEiISUTY2B-~%8oTlUMOHl}E-JaswNKiMv#X~)yzk7Mfz?JT%crbJy`H;3
zAvFpZLfD$HJ4nK@2-NJWY`rfnM$_Wll31-;IPp_VvBx=&QC#K};EyEAb&+Fn2VBx#
zoox-AYZ*cNgu{B0IV?_x0*xGRs{9yzP9^yOITF$fBJK`pa{r{fMmoB<yxD2D(TvuI
z^83sEZpmQ6!)#)oi6$DkuZ-aUwV8-3CdpdO2mf7zXZKL4o+P)V6ADM+xB6FSZ?*?q
zxA55fV9*t~<Js)9yR#vz#Rfk>2c4<8Z0Dm2LrvZ-m-@6<_@<Q}z;Qp3@VO*)(+Q=l
z9W8RRY6*C-&S!Yog?J}=x$RBTp}7$;*ZTTNyxyBEF;i<NAnpUgXm9=s%SQL<OJ4T_
zULZ$^CZ2KV*TJOSa5joh&?}rSH9`h%w-|GY{<{_Bq&k22J?Qdao@zL0Pv<3@)^3R$
zyUp7!sTaTRzr4xJ&nG+r*#GQt+nq?E)To)ymUlpP9o!SH{pXF#HQf-}jTEQhhOJwi
zKO4_yKR4rzBLdr{Z}UPO{4rcwA>AEyBAsYgLfi~#!YIGc5{03GwSN}{Yivi44L)l}
zYofDzjJxlH1LmixW+)Xz+2^J{;?euUWhgrt0R{=iB+?!w<nuYgws;wG1KFxqz&^tN
z_=TeNJF@+9YoOKnEBdP6TRXwgOwCExhiuAJ2)>A%U2OC+0gr(<I}XAO-3YH2wBozV
zjT$anPA0K`e;4`LNvPoNRj43I=HC~vCjV}$4`J}n68-=9lVB?FGYvw6Bc$aHR<E_m
zk8-AJF1am=NJk^Wp)WUZ%Ks2_DFQiPVOCQSAz=9p`nHB%=Xl1BPPCA<Py!^h|J}P|
z0c+OVm8Wf`L4%HZ(ErI`0wdf#_-s!>IpAG`0y*O2KUo7|kjOz7^3K{g7GPTO71w^E
zfwawza(ttBV(z@XxaRV=2znG!G}8a_)c%}!;6*7CT-Yz{lrqB;dJujGn`x^?B54+|
zh+-+%9dcC!e%^IS=sU@u(+EfjFe+r>KAxk1GoGz|J0WHJG|K7|vc{nR^$Rh9SNIoB
zOj`Y~e;sYFW8Dpl*VVKgxU>>uUv$WDhFSV#SJN@G2#?Z@-%>v0Ng{|vva21HzK&Pq
zMT@37<k~(ENzE-$Gcw+-PNWse6H>%KAOE$b{on58!wD10BahtqVY?(gp<99fvxgSl
zB)AmE-va{-*W;9hWfQ7&EeC-u6<EbBUmn(!48<pu3-`kh`}O4m%~R#eDE9&wautC^
z492$ahekCY_xGKgWuA6kuU=;ftB@DPgc!y-`>oz}+ytH@Z6l9hq>*2^1)NiDOOM2)
zk@JC_{eIu|->7OMHwdpn1NM>1Fdn{8Z0-l0c9J#RwhO4q9j4TUCRRHT`JUe$*}I&q
zigMa75-6r}OO_k9{%Ulw_&t<(ytEWRlb`fx*pW`E4^{Th;vfGY(p{{=Xoyeekg!0H
zQ4pOHj`98cRqRVe0|v8pg!OHD#NTGulnMWFtm{ty|Bv+sBKlL=rt<x0j2e6CxsI8+
z44N$}jzVO4#>%cx2nKm{qsz8+0v(zYGc&X0Y{i;DD?Tbm3~;Plr|WZnZ>yEe>Uu3I
z!RoSnqgf{U{aBuRB<FVo2gyI#@!uTz|NeQxjB@9+%}+ZFQH^kRxex>98!k(ZS-t7q
zpD9;ua5U-e>DgW4p{9OJB^4zz0X=y)|JX9bOychonWu+Xrk>_>abcp{9$`*zav_4z
z-G`)G6#%C}?nUTB_+PK5KY@e-X;;K}(lxE!N6QbcSK>_i_~v85!f|n2@hm{o87tOg
zXjQLzJMaPMW2|*yfPa2ItcsVLK(n9|8gPF&o`7JiusN(x0Vv>n&6iRNJ=>XI6->RA
z&8!k{jzcEc;!qn2<kW{7jP+TWqa4gSPCNg+L`nar;H84BR*dA04*!_XraE}lt_ovN
zDSLZ^cEeN;%!>ZXsGFmG1-KxXZI7fARw;hfuu55DEaI8NdkAS2n241a<XNHe(t>_x
zI(3Bx)zLc_iae-Rq^Y9++gp?z+5wJ%1!{M^7JPpx!9D*^WRZEqV$}Arw3LYqtmT_n
z`kqn00TbydT0d}A@*b{=Z0kVX->ZUof(q%`5+Z(UBYa=@x*!_8;`7%^!oTb1fByX3
zh$2yz9D~&7_w^IAY7O;QWJwF1NBh&I;)%=#7kili>!iE4_s9G^e^=eG8od%nTK)y~
z!k@Fl6Z==psp;+ULo3!S4D{@#V%v4ZgDH|lA)8MaM<eecL1NE?`00co0bU|PS!600
z8mUt%7)7Cz;t#u<SN~1GIpqwJK8B-E2jtPeec3e^a$K+9K+zRczk$uH2ob!n<{Ye8
zu}7s!W9{B0BjtA$QBY6_u}h@UOnb*Gzo%Fi{1Am014JBVm>c|VQ*33sQWRHbhjqO4
zr_9n$gEHeVHi9;~2oQin3WSRIW(gC}iumRUs{yfSQhG0!5;2XO7wq`okGU`K5C4io
zu@8YseOr1oK(*m~#_oNS!Nlc>^W|{GigjeEY7btKX^F;?8Dwv2XS!4`#4hze4~m^P
zFibOBr-fOq4iuwFb{+FV3g+|VLN6aCn_78!<}#(c+~3EF(2EQlU=VO<dBQ}Xe~XKM
zj_SOg8R2gX{I6j6=S$5qN8fG;){L}1d2*w0#B$wpV)xQSZQbq4-oelR4|{JJRpr`t
z4J#r@NK1E%bW5s~AV^6`H%OyMgGh<Q0+EhINH<7#H%NC&cQ?G}(yjO2`+4?z?;qcf
z@5eh9gMmZ1;>_cm$DH%PAH}7Q)U6%TDl|aa*xWo^vobXFE-T~xaJ}bn4JNwtwQC*W
z&$)i1%wZD_q3l>eb=3Z?=)M!vkYm@W_96P+Q|%#iv~~|-d_-cze;p9WA@=8>O_tf+
zU3fng1}yZ5tG^W(Mxb6!o$$=Kc#`nHT`5iknN`&ib2T4J&|2|->OywbU%`p_<CMZM
zw&8X}HBJrLOu<cTm7<Fn0a$Fbmd3w!^N%3~0w#rjF`#}9u}4m7=+fxWfY%YvJGb4<
zlyW`Eq2Ew1EVvDpN*laU7C?<Pr_IqSM~?BYrlE-w#*$%a1U;XeLM<p@0Su}Xxse1X
z9OcF5aajT`Y?_bZT?sk5#Xt9zpP|+N?K}BC)~KTiKdu;y+zrvwyhr7q7CAVB0Rvoj
zG+)=rBEO)Z9w2vt<^VBRIZ{(tH1#R^GHsy#V{^TQ^}+SzhM)%D{BCG#;SovA*8AKX
zXM?)%{Eq)V$$#z<!R1I7`zsWaspzetX|mLJ&U?_TH@0Hu5(o3N$$ZgB`psHr>pd}X
zal_}pi{DN7^NT;NV*&c*WAk7xmNOf+E^GAmcR4y?e}C~hPB>7AqrNz6B;RUVh*-!4
zUL{u}#NB^kAPJiE<a94xJ&&1?%hC(vkST#QEA?Ji5j&~}nVwr+8)h!;Na-UFIC||w
z0jUV)z?a2lNdRU5UJD8BBgTW+`eaS>tz>SZY0PoK!tQR>Mkcm63p57LS)1~9XBHDl
z^FW2I2B1f`26P#axRIMqdR3g6u8_jMF+L73s?4_dsVdYP(Jyy86Sa3Hyiv+tp9e9j
zPc6@r1`=|KhGi=07#bcj$Fvmwy#vXI?s(<J`d3TXckVS?>U4T0DQ;?+D(!4el*(tT
zvU!$HRNAu;a=rSjtTg8;P5y`@7Ae>P`9~!9Xbpqc+MI0iQi8#|@}WcG%Vxs1Pc$g=
z3k#>(wmQVt#D9ybRRhqfTu!mhk2hN{@~w#lZ3n;jW`onM{-BA<i*(h;(eo#<1_Zr+
zZ0q`6?2W4k&$QO7RwhbWh!Gz?Rg6t(Zr)$+EwEVUhDY>HP%TKh6HkJwuC8AF_LZ&q
z<d=}XeT9x?05si+e2B81G(y9#)vQXg$&GxkHapVXy~^#9u;Bt13+H_@D9c!4Zu6eC
z0}cNhnV|ErC@9y?B#w7_<-1+0z$fGwXV~StA#p^te6t63Y`tMu$8I;<bgF;n!+~PF
z-Nlq=UxIiG&qo=k5J`M|yi>2Wm_KfFcMQ+sYeL9@2#A#r2Jo_xU@c_82Uo3GIeTWI
zgy7Xu9HalVcP=XECXw`C_fDNtc+rwHznai5^{tbQ<Ku+-ldTCV+wpBt9BJhd`xkGX
zk{ey3TN7;U4RdA5CBnhLIKW}7wo_ZpxFI71G@mUS<rD}UHBnAlQsZwI_wEI}wQFn-
zr7LF6Q_f+?-d}7iGVTRX>CDc_uKwa{T~!ohG|{*UyMXCEz%WzAC-8nq@%n2nS8-6z
z%hCq;ncaTkR?i<X?~j1;=ZNX2x>;)4XmdWzq{da5x;eMF5ks0~`Zf+MB%V3yrE8iz
zMqT)z+->8=9$*$=B@i&%off|NaWJ|`^$RCBHI1Ki#9R?XzrA;loaN_#{^xr>JphYh
zy&)sKOp1Z5Wq#5Mdd|(jCi1?zmQdOC4WROE04mSJ!d6Eo$GX0DOrTLIr5B1SV2h^B
zWvMxV<1p@iwA=fUNQ`3s$wh~T0Vei@PiAR94ql{QU58143|D!SQ+)ZR`_Al5fvIYo
z{mjwAHy|^&GF{D|y|4D58A)U^99Toq8F1&)n%ilB>VI9x$d6XpQw}X3utT#i-0lgq
zIMGA>3TrpNk;=qyiAAv+e%=!s(>wnzg6LikblSydRf{}KY*X%p$oH5HEC(EfeoT}<
zo&UZtcnrm;kQ=qWbVOel7*-{~x-|iJTk$a7RH@7%b)wc?SUFqOV`aqzL?s>SP+l9j
zd9`4PGfHsKJX=E__W5zVz3t&Y!mRVonI$C$Z3<yw9U+H9Y+~%`e_Ha#akNWhwy=wt
zi;zy1+9+bJznDY3GEf}IdIR4(>@Nj!j!DM}+y*Xj^3|dxB|g_(t0vbL=UunHJyB;U
zIVNV44kLJOV2^Sa>Z_0<6*c%$cMv)Do)<{iFf5N`tKBME?s?9hD$O7e7yG4M!oVkn
z!>S~t%i51Pl&^QvDCcE4vc5PZEaD#1*?l-T<3z2?3h~mahZmi?Wff%%`a{h_w{w%l
zDb-8NUzXN59Wunm`OKHjxP@L%r`u3T#|p?3>V@)&LVF{HuxIj4v>tZbe_}yE-RQM&
z=-SkXm;LJm3VsEjJ5_<)8<qTd?~kqcMdi#nvjfnJjh4Trd-CMTw7rsp`z=`Pulh3{
z`10`AE?l*(qW!`HvQOP~3xhvnzxI0_?=0`%g6e#D5{I+zNG^2JFHjj9y14be-{Iv&
z6%C4gDC2Z>D#kW88Qv><xHi(b@sQ$VWdR~nF^$EW8b+#7>2yydHa10G5<hFHC&m+w
zn>(Yyv6{0nFxWMdz0o`AO@;!kFuJho5vuNf0|xH{ImdW1Aw6F4^1vrS8)Guebk!}s
zv`|{Tm_f(aI%fLGBg~4u9BFZoYgc%lQnt}5(;@>i<Po0TUDd-&sQ{7%=DZB~WP~N3
zp87N~|5weug@&DTt6TSbG1X?5x}m<vi=8op+Q-*5=*I<HqrmE=sz$fb2P~uk5TUbg
zU)Bd+5$_7c)^-zE1>m)yEO__2{tk)q8A&1u`ACi3glV8^uYLv=oY!d4Jk&S2Zbr}H
zqx{^aO@D_F&S#LjPuk~Gn9Pb>hxpG!Wp`aZru`Xc30}3+IX~KDKOX^<u*H1{$YVJ;
zlu;Sy#?nF{Cp^2VWvs~9Ei!WPiC_dJNkB-W;<qc`W@uKwz2}K|x87>mT?CYDW92bN
zGe;i%g1mgVuE*-KVjV`1t9QE#9bM#p2uT;}lmo{3jV;Tq(F(g5y<B9-9f@+5j%cNx
z%B2(Lj)Ct3Ure!0rN0&w_nAdxEJrmeaiGDE+6iZi#Eln+27034G1v9!VgRs>`Fz3m
zE?c@<QL63KK3*@IO#?NfT9v5Vbe_AgbR;WjdA4oTq&B7cmxV_Xj@7Iqxf+LpphAVw
z-r<t91KwrZ`K}5yfa+>30%UQ153WZ`eI6v|rIdPhNvBLd0Uv1U%;S9{>IBJmM!YRB
z{UTEEqiD?Eu@UxA#N#-1SKqC-)uFB#UCp*jYK;4Hk?5_&+r-<y`zirV^*x%v={8bT
zwM)LyA>rwUV)08QuO6uHeDlYc%Zu0B(}JjE+;G4M^ONCNadBg|+Re5IVW>QI7pVAb
zz4u+xd@>(XB8a%TylDA*vEGB8HLlEL3w8d%9Bn*D+hxf@p=hIIHvK8qemm-UOb9T2
zB&S_M+?XjPqYHut8@j6jC^t(5edh24&9L+HV?fM3TuU=u3XyJDN<1-c4&~Yxa6Ut=
z!FDMyU$u2)k%!1z=C*}%#f%9AHmitpH2S?I9<{%53z3)J=h=)O_Uy{{NXFi>QQnJ-
zeJ%5u5>3lFXFz~4&COITpacl1Cpcgyn*%4uXqqhf8XNQlR`WJ%@VmVXA~Ww<KQpKy
z&(FB4UVC<B)3OtFjGpX0<GikT$oF>R9Z1_#`KMO}nQ9bOebG`di3YCibI7AOb(`v?
z1#VML8<XXo*(W&rkPn_}9Jeruh)^)?i{9TdsQffQqlrD0L?dr9R+wTtRkYkY%in`X
z$sfxrid!q6tPw(E7H)9<OpryZ+V-AjoKIax2lDCOqDZZq>#oeY)z@4e+*bRo>Bk4r
zCLkVG>hr9yTTf8)UphIOI6C>dl14iMawpeS$8>MEM*XUtkLUJMIx*DiJ|=@0lARQ=
zXG|4{S8c1WXN5{Erh+v2?Iray`M9%Q<h*YQBwCmV+T1k_kqzBMm*0vRv9jK}cxXOR
zNOFEm_bTk$<3yDimjSHqmz*L(>n^lmt%pb>bnUA*gqL5V+P&xdg(yTRA(9Gx^z8gj
z0W711Pc?LA#%=~p#SLtn8&7|LN5uqbOq-u5K(G<~#S>Qm1R*%q>OpZvo<g$RP;zsT
zua&F4Y0>Z9Y`L}IvETvD$6rWog=Rrj3~&%8FcSo!tm(n}0J{6-q452=++;6DGaRs9
z86BN*3z0_hJEH-ML0^QsL{g1o#!x(oqm40a+P0=$Os=Zr#9v$hGBU_Q>vNr|UWZt|
zl=0}MQ402uz_&!+C7a_E&Tr*eE>D&5Kl(h(Vvh{8A-k1U?+GhwEGjO_fWRoSIB~q~
zTR^Ix!X~a*YdKwoWSo}BYxBe~!9U>RBO?AvFRSi=-ebUm@fb*`?<b~`R)3((9AOk=
z|4<k?IpmC#l4z_dzZ{%u*eYwK`_>QFhwp26+t9nzyC@m)azNd$98b{2%Y#~mnba4<
zEGEV%$gOdOPROD4S0ImbyGFYPi{rw07NGEwEfUA^Q*(2xLFmF4Z93Ngw9;J();)2A
zbfD6WlXtu!Tc<d0AJv_iA=MXbp>7hNzsOb7Qn+_+vj9m~%kmKr;}8;qb|a(KtqqxB
z@bPep?V2*Xckgp8NDRvue9lsr@PmQ^aC&P7!i;yc$yv;%Qn}zfjVNk~skWcb<Z6{8
zq&<F~iRw$4rM03#e6@&o5(K3WH1t7ZOO5q?sK0|jr>eI2yimw~`wHpWbfDm^%$rZU
z$yfP+MQVLa>`y87k(yuzl!>hexf%>&{8$6?S0h6mD<?QUQ|+O&3uz?$q1kEp_FJS|
zCi)|UYDGpeVY^q%=OZ|-bC=Uo4wP28wGZQY<FV*4gAjjy7J9RY$#x@48S$ObvWBO@
zL9}U)d>|fA+9`cjk~;Va!1_}#h*_w9WII_QbsqZXvs=On<!$;q4ujb*VuU}>p&PQV
zqZ^J|UdaWYNn9rX+IZgGUa#v<>6{g6(G<E)?^1stm!;hOg5-GDYFg*=4Bw%``4xeO
z7e&r|#Jv>9=aITzrV{%*2Ya8@%a%H#6!YBHzYthP3QU>iW5)}-2Q>q$x=pObkBmy_
zqyln<G8vtutogz&Ef5jQ2M<yv@%k}1mDu3#z9IgwyU<D@(Gx2f1P&S6eto&ghJ6g!
z^5n=kZP(8owJO0H!6_|DX{LFTDCPx0TC~Q4j_g-+^qLi)$2J|Y$sl9(6gyBMk;2o=
zJ(Wl@Au?#hiALTuD8*OLS-aZbmhSigRqZwkZW9Rft*<5>b4D?J8WR#8eP+~@%B%9V
zv9Ts)f5Fyx2sHGG`%msQD)l9a$tM{gePHwh#l=S|mU(h&>bu?%(7!0lSJ<jWzpe)!
zH#W;XBsmWK0ow8(A|mJ*l&++-(mD&;(xCHcB*OQPGqtMuK-e5Z3^m$N$ZFMZJtE29
zHp@HPOEW+cWeyNMMP4lK83@p~iN|+tco(j55u?b`&1*TDPVw@qNl+9A+-J9YODugn
zJ}ud-3_(QOz#D!KKj8dX1WTixSC*cU(t6BGL~ul=(SJH(GM4Mg-xd4t78y>PRsQ7l
z<t2|xm5Tu_Akm$IB2m@&Sayd5+n!a9dn<uNAtBZQ6ZlLy&!iB&#ZejylLyFw;%I8a
z#^m|+3bm;8Jy%`V$Q6i<k!{kmAbTN*yf4#<W57PevPh4wDG%9|)6t#OUg#Ib#mgT@
zJ*xd0(WNo!U}O+IKD(XH_%j>u2>J_%yGT%*%45rhJYWm?Q=AXmxZgEnVh|*xO2@z7
zIn6|s<1!l-Rdq<`O!%@?U@^fRWw8MU#U8hw`eRuZbbwK4w^ITnmMB*6HSW!uMPH>w
z7U%*(jd+xnDw;$h1AWj4G7qivh@^DqR|hkaCo|=VDXHQ{KiaK*&agd~vAN^vH_A{5
zqPdqN{+r0>$3*1oYB^3^us~>czBw=nlOS0@uqxka=#^v8wy1VgBJ(rc-NTkjelJPu
z;ACu2)Dw~_6GQ&S&E$zunY}%_on*8|^Y9lG$_6g>#whz*GWQdCqpyl&MBEm^Ezi0|
zB0n$rcDH_RR}8DcG@@*MjCQg@6>=(LcSmFrlqF)b$Rx|9j-*vDwxkI(0~OVW859j{
zUz2VG2s*okyD~|K9B<D=sOJR5u%g{1q;8qgcE9xA$+!0M@wsn^YD3DnKyTa=3#?nd
zhrZCp_m@>sQ90Nc5_uq$2h>ly3$!}RvYF<$TjmD@&N7S01fAfmWUaE*ij=RWQf18G
zZzEE;dtAJ1zC2w9_C<WCFx~J~_4By#m5Dr-XdbIvdpD`IB!brH%YHyRY_IHB2{S<8
zWHS;=ZCM?-z$_9HFCY4TKXi6>C4r~r%>;c{4Co4p0VA0Z02ePhVVXTbG3rJh?2q*J
z$8G4z4b%!1?Sk<f<x!G9_cop2+F?XzkS_b^JvFFapDYhm)@%;XVP>WZ+qAj6nqX4N
ziqMM|m1ADbsM;O53JHqit9fl9DJi^1_?H7*8$G-REqY6iuvN#}IF|)Fi}6f=g5B^y
z1IS7l&yjR3!G_x-CvuDe1|0v93^&(!I(x0IxcnQpg)k<%cYbDL|Mb0+a%a2tx+CWS
zkHsW4(Qt{TdO68FBCub7b$5ts?9Q!$ud4$)?Q~>BRcWkAGEtnj0lzl{iZamNkpA>V
zu7cXSi)K*>uT{_F7=XEQ>uv1<>63vC+t&#brqDRnVpu$RezX;8h|kfUZe<Oa3cRRq
zyyz<d`_a~H2(YlLUWXG5q1qp;<eAKrR5$2rjS84z^b4R}+o}>S!8_}oMG5IwaB^gp
zcn=D{Iq&P7z!venJG3St!x38nuWN~lgd`5MTnETp25lS$;W&7ygQ#@#iulo9k2{0a
z_Qs3N@VVcV>3t@gV~5)!)jyO@Sb~pWRYu|%8)e)2h^Z8C9AE83$hZ_CQrmXI^VKi|
zU!_5S0o9f=E4IzZD0fwnBWUOztC0x~6s2etxSkczQ!FH}RJM6z(yQdgDqnKRWH&r|
z$nu7nkkecTCoGp1$up*xmyBs4woJ81W4E-I*9w)CT~(iGrAd8dxLAHcTmdrZ^ISKZ
zNZ3*!oLP;QMd~@Co&KIhV8&9@-Cmj>Ti37dc}Bi-*^=^Mx<ZP(`FDQ(@iP;`37wl@
z*(rMW+yMIrOe%=|3&+k}8Pvyr$z}OK0rAFs#Ki^6Lsm3iw7+FO8J{A~&y(u4%~FgO
zIYV5~u+*TUB{W%+<%&7bTJ0LRTq=q7^LtFnIP&r+^5uZ_wtc(|3Mq0M(4$csWDs(@
z;0p*$hES<IU#ze*OI}otT+<(1k+&vV_c3n}HPo4`aXXZ<93#LYres+is>B0zml;Z>
zl%OmnjL7IQ5TpgAT3p#6$f)hxt`)Yu$t}T}M$YTiHY(K`{i!nBLk5Jq#9um4N3!oX
zEcw>Aq}K7*LefiVRSKHFzyENXUNiO0$N7bWD2|nZA_!-NpNPTNP@^(<Ay75JJFOB?
zIp>ApFwwjKe8Xfohe|H>cB#e0+o*eTarG=Nk9mQ!X6olGqUDBQ>XL9!k2SH+mhAGJ
zrAic`c*x8^lbD2r)omXBJ5E$H=b2(CB$MTduD%&?5Vdp(B3^yZMyem8X~4^g1*LqT
zr}HSG%6-IAGllINpVc(Z@jlqWz1E+qH0zAC<lcod>K3hk;*DyHvFpAuqCK}Dh?e#|
zUTV+oIk0lk9T`TW%u)|~r6lvwxtkXklyyKxheFsFI|w%;gtw=|)M(I&dCAu~Y(BBX
z`k;oZw;}~qER3txn%2x<Zla>0eu`QS2n6R%XlJ%HCJN>Ehf_7{2(_zjZN?TBUxk^q
zs^EQZ-m+QkH+1ZdUg0vCEDI5E2SUdD?Gf#k><qq>9jH62(PtuVYV$rmn>|t9hj*Xg
zCfeX~vQ(Q@8#uZ^+jr}a$NzM0)VN$=pG>{05^p;BGp9ifwK<_c`a%Pa`2C-_w|``2
zqk`Y9;}{bQx~i5`Nu9phvD*#3)~>=InjG|T4@F#{kSFIQ{rMS$Pe%ki4_GQAfigJc
zfR&jW&6BIk8n~S{f5`|{^2zsF&@7R_7oKc&(y?g34Gi)mW}CJbur)OoHeGR>N6Dt~
zk7x@}o(ExWDrff{Y`?P$yS{Z+gZ!ck4FeqYVkeqcL%CrF?j~AP0n+=+zKQR00*Cj(
zk&&)1PgdA;_LEV5+iiNqim#+(Qzd>mbk2ucQArLCR~rN^j*!MrG|y(v+9#(PNs^~t
zEH!>^GJbQwSKge?`{`Bf{YC9^+uK2Cp0Iezv7!`$Cm!%d-a-4~k`e*Ux-R8e#Al&w
zwMyl}-UaumN?(fSJ-iNb#VJ7cg+@}NG=45e*Hd}C60~|X5Db)+)aS<gsog79!jB#-
z2{9?Z$P$Wn8Eg?%pJ>|H%OHWK2FdES*%AoBwL+l5n;{6<V)E>bdbaz=19Y-q?#me%
zR02%jKwkl$!`rd9y&Xx+ACKc@jitZ`UJ}6;8AA+CtURmv1xz7fVgQ%~8-F#ML3>>9
zA;sN!9WMZmn8mzrb}%fbE4)3>ZuO-+TQ*pj3C0~A9jD*5IUlwFgQ6u*!adB8=_N5u
zx4Yyn-q?&+h_V3-POQ4i6dVlJiW$knWaYR~gcT6<#GXzzj<IxGCRf;-HvKNBb8yME
z<kRgf;{n%R())Qg3d{9rZp}q(FJk*Dim>eG7eRR71LFFgk^9bCiEawQGJUc$7%Z_B
z2~N0~O66$pU7m!x=FTBe2ZzHVtY_m@ct5anb~rZvButD%_F6xxZ_dn8_eN%zX+m{%
zU;iyD!x;IFvk|<)mKPXT8|S)G@9$G_MY}%3mqK`W-{;dE-Ir^l^<UKEgj@-$oeq=>
zKUz!|N=8o&N*uh@^&z6X#z|!=a7}7c4jW8#l;P0zL4l&{zUvccQZA&@)F@!1APQ}3
z7)~)uDi?RnHM-1H$qTxTIY)n`W)nJ|>imhd?}NEU!;O;?U*9y=#=z$`Z1?A0Z5TCN
zP?v-##mtG~J}Kz_ayD9GL3DPo_TV~uCP8S)*LQ9Ka@Pu)<p=@d_)aFeWL)M?*8)Zx
z_yGj#spb<IM9Vje;J)>0*+IXcO`Tbeo+M9~RHilsmq|$Rk_y`5ytNqbU?BAG{XR}E
zgtuu;z=oA1U^6;8TIb&K=HXG-*!TfRN0hx@W`2A16+0<U)Ug5n)dt3~&G8!vs8Dr?
zV&hJxkaooOxSg4yQTcZ?o0;tjyNKglxkLfVWJ>weu8Hbe$y~DL@Pf^W4AgKdXIV`d
zHZIcUY6uB`P;Ojh%|HoVYYa5>&*APPaoT*lZQ8DDx1fG6`s;i2DupiMmRReS9<cqZ
z?zv&;&;1MIPf*sgNlOc-q><7~tV~)rMRGTFy^Y!B?~OkwH2*Vs-9K=lQ$w(|dA5D9
z%CRz(9Kre3FNZ-TTO$bYInCfd?(S`ad078k#F|w%>xHJ`OIst{A(^k8Bek@|+$DZ$
z;e+?o3tYT8oBWb}+Zed@E4mm`9lO9bWh1n(DC`$tU!(U-M(|d!>4icss`2<G*Mwob
z--FdX2<hp)w-y_p=k3l}sT}|XF`iNwPn<#bRab~$rQ|L}KzM@&4VtE6nq}-d;eBkY
z1Wos;byNb1V*T_e&Ye|42C>uoezL9{v2!P!%I>?{4PRD;T5f@3S45g$wx6Ea;PW_s
z0@j8oJTg#)t0|CZWVbW1akSbD4}e-$--vlEFL-&;N!|ldy-BrpA7^6*YIX?hV7R-e
z*l;q$)kkxbwt9$XSL=Rl75nftLZ$s?ID#;y$ULiofVAZGDvzjAEg-`*+RT!V@yw?m
zDAC|BsLIJgQ8lkYjnP<s=jU_TB%P_HGpq($(D)@|{at;Rclp<&?k-nLC#>S{eozVI
z$k_9Q1TJ`RdlnGqYtybJ?;ORE%#Lb?oyH;}eW8OL$J3^DG+|v__x{d__nF>){M|ya
zW`7r-@*2!Ze6F!T1!+lvY&0y|OrxY%Ezgn+Yz3@;e%xMQCXm81+^dh#p+c_6Ea<X~
z>LnV=X|Cr-0Gake7i!DfZ93j+SqIU#&d5tcpR>JA7DI8muv3~-hYx{)&9jdw{iiG)
ztJ~(mJjD}}O&SsJWVD|*8O|nHVu?goSD&ChY}w(W<cNLi>IPVm2&3r16ZYXu>bUa$
z$ve$>+SIHcTWha{8rm?a3Uoc!%56_SE}9Hggzc5z-y8_=J#m-?rBWOY*TU-QwuP@w
zZZS7%?o3$KCW0U@lFJNt5AWMXQBx<dF!ZZ(%u$?#_L+@HVe9EZ{Q{h$)ykwH=pmX}
zniFrOG`&tUL?l8io;9nRlRR`Cf8%Mvy-em!V3|4dqI~m9q?Mi+LX+2M21xSZGQUg@
zk?eA~v}^SRVP?P^yS|U;GD71vX@WCp)IwaX&-BL5c1`JPK(=*Jb@%8%S^9nR$9tY_
zEyg1yIX;*@2UpK5uiOC8^gNc=Bi&nGwbrt_->~5wU4U24#Mfa$NdF+W4mp6>AT2V4
z`t|Q`)2luW82fTKTYssFO2~P~*?s!KcH&1WJ0Ho-qbesFRA=|ugN%VPhh0iwV(Py>
zlKVg!wK<I9@(voZazHeQOMF1VwWs>JQo!FLSHrvU24XC~eNSvYx5L&_PK<oz+f5wE
z_UA$2@~gS!_Fm&?Ct0oU%?{t<S=v1XKB-&nynZ}8l{3>8D!?Qd{{*+UDOZHwAU%Tq
zB^F3Y&p6%ffqP2M1*4i0zEGq6u|e3vAqH)FR$f2#ZYa|YmD|s%XiK`Pe}VhR1+%^u
z8(PrjxmyQ532w_zelT-jp~#T^5n$%%iVrn9H0k?XOLy%+DWHC@hBCda3#yHG^*Z=>
z?FnWif=D07DEk?edCP~;D^+}z>{Ty7*PEj*c{3Ge;<G-U<IjFIoZH9R$YtfTvvZoy
zWD@iJM1zQgNk(BF10izf!zn-ohq8Ksz<^cvJqxZEOYpK!7n=C?bamLS`(6{$!Pbz3
zQ5AM-eEw<T$x0(cL52*#aPK{kj0o+jEiB9x7u}*Yh7elV4W-iPg_QHO?k-b^(1_?#
zzr@211Rx)M)h#$U*6!`iZtJ7VbKSMUjQkD@nsZ@hX58mN3{)v5o*($jy)D_#E;mj&
z92}h<Obj_WIqH0k8Fk~bxf%)N*X?Ii{iY0&m6}(4T}6ElT5-02cKARv&(8<a%M0*F
z$?2qOAaY7F9WZK(R^k*mT*^B+;Zdz+Mj+iH?YMM4*zh3WxC{a1SiGUEjjJsMiMweN
zLfuz57T7Jo>a4=#aJeY0cjvMr><&7LZ2w7WP&43d&jD<jfr&MN>?Q{sD^q=X29fsC
zTJ6$_>x7DJ1F8xrllNj%_Vq;T{;5lF^jX7+o(o%zf)!YY=6`&owod7jUEL14SRw^f
z!m2fiueRz<<dXX)nCl${k-i{<7**l+0~FZN%A-prq0|gdqt;dCiHZC*QNV_9OTYV_
z`lDT6#fv#McZXQ|3t~XDfqh=V{RmOX2iEF=N$|HUW0sUluwsgY8p$YhEnJlEM+!J1
z!}fhZlY4A1kCjSoL;TPMK1aPc<XSuOQxY!I3u(!YyMZeqZKynLeUfaA-YB=EOLIH+
zyT`fA=Du*3fMEh+hRiN&$(5LTGc0Mtio-D=LpwWMrMoUqV?BcJ3RX{Psr8dkRmFFy
zYHjar39xBYg%q@fGsR?nc=1g|my+<Cw!t7DT_C@dy5gK$bbtX=j7fC0Fr_;PGLom=
zXq8F!9?o;I7}!ld5&aH_^Wdd)-%M14^>-c_H%ovUyA7l^FhmTgufI0e(A%K5p+j|T
z-*LJ)VQUE>^&H9?X;TC;mwKQU|GJipCxX(Df~3gfrho;FUJzR~rrOP@nF+m)Cue2F
z@F;r!nc+pAJPF4wqa;x3BLfTKgH`T#6;HyB7)mD-eu!}#6!7xHCmJdGi(CzEjMnzA
zvH@w%?o*BAh;+of){nxg^%jx8DjVwsAYNUaTLqA;2LWh_t$>X43fdM-H=pn}?j?Ix
z^l0iE5HL%w`5KwKja0{czt7Q_qT^x~q%EHVJo*w!Sihz;oXALCN<SSmJvtd7LBHVl
zc!WIfdzs7@*}k2#EYX6!;O8VQxO7Hk7eA7ow<@CDe3Z_!eMRv=f8?uRRjlh>A+jAZ
zNpe9zG-eh&fLs4TBEl6V*Hfa$&GY}v)DgqavLvgC8iUSR+9zMDH0Q2lrPB7l5XBf}
z5OXzZvhlN;iuxi8Eb-U3f<Q`#)?CG8x=g#qoIo5EPhBz-ElgglSa3c=F|9Az8U4a;
zo5hq`jjRYr4wZRUg-R`_ypkKq0Vn1~couu3onpKquvpZu>r!BCU&ar~=`;-vWt27K
zetAP({gC_f_~vkx`<Sg^fkQhbkTEz4xZwL@HfrRX4$zLJf?OsLEyTTWb*r#FWk5I}
zOVIV<(Aq8yRFTQ`6K5}w$#k^Y9RbID_E7@i>x-iAIqa3E^&Jcm{nybx$>B3!KdaB-
z22a%1#BF`BP(XZ$_DNjo_5AaFKBX>i5d*5hY)FQU@tR|_283AX<8V&XsHBikvz?6%
zI$NJTB|thd18RXlhSe^b?Bl%$udrH|d*gM7vmX%)yAgxed;%dAoaO5n+PK0!?8~lA
z{|0pJDK}-hWjTWG6=}3%wGS)q^G(@8h25D;IpTQi;~O3YSYTQWWh#jzKDiyjk@n%?
z=7~(M`=|5C*xdy6Yj1g!zJk!vrNhnn*pX1^Lz^}S-6~eCF7?)6JV$w+jF>@>fRtOE
zdP(_C`^cx#I&rOp4@C5I)E`xm<yS5A3ST$TfyxJtcRLN3k_ebX?RaCJ!Po_WaQ=gt
z93Oe9V*(8^8TjJ*YWkDmUt(Su{#HRn-bIDoBKzp%!dk5q@TADu8L*(PM^R&Tt|vPb
z8k!orfb5IU2((|s2z>NDB>*x^QcQbJJGBgoiLG2|+BZMN+ZRE7HJF{K@qpb5sWFrn
zAe2G3LmJ3aLlQgdG`q^h4J!7aTaYLYqmEB-p11mH-QA-$bDu3UW2iNamzenjkEsNZ
z7Njk{pF*dJiy-bCM$j`@l6s&O9RRWwHaH!hF^zclgFs1`B&+S34zZ%;E^%m$>sMs8
zS~o$2BQpByoF~o5s#Okb9iO?bwJNwm)ozm@=z?{8FL_(V4Ki%;g0|>O5D@G%PEV88
z#$C%que{NJu3`2H6%cBSTt&@`&r-uA*)D6ZFVVO3C&LQpnAePF(6L?U>RQC{1!VZ+
zg`Lr?tgILVH0W7Qa?kDWMBYp2cyF{6G8s7(umBOdx|5}x_d~2y-3dr;IwRu6Q-!?R
zD^G#f@8V4KF^;>lqqDA8=bi5zG^o}skZ+EOJ4bum9ZsOS#LD2~6s4trTbKj;BFo7R
zyYuNY2{qqO=bKSX8>Rrofw6b0QS*w$e-avm_*XVg-2_J2UMo$RG+{Q*+z&BHGbEaE
z8>82Q%>AOE%VDa>SoTyu6>MB;S|S{e{Y`v2;p4h!Kf(I_C8l-Ud_Zy<dgp{vmg+4<
z@$RQt>ia@}=PCy0-r1O{#H)5X3in~QK_wDGU1&L_HY4vd7j<zF7;ejvQi>XGVtc!R
zZ@MREMR4}L9lNYNu>SZ#O#~F3bi>Opl$O?+OumFO%i+9j{nbdWcbj(z#so<wx{Vb8
zMn~UsJmiQiA2G3A8_S|I>e`V|9d(h8V^>mc?6qQu_2o-*nnA{vcfZRKd*4FFG=-A@
z3Hw_TPUIYryg@>!Gz&}@Xw~lOmT8#OS!ZywPgNdA2v`N(Y>Z8Odk1^zDV~bJ2Ul|a
zB$@T${*BZieqA4IwcA5RWgx#jx495k7M6p_+fX+`S6?^sMG`wIV`~8-+|M^Lh|hi~
z(A|K(9yu3qSabqskR62CGca%q=m0u!o0kr*lf?XI`D{N|iGJl_U_b@{14r5&$6Q)r
z5q)Kjpyt!ZNC*R8EWc#k)+0uM@A|Bc>rX|qo5>B|E6WOx;+es0HdT?!Sub}V_3INN
zxLZo<!^jpphjHZBwRzj38K(2BxavN<cTy=m=fTWo<d`~Kn?e!u|CsGA@RP~*`#bzi
zXk7uX%~hnA{13W*`tVe^gO`_=&iA&Emon#P>v?Q%R^(jHH%r4u4g4*zyRC62ip}w1
zgXjyBM<hw#oew*TmVCxfc%|HxpBKycoX}WSiC1XD=;6Lin=gN(U=uv@=zjoVbxHAL
z@j|W$va+&_A74^KLq&1Z3%2yx-i*oteKpgV1Ky{6_ZcB@;`<291KZhavJYr*S=0j9
zQ=XY=vXod83z?S*qyf1c;1=J{RL%*f?ya=H_8LfdJX&g1TV=U3TOY~kNNKC4cHgPW
zsQZ<v1C}pq5y;2Av#5VBxmL{D!^_7@tA%|JNj*yJfi4kw-09ZiCq4y0OV@GKjU4M8
zb;Ut(#jKQx6V9aV0~T=V&Q8_SR3V5~jf<EOBfq^e;DzDeT0F|lU=_|&i)&LlTAdf8
zS4_=r-(@M8isj+K!KQouBECNDTsA_yjZv-W)nRWfgj9F*o9@Z3FWF)_;G=$055>K&
zK;3?BF<m7_NHJv8v)CDtta)C7j9xI6(e^Z*yYc=ez51|;%ZE)XO%3sRUwox-5jIi*
z0l{FIK{YBN&QngSKH`uN%oagLsb5@x{9>8Wc$H*f5VrV2i0gT@-9rsrgig1G1BiG?
zLoh>-wd<WmcM?IhPKl4h5Ac{Tg4|rs=jNOJ3%jgpP7#XWZpmdk*o1bTuje`A18#4L
z)kHk6QJVCV+}MZr+#&!Ez*Sg@*gqLaa3Tfi;Js1YO}^GzUG4f3FZ-frxlPiuvmI?d
zU`NZGUWP~LYH0Ec{ce1hh+9<DK(rzJam&nctnhJ6*Iv419>pc7;+YXHezej@!eM-y
z9NHCWtjcm#=BQ|7E$~+QP;7GVUT0sfrn=)5IHi6M?_1M9+Q2v{-lluTt~}l7e&x}5
z2akq!qS(y0d|#ASDpAb^8b)~5dhF>Ud{(m0mB*6}A<nzF@;H-nO#uo3U>sTDcnpcG
zls?qQvH63V9ahV<dHo^48W5|7%!c!I8(EEeh@Y9|ik12r@%nFRaR<S;hW$SOQJbZ+
zlOC9iAiDCSeRCh9*M6CGJyzF<Buq!vBGJI{pw)M}k^c|rL7V+cnC76TQwAn{>OZ}(
z-aYoWCY{g5=lQ;`&`Y|rw<l^-V3-+lhEg;zzG8CXa7xb1{9MQwi9^%+DMrPc%}jQC
zlcsB6XJdF&27rL}*P9#T<bXi&(M6$Y_G0&>k1Bvn`0-eYyulULQXZe(E4t{~LWBC~
zA+xW67STq$R13WZ92yOWyB}Js%J)BS;)o$suN3!{fjXnWlG*!R-Hq?<ar{G2mcuUV
z{wCuXhW$KA@K-L|!Vd3(6Lh-b<4&-&4(zR;-VwUU&sV<ASTF<e=*yN80HLy(?oq2Q
z7juu+_}5;efs<KSnZ-aVmr-w&C*auAM9<C7&$5#Df%Ft}!DQ9T-TO=1DtTJ^Td@x#
zS?g6=Ov-H+$qQ4IvR}wsxy=+l>>C^eEHp+9d{7e**(@eh0Uz2PHP;jzpjD}C)T~H!
zlkWLJbN*Z-@KERWbdgbm0xmPc>*b#26PtxFJT6L<u@j_)zI@$V2EFkjd9G_${_gt-
zcr03<)LE=1N@TQHr1L^TLedpWgK&j4IbwP38@zV$`ONVFUXh`4O(zT0iS6ZWK1qwD
z5oStS88}<9Tba?WT99A;R=D`AFWd2T@|oTGl1u>UW@_Y7lbw@Mnb!$!6ovEA%6&Ee
z!n9LakzP_kENS2IZ3sx$`z0jkNJj~X4U0ATu)cM>2x()Cyj@}REYU1cH}R8IIAeo#
zhOTAjr`vg2HJ@B5%QDe*4}cseOY6e+viUH=@?z(bc4j+MpwdMMHgq!}<LfA}LBwd`
z_TiidEhye~XJ+8~`=_xVLZURcrK7K>Z`H4vK?{qSo-T-7E&!5$x=QcG#`Ja;Qx(f2
z4&vmWH0IeHq+iHXZ0|>xAnRHhLKZRX+9J8@{D^C%YEK%(|H)d<cOl`m-g(EFnNd`k
zahC=Vq@T0~k?ansXj5p{xHw>6BaGx}`a0BCfW$hmgVkDBlQDPY)4;d2HQ+fBC;-(h
zO*IsW95R*iXk|$667h&0nU5{*x57wSX9uRz`!T7JubN~G#;ANAEN3j(c4u-~TU!So
zVh~i|GHcwKcwZtZ)C&^Z;zp>7xqdIxbs0D&XVRX>J^_*Rmz&S`F^r&=Q<Z9y`=PWW
zV6oQIS`l#3ntz?VS>?LU4Qw*dH^$3R*x1-UK32%jlrKVPXlRhz^0yVwbKiq@?BHNx
zCeGg5U0$8@>ri??mBor|=)sA&Ia8}`6w7aO;UzG_2jvQ9Q4ca=iZ%@+Bf|6dxP{Qc
zXG*#`Ug}v|%G04j4`EFD##mP^ye&D<w-G$nk#WUwVaVOUsK({anf4LVSW7J4nWxUM
zjJyy(TA|ofkS^Op+%S0i3$+kB*~5TvrWb~tk;8t#wbiQQM9EkZo0d7mE-#;zFnX65
z3Tk0~FX_lRoFqFZ$JNr(>IiRHLkPdGhyaWUL_HDi%B7kUELPNBAM-qq7x23)Y+2Xq
z?B22<kx53*pg;VMRHMe0;K+5;=4eAL$8A+OLAa$X@NwU~m?6Ix(jkP%T?b_U`&}}J
zX><{aii!p@kG*G@6Re;$QsJdiVGO!^pWE3x6H-Y91DrpAHJ^f>^^P9Oi|%C8fz(fS
z6uRkyvnLnLucf1TkzQkU#U;6W2i+y$gC}!8C{TdF4!n~O{n)5`ut$I|j9Qj}W_2;p
zzl;{>_l?xB?5QmeFtC_wMRwVyEinXi++^p_oOjenvxzh6DKiGg7krlR+T3N{pPycH
zK8!MJKER1&h&cMmg&HdIOd(ZFm98(dSvZJR`D!9RS&`B)+Do+nJI*c~+%uDconlr$
z{iH@g%k0lY@q`C6UJ7dthlNr8CHV}{lCQKFfJsLs5v22Z-9*sy<9$2KR0t9_ZGyYI
z^Y@Oh&J`M<IsHUSWRH%{G!Ve*d1qfLi1;CSfKUbPy;3$y{DmbSV7s~!R_$b$RFLC6
z&TC*6V^FMU8^~%FNGeF!8IdQ~flHeQ=u$9)ACB9KK710PeG@OLfW>ko0V}pr=8NYg
zk>|(TjhzE1psGnfp9+2ftT-Z%)ifx+^q?#%pC*r2vMK)hV3mRD<<Z8=k>(bFr|@HU
zS#y~UH35Rj=g};Ua(2Ko@Ny@+^w)-5W!hONu+crQ0u`Mm*Y>}+D>`jr5unzd=f+$d
ze{PCqEq9I6@!`N3(CeH_6!E3$<S57HT+eXKZ9Cv~?8jl!*RwS=bz?qja@|k`gcLnF
z^rn-oG+j{PD2+RaSSSgH;%wE^h3rw2Cu>V4x_oY49t#2*al!~dWBjw-Ue7HgU#7?b
z=8atD3sZycXr(-X#D~DD#Ao(?K=u4K%`-farkE{VM(!ayE*z22K?JS1G90H$n+0*x
zgTcVXJO)>9;4DiN(5{@*R8s7LVKM(V#pMlRAx8}CCsjB&KT)s{v1ClOInlSgQ;m&{
zw+2#WidIV0^%g+#rAf;$1{U*@-$nW*VCF9|A4RZ4BfT(vYIJUTXQC6@v`c3*DEJrI
zcKcf0<vF^65tlB5yp?P8WU@L~r)M`8JUhZ_r6!kH{C084t`Pu(J3^feI9m#|&x;{2
z6lj{Y2F1OY5GJIzKU`D7=Y15mm?P$o`{CtRjFOVup#k=~a#*w(O=d|tW@e#`P?`p>
z?*cC}idfG<)eeGk5VuaA^nZ`mK*c}Pz$b2gTi8(c#hmQbdo*F!Gk6rL<k{_Qs5$UH
z*CMIGrc1C{plxF)EM5IfMnvR$^Ab@9s8|a>QZ_0B`kizvCsG89sqt;a(#L}nR?*N&
zA9lc!xjehKbZ0-T6r#27+HzXg*r+*7N3-jA5<$ftez!Soc42Poe(UN4DJ4pD$JdB_
z*XS1a=C{yF`|Zb%9^F7A5ir=xk?y-yFqPLfpW1_33Y=g+SXhA!uin>QZ`=<oZ(o*N
z0E#0r(>Ga<W00uoiWelRe_ZcyuoBY8=g$}s6oiI9H&5BQY9)Nb`Di0#*=qEJDw@m8
zGXLWG%K81n*GuHoywvS`GJQ-oGaSY}dV60aM=O+ybFvFWifX)(kt_zbS&9_k_MH3&
z%#Laa?z0hG2e=ENTazaKpg7{D+zQfi-D!;@vb}Hd6X=q0d)xkj?jsqBQieja&QP;4
zWtK1`Knw_#0*17+nSLB+9>8H{Jt=U|{F1FWEin8cx(1H>G_(gRIG3=l<ELQaKi?k_
zVq=JZN|u1=UDsUlo~_FR#+ou)#s^eQO3zQDZu!k4Oa{0#A$IMPz>R8|*XcVwxA`P_
zH}@K<T*Fg^N@a^To5S~jmNZoD(<eUAlNNQO;Tiy-$iuY5AjsjA5lGtVj^^ShT_u=z
zk;zScq1q${n9~%fuRt*EH3|oogsN1HS39w_*H`L<pG(ZgDD_1ww6Mp3FF7#&0a>0@
zg6D=;XU{F|D)pL?v$adL!kN0nDLvyxg^ERd(eh<zixMgtC<o@YrYbRU3293im5B0z
z#!PA<JS^-UI6dg0D1&qzL#o}xXT7?2lUOfyp03FmE>wPm1)`g#R#!HPuRPD5`jM^#
zQRyn<jLlh2mU#g7W!q+mm^+HTUTCV!j>WlEUxkcW=)9zXX62hztl?}ess#y66+6M&
zt248&1!kUrHrqz~A;1=fnx+cdq@(64e(JJz1m&vC$Era2S*wUk=77eeTuL8N!R|l`
zOwXh$P+t03&OICI4s87lV4Jrxm=W3)5J}_DTEF{?N%RMW3<Dlun03cQ;Zpd+j@bq5
z)CeWFn;;U5#~5QXP3Wy}AVAbJ2lA%j6V@n}1mIvhS#D1FV<gc2(iSb^Qf*9q%T+@>
zEp_n4XK~|=v`%+d$-3z7ZTAJz^T)1otA-8Rw=*CNaIwO??zWe8hu9+!Bp`Wyd=(!H
z6BOs7_ivv)#cw@3-iE{pXRtSliHXfsZq3-Zp2r`m<4>I(tR@@yOcXl<OU8`OWmzHj
zz0=U%nZf1b>|U4a<c^}YL&1?_jhVEWL%~+M)-)#<tqM2Ho&}|`>E@q!O{GbCnc>xp
zUh~H}kS%oz#1>xJJvu7s5?6P~0xRDRn$iA`m9O|t*>jtgz5!pu-J#m-+t&pinMzfB
zcMDIrV_=Mx$A=YRe-PARiO5)VUNYNXdTrHjPUPIoXqTItyP@N#rS!;CmHyHuE?17w
zUiSzug51(n0!zHL;S7<>lq&^yrl;<wHDTM((<?#Ub1fll8@J_)p3}H}=S%_j@^eVj
z?(5o)-@4qvv&@H6pPgaO4kq8~bX+Z`D!c(Sk{VM{NSTTa^r>%@1ww!5@W?$Lbp@Hg
zi=NyUXrN_zGKRI-=KFVn-j5f>#=wZ0+c-z!n3OFn_Qz#;a=bOg@wjuq9=n-4wE-2>
zz9UfU1J<U=mQzh#)~l6A*(86%#LlVnbyZNG;C8>r#8|9_Ta*fShbl>hI-`^D;Up%W
zrPczb-NKHpk{&)O0nE&SAg@<dA2#Y7^5FHzgdU+s1UX!MFUXe2v)W`!SWh^9NjDw)
zzdq|GESvW>ySp3!D@)(z%m0R*y@D&aZfbCow!etjxy*AlM&x~KU*hlx5J1J}8oV*_
z@X|(xXJ>WF%F0v)ZauRg@3syfFEanZ9r=QFvsO;qe7fjy_|<6|4F=bOa{<yvY4L2K
z6COuaz)ea=WPtlaq5t&Hs)X<|(hsf}{%)6N_oL=;S+oMS1S?)&p2yEXG-}=;RXZIj
zvo~&z6(QQK(-`?AlgdLF<soI$OHMKC7P+MlYi-;XE?SPQG|;Oqopi)caLgx4)o@=V
znMg9Ce|(-|6+NI*0OqwsmU7MkLm167+{sFNbTYRS-Lp(qGktcDzLFMi#{NlfK_i|(
zEKsRBi{x{nT~oq#PR}Ca!M=*2LBU`l6@GK2dC3DicO$kU9#YJiv9G1A*^w-xw|o5A
zU}P{B@1bz!yvM$NMJyQ^K5OIQ3Pp+J;@OdFt`&=BMF5Oj9X;nZ?dfD2^P&dYw}!X&
z;4o|-WmzScrGN44{kHFkTfn-z(Dlyg_N7a$(@N7#hhD;Dm+SRy(f|Z5Lc@<NEzOXs
zaAwWSk>NiqB%<DXq;P)%YWGdG5`S2)D|r0kNi&w)EtFW33G~cXit2Bx0#l3H2`W5W
zwu3mjh4(h~Wvo|mB(30uV4vU48`F7c0D^o6fErzJ?Ajx#2b8UE$r9+v@ehtv#t%kP
zIwQUnNys*%9`8>><?}2BkNh!N$QRfx=0C-@N3C<`PW)y>2G`itxtZHfmbq#QO^p2{
z)V??ge|UKh>@Lqn$H(UZ9os3)KOmr~&g0gAc8VHRD2!^$&Jh0;7E4FP{PI<yPzCM-
z{c2`9>IpiOICgquhqXIaL!Cx@a+o>-KZh4=a4p~J@%@wWgKC#kT>?R_F}lC0wzzA2
z?VI~Z9K@|;3!&>axvk}-%TBRBfB3@*>CNyh6qVlALNZ+7QIm`6*LQU_^1hZ75)>p3
zfce9hk-wcQ#^0-3oysH=^z5BSGEu{lpQW_cyW_V}t+!J?bG%C6ueq$?aBmB#ac`@?
z(!Jh6w7;k)`Q5_5+viWZw7=OPy^BZFsJP-M7CdK|0lF!tMEy=?tIKLz@%uVWA21_1
zjDxi7_}{F&i;Rp6=tmtMxC7Gz68@ad?O4BX^8HtCN<>9fOSH~+W&gHY{a^n1)E!>@
z{BC2OaZuaGp-c_L$3ls#d7m;gl0)fa!Lol(wfNONpx>t2E^#`1^=5sTDQvrIUNHtN
zOc^qfi?DRfBZObJOi~Xh)0jkHt^^KQs;++_TmN64pv?u|?c_dT5oOwgyO2HYkk#7j
zscL7Q;cPWeu>7d1sTtNJ<mcl8&k|jdjwM)_R98ujalUiYsq9VD0v*J#z?J{5*Mhys
z^dgfg?q641@XP=7^@tDyW9ItEb~D@M+UyVM)of4IZKN+GFfL{g*S08ZR1yUJ;l>m^
zx7LmP_w>|1UHk?TKAqgpb%~Y|E1wq-c+@l^3LnGzY{&8rFr~!Xy2{;7WetA<yUD-$
z3bwp+^3#rRHB~744pP+vaMkJupBR7aREz<?)&cM0kKHi0iGTA2OXJG`E0VWI-IExj
zPX54eE4KM96Os78d!)ZUd6X|06;Ww<V+#N7=6`*_d{tN@nU56e{;Ssb>#eDS9bucD
z<5@J*Uy0xU@=5;Ri~s+l#|fh9K%!)4H4SR?oi=~#UyX#F={u7Zb|_%2a<)2N4xB*;
z({~!(emd+wLDAPc0<LE`AT(%&2w%alrbOUR4lAZ~|L>Bbf0)pblC!|sg<hi!_+H+B
zA1~;f1mMt<jiTNzM(;FZQ8021o3!-feluqoKR%i$v4EkI)!dr3ZX`ggST0r^7Wl8`
zrN<y`zF{Zs=-8O)R4NFS@=Hs7Jv~KU&BQO(O2Fp4#7Py`e-^eoaKO>E6)>_7bfjar
z?~;;|di_|r{(Ix;Ab3nyJL`j%b5i;GY6uy)$a5uA$4B?U=2D+iz~S$?1XeH(qpo+A
zm6ckRpibp`usOWca2_N7SJUe$6C6k*dj<yHZF0AR&^!xIu$rmKurmvfL4(~0jn6~r
z_hQvH2@i|XLDGoG%ldTB@H4X}-|Jg?>%UtT7KQ)QSDh$amP(#D2+0q7dP1lW_4M_1
zEi5ebBn6nLV1^k~M}+IYy+L$r;K8ZL6JT;H#byAg#OJY$Z7z<u`*-Z_udCD_59%>U
zzOz|6qu1zj*GbHNwf{D#N+A=dUfA_U6z1hONRa*0!37MbM@uvqI5=MN$>OUwF@69@
zr2oBVeynbi*baa|WNu-x3*x70?2#OGUjXM}CVIn;E)D#V-9P1-;E~$!;jhltSzPwo
z={`UpBUwAZ8BVT1uQ{L(WB8|>4w#KkA*w2=R0v<O$YfB$0sTiaT)L{9zW?nFuD6(6
zp^<L`{q)~#T`Wnmr%pyWU(2^P@wR^<y65ay{l1z9)N@aIpZ<ZIB;^uX%#&94ur3k0
zH+5W&F4Dtn=6W=~trow($BZoZ4h|>b#%=mXH~+_vj|7omg7HXx``nhH=Kgd&zZV5B
z7<d+Nb=YHVX0(rje*NVwEPPg-8vt<5=+i5Z&QxU%Bl=%2_}AOWl|0hXAqVI?=Ua!%
z!;#(fyxLDa@bHnS#&A4QnqU6-<vLdQn#&{gE~v0}a)AcJ5bGA1y5#?S*1uia9ZphG
z(%RmBqylTe;}#;T`Dp%`8F8M5F`Tqd7RM{C55FEA`7?{$v_b}8XFdiYq~z28c4g2^
zCYY)Pdbh7GmW9o#`oEegQq$6cmw0bUOI~Ndi(S^fLHX<DeW+Ls+AzF)ea$#e^4u>U
zT%Gp16F!$1RrnuG{ny6_HyKQP1elRFR~L?f4{ENj&Q(D%OOg_=h)mg=w2eb)LU}U3
zpHvD7Jj!)_<#tWhTz`@nih#q;JFS1}(?8wl(R~juFZd7J179uDWTE1KXKZ3b3EGDv
zgs9t6;g4aVgAbRHkzs9TC-(TUhqz}ike`Dy5`nnHsMNoBiIBPoZ8tw)HZ8jF3itwn
zfu5JFyu*G5Ybn<<6!TlZS_(%F8>E+c(MSc-bqlTMo&dIUfbze12VlGjaxP%e;F7*N
z8>yX#USH7VNI!koZcY29XiVwXBch7Tf(NOOv8u`FQ{~~|5pcU0W#y|^{`Hps%i#Z`
zp%T#GDJdzx6`QH#s8qb&B;>Q-(m>3ar8y7-_c*hq8%6td4|+Mq;P7xCK<6|&EVEn3
z9WS+#eenA~|M?C#>VV0MKR^Y|R8P*%DB0ORiPO@;MpEV62J^4>@u6S^Rze%#n4O!O
zi_ia-5!}WX2wUr0T3T57>i+RXhB6yae{lhR4|%zSAb34Jy?o0lKA<g~2Pjd$()F_Y
z6{7q3#LQay?W%y9JFEoHwYT6W9{(O?gYYmpERf76xnEzsN&M3|`llb?-~eKD(q|mH
z%w|VN$5jU$dk$PMcnLb=421u;<>sHR1N+^`{2QTj(kDwr>;Ga<{N+Lb)NQwJv85YB
z{}+Anm$i_V5!MSeryG2~dg0&RTOSjwkOh;HN1p#_^S@mBpDy$c35AuQ$ID8n^7-$E
z|94GQrwUs*8SOt^{=;Jamqz*XHFX+bbAB$SH&^jrJ>!2a#QOib^Pi_RtS$e)?9Mua
z+x`9h-x40*0-~nK?3?`mb%f|l!uk36?d<HZZapuw9|F6!Mlq5j-0x{1cMMlvUjD-j
zN<;)yt`NEa?5;a#Xu0P;J0JgPg#M@Z({FizWC0Zo&2127)`6vIZec+YQDu<^RR+vB
zpSP`}TEA{3{gRxL5>DIo@ZQ^vQIG6w#+2rh-G!9vmcai$KU!$5fvwG(m2Zinr5Bf%
zY_4Z-E)MhDxk|PQg@-V}J8|te(EpyG>1H0lcujBbdplu7abaNt@P3$`os}DrnLqsV
znE&6qO)d)#l(w!nhPtNO+y!K`rq<SxwPHuxMl3L$b*ETL!hbz3HeOqMyGKzGM^kS<
z5WQUbfE-n18Ahe-f19;B=5W)~(?FmDTth)A^Z;Pd)-YPpdG$gBwAoPILgMdsPY{AP
z>WTIH_DwGE8HD&D7O=?8fza+o|Ne!4oUG~hJf1}`Q;?HCS})hv`#+4m1yq$=7d9#&
zA|c(Wq%?wb3rKf&NOw1E5$O_;?(XjHF6jng(@1yU7d@PFzW<&({yWwf4u`V$cwg3B
zb3XH#&zwu|N43SgS{<FVY3$=UAH3;SdOGL8$p<DVb^=?${G5h{=J0e|o{gP7t`zbA
z?tOG5Ac;IqQKIR!i$}>oAdp6r>tt+wfAhp=z)zb#3(xt-3pr4J14d=gM#YDRr`gN`
zDpC|cr&Bbraj5qH>s8XziEA4gNGK^q!M6`LEe}Bt_XiKfQM<>PzSMw!?r3r{XMfs$
za+*WHgdb<J;zIKC^RwA+eS;k+{lABdZer_SGw*;79}zPrpm%6xd)tIY(}36X5l1jT
z?;HOcBYV}O1Cm0bk0H2OHllP<YB5e;)9L>liGN=D->;S{>E*QK0>gWI=-o{Q48-#T
zGO?+REEh{C48X3ume^Plp6qHI^X=QWp9>18<>cj=U*F#yfWS=`OH3B8&y$~i<mpWP
zpXS_}3f+Ih^9!)XF5q??gnE#<`yLR*iybvwl>YvV%r2x=(;24c^)^MwE&@IqG9Yv-
zI@>+@zkNAiYq_*uiw#bne$ABO=4^_IR*_laZyEr3z=H7km#2@Tp|iPPGWa4A<d51h
zF)?wup1!oSv?R}Mr+YHF|2}xYzhZm04#qK|;Ns?wMgil0K0Vyub^j1LPs0SlnlgMY
z7a6d}Uh$Z*Ri>Cz$tvpOIqO&d`VBmL?g-bp5{86VB{aBf^u~o~GxDQH<ShHsi~pQq
zLgpz5cGJ4J>JPK|d8xHz@$_H7w~(N|+**Lw23<FyoaGwUM<gV`Q&3R&r)1RtSvsIH
zoMpCqliaoS7+~#dz7qf&9xVBqixFR-8C%bMQ4jR@gM40k`4xm4%HQmLn$69;`0Z4D
z#osmTq%Yp_pQru73$c9AKAP)yG`veEtkro-X0=Pir7rl7=a8ZGiW2wab87i)n$&@f
zv;?Tw3<qP=MwzLpsWUS(TLA}x-7v&6`R}79lK{!D{bHpzgrv#^Hu-ORf{QKW6L~-`
zbOD9PUhukVA&=2n^O#Jq#wBI^Wwp=fNCbY~#o3JP^4#8PYBYK#JN@%%nO6VjzDU?U
z(!7^r{;mpknmB-@!Et{f&!!!?TT*u*nwHt=O>*D={ANUT?()uFYT*?Aig2N@LuaP>
z5*LRP7<d_ph!5kAhz;|!sPJIE!J#kG$_<yK#|UR+2n&CT{}8<;e(ZJTe<jGAa=NRJ
zXl6DFVtRHf3KI&&0}~AOXZhm6$jN<Xt8z(t$Tps-WLa`qOpClJTib1++;mOiaQNaN
zVCMVoado`<<nDQOdwjKiGQg&jZu6Rq85)>CtT?>;LD^iOV52DH{G>16k%oy0O)~yH
zzm`^VH)XZuVj#fFD4k7k{ZB56pxkN|PLz(Co{1P+m(aE_uCX7iXPdi)3yOId6+sF`
z;vot~3ib>3bLaheqYHiF^`#f`weXV3=#0>%>4vf|r|Lx7I?DL%H#CY7ip7PYpR?)C
zJb4{u3qDVW2X3o^Z!H9J4Rd6d3m8bS4%5wD0)KU)-$~Z19bezjkO3%0B+_+Ip(Ec)
zybt&Sn(u7cTV1X`vbx2FLEuDO@Ne`=#CIXa<FaJg7*b0e8Q*s`n+IsG$HXCImy?a{
zH8*$D(p(ojo-g2|5x_iUQI7}(j%K+6pQz|cWWSC~XLJ_QFuL$rCmdhlGlf|=zJ(MU
z@89dQ4Ic^&91&JJL$gf07W*vw8%qOxd<%jQ$Ly=O_U!v5-!>yLqxm-JFiT5+)(G8B
z#V+HXsDJ8)__I2~pri@0h|`?4jAnab7acp>+LxR+9q=4Wy80&v*27?FX-Bi{uvApk
z7!uzB3&)cUY&W<WJMu0C_ztuns4yblkM8b`?@R1Tt-jJ0ICd;eEgAEs6k7ItRW>%x
zJ$(Qsmq$ol0Nx)wtqm&L&vR)Ee}vM#RtycofOpB}Q**>^+;DrtOs#l-T%H9zxHysT
zTzZT`O;mOqzO%A%(z@w20+j;e^9&3D=65+tMeYsig*sk`s4$;mqygqWFi-?>RMx*{
zdZkS6{DxOES-RTT#4CG0|B93di-*qrW^JBgna2R4Ti|hZ1o=Tjx1RZPp`(xNs?n<W
z8n$(FmJAFIcdog7`W8;9+yacqBU@Nph~sKogC&=aRXKi6?Nx?~P3?u_NXD?lyTpsq
z0}ls}jDq5I>+V9#2^#(&L4=mxLE*6=^qV>bWJeu`D;=XH+Pvv9sTx3DW%KzlO#Kk>
z>e|Kpv|XHa40FO_Sh}O{LtN#+H>A_Oyq$($hmYh=h)txRt$pP3^Y<%zQ`KdxoxG-M
z$Kay-rgeUHDzr8x!kq{`r4dQ!UV$B8wtkga6@qiip^gFC={?2DZeh6d+m@Di(7VPO
zZ$Y5H3p!ZdIar20f_6R@DLyv?ufsot0G6v>WeEMj>7MRly8A^Er$cd_D9OmbeVh?w
zI`KkAPgv%zi%WIe9LOTC!tCS&Is-HCCUJ8FMj%Q)h$gJsIA=5-c#rRUm8NFUc1BhC
z)~?G?EZ_b)GBJFzmMB_UZh7VkTemWUTdV>txT@YoeRWP~DDsLYqV!hO?^oQ*U--CC
zJvsax@z<<1FT+vtU6L-a(yVbT8&B$QkPhBBw+WKfIH_B)adJ*$@s^1H%OMMSEi7mV
zY6Qs;Hm!Albc`QR6uIz_02fLHAVUqI5AI386Y&wi4_-S0(Mf|^bFwj<y^urAD=JiF
z+yC(o5tKk|q^G6t)!*51JL8Vd6<N>@{U>rwA0d=j(qo-msERBB9)bWjE~6<d7p_N4
zrto=}m`pxJ_@j050Nba)M8$*Td&bq-wef~bHF=u-9O3H-6*zg?<9?`z&04QyV)<hT
z{>pe+#$0E_fswoCtKF;hyVbkZTkjL3E((rdm|&P}Y$Sj)3BRRL-ErmH)VjPLKiEBo
z63_~P{5SQGx(vAzzJh(+g=|2Acm4i5>T0d*Kd4hdGL@kTnLb_nEfV4lDT4Ri3*|li
z7uE{bS=NG0Iq}kthZZkyK5j5DK?MdXP}<xk5k49pIwkt!13WHYR3t_wI2&p?=8}^;
zI)o5c?1RRA+clv<?c-|O7dyLPKj1wk?tA%NHeXX<96E7W;~%o?;6JS&J#62Slf81c
zMAjQE*CU92EYi0=<1x`kqtvN=M7j%ZJ<$N{`1lt2)BrhjfsOcVu)PcN)f@{$yH>gQ
z-vBLj|G6M>qKH45`kz24xv1GlSd7k1|I!3&?MfdF3*Onqjz<-##w+HDl1#;XZYkWt
z_c5NU#101s@?hi<SgdRwtgfD&*n%`QelAvP+KeSY#zd{CsQO3wD5Uv5ZQkYmTX+V-
zoljiBK+RgsHQwIuulQPyX*lo*a#HC4P#X&KXsoWprS}uA&r1-Ls=McwaWPpikq1^N
zqanc3E-r!`KfJ)7M<Bq@ey`7THm@I4d9)T}6_En82tz^%cf`fF`vgmU<HOx`xb=8s
zhbZpq)jH<7KU|L~R23H9ro9Cz(=<pTg#tG9xEXp{S(jtMz)VVh?H&O!Tb|X44Tt-L
zblveez)$;0?e|2KJY*-5Ii4NB;o+$(ef|W^rSwc6TJXukcSAr0LG_I!uL9lk{VtkY
zffmM-jSjCoaC;xovjHpvPUxpcHso{i%S+m00fi6WCvd5esNWm+Q4zY@n#~|Q9Zv*>
z*A&DsKJ2z<--JG4trWy*;}}cZn>lvA=UwaUI}xqfP&!0E5iJf)3oH0^*lev~t6_Ua
z{`)b!#JatDso8sJlc!z2(tH>|0`>*#c(c!6yq0=smy0VtLvcW<Uh-uf)>Fa&JdYNv
zMo4X7VxoDTWE<G<(^4U{>m(10FJw0ruo0h{Dt-Y!tRn+Ca6!`}Y{nBYXKrZnqG8|h
zv(Ke@ViZW_$}b9ZiO+!ev=qn?*SPyTg4}Es*(f+EI*Wq6Gm0|!hft`;XN1!RY(PF-
zDV3$~1!A6Y#wI`+8XCT(qFR;yAJb--Vf%La7U{_5xN*2V9IbxO-`l4NK#MHDw1T1v
zVg~N2Dh0##AxcY;Cq3)_bs(`L_*OFn7aFiAm}lFXQK}2)Ub(5e3G6`5i$QolrW$Qa
z94v{D=nR-%mA~FQ4_$mPRmz{p@-4zk=Vwy#VT!&P@O;X20B%4CTBx@^(I?$q*7msy
ziy01PsFupbtGo<qO#SSY{^#^-IRhu6BU7(drWzrrhZV|KgT0E%+inJK2Kx01X}Z~3
z*>q5CQeU^coTfM6zeQIsfFAfkGXq#t+-zJdmlXmd<9OD6wdEdkm?5FV`+_xtzhU$<
zXj8G&)YNA&blQTRp3V0CS9X|s*RkK?Ik-mzAm9`p*<(bMM}s+|DRo}dqgkv{F!iY`
zV*k}JH-NDc7`TSSEL6pIFZE>Izd;$GJ>eLEu|3#aUGIpITR8k3>Uv$1k0$rmiJn4T
zND;ex!)$?~Cq$;n*K%T1P!3EF;CF?(M7p8=th+FnBO7*SIj)k5qhp_kkAUMXQmcSW
z;{?TQI8;|x19Q&Y_M}K%9G?st$XnD<YMpoyk%)M{5AUtdzA$^J%2F!shy>W@Ip|n~
zp72MOHI)$Bl6>jw$Oj?=c=P%B_WrJXLiJAZe-ft-%(sJ-g56OF1{5DQ{q#Pz;P@`)
z*{Y@39gq`e=Wj%65txs5MxBUeG_Cx?{$t>+U!U9B!J)W_0v`6B7@e*o#7WEPVTJoS
zeQa)H2f>`v=b)g1)0%Rcug0i8$Sx^aypU_on<v9It(%{1KHIb&g83|^5a~4QG83n1
zuMUZ@=RjX!ht>dIeWC@xkUI~ArO>}wuBX5UtHPny-W=H|#YNp##z;sFRnACEEzoa;
z_=t=jS5%1KtW){Xah?VSU8($QM3DPTu2I4}P>Ro*$FciZS{WIKee`A|@fXZAk+z=Y
z?Y1KTmmTq4S+-vGUcfZ-Vqy_a-(2x86q$30M{GKTazm|D^w(}Wq<~J|AtugHpVgFo
z0!m^NaecB9p?bF1k>IC%0|3bkhzrdg@S@U@nJ$`tMZJ*g6H&)d$m_R&NqIC<o1+1T
z9rf?1r%iB+uQNL%cSS+JIshuu5LT_wXqF=S1NP@0Fa|Gsfn{8O^zOigkdcHb>ZG7l
zx}9Ohs&sUicAVaaWQTjwz1#!Kjg2i@8AIx*rpqAw&|$s)ms-G<0FmOg{|533Fxltl
z<=dNAF(WW3Vg0;BI3`!M9+HX}!xs$<?{{<oa54~%^$yfG-`|sKmqTa#YwInoQ^_?L
z*~P`|zW8z+$<><c&_#z!hwIw|dg#$NV9BJcuAFR|)DpT#vX`jWt-;t|z%$H^P|UuA
z+bRI|?euK8R1?^@QTXQ`IA2LvLq+b)e^f&Jny1Wn&Lwkti;nJ4cGWP~Vmayi(qk=C
zG!#g0)^ms+!*?JwHW3by<gdFoVAmJNW5<~=uQNd}ry5(jb2q<_o_zV{$e%3D!D{h&
zieT>w3w#@7bKg}9ao@!j!@tTXqVb~d5crFo#>vQ0kr_^}Gu_Cp#z)`Q8>UD2a(UHI
z>wMQZu?)=IOc!Vf{dWIR2|aM*f1RTaK0_><^<Fpg<29D?)31ma;8Hvfvt&U$eZ#VR
z+Q4zjDRJ`Md(6?*{?0@{feAx&buHSYL08*0Lm<~dE5%EVqq<eQu17CLZ1Yp-7+!x#
zPSGdmckBeR8mYLy!zLtd$J|uA@y$T!ldsj`fTR?qM|P42j@2^ytAr^4+jH96@<zMg
zv97;D+(LN<gh+#|6jXTOw-Ik9s349UszB%C!cv&PZlB0&Py%sw1epJvZ%g<H=^T!@
zxIE&&Ce&C`Z91e;SEnhSOl32Lm*@ebY+WE(S!<{WhdSP-Y64^wKF%X<+2Mc6MXBfC
z*LWqxX!a?6FJ6<c4{>kO&my|EzRiU{y!v^PrUe^x<OM8#GM%-fcW1U#aOLos_jI@o
zFSUnm^9c=I9Gv?6O+0%8A~z%VY0ZDmW>2`6@5i@OKfNiR7cPc><F~-XRhJ~;UgFBz
zR^Y6?0VN05er<?IGdjUEp3e?Bo;MXTU=wmg1I19Uj|vPsf}MXl1KEXht8=e&?*-JI
zD{rIq6=ZouMV5=^VeM=zs~LaDaAe4tKG2aN;OxwnEtM1qTtb;Hox;;Sr+8j~2<)P=
z-%KO=qZfNq>ONxYUb#U6+c7wXAnD5DRp-MVDD#{t_K1Rj4e@_sFC(Hp@8Cxzjp+9+
zw&ivo7W&!P<~VTo)=p%%Y!|fOj8M?h4E+F70~M8cJ1-eJ<bLQ6butW37^<zuivo65
z#6N1&Avz@ShlMac%`d=V*63cS6>2nT?3ZMbxt{%cMJ+Dwt@<K(pw<>goen&(9*pxh
zr?JL{obv5+gH733<M2NbIOEQ}JewmrE)Eu$*t)i{K_$JtzV3s=WT*qQ!T{q5Z6q}V
zH{ZQ`Onhcu897n%d0k_O3`luTyVGYHW{JHjv-*TIzfpZQ|5ocU&-g>!V3+dA@}ndt
zfqdI<KjdKz7}9pMkHWvI6<%j;`ky4#45<e5*j=t{!Fv_qR#OYnhp_xv57c8&1L7Hw
zs(OHU-}chCS5=J2$Z(aIRq5cwL*pGZ(lK$OMQ(=he>ALr&o*Cu<ZEOnJ$c(Th#(}i
z8fFy4`gd?Ge+GW@d!;3Jb$t2ap2!Wtyt=xQ8`v08_22&VAsFCT;(q~G4eTPLrkfa8
zScFVWDEOx=yp_QK-dIs6q~{TH<fa2x9tM@!Ps?o^i>r&XU055g$A~F;>OylXA?rg6
zDSgxb3;4@~L8TLGD1Fi3?wlUS3Ku6SV3TJG?)WEJO(MH><-l`?##W4vCe{q%U9Y75
z2P}U!>4K0&*%xj6kMMnMW@P4<4AG>b=9^`pq8CV({XemzvwKm<y&cL#Yx7SC?_%IF
zp72%*1$wyvZhq3Z?{s^-%=6Bb+xSg_PVQp{GQrv72*mFf6xK{f<nEZ)Z|4_lKYdbR
zHDGCITNGi5ecH?>l-`kH;o~lS1Gm2-yX$y!yrsx+Q-XhNe!O(`kwq}MWuo*{CDRB%
zX8guze2egFhMVAS_Q}TKn;@B51Xh3tyYst1Zq!LfovX<pT3A?Eqi?~e(#q5jxt9uC
zo%{Sd_ea7ey1IYeV#P?vNVucQH8|YgUKh0a{iL3Z=pIVoCM@>`whBT^R>2cZJQEWg
z-MI1|kYBX|9d!t$%I4+&t_x!~wH8_#>F)Ja$?@X7xBN`58ISFeGRyxr?g2PS6V~x>
zk>5!)iT}j<S{8ccgU!&Z8yeCBz~vxHz53Biy%~AtDL?b<>0_pXuj!~>QA)45Bz+p;
zyDVhfcUjzAb`OT4ScAis3Q77=C9d36b937VS67HU2-o*S3kE~IK!-`U{J1JQhGwFL
z^`j1}`Tq0<H;yQwT@(;dxHVoZx#4P7)zk=!iNO^X7D{+|c@&ZXCOC;oBmJ1{^J)IY
z>89UG58n^VeMnT<C@PdJyT#I;rsE@Wi4%N`gBFnPJeTeaJPbVX?Kd4koAie?t8wK8
zx+!0EOP6&rlD=kZ!wFBg<zMOhSq@uUo2sO@1V8*)*~ky<50mIRw??<Dr<W;l*Qhnc
zxf>46%RXtcNZ_Pt&Co{wEAb^Ip25Z2260%M<7_3^(3Wv;Fux*f@KXXQ=1$apA#Sj)
z(B0!25|%e2Z$q|PZmU&e+lREZkEK(YB0Xi_12o<F1NVoRB?4S+-Q`gak=eB^SIW28
z9{ndRKmd@C4R>0yH2`lu5ytl;hM&zA>~C*_$BM`1py1xB$h)_q2v9J8cI9&#JXgax
z0==BJ$l}XwqA$z`CGSq@qJZnCUHC;BUXM^D)XujKjEJ6T=B+a6+|y{$`GEK=_&|I^
zHiMhdBSn>iXkbimv`LuvymHDL5~1Yjx?lva3nPFhx;Z>?vN^hT$!h^FBtaI>v}NM*
z!o$ia`*cQ_w_-V@`R*vB%-S-4Gb$nI^XVo|tsMzw+33aLnl9>?+}L{A)-1Z^P(nL>
zDv4V4+1IQngD5L8UEMc?Jrr+PI1rHzxZB5t`}%j;8kob2C-XJooo}l7GrAGUPT2F^
zC-VuM^&W8W=k3HeYTbB22np|N)ja4I1Jm1y@fYn-a}bjS)V1k@Q<&kZ&sh}B3ZpZE
zXRn3!#XEb*UtjlIm=y9zi9|fdPk0~EonNDkXTP=jiW6oYjvIWy)42mGQr>6X9AWyA
zE_nhM21a~z+*GKwwtH|uZ6|Z_DeX@exO6S!laoVTSX`7PrFB!cg@=dl_}SKd)AJ7h
zxdTb7er>(Yh9VFM(J)m&rrckYUr<9YxYc!xRqqgy$1;%pt5!$q^()gnQ?QTdeE_Yi
z%X<!2`}b-|Yf!r=Cjq8Bvw2~hg?luQgd++$%DVu60O0=y`9HyY{b2KjrnLC!E2}_X
zNi)C{2{`JA#GCIEY?tP8l(lE(loES&E73He_GixEWlj66CEtmXcGPWKF=|LT(n%3<
zgauI)|1{VK#{F@t)sa}=@>N<+m&*m1y3n5>JFH_46V~}>cII#_F48sKF<2|`zISw-
zFnNey+U{&-k}Hp&aXrQAvlcLYk3+}F1jFgM>^+mqc8Yfn_abZ?iRq-g0B3iO_fs_{
ztfy>@)yie_xbwN-*?sCzclQ)7)W=i)FbK_oCM#l`{hIZ(e)+~`hS)yqdPJ6ez*&B$
zJS5$IXDPTS(@a>eho{DDCHam38tBgm1g`lq`?Uz@w*WB^wxQu9sAZyza$!gW03Sw~
ztkH`Aggg~~0$MNFE3H3&n!mpCYr6L@C3}j(GO4bR(sW$>T1v_G5r5GYiU(g^J?;Od
zE0hoqPaXA-p(2S8c&@ynN>yNkqgo{zP_Ykq4_a0|ULOyl{Rid05Vhe@q*nRGY2MM5
z2hHH5AC)0fzqT-#=9el6xu&@5`&MW}4xTxIxr~a8KLxBPGOS%-k(UEha54`<lI!Kx
zcjODUd)RX;9>{Qeg;`03YdBX6WcT{&k`wx7GftfSQE!Tw!<3?4iGlV?CoO@S;iahc
z(V6AgA|c|1SrRPkwXg{BjevJ!z(C^nJc7&}W1_YYC5{3=zIv6K>SRsE0L5xTM=l6}
zuM+zC%ZI7Azdu3?mXH<i#E>>sp6t2n>6X_(jH37;6{2RT4#YiBuvc7_Scv*}zKaPR
zeV`f|HswyL*|so-Pw2s=15hhRp!8n_mp$qzftQ@zY=oCU?+63o@!jVGpaX4-ji^_m
zSI|@<|6lkH_kakG3r;iihI;KPGb@Un<Kp78j}Ol4dk#4olGYUapKL$vS^=2<xO7r&
ze)f4wbS_8#H^kd1iORwNCeR^?ix(5h1`!QS?c@nDE}Snriccepvq*W`C2A@84vY-+
zr0ogYti;|LREXH_UYW9<5>7y!hn|a(gnzX^IEVVW$K};|6PEcvhx(0tkdT(vL~TRq
zd8goFSMn>EDDkvF4~t)gqm->!OU+u=78_rDr(R{sXJ#4xyiao0VD~5zJYyRzen)&o
z*Q%Ve)(~G`mwr7nx)|7Rt@+N~c1X#W6VDL^rp9-25c?g(Zm;<042CP8YE!94PPMax
zJ@0$7m29!^G>~#B+&yJYIW~KV`NWaAhBFko6oW%U%)WjizBdlD9P6NmabYG^oZ@Uk
z;$%Naa;n4E_dJ$83(>9zI?FHBmG*p_igr9E1y$|$9nH;Y8k?$sk|-~TklXEi;!0y7
z{9mLVA1!RAK~9@lGnQOMU`yedV+2))%~OT(5#7b}>3Y6F9U=X2i#}hjI#6nT+Mb^@
z_$-vFSE3^dR;yP!W-~H*FDbL%T93MbAwFH*UDIvl{dF%SJ`q0AoD$ahN!zA3kjRwL
zEfXD|z}tZJiH}d#Tr~fiQ@XGWi>6h>X$}4^Nh97W6S1;tx!BW9R4v{oBv(~qA~2el
zV-<%n6G6n=L1cEmy28!=X80jW3Y*!&I1ZymP%lB{g@VE^Y69_Wv9yl#%v?;<b8CK2
z2)d+TzL{che<ww_JL{43r``9Wc`T7DU$2%z{H*ufN$XFKBk_Ht6<I&k3`-1jhAv;M
z?hm=4%C<hMn^RXc)XP%|-*P#+#4rJ2IT=aBacYS*(=@%qDbCLb?Ew<5x#!0xCpJ5S
zqHF5|1xkY{70BZ~Uy%z?|Iwrf0{V0mH2Lxf3o@wzO-h*$j~OBIqc1WQ)j6u~QmHq~
zvv;KCq;@_%{-JdxDvMH3*stQQ>6n4k6IR`?`-EJM@7gk#h~)I-zTNZPM|Ayx(Pk#(
zIYc=gheLX##ZjA)EaysxySr@;7<I*j)05!q{3o3C5d{X-GUw+Jw~Ln1|5@xgn;sq}
z^l-bF41=Xmmo@O(-OMQ6tI-@gvP_nIw?1^_s*EE=nn<u@gJ9G_C$e{Xun0?%1D%%U
z&|!N+fJfjr#Jk@bv=AeAR{QgUJWW;HDdZF5ctxq*9?wD@3N`P5;}kZ#!&?sM%dP4K
z6BO$^J@oTJ-q)H1j24B*A=C~(kD78&ns2_IYP%t91)@Yk>6$1qVc>PiB`1LH1=PMc
z4hFA`T@}>FeW_CEcIvoLCnFA@VYwPQvNqdSqf_lZK`_6us8N`dF;6dFrWSZ#0Nd3=
z{u8UJ-`dK*-`aw^z?IV(A0UWUA)WLr!Anr=Hq!$Yucq$Jf&zjRns3sn5_1uiU2oPR
zr4O$*Q<U$&a(#hR15^V&y1xlNpla}k;9HN=ywlK~o{(}N>q?oUx`zkFh%@~Vzyz}9
zf9K617afg>TwGi?GB`wk<pos?k>g?V8w(8q{(1RVnaWto?`@P5q<4gwX5pRy^jjGd
zuZ2P`!DP9i-Nu%eJ_FW3KwktoBelVkQzdcSICwQA80yJpy>a1dg2Vm*gm5Lq>ic^5
zGFh95$*=4gSzl`dtAi^g%x@fG7~70jcR!8R#aQ)q1b(&?rq_GeAnAYL(Gz>^iUgNC
z!I(p=d(b-!X0y9&v?nY}+^KXTZ!Zv7RFEnh(q|L0Rd#{(!B~Xfc0emrYQb8(xe^Gb
zD0rc?j*5YSDLM%)O~^*Qr>yjWByz<b*JL38C|~lwrsjmt8A+DguluXi#)03bi4F<W
zcR%p)eO5x2#gJ0)1Ap|2f7E_SzCFf&w0|+!CkFCP8IJ<5)(H%wflvBRl;(hx&PXOc
zu>o#RXqEC}yFFUjNt8MW3wg>!LJ}>0@O(ZIZE7(wfBjS(;DGJa1O*xe05p7=(Q9`l
zoQuZ&6Dyu>FEqETdlr+`;LnzGEw|&aVVv6`pH;5qNP`207Aq*pN0Odl)sU0T#Uw>}
zTeSQVlu*zJ@X%UoZr$50Ll!UUUY6vOQYyAL7eBz<nyP{4YFfkih(14n<x=ag7utPq
z;W00A)4*&6SG5?`6JV;>bDk(Dc5Ii)E#6V9n(;%<&s1e+50e6qhaFjOR{l(qbkO3#
zqQ+KKJf(Oq>tudLP5GLT+a>lZE$^LIPFRX)PoNpJB=&Q7@gT+#GK7PL+Mun18bA|>
zlAMz%dIQ(MLZF`UY5LrFs=cTk&Y|Z#`NJh$9b=+Ov|X)8Yra>%wU3X`%Vh_Ait`jN
z0Ok6B{h;oaAU5w0CDEovDx}4O;gvAG|5IofzH)a06H54Oyifd&@Y?af-{uX8u7!O0
z^$L_Mj%D8MfmvBo*;~Qgid0X!GW`;jrD~vp0c#Ub>$<@~*x5;PB^syzT_%JlZNI-A
zgs3@}#xI~amaMY!@fT?Qv<mg6X8VVlTDr>VP<dimk$g@aK-J1vK~9sDfJecCR~{L5
z^zL`t7iUwL0!@cN$9SpvEMv&V&CLxMxEI!Uj<#AI>5&}tWA%{)kR5{hQ?mC<u)=ak
zXS~xls#s-PjJZ_xD&$YD1qbnm@K0Lhmoh~p)O5*QVEq-zfN+jyJ`>pu5bvIp__7Ci
z8<Bf%ECLziEE?#;Z!vu$JCccSPbmIC7oaK>tA;Nk-Pg7U50sk^oOF#RI)DmD3<<s7
zeMC(ZQ_JUl?$OS^xB&vZw&1OJ74oy`q_L)`$Q3Q`pal7VI4wsdmg>85_;<2SQGM3i
zc{psfXD<mUNTaUkzd(>&>$LiN8vpnqCzd~3bA(K|q|D8_0~07_S`xbB$K%zcV@bE9
znEx=GzGP`rG`$h*^+25@9;_~VW`5_CZ?Mws6a3-AQ#0TGuCb8|f5}rp?2GB{t`DX~
zLdR%d8CzS4KIk#lUQH-cdPyoYQ(FNomv{o*O;ocn>$Va!k@dT^i4(8MI-uM7#Yk5a
zN8XXp!q0=|$;jXXZ`hrG@x>+t6)4;{))X&xUmvWk3v`i3G}T|Ai5Dx~6OhuEJoBsZ
zTM-%_&@+=}<6v(clz3mfb_ImMtmn)tB&aOZ)r8NSq>qF(=H?yD4@7AL{D=bzkMCrh
z6nIN=-?G}8A{XMme`~V@Q-*XjFvi?Z*H~tZsa9=KY{NUz<1ONAHd#!DXoVDmDmyUO
z8B4%yft{GZ@rIg7cut$NJ*w);;lMw9+_$Yc+#;%5z@f?6{j3}BPGX*(><ouaD>{Y{
zYrdb1H?@K2p*L-T)nAT2@T;zqlM^YK$*?ydK)go+T@7ec*>UWPd)WeZ)dTSX$dyxl
zp}#aEw1?x_x;Vo)!Ms01dzjhy_*W7Q@OCnaUY*ndy`teQTYu8%0-ny|9rm-U+NVT1
zi5vgtq|b?{N3*v1-#s+-@C(h~{7;PlM`WP?vXPewphWyqT7rV6*Vl=-7r1}a$xlZ3
zT&_3VdtA{`by(a9MIo@R(}p8ZP<7-*BnisRPNe`0Z8*ZFFX7#t&CXAD0}LkFy1}V7
z#916wXY_~6PH)nP0YTf4+~(|j*D3s07Pw-HOA#r(iW8mWsH^1nCiooJH3i{o@trVX
z$n~Zb3gHsw6}*X|)jSs8`-IZvlxC!4&wf>&){C%=Pd&sF)}-Sur0uQLTK}?zrN&}4
zV&mXUFHPkWV^oliFk9%9RU@o+xS3*L8pjt-M~&SInh|7N@Xr%4BL(n*c9CWs1ih^`
z@ns8nhCs}@qS!SR&GU9g!YHc2zA4Mt<LKNEDh;Al^tLC~#%?M&Qz1YOq!Q{-uo~!2
z)+98hCTbS$tkR@kyO0<&Q)h@oOuziy{YCH=H5p&Lc${DUZqUuqiE;~d&rm{;QA{{O
z4la^_HeT>$1CkBXYPEeLwz%X%eClg@qA;{%kqf0V2ahZ!ErMArm>M7I3*}((@U=MP
z95q}G{Bb8^tLLFaFU7?Ns6vABNfoC~PkK+bmv2yegw`WNs5wLFwRpmeD35Or4k0oQ
zw_&t*cwje@Blr{)&~I;VD@}n$ooBRam8s~^rV@`z(~o^<kHq`4sT8%apHXdMfzcjn
z?fp}5h71BgI`sb%(BZ%^rs%Ksy_&=l7#+wM9nHuXmEw*>Dij{10I<RG=jhUE`5_7n
zB1@``m|M;J(n!MN7S_l4n}3ojfd4zW;or%SSO}3B+P!5J>d3AFs?3N<3t?p150dOw
zd&?J!GqW=vsGJlPI#XWkTkm~|*00T$6kC}7HE2$$D@EFoXGLsd_#+lV&Nd{j4kr*U
z=-cJbcjuN|uNA+bC3t5voywDDp5e{md3<s0NDEK0zfW)fz*T#L?8b^OqZEranJoyJ
zRHc$#i3=_$&3(DT83uN?Nl#F5v_HlJudMB4a#>kA(m56{7D(_#6my%sFD$~Fn5!M8
zf7=;OAZ&AeSceKLvJ;xV>~aO|UymOGc9W(?d*O4E$=Z`)kg$Qf$5|~c4urDy@_6ux
z#0>PWWp@>|>Z(d@gMDdB%Rc>*`1+CcGkRWnB4J0uj%|UgLg5SHPgoS&+NZ0RjWI6%
z^|tUm7s*s6cin6zeamBa=;G(Td5SA|Z}JbBOY^6kLqeMULU_#;$+IOvMyQ<;O<6Y-
z$yg!!wQ$|uda2A-lL%gDFWL$`#ADS}MD~XI`@_4zS&_%Kp2Tk^r<O4D`bM$>-T1*U
zeFLX~nT_ehnY<&ND+DY4Y)%4K?l2G2=+2d<N=5g|PKn3P%|y!NgmKr-E~jhniWg;Q
zS8)FD85WP6pJ5Ns9Pl{mx=_6a5x7UgFBfwf7>S(7_J&@4I0TZJ4B(OCBi39VYwXVs
z)z|!TpV#SE%TrU-R@$Y(KOr7K#Kio3Kcy}#M)g=*gZ{5N5w6krg8)hP)f)<_Wa)Z0
z_O=R=&B$xe^vs7R$Z3z_&t*fz9tSum5`7dkPo0I6`ft3_A2q%tK3|{ghbNf3Vc1T?
z6;-=FCu___s9BHUU2_5Wi|pn*T#ZW7jsOomV#Pk|rWa?erPIEtULyF^>nILa<Se(s
zRruL#HMxY$I?hA}sG;@V*Z0Gc2|CxDXe%4rN<}t~2U3ldw*~3Ai>ae`?~G5M#j5v;
za<5*G0J(2;%0sL7%=5el4=6X?t<E>4$Ttd10fmL&PwC*&Bi{VXoK#rvQPx~6nj)I2
zbJlHHM#tM`^*D{K=?)mMVd&S8B*}ogFtMnWP|>?BN~#V?ycY$g<Q)NGIt{G&e4Vqq
zTUVw|3KZMBE@B+jw~Yb@1QU1y6bN60aPEp>LGQFC(FVVeBYap+D`_nczDYP`faRR}
zrKrRkSt#%kxY<iL;K<s31lGWycBkIo(=gjWZ?oQ`U(DRXqcIaMATU$*Y*U=H<L>mK
zNU2$I=xzYNs4sj6%>V1x_9TAE5YDS<!(v=8&6x=m%fqd3K)#ip$r5bE#||yhifGqn
z@dIAR4G|M}2AqJt73QM_Tl~p;HV&7K=ZqC=fq8!kUFs!3b<)s~u&}Ula*0429l2z@
zAU)hfE>I#i{T@jHB=8cdI#<c-m&QZFs9ru%&}tWEui(YuL*V$}GRmpCjvQ1!^UP7h
z+PYi2A71IW>$vN?Pbk?;sVo!85oNA!@8!r{B>qZ7u3<zBtq%Rjx$cf$kAI=udm<`M
z4R4P-fg3&ME+>0XP$<^nu2~H8k!TMikDZ6@FMgeo$5g#6)pD7La9LdFLnTnKF<<*?
z$+@bdYXua=Uo}kYt?#VMfw~w!L*2ZhlbVw)(#z|`kyBT!=rq-I*BnL}RZcP6v@#a;
zwNt8AUT7&C#5cb(@MFXoo?L_;UyAe39@2=aL)y7>yQe7<T#XdzjAL|yOiaWxWAi`B
ziJp512{TM{xQQNP%kRGwl^$x{Kf8(Q6I#8FT*8JiQkFfEs*3+~3^tlM1v{GLiIly<
zVMQi3k=i4oQ>^z??P(bt>PQZ`u~|Q5-gYIPHEqoOV)RLK|FSkCdV-9~Nz&znx$<CD
zvwbtS(msb6r<n7j1Rfl<_<(7Z_M)(uGwu(~-MeuEcM1ybj%*?^)vz7yqVp(L9<AoV
zj^s1oYlHh;7Y1cZ%g``j212!gG4oM<fe~j5(?6OcodP_GFrrq_r{p4j%BV0S_XWp~
zxCnu307*>_Vp3#&QGld|okp)Q0JIAnZh31HWB}^4sAZ5ph$R{KLdb*5|3oU58+Y->
zankkS_yo?{SlBth)MtD->G^UoLDf>q7r7fGLZ)Zl-HlY-YFS_np}Wo9>%05bn@{c+
zn@e|W3ylP{U3f4zA7{!<S$|oq!xb+8j_2-CgLJ=yR~-y!kbwD$|7ehQviA~zbM}9#
z3tNz#q?ddyTUE^eDsL7zGD5NC|GT_-XadJdKTPrJ=Tuq1Q@W?{TIQR!|L%&@8&sIi
zDbqH__way7T+`5zoO}e&J9yAOuZ4i~vb`WH=A*6PHvPbrTOa+%*`fexhsC{%pM98+
zLG8Ac^w)|6Aljuk{gl!#tOQ0UE;6`YGf?u&sXt_{u~*&Qyg2J-$Xku~k-Cxg`~-Qc
zy`9O7lS@1TP_(yYhDhlomf{u|eL|5-v9IabRjKNvNZl-MbrTqI=&Tp`{5uKH)`|m*
zmE)#~dMI=|K-E80<pL+FaSo(>Rp=_ych~h>9?k8$xV-5JSN+BF&|BF=gbP!r;O|r&
zCP39Ry9vceD18mlByaxPMNV6WK?N1>L!UHYCPe=H87~yB+ygTU%CAo@bx~Q|wa4aa
z5)5sXx53&#5?U-1<-jZ3wqNx|ZuD8#CG4`t2Q^Pc^K|05-n@?Qeb!M&v@~x!^8}1+
ztkq+W7jqtdC?m7+lM)mZ=kV|lb^@bB0X_t<SA%&>vIym_`yFswk&C>9n-QwUV&cH5
z(sGM|x-=!MjM{yz^V!hYQq{YqN-|cZ?#Lc9a-{K*(xou78&3@MHIiFF?BnC(%vh#B
z+UE8vUVvJ8_E<OloX|-QXrr#_{}y){Lp8iG9<K>eh)XY&Ji_X8*KBy^vDPxP0cdA_
zkhgAnkHCPTz!`asyj)Eu=#&?0^qlV89w>G1<sK+}CXH9TtuRS!iU8$Zwbf^-64$6}
zC>J3itRfH_l42KLb8*w>mv?LZD!<7_xZFqSMB8QSrAosA5|R~97q|h4Yuy7__jn<0
zS~aFaa9{UtHFrvylDJ=-5~zfz5+hA2d->bhwnsj~*vcA;b|C;6S<I@EMa~P3#j4m=
z7Shtw6Wy>7bozi}vNS(R;lPErDo!a~-^wv+97six4@iu(C%*!&u8z<+`$*az;23C@
zkdR=OC{4m&@WoTHo<9q*7RLr}_r6K=x`ePOUUzOGz<5ldh8Lg1;}ZYef`^n~ez=5|
zyoDX2S}$*6Rgz{gFFchj`gPH&(l_nT-t!%qrn+NiOV~*2Zu3;huTmXE>)YEu^eHXW
zVAUO$&jS?$ozqj4QY(GsY`eHV>#R^2BBZ%)@oo6r9*CBGo>&Gqgu<@lQUh~l?xlRL
zQfGi8Nv&ud;vDsFP;UIyWQ13EH;tJTwjK%~h&wo->&WxY#Q{#H_2m=gWgs2zWD@#u
zKO{KS=ojSQcq5m{D@K(b#Z9a@`E~V=tF+irtB-f@vk{BqpHyuY^~}O(NW&DnpOa($
zl0L|!Mnl^&X|Hy_`;XeUT$NeO{-O0nu7N>vGtYD4VNH5P97|M<1?I>VmWN7BRUFH)
zl<o}m6gmSIWzvql#kUbFbfPyBF7l)`4s>=PARq9&4?LfCBrbbjkh|rudrh!D&_Vtp
z{h?I3lDs`2NzkD76@d5@eerl-dJl*m7wq4$lcCPB5PX*+K}`YJes#<C@ar?Z^!?Qu
zjNQx!QsqNxmy~v(j6fuuHSN|s2bBSQg(+!=0oO=tc1jJ&{**_a+S1Q=P`Z|Mv|m-T
zLZ+x#j@%PxK)AXMYdf}jC);xnV1bV@2`?$m8;D4#I_|Q$0B_wGs<cG=em~&a`Q<k%
zc0@l#r~<}=Bv<?&2DcNZ&IMzstNeh(YZy{hHIB_u%rgvmJ_6~%EA38r%bm*!P%`f*
z11ZPZ?nuF-M*20w?!2msn{CkqS_8SGG0u;%m#$rOPhwC&kb)-CnCxFEY$68*$qBuB
z*y9J>1K?o29@{eV;h4vTXSTNsanH4eMz&3S^!Xk}k;1FB-8m|)Q9rOSCUBTG1=QE+
zd5GM3u!&50d6t^u{lEMDC#zP^){v?krrcjVsMZ*)6&D42RbPpa-F&UV2_i`FlvDIg
zI65a~Yv3%HGkf80AT3T&;IOx-njB+X*Fq1t?GP9nEm|C@I?(o&1jewo9*_PDuEqzI
z{FfA{KAq%L?XTP{Ub-Uwf`ErVVZ38vDlfarvG5NFjSNK-*F#LferYEjgnCz!cY~E3
zT{^WpNjN5#{R&rQ=RM%}I>COcq;CQ1Xq4`{PnVn4t<tX@1^9z8u__5ugPxSyr5&pe
zS@{Mf^sxbs&3e!{v|yE<O-@b{JUq~xuL`qSc;VY<s3J=3v-7C!G=tly=6mj{uN8*0
z3k-WjqSg*hO6M(|9{4i?YUEX3QCU(vb*k)ar#-TXoUXE0NR7oM_HDV80=m081vnI9
zSNUO3HeYHUZX8VFCdMvm%I{hn_c>TF=woOiR;X}!qqsr9BiVuoday>cySe;5t5CHO
zY}~b)JfV?@xWBJT@hQtM(mrVc{?h;dl>eXwbn@j=K58PqlRheKV>9`j_1S#g-wDoM
zOmV9NnuNzV{)GDT$-hNK__$&1Q~rN60BnupA`Yu<L<UN~8rVV8j^XXf4s(QfR8jnH
zmI@0Fn|p40b@zsZ48BIH;6IcQVi()9Sd#I%WWrIAiJD4dEDvt%fYJV<r21fSZPH-h
zBD6YkT0<cowK5SY4*1^elxj-q4#}h&+<pm5sMirIKq~tAq-U+)8e>m;_Jq1yUKa%!
z4aa*C4@pFEsnhV2k+-b8U)nr%lCGhpmLqDq@#X**Wx#%W0QFcI;1u;%6U&T^$4eK$
z`5eg|jz9udc!v8nzvdM^XF&c$EwZT0!&Qs+K%?|Z)-bp0oHaX6aUgj2G_7K$VmtM_
zG4ozgfVHh7%}E%e{ZTI{&QBBX9Go71W@gvj4&eW)VRfa4eDIry70<T@nm|Ymym)t}
z8h@%Gs8yF2TvCDn$M@rF+gL<RJ=^1cM>yTzNE>e##UUuq-WT*hyknvMz0-<4Z^EcK
zXSHLFKJZ=#*hEAeBN)*oDs?CQX(D_Eb(tA12^@HIAhd#29(}0HE3Z^~Z|k*glx>92
zA^eruE_nx*u0pjV?aPMSZ47RGh{wB0fA<CaZfN^)?dg0l#;wmi5Ya9CSC<5{q1iRa
z__XnLrY6w(B^l3|F3BpL#51_*S8;0{O|-gu5CaA-ns7h9T07}ud)?#3X>{<08)#Sg
z59P-YH`i#99qaIJrDA8kUPO2^x3X_Mg0g~&KTvU4`FTzvcg9%8?C99}_1oy-`H8Q}
zsC>-DJE5@{wx!u3Si@#cXv<_vMJeK-kmjIBn341(ZS~5tqb~&s73E`zX`~(NYmKev
zlMU8jIdrOoO68?X9+xeeb4+?P>O$}3q%VfamK=)|7Mmfm73b$iUn&xqXlQ9OPZTLs
z3s`{K;U~1io_cP8B@?^~yJ@w7I@<B~j1p7RXH2vlscP<0t@gW+jJe2|!=>1yOXfeW
z(5E~{_)C>~59;m31dt5bV-X1AcM&KMa#y-Y_^@AReRS`-#=EMCIWRC#$I!6*e=F6k
zY%+aphg?VOt0jHnr~ZG*ISN<JB_EK)c}fZ2rQ~izCcOX@_}gHqu6ohHlE_QWOb3pw
zwxd6lQa}fM0Q2I=Gxqq9eP2uIzX%=(g7-Qv@qhg)NGy~~(Y~LWis|omT=K7&&RVDL
zBB$yYayAM)@3-c)?Aq&9F=@%OqY#Ul?yql=10vv!jTBHhzP-!pw~pd^2=K_UcHWd-
zSFu=Mg#2U_EO??cM8&KnWII|yJ+&gV55KJuxe+W72C<m`9YI+}<!fn@@SBLG&W)eO
zVeCZDJ2DiRw%v(w%tyMR#>@yBDm%W{q3N2L70J6<X*KyQw=BwRc%b%a`#F9UF9~{H
z{cc=XfvQi;LxWe$`OA+3mN!?TaO8uyBuORY-Ay8lH<ClHtBE#S?5U=G)j~c(Y;ygs
z=P+%DG|dk=6NQ}iCVnP%6zy|UYB`>QnoT~;rc>j~abL1c2Qmp!qE-Tn+=Adh<HFxn
zSELsf=2`VhKkZp^fyn8rX!Ef`<5=+TOYD7p>guE}v{&f6cP}1pf@~LBZl+oL3lWs#
zj#3F>>~VOR*#RSK<896jxNNt{^uQ0ulBE~!RaEQ3{KMH4JrKPw_W{5XEZ}Ty`2Hv$
zKQ9yW3X%zm3OBgcwKp|!QfrotMVY=YJI4Fd_=OJnr*Dd*$9tZ_3M`a;YtWxw7IZi{
zSd#x+14;ZhN7K+gp5x5ZCrxsd$HF_gk<!ZaSAz<+Grm7UOGqMui`FJe`Ab?}hJ6>!
z-V_QFx|QOI<!Y^^!@ZvcN@$YrSZSN+XhHOKp>?`08j}$S6wzN5EDU>x2fII3=`)9`
zOq@jxcQqKFv(igzc`Ae>D39Jpu|EJsEIH;{dA||Lgtw>z-6Qz4ioWtqM>{N9cI9?u
zXEnUYE~;AfV<inp<V`$?EY+)FZk^dTcM_7Gp&`5$J$JD%U7I`pq9tL<$<%7JX_BOS
zfDtk#Ccv4UIsB~BF0m(05)t`6|Mq|$^$@*Jnwa105w@0>7b@If+yNI##eaXDC$fQN
zp-6;OU#UEQUw$Sfs1-W3BadS>>x#GE)7#hYxri~v65;2#OU~hv$YuYL)}*Yda<Bw|
zQS%-WLmbb5>#OLi>DaXdhi|Ij!)NL&k;*G<GaXqREol_#&6oSsuN%txgXzm^wyd`G
z*}`9{%{7#R`MD9d7(J=n?&Gm}qi~2ClP!-*%E4hbyZ4@{^x(6tJ_db#Ly@7s6f!bF
zShpzvI7<Wi^M_^j(dU7AR$Ec6$J_%)qw!&xEUGH1tw!IF25ASg;P?W0MdcLy!3W1J
zlG|^T5!U{<M!a|6Cw;&xq-UT0FG)-wy+N>@+rPOM!VM^Rc)PmQNIkS=$U`UB39$3a
zrgMBdmH}<CoqFc7GQJo?maVNZWD?HzA0DP0jVFG5DAqT44rYH(tZ2emm;~jSG6pwX
zoqxZ+hOwH_6mWUe&J<vN_C5OLxf5gjJrfTGJa(V+#@*W8_vfa%T)cG&!IyXd#QTr_
zCGoq*!5lyMJ9JF)XydX^8%(z8{jOX^A@0smL>cfr7pR!)RN<YTw{YSV@6(xTBKKLV
z&K!bS^4g<jvRlX^R_f6YGG<!<dUmF1<=HSnL#_d6=qPRXO1s%P0BXM7dgoAeUDUN{
z4vK^vrfo`!l;32Y>b%8zHR5|_)=c*>wpX)P)A|S>bHqwz*a46R;V-(O!qoUtIw6ZX
zaKoI+kT-CPO0?I5tUiuq<c#sWImu3wcM7-8Pe#abxljCMZhoa0B{&6AAqq}J2g<7O
z1~;L<sNzzYKav_VmyXrl1r}hgVd*SSb+EY{27kw9yOPe=Z_dBYzc^?Q7yu~dbSs?_
zJ+sAq?b)ftD2gAJdl~DHWIe)Cj#BJX$14Wv1yb}pOlSetph88m_9|m=0PLHLAB|DM
z0h<NsHzoJ=B9*uMDf#`@2X`&3vWz%V7cBj5{dd5S=T#gm^#d=6T;tm<y;RS8M#5pS
zFzKNag<E&!?wHMK!e!@&e$*caM_6w3nBk+WncbW-tnJZ?4Z2ol+UI*~)PPAY0m++`
zmDMXeplY>M@UNz6({y2Bq}po7$?1)9pmzj75syHS?*O_J3ud$@UG|&u>nexYLIV{x
z*R-XYhWuAF|CaW663Q8{bvcdR=BYu;mEjVD@@+A8Y~2f6vc+!mcbu{MY(RKxzBR5K
z9o1fSZDA)r3i4trgrtzs`>YEnYQ6*|`IKwV(!-Y-r>HS5tVO@v-_bw3D;}`v*=zy^
z7SxY7WhbB9HbO42p}#tkH!s4Z5vb4sR1XG*MtaAM?|OSBVy<_m%h2Xrs}!3p%4KSg
zFyWM?zsLd_#@YrEO2$B1@qC0FG2evqXq9uondMM@&YO8y7njA6ihA^H^@%OHrN|VB
ziQUb#eo=cnOtln`Nqm9tE3wL&axgOqH~hTv!B-VRd2`}2OLu@-+uo)iGf}i}Zciun
zHz0Q^T+Gn0oUAc}@i*Yx!<(fH11Jno;e3#IPO;eZHvUh)0K4wixvA`h^PqK;={vW`
z;{{vPB&ss;O8UY<eYT++c`eXbijhtSlDZ9m#1Viz=0zCc>${BXY`89?Y$M&>og}$+
z7N4_P03p#=V?Go$o(GUtP#cwqiy8)IS$Nbtev7Wbm<j}+^p{qbRZ(8tIuU;=tNvak
z0;g~Q+7@Jx#VPjs?)=s&{m!OlU-x8$^V{{Wy3=Jr;)U<5-5DlK({MfVE_c%szGwrP
z)f2V6Q3OaR2IL(kEkr=FAsLbq*bSpyvTz~=c)R|Qk)E^+m^V~affo0PKCwLHf_q6x
z50L=tM+$^HoUuhc^~dgRWR*tCchUQ5R8?JbOSl9u_7g#lWhj5!R;vs&suejgC7q~g
zb4d<8am+_tQ)PS8i@&)Xn^IqGog?@1!u22MM(ewCF_M{KG&1mD5(R+Ct1lU^EjG?^
za%=br5_WDK7E|D(*_^K*yZcdbLrvCU5iJRb&E$cqsCX7=c;JJ3<-r?IE^Q4QOVfhY
zerr=`e1zHuT(HE*KGM9hY2wlSgqr(A{9@|nQ9*HX*jKy&XMB>8UGqlwm6lPUC8hwl
z%X4dZIqRNdhEu1vz>91NRh+`;yJySVE?ML{Lf>Yf$d@aAfQZ;|lNVF;g%SibufnRp
z#qFZ&KtZA3x}lRXPdJFWBWkXtGK_8$t4}DwbFNr|Ziguo>!T!2;EPvy#-vz^T`i|q
zcxtSIDQ~!sb`G0M9ZIQD=q5+#Q5QmqxV*tPQk0Pu0D&zoZdSehqrUT0>6RFEeyVsb
zcfqMM^Ytmw5{!}@#>b((qYhH@t23m|Wyx<^=%UtWo54Ow-ieo4u{=R*q(X}IOWu$S
zN;HvbiK&WloARXIca;*M(007>A&#CFW}V91IAWx8dCTGsW@nyJo;O%i6o|=fWpY#U
z8_mAeT~j+><_cjD4&*hC<f?Mog;lKX&XBl(4CnejNT1DN7eH^{MSKCut8;3y0og%F
z%T<}QUDCtu%lDT#76pTMbxy(xzBj5axOt2k4H13Tt6gIj3IH9Ij6_LU&GskGdY;{A
z10uB^GXf6WqpW&xP>&g$!p7AEMN4l&xPP6SuxUihZ&E6{90o-j7*=<x1+X2ipr9E(
zz-guY#Y#<ThgI#*;Ytl9p5_IZ|5F8Hcnb)qGz5vll&ab7J>MFhCg{p|<Vt+9+g}vy
zQ~)~qB>BwVAFa8UGx{#n>_5*W<7Zurq?_qme8Jm1g{mcq`+fd%Vk7Z_)NmJ(tci*p
z88i<4T0ftjI?o*1!G#TNnVRY1S+-$urX^#cCW#`W#?kh=ft(PdM)|3>lC%Xa#UJsc
z4kTn<TwSe=%71llRzPUFrPm3cFqWqb^s8-xrthZ_uMddx%`GnZ?;|u6d=svg6l&`^
zV`?AtaK_JY`>gK`?lA<G24ZH&s5%ImUMD%#16ch09wUjDf*Nn$bD4~mlaeJ|;r)6D
zthJ4qT%L`oiGdejc%~LLxl}kd`?%G2Ky@sayqGM%6JT9%2<|vr7c25Azgp|+LPVfO
z+E7qoDi3SOlWzsw3F%FD&VKs2+#r6WA$wN>6ZkDf`!kd}68O?<yx*u;c)9o`E-$te
zT_^(+Q_TOv*IP!#*)Cnefj}TgaOq&dEqHJZ?hquny99T4cXubayA#|sxYM}1%Xf30
zbI!ao>znz*TCAoQ(AQm8)!wyh*J9!=BLOR|(B?on*y%wzBXUs80d4ZP0Cl*B85w!F
zpW^##g9wRwfh8vzDk{PG6EwEQXINF|hMq7-vOj|=Bd?|oxS@xFX)S~+@<R5k@%C7D
zSC&YQGOU0PIp9JLD;6wcO@2UI4wz}_a$~-<aZ;}QeT>h{41BNFZS~YyBsc53zO|Xx
z%B$PD+K5g=8WmrQ`7vDc^D}StVWBFcx=dJnLLSCzJw;40MW}(BW7%3k>yBAXUzWtU
zbi55ZxO1^STe#z7CAi=5^-0Y6<o0Dv9nin$EWVY61A|uEkpvojOVj-_RVOW@0FG2P
zttkA4&E_DRkBh)uUcG|RpPCXu&6P<dZDeS$F;FEi#8&lh?Hd`ruMrldcIz7}zUn%h
z0muXD{>#<b8NwLB)E=&sW0rTx11R&~H|I1;BbGH1C6avJCzyYEODGQCV(0Sz%NHq=
zE7N@I!0YCjJ@Z@Nqxn~NxVDrW$$-PDkq?IjuGwuRb<!(nTE=*7soo9stNqWrboqKN
zb(P)C`<lkb3LgCF(pZHB>3^jFExJkB_K?*-bkI9%tkVq@V5G|S2n3s{_%_w?lnU4^
z^=KA#l+#L0Rye-GVN#*H{;ahk(PEABR}}J@bOHFYsMnO=2?*pcP|>BB4+KE6EoWo`
z9}n)$ijXu=*2mnci3g~45EzBVZcr_D<XrX(_Uc{_U6Y~~6r&agv|$EiSsx6tkuK3C
zthahFpt`c)Q$*Y4AJz$0FYcLiPqKG5umy{?$N&;5fFbv@1!y4WkWrxbCrumvLsr&H
z1PPy*lUn3Hh3$48L`+DtKp8(s)pWMCk)ETvfMW8h^x-=0Mo}SP`r3JPbg*|v%R%%5
z>G+S!@tP#Pxv)Crp`X(N8^j2=zj%}p%|_)00v!XWV>k54fB4J1IC?7^I)Ee9@&yME
z%#S(i1kN~Pb<8Z)h*kj+_v=f4<dFR9+}f&~M1XvF!%zfF^xvYO_swAh8e?>*@N+M&
zxxiz0Kdz;6E;x5S46f_Pd6;o~n!3Hcl~AH1eA#$!!Utt{;$Ol0+|IEP)DN7!%iBym
zcjHHMfLjMFTmNI*p|^BvK}Xv2e`vA$Y7mxr9Mzny=E-%ik60T`$(g)zr~Wgp-0gK^
zWwqfp_0A%FkmZN;Q%LLs9eH=%Bs)kNm$<$3@kSCA^CUv)b9u&r|9W}O=U^R+12P-N
zG%UbW=<;Xv41j78;q^(3)@MHg{QUnp#Vp?YQi0$VTBAHvQA!ut6jA=YBDUNheIrFG
zAhus*VbE%^U&eXRis(Cy`7HzPJQ=x*q7hnrkxyUnZqw6wdP-V|-lGh7s)(9{eEoWF
z4x^h4A(|>e(au9?n>G5Ylm6}`s+i^C!!CFYqMQ6fS8l#*@?p?nOj@1_yO=fef`G!A
zYK`gpm`twjEnX?U3ME(cQ`w&ejWPX7b{@Yf3dnf-tDecN_y5j_tHOu}`(FI?c<gcF
za=p?eElVw035nUI17Hbkj?L2eu;*M8ok6(K@u1bk{jhX^e4YHG69pj4wtuyg2q;;;
z2RaURt}k}Z9<}7e=g))ltC@Dr?bI$WQpeX@Ml#kHM;{H0RsNaIE*w)U7$iuIiGfs8
zQ&Uzuyn%@)LlX8<1$;!+%PKGOfXbW!?w?`pEhMD0RL<^_5{m25|6!k)RSZdlLd(sr
z!NK*<K=$wA0>&V4&Mq|JF+Wfw<ukQHvSoe?!ZGR!M`(LMuw|i&*)Xs25Z;d2fLbXl
zXiCEW9EuaU4TqNr<s$k@r<6nADz&1o>>7<Q7Xa8fI^rzo0s=Qc!<^X~enkS%2aJqR
ztC??_*zWPQ-rmDqz|6sT6&3J^E?QR@Ee`>dP*uCzH7LrjnWIzKXT}>5QO?sF2Rm}Q
zxcWx=#8-i%jrJOVFchx&_L}apeWz`pdkoRlnnhic%)W1F`&4W7{_^6YlF6>3dPv<K
z|8XOSary2KM`Zs<?N_cw9J=erA6;l9JWhF%Ha5Tr#SRVkM13w6hl2@HKUugo8GeR1
z4{wfm8k^@<e5w;NPDQEcT=6gX+RI<?udC@BJCo>!JBHi!k7wmp23=W&N6fTNakY77
z28*`$wS}mSAX<yDq~{leZ<lrzf6r%vagqU{95Py@TD0RCMSEOb3w%lgRsD@6Qh#6g
zi|0)_XL$(qqi1}M9Q1yh5(=_wIQMKJ17=TxZb%i$!O8FPmO(>DPli5J)?hLQmah!K
zG&h>MiDYu<ltexhqKPk$X`HRj{<ru@5dQsbuM{KcoVRY{O-8g#Yxfx(=dy;Zy`XRc
zSB;OeFDI}^DDTW%OE{jL%*VxW6AI4|4<fAPMZQ<0<6lM|xtNNUy1+OQ*_}|+)3Zcf
zx(*#Hls+dafGk!;EIWI<n-Ve)aFD7mXGuK1fp{NZ1kCRpI2ot>7>lkoJQ0Hn9(uSt
zq}z(6%+h$-%g14qh`J1T-M*AIy#V?(pK+X+kh`XLJP&)s!1~P(z)oX9V`B=y?k}YN
zS5Y3jC9bzo;5+heZqCyCcs58}Z_KdH`pYHcW-^7^FWTDg7ICauT;UxkGdV=Oh&xf%
zVXdIpAjd)8X<@h|LZT8`J!L_@k?G^M8H&>yFB==b8lQvhco&?Nfj3@0>rWGoiC7>&
zL2b_Au_$1v5p_4sNq{28Tou&4lQ~b;MRoCUcEd9g?o4aZIQA$#VO13od$=k7jP2p#
zcBDs2nD`%Y?M6K8ZRq<|wc{}?57(|j)rJ(Cs<*58Y30~j(dNI42KAb&Q#~#&4Z@;v
zrs`y+enhyQ;Z2M0_1no74WOp1=-T9+s(g`S!SMv6v-*9okFgcey+M;)O7Poz>>a@z
z#P*vxWwoi|Jz=2D!yqbX<gax#RDwsaWBM7Kn$sZ`Pe<4SwUo1Ejk|6d#sK^g%!O#?
z;n!PFxg!~ULBu>CfR82X8Fp)V^iMkVl)J|I7@M8d<f{A1q-|iO`AlGrUSJzNu5dt~
zy7F%a#hzKRuigI1jr2rVt1%V)C28msgBeofY2@nJ{*jq6Ng;Jwr4>>E;5!q4Pc;`e
z)9I(_)3x6C!!DRlf4h`kJ{Y-T9j|q=R`~-CjiS<+Jmq<Yz-;zlEV<tPRKAs+o%v?T
z<pw(>BxD*wKe6|}&db=#>lM{=S+Zl$WwCjwYcg@Ikl8ETapPYHi3<_&?(H8y^-h{*
zprAyK%Si~+@S!X4ESfPEk~$X=tFFwfJ+JoH^CEV2pfHJX41C&63B+JNrDek($3hZ*
zTBMD_!3NR(^^NX7K&uEtXy6tLKrr&h&`X=p3_u_UMJjB)xO_qM?BV{IiZp^~cfO&`
zIdW{$JD_v0BalZ=^tY{n4nt`=^e<2=x-<U8uYgJt2viU>O*y)&PNKh<S8!Cu-U~TQ
z9<Y+QDbQIV-6**TmtSGvE{t7vG16~5TVm8_b4-(>>=X$+@rfuhbx#Ygv`J%XUc@OL
zC1~o!xEn8eRJPt}AF6`w-p);@UtU75I6)r>89K^_H~?F&W`cXcLQc1<izixDB6_kh
zbVr9rT?w5g-VUdSU8|{nTXx=GPxepz-*wQcpXY<98?Ez+%asS>6EY&FQ2X0-8rOnl
z{=B}l;pdVe_ODmHzjqH<Gs;Nb)HiFf#OHK}x`*~@w%%(@PGZ%IUfligq`d;G-u$lx
zTyl%&fQL8Wc3<I&yy{Cl0Sjxzs7~xP5cn2iwK(1wo~50eqoK*s%`VNbxTvri8ZHYz
zM)NxEJ8)s}VchV}xH>Q2jlbtYI^&Y<;IeuhbI)6m)O)065?~^!5DO^sE}FHVIt{Gg
zI~YoN&TFTI)I{)$$e{J4?p=%cc62<z)LFp-8bi}eI6h!L?Kwh?z7q;{TcC$=`_`jL
z`%+S}wEj1ptF?+6RMeEOOG|H2;LFGP-frSo=uwW%k5Q?qjZw|nyIWBh><0PvjN;-Y
zZPzV<))r1^==}K2MnNcIn@r6m7EyO~)ljdS`_KsGzvCZ{?Zil>U&<t$wgjTTP!LAi
zGftpA#>Xk?o_ojz#Kvs!`+nW(6C0uJtvyo<LtyRch9Y(!Ig8BQ-^BGjLeMcO42=n)
z5;Q;`9&HQEMPZS)4h;0S;nl}T*hpWKtGqQVm%}wUCUw&~wg@)64Y8ta;v?xET<l(J
zK5wviDpTV=ys4{i2sHLBRlBujRh^&PPe~$la%8j%t98_BaDT-tU1;=KD5)rjL=h0?
zmq|Q%-)aq9cK?cztaD&-?FFO9N+XOdX22ib`G)fkJH&4<6G1kdEc1_yQry&3MSpZC
zTR|dzPaGlp9!0s)iV#=|sbLlu6@{$c=2GXKdpXSxnL^w-XqBSZ>uLkMxeVF6b(5=c
z%u&c9`i$t$!WyXFq8<{pSbvR_?HoL0u))s5<#G40aqqh;l0gt3Ke~P!_|w+2u(VXL
zrwUdfPKiFXBjp@pmEgI3_T(Ei{3SBODh9@Vm&X*t1ZXyc`BAK_E(PZ|4XmFC^7$bP
zQy^ux6gi_hO!z3VVtr;gSEpJB+yVPvkJRn(Di1>&J#`7Xt~i=tA{3y+^K1H;KH6KI
z_3fv)1tkw>Ld58tA8Ov;)qOqR*DyIvhj>_>nuayZLH!ln@UsP2k=At!$d%9AtW?9F
ztt*1UA4SsQ_=NRn4=kw*kVWQVd-WTYM9fom2$ibi<Hc4IL#`96=bu=lulbDWqc`%x
zQJ@;&^w}0}gk`9o?4!8TIO+vbQ(Q@--z8=aMaer8oAZDUWX&w>jyv!~yI)^M^t(|P
ze(89swJyyGr{(sn%6ux=FqYa^9mI`W$gQez76ERyEJY$&70uuzUz9=$v>W@e&~~3+
zg*PXMaG~%FCac)!=x_&w8c**x_IGBe+{eRr{<#yqe+y*g6+&@(JR$nu`@_1MF^<q|
zy?U57Q(s22LzcmKUxe7;ayYy**ftp1yRTX8|I46L2IECmUEAFJnT3VL>5T7=Pqv;$
zs~tBK3!VV}vJb@x6Nv8R8op#LU*J+Uu7dr#(u_WQOJj6-s=W;Hoes*Hl&w6d9|NL^
zi0Ge~{#S4#4!qBUO^lBFbED+($s<Sue@XW?Bt??z*R|aKHpK9Mv;g|%MGG=1c3!V%
zU&>)`5zt2)V}5`sK>i}K$$!<p{)Gn0eCpWZkL_3QuKuRGwVq4Fh}z(6nuH+Ab7`BE
z(v|x@O~HL#j3q){c3}hQtPG!=H4si|{Nfa+rmDIyC*$EEDMUuuINUA5$L@!9n0mRr
zkYI3iMU^JquJW*RyM|(&PmOYsIktuur#}1R{msqu$CoPh)Vl5ndOANg?4fI<$N|ic
zK5`dHK8)%#jA{8kjLSd*PN5J9c+|~9TgB0aArmg3@5re(eLCiYYj627A_a=Ls_PA2
zx=*U5^p1XoO|&Fy)HRKQUSu>(Z|wrfM?ErwTrGEkiX@Y{)j<)|-6qwm&S|x)Y+{kp
z0){!26^H8wft`4ZZ53)Q&ZJ$p7P>bn*s0p{Stu9w))=gTA@;ua*3M9DKy>LJieFy;
znG5_1S;od27oV`P^{0l-{S23E6hCrZQ)5E-B0mf&_h%TAuh-J}DM7f^Yfuv?yeAI|
zQ?C4z#r(xJtrY&{IgkC9f4NcPu$2C!8`4WoP6irWo7gr`86Y4R-;QMMU~K|Vm+EK8
zbI|#jyA~Di@yrZ7<|jt2D_aBNa`T?}@Vw!a3Yd2MJD2st?>_IwVYmni2VqE|NrTux
zgyo56_ZvyGj;az%WTsqeF>y|UQVA#ID=sd7mc`~}igTzad6s@2^+&?PT0O<d%Y5eK
zsTb1M+hO?#8be;{3PPCsPWqC!nxLf`thPY<Womx*aAANF^RG|gue5_;lRpVxey|K4
zZ{8hqNiJg3p*DJC90VsmjmFE7k%G~=ih8=Z9R(p*w)o)P9;-Pe7Y~EZ&!@M)j;zG@
zco_E7gH~a!k%!U?Lo!{R4U*cP6GXGr9(=eH!mao@@u*;EzX=DF{PBzK=Wo;2$)d(5
zx34FCjHxOQP-w7DA^#1d7SeBm@Ef7Qf51w}T+(}a=c5`L(MPJxwMn%|2AgLo2d>V(
zcS~KZ8+tR}d-lbki@NeRQ*CcmGFe?z^m>(*YUj+<4mF~CUT|(h8Ge~TZW6>*+M*wC
z+>3amGv7RH-r2YrP}<Qjv>R(Vf}`xTcF$GVz6u+kw$ol~!trBCp`V&?3%LB5Torpt
zehvI@ksI#K_g$0l(-B7)?Fwf%d1na=N;v~<SN}R}tql|$9K6+C&vz7m1Ne^4N)Nhv
z8->sPPji~Ch`i5N@@0SC%sk`^q?yNr?kI2}5K-3k*$}-qSKbQj*}%>)OGEA9CYli|
zd3gD|+nZeiLDq;)sQzg1fd6~d3afWouWrwm*K0#_uVse~QR%dCthiIouo3QVk|nIk
z4X-@!k&xgX#n7nboe;0)OkJC-7X!V%o|k0Lw#_Vkcm*9p3b>QG_o*5vGQW6i(Z0A6
zr9X4E*RapFZ9>$1YPQ!rk>#iDdja;@q6C}YniJTS5w=7#F8ZvUY7g5JuC*?WW4?m6
z!yJ5$Qig=}+wHNK&QDJHRpwU5`sPs&pvh|-Hra;;tkMzkE-{Ak$n@F}K(ybI;x_W6
zq7iN*QKav*D}9weD=t7`jI%#Wzk6uJ4%IM-jHOyGSK+}v(kY(R6nY}nXH#)Ku!A~%
z!xXiuP%(6CUAYQ`<0%dVkZW~7;6CmiL|We+K^RuV`!j#I0?nJck0b8>&#HAI2Bp#M
z8Fgw(KKbJcq>J`U!CIZZZOZ8K@bIvXg+&Cg-$&K<;pOp^!SQ6-kV%Orde$32I(46T
zZK>ITgK#0TV!R>w;K`!QMV)2&PLeZumdskz#mXeah*;_mZ~AWLSi_af?){V5@ciY|
z;AH`$4R%Kd9IWxe!+>jCVA3VS=F343IXV4+r!2Rl2~C0POI>W^eZ3v;J`=?|;RHEd
zI7N2C&2a3KC!+`vrm(=X!#p3dI7H_~;hxfbU))M#FK&*#kNV%lo_4v5I#WBA`zj`p
z7EXBy_I)~Lqq+`M35wYBVc}&u{o~vmPWnt%VUMYz#-{KSiOfr`l%0~0G(W}_yOpSU
z&P~HJcyvoGkJIVtrDx1RyK6}G+{FPIw*kcB@6z2RtHec1(Y^Ag89sTl^>?81P_N@=
zH>%_kZsZ=qcThy%mtHRW`!5?1MT27F+0NJ23@GQzIH|tDI8fV--Y37!B{_qq(o&wP
zPk&!=8Qcs!gzi#9KxA~>KbMq0{rwz)#Wf(%MY-QTcDjT{|M3m{T+f#zbx1y3ocOo<
zYMcC#;Nm`B^~&u3H+=k8MYZls81;MxA)U@ksp&CVZNGQ0cI+%AQc~XHdTSw>M4z1?
z#lgWbFg6wdT&6!4hQm!>-YVqFTa-*f0J5CnB4a<?%i?xer)ew<{ejaY)Knk*9Z$VL
zQ8wL!k#X0&e&|EpcK%P1{CO9!r?FVKzR|1ifIGwr7(KP{_|(#B<T%*qED-G|G%w>|
zQ}1I&*PglcNXbkxXXwJKLuI^iP2P<H5fP*ce$Szk=1Hy$I;3pVqKKLGNCgL0dF*`N
z#-NICUcV(VZ@Td+t1%!>JD;Q%aY#SBg=2Uj_nzoe0Bc?gD7g}c_l*uvcaNMwrAZIz
zff-{og7Vp3&&=d}vi!Ja2Z6)UloKAZa;$tqOs+hmGSLW*4=c2LVDlUYk6<9s%E>nE
zmD1b$hia;EadiD>(3W*14fGBhWi&ddNna^z3d%3<&l4KT6=RN5P4G7_h0+i2oJ+fV
zrdgvhNskV{3J3fF6TIcZyYU7q<{ieGxrIL<ZhjZ0{#uCm$pkXh7p8vS{`Xo)F^G6?
zdFRZn^Achlri%c3l!hSV{t~^qx@NxJLh$v+o!y_^@e`fw1TB-HI7(m(i4hGou#F)6
z`SI4Q>3X?!CkO~IW|4|B_P5e_9K-w9ZUdp6R3TfSg+TM4&Ur8DTdQsSa)g2S`$k3r
zrF!s?Lk!oKCJEDIo8-Id_$Bf$6G1QTR<{tF6}_R4C5;%?XH}QZ<*x0&TkpENm#iMq
z_C2aE3oZ7vbvxs-lX9Y<G=aZ8H@UyNJ_shMNm02a{TJeSH#%LTohe-&490m$#Wbd1
zz%<FKiplsWtB`*!)9HuohLYgLbJD|t#!@(m8#0|0|3`C=Vd2Jg@P@RxyN2jkzn3M^
z!fDGqNOuW=*Ga`;f?^dPGf{6R!3kGvz+l%h?ZkIqqfF4XUenV{g&BaU%1KCH5w3c`
zy?F=?A{=zRmnI%z^S_N7LW39YqZ!rrc2a|if5dV=TSR3kE5*w2@Wuof1j^5+?cUV)
zURa;dPgtrMwMXMAC3ACW%`AU%Zl5)NK65g`JHMc~hB#4=Iyr=N?928I1xnY53{F3T
z5+3x12;#s9pWtXp;y+@VCl{X3Q}F%X;7+ht#rq5xC;WA1uU+q*G#I&Qd8_jSXlpI3
z_6t_FrIv5o#9cp`Zh&PLNf$WJCG&62`4h7Hwo^af!?9j9y<wB_6==6N(8Fgd(CHb>
zYw;U?+W?mfcYWt{cKP?fZukET;w&LTtLq*`bS(55b~;cCl$P{swa{osdu({hW@L=M
z4zCFq{4S7>H5|PtBd{7?Y1t$rgmbi)ZUaSbGx2LtY1K}FXaKPf7WH$HH$R5F_EM_f
zHKTrGWB?~a{eC9pJ$9iVLcw5TWXKQ*LYAc4|KZEG9&usN3520PJPEQee(H&@bb#D8
zXRq@w1-4AeZpx^4%vEa9eEnn5KLQ>T5kM#5tMzxLX`CE;4Q28_#2wTPZq8bJ;@7^Z
zuyRSODy8+{H-^Nx>1Y{Y)k8_c*Scz9-K>(36s+moxVSzLP4HIQv!~HnomhGhuY1z}
z7jhem;j`I2y}N*i!nm8TW@51H%X#3r06!~!dD|yBlEjcZI|+b$VpdkHE4nv^E&PpP
zGwZp(ek%YB!#hKGe`=ra^qNNi;C_#}73h80-6e)I8xW*iV&Of762J7KM9REu6aIq*
zCb%y79**ruRz!9G2fjC{Sw&;fx4G?B`eJV|>LI6v1p#*XE}{hUYq9ow9Ywt2sOOGT
zZIBbJ5tXsExtLuQQ&)X5s()Qyu#qs_k|1Caa0Dy`fnj1yBRkA*EJY2yMF!FOrnejk
zVYz)<_7|Jg5FPCIVn(Ukp%GOp(((ij0g6%G?GvXb*iK6lsS5KM!-Gbm2RJQu;%*h}
zgV277Au6-D4L@hf44FBl+Q*S6?JGkPzL`sj0jjRrQER!<AcWo7AGN!e#FWYoqB#gg
zf!lN_popQo;s#cR64>liFmn74Sc99t_B^~}mrZ1c@oTRnQ$94jn$@UN)~x*h55&$u
ztm%`HQ7#%H`}Hw@U}vX)dzi$=ZEN9$i^s`rAYpRr(6<6<p+ecoWBuNH7sepJI4)6R
z=9?$kvNWImh<$wp6AR;?^14SIeeJ(dB3@7GAgDt017HcEsScz4fr9x`eYjpPiWBV{
zjI;Gl*NS_CaS^av7`uP$+o^3r{cPAy(pNV2Hs)~)hIV`Y_Zayi`$N+^RD{;&CI4n7
ztm5ot?^dm6InyvE-rqS0`!poPO^0_M`{<JyL3^g#=dbtdqwgV9HWNhyY8Z`c2lKai
zzR&Nlsf(htkF0V`BgOibe*4UODh9o}9POo*TdneHK-avPyD@h33e*AgFM-%^bpWXv
zQaeUa#u(Y>@Det1q8k4I<<Q+(vhI7hAuHyxyjHkS$2Ckbmyligh8~#SIx(;EgFzDe
z^vgA{$=w~&_a=(cROX)@N4@41CB6FSx<h@+%R{t9gx!Nn@^kl3Y(NH{a6vBoTERE6
z8v1_$*E>wEzL3pb{oNuGc0Uf#KW#hM;8B#n@3qhR=YE37fT<-zii=OwdyVF%pr+P)
zZ2k|LHX>f@Nr<^f&Ce&LqNX;}Fah?O0((Q%(Z;=y#!+M=fN(k=_=k&T{x6p11%1ZJ
z>7T$SVkwR@B*DA0-WLCcO8^0ZPr$_!N11RApgfpyBre|ifXD7*oQ7xz>Tnt-UKrjU
z+Q^xMHlxLZ{DFC*BXZ89&t0B&NTiSN28Try#XX)H&z|Y}MVdDEYr-Rc%~jLehRR+f
zMsFD2cMlI`O+LlG=Aqg)^auOZkBhWyR0yOqf97RwfYM2|+~lwJ6WxuZ{PXij$6eKe
zvWulZq`%AP0UWPQot(&C3_e`cAes`A2)46?$xf`Hgz|i+CI%}1fc85NbBCOwdj^+|
zwxTO$rJ?F%+!diAT*o5`UzZra;J#`q563cy%F+naFY7It-IH=r@bC30|6Uh9zVa#G
zJ6x<XTxOX^iT2-5^~2?y=tfBGa?`O~)a0T~&fI+9NI8Qg?g-W0ov9Z=DKNtEth;>I
z!80ED^m4{^8VN8CAo8O_j>ZQ<t3ZGTowd*mK5I)&YWv;y8=Ku6fgIXO&TBRbQsMB}
zSfMNf+$ko<MR4IZ@Ww{cP$3B}2l2V(oSApe{6SP}U(KHxC(o?2UUR?2TlIK6R!wAP
zWaO8Y{&aPfM}W+8*mQnaX92|qDz|^kR2v7v=I<BeOcbQGTX!L@r<TX=#qST-yf7jE
z1!ez`aUQVfSeW3=$<R%)jmvI^veViJ#3>(_#JnD2(`umno}DAu8FS-Z@KGNqD2DPr
zGfO5qm-y0cZ>ZncXBw|Kb9ZFL_+l?eamvoqiF)|^p=-oTe4@y*L&Qk$A^>}Ka8{6@
zMGHw#Nks3)_J;~7=1WRYc;abI<{I9**K61{ALmNFDNMe2Ux{>q&VI&Hsg|XZym|5l
z5_!$dlscJjFdzP<QVxP6{Gq{sRW>1u&71+1C6Lyp;rS;-&<xm^cb7-LcEx2{e&Hzz
zqU<_$se}LEnp{+;gyIQYe%mk}X>Yd+9xT@&U0?X~lzV!=(MBb9cH3j~Usp<4jc-5f
zEHY2rH@5bsax@=os?o^p2eHz(oUFJSiAcsn-x%Quj!VAp-+5lxnq`*fXw>eY1Fl`G
z?9DebqvPLyK<HekF*4IQzMj+S2R2(}PJX>JV1ulC!>ITe|N4O6{Rg2Uz}r&WLfyje
zG`_Wc{`J~FU7&C8yyr`*04F+qVuXDzsD2)vS+B{I*P{OgPH}9VkUs8ciC@y?t363v
z{+eTCu&v2@n(oaQu!3M7vI@UNes<FRMD~mw?lUDWy!pUBLO=(tT&%$hEklL9@aq!8
z`-r6-K6>v-+ailP<950i;WvlK=t3*E!^lTIRuQy1u;Es`?ref`S~F=JM+R>v$}9e^
z_f7M!-phJaAc|r)c+x9P0(ie7T=#$_K_|C@MU2Y`ub&Or_J_&65wLvsTz!U?lKt}#
z=6?KYCE=nF-s{5_oJq5F(Gd-2Ory6TM0w{(I)MBR?|jV~e`4-eiH&-%Geu2c$G0sy
z#CwVEzF|iiXKvD0P^idYSNmA)Bl2DO%!&I7uaak2w-FM(Ygw5&ecr3=%x$D;mJ|-N
zbyuXVo0Hqwb?(6_Q3k=kQ&9#%+-uun(EWK={H3UvgyCbe1q65H!Srd$t3V&-x9+DW
zH()_^c41-RY^h8doA-C$4|QM6?iA_}A*2JQSt4y*-)agM;{Re<I-r$2@^!}ZyVhyw
z8QTALYPCdAHZ``^$4AVQn`oJ7onJhU-S473k4Y@rX;x%hkPyAAnAc?LbE4){<&zng
zcW_49X>a;~Ar}V3F;6*fp$@X3o_hST-;#J^YqO&J2IADkl@|)1c~<EP8&JGYmMA@v
zrW7@U&ownE4K14pQJ4Clm36O42<B7k<vD1h&w1uu;ny-{jQWH58cc+ZuB?wK=Cm-;
zyt!tZ>KT_$KLSGvp!95j&>OIVHb0t%Z<%S1*d)pCVV>E=4fph;#L4UFs*81Js4GEq
z9W0zjKh6ga9hpdKT~@t)1_{A@-+$7gMu^?od#q)6i3I`Wa(TaI?~{S8kP#w?9$_{?
z{kXwEm@bnUNMHZyWtM@wOOL+d^4co+Y5xO5%a{eGNa_>g;<SOR1B-;zpDhq5tEu^k
znwnbQ&@d;w8U|I05uS%aB#Y&(&z9@+7OmH?2+wX55o>H?R+aYELc1sFYS3ic(Az;O
zc)lV(B9q$i+BkDAea7~*7qVAx)_7eBSZ87B205S|_&a^x?SXhRKc5MKgo5rlK0p^i
z(>t&Tce+B**D{qhNL7RXKI$`*vmw}e-F~&Kol(-tDPpJ?>^J=kRZLy-BTCLIFd8X3
zF)6U?il1xHql;FU!RW{ofr*z%*{+Vy_439i&o80A`iW6LW7zMQPi+sxlUeuLnq3C(
z+y61rdOMbSlt_lhvh?m+4F*!sZ`e7iI}Gt3lss$N4c<W%TE)=eT^O9fv~XAJ_mC?|
z9}3rl_0wvLOlH<^<a5>Y8~^0hp!j?{YyZcYhK3-PRgjQbWL{#DnFn_xB-nd)FN^#r
zwaCxJM*1TUXA%kO?%u8Ct!<B|=7nE&yL8#`QTU*te!I^UoI`19>_;8~us0UhX!g?2
z`7}jECE`sBh!^L*PcK2EnEiLUrR1vwM_+GZaie<u2J}*mK2VT$mz5CCps3K+rj{))
zqvE3zUKz<bqXqpk!gLTJV0E5K&tXVO6A=6(bF<bL5?gd+2X<(_yAxt<o$SXaJNBQd
z@dJ;J@IO8Cp*o20Mzw@7x{~0%*x!BP3{;ix<JYJH!uAc!;zJO1Z5zQoa)d_foZY=;
zM$DH>8ujp|kt2J18=*}6=g)XofD7$Yt0A?8aIYCxeQqfC2+iom!9D_=r_99oFZ<fZ
z1vwQH5N&452TM;G<wfPcT~yRy2oGSz=dB#md{ysj5)!MWQaHm*o}RruGyd+@M&EU6
zBM20D^%r3a*h{592y-)bvp36%jn88-F<tR`8AVPiW5xOMNtbxjij^I2u))u@QQz&&
zPVnBjlXo5qZ&d0Y-asmCcH}`pQBmi0%$Ke8k?qOHjoxh44jaVk#yhH4kj$>?yC0Vr
zPrU6wMO4xA<=P>oaTPHfVg2X?ip~?s?idfs7M`ijEPYMj+dCXSjay_PT%!(7yp>0P
zV-~j48C~QrMCv>5zpps`n<JnxJfX4A`S+qyDk>^^%hlEd(+`;q310w&sk$j@@zys#
zvx7C~>_E&m<L`}q{B#QcUhjSFcbtjtPr9W_xvF<6=WCcAX+b!G+OF8wvbAdlrSQuI
z%>p7a1)4WT8KZ(DDcCHZqBokq?m=o0uk8A0=?Lop?S<&{A8pS{P8G-u)X;JodMrfU
z_ahhIaoSkMp$Qlw)SRAp#;;WrRsBcclSNlDL%P0Y*D7j+15$)o?upgvwx8RHLl@#)
zjaTm+9aeM%Rf+x{;2h3Td)8h(scLA3mrAH0yuYX@eo9P?3LMUarBmqPGAS*IzfIh2
z;!*2<ZXsPyjnDAk8=llP4O2sbA>hnxmNN+DemOg1S$+|0&OG8Or&~<LK7-7rY3HM$
zLlA{t>#B%hOsgC>8nUWssbnk)*I=t|ZbtmnZv<XRKNlUPM7e~BZi2)M0};};R3b$0
zU$|aYo;cifoM%*iBdTjygodRq_8Lt*)j78eCL)aZz{)zIceA%&s{_^A19P#>^};fO
z;fZqsC*V#LH5iHG!7QLKjB}#rEyW}TlGZv>lG6+Zm<uI=fTe)}D+e1eeSAY^u0h;}
zOIBoRk*Ycx0lMzuc(tjGU=}g7`2D>BM%-pPp7lHqv5ei}zU~mLb@dHudtOq*b+N3)
z&t&eH^H+PekHe+|R{URO2i4_H%?AAm?m2c}z~^d8JmX@6UFo*}KT3&BcjKd=L~V0@
zA`0_aKltbnoCjO$AJ2nUA*lvL(qX9_At>NQ)ITjKyK}OSpIvxBvB<XYu8z3cfWx&K
z8`^CB&#&qR=i|*<sR+9V+l$(-qB~j&3SqgfGV4c^mrinW#W$xI7yT01xf}b3+0ndk
zmvGjb>!3958}bnNQ050Uib~ganrZY15#={WDn3N`Pe%85#<Whz8-`$3$b0)c2?2VC
zA<^1j(>F!WN(!<;XfR{z%T>HT6e8zA?AF-CmZoa$#>VZ=n7c|!io^^jEH{3y-hcDH
z)g89E2RdNWm(m49;2^vgc%Oe?T9^gPJJ%B@JSZSW?tqlc2)7E6p_(WId+Il@Hs8{=
zgI}#pR(f2vbwUtF00*CE7W;)L$29VVP}QE-9rJu8?UtZupOVdUzqAqZA6-adfd<PH
z*IUmMng|Fnnh6uPp*Io-Ot{|BL%=$1Sq-7l3M{$mRpV$ZHp+>?r+IR6_F&JM6J%1P
z7MniK3iY~&s64R9e@JdbeLT%<iQJp%>i!Y>1zz~WB-51qi)3B!Ik8YfVMT6|Zemp^
z10)u7*Tkx<X=K10U9*y3VSE=`_yTSqdzWiHD={vWyGZHAlR~@g<}N(!x5<Z*j@ZaX
zL~AG9*Tz+y@A?cCAkE4X0~#LvLgsPIosLJNlNv6fh8r=A>fkf1@0P+~vy@j+(8&~e
zB2+0P&4fXA4A(-})7lg`aa_yWb~5PP5wI~w-+|ZPAk|A;AKurR<ut0#L(BPnmrJ5D
zFSB^p(pgo)S`hG!L$SDA>H5h@{)bC_G~+VahUPe|tS#h*fx*1Kf(RAP!#+X7SwCVC
ze`>WGpME`(NF#XE<CAekL9PRj;!Nkh8vC8LB~khB?IJ0<Jki^8QO8zP^Y`iXN6fZm
z*OEO|0RM+@OUkH@rHp+L{2Ve&UA5H!RSN7C(l7&$&?c>p{<Od)T9aBr24<kfz8(ax
zTHwtr{)yXQ5ezo9DZsiG)M|(jzZ~uEVsvX{MZuw6l{cZgKM%Fepl6!sj+*gtDFf2c
zQu#O%3P_qlYB+8Ka07zol_uT1FDSY|B4}`y#n8LW5E09$tf+xkgDijscwTr{OVpU@
zQ+n6u)~88k*$25+1~mL=J(ki1IXVM2#gEUu=K|-OJvtL}b8{(vadAmu6+9ZFk`7k&
z-DaVN!9Zt*m!5HXC|t5Ed~ws0)~hAaEO$NXwej83?5W^i#cb4oJoaX`&Ve4}_;|&}
zL>HNeFhP<8lL{F_ack7-l<LMNPdc=XLkA*C!lqTvn<$^Z!vIMuvW8qc^J|F<6<TKA
z^f@ZQ4QuV?7G{fktY1w+OCZrAJ#*3W_SBHb&HcTuQJ<6Wjmgr_FJNXvO)!|UKbA+b
z!-WTO$zz^Uq&UA$a!U{|y`Ex(u4^UEuE`B~zKrQN;L1ye7jAYRJYW?LgcLnGT?E00
z=bHyrG$gU}!MlS~#@i_W8f=g>Ad#?6d$vfz1w~XuCcJ7JtwpyD3}dZ={UTC>otZFa
z9@~dy-pB>jxB70b`nlie4VeOv=ii!~1y~Y2UhS=Sd;60TVV_CaXq3i?fL$t#xH^-Q
z1;mh8P&sWvuHo^+g_6Zz0kxZF#*mebe&a3V`L`h}MGVEluVql~VtL|C#W{_TRfoFg
z6v7nDhb)tbw0i1)sXV{KOdQIh8L5_X&7;PJ9^^kMYm5fvvd*muiq$KW`H5!MKbz~I
zm_z8n=<V5MC06~CHQW24@gFU~MI}^|b(1J>Mug8d%Ip}}f*0Foo$3A$zZUD=wg<y%
zKy^WtWl|vN=_e5=aTP}O^)9|W8w0lTumb;~n!(^f(I9E<TcQ7mh#-paOpICx+98Xm
z{h#w905^6?+;@1~RFOUVgipsI4y_tFF<GdaiF^U@rl<yN3r<DF<^1|Cgg1gl8%|xs
z>f$K)nw+VQd`L#}t?0(#(H;r(CNc&;nERV*b5OBjFDp4Y;CEweP?9so`$*-KE$1SV
zP<2B+P7Ik1R1_nDryYI3>H;tVP=3Op>=qle0{n(p1c5J@*QJ!P#ShJrAn5=;5q&m;
zA*+}t0k%E^Oz>{#KASzYNUft+AiL(L>OXPFIKPdR61f2-YW5NlNLoA$a2DWNyL{Ae
zzmSy(O)PEHUv8s1I<%7-L2l&)aR@KAR4lc7!q1wk|2ZPnUGR&h6NrR0lnjxqZi^L-
zWj-<jTtPz70(1xVu=G#NV%ln~Hin9+^w&It!mcEj8!<zZ4y$ZZ4<WiAbvLi9CoI>>
z#denFZ|0aYOc~R-UQ}*t0A)`mzzo1|oNP~Nd&Vawf^TUatqv-)$cQPJue-B;yw?%$
ziA!-R)IAUgmUI`*xe?q(6w?@(p3{gz2G+3jN8*Eb>(rvLT_vsv<CTwR8>p#c=Yxlc
zXX9+f-QxmXO+bwvgmPoMa#I>gEb+}}Tm3lonm8eA!S4&I%%&~a*p}d_0*NRKJnpxw
zEqFX$3_~yGt!)L#OQd~5u)UT8Bdt#@$jY^XWC01A&tw3swN%vF=a9tubeVhe?t@WX
zXM^D4f`*R7tSO+Iae<GuzU4j5W_y(;nxfalZLW^Epz@h6goKtz2nobgNH(#3N^pf1
z?O0QTPe@GT8}Ihh1?i*f#|MRjrsz}0aXqeWeYQi@b7`v3>O6jDM)l{}f@-(%IVJFE
z<^h_Fb#PT9rhPmOtla%doda(e!rrOU>Kojf^#fSIIx<UJHK3Ubf%0D=)eqXobeSw9
zb#AkE+_)=K;UNK=Q>V`+9(f2@O&oI4AZf<hLpCKRscU{UcTz2EswR)VCIrE<5M2>b
z5UsJ!gsoygs4@y6RX218)|0}JRSJb#JXQj+O+325cu~e#(iY*DyzTSp#iQZi(eGER
zlYm%C^p`Yo+B(N9#=~z$tue~d$l_|MWo{3w;EmuyiXv?hF~SJgT*G$5I&fa}>}}13
zOkXna8-FT4mPvxj*|&Mm$@(vmlvWk`@|`~1=jXIS7JN$McUEC<SyW<b?|RH#n_I*)
zHqgN*>TAg%(wdjyaJcji&n%f;97Y5K_f#z!1X1_sGR4tum&?4~D76Q@h0L~&8S78T
zn>i<S@j;=xl0TB52v(pv;NwJjBW<rL3S@uORM=`>VVDclIcg~>s|Sf)_GM+sei3(s
zKlr32!wX`Jm2{UDOCS62B$1EksoEx{5P-vJfG#YBTRwSRTvCjT1L_}ORsl`L%vX%0
zbPlLOvKL$LK9=?qUkqHx{AAEV_TIa`H3ar>RVQc&20U(<5p{nQFO=XRX2j)Q`<3*3
z&T6soxXnw?AL8dU+h(Z*x`=wZO9mlTnjHTmS(7%n+<WGik#xQ`s6`tL=6@|dFvVT4
z#dWPeHhMNb0@TyFuM_PPGx>0+L%UL*oBf&{!M}Kq`K28`an7wQVc24h=x;RzEDn!b
zDz_GPpNgHMnZKG~(WW=~WV~<IJ1B{rSj}py%_s?Q&MU4)6}icf4N7O&<IQ8W3+~L*
z#T9FxzsWT}YSbBfq8HAPx5s~Vm*O@@e9ncdm?F=gKcB8>v=T2wN13O2C$HrRjW;S8
zSH4snB98kACh=HnmKi~@q;<yQN$1{vz}m*zEMQK}zl|CsW-y%!{d_MUBjfPvXJg<<
zM$mV<8>Qr3Rk;k&kPhGN%%NogQH{Tv`~k7$jzYM*Bbna#&n+f?dJ<i^%;mcmA#hJO
z{j_oMA4$h{$6W~I`~6IIf2XVU0{N*>$SZck$?J`%D?|0pS4-5rZ{HqqAVwp;BAC>Q
zi0eYn-D(euA!jw|ww4Q!7YZQGkx9JQ&|dL<5iO)?%rf?D8~Ac$IrzIfaVS;T126a~
z&g~~z9VVz8rJ^DZ;kH^Nc2VTTSp4bkuI`*>lv-%yRX56AgJMp|)XJmco<WGsz<|#5
z9_C#ciX2KD^5zlDV)v;%@7`3XF!YqT#Dygd<Fe0}z0^H3_m{Xf-fsm@{d08Vk>i}6
zBt0bS()>~?ly6*oqEhKZz106=<L$oryf#9>Y3Wj{eigebj#}^c4=3*zVTXL{>=bGd
z0Bxztvqa_ExtZ^*M)H!d)d6+ZEh$!Vc)TUjzqJP~txMOszSG`ql3MNoRwb|k8&8|n
z;;pY{xI`yA#BS2a%QT1wxLSLi;KY8q@l%@Pmjn+Q{NyVBkNuy@Jz$r!S00<|Qk96@
zPRR@_io`_1a2VRf;yps)Le8f@gMhi_AK2FyXD*Z(zgNV%ZwMTNp}yj{nzgnYvmB+H
zA2LhZDtyhm$i5H^Sg5S8DK${~Da_)b4YXkN8$D_DNvZ-USbi!fXgl8V;J$*<LjyH|
zM-vVuiDQCWvChX#-41uQY`zoL?zbVC(=MH5=isOqZN!DmW%3%Do_{r%P8vI)sv=BK
z4NtB`v7%N*L!(!mCg8PFKp`B?)%Nqlr;~**u7}OEysPxkf8R{*I>ql-1@>@aCq0cb
z3oev2PxOR6Vj|Oj7?%hz6lqc>Yg(&wdm?Ty8gO*4riHZoy3T<dmtPStdAIjppVwR_
z77K70Z=c!EjyFp%9ft@xJr?~#jmj&zz87bM1^W4w5_x{2><EMy)(u$Ebtuz$0v}5M
z4T0wm;7Rvsi*;}7E24hM=s(Yc0^~oOhpG;gn*ic`Zmx5Byl*b+yWTSkl0qIzYvE|r
z8J=sYJ&q=6c++D|Aiq(a$P}yLZ{OfUZ8B|cJn}qG64tI~{He69kM(3JyDQa%_XPNT
z#pr<ubXXn>y_G1mOyB+nKSn`A9z$47<giaa@k~2fWIMsD3$Es0O%uK+z^bA{xf8pa
zOmGQ2g)oL#x%b)&2QXPPe4o2*0g4O~Ceyt8IE(>3QB#+U*gXlF_~NZDguC;En$=?O
zE#JY$`)MhLjAzYj-$PLlsTksu7@YOTB~&Y;3O-cAE|UQ*er*)($sBO@?4lM$)W%jL
z!sqqrm`7IN!M}tbuNc34kyDlpnhul>ZDMruVEcWY{+k;|6k0-DZWD!Br8(1%02#rS
zm9q*iK(5=M*!&DEj=X+%c2`uC(>2#WT^5;^V>7af9{ipaDn3g~6X}oN7o77`q`?Nw
z>fBw-L~wZ7<UC&dpvQAZ5k%5bq~<bpZOXMX@99UwC)|(l#O|f=$B@&XnBi>Po-eM4
z=PI)L@=NPFug^&U=A<S6kt1@vONW8q_13=VRt~w@IA^czrR_u4%wU<ARsb64q<;oS
zNHNno58?JAh(nAfBoC5E<d;f=6p2W<$GuRPhXK@LeOapdOqx~#@YBOGKN&Nl7?*KQ
zM}LR~sJ&i)h(>6*su4mY>qcSU?4QYNBth;q8iyCC@a8@UgCU}xbdI?ws9^NZKC#&)
zT@&vu9-j*dv4!W&fKie$3V+sASMZTq3n0F*`IfrWiSo=~vLi7uooiHyNPK*Dnfjq)
zuY@5CK?rKl+0z|O+TzpaM>U5w@-^|k;_+8<G|p34MT>7&uGT+LSlnBr$Q&MYrba#j
z?{7<UVZ`Hoxw=#5arKX$ExiBWRKwULOx9=raC|grJKyX<&N)sufmK111R<bDbqbJ=
zgFBQP_;S3AxcBj*F>15F5}N|UI*p0;s^vyyEE|4XFDN|6pvIi#hc)lv<k1%UAdrXV
zR=2LiSQ}Hhr9L*^Lq}FUuAgqUTP$?Z(U%gGA^lU_ya{RQgez=Q0mU({H)w67Gr?93
z8xC_86R^lJqRFQbGQjxt)by=GAp!xb_6saf-+v#uWyo=fUt(i*lDQ0sXpcynlBY3f
ztN{fnWBj29FzQ*3b`wA^`4gbrG13_jUD}Y@fNdAJI9(k4Gzd7OrZD46NZxQvT6A0V
z;yv&2(-7P+a+91mMO@UzH@u3^GXnsMFCA7Hh#2bw3wj5MUXl_KP+qFo*ZE|4XY}1o
z<F;y^^V=&j!!f~q7fctKjEqX9h_hm{X7&~`f@edp@rsT`k!L9}YoEcG84w%LX#^T!
z-#w~jS^S3Fr?|>|%Cr||Q1QNURIbf2<KtBn7j}V{BBQN&-UE<kKw?a9_wypCYk3Km
zrOkYTG7Bw~+aA2~u#Jh@KwfQI#8u(W>$=sm{l(dn5x<X15GS3ba!(hc|9b@esre^7
z?rSh5m0A2w9FdI_3BE9|{#@n0iEpPPWVRWFqTO_lrX1?9%rH#+8I2^JzTS^u#ms#k
zbLnbvy=rhgH?6>I{8xR=K|u?CTG8;RAUiy6j_Q~mmX%-H`bJP8q=QUwz4sIBDAdH*
z=kX|lSdRYLWrS42R15PeiPcfA)QRz4q%)kh1h(_nb-s#@)W6jWEV%3&VtHN<-ZNP;
zIe0uI+eOnKa?hm3qtS{|YkTB07MGVALa7Hre5yRLKuo_;z!!At>n_()439}e<!XF*
zhm?z-^QG;bpj*DMEm_N++Yof0^p&jG!^m&iSJ8_h-$Nwq-~=meY*z>UD!sKed|;q8
zF*Q>6MZsSnbNsP$c`off@#%I~#52TMj{oyd){dRM*EgiF9WSE>umy=k))m6hI-<LT
zY&Yi&WvkY!FOLXZTogy-Qsy-4%&i5I3#fxkrpXO!N6E~uD*B;TTuxkP*Rt{+1?Lw+
zDng_@%V?2`<WQsqAa1-$n`&Ok#mV+MhnJ$l(ym)0`Kmpv@5h7veXyLGr*OPDdhzG?
zt_oyi6&(Q>u}|kGd|0)F5bW(4_bxk!O_Ot%m+e|)nZcd8oIX<tCbxuV8YMtl_AGzF
z4jjkVzFAaV(w*=8?Wv=xAnQk<lC0ouMM}!}*AiYsRkbh9orLsG6yvxKx<=c=q}<^P
zN2I9Q9G&o{v~IPs#mvk!bUFSQ>^>TT5)xKzSw1=y?~%6_lFsQTh!&;YasBOWsv+tq
zWLJ34jMkvLBq&)i<&?;VOc>6S{X+;1;m&BwZG-LkrgPRbZWqntWzxB9X$7)+V6p~L
zMm*!o!y=+OyJtG3lEe;Q{k3y8dFvB*wn#o{eC<>J{hi(%p}rA!Pwo-!h#M#R$*N$4
z!Wz1}xPe(<7(R(3Y7vH*sK<@8L)ex`h{v^w-+dDrmd-D|VCkPbN!T~22neL&WnH4-
z`&f#$#Gb$%sk2pF8)uhSkQeIr8IS=fFR2g~s$tok90cCrEt;7V-Jbx>grL^X;Vpyv
zHdXsey{Sbklk-;%4jSPZy<Cl%;f2(bjUc?pxTPZV2=vGZpwX~AL-2j&y|%gPC2mnU
z#G!L*aWR~$<CC{Q2Z6t56y459=0{?10X4(8UV+6~2~|my8z=PKOPIE)5qjod3eDwK
z^~qeVprNN+b34^(U{_60GirBVE^2Sy$<F%bJ33w@yTfUW`cEvTdgqJ&u<BgyAW~8p
z?Okp7A^2jO$Rk|Z%(fu~k|N6C!|_LaezVY7qpf)39wG7p9jQaD={h>If-N3Vz1Y*S
z%*LR`B&dLv>}Uu#$X1kPVq}OFj$92+5K6%3^^6_-oS@vT7k<4-?u&8xM8UOb+~w94
zKGK>NKf7`z%@R%*=g2<6_2qn~#@LjefFsO__a!{f<qUzldT%;L8U&#JFw^-2p56w~
z*5`oE!TDJ1rv+{6C?2}2?*MA1QB4j8U7`<32)=7-zi2K*n6uYW6259`+9#9y^wIMx
zDxu%sjlJ80xYQskA-x~OjOuf@%`C-?xF`iv5x)sNT)BwfQ<@2Wxy%G3y*GQW{&<j5
zh=5hi<`+Gj!tWdsx1rBK5`%;?`RyDAB?tDfv_!?!H)ZDp?V#K33(ww<oDuRrnLFz*
zF`O39F7Sh%B%wXR23Xe4lh8cVZdpd91*n;=oX(BsV8q{_3(J3$PB`I%x&KCpcnpY*
zd_R(@3PUI;k3QCnZNvT(9F(FPTg7{xC`e*1qaEmj^?@lYrz5i5fEA^6uhBc+Tlgiq
z2R?eD{7ED#zd%@!tsq-k217=WCv*w7f5#AHJSb{(OtT1IKll^T^t}SQ()!J5L$n)T
z8<L6iCptEIUh#pTkaLl|(L*hGvvg6tS~6q2XswxhzUtpWM-)r?{qnmV?XL&J=6Uza
zobE_-N*~#MVC&P}I0<#7polXDS%{P!9wk`q><%ejnzw#@-0TKNdr)coY8Cx7>mLY}
zIW#4nhc70ySuPmwEEEGnSssilxh2V=#o$?-g5BNucInMd+(-HaK75s-J%Dr9bI(7Y
zPkXKq^Dql>D=~krS>x_~jvSXE)5ZDN0hkSJjLSnV#4ZNQNNY!RJcSa3V_L!Q$Ni^c
z{t$NbUN2R;>qzk<oTxdtKi*S$^nByIdrl^ID-taj>HOkWN%kzyrxzV%lLvH+I4AwV
z$5c^jwIlNt%1oLJyu?JQyY?!1*6k{=oFzb)XyE^aqE7MsoE(APpD~<D29rQbffiU&
zWrH<=%cs7}R7MDtj5RfN54r&tzdK(*7_!ztu8o=l2ty95lo&i0)UB~NbSFgTm475p
z{{Gbke-|DU4&-iOy{o(e;te77kS@nzPh|3^F|A#qN;NOJ_d)u6&yJQ$NK|Wu<&(oN
ziAG^c60<~8*wOBVZOz)4Pmu1FX5iCCr=9Ad>=pNAdiyaFwkGzWYaPLWOwZorW`M*p
z%r^3*(j$-Bn00?fo05mR*wOm`*!srkI^zD%HclGbcGB2xY};vUHENQ3qsC}(W81cE
zG-zzw=H5R4efI30v+w84oSFH~_xGWYH;Phb!uO$Iqrb~VHpoa<t+`4E?fiL?DvEqI
z#~E$fFIgAbBjPrWU%iNW1R)7rbP5i*3XL$CMS>{EA5UNQj5Aop5907KKmkZcMCFP4
zGL)kyN#r#jNn&o%FTJ7fxZG}t6vM?f;x_p_!ockST_5X6XEn*Zjo(7DGO0+;$WLB<
z+-bNvR=Yu4;A=US8#Ndus+O=kPm2c?M8}t#Zaz`;l-j12`-Ts=4t<!lM1|HL<AYL(
zgzXXK`BJU+rIEZ1$k|qpqsqBsMZoM5!60=$B2E2)8%mC?<@#7}Rhq{_7-xlVZ?H=`
z{dH-1Mm=&!G*M(=7tvtPMs%@w_+SRxfw_BEEYXAYIaxT*jPLmTI{pgxsi<nZ@|uK>
z8Ye8$+E^PW8AJ-o`Of7?;5ODgqh3hmdvB)yKbR)l?H`F#F?+FU6D?9EVSIQX=KMP7
zY|#~12gwk-teZw28!_y34|$qmqp`~!D^C}mN*E>$t^F`59meEB86){A2{SM5o|?=d
z;?hrfM&3Xn1)VTflCsoNJpamb1hQmEb#KJUeMT6Yz(jT%m43v@O@DWKLMjy8GydxV
z>KBVmzJvMk_i(OOd)#l*_qb>ip=JkMgHd!DbX+>GC9C!tULeApSU6?AvmeTdTGaH_
zGK1MJS)J^@&~Y)MYLU=<@=$cWDXFzrp{NJkvoQurHPe0Y<cfpk@uev9=KMr7cN`Bj
z=T>D}FFhe(j9l6k{TF*`UTg&#eJ0%1mhei}R#B_Ff_+geJ|2jx<8I}vyRs-H1yy6A
zosvr?sf911ntZ%{Trv$h4qY-@iQ>$-zyz*}lDb4ZuL_sq2+TpwWOK(sqqKPZV*=)<
z8`bOGP?I2v*hbuvibAy5^ZoraS}$#K9LQQ$c>`ZDUz1~p(aE?d74n5?++Btd-dh%s
zoeSY9To)c7w+b{sS^+B~&ydD$5c54}EX@Gwmn>Dqtuq}3q{VAW4yNf+JNXeQ!HQC#
zzbspN?`vw?M;hDx)8bbg|1U+%n<k0b*?9BJh-n*GqkdMl5oz3*URptKW;7nK&B$;L
z=NjaXuc?K{XhWc#<<ruAb#<+slRTM6r<m6bMZ<YXCoY`(%MxBH>G8OCBGIli;bEk#
zs2hK>vn|jg2+(ly6NZWUVB3*Gz(ghpaQ<O&@PVyA5k92GuB#L##Y8GBwz!}Zi5NH}
z4AwR`U1m9M*+h!0y12T-f~0-1D_Fx2!-J}?a!m2ABx|NO)`s4BM%1UIedT5j3jj27
z?CP4DtZvLY*RVusSNrOo&t9~@BocpcIzRJJQWdgbG1=t9>>qxQY2@Q;7(6qxj$XQ$
zG(MJ%(=;@UhU_A)RJr&Wz#rBd(i-PXX(KBhgmAk4i%6}$<G}A+en94XRzDZH%&I~Q
zjU{)<W-P*QP=$G3vwm!O{b9Tp7-FE{Ze-*QNB%=7?KL@B;|!k+UUhUJX@jHB`J*=)
z&p~x^0qk$Jq+P%1BuBWe$v`9tj|yl`s`RTEohU;n>8gB{Y#-CwCXR}~!q7Fi!y}?L
z96%Ix!-KK4yvTTW5QYduW$Mny){sHN9A#aPi31sYai!EdH;ZWV?ceh(C$v7=qwR%H
z6UGe0T_~Zs=^tc{{qnDAtw*n8c9ToBDI3_y^XMWZ$QLJmdh@Ae+eT`tsDy1ao%evV
zNS460D7W493r4YO<le*&sHtNWX-zdxFJ~@68u0UceD+tg{95O@75wBsQHTVk>F_iY
z5xIYi8>=TI?aXb}kw-chGFdHj8hj<0X`Ve${PxgEMoxDApHULGOj0r<v;FJ)uX8jF
zjlZ%NruqG{?*|o6q)xkYWUEu?nGZ3C!U$PRkL^$L(NjTS$mIW(EqfCGAJv5%3Na_{
z;NW0+J1HpK&i*42JAqB34i*gsms;EtoOm!i`n1OHfugb@5oSg}o*c`E$8b;##Z2C+
zujdsNB_^KqTY^lUsGQYFK{zS8*s;JhC+?ScE!i(-O#t#=&+AKgbayHi5Y3Y0E}Lo(
zTcabsa1u@edm9yQ93M{X{M=GTRwnarB-_zM)fTy~Bf#NY3&u;{?=8JG7>Ol`=C>95
ze8Sk$Tj)-#;ZZ-?(?oqKY+%^iv*hz-#`&16KrLGt=Z=sfU_A5}3ORHaMosaMk$1Jr
z@7kM6+6v>K@_G#G@+^BLu<PYE#6Jqt<Jc%pFM!|sg2e@tfA=VdqPee*21^g?ZU2{(
zwQ|@!HP75q&JW(9;hTmV76TDM7lolBXAGP>@s5CBqo}Y9A@ysV-=qE2cEbp<>Xi%0
z0rHxAYdHysk1EcDKi>UL_dq-}T%RB;@=Gs?vU|rFDi1U2GtBcHev#0LpZbE;nT)iq
zS<Fd9r4!LkII;y%|4AZF62+M(ekh{!fin9I8)U3*7!+-rBG!&N1ol%x0+d3!{9=Sa
zD$P?y0vU!obtu3E|1`*t8$w*b{H(=$zp@xX#y0C(ntx}RXY4*+KkGF2MDA7erku9S
z?;8yy*T?9#toR3NR9n$`tr|E7-r1PQZsL_L1(yDE_(^|8*l;^N6>4dCADFD@cL+Bx
z^^TJ?lWUrGh&-gzF)C;&XzqC6io1uaJqE8z<s<R-`4316HzXi;p_iOX5h;yKJL`8x
z`v(Pws_Qq^pSpaN2B`O+=C(EadxjL>3t~K#zr5f4SM-9I%!7NzkH3&06`w%1`+=~x
z#)JGX#$d{nC?Qen`^sPM$FZuKVLJ5tu+r3)9HGRht3lg0{Xfw{ty?FF2d<f2dUt%o
zGO|)L)<>O8PL_mx&0zoL`C*HgcH^-as|2_P7QT?W7uG;w?GXAEr3`=#+A>U)9d)9}
zI<6g%7%1Phk&hyu@!NW1F~1*kDR?~8*#NmNRGY84$D+=A8!JVo2%uwYW{9(Wj?Hiw
zpfQ<h^ldiC3l|HjSi0%e<`o&ZAF2#otTRv!X6Iy=FmxcMu#5CEr-VT~VTj`n6>&NH
zPD!98%656_Y)WOIGl2X{a^bxzMyq;*uX)k*{H3aQu+p7(dOVV`T#rm6I36)?oQs^2
zXw3s6pP0(plsd``SD->S$&9c!N<AO>)jjzwoFxyvIH<yoLiqRg-2O#lSQj<93>X|s
zW=pB)P{=5-=7f(cy#4~Zdvkb-`Ht9ex&0|Jl_&klP%|g&d)(+QI-QbeC8b-j&V_uH
z#tcX{EY3*UT6P9(Y8Vdd!;?{FV|t2FhkW>OQEq>iQyXI0F0$j@0H<-<Jvk6CY)D6K
zTg>I-pv^EyJ1Ai8Vr}l`utz#kk}0W<ENz*FaD<~AE)U&jbx&pi)$mYKA5~{Uh7`{h
z$?RygUYn0CS*w&gafQJ<TCOZyeZA0*Bc<`e{#tIhnTWLI_)mUy#p$SLBC^_pL_tLN
zQFHN5BDQB}?43BtCq`;=?6#(J$URb<Z{m|awyCNv7QV0YzO!l<HP=%Qv%f6U!CX^l
zjw#GYMUAB|N{-hbAAn(-1Y_9o_EzN8>|k$ET~f0Jtzt>}QO`z9D1BCt>Ede1;kuHn
ze!Q{E(J9tzv#fG=p36Yc*NUO7onYg|#UAqWQ6Fh>PSu4U?jEJkHHH7wZB&r4Y{!cu
zdC37Kl(5sQ(s#E>QC$(dc~jYj$@kB@l2p_4+=iY&I*5@H52xn<5%L1;|1nrOh5q9v
zHUBvj8m>QW?z=+iVnDphy*jhMQL``g4MD#9zgLXB>NdvdCNRcEz}X=~5^4*pW`?if
zM(m0~=mf#>l6?^~p3S=tWai5RP<LSCMT0(UjVXRnrtb6+^lFWWXI^FPRjIp0&i#@8
zcD}FKGc*VGciTI2=g8Y&a7^5NCjGA#Kmsug@!&1IX_P@FWt>2G4He!Y$Z~U4q@LTc
z>XuM-fPxLV#HpkXJXMb^jjg*!S9BlsXo;G@s1n+ScxaG7YTW~-V~cA3Q^`~FD=<mr
zM^|WUd27c)6ewF`p4RX?ApG00ihtxG+bVR1(HUbASJB~Es~yc>_F6skKr^+6t{+lT
z9@bCtt%+dXeW@X9{%&Atp!xoevqqc|tuWap<$M;zQS3_m*SIt|c0)4!Bzij#lK?b2
zq7_(ybWJq@!bj5zW27(3Ey68wtDqWaW_M?id?;8mIknOsgHl-N{2sadqv_ip{%=;6
z%ZgNtOVdluuNz~L2@C7xf)q=-P_!}D_~NXC>Q8d+`g}1yvf^^)r4ivRe`+JBDEg;;
zC>7KYrSTNOm5WI<df6W*k=E-u8l+C!oA8*sI4pWe|Bb?Nm-<f>j+0R%{H33m#}Poa
z=D3_3^|jeY!znsO;6F^%Z<d78dycEvk9-JNM>7+nIbZSh2{F+qe8Q4`X#UjicG+Jg
z-5Y~tyw`L!hMt~xPt>sei^7BfotXZKK}M&JG|Jc7iwjG?zWUEXoUJeX>fnYSITFRI
zpA`Chh&{KHa%M^pKP<w0=%MU2O>waJ`WEPO_1wF}!oh`kra~E|07o<w3IleN$=YjD
zV$xA8PKVXzGd$}M9r;JothfI{vgL~C$I>w4PzXZ=J4(fkJ)K`!c(>#GrF+BSSZPo)
zExNAnAI$b(`BS+gp-M^&ekiDBl5a17D}`)Kly;H90@g>^=YWJ_MtjDbGQGBIQ3L;F
zOhArT{4;j!YVjS(GEV$KdwCpGuJk%N+jJXAJG>>g*S757+RNj68HvO;3fgJ0fMc6`
zLly%=94IvVX@7e`++2`+o@m_{u`A@us2lp8pf?SF!Ed@aaqDQ;jkG*f_!_2F^B9aA
zn%9twX3u9f^@^<L$RX1r_fy)ze7^0<+E@0Qv!g@q{H8+~RYAQ6FsU#{@H~jA8K<KW
zrX=BlyZ0AuCxk0hzbu&q(^Q{r2|O`W+NV&Tm~W#-4y~9z(hb>!2O3Jl=NRRJ<I~Hc
z6G}r{=-E=i6TsYzXVjhp*4#W9YotxHQPb0Q#Ma5FvFgHQ|1Is4{~9cCV5fb^Zuft?
z|EQZ0NaDbomX-BCkFwkiZWi-DeXBZL`J*W`wxi+rWI<JcWF%1!@#CCSjm!$Bi8Nkx
zsE`g?FOEFf#jkE_zxZNW2?UFU1$`WNH89+fvyuxox&F<;b`XbYdA`%&LZ%~j16FzG
z(e{l*%YjAsazDGl>bruCV!MO84+ESrxf$NIq#ur81wDh25jYXwr|7BIyS8ccNiY%+
z8u8IGJS-@FPQFL4Dy4=Kl8-5|TkH*7H+^)^oW;vD6;%0}Cr=fDad(w++Ba-}YhG5{
zEP>_3{z{DwR4_AbMRfhvZF@axgt9{drga0tL1MzNP#r{+8p&$8`N>$IUukI$Z029b
zKFRSQIR_ZN;7jn;e;cjA6fci5-jT86zF|=-tG?fz#?Uu#!$L8IV#_~(Qj)x#nFi5g
z4qik`t#Vo86$_0$ck}&eRY_P0KgZg-z8)+8LEF&!y7I?IRMyPQ(OtdK_(CpIo>vJ(
zPSIxQ#!2x<;#Dc!z6#}MPyb)>NCs8?_$W2<raSHBK)T+NQCl<iSh6n{U#RMzf$LeZ
z3|KEk88W6`LHKlyD#R*<h%_rP&oCWWPzz|4^i@hIINQBD{kU!<JJFNu{CQEC>S&eN
z>woQ0yBIPvq@o6CR0}IRvr(nr7{(Y<==Em+-_9nSe^Ue)q69?yM~}K`|JL*yGctsU
zl?G?B&Eq8cCw>Qlm!;UPNgq;yQ_Vygz1)w6v(DkOk}g#TRZ;Zs<lnXR)9_NsycF2v
zk?`-y!;t3XSKcnMqz&j!Y1~$ha>jPi{!AC)gvKqbz8V=GRO0<@*>Q#^ju9aC0_;Cf
zWx{eyL8frWIX%QxSIu&=id~+Y9pYMFbUM7ef7Py2Q6{!6+CZ*~OoI^Y2nNoT9C0=N
zPT0}}3ym*hq3QHvQ<rD!!^05@&BTv&=UxL}FQ0RlSW3zIrZ*PygC)at=ZX<mBe{!u
zx}~3BW8!%8DGRpRm>1Z$@neJb-`toea7MWBU8nBfwyUJ6q*Y4f8igsOiC!?$I=@(U
z6<$f-DX4sl3Pt>QJ7ky-k*%0Cculo2Z>|EEQ7M<iicZ^pu`HpqQqBTnwyzgmC&{%j
zoM?E7o$I9JeLLwxV44k<ocRm*sn4E#vGkWlFU2=VONdt_1S{=YHEM)ozxYD7Z;_ma
zO!fd+npo6i#;t9-e(V1AFnaeZe?d_n4BURj(@_}|2EHwP|4p8quOc8lH_LHHQi2kP
z{$0ObnTq^$!Wx&So#aIr?W9xh9g<=m%{^ZLcTUX9OJRf%jQHbvoe!_?oNDDma9IyR
zfy7lniW)qmwr!_`oPt?H(a(!yKE>&K%O8~(BewO4VbZ0gbo;&Qkluc9j+hMMqF;=<
z#3;Sv*zjUk)X{>e7P8$wc8;U$h{{GSQ=il4A!GX*D_6+&f61`Zux!`n{@kl;524*x
zC>(CQMQWy1F^g|#Xs*Uh>)IF`T~tjINW(mb8QZItv#ewRsmq)QmopcRr{rr|e@(=s
z?|St&$v<AtKN1v2lOiXX5tJcnQQ(^T)4uWYL+BNvK}$kw_R)9!?fZ_Y6Yn;9TJV2U
zssanqhCSCCuTTED8}7uw;*sMJoQnytlB8Y)>N@;plWzm7MWV&-{&Uag%gcK!|3r@S
z1TfJHXNS&4NC#9&QLoW}^q5+g*!770-XczZpe&=L^Ox>ei3%0W?vwZb(MHk;roiG<
zC6E(hES~l1IEC&~7UsQUGY>^&J$%;PieZHJucQP01Gj52(WVm`j!SJcpW}!1=0b}x
zE^YkMf2Pu-V8HCc+HkGe{joRUKEyi;TQk3+!Rivqy8lL#wQNppHdHlc8M93i)=xxU
z70Z@QU(nJ@9zP6JIm!L)E=$`&PsT+zs8Zov+Ve{C)M4D#62sftwUH5fM5&O7Saj-X
z;Hu`)erfp6H$eZvwaXXhN+K@U{bYWy6q7F?na{<pRL2*><fR_Kz)-mf2FS`CyVF1A
z<l8Bz>7S(b8)iJ^Se&+QOS8c)0nGF+Vty{DKE-p^U&~O<C>F8qwQa)=ln0gQ162KG
z4Q|Mw1$)Cot>=EeH_)8^r27!nc7ZQEe7r%D45ly_O+zKn{*g#O_t#Bfgs0Y4RwvUg
z?7fPe#|ySXGP}K<Q0kuOn@);yz(pCJ(bn7ag>A*e|7{;$WdAgu@dQw{C7S$?CJelV
zg4N#)qS<k>oo508;i@J^2knOpi@`P*ot4$isPX$z-@cg$)ip^uDLuYG38=+zQ^)#K
zGs(&3#W~-yZHx}aE#*rSlmBf7tECpKnbjI~yC88CjVyd=<ViwQFr<bvXuudLn(_!A
zy?sfgpwDF}h>ME+OM}P9JvBtz7)Bxo6&r0_MQ9}V@gABDyr~;n{V|keq1gKlgx$N4
zK4q`9Y<DhM=wn=Fs(;=~G{v@Rg05=%Zd_(3F)rUe^bdG=Ww{w^gf3Lx>*FU|X*ory
zm-{o8n=@tHlwulu+QF9Zry$HE<DHMj?PcgKNF{SKJP9DGgL*U<7di6%r2FN`b|0K$
z{<;K}m9>D4I58!gzJY6gCwJcV)t&GH+>?^Y+Z+wC*^8{SkobJcM~|MY9Iw#XqJTTp
zZs@1>n**bzj}pX-Q;*(b$|qHarT1Cp>c)nkv?PL1RH=Sr3G%((n1c~CZjGm7XGLOw
zcEe|QZigDC1n;&7q>?7vk)ot2gMq}0f`anDXh~Ung5t5_(z|Np<uapxO4;vkV_g@o
zJCw0gWn*Si6R5tA#ck07J8!@}OP|Us(MHY>KI~1r)P@O|*?Y6?W^ez|SJaJvK?}!c
zJ^;}`G+JinK4x!^?E3PAC{Ez~OEMpIdC$)E=YJ3VOa%NT*Z0M&BK(v4AwkGBF%4Y%
zj_IIas9_)yaVG3Q`wcURBgFpRlAf4w&PK#T78!^~SY7ml^gu>1rwFx@c(d~!7G7-i
zer8h3%$1}>A!p=FK$N0Fu?8%Gr6!E3zbJmyw&K(28a#K&(T>xCCAxCnEauIHR^hnl
zI!ahNrR~jx3tcTt^zG1HMB{MIsrBhWgz0+j(Xr?X=?Z^bt(AVaCgqQ8SCve-Jw?U_
zceX;1Un)}|T6cm`lRrJVfs#hKvin6miG`i{HzYKIgIQT0VKgmDi%yUqf~fhbwASLm
z6d-<^d*9RPyR}OBo=R5T9>H&D^0|mblU-5Hg7jeoZ}FE@tg*yx%*RYpKh@o$O2g?_
z7k>nT<_aVM$Je?zu~s@+JRp6s_*W6%lW+=iR$1|h^N2l}s3`v|>KiZI7gS3j^tq*k
zih`n)kV0I6aPOd561W2GguRi(!;hk#h^#z8xYf{h|9+ltCUa9gC^ciI4MRLsiTM?V
z;gMf^0kwfC@O)k45AnKZP7Z|s$8V^M2I0ZHA!1Sg|BEPd=iuz?pN`LLX*x20eg+Ys
z5|9Z@;R;73Y)J~<3EY3Ud}?cgN3cH^a!7tzFC)DpwkeUZ)_xmkio}$XwQ~H}Ly7|Z
zL4N&-+hxDjH<&ZyJZtC;{miz<U|zKX#yI4&7>XH7Y}FkcUzFOMo<o(Rj8_w;vu6rx
z3*a1wXtsn&WD~Xswdv#uyiaudEY`=qQ*m7s=GU3#hc4LqJ-al8FMgbTQDtzg;=X^1
zxzWLs4?9L%QaA?m4IzKi%V@mAk5MmCbtu?;E}$*MmUb_(^<RUy94&{hu4z(Pt|)m`
zWzPFV`@P~YYur2K5e!?9{^hj+KE3*SJ_)YDO6C8S-m|1a2on#dbji=nO=Ke=UY#0z
zU;voe7UbMy-<ThW<_nO2wh<jIWA2>mY8X_+1D(BH;iL;v`>2;Q#>~M^4KqwX?r&_=
zsI~KZZ$(MCa?oeVlfa*^UNR`oCkO&ky{cV04ysrsn!z?bF4$3-x`SVI6W%{wb=Ji~
z>z{9yzmi+K4eI3&E*epIUQkk-4hYY))H0V;sK}@+wYy!7n-|9ElB~j43A+>Zp4-1H
zdX1hdq^^r)!blBxCsRKVMRjjyWa9fXpe3PVXyr{FZw7xgy7+lq^B4vm#(wu?U`}WG
zK+ORS{5p^Ziugzh_-rJNnm!!PLPk$s2UX1@;Kmv-RW&n81-s%Os((RN6>DvMrguSy
zrlKMzzw@kuMJ&%h*3aLhlnXw&0>pqV&WEi@o$W#G>N(<jUp@feqM-^#9mR7g+Cs&p
z`dK6iZAIDA8H2?WuQST^qccVoR7WOODZi-{=IKe*vs)LbJjSy9-Ps+=70n{YRtf|f
z9Q=-3s}O8Lty7_mCtrb6BELa3iIB}WCMaR166KPEv(HDEKiCmsajnw3imaqNQF=}p
zPOf#vZ(;}c%C|&0#Z&lOLsh0KiU*srYO7IeWB>_%eud9Dn>XB&6P<F-EN8URGngRG
zUg!0JMcY6ynu&8XTgWc7HJFD{m?~k2rZlx#uIl7LcPQ{Gh~!AYiwzquOCh~3U?V*A
zP+&z-7@}lbEkRMF%g~d%1Ny!Vyx5{ZKtDnATo0j5Xa7$`Ug|w=v9XMdtU#{W`NetQ
z&+hTHSe%?)3ECX5D_+C$SRC8k0sN<*TmK2jNr9EUK|sBU`2U2$d3XmtBL|n&CBPNX
zD^k&L%wOdOke928279MojbRR(tE;-6=kkKj-)1j=$G|UDFD@s!9d9e2h>=lU4BVG%
zgxEx%q>M5irad>E87#Et!A>^WqjzaE;ia2Ub=EtLZ|2h9fvP^}?-k&Pn^2|s4n`ju
zvEY}@hykYu=c)#!qz=Gm29>tY+%}=1wPrjDS5qUCJ&A5BXufB`xN^)>xG^|GND>4Y
zjW$(PrvO_1F0elO_ca<_4b>p^_VR#q!7dLsGfw;Sx%?uJG`w@%6T>sOkh^?LNCPm<
z@vdSJ6&FV{+?f>>)Vvbfi67W}2!)o%jU7vGf&u~nY_Dvq8=C<1uW;=2O|)0=u{2ac
ze_G#_E6R(_3f!s0DF?i4Y`UN}+S3IZJ24z6+2T%`>GE*ts+e90ex_eCGcro*s^WuN
z_?!(CPRG4Gs80#)7Jt40?<XP4!}GrMWnrnHqJ4n~4<#psqsN94Kkd-6tMM>pqywfN
zDJVp$<b{P{Md5p+^K+?J<3OQFU_!T%VvFnnCr{TpydSX73#6U*f$ubA9v<%d{FW9;
zV{IF;Q;unf<r<=`!GlE%9#&Z@_f^av@J_Pc_k>Lt7uzoCjFa>(A$}Aa*b4b?5-ri7
ztg$Ln`Pbb|N8SjRask({II3D%qI+IY9QJ2TgO`+?lP;ep1e2Vh%Y`fgN)Kt>mFI1W
zNjYAi+P#rEA;<{_lBWncq`;iFun8TzVoRMUQF9QKmH#pu)eL3rP`@$n?3TFgsSeW@
zwjAa~5v`U4xV|}D7r)HqG32|bMb2(b_0^~V=ULIh&bsZzUOzt@H>Xz`0XokQcFX4-
zT}CNhJA6CpH7z^dIG3M2SZvO&1Ld_g!BNkMkBzmpY;R@&7%kE)%uN=3A^0igF|195
zc4Fw@S2n0J<f`vDW#9Fnzv?y^e;6VT{4TjVN#6Hd51N`KE+X?$Y=-JX6QB^4#J`z>
zsw8QRR{NqP!;&b6g5!yJ|9i9m9Z^&nmlMJ{RlzY>HSw5vem9FC+y}Y{0nw1b(TMmV
z_nhqZP1(KqaYLebLfHqmpPeSapLO(R4rQtZ1K66V-}kfUZgYM^!4eQ=jP=c)*Z$iE
zjJ2P?&QVo6@evM?)R4kLp@<vwNm9kg1twA-D{O29N<_}6Y+D05`df6hz10_X2KI_4
z3@FMN^SYD?<KiC`0-dr^bV@{!L<M&QcUqX_VnI#wwr<7HRk4P&I<01v<&FX6w4Lgm
zarmEYGW!IUlE_**fT)+A?ngTMYh=wV;Vt=ct_DxcQ=eC08m-t2kn-s{?f5E662IiX
z^38SCyacj6(8i22PSoGBatIkbmCSM9m{Y69tK!K&KR%nngrFPCp2!;Ibx3C}{LsG*
zw$j(8^U+b@T`{(U-1jU=@1ncnbblLJ@OQEM#rZzCZ>f08NiFR8VFPWra3j|r3#Z>j
z-OjnOoX2_ZQ%_UW)Cl1VI4u%$x)9#-s%`hhnmeNuP_p6PTWDCuWAy1Ls=B&q5fTu7
zeBOM4+%S52`?X<YbaB)7_s>o<3BeST&EICuRcy?&gVrhFI~+gC+t|LXrEQF2MO*cT
z(O+N?b|>@bNfsXB!ww1h@4v3bT0h7<IQ~9>3iXA*enqlPtnBMq3lLNFB|(V<c=Iqw
zNWBpW;zn-(UfGAHR-mLKy(}q5byT~0)mHA^K1^m)E6Oio9Z>>}<c#ue`CC750a;D@
z2c9xOm%FEZ$R1G3PX*Y>ELK@)8SBA=`w4z+ju-9`yz@Xui`tps;9Kuk9;~jB`j5>t
zJT@MRlSiXA{20!zs=mEq20I@p+<W3xqOU7WB4Oyg4-$p_%w;=9=AFIZztT9b%?1#b
zqq})<f}qBwV2^RMloZ9}b!Fi#LV<pPl8kABLui4!Gw-DjPq}b14~^E=q|#MLNi~bj
zkMaH3Z8fqhj9<n>2VTR)Z)Rb*=C|x^_<}ItkrI$`!_S|-+C13Fj;@I{W+^8O(=fc9
z8XrV*GwFm?pX4TV7*4Wx5tH@NTc0dsi*W4Kf>!`^pYO_f&559<DXPA199_*pPwyL4
z(@~jtbTD{`*=?4HcJKU}hL06f{Wess-cZ=#BfxTewnbVjzwijBPsu%WqJ`l^A@tcJ
zYB*#10x`rfqEHnGHSsm`hCfiU1fH+W()-QcK<Yd-U{+gE*ZAcVBH^?a^aK86R^%<t
zVv9NV`%1l>F85vg8mp%~P`yNRVDn8R<zhErvBK!UZbF#iE0|F64rK#C5izH}wbf7U
zi<9gD7`lYrE;gVI;0eD<ww8PVifp@DF}?6Md^wVk3JP73A~S@-d9%WwBw5m5iAAZ}
z%a$-+OIm~(E?`A7!~PVUKcXQv?Z@2lRI530PaFYa(;gzq&uJvCZXl&Z(FkZ>&qjyp
z;P-vWc?7N#xe*xk8WDE}qr?s-GtdYLBbf|ZyT`|Q`eNGg&@Mm>zZ)P3gU56!08m<f
z49e12a8Za7Wz4!Vc_?;7q4XMM(!#R@QZ8TKxz7&j6F6YR&^RQcR@Vvj_>Sj7vWs2E
zCbvd`y*PqUF@LV^70#BXpaO%YkSsqGNs&L(7BI@nJdX-<Cn!Qyq0vhGS?K_$nLYa>
z2B%G9?h%bz;t8X@qM7j2bhU>%O4`5#7KkIk8@pUwOb!;e8_J3z^je0nhtd>9^*B{>
z@!kC`=&HQxzlf3QkAdhO-Yl_J;7C92znT>uF^}<U=07->7Rj&VRW_VEQiZMTFi^>r
zhqv~Tt&Z(+a>tge5cU!{N`-PktEY>;*fzr9H0HU8T`w1~a_E22cxFg{KjYYTAF4Ux
z<M8mCG9%Fp=Hz}(x)5fbG2A>QNo7e-mTTQEYVXYPj0+yxCyejbg6V50c2EYIKEN8-
zX30{fu?dOYP)MIHuiG+yAo@FQcZ##@$5ZCy`!!?l?u<nEjM#i&97s`xNYJVW1rNB)
zfCtjH(C?KK{|Bw=ET!|+&Ok`-UvXV?_rC2Y-|hlje%dlr9~7k{4h&o!AB?mbK$z$#
zya|*5JT!FCPHpwvFgDiqv%VlW33>u9?|L3!J4vqJes7>4jJ{p6ws70ytd^CuFfWH(
zM}?oeu9Xa9Bn#L^a%?Etw`S{VwE4hG*_2W@i;+Tt5GE#f&|8|mJ~`7dWky*P!MeeT
zf1q`!lqOup9oImjjXzNrMHx3wUqt%-lHfgZwn#%HB@6Wf(Q<A(Uo7Up!T*VWZeXOD
z5WbJ&8YTtFfbNtW<hw>mks!X>EVyw!HiJ?{6-8`NuL7$qvm4XO<x4ZKW-m(AdlO{Y
z!<m$!UJ!BDVJ9Hciska~SwBwSTQ4u)DazkWJ0z{=WiFzAr6rXO?)=(3r}7irO%~sY
zPxZf284*QGpTlxIrr26mcx`j0=|v%$Vk@B6td(58`O&b0zb7q%K_@h*Kb7o5E~wMe
zrysDDuRpneX0y=PF@6-$Ee|1WUkXPRQCHU)l>yXzI~mV-80wvv)iCX_s<75+M~pJE
zWVBWoyN(*&ii6Z^bd_rCI@EtrKj%q1Q4+df#8~P6JiZn=Did*@_H$5-DJV^}k?dCZ
z&VvCd&C9g=vQu&|f^DO~TbgTMNn|murPf<|NaU8JvNgmELZ*r1VJZDEZ`E<{frbl~
z9><}-Dniyc)tK8oMtMTGY-)g)_5ybBH0cOW{a6pFZ|_GJi$$kS#Rd1>O@F@Lqeqo8
zdFJw~j@+N>OeICtgsik*5Jdd${n14H`FV_tjDO0sn8(w3F^GwY^@S@%{MG&XoS@kl
zeU$Y-SBf~${@EYnTJ$zyJ{g@JK~wNeyKbdq>)N|b_(7QFjVMzDg&X||52o*^#3<f{
z0($6x2;3w4a{4Ii0K&c1taRb+CyBKDw}vREPFPVVOsX3wy|?5nq9Y#g&*CH^Lm)O?
zy!oCwF-S3R0GoIj%k(UCw8Vv&y?|aSL-c~2L_-!N@y0wc--w+q4l3EFice3bNup(0
z^vfLea^k!T7!ORxo9kwbl2C+u)X=*Z|6orPa*7>Y7a&E`yvgtluN_M+BCWPK!75$A
z?a_iN+>4frh08lPn?dla5uG8v7w|P^tX>N^iSCJ#6+`9Mx7*7R&8Xx&zPuhzFo9d+
zvb*7x`Fi7txam4SgYT3~?|r9|CVG;a4`<%)Q&yVmpV0MaG&;&^g(8yZqm!(em)t5(
zQT^0nj8U3VoVcI)g11Y?)pVPpzY-!HXMxVQU%`R>rZMp+&jL@p;G)rqYXz-5)d?&c
zP2-c|I9n>spO5=BX)rgsnasP03s!a%35*Q8*6t@n_-*gICn7CM*z=DE;!UT=6H{X!
ztqD-QYl^KCiF?K$J=(n1<_-dlSc^ir)Q#V+owvMZ3Zp4(pgF%*o9;pGAUHucLx7*>
zC-BHN4u?p3=k-fS>vc6Khz?YFsmp!S2ZvYhIBmCIsv6rD@mVdkV|mL&lG0o~L@wO1
z_c%e3gck65!+4mYinv!QQ5|BxR^v7~!dKoysnwsXr@m#%ffo<dN&Uk`4<w{x|J4@8
z(J#Mc-79GpBC^XA$O702?@kC7=V%2VF8G45A~fr6hI4El^T}AaxswJdnx9|0g)AqX
z*rodBD@*4+ttY<#waIz0-!9TT@chhZGh@`CJeB7%S9gz55iLbxw5klrehhtk?Xaoa
z-4v(d35g<sa&e*8Y}c+lw3Zh@$B5%tz5I!C?my|vXQgk1hXd0G{&7LJA)B-?GN}iH
z(l4Z^#<fz$GaUA|Inh@OR~^$&<^xPCjuDy4cXSCX%w=tPNL8a*j8*S^SwE|PcW8}c
zU#*X=u5BEh|4CUa;iQLIDwP};)6(*uCFB(@j8ksbP2K{Hj601}Fd>Uq(8UTq;{giv
znSFzAMT!iGIBwUU7lh+h^4JW62b+LHWaW<<jfFJG-<Dj=mjEIk5j;NVWr%yx5g3FS
zL`%;VJB9=Wa!D4bRh^#HRKN97mRb9!s{gA6z;bymZ`#^AbC`S2V_08#C+C(hVYXe$
zl;wU1Oi8({b#&N=z4PF|LT0FwcdA%~Sb5I7s5X>mH?1rhf5?Pv%hCIoT*7Dm7XGDE
zZ+vjzoQ&c**I;3O1|Y|>noGh##8Wb7?XX++`cAnH8<hgwO`#o6;Z$nQYU8Wx$Wz54
zuzufNEpbLkVa5Mj%vepHIr{Q>?1}MS4yNy`LC9MidwO-%UNEs&tmELJ_4r#3>EuP0
z;EEH&5%pN${pXC!kpyN1FU0+!)Z*^G5EPXb7xQ+XC)Ja2;Kv&d)acb97Huek6K*A`
z<gb`8hHk9k8VK9oc~Vb6^_)^sWd^vkIKWZ;L*qyw&&Rd67^pa@<boS*OzS)8Bwwo~
z@|xqBuI(x60$a9{$8?NR5<?*6z;ZxmqfS5T(OMPEPWU2nWudD1jPB#H`$8cdtz=Qu
z^xoL;DaIU4*rkV$?|?S&6;Ag%YDNcQ?%2R%JbSw5*W*Ov=AUmT@Ut2`DCmeDyCD?l
zh*E!yqWtiW^|@6@ygt;g=hWY-&bkUeBRwhJU>WmJ+~qJsWUYAz_s54x^TheBSpH)E
zQ-TJyaGO!+WVPQ70Dw`+kWQW;OO&jdDP6!iI+g*w4Nv^hSPYFvCd=k;5po;#fFzKN
z#u34j79~@;W}g$1B<G(aL`$3Zu($ZKHGdu0{B%cw7FL^viUKSn(`QRZw5l+usFc)}
z23`muWIs21+xAh<U#<+3TQiArr|b|29a3;i9)sm1d`dz0R!B2HxfqzW)$2G}Uy7mZ
z<AOPL=U9sWot`RJXAcmiUXVEV9)C>45N=IY@lADqtC{wiJj+n2QIyb(ZW5;haoP32
zv!r!i;l8xgJY{9vN%6Kn2!1Xm!=3a9u2N5Jt{yb%FtNfnF>mO)nc~dYh+GohbqwDx
z0;uEPSuLv2-Kx0eZQ9VdlKmE<!~bw`JizY<lmjwvPZ!`Vuo%_%WN%sP)!oGyPjVs}
z_Oe`kJpqT7f1aK<|7E{+3O&AQ+#~0muEnozkdO=-4uO1+5BRFwR9LEXeq^1vldSve
z+`Q8W=HH)jebU@xfCEx!`gV~fpkQqB+)Ei<Wt0d!rpwARS+JyEzI_V(1TKH2HvHH;
z??dd&k=&f&&yo+0(6hIGby>_VtfLrI;pOG0rzz5P<W1gabo-EI(Z3)Q{8Qi|m3JH8
zrwf&dNVr5&Ae)qM8xz$TUM2jPIT2>u49`B4UD{D<mn~01zftrhm>pC2GeB8P4XO|R
z=TS6JSC`M-Zu!@Q?O!TP_{=ZFoPMAKb}ry{Wxwgp=!CSA(oh5rGhX??=yOW3a8!tX
z*!_jT@a=R#mcCKIUhC~qu?OJ#B%VJJjvIZL)5%OYfnFo7BBXiIR{L@deYK0Bol>FF
z)4r`y^i49ex}gk$-IX1aSyP|lT^-r;YITF^QaPxs|FK`vU^=?a0AyDo+_NT2q;=-w
zBY2{RT!hXjqiE9WQwQ{=muNHDvrzUL#9=lxKINxSP617G*ni~KrrRdIQ+*R4FCWTg
z{`<Jgk`6b9#}1&cqMyq6O!3lFUazb}@jmMIdc9WF41DmHJokL^%ni3UPfp^rOpEpT
zC}7KmU@1r9p7E@GUY&bW5CY&Y6qoD<&~^I^j&CeJ4h{^O{*B{o#-~sx%Mol-o4U7J
z$U_>JEn~1Z;{@!z9ufDB=BIjoWbdGy2?huQb;zPDmub{+|IqIzq303`n{wd4yd<Xs
zThFVmU)0K?zot+S`Ty3LF#2+x$pw*^h&}Uaii=p(8G@=jF`7?*tFn$l89Lwn#!3-5
zkJjVVOUwxIqw3^|tSU<hLGDAVYT*0Emf(!5!LSm+mX%SRl@Fedh~?E7bC>b;?uc$)
zU1Q_mTAPPXiRXGV1hmR$<UXwa;!9R%@IRMVGkZPqXLB-XyTQ)!Il@KoI<Lz+UAQ6#
zqI-%fVptJ+iXv1so=b4vhNw1Om+;i~byw07Z^5`6>tXO{F*>GXg>IkCINs!VFXM`$
zT>eAd%anGO$K)4BDm-=}@W*^eY~bB*%+uxZmOQI($b(x@P{^t*H}}W&chRT5qsRMH
zQ0eOxt^-lm!2S;FpTM^)&b1A9g0~I$Piz`Mm*W{8%T|QU(P%Mjzm)?AitPh9)ON{<
zju?J{5YJj>#LQLSUi~Za?5Ixz`0t#NsT+Y~C{BlCVxIS#cw^iUN5$ghxKqOou{>SR
z0&Ny&$PGQl&H6ur^O_~bV?=tVq1=&RjbJ%aW}phjN-|KQqDW><?p08YY3?{%!aj+c
zLnJjC_8x60)lHrHoZ>!Oj=LFl`24CHo0KIr$WU~H`mI<H$fmTUi@<S;XnFc>Y)6de
zosO&JwJL+6tLM<Mrrd7k9Oxx>xq4I(5NyY+y2}aTSpNu9h8a$H+j{OfVHDacO%TJH
zqAy;mE!_37I(^Q2-MFc|lxwZ@tz{V$X1>=jUBD1#0upY&ue_7Q#Y0Q`N-p3%n*83J
zS5{wHdcV75RP@ZOzjKkevsjmKEA=t_-#b+IKbX!0L0|iUBvTMG^W+p}N+gO-;-$sm
zlHV6K$q6*k)Y-<J%6SdV9!fUEa22yNGqnwzYMUw%U$1h7Lq`{hG3;5JZZNX^>lxW4
z)S~>x3glHJ#sHvH9_$U0BA<k7&b$`uub%4MP)gN3bfv;;#qhPT0Ry(Ug6&Ke92P2&
zo{>&|p;N(Sk{h_4NSU*X0gNIb+feFm;F^zSisDZ_NxAs2jXJ7e@TfcaVKmO{8g5&B
z8Q{NkYVr5^K7Qrx`^)9|HndC1_27}NWv%4k!gSuiS;-tNEcg|PK+}Hk+`c!$ehw}H
zkB<u%i(w@>K^W?BBJ(aDGX$@$FhDjoH5CZ;-A=m(jlT(DgBewX%}LxSfl)&?waM9e
zIx@yx|0e>ffd#dzzfO}6g?mY*PT$DqyrO9FyVQG_A@I2MWlX){r-dzv^;WCHRw!_$
z?x$gq4mTzU2q*;dQH?#S7#c*ij1nHxYc|60sy{W@M6;QXZdRF})I1n%@n>P%zp;AW
zeoCNMXhq#E)z|W>>x?<2p&2YFHn#s&LZTF4`gO~EE?qyXyM#nY<6v@ps<3x$Y%Tva
z?;jj2N)cVb-yr_c^qkG%DuQvlC_3p>%DL95M{tX`91qgx;s6BPYQ8mML+vBjw$kuD
zoDV;b4G(t*W2NQ$P&I?DyE}tBdagFwT(iz%zn?3&EbV_C@lwS-s4JSPagl+G4nkmF
zNARWvr7~GKy%h;jA9P}Q-rwRyC}4C+I+d!$gB+-ecx{lQ?~_?v8Ba^cHcRw}vyvl{
z@S4+qs1?tM@hWwf0L%7i4cnV#Ub<XJPViQV)>7EhuO3ilVn-{`;b_*yRy$U4SIsqg
zWW%RsX{NLlvjoX}KGsb|w0@jyKo3NWZC=?OyPAElju6#4-$>#LS<{XtvU7U^bD5{P
ztJb2w(u?Q}oA&stqyxtXwBhRu-fUhYS?Mi#6nt?JxjjmL2rKWWRKBIS@a=j8$U@^P
zwp8J@ex%@&s9m5qs(uhORt)s*hA_hPtObAH{#`QToBAQQPP=(<5JDjpZRIB9u*$=z
zTTkoAH0r&Fup96XGlE9*!ufZo>`^s77j|X!-TT2s`!hb@Hm0(dfa-P++x1wvWBMle
z<5gg<1shR|z1kx2nQ}u~@^Nab87O1G^ml#?8)qKjTp!AdWr!kThIi)0d{8dkpc2MM
zzqK{HM-NH49P>@J?AcsSO2Cn@!@q9>uFEH)TFxuv+6x^x{Yw$P?+8E3MW}JT(du52
zJ)wrl)JZ6z3hH}oTC$gf>wfs&>O9^)Xq?Iy#1hT=3|;`ygBeV9f(NqnONXH^zl}o*
zgad<?&urP1XZ}2^OvZHA2j}7EdRpf|9flbxkvdde&JzsGynJ=sKQ(s}w7?d-O|(jX
z7kgAqk~NuGWh^aSzZGTGe>@2Q-`-5QS>@o%ab_9fErwp-G}6}0fW(_cw!aVN!Nc(Y
zgSdTYZ_D3<d_bzaL$CgRzfX(nJu<SdFC5+!8`MHG3M3`0px6t}=g{3;+C&!WHG|d`
z>8o6}5~wX}zbgOW9qZF~r!&830eT}zIiT26p>4--OwxRfuUFtxTuN-%?3f~jg<fPm
z4jQ7CBv)oa8Xeq=ki3f(;v?-RWCbF`{V$)54m30S29O|7)tbBwJdtih!{S8W!6#9;
zfYqB!{9p>&SdrmgLLflcw=R<=H!s)uIiZ?FFR4y)^_52()IhI|;@`rP+W)&H^h3ro
zlhIleKeNX`C9Wff(K{0}OYBj&g-L*UZYs=2@ba#E0bE<+LqdD`A~MrrwaP;;umECo
zjDktqQSQw>6k4cQ>qNLonliuPWUvqZEC^cz95wm6RrUV7iF*~e`jVc$waY^2kGOlB
zM=&`Hd<*Rx!qdNcU2^3Tu4yOo$J%~9Uuy|5I$#84qqi!?_#3E{C&L*Xo@eRj&sVL6
z+SXiEH<o+~pe#l#J;aoJ8_QwPifP5=HlTGaf&Z|&2htodVNX2~X?@?uMPz#lyefBW
zd57u-2Oq#+MEh4j^@VkXd~OBfl0I2{hS!huTd)CJ#f*7u?^9U?COgLyi&T(X!+T{H
ze=@6$k;PkI7J_ktAW7DMK?A}}61QGGpR@E^0;KLNf2PI(Wj%k1G`p{@*hFR#u8T{@
zf=Dui5U|P_aL23KWFD$LCmm~><wXRUweEdo=y~x*w<f-=;dRsYvOEux;1Gmlb?Jsm
z(O+u%N%p%0KyAZO*-mAlXDdSl&h6pdoyeKHCo$^}Oy;n)QBb3pn!9`0uk;b76)*9s
z_w_%(S(PRhbZVJanGE#)t*);%Rty4s>%^~4thtKqvxrLNG~OYv9$LJ)s{qb8UZod=
zc9QNYD#OxuEt6OvzOw>=s=al&>(^KIPquBlm;1*$NiBags%yBKh1aQ^?C%}N&m!k<
z6HutTD>@lY66e!aq16a<X9WSvB>TXMNg~dC%5w0s#a83n*R|NJ>*ws7%H2CWBC2B~
z_A+(XOrU+)p{$weH_#-U8k*W*TkGqvp427XBCCuIVw#XO(v2CzsSvf1rMD|x>~3@R
z<*+-hXy`I*b;29NsT9ve2&yZw>l<anK-DYvbdJ9eem3E`d#Bxx&w0&L9T>x|RNW8@
zevo)Mw>MKZp*C#y3A<y%)x>D!N(ZVlC8yIfsMml0L*@?@<yAA$88+c@Va#|u*Pe4K
zb#MYw@hBnttCj24Uo9+d?6T;)(ZI_8HU>}9geMtN<D0|@M8hmUGhRe)i)HCY+22*x
z&z2ygSJ$;CxAScsLVh|CokZ-t8j~J(?SzKm0Y@k=Em>%Ebads8pKckwgaV$ZpFe-r
z=Ky1K8vn34E)E~#e`C}gQyP#8u2oU=d=&R&$TrCfc5-8AH7Bmki}h4o-r*oj@fC+K
zOI325T#c@^OX+S1TyXN~afUggCFpUQu;d6yDNxR3rHLA?B7WXSJqaL8fmlZ|W=ZZ$
zyyAeQ!Ojv{75dy->s}kY;SHd5G!mle`{P}y+V>oI+-K@HdZsX>1HT2&^H^{>kJ>;T
zpPKF7wj3}E%9JEk>koj_7xu{@Qdw)_OI5TUH5RKu<N0A0efh<&=9z$x=0<~hvVt#U
z_^vle1*(w~e@)t;R&7WdbkOYXVct#PdJW-Z{(uFef=?jFcYgF%#KZSM>za|L*<@?L
zfxG@`Kn~tll>$x;YCICJ2vBebycT$B*+j-+YVJepp7Yc<<$)<){absO0D@X1iP?#m
zvkm^l7_c>}4T_zlerU@RCYG4N+d}%*l<9DZzg@s;R3T>$QPcJv?JDBa2^!1Nb{}@N
zb4@Jr#H23wJ)x=V)Z$M6H?{K&1Lwpy?Z0*OD{58_!T2rT&<`cMnK%t>{YR@ZU(YDN
zd-T_urt9+dSeB#6g1#0soy@J2z`c;kL-;`o!HL}M9~-?#W~{6gw=KS%J9L;bL|#oz
z{@gymRD?d9SooUOSRE%cw~v?-Xq{1nH93OIdYn`;!n4)~&#W4BFmgOW(Ay{k%;XT|
z1Hge{ynid`KU5D|-ye+p@-+SYVwev<<rb}Wkge-9?OQ45n;I^J(YEPqgOFBL6+<Gh
z);=dWn;V#7)&80a09^BcdYp3k@jjF9)>H6~;<9EETyb9=FpuH2i&sSJ*=s_8NWYYP
z`s`})vqvO2L+{sy6g)CHOLtCTV9yB}Y5iuK(Klp;Ihe@95M*-79Q=)Ww-qYfU(q=y
z><7$(J2cy-3yr?xc+t9oj@<h!;C5pw+?X=(Avt^?=>j+uO~w*1n!mg}mcFI_Mzbc{
z+k27Z_X)>_#P=a{lhHy%7)E%w{^N4xVb0uKV}S<t{e$bG_ZbT)$Dn(bYo1u~>h*Mt
z<;q6B7>fUmAFA&Onp0{7eSmY{QCfJ~EAp8%*>W@lV@rQXX6Lm6XrE-vgPmlZ_lHsW
zY`E|=YdFSW#oYPIrSH%~QWiPT1O1}frGBBZzpkc>(}YF05rrdJnAxzicb`)a7mk9<
z9H1;@N9E*rXW<^2R>T1Jv(!rMTNUx{>D|z)d0af!*qDR8_I1Be*iekoc#>aXr7wTZ
zQ1>~^9o6B@g+`iodkJqF4=YZItOi3$Juw|Ho<!)oq9>HVJ^mczns^Z##eAK>2k2_V
z=MEul-T`tzV>Bc7ZEHh!kJdu`ql%5)`P7&Pa<+ff>!mNusT7<%IHfmlvp=~Az9E6?
zl66nightC3p>pg64p(f99BtC0c_PNRy~%3$!0A%o>>i(@yC#`femlZk_H>)hT~fMM
zL$>PsH|OF5gQ?B;I@2-Y@;9+h)C5ml-Jvp$H7><USwejzU5rCl)$yuK=<PZ=xuxli
zcEGrygWVc-#6wK~-I0r!kQhg?s3$6F%Iya)psY7%k8$2Q@n>!&oX@Rn<bcuF?AmVi
z0BtClBjOof5uEtvHm^78z|q|Y)R`8x9-Hu|%y6CvFQYf|ioGq25c2Abu1^z3c)~qN
zP_`r}{5b_pi^6A?94IxcT0MqDxTM7w08hSGM92sk*VZhBpB{v#HjCDTBp4Fsxv3Ao
zZIE|wi%OcV0lzgmXP<V1aBFXwgU)%ngERiVr)E6jFQcsAzO%8-pw&84&CJXuzsH9t
z^!YIb77%WK#Bn3|yqq;XZU&;%^LaQqb>E#V;028nXk};9cbhZ^;NuA~fl+f(J%n+6
zL5#Yl35x?=9i1Gx1tuN7jE<5?uHD!CpsjS>nIsw={<Yt;3njveGJC3fJ>;JT91D%2
zGq&JJN2Mo39s``<p-ewb=Tnb!${i(0OzKUi<vRatxpZgeL<_ES$-YKyxj>YjbTX;y
zqs&$g+PdXkpgR!rGpcJ!>uKz)*iYM`FJf(nVsdB=WgnZ8b|tr(rW=+q721`T*vYmi
ze|OGt*8P-mh$ywwh36dqSp-Tz)k`%1;f!FJqAvOTRr-Ii^_F36huxMpP@Dh-ihC(e
zaSKi<P@q_Gw^F<~6bJ;TxVsj2m*7x}7I$|D?k?d?&pFS$GuO=bT={a#f8T5W*4iCe
z!%<)oM&DPq8MPo!r_)f>d6^npkq)k5yJ1(Mtpahy{#0{|d<wdXr*C(^;>f*83*1me
zI&a#LduH}k(6$$omPQ?1+LxgAJSvK8uN2az_w~`fi!Q4a`Zp4!rPz~>@uajGOs}cp
zi^w~ld-oG-h?ZJcHVvCX+74@x8z=Za<)uj?<(KXZC+h&@pz#j2t6Wz4+uCyuAEi*o
zL;DF-<ZAFTH$KU-#9I&SkIg@C$lPILoV{z3I?zj|(%DIO^2T>iFec}l?0$I#p69Ec
zY8A(;mN^L!h+a#W4MXo4%-pa9(?Q~L)VZ24uHG?K3O-)w@O18D2|5J|x5^`2T1Jf&
z@exx{YD5v&YvO=tL5RP3Q})s7F0OVUg#^sNo?}m~JEz2oDN-M@QHNeklc&gLBNC^J
zZK#EJ=X<7yOrR?&F$d<!&l8-JpBeAFCZ!t;UOX7SSKwzcj#N~vbLE7qL*<UIXu`A0
zaZq3M)reJb&u6SFapz<g=l(!}9X@xk+}hxOBnb5#jD|yiVOz>@S)3p$Qxg6)Y93u4
zq%;H*!dVa$R`7>`S;F7$A2kUls^H^`&ZY0*{r(K-;nTgGHH%N?`+yBpSJ_)1g9tL*
zZbsrU>5VFJflseV{AiGwVh@wF8C1lwZL}&klI>@V$$`Y!l)~t<i7b25bo4-!zeMGK
zN*p#GneaZv8zapaZ8b3&3nGOEE0gsNSsYCA2yvq-_&mJ2)$q;S+mpgEyV%-MAxLD)
zxLOrduWOo_WK^XFFis#Xo)c#HPQS(o*iavJFcs4ei|Do&f2Q^G1q71MqJ7)7mhC>;
z{(JmK_-OGh=!0WkN68=CAWTivt?aBMzW^^B6SDzL5s#<9-U3Ca=H9I)_3mp1>S+8g
z`H+YdUI0#cc?kmLP_)))a<ggd6h3PJ`s?WruV)qrJ~X`Q)a)SP79!c**?Z(Xh8ebd
z2K9JFa~gjP4BQjJ2#z-pr-cy_fnHd@uYCIIeM>Tro$2K+vHzu@{wnA6^)@GsA1A>Y
zB^sKnFn=~N$H0gKfm?AG|D^jKL=uo|IuK>rt_GI3VEgsQ0u3&j!Yy%3*)KtL#OFPX
z;<X;bPe+s*r22N=2@m9#m&fVu)pxST%T`Mb*a1Y?J9O>KZVq#_w)%Og!p|YMCiE&h
zb6US0*xVU#1EjFMGCe<LMt7fHPTNQa=6`I*Fnl9#Tk6C(KrGdV+OymiF$yzm`w>(<
zPB-kKK<8c<>0ed;nPo%(PWrGCg;eY3?e3^$CvsM*&&U)}HR0pyFstKb6?lY3nNF{j
z{N!O9>plE}vFX#}+0$lA+WtI9+k~pvHV?DV=>8&<MZqxlevKyal>QdNr6$|ekQF|7
zPd{ell>4Yb4X`kN$paO1OJ{}tHu_tCN~`!Pjk90YTujvC=?%OvhUY!l(&gB1Cp1Y!
z<+MbB?>7nEY!`xD|ISF%R9;sFZhXpCRa^B>mo>0@iG@N($V0VRUSVPUptqR8->ziT
zTid>3^shOhx;PP3H#4{Ao6EcKq4|)kQ^7c-j1im;Cs%rQ<}L%d9%bE5v`}2dHj1J)
zZS&7=#Ti3%wXe5=b022D`lTf&Clilms%eSfqX{_geg?g?tITEzR7yzxS53^v4bq3#
z#J+==AJ@Wymv{yCMP3lV)8$ryQ>lYmEM$x^mjsC1O-d{H9Rn#F{p<8*<jowE&-+X#
z1}#O7VFpsdh^T`o7J>(NH$h8++_yiGAJuJCuvewpRcqX&_ak39orEpE2}%K*@-82s
z2V?!<Ku~Ak<4F4=Rn={HMzL;n_GRckMv3ZKG({?7IWtT@H@8W~X1GMNOm{f;^S&uz
zlem)qM#PMGnKOD0<_kT=c}P%`m{o?MRQvjyy?jUShrF49c{IVgO#r?!mXW?jg%P38
z)U(QLVR<kSMnC>*ep!U0;)UrNlbUxV)!G;v$9OfJ&Nic5G|IN>Z|B;JT3s^k)-Zwz
zBnyNn54$Ytqb1Q=ZJ~(3w^We6_0P?8Z}&g!1Rawv<@uYTByh6&4JX4Jpls)#-=ZTC
zAPTkc`xz)Q@|#2VoRHNIl=@p;-27|=w6HCBoh0AM3bGI=czmk(vLb($d4N@RkhWu2
zTQ?gi3O(s<oxeX(a?J=9@O8GaAYvqI-2<d(bTDsyaM<a#$B}2ZuTR3bh=*ckN_nmb
zy1j}ee;XS|)3(piXuXP*!s@=3H>NGOMrc85$<p^Rxvw5Ik6(6UQ|8)Hkx@W;gedKn
z(2OsJy2w70sPXRamb`A`ONS^+tJdfZ_ZB13XOu5rzEr#Z)wFV;jhBKC-!$%~uw{=p
zSZJ`!I4<(Za{0a$9mdXAz!A{PffxQD-nr;WmHi!G!q_Y=PZ0=N(vskT%I{4S3oV!G
zSb><cs68icpDZ@1*M}oRNKQ`KyioL?`#YB@=AId?NNNQEt?cMyv#fo0w2)M_EF-OF
z7BPnxN3vix)#zUynM1Y|e0FS~sbmLg1OWzbefN(=j;fu?scnZ+Q<tMbpym>HlM^jc
zMzh(JAD?@jz7-p_BP}15kJogFJ{gqefJir@dxIp6JPaePbkLkpAG+GqLqG8N6pU8I
zgpP;H{YZo18?;lpjM8lvd%mv$;-2|6>x6#5gO8Qd0%i17y+Co*9DE`lnk#RJX^W>j
z;F3x2T3hgXNqfFsqoRcHbft!@HrO;C*j@AT?xe`$378n7&kGe8oGH><8{Hevt5`j>
z$T#<uC_QwXbKKlx)(>kdcYLHV&NGX?VUj#~653@H%sie(Ah4b(2^Ev_>YnLI1X`%l
z3I3$<SF~x|7<Rm_TSwRA=Jf5xxPT4r9N8z9B@5E{&pu!u_jV=36)14u`DVB%Q2;}+
z%lPGdbqK_%U%O~D>g!z3(Xl+2B+w*szxkDTv$B}$-NetNHVY_Fn0i5RclX~Q0u_m!
zY@M#jQqC=+ZYjJD)qjlc@X7}Egq}0*ZH=L@zS|4%EmKtshT{Waa`&pW>dXmv8$cbg
zOcBqtX3NkhuPZ!BJ$26*c^{!^2VX^V8#mPE!IhS*f%X2MJ&gJI;a9133hyJbgeYbx
zO0D*b1lZx|YniHdLjGj=Ri?ZQs`Kx+7QBukZ0rOfxf#+r7$?S~Uu_!YE8Sea7iAq!
z{E`*^#a3zUW7`H#!tPXTb@-8uz%)n>48BLZxBesRK(rLD(TVZhZ@M_aG#b%Bb)hkC
zTjcs^pC|A{=;w${<Dk6IjpXCWG4ktRT`tacAG7LVjhpOGOZ=Js$m=SBeF!RGv>S?V
z&iET2Sj7DQcKnA0I0l|%s+O}=&u8zwqlcB-zvP_qG*!N!fvXlHgepf_+<_XvPdZP|
zwJ#X}@yMCZ{XxpL;Njzm!Z;Rp!w#fwqgabP2Ys!RHc<*UzEU|Qoy{lo;HlO^@;TYT
zxYFf`hQeoWPLMx~yE;5;iUP?Miqk7KogCGgXiVc?7!?S%F{*ezJ}^h}CYHPZd(%MM
zp7NLGIAmJ<7_17-t=2>Z^Ih2e#3Kml%55u1=tQ2Yc#iLCP+=x+C-?dV?W!!^KVmut
z^UfAuPRv(3(d|8$W+~nqi-@!xT5Sse`|%K1k=*L%)Y!RwhX)>U{ZbS9FP;scqWMKC
z>^j!9FpR?XOc%7dDl$mOm2H)wLh0S$rMS=qe$rDhD)nETB(v87iSc|K3Ol|*P&}5y
zR+~erqk94MG=YkJS~Ij-r@o7GxcAXXE>Ym1>2y5{(DL4nLC_I1O}6Il=Pk+vc7qe)
zA|vnVn$Rij%2Li}Ql44H){x?%I~YqVF+ug5NOfxWb#_)(_l*M(&_AjwBF=QG+>UlF
zaSJnW>0>N=4g$0PxmCTOD_?2TophL8GRms}qeVV#h;Mfw#m2^-osAn{A5NF3S&rvO
zo){(E4iFAVqX+_$!%Z(L;joV1lZNpChg9{!84EwUo}OO4)e-|3O}(q?JKIsNJDKM?
zjhP8s-p;Q>ik3e4&tc=ZnF_vo%|kmRDJ~qk*xPu#@vwK9h@VjKg7b_ptEk(f;ybga
z{qxIhba-MEfQ)Gz-I|Jh0wz1Ai6iRcKSu?2DZIPP=;!1CF1gM5%kmQ`XpzwwbpzL*
z7trxrYi}`G)k=_uSOhhUNA;CY#3uaHL>@eRVMetPXF=g@?ffK4C<g|(t)u#QJHF{P
z&7T)vk8AW)1xp@U9tfMN1_u8nHOR|#mfjVVDG;MJ^e#s6;ZL1MPBdtc%s(o3H5>ED
zxFZSa5dP!)$vpm5W8ZfA6Nsf%!S}F4T-nM+fh?(sb<77*NkOhBfc4;T($oF>;V_LP
zORBZyGl7Jc^**!y3L^-}!q3&@S|RBzQCLGp@~US-5L4|52yY!8a4h!+QQ}DHXQ+6V
zG|1W<-CHVeWT_D4ML)ZspFA^)C?~mB2Pw?)A2^ZM3Pkq8L#7VIJC{AF>OT8!B-Fay
z)b_hSlmDu;y9Z62Dh)Y5)$?G6-}D;G$htpOm04s(Z~5Vnd^hGP66vfQ7DT_LD&q20
zM805RpAv#>T^brPsg(v^F|=`cw923XWNDI~Jo~PCqP2u4e|qkRsky-e{#C_~H-gr8
zm6%k^wf&-_MXF?)$^)8Td24?s>Jw>p6CljC1BN7;Xpy><+P7DUk%gGBR3{ii7G8l=
zk_>ZHfof;q(!X!vzh$yLZ*5@&+-lJCRgg6WFZut2yY9!RO`x>DKIVDzp}ANswRcs=
zt21ced^Nunlki1OOgNmc6O~V7H+S$E%Pul`d;pVH=bd{w5R$7mlnr$h(e(5Cqi<I*
zIv*Sq*MV7<nEZKe_>RsldKKQSw2UY*LTL_?B@zzjzJKb%^WI&^p>htXZJ_H1*(Kds
z9jNVkzH5;i?3dbM^tvjNr;&(5N_)I(y;WUO)MU^=W0i*w*E<V9?+M+a`C+*_hsCR3
zQcFO6FSs)-lvvj0AQUI5C|?oTjw4BfTn?P3LG!kIzIe<2M$2ggpO(k5eALzEm#vO-
z+mG`vtfR=os;kG3_Ts`B#;-MTwDFR!pjPuOFe6*F?LxB(GPtDuQN~kfvubtxr`(~E
zB*~d3bpksbt6tf`t^5@sn_j)V|F=Iivz2;S=0j<RN73uQO}0?GYxp(3aAxWhai=Rl
zIm{ePR6-<>qsb8JCJUUd4NyFP<V+3$M>65}9Hu$F>##FV;=Bf8Wql!YWmBZ~6czRX
zmll6xGSfrBrRz=<3GSjmf*10!l~wIBSACWerSOP8`i^f2<VEFZ|KVbcX%?=d;5;78
z)NjVeJ(1YJv_q=L^T__j(ik1Y&sNon*K9jVu!`TTGzx-;Y@8~E>$QG#GqpwnojZhy
z+-}r$WVBC*yxCg4!*647M3*d5s}jJBQJH;YL9^2#aKw$tpoN!Jz-uRN*S%s@?3(NA
zGL7rKhK>&fFeiB9laYU(^n7Fvgp99@M=>!ip20p6{yNaI$n?ladw8R;3yb1mk^Rst
zcH#Z0BoDo3e^IL!g1=l5xB+^VZWSrz0iJAnF}LXC;_b*?UR&+)a=1UBCsLvwV@SCj
ztXfd?_Q}G)1YJ$Y-W`-x)sH&3xyjhtmSc#XOEn|i=OKh_T*$inX#A5~lK&S~-ri_4
z7lr89wy))DoGi6Q9hb53WO0|Lsq#Mlo!9vjiW1BbKN*Hc^`Qh0Kz#OHq7v9sqr~Hv
z?<^<SaP;P+*HUO=5KTYN1N#R&7Ny|J;CYmBO4seH%V|=j%3rs}_S4kbYfRYG`rf`g
ztYb8^Z@bsekuztG^@&f~mm4T5{}5CK8}@E2geXp~%{#{t#5p?c53=>k5xZM%BAn8$
z7HgN~!hoct3J1QvvQ@?$(^pj`JLrstz7*S|<@iHx5EmE8Ew0GRK2jxqI+j7XN>i;H
z{RxCg07~5QWE-kScW!Br>b>M){PL)!-C(z*0i)VlzHs-i7G1a5mKx~mdKukuu^iir
zYh&Av!AF`s4kxpBQ{6GB7knR}Q)%zC#H{T)ZLyr<r)Ml*a%*H^c956y3%FzdM87S7
zlE<B-K?zIYX~(f3)b;Myt5S#-7fb)WWRR3!e*U|oMf!x>g2vM=Z}d`PbwSkZNY%ps
zp~vOssE&=T7!%BfW;#vZ_`^CDop7?he@V6)3(m0I)W=Yez3Uz^G-!mn?%lC%(HDkL
z^v6AO+0JKQU-X3-96~)PQhAoELZ(L@%%1U<%QfNE@)GIEYqw0|<^5uvB<7$Uv!{s)
z6W90)!R>hsfB~C`5w``z5pkK6JsEaDBiLir$1H5~nQ9BSb*_Kf?Sx70``>2X(Y8I=
zEG;wL7Gouctq*V<ZpafC=tL-$jBaD~IR<MNU^$$*f!XcN$#EgxnaxPGCjX(n)P-gz
z?=jb#X{{Ju5@&{q_zl+eeOj`)y%AL%nKh&XY=%5Ej<h{x7(;^U;9i#xjrNFH*#8B1
zK|E)r9a+RHeREwRv%Bigs<-ATYWWmihp*$Z=qx}y=-}AZ7mJtlx$kg78Oc}I75&dz
zS(7DXh?()zYqit$6M$20p)#1osQD^+M^}Z+z~XEV5zt1Ha+G}Pleiul-T8um92Tl%
z-txeB=O<4ewm1dc1~97Q4*yAV-9BkxUUI}!es|`#JaRR1x)C5d-(ZLP!JTRv0R`J&
z{&y`!z@TmeZlDe+Py2)2prk%q`LvRb{BtqreYa?G9^mD?Rg;o&7iaHeC%EUAtWN4@
zYj`h@gx{xH<*nMxM{jO!!}UEsGxU42Oi5An7+R7zdF?MadsDZtb~Jc37OxZ{RM5Ar
zJ7c5b9@CRBpIvhJ^3Z@h{PK-9K(b_*4#-S4WUuy^Ja!q&;tt2teyVE>E|=qLPFI*e
zO6*1zoIYIeDIPYPhe+7!wxbl8E5#1k#9mS*ma^HKmuPiDSkJW^1d+KIws8=x!p)nd
z&f-%&NGq-t5Bb^osmj25@#ZgYY*>P9)a33?bhTiwL2?L;&8XiBs!|TW8-6Ey<h-s}
z&aMrM?HI#&O1oPX#3qT#*l;?e+T7JiZcn8Fi|{05Z?ntd{0w&TAGN5A$P5@Sk@fod
z80O?1#AN|lo8Q^X;&=EcezaVRh31AeHrkQ~SG3*q373XE=fpFMPN_b4h<<X#g_B`@
ziqzPhNpLXY)xD&|K?fz;{8^teHTGH(+~iSt$72COG{9cXu5EoGaefit?+8B0z6W;v
z$xolJRr9vb61Tx1o4p)7sojL8=15#0%*4jzJPQnui9G_cK^LL_@5|5f(wzLLy*l|?
zw@mZ>-Qy$lf-ua`z3(@>&Pc{`&u<?<{7}Iz2B3%xxtkR^x;LbkZr>K)IgGA|Nl#f1
z^+Myk7)eM5c^Z=&AglJasQyvwgpO7O5w5zvYvN#Om!tn!=AjW_=1a#`S}d=(k0IkG
zL8mtQ%W<{cB^j8w64h3*Aupr!RXu9*R|#A%w!8XZGxSTXWT<t*-EnIGd$wrrOnM#f
zT5d;s&>$+~0?_c}yYc5-1bmgT?{v8RUZm|tMT$5LP0~+&c51bsYKJ9pN{5KA?E3l&
zrlyH{wHulaA$LvQ5$QHk$DgV6%T&FO&20ES1WcfAQd%=L_r|!x@8<It^p?Rew^1Av
zz*gi&MB*Owi$D+<BDJnRVHMH{G@uh!x7Rda8B{PPlPrYJWm@z%yE>bF2hkTlC1n*V
znCS*2uCUa#xm`IA<Dc-!yizk3bSy>@g~2ftYZPMGnEUB!<e!g-?3731B&>`TY)|pr
zoB_q5b}=A*IPv9J5~OSQOb?|x=bb>5d79OD&~cSez1PR-X>h6v4(dtkZqc`DpenVO
zwCYe-Vucy9<PxY!jE5{ImdoamX<upbs5icWtzQdWc~oKrA2PI+$0W~nW|w03VB2bj
z0}b0NpYSO9>uJk&T}zF7wZVdm+UQ%cy~=BMyvBmcEb(Lr6QrXlmj$)rRgN&Y&ue*@
znjUbgDPf0dmzMa3V)APL@j`=P^F*w#druf0=t&4ApptKv#&KlrL+u(X2va>Dsk$qV
zl}yUhYcXc;b|{5D|Hc;&i)JTL(mf5kmaf6?{sIJOsh?7Vz^^Uy{YlnRDq7*njL4kE
zfgnkPIP2n=t36l@hjCQ4`o3klG<+7(4r|)k)=dJ{Ni{bv#r1GTQk27eUcJ(C+;w$4
zu;k;rMBPYq;ckiAyRx~34dgDM(40hGA@=&t7IQ)e&v?Q*01<V1^kYVkpJ>QyXRf4f
z+wHTR6^IzCxYkJbY_X%-f}9h{VuKyWXm$sUB=htb#41mRRM~(tL;XU143fAQ&PhR;
zm6bKB-}G^SmQjxGw}Z~LWX;cZ7eZAi2knUyML@|myXG5rmWfa17R)y~y@8d<$&R#5
zl~Z^YI?aYop#=T%Tje`nu#ihtQdMf+97ft;-07f%s_G1vP%8!3g~uD}Fo{8xr2UsW
z{=hIwe>o4RR_lZpvfV~nB}j#0$O%$9qQ*jVQ#+6aFqAa;)(<xPzac70;8<=)q{PeY
zPz@f(F5?t|I1y)TgtqgA^G0`?^O&c?BZwgs-qITK@PTx5OtY?s&0S5>1TpcyYHF0^
zl)8g>RN26rRr<*YrU%god~02Wwu<-(C1rSbx4o@vB!mi5XcG+#?E+E9W2S#kz0eci
zM|P4$t8G}k_Ldw^yYWPF%RCVLW$3NNi){%=AL2*7CYr+@3IB&6%fNMDosUM4KyTrj
zt<43d#&1?9SIx{^yzmtL>DaEuU2YO3E&YUB3~wm><=VGO;FQ5BT_0fA%1=kDicq_V
z=JK0St%EI>=e_n6CN8dQ-8X9qbul5QAUb!!f1InqKhE_E{rvnN=c>a}&0ve|4@Fbx
zzdo8E(A3n_$?C`b?+!IwmxY&fP|OSiXQCplESaK(ObCzu=nfmFsEXC}E3_a1gA(bH
zNdK}-8$VK`@3`BmqkKP~wVy@yfyInN+J`@~)H^WCC_drcKb5_=IN$G^vUNKaIj(Wi
z$C}qpDKW~NBk-+{NWOElcqgV?#qQkh<Mi6|K@iiT4b>aj0)ozlL6{LPemH|b?RGnK
zd5FciEO{gZjyJIBg%F|NXh7JZ=TD?x8?;4HJyHr?%R@Lk1p-qMCG@GNQO}+{Rd~>b
zQb~g!FMs${(SdmlXU6ibD~1!fBwe`}A)$-3ifgEhQ@7`r08r~OMaL3*8-j$Llb!%b
zxGO}`RY(u<CvE_nh6nwK9_YRRA!2O)?oj_UmmnYSoVECz%MAv45_-e5e*Z|rfTQJB
zl6}}VE8X4p4c@<B^8SJLZR9cgP4RkGfNJ-UP!lz^2bg8gI$%gTGNa{VKR>$?j75Tz
zB3c2~ms4$+2~{$Q)Zd^G0QS63YY)Wk5bJi}>F@eI2`Oa>GVkLe=GeM)@085$7ERS-
zvM*$H0J5XD;vOX^0q(8nD2{E?R`JApzM>Q=WrlfN?pkF+x3&Vd`lHjAsm1+RW2oV-
z40dsK=-f6OI&jkP^W-_$F#Rn(TO=!rJ06)L0AJz%API&6NU2?^F+An@cCA&5To$sc
zVZli0O}2e`veENjzqw@k!N+;3@G=xy5LrCV_0$Q%y-JOKXcSMNKyKR9a?_tJPkrPU
zm&tmgrb&Qcy6UpHpj1`Qcl6>tdI~ru_W3HM@xI9m-HIxWE*N>EV<oHG0D<@KkU9sA
zAuq1*1^yv6j@T2br21tcPGppN7f^&2oaa>Xw+}hTE00WO=aP()Gh+{M+p?>sKDX9B
z&5Ifsy1u=a(~?5IjBGBCG3wBF@JPc)NsvU{9sJYAwt}zRDj){K&!JLRZ0u6B^^wjw
zSgqS(vP~hGsd{UMINHB@HT>X#lmF%ZNzqHJ`E@Pn7!~@Mt6P9YPG182^(}ChDfUwV
zGPHYmJn?$O`zh?o+15uXDuCy{qyuJi33)RNc@wO9zRV0zWf3w15f}?PjXqvH+fUeQ
z+q)btHSZyx@{N8y-ktdYuOC}Y9h!cn2!1`VX?sZbVs_H1Z)oVv6n3xwn8IU$Q>d7J
zvm4f9VOabxPv93OF*O@}S#s3>gInHYj|>lwHz^yzOIOTZm54;!&6jPVGm-T}-_uth
z>h(v7)#FzjS?yi1D6(gQ=Uhl@Sb^q*3O+CIk;bUun$=H1_{h3ic8f^Jy<;_to{NDy
zfz<9>7$e;Lo}Ws3&^bd)FQ1HmkWo^TAaDS!YRu>>0>crHKD6I`{_>gfDgK6{+&tkM
z;Uh(7_k}#4)ppq|e%}LUo^>8rKO1>}J_gLS{u&}>x8L#*=cAM&I!*@mpuB|u;r*dw
z*PZ7zW$!53aGXzOKZ>M80}VO?ezJZ@b<|{c1u0DDUmWX1gv+nuJZm`M;hWCm^Ao8f
z?Ol)br+`=^wp9B<f%(|hsF#zyinGZS@W}}tA0PPc*ZVVdGO$Y7FWQC%+b+F)P)yp1
zXRf=UiPzq<(@)Aml%eDyLC}`Cor(!#X9{`kH!IMVh9xZ%*hO>cDaK!8%j~ABr0T7;
z+FgA=xCSnd^R;4-^!u<xrTeJ)h$F5DJ22@i^4J$TN>2U&O8ud|r^=Q9>`?@{+r~Nv
zFFnG2`saV0KEHg$B;&V>;SeI;4bKV{iLCx9WsCX0c(C`nCw?e=VqUi=(&_wm^}(_Q
zh4FE5Dg~I{r(JMLVV(6Kr9e$5_5pqZC8CiAmE$*f7cg+MMZU`NS}g|xd*Gg`|8)&_
z9jdOM6tvDi-WyJ1mIlf6ko^k$b>UTI@#%>Ls7@nFM|mhvE~~3cc4##k15BB1-s?EN
z<>n(T@)F#T4m%UkS?R=Z4)&T_($X7UpXv<yQe^#lXxLKcLzMmm;ulh}7o7)%kNV94
zBaB;)LNKw>K5>LZ|NPx!%Yc@d^6?52Y?rIjsEb`>R_jK(Wc$2c;ujJMGfihXk0igB
z(bcY5wK+!ym>4@fLmez_w$h*f`iQfuiAcXp!}!EtsBehDy4zti$Aj{x(w_^dWYPp8
zFbCmw^{!qK=3!NxJ9g$Jm#*!C-w+oYHSy2Q-Y#Qi)xTDhqG!@;G7kz?r4N6V0MgG)
z!JGO!53#J2y|oZ_MDxoS=vtVgAIaWfgj8}Ir@h6y8|l>#CHTD)rCF4#)~aZJl{3bM
zEyxK_WC8cS@Z_r3j@Rr{u(N_Cwk)UPRz8=Hz`D4}L?iQWwT>?yoHAQKT8Mf}^OB{K
zT1W{5%rp9sn@vY^3T>$h`8_iB^bv#0tgFe>7<X=y;^%4@FNl7d1qrlUMXG`T6+GUQ
zYT2p+)FuW$C}2S60JtqWFn{Is%VM0{#rZ*clSopZHJvz=2er$Sd|F~S@_NJZC&W`(
zD&Si~nQ@N&dt}I>b+2AYQ)M_)zwL0w&htRO+uo#IVF3v>ncit1AKsp(BAvTyq?C_=
zDuI()xDqI^LAJ{JPu3>Ywq|4nv%n*!PNk~aWde0J7_;rOa(}b=;rWMaeW;)-PALj8
zeD6;emF)4<U{Mtna!x#_A$PSJN)Zlo^hr`rkBAe$;Z)34v?->5%D5FG^q}O`k0#4Q
zIu<6hC~>E*rlP{pnbU~Yjtq^2qd$|2iyM&^%@wT`^67o`wRw#iO-l1h-vhQ6d}Cw;
z3R}GH`4Bw-uY*5>l85qjSqy)(wmpc+DfY&BfZy(@VloVqgb!)hxhz{?ojwJEhFi~z
zii%8ngUvrFc#X1id(86!_P7OKsLI!>S)(a{_~57WklAv*M%&Qw@o|CA*kUWc%i&@Z
zrt{U^d^3_g1c^s$?9Y*1T`J}*VfjRv1;+_;Kjk@uP8HAL-KNs_<-thd=AF!Vz3-J(
z34h*D7`VI{Q_G@r(4{H2);@o9z=Pn1P~-OBV?h@#6I%m>-&)8L?K6%ZijV!N(D;3s
z?;&mT18|~YOquxE)qimC+{jLhK462Q+6k%IpA_!4n|rqE_ZZfu9CbO~HlRc{*GZo8
zhJ`1m&cAjfG_n<X-)#{b7v-B7?Zq#V&CsAyi~KApbC)+9n<8v^-pe0wzc?Ddi=V&a
zkaV+&!Y}h@>+eWWEk(~A{|QZXe(nd{-8T)KZ(2tgtGQxcrKcO9w6$L_Mg3Wx6=qKx
zTkakiAFUG#BmA36_u834@6(CP0hf!tI>Bibc7Lv@sx1JS={+^|&+phaWG)Y8G$Xms
zM-<*Iq2LJG=cJS9?($A${~=j2>rC?LE@(UX@4q#1tA~Ft?e8H;EIAyXkskiQ(p%Lv
zfGaxdWh2NASqb&H#d_G(PlSH(owrA+q^IolBr=WbYWv>t){zW$tAz{`d7XEAC#IOe
zX>aMQ-jJ(5+98$2tz@Q)2ap|gK<-JMJ&p6;2DodJU3N#e<~E1>SU~pl27lPwA;W;)
zE5;d29O9=_=%VDpT+i5%tjzb!Vus+~Ic}mLYwmRpw%x`At%u$B(VY$eKY!$|10(8-
zJD#($5cE#V1G&*ei_KG$S$G<VwZuK2{Z9_EY4w*0`409gjN%NffmrR{4A-%cp>c}v
z_5gQH%AU|N|Ma|$LVN0jemS3&F;9Wt(@|7+K1;x-woCt6-8dr)5V-(ob_tEJnIM1g
zaANxUQ7Ed@QIJBeC=e>-%;^5(S))Gtkf{6#gH}(g`d+i#IAGoI3QcY5M#HDg$}={6
z`XP~pq*Cp5@Ltfy=fmWFS_{u}$oA(>0pJU~QKvtJ{AE%>5%?$?ls$Gm)li8ZtasL9
z>b6$7Hb4;nEh<RbgC3wIIOhz{IIw+rG=c{opvipp!5Vp$USe3vnsGT8?ag&`N1GwW
z@f3Bst%6nNvuu3k{tjkx!lF|5P4Q2Ezm03l*FpRUDp4VKjy|s~%<W(Qd>m4a9&`Ty
z<W=U^-mGNbr<U`vkWo;(7Y2diS%Y_{PEYQ@EcYjO!4BLr@Ab%Kw75}5+HdXvGJq|@
zU+W|g<r_cOn2flHYw|gTF|Ge;-LHy{?>LW*RZ4H!ieF#mJBQF{w+hyYr-ms44%2Tg
zwbDyIm!1{mkp*(c!;nosMHg9daGKxav5CKvkfy%2^o1XfkD;5$URX;7^W1QFGx{a7
z{{<#g8H<hVIbHlNp#zeBq3<MvkBjeHnY6b!u6l5)zAOrYzyuB9?62J5uwP_aTNN2U
z(Vd$Ffp#*1{(+?kuOX5L`s-MMa}g&uVREq1E2{n30le_w$Got_W*6_>8~8bX;IHV-
zIQio!s97svr^ng`N(^<LJGGH8yIL4JmppSVNy~d`O->8`Rm&A46OcA1GhHD|(!5PW
zj?0Jl*17y^$*reTjLfiK7|}P~{vxrlE;JjuvqM9~MuPNWqdtX|%V%#FEaaAj5((A~
zHZJk)d&W7{>6l8n@886c1$w7(vW}uLFPuA0)>0&;^^`G<mcyV0=9c|VKf#os@fGsZ
zs>qlq;(jG1_gtGa!78gvk#lJy>Rpngm1oIR*0&-Wg)VaOn=k8fWDHM^B;;>5=g&Zg
z>b<2HlC~XAqP;cL5lX-B#Br<;vfoTS@Y$-Ae&k9j<6;r`C=@cg5F$!uY>kGSz^1AX
z1)aVV@3|AlRagv#dH@$WA*T=R>ieGc-v<^x>!0T7HJh>rAfUL58JuecnA2k)ypI<5
zGvfVj4zje@ruHF>PY$sv7gv3?n78c}k|GR_!v2UgXI<WDSF~AUIVR0?nU*oblFaxC
z-||Zt=3nNzFZ9|X4JLvnqZj9~b*pJ~TAWs?3|Cf6VlEXNk9>W?l%g)XTL*MCB!x)2
zYx{el7Ni{Khgspa5pVRkWLj?KL+w`1M|>HW%EBTmK||kR;r9b$b8XbSK~z?_@s2(P
z|2If+^(A0@vEc|kS(hFDaDECDCf_Nni#Z=BoHJ5czuwMv(}(aZA5M323NP&kZ^)Ms
zogEIY{Vq59eo|%c&y(lKslU+%ekU&9-LgYIK0k(L7`f*@yXl!T`5V~THaZqE7+w8v
zx-;p#Qr^@lqxXU{Ire09Vd(eIdJE6t(QfzTYqVYPFH+9V{TPZ5@w#O%ub?b`7w{hz
zpyJX8;R#dUPyY+tBGCNq-DxnK5dYDR3Ai^$*Q7M;Ww$mOp~@6;+XM%4Ab|s+97h+#
z$9LhuPra`Hjw!tUzi+|~qyk7EC2HS*X8t?j3VJQJ-o9Jb>m!7G3nR#4{HxU@yJDDW
zc!`dAvO}U>yMat@-1?B4o`&E_VvRq34mClIj^)uE$e%qcS5<{twT?07thI^aTgp`0
zoVwoz8w5F#6?~!=lk2$%;JWktYlUG-TU4?p(-hOie~vUZX=MCRYNlPQ>{};xxYg4|
zVi4KI>wms#L<gByC0{>`z-<v2XCh3wV596j4)04&SCTV(Zi~iUJwE8Ap|$p3(qeMU
zB+y^@Bco)QeP#U<m5)`f0L?0*tjCng`I@Vw)8xJW*(GNun*uf2q!BQfa4X>Tkruf+
znd9+<HtIiaS9pdvhh221LhARxh&+?`FAg+xcDfjclTK*937-B9e3Yg`sV#GF%x>%X
z9Uthu^QsxbdY~p?SdHHUHerE$GQ77LR6;3A@BLj2^cx{ZKIwxY%h<1%U9~dqfFRu#
zjrh%v>tp?g(Ob9UmTpwn2F-22%2hc@J!Le|!k8?bCI)WSdY)^4&B2+u8>j)(yvE$q
zi`V6?Me7kCyq2tY>dEM~W*Sa0l=B)gAWNSY+uMy)_7aHOWjPH&j19+y=NYlvF)44J
ziuSoLjaGw*uXO}}Q@;Y)*d(GALhwIAjFZ=SrFDI#>uAxPSnb)LCK7G}s}<l=s#YA|
zhkW$DOzNN1w%dm8>0NG*3x3@Y!%(>{HR-ttnWb{WkCCO9m*QCc{b=*n;ig3;t3KWx
zn~ih@sZnz`F8^URt#W>FF0SMLH3Vp@T3j#Y+8a_Y;(bgdrUhB_7OrMSccV*DUfrG}
zd=c(;HM+8k@Egf2AW>Z&d>>6Cy%^uC4SQD7t1q>%|15N2hFOfH<sQ80Q-7v!@qcHe
z>8(@>dAc0+fNMOlJGJcqv=)%g)Dyb_((gw!qUK-j^%ANs!)>UDv+Nf}nj&e_=`93J
z6xH_4P5{XUqW2!bFUGDZ85<ohi7Qg*B<WWtkhfcaTbT~N%<wVFR-B_h1=<_Jl>gPW
z42@xI+lfX}&JyX~9Z7e*zc3$8<6{M*c2^A%4ix_T78W`@@&1#yUf%pC2kX_|WDV9)
z>B9NidqF@yxPN`^`oGTCzxOr58d_~L?y`Ja(D931a#q#<WH()$`f)^8WjJxaA?`2B
zW;F0}_bUi6pfYMWa~T_YW8LLYp>F=gb<8iLW%ER&@ANURdea#2N(endNARr#<437>
z5fdDz5YlZKJM|dWUBLHBVpKX~vwfIlgq~*%C5MC@Wqr;&#9-l4-;Lt0xo)i~XRn{Z
zt0Ngx#ds07$6eoq`Xyojl1A$C@nn8JU_~n|ceIU2ZN#XBBE+7F4EV?84RQ`sI>KgV
zyhNK&qo%e%L4~2ic*>D#lu7TMCLoyL+6O$swQDV{=rD)%p-D!aU85C)W<>195242j
z>Wjd0l7x^2zn4u^qQtt`xi=7HUJBic52xyhh$7L)k;r!LW#@<y-Wp+OPx2Y=02MRJ
zjQ6#FsaMqYu6RdCb{-{maX#o{izJR*{1~z!Q*%q_R_Ys$u1th&5~2g>Lpr6s-tsw(
zjDr$w{7uDP-iy!cB~JlS%qt0aiMX5j4m78xzfirD&kA>k$0rAw91hw{@?DSUQ7M=;
z#AefVIi%SbHm{LPhNJ%AWWnnp4v}A9j#GWlnOjiC>n3HR$Xp6~zn_L{B}KCg3vZe)
z(diAjnZun8R$#YBPO9{3eNP6|n1$3owOk=m-a_YT>K)1oc}|cxn~ijd;ibtymnQw!
z<drrkwA~PeigEeF)fzmo0`4?Z=}?91e8PktNbwetu^&%nN;r<nfNh2oYvW~hfM0Q8
z@+3+wv7S=W%m`Dvj?4&>O!-pl&1#=sX?+T*!0w3*k@o1`Nox|&@=FiJn7^x9T=@Yj
zr?=D|gN_gej%ECdPl^4vMPorvU;k}U@i+hJD;cyX9Sd2`<PwHZ3qEyigxLPZZNB@B
zV-y0Syl@?uqg_M8JQtIoSsu|7Sh$?|pLii3kzIH*UPjS2V0#PxtckRtoIc1w-7;LF
z!8{A<Sp3MD@ooQ$TCl?SF!Y%$&%%sx5L?j$c7J$F`pFf@oa$#sUeq9K24z9zaArQo
z^9Qfe0r@GP>V?x<=dCm7{KUOb_73aBBF{~f?MR~5Te4xoA*p52wOhI^OD6Y$nx2TX
zLJa2<YHbZnZ@f6rPvih$*2kl;%x*3$^x)QRAGf+D&F0O8z07#ztrgHf)WW%sMg93O
zjQTXuOq<Kk7$P)zo%dbr&j9@+9v*9wLIVn=T3v0yEQ!=_cC)<?c)?{+i3EHmf~dsr
z=~B+|g_Hg4=o&H~5-53m4~h)%Ih#Gdkh@r`NNsVN!$|YOFEZ%!Uh%Ed-=m<jJgzM|
z5P@qdvKOOv$|?&*VDW~LXO0}y?~8cN=6=HWYF5#ypssyql(ip%7ThG#F_69|>drWo
zA-~1#P@2RbSa6h{(|WFeYbtf8aWjD%0OD#e^TuuxN6jpqC9Mc3_wALM2u08845^^A
zei)7IpG@&7;kXC3%s!p7x{uqOD7o1o-Oei7^=BYl^Z9FZ&j0Xlmi$6~bn1WK^Ssd9
zGTZ7@G5DBRMQ?C<ZwBV=dvf#5G_?hhNHjTQBae?9ZAOkg@XPGS`K!_M-Mj8vI34G;
z&7d_+WQZGpDb~J=j*L>1Zii8aJXj^jk62Y+TpDG=Go)6d63G7_8Bz#kt+Q@s^j|#w
zkJPCyLM9NV?SuQP@5?QtsH<W*1Y)Td<{RnA)Ld|hQpNb6VCjc}LjBL$W*qUK)E}2U
z-_=*f=H@}BOu?VOmcHnelW_Bxi(iiYcdF*=Tf`3Wv**|~>iW}kkT?46pnwY);&hF9
zbEK6#8kR<mq9xq$DvE2&GH5@=OL&gv#ti<pfL0?inb{?RudT3*z4xm4Gl%&BS_N&u
zzIwNwpET~%)wEFNm4v;)vfFAg5u&21l@RIFIBk(qdKR6yfHa@dE1v@%?DfwGD*C)b
zIid>S(EYNgPcczs5feuO_g1KVPmR=I*`L+Q&YHpTH;CRY0?$6@BS}+$A9<5i6{1HI
z6?lkX<embP0snS<KvA$Ede(%Z0yz)08q-$!L*99vQ;7oZ&`^QrMz~L;iDqejpI~a)
z^an%|JFRSE$ZWL<j=R4Yv`So})1#GS_><l9^v&KvivBJ@tvmm5+dPD%SSk{stO|`+
zdOwanAO0L4o9iViV`%g@E4V=YyE16PG<HB(^G#kDvbb+Fcd_MdP-;5!hkqETuy)sg
zSh35=p=bstPc{3PTOS$yblTH}cm_(onXf#wflxp-=MZF9()eW44)z7|Pz*4IY0u$w
zS1b@iH#x7iJ{6LX60x4`u3&r+b=#T~$(T4jUDY3755V^QLK5B>b#X^7Y-mSP87tHA
zbhQQ-JzN7v3PPd@e8{KpXLkmgKfU~j{`1cCAdHS?uJ4;`scAKs(h4?o-C_E{CJ`&o
z4NPmog|A(jF}4Ne;JhD`(stE$7|2*&B~ZUnZ>CDz8Xjf~?)9SM+E=C(YncswYDr*y
zX`90sF+fk5ez}p+JQNeYA6-lGc((s50&zjnf~kqmc~pe^dyM2h>Nh=};S8tUid(T#
zPTs!PPOci5nz>pcFsiHqH(!YCcnr-^8oB?^NBrxf#PE#e5jx?}sv6dR^-o+|7c6dX
zEq$oq`}gb{GhCGwU|j_@>id1Dl4vKrcHok9@Ho4<3brR$`ykFJ_TN!#|M76%mvk+5
z&TVecb*EbYR{+U4?maLA?!Nl}BS_9=t`T4&C2umjnGR7x5|8*5F2AHK6d<ndADLIT
z6)->Z$W|ELeCk>V3$^3T7%S%31vJ4Y()3*0`Qv)wljJSr+Kc|~@@RE{^?Bb&?lG?5
zt*1SIX&$QaE`+vRs514J+K4q*;3CSMhEVk6t`KX_y(!uu(B&ysUQH>RMnWt!=3LI_
zCo6Dj%4`eHa)y~@3%Ha<#4y{bwz70v=rm=+L|OghS{{<1)ZK;(GEIRRIz9Ar@KW+X
z4ajo;6ww^q21ez@KL6`x$m&dQ(>pEuW3C*TqhFqlLGt=Y@~q@n(i5+NuUJo2bq`S#
zAYmC$*Yf?xk)6xkH*zZRzrGf(<mDde2vWb)53(*Xj(F6zl$D$ak}%mVA6$l`W-pu*
zY)-8fvD)f3Ig^5AY7TuSm(ZwhtuyQ_#1DE>+CNQeW!Ve?``5f8x;u}$CsN+I{TSG8
z+=H?v@Q;>ehP8&Dh(ECck0~=|)OwiJd<-~b$pb%I6=@vIj<;}rqcrD}fqB4~987w3
zMOryT<`%HC{K(Y8Zk$)v-5+S!B@vd$XY2tlh41gE6M_1_H}G+i|M#K~b3?59;?(@d
znaW)h17u0%GX*yvn5y_*qZtnRzWs;3g?(na{|G^nHu7tJ8ScGGoLcw@@6~WZ){IjU
z`KMe&wnq8w>+2%7U*81_J)Yf{QM~ndYQ7|Q_t_e?mzULsfpTu|*N2061o7D+JKFyV
zVc^5DWR6hI4yAR*dx?kZ_q7qE2hkJ}_yvQ<ECdk|Ig6>eYUAWjaCt=uo@~X6i^i6J
zkxXQMbr65Si>-B%zrNn9r|m6>iu~iJ9Y_xsM7|8z8A{;SS%c^2eS)j@tWDG#;{_j|
zoLQ(I)#>F+z=%|8z@C!L-i!@Rs@dR;H|HYw@qt8j{52++IiI|yD;S*ddAMM=+4&{|
zPb)$aqBBOcE?2i~?_W)6#f2NZBM+&(hbQaZQj7P*7!719F)`;>brz@NaymIOy*0J3
z+bW(u3X`)_RojgyMCnodj+O3YZ6V0QC1hq!BAng*)y&`6T#5b11C+J0N)&4UiH6Y6
z0C{qzp)Tn{Dhq6;JJ*vzJ7V%(bI2GrU{V|_`TfpO8)0ZDbLq{44Whi%vx%Yn?p2|n
zA=dcN<`1dI&~-o)aqB3iqbKV)VfM967<WPOp8{R+q3E-}?wnvKkRnRyYrkeaQ^3o$
z<<A;BQ9zVdo6phl8|0TZEMz5Bx|B*MB7{tc8^){%IXJr8VH$Z}-hoKZg9){VcrfD?
z;r%arnA1-sgq|mq5Yg0CV5E4!hUfjK``qdD3M?>j{4L%6b9ch^D8-_>aS3;}a_@@N
zEj7$Uf;=&NmAj6IOh|+Kz{llKlI3sSE8?IpDMQy(PTJlpZiam7CsD}C2c2tHQq^`S
zvY!ZUrF=jzQRR1;LILDxf&cHOmTiam@R0G37gZ*^Z+$O!2Iwz$nb>(6R2EbgIKqze
zq5sn(#9KsD$`%d{*mzuU&Mc7ju+^@nBr+dt&;0K$aE$jqirxG8E)}j3cG;-Y+*@&H
zFbHpK5NsT4b{6`;P2vB0yZ9gKe#9J;l@&z@u8{cGdoup7-jj#!jSwrcMf+#PzI`gC
zAF*#mlxHpXLrSsR3py7LAN4x#*4FaJIv>)FAr6IVUR7j(gwn4%GK{Rn)6kPGi@P7k
zTu&X!4r_TghkG;;);sJ^ImQKlE~!NodlqbN9fV+1qy1|=GfT{Z1Yj*BN0)+m_|3`1
z)D;w|a0J!q!}@IGNTP6q?n^2uLw$2KVahX0-8!7HQH;{EWHz(iSgiRgWA)se@C8jd
z%Z#(zfu+l8C%=(o4{FNUIZzYk=T%xW3G#o1KB+p^pYYUro04%kZ8lzjyPtiJd3R_D
z=zHlfnsnDCwg*jAkq)DO9zvcC?5*x7#``4opq`JPT6)K2Y?T>9XSO)qv-}IjBvqh!
z{YR^EEq3o?{hhei93oNgOoI_?jz8WJc1=M0Eh>{gHWm@Gf@uVHtDKmZa?KyI$cc~A
zHGNU-x2w`e*sZ?4nDcj8&LOFpshI{zx$B;%FQ85n=<IEl^X|!!@F{6)R0H9F!*?cw
zb@IWE?HgHd0uz>SX_>qm3rn%!v-qm9gnPzD@HOP(Iz*;7=Wm9=Jf1x(Mx?SEx9cS^
zLa$8C=Bh*sQw?-L<L<^9^jvn{2}^nN+MV-@p3~p%8atI%7qM^Daw}kYQC)CE_3ef0
z-*dSWSoCG7U#wl3JY@L?h2LnM`}(k*-)o5X4+UO!qfdT%A7B+7-v`zjj==Y{ZwoLw
z@*6FHRKvumubFSj+abUB#4p^NDND9OK>A!!OF9H43l1QYKcd*3%4}kNN8%;tis5Bh
zDR*~#@Gya!_w4NYVYghrn@4@tdABe5*PK_%1ggd>?VUfihkZS)Fe;;ENuKFkb7r@w
zg>+^trFkxE!YXmVTA`2Bx8kr$DzMf>g8RturqB6_0UJ@p9XVGG*O5MS&XrS$|5olq
zOM5j7K09z79X_}Rg69CA#$|+z)RTnUGd-fVpl|0qpR2Z{Wb~)){vGh&P$z`{877mj
zfgF`366Q=iR<jQzborfFg5)uv|Myv^=gczIA*;T&PI{rT^fLcWrh*osBlj?Vc7MR(
zvud%C+0S>F0XMWgP?LEZmuF|<8za0o;rc^_zc)MB2a-k03@`uUemFl2=F#@JlJPIp
z8ek_k?C+4BXWc+#rUS?PCP=(H51P*7=FZT<mZX<&0Q+0VMIG&oB~-9skJ0))<9bul
zG4=EkWV~^XA5j+b^{n)?+fG!h7JsW#GU#9p`opf%GQHE!X+KAQ6C1<cz@@86I9b3D
z7`-e*(jCHFV9txtnKySe=um(OqGJ+97(Yy(?qTDWpKCx>jju!Q<yrtm9Cb2ZD%-wz
zIwwSO<j+qpgfVY*`aI%@b4(o*U$3<~0-qL}iE87lvk)I4H+%GL76mk<m2a4Ptd3$s
zA34^%VoCaL<oI8<PsC$4QpWm;`A{RKi92yh;xJ(Ji;NAbr?hy%D2ejVZ*S>)(-18m
zK7d$srIpDpr_uvNsI2cfdmX>rO|zUoAKtgmGsNDJ)Pc>m$N@AUP~N%nRW=%0^yZ+|
zhQ7tldLU-m@T0!K92gae|GoKP!(S7$HMbS)wSYUq|Fx3;mYl+mH@xc?0h+_^CNsZy
zh`)5rQFMcqTP4U8zT{oWLGXF9-a3!H7Ju>XSn$w&-T?Z$qb3(uGFDV-Q*HuwtMF+$
zEdP%`4E;#+WG<Lgp~KLCd4?rXB((cjD|xkItZ!K*tW0FHi>fCp>8HT)KK^IeFBbx%
z?rXTS!DRSg%eEi_0EHGR2P+_Clj@C`?M+IR<Jy!B!xMrDt>28XE>K8zKJ2;PJv&Rh
zZ7Lao$b_~PAH`rbDC}5lC$gpT<Q^_(3jNmUWH_tTlCO!5FmAkhO?1>T+(JIr!5~v#
z0{>@WCv1anhY1@`kFB)g%supuU{Wsg8=RHpr?jqCCFZ;vLd7A6+Gh5G>*@FO!!#H-
z{R0jcK+APHqqUhBPLO~-uF@*#w-`f&6^^vi0*yp)^w(G1ZzGpwJX?}_>npDlz6#P^
zS|I0e{qC`o#4cl&C(Pk3`F(WtaPmbt8&Gtxwho%U{D~&fA*QyQJr9dK&R*2~Oq%`R
z^>Bgh`O=P?uC$ia=veD+2l%duLCRC86=%LV=Y8xb*k)myoUnltWyAQoJI>h1i93ZN
z`ZVM3di3|}R4~J~_(i)`xt?V3R4=tAW9;J?sKbtrnO)zD1m`qW2+$aP7HD(-VKD3k
z_1OTPf=|&B%<dehoo9%v7EE4ofg(?8`x|EY(8}gB&_8o&^6D9lVs=DhLPk-B$HeCe
z_TVu_%=K#na#eMruLb&m;M9_O+I+Iw(|SfVFM|oBVjod0O*zRZ5>r0w{6Fpq2jocx
zqz#Uq>L1r{t(?rzKFe&iC{FEuy)XZ_>o{-_m4Y8P8zYoc=+3Tb*M<9u|KecbUSkq6
zAuHo5=SvCw8O-i0Gv1g3iY<$6f)%IHNteYTV277=n8_Y*_g2pw0kzx7QmBnU6_=U=
z^P&F=oqPH<?d2|sQQJI22BRG&@dsuOW$ao>5EYS>_=Up#e2!u+XWn5vZ0c=~4u(hL
z=RX#=1!$H+RsmnN4nz24!&z{?Jo30_{d%sifZV<fo2eQ3779BdTvu&I58Js%<=5e^
z`{lA#AZ4juk5`*q4Q5k3z=sd;a{^y~VtTcr<M}~tuBejufrY4}u8EwsXMZV~fmBKQ
zB)&?V_DQi@R?J>Lx#i;jqwB4r+H9a^;kQ7GLxZ~ncPZ`^cXuo9u0;zJcXxMpcXxN!
z0>w2*kduDr|1QqOf18zLWvx7G@7Xi6_nf%iYqg>(&DpOBAld1|qw_5G`%&3Tn+W@3
zEMT~;(UW^0nn=V*i`5R!AamE7aKQIWm0aaMBr@3W5#mm(*E6-f{T&rai(EJ8H({rE
ze=%u&sU1)3+kHjf=;<c>JFJnA3ZtBQb2G%JWzy2!;3nytz$BqA{=-H!c~tM^y^L}I
za8v73g*F!L_Z|{3r<Ye*+K4K0mIBU-M=BYhssn%Qi4c%oy|?eNHPEo*%J*hSHa@Pp
zit{YpH4F~<`LJeY<RX;-h@&7GquW1&)q8CF4pM583Jtq_?~mQDeqwl=UG&zS#jvjw
zkq@j5?9_xsFWTh#1(IuiYwxj)zn}c!(*o^Ia$oLHJ)uAKIof#&Ql@;DCC|AUld4)A
zK6y3TE(?^k*z0K<-3e}|RXkI<@a?b233lW2b~}LQJT)jBM_{9Cc&43RJ-qjC!g37$
zv)RB7t>6q;ZL79wJ5A;G?A8*@IgB`fbAS<Gg33gC{CRU!;0UR8VWK~Fl6)K>7kiql
zHaZy)j+jzM^*vnwpYlcU3E&?qLHcAr;fF@zC2=8Ez#27)l{@+k{B=>sJ)3^3j-e6}
z6N89nen7($j_JL%@59VP_VtQ<xRPhsroSqHKN(~np1GVH)^9yC1a`=1)vvEW7Hk%}
zqZQNosJjyU@pMa;3ZO1@Ci{*NS2m6{Zs{U;DG(8Gb#6=FM3PkGYj>hs(`fWzwA}?+
z`-iO~5qLXpMQLKrw{`7$>vp;Nysm&c7&XF#3s2$TLAS`5CMfL1<#-4Gigh7E%4=P$
zxl$x-c{rZ^*-(k)I!(-k%ilr7$}{UUnPYtizZj_`qhi?;%{Ye%Xgq5Qay$PjC5ZP_
z3ii4K5&CMo;oQ0Dn%q&+C*L~{oTauzdPo*YefFY1+erjq4S{szcY=?vGS(i(#hi_Y
zgxtlA*i5_qV6q2hDPXm9V#ZNyu525$B|#J|J=swcEcrNP(YSXjQmh$8919)GA*RKV
zUyd906>A#>R8Aq#%=qSym?~2H<orbf%|XPQPh`ZvJ;X12S3eMM4+!>qbP~%x*ScQ@
z-^$UqPj35HnP1hu6m(hrQ0+*UWO{!}9ekpCw7=cK4CGT}HHe0J^jqtzIN{Jxm7)vz
zyLCFi>MS1QN9p#l6N^_kek{>k#keM8a<6mLFoE*r8`7|QY0#c+dBX>XS&d<h(cE3l
zR;c$KRuX|zUxhlnDnoL}P{!aB_*SIP2DP;G;`u#-+9kUQd1T)h?>SE6BA#+9XL{2U
z^7z~LA2`HGuHdoD93}1NdFx`0yrlz6qKf9gTj2vdJ+ZIb>+^-r*I>)jEUJ8>-y|MI
zm%7Jn_DnYAsgOE)p0sd{zo8QDf1%^@z3(;1M354ZspW(#=hZ$MwgvVQb^no{2-Smc
z_vY6KG!g3WV*^VvqF}!En57I2>Ere_W9(d~v~n}BcRW)h1j{|WLUN!tk4G|NMj^P4
z9O~0bSoNy&PgHBAc2Ub#b|)E8YyG7E!F>PI`~QJDp{U;WURRIn>rdd-l4{gf@N)*c
zqz1ShxSt3CbI!C;`BNsxQgNTDfW04P<Kj+`L~Y<Jm{W}wA?+77U`w<gT_%C2zA2WH
zov=4C-DHpjSn^sEC8UY|Zs4tSF;a%FFlTmZCjOrL{gCC#vk3FyvniJkD?tVM*_D_o
zTl@63KjHD`+x5lj2bS4___>?1w}7pb+2gCOO1X`6irU;1#Cr3(SsnBYovnFsFnZ!1
zq{}{dM(&XspiWV-#beaE{N`iuZhYx@VMf|{YeS^65B;lrH_&DyEFNA6T>Gy`DrzSu
z7dG#N*@v!_A&FfiJ<s)FAATSEof2zjW+Moh%D&DGIzz4haN0GN+xUy6-{Ro6;&gEu
zuBfLHvkX~W<Y_9Qz+4m4LSG}W#zw}gebZY642|)@#D-u@r1;fmFLI>6>oo$dETvke
zCZi<Ixm>#!uoO3oMY@Lcmn$%Z)aFPEr_OI-8L5qz<hRG8!B#`Jxz-|77!cECalh)A
znO%bFnC#leY76o8-7SRsx!^l1D6!_2J;Um0Vf~Oje-}jXo7x}qrz5R2i2IMOI!02v
zr;;~p?gqwS%)V+r2?co&pW86`di}{*MwKx+V@drgkHv4LvB|^B_YFk&!n7Vk0sc$@
zBls9@!^9w2G#lErkZxwf9bx$Xo%U3_=+-bThxgnW5iBAmz<6h<?A}<HSZ)14?4T*J
zsWRxh_AW_y4K@w9VPv`CCG}%-ds{zs4eg}LJacLX=RR@W39KOF>>!pf!9P-spqcLO
zZDuCya6QK)5EmKSt&?y%)FV@E6jIb9w@BE<Y!WS?Jj-3gSYI=czy7Zmprx|^Eu{Y&
zZi^=o&X!;PB^cp@>3@oHEY!bPbA(@`9lQUEIWiPobmgx{8~d`mYQpw@5dOwfi0*!h
z*L%&iIQEGDLGQ&c@TVAPK9p;&-UFe2N0KKq=H`nv%B-4d`JV9kXH%1W1?NhcCVhi*
zP5+Dnm*X)f{`UU)<>WWrLK#GQzUTwL2+M#4!};}tcvjtna4{TQ#*!vX-jIl&|3&RR
zFw-M!nO*gEPxQy!G5$d`mg*@U%^*Xx_Pdkf;`yjUHn$TrUHLfw8$gVg!nrP3^qrvN
z1&6FDGzjCY^NgDo#Po6ji=@?QP>iI3Pqq%2lW!3KISc9X$m+X86=)EdZYiaY4f)zp
z|Bx|3v-N`WA`HCxAP26<Et??#PSVwc+g%*tBEo}aDyvp;6{_#9uA=1=!&%;swHP`j
zlWa$gfCKR5Hri(_pY0vHMX!y?2kzX_mke}u3DgPOhT+tW+Ev>YILai)9f)e1nq!wm
z%H4bsAEn8&-I}D0wow^J<B~=>g5NAxg1{Dnwh?U8>3IyZ?cyv$@1qwr<_EWt%t;i^
zlm#FbONKIB1h28g<#zVprbkF$ZVVhx@Vw38>t+*Q!Y+1F{S<k0ijlB;6pABjE+Ofl
zUpf5ZaIA7BzEp&;i6j~Qe7N56e!<fIXqH-^e->=b50T*}$ujiwedfXHdS8!f8nLT&
z%U+5&xDtpedh-Eal0h2N*M=5^K@Jtr$9~uGb#@)5h6oH41=%MzYq*L@TnsR=!wKy`
zh;$(9=RLt*!p}}7-&#L97&pbhmj3Z_09-qhj`Z*si9q6lwZn%>lsM-VBukVPgdpT~
z*r5<XsF@#{odJ2CZ(%p?&z_Cv&pBNwbI#k0&H9eS6<B0JoCDxH>T`A#|4m|I<2OFq
zp@b7wX>gBjaDmPckiL6WO9%HsAz7P!n`!jw4eZrBy&kYbto9hw=%NqTffuQh;d<PT
zP*Kxg9{QYp=0><3UH2_S=BCnX<jU|diqKDfGzanaIRCYc`F|hX|6-sGGznTgzrp!;
zxbCPxn^gw-_yfy8ud%3;kg}di1pzhY+2C=FApkQX%s3K5U~cS_?hR&nzW$V_Ssi0H
z-2Ve<%7S|5mt+8VMB~B4>qgYYF4^(gD$n!>`X`EC=G=eZnQWh`DV356Dx^cb7B*3l
z8^s9)J7XvA$x@4QJnz&Gr1snesCwnQ6CzowA>&eWlKr|~VL9Qk?nrE(Yw<@n^Id`C
zP)}o*@c7sxOD2Nrj~#-=FsUkf+e#XD_;0%&qcSAT3N2|)pPgmA0!t>Olrf-v4J$13
zTd@xVlC7GWBB=_dvJp^OW@QaZYNRfgK^@r+rR2NYxxFxrRYqiJu^^{w&RpWQ<IN{r
zT*f_);bsyV=dWNdLde(+ChqSfZ1+L0!L9e#!@i^mwQZaUOxFOYR<VsI!R9^rZp%yg
zN5hHV`8tR(;sZ}k!Z+VxZWW=Y_@(6Wi_<CM_RxsPXq;IaHBNn4U88Q_jdpx`OED#w
zeeQ7o8jJXfzAuAo_%*%Gc`9UnR|NT_=NhJcPxlpm7k`T<4ilSLfnb1f(8u<jq`Ixs
zaXZdtJ_4(mWKGL`3BZ09T(h<6BCQBhXk@VkdG@3R*~?o76-c_Qp0Br#<4ziW#2mV-
z`su;~dzTd_{m(l-`h7I(KM=r~$`jL<dy}=Zw1plg-;C)PH+`lVnNGHFc6d^f%Eddw
zOktWoq7hB{{J?Gg*DJK}YwH8G8uZnngU!iH;dO18wNu4wIXuJjc+}!6j9CUs1_IS(
zj(jw>NNe05uXiL4&pAb*zNDFFw_qf(z4s%Z>O0&;6!-6T4G-!Sn$XBr2!DSJrR2Mm
z(iI^-yZRSS-v9u+AkviP!}q?oz!6v?;^#O29by<5fvdNh9W_J%-g`jQ`BpvR%L~dY
zAFPD{0)#2LR{NJaYp~>YH{=A<O#o92D?jq}gVd%!87ui@t__yeeA))f{;4y<C*;lf
z@nGLf_*1@j#I40*3xHMgWIL48+%)54)WRmK71Lb-J{<FsY;*7Mw9v!(c<zsc=HHfh
zJtpNUI6iUD-udC%gbP1elSm_ZLUVV~_xk6ik%)+ahK!ljkMXt?aG^cH++)mjti$h2
z)V{a7N5DG?!HgYW0s*&s*ygAjJvKrj-GAV6Q0<s$V5>_pjY!!eDMfxDmYpGLf34{z
z7GDp=Yb<&47azP_tB!KY8I7tZSI;R+f~UJ*`3f-rBI*lmZTlU0dVYVh4<e<-spC}C
zDRK5dIe`rFHMPM7IGe0m%$RDv_!msurEzh^X(>y=Xh@(+31{~H_L!Fk#ywn~y9nS}
z_34I1B9CBH1&E5EP9K`j9uCy0wvkL>Gz=<LV3|qOe@t1-p}DzGsR`L1ertHV5diUr
zFEf@~h6r~A*4oDBk_bQM2K(Rx6M}ME@5+uxzzTXjd|o&7AmS~i!Kg}Z2Xhdi$xJ*y
zh21C1`k0YV3y$`|4%w3ymSG58o+MpCKQpNWn|kDRy^xIeF&C(-_BV-j4KtRMi&e-^
zli)vmLn8(*r~q6z_y2u@t=XUV+4HAN;weJ>A?4@JbvVs9lm<~jjQUTFQ(Hl;s`io-
z<v6SP2r^^T)mK{~sBpeG(U$lMeILJstP)fF-qQ1B0$Dw8WCZ+10^Cy9n?99e&P2HU
z^|!CrWK~8UDu{7(Ol^1|`101FaGDIqP_Vx8eQtAy`44?LHh?f9sDsh4KiwD&_lk4p
z#iko~p~U<kL~N+noRVe$-dy+N6VZvP!iO#}CNO|?)0ch~w+4x`moV#$W@acXETa%<
z#q>yB%F)IW2u4JJ7mI1*b=V&^{yZ&OsKxOM0x*wnQr)jUe_t+M5WgvKtrqk}Pgo`z
zXp-2gWCVCveBg0mvbOv2XM9zv$}2=NheH)t*P+RLES<Wyi%Mn9%zL3j`8?o$q@Hm0
zZ^!``Z&RrpRV?j1j2ND9Sa%rpB!jTe8M{siJg84d+o!fE0V_5QT1Flncu&(se3ezK
zQe6D-et}y4<~TIUs_^7FsY?T5N-;+w@6RJg-=20@N!6$@n}`$o*V}j%SlpJBV6p*l
z^hPX=m$fT%0Z<3<{ubiRta&PWja6z;$!V;tUZke&Ti91d{pkl>u6OV<VA}G-s54D`
zvW=zrhKhxj?A*7-DYP$Y5ywM?@~(5aHo_+3vHnX`pU1mWt`5vU0?`1|4dRZIA9?Uj
z<J~dCoo~PTA|9o>Wya&tyUy}Oy9wykY~ARjY3K)I{)7a~nwoVY<bCqm_CCcTGY?Ny
z=}ez3L%~a8n0zVt0DpZ19NP(wJS|DnQ-K!;VzJbUJdIuxV<r;!zIaF?<P<*zk@l3l
z|9Mx;lVq&K*(9L2JKjAWov9qXy<gPY+@Z@nV_iQKDnfRTCCz~JQ*}G7;5C`zhKG6Q
z{%irCA8~0IGC%hqXi*v6zjT!PmE|Yabk@9IUjJ%mIF`FmH6OQ9aR53b_AwMH=(gcl
zv$IeDsu0iTX4AAYDom_tT}{RByZ09e_xWp$)ZwLH>YTSdq-ugE19^FRE5Jsp(clYf
z$I#v9e{@0)=06In(&4TTLTG(b`5Nkh{s~qP_ID4KFSVs+1!=izX}%YWa0-LyGu<Zk
z?K3#_arN<7fS*-YTsik^y-V6d&m3TGl}Ab<_zvgt)5vgixT|32tzrz4xfqR}06<_)
zNjiWit{owfCYO_F>h1n?fve|+uiUjXI9<=mH~#JA`W@PixGSpRY#MlI6iQ<^aGF@E
z0mbfZ#jK1$Cw{Gb6*v}%Q&t+Jv--wtgIl8zOsCWWlm5C2le~5kl0UfSGg*Oi{^~<z
zM}Xx^f{qxsN1#XW@WOlPryGMLVXGZ&51%e!G*a@yiL%{W%)*0fe;O0FC+^Wp+o2Y|
z{uy9VbJc?a{I9#UiVPcP<68-(eCa|Lip}jRxuwq$QHa3Oi)@&^R2`?vG5#|bngkZH
z@a!~+S%mhx(1F30E2a(=BFul{qn8VCxD4NlOSgQ-UDLy1Yf3s%2Zx7^412#14i49L
z<s9)M2WBA4hh67Tf}gHwp^o_ouLYF36tJA^LOYkr3&+}uhotJ*l5Pg`4<~C>eRU?U
zmT(?25l&s54;;DLr`LlMX!q5k;-7yQoA!lrU|$J9B!xb+vafSTYLmRLm^D0{)y9O8
z78)uwS4$Grec%`d&m441vL6o4vz>$A!<Bm8-Ju`@S6->X94V*Mn#rfK7QI%mY@Xyw
zho#u-sSdFg1FMuR^#es*MIlBJG~d2TE1tNbV(<dMDzp+<=p&~sHBN9y7Wc9>KA<sY
z-e))DnO45|pL9y8szX6@qu@XaVSoP?RGtF;Z^u1dj&!i(VR6|OUVnopyj&B6?t?dJ
z6{n;g_-nJA$U*N4p~+Kbair1`^7K@%5@_z0PGywoGAT9}$`%>*{u2$Qi`Oah5O?Lw
z?DWrzc$6_=E?1xtp~D!Px{ivB9tt~?1{O*r<)>~94CMykJ<pW>8c4;Yre+BjyFUQW
zkU3t&#bPB%S1CU`o~K{4hSj;+^SCAq!Twy=ug}<jcG?{ni3wkwt=D$+SaBB}vCPow
z3EmGNMH=)iyO<i;L+~z784VHh$zY|BbbWT+f{#R+EjtcdS$a1LuVQ(ac+<e)R_`_g
zG9)v_N5nlWYR-FnSgSw_wp8VXitY0q9^l;>KDW!-l0Cv!wOWmWo{woldP7z<bu)Xu
zUK#)PDtzm;96h$}K5o~r*U!`2ISTcp=tNqsZAL6tD`n)pj9_{QHSl}hf%=<tK+0#g
z54xtgAZ2Am`$jn<&xR{c2&h=Dgu1#UR4i=u;cH9Zu(w*hf`pOOT8-hIZ~Xi(Fc_Sd
zEu+^>%$I!&hVRK47Ce$m16xjRXufuxEBt*fG+607S>2@igEHP(B1(GwbYpxtk;0^8
z>tPDfN`C*-N~Ai5j>iE{(qD0!y>W`|vd7h}QwC7BcwJTu+9b1CeoNtX%Ukh_26xm-
zc^FrJk3f*cWh@AyCse`97R+#AJ(3ug@=t(rd@RgOpR*DD#??;icTdK;8SYEB>I%eR
z$6rG#Y^F#uiL7Dw<0uL9kJr{K;m(3#&8)2T$>>1E=0F6Cixz0>Sy}4*fP7iFv&1`h
zfX;{u&2O}Wm$U{G@dx=OrKO7U`QpQ@Lg>&&G6QF$VP@5qg>!joMRjptA+aR!Z^OIR
zDbH~St&948=X|Q$q-8D(s$V+?yCbX2PSE-x-RiPp^%QDgW0@Gs_&wT|vdAx8cqD;5
zNg7ZG=vp_uozRh$)J*YJ;U=vZp$dgCE$++ue4Ve%b2_4;T-pX>Unr97S11Pr63#x3
zwzm_Xk~5zykpx`{7Q(BHynq8r1k0VUj^MA7{xrKi>3W60ZrEtz-m<%r)y7N&p1pBZ
zQnk#sWN<_9-W=;5pJmblim8{&F~O9m!wgk5sVP@cxs{RaP>UgM<X=TzeSmyMh7Sm&
zf~4DsEfEm4b1d0s>2&OAp;^rgg#yD8L`N%toXLFbrD4(l91<ifdDJ?@zt65wNsga+
z+bxujWG1g6zPp50Cu$EjlmXF=x}UP6vzUr@XrKI5|D?s5@IA>+&roUZIt@Ur5^NnT
z|3DbEzJ(7Qc#|>V+VgI65&Gs*eU)m$$fQD$@$W4%-hTFwbgq&jSY~_hPs?XnQmY@|
zR<(63t3#ISevEuBex;;hvRadEC?8_(HYWVKHap8J<ODujD}*X`wJ!{5dC?31{}|N-
zNFxb+s^Rj$?EV(yl}GacQ2^@P`4_<^<4?gaWc#WgL0-GWIqwO}cRkJQEX^XeCZn<X
z_HWpid;=#*5??*mpWKmwV}QTT4in|0mI)NXyz&+wkgbF4C8Y)v0XEzgO5shZIFqd#
ze4MU*=4$@MxW9`WOIt+*lpZTX7gS72e!n4R#!&~dE=gQ@!OPhY2m{i$>j$F1ai<QB
z1M2l<E~lUMymJw3nb|F=D4R%IifHGW`06a%JqrW%Pi>}R`TZiHQDSMg>I1oa?!x9`
zb;^G+h&R<B*Jx9)KWgzaLKLc1dVSpK>kd|@9_FJ*CqKoSmC2n887FUfOWn&W-d|Ky
zzn!y^0?}zJS+AV8bU$$6RvQzmI7@8hU9MIaB<B@e<C)%DSpVXV1Rjto#HV|Eg_4`L
zN);lN@bXfEyaF`^HL`VauAK!Ft8!)dryuD|A183K3vW5~no|OUA0>QNJo&cAk<vE~
zjvFKbvJPysJ+1_sv8iU@wV((Pu!iadmnk2b`!*=h{UHI71%?B`kTYSyEiIhu&9;cv
z)@LT5@eF5}c`V4tp3)2ooCzkoULyPYN{tg8=sH+>R6Vt)scZ08Dn|Cl$;v0d9Ggp4
z;vVR|O+z@?`SKV^&a{<{%Wr65hkP<7u9oui{`44cB|Cs46OZ>V&A{h-#IGoWO9Yxc
zGLSct|Emeg!HAh(LwQSnbN`&VQIFr7|8(k8vC&l>jg*20dH~R4o;L3^OO-JV{lTnI
zL%Q$bqqt)(s&T=c_g6OS_P&p&8TovH(0{iv&Q>V8TrpP`EObAgnj;yEz9LIwaOSI_
zp(Pmiga%*|5nuCQW!-GhX}_;FYAF;KD^rf!?yb4gKJbH)ui>A9D?0sNC`E-$#b1BK
z@?p&N8q;kBT_kPGQvSLAc&554)bz(2Kml0$^VN*6yBu<pyY#a6TuV7vhTJ#Y=LPPF
zjQ3-;tQJ-pTh_`ztO7eItBa7;d}?wT|A7qwm65Nvs44}cS;%YN!~va)hxha#GxK?D
z>^qI&M|zL?u);Os=}zJksO-|<lxx_iqGFQRgiCoccGQ7^uD30wp_`SU1@)6Z)98ET
zR?Yyhzt&<1NMrNm>2t11mb{tQWn?#gB;tkZ@t6@q#=tVoUn?Cm#KgEDcv09luM!q|
ztHjBf#ZtUP8l=Z@F{RmQZIuU^Em#^ETW(iGe0g1STF1w1#B!!Z;xg8s%hdkq6n6u_
z>`vdm+s=(S^?*Wpi88>-+alhQ^hv@Q+4kUm4u4ARtI*f!ou3dnqK{#=Ti$Z2DAov@
zro)!!r${<toJZ-$FlKnJ$kse{rb<~Wh1=j_Hx8~mmjYVdR7dw^2w&BwG}g@?FHDox
z-|ILVfrXR86+IQCc~5rH@=Y~TVz5uBHpOK~RQ5GNIjt$xe=cw@ql3;h5fqHY?EIi`
zr@0MHxzeW6=P7qdimDUQ6C-2d61j1;3p{gV(`r~crxjmjXz(d^NP9GSa=G!6=qlGD
zRYxH|H+ZVVZofAPHvX10j=utHW@P>6660sE?wZ`Kp<^tV30QaPz}%eG6&0vKR2xIb
zb12+ED>he`O(u<9J5Xnh`mcWQ&m_8{tDi+1j3h6#IM6?llIfrUB5mqLvgS+mS&17;
z3i{f8M_el-0Qz6_wZn64&O1=KeBS^vL!Q*sqTv8G1LgX}M_VMYgkcS&qH=Rc#5GE=
z7=taCum(!WmzR;IqL$_AVrwhUG?|5SVq>h!Q8U)XFHsBWahMB(zB}GociQ_-vyc1^
z+M9>kYO8+w{okR>H;C1yE5v<|zcvcSF=9B|f!+`zgmXUz#tZ?T)&Yv>rR%PCf$I8G
z=)eCAH7IVjxngf{T=quoW+?lN)E5bpS8o1vrdWloiLHmTh}*+!U6xrB_OSowsL+TK
zTAfg&M|zU!|8f3n%tE*2uSS%V7322I465_dq>`*ZoMYe@<nLKLuFB3<-Zs%f;s1yf
znlBVrp%CFl5+26;_{nH(&$N;b8~{3*%@1^IDJ!k)r%a5zkutQo+U(-$`G+iwCc+&q
z&tBDN25i$>#ef6s#A~`FHls4E*K(KdI94^=E$utYVi$}}Dw6_XqZnM=io)KsXX;o@
z9m~U<&)rJhUkdw=4?<#E{XhbsnH|(P7jbgxmOMKscxk=;q`gCMZtVcZU{2y5N7NSi
zRD^PC{5oz@{DXA<-xxVVOIyp^_Yu*dAK<GE$G2Hmld-tZjZF4Cou(HQI?7WYC-Vhp
zxQrQk9O5Je_f@KVBVPF?Z}F5d3YeDX$ORm=9E&>@vl{0;VMr3rvt$F7yE(D9i-C92
zUb%0Thg$GU>l(&q9E<M*9jL9ff1ps_eH8nH;1We&jLUt-A`LpbZk`lQZOSlrWbrbt
z{OT`v4C>!tFX0E1&H{`=Wwfx*iQ06)@1~SF+XM#@*c-p^3grM^iI;`E3V6z9fRyPP
zYiA*D0Vt?VfmR<2$HF#<>EH?P09vK_)tpIQ*Y7u`h<3ioCt-L6LaH{u!cM=<ZGzC*
ziT*Z5FLrFMX?92-deS1RJBW=%S?~S<aIslhla;UZ#wGZ4C-e(Wb%Am^F>$Z6!(B9l
z@j_N#_ddgdUaK}TZh!FBHmBNRt-6hM!hekaHe5<fs4FE<0to^uDKO=>qalC)?DroN
zmr<=w7pxtsO}eg=1pGPBNz!oMPuHK?oc?_G@MzseGMml}dw&C$%uhbm68I;ILEK%+
zSPI+t5K=A{2?nIM273%Yt&Ezb?fQN1b=|&!$~aZkUivVL-}#Yx)Pd1Q52>V4_5ru#
z94553?&rQ)7}AQ738D|H?l?*su&eB`YP;nDDPa$dd;jcuWtGJi){daNPaEx~%s1wx
zbTvQU7eRr{CuAFOXxX4;e{M;lQaz4VY}Ur(oA|0S_JhDz+)I^J(zrd2=q>8Fy;GN}
zzz*}v`o<i#8V>Z0Ul{AM45TR{wMQfDktb?hBCd+sTBkALq&h&>%{)JpIv!EY2|apF
z`AV+7_v-m(1iS|;6V_=b-qEX6VvyQ}@VK+l4jZhPw1EmXWyK@UIeW9*8fDQq1$*o8
z8~!&V<M#XEBvnmLN1@qllJnGV5#S)8Zoa99j=o--+z>*omcbR8dCN@I$qj|9$PmI@
z3bw~!;s-4{s)C}|Bg9Sdil$8j21?`1L8PTEmwa07&QzM%96fi3=D*@m`ykx>|CLn(
z;S{C+2uFZ+H?vCMa5m9~?*8mr_<VuMzub|ycG_y2w%pcVRb_!xnKinSJ<jHm-!@bV
z#2Ab}K5%1?ME-q8F)&t_l_2WweC%6t_TA~99~?G+PLoo3GUeQ|%cmB7B*Q!*fC7#B
zR^T7^AOg4P=YdMc0moFWoW}3=i8iHd{F4j?OtbwCio3ge*W5ypObC@iHi`S?W*OTz
z0eT4%Q^2lCYvNGC<M2PB82W#O;<G&2F;<Jcbs%n6(N=M!>^F1NX<MEVO06<eiq?B@
z`=~<(rHyjz-Z&d$f*U8jui2>&87l#xBVKR+*TP5E@a0;2OB??CY%HK_-aPj;FN{vK
zbc4^bK{UXMKaX8{iK%D>9$<b5zAihD#aQKt+~}3GR^{`(?Y-ljakP4%0dbZA)gQ~N
zRlC169&+uLekc*>y2JwhZiciS+_x{lug}L(NMM+>TInz0SjJ*(-SYmCn3!Yu{J@VM
z?#-{Zp)OW!)K+`(Xj@4ExX_NaO5<bA4CZfrX$u%OQbwf2QV!+4e*BcF>l)ug5<me!
z4I>$wC=n65&=8G&)xBY0A9;5Pz0FpE@=wZ8-?s}D!UL6^L4WwZ)6ChXUmRE&XozN_
zO((dSPUKolV0~%W!W$2m5ag`<WhRLm<!8rL!P-J)r|{3Glxj@ifdfL&dtZg&&7gnk
zsNuD#*SrQ-0Y1+qhC}y#q67SdoI3d9@+f@c2w~vynY*#&4bZw5l4;aX7<6@Rei=o7
zq9`Kx{g^=e4!{HiStgk6o=cHJOPebPpPeR^LjNGxuXVbys?@Ace=syyBod6m4F~{z
zMq5Jd0Z?3tV_;$iS)xNeLu1Gk5=rpu3*&>zP&JaYle6G4xbG^^z(!ZM&5-*tc+>D^
zs45#){4pohEQi4t8U1X%i7uH@ue?PS8yow4+je_=!C>J1mi_eWt_hbNGNg$Kr3?KS
zio?9lzF?RtDsfj>^)XJNJW8HVAW>er<nF*U;ac&nh76FR?OK9%{A2*kDt%SpS|^0m
ztN>T<zwDjKeu~7Xr+z5(q3}|*(FeE%U0@XA0wrNCQ06NPI@KLjE!01)-I;N=iUI+I
zug4@E+fF7j;-@6>Q(exF4y4ymmswq~r-EMD=hV7o%pLlz-wg|Bb%2Hkps|f`-#-a(
zZ5Y5eO(uhmx-Jr3xP*)mJJv~Hwm37?*D`l*Rp(y`4d~Q*h3ZN(Ve=1I^ct~hNi##U
zPsP(p75spD4u$Q=j*=kQ^+qNlK~Itj+hy`J{qr-;^Yk{v6gOR?i?G9DQ-f1>{PL7S
zsZqSK3(ZC{>AN<kQEBxE<suGnS(<OTYLtHgS$>`6Jmsc<V@4C17^ijxi%!1Xp_{DN
zl8dB@cxgwcO6;NCx4bAyLq+Qg4L7~MVhA(sb~v-YGgb%fNZ-Uy{)PC#o5qw80FYvd
zm3{y%&DENV6ue7;oX?&Zk`X{kvt&mSst8<|LUH9{iuHC<SoV=Li(|HQ?l+qj{Kf1x
zh#UKQlf0W;P9k%hWhY86>JT_?2G$;TO~m88ZV_6JMxen-m(C^s(0e$b#iF`6znC@C
z#XoczEwfxULTYoo|DwWJ=sIdFImSa_s581uO*5DHR1AXu*~xNff6L({Th_fviQJ?C
zZt0Ao=~gI99V0r~zT4f0rdeXyvWKvizm$=s>%0^xCuGG%g1t9ic9yZ>h8%@X-V}A$
ze1eAhLWEJZi1&S}ioIko{?`l8De?2LamF&8*2oDu8Mhg=9q>D<gF~9GgXfJ_`k2W!
zm<}`f-e?42lc0!KA^B_xJ-NC5u8;?r>q95aE#si?CYLy6u>Zn;1NxtPMtr~T#du}+
zVS`8ZV;a+Us)A-`WAEl5VKAmPmt=E2mv==?(+5*sZ1PdJ<?PxmGK*XHb|5>;Yalrr
zecTTQyXKbfh%bhFW+Bt+XE}2G9@^g=SU1AKMyYQUIUlyu<Xp)Nwz2mBbo7XcW2S@A
zcVoD2Cqh(WxEVBQ?j%vE5wy~cQVuYu$P3b!w?LY_p@bfSaLm3K;}8;!>de&!XJU}e
zM#!0leXdX*RH=Fet8O|^fkWJP56lN?;71*Yad58C-mAwQ4_*o6CUzfx2*8^HM=7pG
zbA$B{;oj+9q;fPx&@tZ4P56=D{<T!sOHzf}Z|iU77&KN)g<(sf_K+9`lQPRP#W&J|
zFa7K5>lWUL2OEv1==UcZglP<{4ex*6a6L6=8^s{6tPrMsGglYDjwoQ<zdq?R&g61J
zJC@h$eb~j%y>Dr|JNzDJB&JxUj=5N=L%8N_vQkAA2kN`44cz&7di#g~<>caP)LI^U
z(hR~(YQln{UW?uqJ%zxRTfD35>-$R;8lEbejTV6JH_*mnIv#NfS?vBlcSpF$IPW2(
zZ^j1*rRAA^#XuWOU#?U|pKfyfTNOWVplv5|n5Ng@PBZEd^@evAw--uKYk2;Juj@H6
zZcjVXqHi7&Mo>!(P=E;tZ~y8TILer$qbB}x|JU(imR&vPWu8`4)Di!1IMhhHTqPuh
zz^srIBCD!3YAc;*ci^YE!;u54zLVEeLZ_41k9DT6gwfYKzKRq2TdK+=+Ulw+rNGh4
zfD!B|X4iG1>`WDHBj7j+>)lQM0>t%MRN*qTx0@6TM3DI_mz+;oTOtCGsHTS$ci^Yy
zQu+0HJ&-@5UKPsc-Hdb+!nA^4x>_%Qm)Ttr`|J_loPdC{#Y2jR^sui5rXgbw6m5hN
zZmNaK0GrI(xHwv=y{fH4x5&m+MN+T<uMq<o2cWM;cfkpM9qoap#n`bX;1Sk%LsfXL
zO1s0#;qoH$N4aW!6QSSve|tK$Ho|5psEqtS5WI?4t-hZXoQY!T-1V?4s{s+vEZJ#-
zjHb4qy@X*nwK=EOC|jZ^%R(iqk6SX$-|Z%2EQp4zu_Xep@R;l4)#EnQv-f}QaxbHc
zT5NDTQxU-ma<pOq)#H%obun+KQMl?}U}Tu&f)Y{WfQrMPBwk~7d;6g)vq<r&ym;DQ
zsZ}frKRDD!kMQm-dL3jfF7mr74l>rM4sV8MS8q9h#gR`)ySLR7@yK{qHToT$QA#^8
z(BG?n|5&Ni6Mu{zE_z<PP{)prB0w(J9v`Xt(DT8HiDBdoSFu*W9qN`oqs@I8lM~Lw
z5PX^6px3PXkEvh3))dK>1w(~L;dwsA#N{2ECsFvGaE~vo#!L>w$rU2+r~a|;uOI?B
zfBvufzh2D$JopYAN#q9oTVs0v>cp`8=tDUii$lKi>L62yuG|z54D`{5Y+B_Tz#X^0
z3WzAR_WaB_^ZwE_ko9$sVjM5Ys;{k@uBBn-hq<m6-Oq7t12SLH;0&I)#J%50)LA_j
zzoWxcgUL2Q=(u+KNPofGSI(3Jb(sRVjF%z=(L_8bA3>kcK26W23|ci+>~HAnA=6IA
zp{=U|2eSBJfz1yeWlV0iTY**h&1bG_?ILtfzE_ajl`Qj~gyGnR2yRT-Ax3kXL#@s-
z9&=h1jgOXyOgnD1jkUc^6)QI@lR>QjZ<U5>H7q)p^Zg+83muF2XVBHMp%{&6h*jL4
z>dp6PrntTCD=-7A4uT2SPru-!ne}_Hx<qiql_1|e#Va<Wc`Gj6GRIn~Et}<0xNO?D
z8;|LOpXppHx5tb*S8WqJpYCpe=iG?s+YBAKHCI9nz5;2>*ez}s+8#fD{+0WtQY;IH
zh$sR%1T+7F_VmgHMM(ykWhq65=$!uPX9)g({TEmU{ou~F=ljuoI_`r^&1698N<4+%
zI{bEqx{YaAX6$$-mSshMQ~Juq_5lNpx|l9|AUqC=8lhsP8B6gzWjwoN)Ce%*vx;4O
z!j>k_prh?qLd>aA0o5D<GAOa(Nn8LHS~1B&9INiejcr_ui#~s+l+sx*&|h-?Bt+c9
z;}5BdK>-kyCSdGYJV;43tC+(O-15su%&Y6o3;kI<rdPGvkz1@QB!$RGYKOvx066Rs
z4f<UaRkZyoIc7N$Nhn(-sehY18bQBfB7B~q$zlGJR5?h-9^0gGlKKG~9XAm9fKw8t
ztJ=9EU#Jc0{N=zp1~-qHkx|~C5FjWCh(Vy42BeLTuhS}lUT(G|&flz=<g2mfeM4a?
z>~DP&N`2$^sZM>EO&~Qwq>XR^FHBWyxriRbv^gWmtf~ucXI*0TJeZsHAQto;%=1HK
zF~SF&tbeR$7jYzA=~O6HrSjzAp+bra2;-R#?M%GN*h{+HylS}dd$c!q<o^5qG-4%w
z7Yfs5*N;2b_dZ-<aNzx@gO*J_x#h<-06|xDUS+zUJ7F=W%<blKiuXLN2iBYujzx+-
zKg#dvT6ke3<6|?k%)jhW?oy#KIGllY|0(=Hc~h#?U`8oB{v!+|<lFjkyhPtrC7DaV
ztJzD<SR-QajiQgY<nyy@oCWPj(D^U8Xd_+p6?xH~${mx>L)Jj@2`MKNro>Cm7LiNx
zA<p7k4mbX@lyJUnoO-`D;<AeXWwIPCzb~ggXa^)waWgE$r*35tNNGXf$s)ZJ@{X#_
zdRA&od&&se@P}A_{cRuIjS_o}${aY$LirR@kuP0f)DDe%o56WL^T|XdRO07W(cfMl
zApVSw*X-|i9&fizg#uKUQcE0aTA6Ay3#z;ycxPE$eo!4TaP=VW`lH^gHTPKl2t#&F
zvHzZ+K0FAR2_Mbij>jUClc{%Vm{>J6Zw(rr7Pe*{T^_Ap9}+gOTbS$~4YquOyRmEd
z@gfiUNT1d!P@L8@9R;6|e6J4&;rT)pG!o}1VQFn!1CiGLp%JGE)=7QVtgJJ7hRWHJ
z3SxRz9dCasRRPoJU(aMhr&gXOt&A^%wfG_dmkbptRYUb+beinI;x{%oI2H8-Y|qO&
z$0@T)-S%Lbn6RMJKhqhA#M?dQzuA;c6AaEc8#%|-9p*Cvyf2>-kBqHsBv6UWAHtIz
zgh2N72Wtt(p}u!t$SB_@6ORsUEwb&V@Qt^|A&C^KBtJthfcoOQAmqKDiJBHfY}1hj
zW-<>_V~jG(^fm-aThXWJq;{jBpGuWV+P6C&Ald+I|N8KHY2Da{B!q(vhEC1b*$uw%
zd<wJ9<DrmA5%St`8<iDpe)@iw^}M00+eJX{BYwxIz3E>zZD}}Lujl*(Izf_=AzGLc
z<hjOlf`NgN&lZg44{@*iXr49;Y;tZP(ZP~$mX-rsY&0`0SLu{+epXOWs5Gcixt5vY
zzOCbiqWQPPRD==VNer>Ykx0Rbc82{5%z^GV2x$OCopFdqg|o*muB6O|E~dLpZ-+JD
zf8K{LYXv}bIWh4ew9wsVyGfofW?}5GU+PV+YRSM@I8xPpf#JR7Fr8HnpA1nB(j}mB
z-9It@)V*Q`&fVxmSlS)in*_(}=pycP#lir3y6Zlv4uwI0JZ{f|512nou;??NUQN^*
z{^ob4dWMihH<;KEx=O*6cp48U+akYhM2vn@WP@JBN>FCY@RuxhFI<gjc+^3C?JsFZ
z^=xWmu*yePyg;F=>RoxT{jE0{jm%tT3}hu@l*@>89s_VK`atuL+5WzU$>#UqaqB~7
zTIsoUN-=SsuP@3%FG(7OaKRvBnXYAtIEFVUWc-khyk_V$qP|hry3?YtCEmOH-!eaq
zPt#Ld(41;nV+=l(oylOalii<Pfz_;{bGgyvQonx%C6QF+%h5L@o;5f0jXW3V0~#|)
zZrcfaoz={)60dJXqS8(6hW%d(nyqMq3wN9UtT42;8<4p3aJaw#rW=(*($~$uKfn$6
zAcFRuq_g}I?X`&>+E<JZh2|tgiK$rgCA%r+>7I+%H5QvGru{)}eJ#kb6Qn7H@)C~W
zql&QDXQADge6+cLbBn_hGgR%t#2b!=XP=k#4t%48->Zu(XrkSZCkZQYyMqC3p}otV
zHrC>~7>F`GJQRw+1*;*Aav89bz%?J4QFx(&crL*bLi1-^%&!?}JBtE&nD&ccN_!C#
zE+AjTxPy_N$}ra1_{DPU6}oV&jSe|fyr=QN6@I?tA7K0~{+3_1u3tdl2N6tB3TA{1
zG4|rQqKZP0qCmA~%Bbri6U~WKGZ8Fb+s*H|pdq%=MnV0Au3g0;QqqQ_l`*@WVIxrd
zT4-35p@-~UR&d|cmkb6DVtujQ%o6Br354UZ{wRDhg($oqTBiFZdEY^tUVsdP<0l93
z?3Q$!pX7V*0d!2v>wE82j^*Fi>pfd~#>w=k|NpH~*QyPeCAuZH4QF}-9En{SiC?3J
zDwMCJGy~7frmdhbZS|7~#bnUkew$3CURH3^+{Z}wXgYhuA}Zhr<(&I>+rp8ID3^JQ
z`1R%Tr=Pcfu$050&c*t2Wi-;92;O9Z;LM@@Kq92zfbgGX$MV>SSd71}UR)%W@RHWP
zA$wllFdqW&12LQ%*A;mmbS=6$Yqt7d)BFy!W3P?cs`OAm_L>%4>n``d&B4j@=4|q_
ze<~{9voZquGqF(axyw4`vYucoM8+Kk#F3>_k*KT>w&&#(e@0LZgyoj5aJU?ju-WJu
zI5Qap_Y7v1-8sBEydrJ~A?60wV4t#;7M(nrdcZ)f*bfD>XB7FIfdZK%Eo1cS4}&1o
zDziPz$@AgnP{UVry058k?I_|SM*>G(Cv4q3(oxbh$~0=_?UmY12p1bvmqgjaDHA1-
z#5bw|N3wsUpE9Pu;Y5hx=~0;tN}G*yUW$8-Liw*mtVBae%h374>^17ry{~}2sTnyH
zTzR=*=f8gf@CdDInIaHA=e4?Ub-!@AO#!xRzao+mu`^cGv=)^soXU$@)s0yeoIGCd
z53KfB^ehs`0r-MFb2F{E2a6UoE<60*1vuo_spHN=>D7C@e3I+cGOsd1IXz*Do#Yb7
z9TF*B9dm|kbb(AwL|b^tYY;bX%gknSD5Gl3sCN_3^7nS6#tG@5x<{6)I`GUESJyk+
zSwbR6JIdy>HLo>CO=>7kC2@~F@Wd6|fG?9kdnM6jFB(jy{KF}6A$!X+xQT3q)s>Z2
zG~5KiGDjGo)P)*+&dhyQ;2N8sv>81{Y1u<&ASx{?RD2JT@xoDgiNA7h$eU6uQ={ne
zK@W4YdOdWbZz*xcpCNUyA{prR>^OC}I0(N_o44No;5e!h9gYW26lJCSLnHN6a=47Q
zQsm1~Iw6o6U?UHdQj;;HYXbwDsgzIGmA;>9(H5SH`I^9mn78DxeBT`)5q6AIHsT;&
zCcs+8;wchK5>(<Bzdt`HS2V$R3{%!6(D@lz<WdyBAV9S)Z&6XwNl8V9R#*LJh5_?n
z{4pyLCM;J>RFc-!`z%)}MqJ1fI8<mX9aGt}^Vx|k!d8S&akwcGs4vAeGQ=`}T5P}}
zbyH<Fl8(gm{GN)oBZ+(2!}1{|YF_10ux-#r9A?7RWR%Mjje84Qj6~J2L0eT`d*A3Z
z#I8I^c1u&FQ>+YOZ#<AdLe08sY;S?8T#wTm5exXgab!fLYT)%f!wYU5hdisq8^7;`
z2eE^Fe6P46Bo+M{{+0@!i+%2f#5c8AIrxulXA{)uEZ$+zKv^z&wQLt~Yhvm~@Nxz(
zSYiUAai~@6Qg2jfyw*GcJ3>C_YQ4Eo2w?4CH&HzZb{&6>_%kt}@U1q$Dn!E{^U%Rk
zscM>TAGf!?fBRZ%S$4fO%Yp6rau1kx4aRsHyhHA(u2LPl*#q5$y;0Eul>?Dhgli?x
z&zW8x@L$0xg3WfTls;+~AI5@^!6Dy**hvtkCn|#$Fx2jQ1w9j86-z_br6?=PWt!=M
zOBBtma(wW#hhAym{~w5uRVTJSO;M2c6tg%ohY|2xnl1F7^$&-4rMU42<qWgz%aiR6
zM0$q5Vs>kM^F;%5&?C%07wyAq_5^l*4pcbfo8-1tfR0au_1<1T-x)tW8B!}p>LV;Z
zR44@FBT0*thXf*hymg>R!tg-UrBj>w8!09GaTr1(5eek)A{xK}nwz&L7L<?Al>_lT
zDRr)Zwb;=J5DLiitmNG?6xs_#T2vhP`JKxQZd{#y;_VI{Ox5FednaO*lLI1Ji#jk5
zIu}Cu9AUIZY&~^TcZl6oJQRUl06B;5>D=7dDC*(Cbz-2;1$Fe#ER?ojciZFu)6EPM
zK>T^Yv9^HijKj`WvAIy0^c+zqUTT3NsE81<!3k+Yz}5%%1=V03Ue7AgRBG)eY854P
z%)P}xCA>aN!RjBUwyM%AOi{@n)|QI4GWvq<b(<EP-GpR$qljhfN<1=kY%kq_)2GEa
zRsVY|0@%zgZ*^d5tf%sZqb;VDY@vRm-78PAnM6>iooIACwsJ7I<+10yz=%COZ|Z~t
z!IVq6Vg2ElF;kbU(??ud{W5Sfiq?X^=Z?Tyys9svgGm!VN6<U99vc&aI%dwQAyV^B
zOE5nkE32+kxOY}?t)%?N@Ot@5W~O6krU<A&y*yt)YEIg2WZ!oIKL-519WwA9vNkIU
zLry_c1+1~w8O72^?}{zaB<k{6sHZqpup4pm^7K5TS&q~Vtc5kS>Dh?19FXMG7SnI3
zw%3bgn#Io-y@cFbAk~pWs<x_B?jMefQn0etDX~vlT+YPf&U=oTilV(&@)OM^+Z@E0
zv>j;{3tOis%v)?uYLJu{j|7TwiHJorE&b&wp_6H#XJR7mF(GFK?gM^S-8D{o`qC~_
zEe$;!EFWC9byV7gj1l6%tJ7mpNj;o{{%)L;0sm88*HK1d(^S2QvMCy1A{6>7Ib<a5
zt#Z|%-l;SrygeM-lg8S-TC-`_bSZq?oH=l0z}avbxymtz_$+hGX*I1hQ&8*M?-?ef
z(cx?>E|h3YOPs<}fFo_?e<%b0ETC7jn2(l@qFDZga(JXWxjk<$FSWy_x^eTrp8s6R
z7shN-h<!M5u-DI~6Na=~{Pnokekyx|Xdo+l>g85Wd|X7gUnl=BC%W|(4~BjWVcSJz
z$;)fy(r~Y)6CnMMdkZpixL@t*d7#>mJ?EJ(G*hd@XJg~Ou;}>os;nHWe1_rjT{rp9
zj+n-fSQG`UCr(Y4uTfGh8vHHt3hYjNeQOHCnB5f$Nf_TBA0jDpYtmL@*)8!MHKZ#o
zjeA3Trrq7a)T}dM2N&8E@Lc1s!>(0^{b#n?Z4a|EcsSQnk5MSvi^u*GMw$mJhRWT)
ztn*F4H(fNC&Y+AyMXO=T(%j+w^dWwI^JDk}#tU|w8AufcugB+g8yQ7fj7G7)s$d)p
zROsfLe>uEZeDKTQhc~60ebm|Fd5jz1fFVs{qZxSP@(lfgD<AgJtp{dt4alWD-ma(d
zd;ze;&KI7(P_Ska#N-uxDFCPhtHkeJZvFgaApQ{z`$$ylQY}hSsg*L{c3NX_{;}23
zqWSXZl&qYLP$cu#?9OR2UMOeAW+u1}i$U}ewO4aR#rF=mFm&*!ac~?ldwraGRl_?H
zO~{N388woIG~RQ{{Dl%I<!Y@{lvRU{_6`o9{ro-_Q_+YM$gZXSiE;`EOjnuz07>w#
zD*^+0UiwbqL&ecg-Pzsar>7w^r_<fDXOszhmqA7sIE1*-aL}><32%}L?Vj`>hLO6W
zioxYCD`X~vpI}o?MjbqVV?jSnz|*nGhZykYs%|OIU#5Lul?+9~z9EM?8?GJ*oIus=
z(NE~yT@`v1;Q}kN3npi~UCN6?8ok6#`jA$;AZ@xyYkd&LPSp~4`MM)m_F?)5OP@ek
zXoF4cF0?9!*N^XB#x%gCGmV*VFQ#nEa<g%03FuX9W861)oT9-6&q{a`Y6BW9%9wyU
z?bk&ptTp=IWuwbXR2)hOa>}JZOgd>T%UDes>l&%xhB~OSx3RGw2s&ze3ewSu$|^v0
zrSf4HX4znw{&$LSq?F(9m|g_1d|g#!(5KIFaqq?kL#&&jtD-@Yk%ihc@!6?@Q94vK
z3|45qYHw#lK}uCxJi0%X5`KimYSzf`PpOtAYCEs}=xb>V9lSjgWu{cdk8>|F`tcSw
z1M|=g#=dMQKjpHP_Wl-zWKQ1w%@2PC#SL^D((vLq`IS5INbawHSgW{Lm*8QHPKoW%
z5Fz%OtMtTHtAa6wk|JUb7qcmsm5f-#4_c}LTne`Fmlg8NbxmBkxD3VDZ;xkZ?$skj
z)&YL)iLQN@aUBWh`wQM18OGi>Mj5+T94IK%e-0-6&gQYjXJhPNTcF-`{P%XspX@3A
z@7aS{u)tIsB>ZJ;koTNwA8$*)hbS-Qx3eC|iqRp{8q%t78pMr8NcX^^ao@$}+SvK2
z7I{{2pfdfYZ}ve(shKyHuOf0NGSk?j5O0bxWZjoqBHXYZUqSOkMmFOf77}CC+s-LE
zue{-RpMmrGJmP)6McR3?)l|Z?4sKkc87LbyNNjEdqw2-7pqGtU9u(=<=-drI<$XNX
z){+~gb-$IJAsV}giGx_p`X8jSs+10brAP*7jv=n|0=BYyB#s-C)-3-Hmb(`2{%=Ki
zNw~PvlpU@CH!qui5plRZkgPR2Rasuy+;DEObFpRHc6{x6{v7zEdyG1&?YBD~^XQ^r
zWix3oXb^srm{FK!2xyV}1e4A075w^q*O0h>iiU#&hl(15ffMB32aweHr!1_mqYB_d
z95;cz-E`nxzCJB&`xcHZ#X_ODn*AXn;BIVxYN~v-RYkc2{%$aY+@x9l*xyDhE`2yW
zLJvkSD@byut7!l%rM*o}a7N;|+zra9mn_s|DVFQg1pXho-ZCo6xLxB_Pyt1{L6D)l
zy99=m?ndcO=~BA8Q;_a1fsqbL>28KjVW?rw<Gc5}&xf<lCqA&Ai8cSY<GOygm_^wG
z!hNlU-k__t0D&BZ`Dwk!@3`T$1HMkAMiet6fPf)Cyykgq&j0=>X>2$BJXtz^h<Z!7
z$ZdEC6UMeSDAtL<?joC~AIJMyzhIVd-sbT{{&d1bKbRmvRHmdl^5LZbGdAXrbuc5e
zua;h0=fW^sqtShI?u3+qoWsS0)1UdLhj<HyUX%%D^eAF$JPKV>J<3GnB4*~s#Ku_f
z`jTLxqK27BEI>qveEQXHR<dQQREQLZCJ*SP+Q^cUx*%mhndddb;=N-|*Gt5=9({&F
zcbzsb^EsnVM>D!gN5=Xa7a^@>Tn1u?dpBHLwF*g=)*E+xS+V+?x~aMaK{qDW<JMm@
zFZAPV-rDNhT2O$AKoIo>n;T|1Z;s>sAVOyIvi+;kf%bmfPqRBezzZ$;gj_Z6+CuTg
zsy?~Wt;nrTG=Hxm_t!gNsA9MaYCN5?kKfH=K6Y%1W16ERQ9-)+$i~_?4{_JwvjjCZ
z69Z8WIt6I#B)5fhPz$XrnEtKX*bNW<q3PvaTBTv)Nf|zNCHkmxBZ}Ml;jO*A?~mk?
zOTW2GWjBL_hlHMWGk%+U$?eNbF^k_P&E=Mwp{leWy%*ke_+H{5V&Iry`C<uo)w#Ax
zxr^(mCK(+LyO#wLKXmjNiu)(j$$xUs=YK==8@79vc!<eLKhFBjW&V1iGo!HHykCv~
zsnek(%cW#~R@!N$QJgT|Ja!BLO&m+AIIeotf0=ivt<%n-o3)(-xsDv2A7$H>vuKrd
zh>={<IEbzIx5A0KrvEWby5Ur+PuN~C*1)v=aW|!_o1WrGz}_-$B=8neCqfhLyX+q3
zC%*Yo+RvR<Y=Z%N<9RB|!pxi61R%|O{E{u;1wih2b4COT=6~hV8`Y<m=kc#}xW6$C
z$oyJ6>7;TviH#c&D#<au)oX*`(dqc$S-d}2facAym3eL_#0h$>KA_y6kpx{>Uw`lm
zn>*1zh+%7C%OB&m?(8ZIp0E1R-2{n==?`3ts&(6TM>eI0Yp){M)hul#%=ir#TdtlJ
zE2IyIma28)gvOAQC`(UghE75{Vq};AwL-pHDDK5%mHk2~rPT~rBNX@EA8{)3lmmGI
zbUw`jT445JtRlYgqJG}UMEi3v`W+7ccS0}uYS18tkD^e*<f;v7hHGn&U)!W{j%9@F
zw|7(rPabSLtd9s>f5;C`pJOBqR`uCWl{<u@m`H$z?pt17-r6Stmz`f%(1`@fpoB_+
zJj+O5JU~kKul>M)!EhG;?|yyxU(rj2Ue_0|Mr**5_P$B^UiafZD5iOysuvqOA400$
zo=C_OQTniQI@0*9{JKnz$a{x(c7Qq+oh|{==O4Y{vtj%$A{Lv)>}Z-QY6wl?!p}Bw
zGyD^t@lNpT2r|l_xeM_5J|*hj{<yL=J`v%WP{%;{Dx{DKuIj4DnHDNZ6M;@gCy4V|
zPAa!@rV{5PJcEO4c$#iJfWDtHfqL39UgQ>Q*oko~u5o~W!*oT4>=ycL$fz_v2a8fs
zv@6>;xk8pTwVI<v(qVeo>DiGnwOjcx_r$5NDT7$vy5D$mCH*dx&kLg3hbhn<h8hn(
zRmfxQ{1BRijf>xG)$>%50-vE~_`#bB+^ofTC}jG6(b%qK@tnrL^YnWFa);c5a+w+&
z1q~CIxXfBY=v#cbc&Pe9)3@8t@!;b60FZ7%!QzQ}OU7720VxAyI4XNoiu!jf#9gDb
zs|?NlHBKSgMNIgpezDzWp4p2*KtCWCs(x2;|3u^XLr#<j--9D`L^k`7h^56tB+&-x
zHr|QikLiE40Fm1P+fPYqMdh%y!iB0xu}pij9jh3u42v_w9|QGY0dE<CSPda~EflB+
z0e*+~zaV*Dzn6_qjv~X=d%XTs`+BKFY6Q-!V6XE*=bNxMZc#UbmQh+TfyqPvFhUpK
z1_vkn0$Q8DFo4d0{$BQ}D81--$Ala50IcWWBi*9+?-bTJEr-IjLZ23>>o!~L;I>#C
zLimX^4uQ?+lAf^LxK<Ph6qn?$M@pdvgvp{BaH@{|02COR2{j{Qa%zK4M8O5PI#?#k
zllYtde29*`>043}Fep|^3e{Ro3Y7QZ$>zOV`7XkyekmRsdGn73EHKqqdPbD@hJ|3%
z2vEoEn5O@Mcq`3705p$5lDF@af76f?0YL`-dcchBgYyg)ck-sVA0*PdU#&4$*gf;D
zl#gtre{nMGY<+ut>?gAb3x3tp+p#=g+zR#vjLttm{x_^-_b7Ek>G4YLe`yO0p9AOb
z`cP-84M_)M(U6cX|5RwRjb`zk;l_>sgrXK={8O$AwqB_UAQ|)g7r=lP^wb4UnC7M1
z+S-a|%{f)XJUF|hh#A`zOek-jScz39RpLM|_tKJ_ey&K2CfD@@h2@%#?eloLNUDTB
zIIMZp0<)(#iwB_BzvrWa-!oRrUBBJ1lt&bnHo?w1?MCOQ^nWDE&;)J&hCv^jub5xb
z;j0<i2o<wXEl%EtGuOKx>sjZX3^bZXdR*RsL-iMIDQ%U_lH5&~Ic90G^MU|0HO!Y%
zL^h-tf6~{=)vS3W+w6f-ismlcd-SY3T-C0_^KosU!@Q5uc|F^QKS<-O??|q0VsM_^
z^saU{F;XpxLZ98|yUFO2+^qFUYZL7tcE9(Qe%VO~D#t1E3XPhD{2fICt9G0b%VM23
zBx0s<D}V{>ftJ9WBt^Ue(PYvLidDAuuyzdRr@w!VEps4PBsW`6!{h%zc$RjoW3QZy
z!+9R}A^Q#98q)RC7-`W_iq=NJmswT2yr!Il8sCx!&X-co{+FFt595g<QB!cPM#w1N
zKxKcGddq@=jYnn7Kqm>6<Q{q+wde?3sO~|zJT%Ii&E#rSAH>!%;6s__$9mI}8kc6P
zHeMP~ev)AFs!X*Uegq6i1WJiwHwdP}R;z_hN_umSyK*9mK+Km^8zs-&4I3GvDhd9$
zMRqR(RFWBi9EgX-Yv1CvfW{kxXyxKQK2UQNd7GZvvzbffCdBY|6OyB@@BCjBHTgpJ
z;&w>l2qj=O_QAqd<!VSK|0?{1b5RZNyQyU|v&0cat1`y?$#Ifi@V)L<&WxX9?O+vi
z^w{_PtLwr6WBHhNiwBZ}H7B=w&E#?8r5&ICG_8yNSo~H2c>ub!UL(frio>8i#=8?>
z`%y@qpu`(DZ@d2@5#lxR-?TR2Us^i>#mrA*-I@z%KYPU2G7iziz7m-~5$t$pYKf7y
zQl&iqRQsE<!>;J5E)P-lc##L_h2oPa)q2Up(P2Vm6**YRn}Y7vadYp4W9}UR5eZQl
zi)kqOYtGdCyR+rCyo9+#eflD3949^x;P9c%dWicGfnK8ic{(%ByYrG)?>Yh<HeU&$
z;WUuARy;6{4f9nTxG`P}P2SY8(Z(kp#}l5p`Z%7O9_V4S%tv9|(p!dY%>QIe`;=B=
z*mG~+4o!N_HP@Y6A9#^ijv5>s{9PqqBGZ-MY2(@9JROb7zq(P_zq-+S?j^qSi$B6h
z=Ao(LDzaEvmN_GC8(Q-x72fUrLrx6<!;o3AT<6d8XfU_Dbb9){WZHnn_A&~CANJ=D
z)V<fT-@6##xSM8Y65r4VWfe?bkG$xDTO|<cWU437Q)QLN#wLov-)|*lqSG~(i0<sr
zM!YZTnJ6q_xbU=1*wo3?SoBxP)VS;7;52~UqEO2{^2)x}0!>AQmapdXlE4Utln3?@
zrzb)kz__dPWtH1)C%g2H=gN*E=5FD3KaAp*jEG|4S+C7nsjFpT)A$3VAE(~;M$F3{
zM1?1TpHbnLkE;t(1spv7xevlT3Rg^gJ1`N4?Ksnnr~H(#T_P2lv^>4YVp+GD_SWv+
z4)dN{Iy5?v`U_xYwR?1<sy_Z1h{N;o`^kQ@gy7_@<F<D*x$>!8G(5sHDaJmh4&RbD
z@3ScvXnvyVlePVwT<+<Ac2EF-Utm6F!P1Jzp7r<j$>Ge7G$;>$H}fm_)|`1Q*x>Ew
z1H)nWNKK{zHQ9#7ri%j-_wm(;ZQIxehzJ52F~Snu*lt3JNWe`51k<%y%)5Pbv5v_Y
z+KQp-4<$?fI~RH}E#W0)Pw$Ii_-Gc9ophYp%uj$tE;}TcqCkfuFm8{GzSD#Av3~(h
zoqy+RKM!Tj3bOn>u}~k@OccqG6T$}_YclvhWQU-eN@-nI8*55U6cEe&ukiR=4f?m=
zuF&`QZ2(;Dj0-R1Aq&<vZmA-)f`*-(WusY!Jn#Kz087abHa;bcepVJSJnYmZBG|hu
zyEIwIe660aIpeCG)X=z>xxnif6Tg|b1fk^2=CA{9oeY-XG*G|E$kL2kVtt+osrtE+
z5~73UpQv0AbrhBt1I?dU-1r|I$gi0UR0_>+zt|v;ft$T<Y<b~8ABZ{p0d~e@6bzko
zeldy$IH$QvR(EVSQGK7IDqv>EQa(v8C{!_5u@^pI>6kT7W_bw~N@gd3rHn7#R8wTD
z(do2zK<rk26K7(B<*F6bQdMWr06@OdILF3#xx()pw8>CY^J-z{DT28H>PQJEN>EQ%
zQ-6GnKEW))Ub%gc)=EPa0n1`fvz0>`&a|`X7;xdMM~*j^5hOI}|N5*EQld4HicFk9
zvGWAyqsd7x-to-C-eqT0EP4sj!pz02A5LA^kQT2y?{Gb0@*!VGz^1791o%>dTRc=-
z5jPUSL{k`wIgtE#Do15Aen6_>ul=QhzA7%c>xN~Cj=`@D<{^_4>$Y$WJ6kh_X6NuX
z?-9LDWeC67UZRqDA{!OkZ>Q8(n=p1@GgCG)?fjIoDxWD@ByT8SLh~_2zL1)P&S&lT
z*jrWwRh^h>OH&<EtWXuBRv3As{8}GWCfhh$x!6GKg@~;dj<j!RuB(%asn{-zWqw=M
zL2VHbZsuv$k6m0Vdm~j?9G$^ge>dzbcl@;e(^L6U6=dm_J}D0Bm68FQjxF#c+L!sQ
zB;OX{6N1rA3QTJKFTG#9LpAX~gIBa9Gx9vD@@vKE#>6vGes63KrubOF#B^r5ZhACY
zl{oTaDGj$LZDwpP8h6b>eUMIlD~sW&EVjl)Ol_Nmbu76~3LysB?l5b)>?}FDysh9#
z<?zMH3`&Iznf{!EORY+uKo0qnC!aEou75ydO_s$e-FJpEkv%Nd#c!A;$hZbC`9BGs
zj$w@_5m!~Je4fNItNrf!Oc}2SH>7YEOt~W3=I8Rj?W?@?k5uvywojoP`>r2;0$4MC
zC}L<VIiV2O1N2IV-Bk1RPBTG$HW9jGM9z(?gsbrMB@eJJpFao5z<C`a0YpSm7$)G+
zP?#z@p)qfDR@cod48EC8|EyL%y@Otu`w(Xu+M^5L^kzLwma&K(L;;@-)|R&f5U>TP
zoyWIaoPUw2d0Iaz%H%P(Q$M7(RQA*d)CU{Cr8jODe=K-?2=|y<w5xSLB{2J)c=FpQ
zIb(Efb@knHi{rfhSncQD^M^A?)-dB8CqL3XP|F}|dHwI{21=|8S|kjA4Ag*k;{#u6
z&SC{q4pFyDzT;@>oyDQ5xPV?^V++6X{_qj#GFsKMQRLCieyvZ;{t9Girq^=T^!ZOP
zfT#g@Y}*H38Iq(o8eh5}USsP@h$`1`7%|V5Wj}bH{SkSD2gRa8dB)$k(A2yUVyhJ`
zWp_zB^ahmA$(wW0Qwbhjh$EkcwCSS4qNjA5w8#(Br9M<i=j^-`YR~l%zrBYsJL!CS
z%M#S<lYhy2+3d;^6DygD6arlVlxAgQVbKWA?3Ho!6WAZ|kGkbE^)VwouB~*;4SIp$
ziWLKuCVj<vMok-Y6EWCrJMuc-&ER0glz^cV)!#<44AFRt;K1^UGf{d&ZP}EOA?t<~
zv~o9yo3waIVB}LoDHWYGwyI1uS=xdpn7W}T=p}P$<CC$ihP5t@+|7HOx7P6EH>Ei|
zz9qLH$dB(G9%j4P5uV!R?hMkNP4&&34CE*j=mYh6sdfGl?6b}iG9kc6%|lIkt@r{|
zMJYUUGjvRy>szra20q7?3ySHoe{pXNgbCP<VyMr1L-5KH74<FDHej1@l;cNNI(@Me
z5_{WUs}ngz<-F#Q*h;~kKb1E96#ug;v_gM;S!9x~{IDOSBGQ3-gsFk(anD(0^riUq
z!!~Y1Y$gx;LxkOey#*g0BUO?Lp9_Sx=_80Lv7HRR2}0LztR}cdck^^kGZPmP)7=<Q
z0*sr<+!P)2AbG_{T6rbF^-zOZ19BRlett$3a%GlZd^)86kzd(gYEZ7rsB3*?Z954>
zDqut{RXV@m90H^JnRo7)Rk$tC!q#hy(i+fjdPT|miV%?=3Y+@>BT(D_`Sq@eaXvM1
zEowO84$|&gO%!kkks#aSFBHXKw%==I;D47#n=<5-^KT=;DK#6`w}zc)<nag3FT^N~
zyM2e(0_5M^S40oNrqlC*h=t>)T20!JJ*y%Ki9t;|Esu(HD3{iH^3$#-mcTibqxVzJ
z>e7CP^`ox(UcJup=(J%1t)@W@gJ&spzEfNK%inAHDfpg$&b|NGHWLIw-H&OyHefOb
zTD_8kR0mDfwZ&zKq&&K$NkZJ-34TEZocn`IO}5ws1Oo)FH^;gf<r)XS4T<suWZpl2
z!A~5Uh5C<UwD`YHeuA~F5&JY!WM7ma@>fhL8#5y;dw=<;ddC(blSq~7ueD*7u@=Z#
zls}||u_cSqWrB+b`N$i*MG%J!3CoS=?~B+#l;QHmXtQ*?ZNc~JunE!QCm%l2yqWfu
z`&LTeiA2t{)^C@0Sk}z{GJ~Zr@}B97Kha3s^TUvOvev)^2@?(Qg`3w!?B$Xp-~jkJ
zVTJNe5yrM4eR79GPDe<<tH<7IC{qIFJG#Sg{=XTNK5dN;IN?KyRK^(m8ZOVuM$DT>
z(ADZee^SC%F$3V=KgyA=Q*;i~t@F*F84^DqT=3MpEj{Auncvs+blQDhVKmpqtcO_9
zlZ_!Z#<4{df@s$DtFB8=LXP7TI9WJ}SJ(ThwM7W4;1=LgXIm%ut(y+CQ96ymP^Wx`
z>7nNRMgUUX@B^V^vbB!%x0X~E>~HchX)Cc!#}S7~YIBEm75L3hyn8Hjih6^1L)j$!
z`D5X{(af}Gc|#&5LWfxzyE7wB5sQ%hQme|mMsjTogU_z{&dae@S=2t?zqEk~r~`8r
z2&*Wg+z9sKv@s3)>nUDJ=zsh3GQB;^P+TbK25~;p8%-u(q>-Z*{kJdt%hz17soal!
zyvR>w%`#r3ERVTs{*hm?pDT>jSnU=>V}0EI?WfU5Ty-<B=V%8J?Rfb5F5pxJ>mc*m
zk+zVRm!m6VE6>H8KXYO=%A;;0>}4j0a&>ZPY?a>A_hyf_v*~C5{LC(Y<s&Ag!GkW<
zy9$vqm_Zah0qJ4|MyTNSQ?8#soECdt)AaY<M=cG`z=Ha2wN19?)6@c%^X>V+>`fvT
ze;&4$*0VkFF!c8rypfWsH9f(wt9vPKh-~X-NQ$h|t%LtFOKmF8mT7OP&fO$x^))CT
z1es`y^SatoWDh=NUwMe@S>cEIjJKU}Ugex&Uv|{Ka~LHU$27H%Ba9xndbTNQoytcs
zf~mPiC-5$ITAg1+s=Tsq``aLL9LjD(JQR5%tT6#oS>k51UvAUL(!xG0GAsJQsAVC5
zJkb5N4!$>*gc;=X_9I6djR3{kB_{kX>`B-mQMrrKWk#U}l5E!c#Oic6#-KT@nm;ee
z;cyEE%h32fw;NVSuFzoZhrXl=kh7zZG2;)Dj!4`re4s%x#0`_ygON&_t3hkN2BN3w
z1PJ72-cz6n1QMg|{IUUuUXdix6SICgd39Tbw}xzH)+)C0$tF>=-@)%Z9DezdiJ-D)
zyz1wo0n4($C!nh_WMh+P5^yf8;7@)DB?^comUucAYsM@Y&--K<<hF@}n*Ot2XFZcU
z((h!Y4HpHIybov~#F&Wr;*veL|98{Iz5RF7mb}Nh7sfM}NyeW)oU12gOiiGF3CCkX
zH^m0^0ySv;7<3tGYU`w=+3-nf#!l;n*VsWj!K{Yx>kPzsW-kb6%h7@V)-~48piw9M
z43`z(VFgLHj(!mzwxQ|ga1p}ap+6`DuO%dH78d>8Y5w`N8(Dvyd{H#~L2X#D_3{1F
z#Y0%9!cG@KkUqOhp6J3bT-vO4z>lPazjU9JxFkr~biTuI%K$7`2{phZ<}z29d%VBZ
z>au*vpiUk(%=v*+)**^E0QVK5@&N@!i|iAXK4L1f%<V+K5wj#vx$aNVYRMOwaU=oF
z(J8uAtFPR#fM?6gTDmMJ@*1=*(-Tnn;g~Xu-eN~lzF?ZyLybciawXiXw9JTM%hmDL
z#;qy@op^sj555FzJ>x(0(CAB+b|YKIo?X(yrF$<yTowbcpflmsgMZ0En=12+;03D0
z?}4>+V>BcsHO7Uu)NZ3``)msWmUOJgUGH<bQtz0ha7vrgRV%#^kAB5=GOvBXzrN%g
zFqS^i{3(QHfwYR8^Cv|A4lB12gLBt-T}dCB0-t5KSgo1uBEL->s1k-1EvT(88QT8n
zKVx~JDg7?yGJWReh8as106xTL`^t@{_dQwx3-s>%mFxTjehVHqH>#3bcC>XAqfdlC
zUL=!Y%EUvWVLPSA-CwUfZauoIn16&P!0QGmMsq+|S2u5~Sya}J?22D?k92#UxqY4>
z@AmSPnOX*Hv{C}gud@{rzWdssQrNh<yrk3K#|;w;LWkY&+A6>v-vAmn6HDp1_3Vv5
z^TZjP7Bh0~4>#9LS}eL@KKULAbdKa2Wk$qrv-t4jKHbMdcD%l0>MBNNQO0kY6cp5b
zy1^k^@<`uI!{H>`ynqeZSN>)4IV<tRR=9MQQ9Uw=$X)kvSnGa-5KzjDlhLb=kSOYN
zOJ5}R<Gw3VCcC_f%+nTDdeZ0I{Rwldvsx9tFoIqWa&uQKgL2jm@0$N1+a)Cck1+Z$
z=?X#YS8S?{_96V|bis}ki+D}q>q>qxo-OSy#qTw9;1A44I)K>)h<c?%-}nT*#pMeh
zaYEKcpNRsLG<q{QM+oV-y9jXMd-|#D!XfqH{KUDp_4SF}CAGHIACBwG9D=HI4T%@Q
zQ+Q_BAvu|Q)SN4vE#`Kz2bo2S?P`~aWW9hg>ig=DJ9{1blnrTu4zI$}d&b(fQ77yO
zF_$dv&veEO8Tt(xl)VYv;P!i;f%R6eW7>v1+2Cu#BQ6O8wPEW#F@K8}16&9pDwBL`
zu}cx#sNQ7Sa^dmDeyls+lk+BtvYCP1$x>}9y9Zm+D{aG)c965GO`K>Q7+&OD$&gzb
zw*mpz`XY`++$Yw@wp!+ojTA%_*TPFBsWt5DGRYo6Vqi{AGKLn_A2{77(k+reh4C*4
z;G)M3?!wdR;58>EZJMAcz}*hO`wHa1-b)-lC{)R+OzVZh>NFbk9o7vwczDRG$o*|a
zZD7;C*8s_c;T=_f*Wy<Tg$vbu!MR#($;q?QFXhG3F9tV;yIWyrA~&47BxO2>Am6Dw
zcmF{zyBt<5eBy{xHf_8*^;E1`%d<<-SX01~^+7)hP8Rjf)qP_tH>&bY!e(|yPJBzc
z1!s13gJxYQo{(aj^26hLBckpUEJ0IXwprgk9Ej}*svWk)B{%T=HGmxvo069jzn_@I
zHcXt9-SM%)G!XW3`%l%!(MQ_d%#97O_}Hw@vWT*F$12kb9m~gsUpFm(ee$@qD6ahe
z-wo3Vatip!zGjGh>LyUmC%)J};8&AX{|O`!`^LDCLA8<E+wI^-&I}Ibq(^UK;=9`;
zG754Egsy*ZK=dzkHpV2zj+U7U2$R9RZ1ScWrPy|tU+)NwZJpcpy4_pQr<goJ3>y1N
z>>AUjpDm{*Bc2o;-VDBIUw+j-?Juli6LYDAkZh=Xz9{x#w5Ndbki_f)LmlpVd#TtZ
z1qWQYGaD!K(FuV;Aty_;*&cvpA<=+y58`!A8;VKPzxPliOX;bQ^~{#kcC4`8b7K=8
zcAmkF97lnsDYwMm-1Bu4ZNw?@nrCaOBVp;zh3zjnZR}0YLyJU&NQu$vWCP?4r{zc?
zc^KMNz}-E*ZCce}>ibO*9{mT-bar#*OG4&f6Kwwea7$=f^;J;9k=o{?JR2;n%7=IR
zu<AE6q#^}@e<cz;apU=qUeAw&;Q7a|H}?5jGl4COf=brr{351=5x;Szy)_>UxDh`=
zGh1W}N!6YwC2q>z$>9>STA9SlhLSd#s1%sB#^jRfmdV*fT+=10nOsrmxypv1Yxi}b
z3M%hMKhj*<uB-rsa-Fu{m~3LT^Fr+&;A{6;O;lm<i-e^`KNk<4z<1a`QrlU~U9r3(
z?uSx3!uur34Z*pfOQO#@fD%5&#-=b)D$?m8?aa~@N3JHzk}si<u0;;lU4z{DEU}Zn
zPCVY*VX1~8^RImYp@3Z8Cpx$4F?GHuGZ7N?W(QqvC;P7^yHS*q(yQg4z~9>&VK`J&
zjOW^w75{cFE~T_O*`evrYSByn=^aNBW$G@)B)-k@1EE?*lm2jBnYnkL)nSH{ghl==
zMa|K`4h|}Gu4KtArATm~=PgSJBuR1X^m?Q%Z>$Ww`LW!<pTu`SWavtVFXP2JwZtPd
z3x<<le@>W|IQjy166)IIk9|poJ8aYTddI|E;iHC@m`B6{!3HaRp{gxM>B~dG`C0kb
zR6K1gqWXROMSq&!4Z3O6$r(dr`4X}@&hj9HXw^lwDHha#)S!l4Ea4C7Zd<c-A1~u>
z<@@F-qCt+>(wUCL0fQe;;LD}0PN~tFB}1DZyJsw#;|_i|lmV|6=IW$ME^^{U;u5(l
zYqQ3yuYM-K5WW`yHgRXp?2r>j0!Q>O_mguT-i`_xZ7fwC^`wES;uUSI<x}I<iO}WQ
zmY{7b8J)e^N_6>)29GJD>_x_SCx$)QG1Y3N%>P`0g?svXCy0#$<jMu+<cT78&p)5J
z>j_6ZNKK<!onG*vFq|bd`>lMkHUx^->fX2+Yn3y_rQLwm{QP#~u6qJyjDq*#Dj_fK
z(f3h#VlW&96RTa?tENHUn<!N?24>`OG2n%M?SflV32{3K+mhNp)2u$}ZSrevR@VC+
z6n^f6FOdO&g_W*-yIhZAU)NQ+!!wBl4?T2t$AH1BS3u4@n$?B{7+oza6^GSm*t?hC
zNk#r4Gz{U7LPhl9LGi$%jC@Od`S}BPc>W1z+V5MY*2^*{Y8DoBn6v2UdaciDCLWc^
z#eydK4q5}WZ#hjITXOf;8s2c4gQNm9g+!&mpB4i(7a;nj1L845PRL<DyZeIIpc*#_
zGUAu(2UkwVwdZRvqBvEnxkNk7vEMbukk4cFZlJocl@;_}X-x6$y_ke=esr2mV(=LM
zonK<<Vy=`0IP$AE#51I;2JP>ud>~m+KfT{R4s4Jb*{hqWBorOe4tZfy0`IC1_&(-t
zB&NP`iDwZ&!B$r2KMIu#{FY>@cnfT+!&SVnY6_DZZB~8aReg8m1|PY%y%DH&uBS@)
zfOSo5*@n9Jr4p(};-(gc#Kr6Wrpyo2(QnB~fruM?k1NMe_n&(=o`Hatryw;*zd#{}
z6ZARs<_mntEtvT&b_57NQ{xj*l^jW4Q3A<mI;iTJ8^=|P_WXI0=ggkQqvPRRX{lLX
z@E$~ptE&TBl+Csg9J*Ij6o6%3s|*Ilmi?vLlIyn^skgZ|q3fqIzVM9}mM&KTYj8EO
zD<;>b>O2M;zUf`(LJ;^>?;Fo_(c2v@ajKDRD)-5$A@0w&kaJi4Lje~SR|wO~d^wI-
zy#CT;+qZ)xLpG_J=3aNX5s$m<zn*6_9I}R#+G1up#SdoC225vzc?COeYV|r&O7)O-
z0u*<>t|MqN^$+<#mDZ7ThjcBPb=n{wYxUFi#<N7(@6$XUCS7R|B~Pq6+oR|zMt_<-
zoX=5P`5%Unm<ht!rOZZtLQgp*3X{?0yu@#`?Tl2ZoYiw>%w%S@G>I3vwDS$Nh@@h1
zQ-g7NK*9-^R4FQ{yngV+KOE8D+dbK+5OVq%dgha36<JDu-d|hp8kIQq&q`{$<x&=o
zx;qRC6tffIWB1>yDY;+lBMG7;Hed>}!3Xc6B!^2hnlolj@J_aW$MF;Us`+Zg*F^l4
zH)|yq6`27h3|W30E~i<{2D26={w~Zeqi<r0(T}uOoOrKRPVG{<Wsu)6{y84m%18>$
zQZQ^EU9k1=y#vsxLN99~`;&;xw_B8(l*D!)o3Pi$8=aUL7P=O&o0`%J(^293Qg(!d
zSWn?F-*2WH?9!La1U(vWie{9knI|8w7(V#yMZRq8EEzW@#FBkMarRyE+<?tD;a9$%
z^=+K1MDFpPJj~gbC*ajNMgROyWmTt~y62S85VnDtIyMYQ-KuKG2Ubm?>sq$`bDGw1
zysr1Ze8uL_yj;@XT<voh<=x6NKSf_(dtd_(mi<*%8}D6N98;|Wi!Tt2g=s;S3LFG%
zbrLz7|27aI-zZN~nj5q=RX<;UZvWXLlV)szEe;V@*^m5REdbEbcHX%Z0r&e#Eibu~
zTv6>00dI=sBDc~YmkORRHjN4J_osgUee;gE-GT;m`UWOgk;+<|VYdQn7b`~ZyR(HZ
z_%khON|$IE@)HLpiL2!Z_YaeZg;k30R9S=Rg@B0Q+MSBiAWj2jX|Td)hp`l?88sgs
zm&EB_G-dgTO`}{7ErS#L4HNPfBw%oL8MfIX_{>rXRa3V##tDJUvfT8E(M*kVg?l2S
z;N&z35SN3^PIj2@{6jd}hn+q`Sq7^iFU?!~?vJ>{5~j464MHG6Ee!$@(R+VdP?p^6
zkkz?Bt!Vr!t?q9~;X%Y00<OD#X$v~i(zet2_^;kle3uT5ih)NKU8P1eYnAI^`=MM9
zuu1uUzRlFv3BMKb|4Y<Re7`jpqcRohp0iY7+Hx~wpCa=&U<Nv|@RRDrb6dU*^**|}
zv^^S}d`%;!CVnrRn-^=0U<d`S@}uvq02Fypz;C02)f1mwn4nALM5P8_IMj=7cfw#T
z?LQr+M-`wbb{f|6n+^)ZCcN1z&!8cH?b-8VPDMzx+N$Ee0Ym?M+aF;cj|sJ;|BkK)
z24Y}iFTocvG88#0HZBB@n@DuW+5?-kRz$-Bw5WOy4uSJp&E<JnH}>7mNm9$G1E@y_
zzsUrgP3U19PL1td^(E$yf*g}$LWj-G$O*NJMhbzZ#(e_kO5)SU1wk;oOYRnW1DS=*
z1<p0)Ya!@aUw@~>-3KWqy=r=@l78RV?(L&EO(UvA=^*mA@szky`B-67n^UNls5lY9
zLE0D=k<yp^o!%RaNOZ5tX_^B)0j@$9-r#kZHbrf~qQAicRYyefigYk7Bae;d=xbsw
zGldz<VM7@~CGl4cL!T+D>EgH)pR_g}{|=~Ek~Uy~KWT6;1@(SFk>*Z*kUZrp8CEKt
zDM-{46ChU77gqH>GioKJ<IBODrvoJ%p=;ySGJL=2Xm33$AR$(2M<i#v-eMStxZwK`
z9IL87Ob;3-|ELOWk_eH;{~G;QD}B$EPY^`nLvYsJX5SNr9Vy-SF_Y?4KU*($`fM4)
z@y*XrcbkhMAiKWVcV)GzN1Q@DHu(mWL)aJg;&>|?tBYIH9=74XVU^j(RY~$7!V`!}
zhQv@R7b#QT!{fMX${oYHQ1r>yt=%ZPOzovDkwf+5*VxMb?o$5IeOJp9i!~9G3-@tz
zp6*A);|@YhiJROxd|~qx16?!^Gd+_2vGUu^7GarISC*t^Jj3PndAF^e+3S9}2`xon
zeUKT>s|-#{6X&w`GK>WFD?d3%t{q<Ba^dp$0+~h{%y8h<#9KRW!wF;o;_?2$keq-4
z@AXz3@0a2}fnekH%g&d}63zEyZ>{SmmgC>H^2vGSF^l<fD^w*vY5MBqpYtJfZ_jCq
zJa1o+2bSc02(>k;;zk+}q-rV3xZPBdheEeQpe{E3M-Smvt)u8YPcJyh?(D%O_v!cg
zZ_>?%lM6jj1%0taaEF|B0RE`1?rYf3nR>p?dxRmF+rZpk{<PbG*_hSy#ts8IkP;t~
zQp9P8cp<+vQst&NN<8_>g7$AnPI8{Mm{4wykza7yy8kUj$1m9u5t26n6C=P$b#DGQ
zn_<N?9G%bdgHXEM(<$^WS#Y8v?*uVpbOxVgYNe>`nAb$_dfBiw;8P>yWJ1xg7e^S^
z_1RDQy)rA)W|R=x_0azB&ySX|C~+wHqIvmH$=g9E1(+HT^%IXbImS^WPH=G4p&GPc
zlg_;XyGs2weosMVXFi&aU1QYMlT=&6SOYtT_Gk<#C@5;2M~B}ATa$B1ne&6nc`VGu
zo1l5VkV4w|Mk&*(2b!$3F>Qc!a#VBn_^+_AbFpTErnBVa{`*+rd7P8y%}&Oxp01Dn
z7uORc#OyZ(#n^<+C!+76DE<1A_c1yRUvieN8}(w!iNwsA^(DVej2-2tBjzPk^OZPC
z?F967bkeDKZ{D1)x3QIXPaSEHyDjXmtF}kS4BYyY-GHU8Ph|$w^@i9aCjuNOAT<UM
zKm05YhKZf;UizxXgeQUI!|>&_xc|wHhX1k65~Ii{27w0ne#loiz0ZoTcha@JPA8?_
zykpfozv(G>C-R-ENBHiJjeF(rkW=PJ;j3uOUspeoC{2b&DCvz<TR06S3*WDZy6V?F
zfj`OGJq1-@g3P>7v*L~`4|H%TnItsuzV39DueR}K!W=Q83^G3oB-$j;D5)KQII|l?
z+AmR~9@s1(jOq>O*(HJ%4808Z>bkOO{ttr)3LbOrgKY3I<xkl<rD`qe7CF*6z$r73
zdZ6zF&pWcA%uEkR#F;K^1|NDhvMJhHXGO^}f&(9#;59>zImRVwIldEeb;|n2id~8S
zA|j?es-o(RvhJ<ih!-=qT5QSX)kdI{fqmSMGs+}`5&hGECvSsqPI<+oiFbDL^A{8z
za}g;N5G)?_L$of_AX@JExz&cM7OQ&o697n!L{_^`VvIS^%NsA6)>q!1Jd|YY(^51o
z^>vK;#-dXCg)&C{@QCnWV<$TQ0^5<QcFLAYz5j>w>%dOS>sKjX=2{Y#NK;frn|3t{
zZgHwL<asThvKiJMN<8vvJr0btsx|e0xJK1BSK0p#hiHU0s+37MW$YCIy%AR|-HdQk
z7o1w(Xl{olksvEkZ}lkoqEp>Q6lDj|DYUX`E?%0u0>@anD+5C5`+<;rSXQS4uTiH@
z|0vn4JD$aLy7@kB$xs*5>M=H4rB#`ravn=ck9vg;-*`cLbUa|LawSE<w*wio$F<W5
z^Gm7LtEbcP%#_VU0SJBTufaE{h3x(SGV|~I;B3Mtg72xTD=+*_vnQ*Y{Sx=+05!m9
zL>C3L(vUdzflcnobi|2E1D>3SoaifDC!>yjra_UxRo|{5R}Zzk!4xcvW@_Q(+58>;
zSr>iIC`LZ~*^VgntOAr)lbfkm&Lsl%YZo%^O?Hp3@_|?PARX!yk!+VimKf<&O^A;n
zWtu9z;Arca8*703KyQhJHKcv^s5k#J6Hk3y;ESRd^nPPPSv4|>F*6Qw3Nb!WBCp?9
zY>z&B!oO`97Xj!&ni`_d?;-R@SX{58=2wyfMgI~;wMuq_hrb`-wAb-8t`;dBltSU?
z&}|M$jZjiCuS#He9~A^7`dZ#I5{~n&+B%BXM05lOl;p*U9r(B$g&BtB?tc%1@5hx#
z9Kq81!X69`>v+leJWZKywx51YJ{?bqxotJR;54vBr#0s%7Kp!rA77{|z%pqVyBPMU
z6IZG8&?GW+(L`&Ey9SnakLm&@y1_M#WBNQV6P91&F@C>(^uFAWuqr9`l(&8a8i6q1
z_(8tG_1kmfMt4>th|k#NdYhp#y7Rl-ecfL|{{iZK%ZmVHPjH3CJ3X>*UjSc>gK_j<
zieYzq2OA8S|0z2c$HzJsE&1eBba6^RY$a<Q7EU)49zOXraYF*ID$RI~BHRD{<u`=O
z)iKrG?V9n|C!LqLcxv#*%i4m+?}|p*==DpOshX_SLF%_Lc;9!-TFY6=%NctO(RvG>
zqKHJnMu!|tVbbROWnyAocc_d<GOv{Zl<NyZP*_4wmY;}EvZDd{65dbQVMZsl@3Pd7
ze!FM%TB^Vjnv05!w(z0j5oE+HwJMa*DAO&{h&xlJ5Z%0J8XRpVgc+5VYGB{>mDl3S
zqD>TMzV&B-zo}0*pZbU<cN#XspNj|PYU9l29uG&SgSFg_-Ft2<gqki0IMQ*I%+wT3
zu2u-S=<H60SFAV!Gk0k`cD-2XdJiH?JSNsTuEkzvl>Y8xEuE?cM%5WDc!<ZKr|FG2
zUHV}WC&?e!M7<p8P1KaHH)tHSEKB3ed=xHfZFSwlsHW(Q{nx%!>zIj5VH$W8-wk=A
zLpWK*P4}D5-amn<m&TpdN)Q4W46xjFgR$VwOoms^IV;GR@l2F!lgq*=0GR?`?COX8
zVHIiVBCcT63=NzVykSP(s_^WWE{A=Zc~^?<acAq0?co@@)<y&lxsm_SIG$co_{1)0
zD&YjnfvcqW-#^9RyMr;h>=hRrZA@c)&1Ea47Ac40%J0`j&iWOmELyahfG()%>L{(|
zn5_L~f@eypCu{t@+EB%g3K|V#ii%I{@HUgcorJlXTpL#%^jkJMRB&w8DrGa)h(9t1
zfeJNajWY0TJ*RX@hsws7uDpkOV(&!NdiD5mHe?7eM!U#ZaB*Eos`qDA0ErXna9Ko9
zTrm24haT~ZR@jHrV~a8X=DwU`1oc5Mkg!f8*BMKM{^-beDLWbz=xzVt(78=@KKb%*
za{T7wfIx3#asY0!_bp_2g0C&lTS#MY#|T4#QjxIwuR<r{m?hTOC;XQrq4UY+n~PD_
zR58NGRiE$8h8#iT%GN+?R2CwDywtOz0qc6pk+&WiAgzp+KjoN>>&;zOJ{3S(x^s@c
zN!~l;6tvXKZEF#9gFA7y1NCs(y6%to;aM-!I&a=Q8Zxp>4pQgc?wq$@!T%a^<hSTQ
zrZC1{kI+d^&%9)`IrtGt0&7UP7uYg=Q94s9NPIe|B^DbQ-80KH?|eMzyaItegE_n@
zDvo$&slN2e?QmWqMiA1ovSMI6I~^X53Umf)4d3_j%xTi#U&8A-6o{3s{7p?w>cAoB
zB|ihypm7nA-3LjQPFq8A(w}^;FL`7cYD{J$73!)biY|u|dEVp-dq<Yx`T=Qi=CMEQ
z^j)+4TEw<YY*er$WB>hp&jG!yvDRCPb|Dq7X+R^t@#l^SwbLsr1zf}dz(m0PjQEqa
z*2s7*1gGWg<q{uHr~?>>Q^a_48}H%}x!dCv>9#fAeuwMZ3lDK0{*L92;ph{bN$_6T
zy}DZi@d54tY+1;^y3GBx$hv^cMi>Z-93SH-fA02N_q|%9?E=Qza-Q3K+sk2fif_mj
znBmm8AcaPXDI>*h)@$uuu76OY|7jT=pvL!Xy04!^a~}$>;)98}n+b49k@HeS!y_Y1
z^s}}FQK<Cd*qiMmA0ndA2Wez%eWS2Wlce-pB)<v14d$D`SG>qaM(BQckAHEF!T)T+
z4QVl*p*O|@J9(4;=@oH+LY0l)*GSIno=I#&lAfQ+AJf^TfH*yj!BY`a<GQoRt9(G)
zX3$nIBaY!#?h#zb4wB^v;TE}qscSG7qrB@IjR}*o#5hsG<a%iNCO0?wia5_qV^SP2
zYHUTG_<Zt7&j1Ea6kXn3Q*Ddq+JeD&DVu+X*q$EcKYW<*<g@75GBI`9Wux$oA~<#H
z51kMfw*=Odl(baARXp!49&z{0a=j|)OqrH49YbIe7C+jzB&^0HEKmRQNA|tPR{z~s
zq1Omp$lLvt@7zOh`_ZR8sP0nK*#a9!dZ+KJ0@H%UP6w+R?ESR?sY?rF1RJ2k0Q?&0
zK>j(-{`jkE1C{6<(hp=A)0^={JL9{nox$|=fqnR17YAgqY$8lKv=M=9`z3z!Hf!G7
zFWpM^Nx}AF@FYc;#Ix(dcblS`_I~pzEuC$F*OD|hL=;&)l*xo9nxo)x(d`Ri?4wVx
z{qO33ee<r*Up!<FLdn^PVxZ4^{QfW-`zIq@qdUr9A#|HYq|3K9aZ9OtIm}xxSch^}
z4fOEoH0P5vmN&0xduz_bq`J-ZzGiZg?P4;e362xn+onwa4)}(`R$z%_9T6Vyf67`r
z@_8Jdhz=Rr);pyuj_0iv8C7hcMq7_dm?Gl$;A$rO*)u*=Ui5~_DUeAXZjw^y<!`Yj
zx6Ro0@RV^gV0!^uKBxbiT}_>PA9M7LT(N?P`E==~JD^63MY-mRgU8vY-I36U4ENN{
zV$Qv^br=XTqTo<Ut1yaX$1dSJj>@<T97Obe;5gsE@Y(5wKN6BQn_=(yH>xQ|CklSp
z;S;Xb&c}cEublX&^(hg4=Tmx86T{XQNP7-)BToOV3UL3c3Y3VC5pJOtQoNAJ3oAGd
zuipwaTy8CpUt~1E5PLhad1l;6;gfb^-c`9;$4e$7$Z75Nq)7cVrtlO~)d-7W$!!sK
zj1h6{H4dZtr1p&2MeqAV;VF9)J5M=u_Q<1#!*A@%g+KoP_H1tgiQ)8k+U<;INi;bI
z2_*$PqGbzh*a`0zQ_-!&)}9_V+mu>S1T|U|RYDs!cFd4z=pa$I=-L#Z{b0^K_+k;I
zl3_r2Lrc%XgxiR2uVeO`NW#%7VZO#b{mJ|{3Y`l5uvCIq_<5nJZ1+o7r_+|=#HYsH
z$`1B6y&kouiOHLI7zTA$c(2BmneVR_9zTyxm~ef+3buPK==r>QI;7R)fqW-jI%QMC
zPqsW`T(SJ}k-gTFs}NdNd_}o;ZEJ9?`Ty-BUw9tQ%@N~|XmKw2Fs~E+XT5Kr7NUR4
z(f8=~81TIv%?(_-wk@*b6wUm1N@>1W>k0^JtFVZ0^@el29Xo86o#nYT`=7(72>`XW
z&PXIq^a8(YIBF}ai_|Z)O~unc{qm+N^Wr_Wk7HBZu+4mFPO_Bq&8MA<(JO(mT`ujV
zB;iu8M-uZ#BN1PG+rdhyu!MIiH9jI{@9&7Hyp3L-V)Wf*K^IdqlWhZoa4|`)Pp)!H
znAT45Pm@x4i#{(aw<=%$g?k-TM4+cF{{d-I&~2)6@Oxc9S8r*D@8FNG;DzhY*eu~o
zW`OQCXIft$rm}SpNw^@^J^$jY($@4=4c06WK`vPL#S4>zi<ykvbK5Zn<tlWQ2ite+
zY&1M}sCR`Hw*z-_Plf%0-j!5y%{B!{_I@yDjM|+Dc})`_enjGKW!xfsxjS0SC0g{N
zVYpM@q`=nIz~>g}s`)YK9bL_~tpS<MvVL!qu3KZZIm$=ZY3hEVgw1qL1Ek@RiB1$w
zv!U?rtH6xT8R;+9FV4EEJ|#-g7nQxGum@sCPicT?KC!63;J(E;zWQ2nfhoU(<E~CF
zh11Hogp4tTF<K$ALHGB)C=L$o-E^BXJ{jcx+0~N=UKrb|c&b9iWGfKx&TVPW)L0?A
zCNakkln18&WjPh1LaN3T2*aHx-?&MdS6Hk~?c|9nb9Wo*Ug56TKRdg)s6;GvT58@;
z>&-ySW9o-8X<Q?>5h4i$O@Kn%yXz%+``28wLa81-rO;_^V_#ok)>XgA6X45k^e^*g
z#k-nE?nu&P83S_#yqf0=`zStpTups^fp8`E8DHru@=?wp`Mf<+G}Ykc#!JzkJ!1sw
z9V0#-dMvSWq!Mj_(}xG%?=+ApGrU?l*XSGA`;n$A$~6A2=X<fCv({y!QAncAQ|flr
z0VOjH-=BL@>OM96tskkkaUa9=>tnW_0HLoPafp4+kv2F;p0}2LEOX~7^bzb;yE!RE
z*6yjl;$fcEW2LkBs`H(Hh~Wj&P<s3ixfq!>G3yEFLl)A-X`)ljZ`Dq;5mw$)pVj1-
zxY#bfn9GjCCcOL8V~XmpFV_FB@BA!JRLW#-X7ToXlZsZsS>{{dod6@SCB8nVto4FP
z8+17*=xS6w)})`+jef1%F8)ty@mv!)zhY%x09!7;{C^v)yaZ6PB|mz*$OWPl&Zs)3
zBUu<>z2FO<)YpzWa|b#o@1)%}Mx3lwHshEAouz93esQ?)R1UZxM~ILsJBW0rLYja5
z8Npc{@Q}O_UW}1XxLHf`(P?$g_k;)vAFz`w^2yH%Wm75=IB%jlV5qIp*`uI`n4^(O
za&j0hY@JXbO*Hk0#543N?nek0Vx+!i5xikaJrhQwKwzmX4)y7|Ef=J(L0Cw%7>lUe
zk-mFLGvARix#yaxw>DKFuXn$Dk;3wh<>rVl+OP}xP0MR@5erP0zVq?7ntR3T5MPRK
z5fvV$Cu>}SP^!e)oprr@$0I$^l!EB{{SB{Hr7KHFao06dGWd-68>-K>QIT!kV;lDI
zHEDSYLnaY3W6hfpCx1}FP(TOM(tOp|y5B9P2RFEtY)ol~r;)&-J<8U%u{G(UW7a+s
z9W$ZfFmA3kk7Ix|SrjtaE0TDa*$zY#vs<0gCQOzPX_pwF-NtxGT_CklzLR{Y*6x0_
zt#$lspqH?*db%T%g#i61S}o*iYiv-qwPvMkdLoBZ!vuZ>9m%DJ3<>>)7dnpG=p}FY
zPp*8rC^CKX^lY=BDnt}0xC0*wX1!WO6yeF!q-^`8)L4~T+B!=;kIcDJ|9O|B;Koi9
ziV23*QcE$ByMvx0-P$EbJNqZ_FQ18xCI{sLBCOb}P<M$GwUGYnBCq^0`n#?|(=T_K
zc0ztv*wV~vkpYqvM=M-+gFm$2d~r}A&fqeiS<v=hdt5BSws`ffFR%(xx|{eVQ2gY`
zL_TjWntx<DPV)U$w@5In_S(^LDtV~6CJ-+fb~uwub|&w%_$wi+7FZ2aps?p<{HN6J
z^x_XQ62}zB!ahax#xEA4O;IvMGLkZQOsWOn;9Dpv(Q_nB^7X;;xAK?E5;gdxCN&R^
zyi^Ok>q!5Wap*sa1oyG9gP_bWtp95*7Hc+Cyqce@4z;3fkyCylp%wCJf)foGjf-L?
zx6>;nQnI)DztJo35{#&mC_`zOMgxEndGff$VknhKzkJkIHi;R-C|c;VN4gxVo8boH
z=Z8{`Id@xG^<V@b2f5Oy4ozG&6exj-m5wWdd0rA%d%5YqP}oNc;WUu%BE`bLk_!1=
z6b>9VuX`C~cn7R$1Bn=j<g${fX$=FGFPQd^+d|z~^qIwYy>31qN^{5Y3Q=@;q7a^4
zE1*RWsdO9pspNwxQ=U9>JES4Hq)vzT$W!Q`Buz$LqGDRzb6(*@20OXsTBWxet1U65
z{J@I_1*1dHZFPMg@%rcJ{>Z(|GRA_P2(hjQM6R*em<)V%XCF%xkaJdkt;;IvU%*M0
z(3731C26(5^!JV_zN|@=DskfeqPmtfUjOLv8_L8FOC;yGAt!%a@1BuA5>)|U4dxZ?
z<6rgfkQSAu!Yd`8?r{jiDqy)~@jLsliUIz7)grYIq3FP@z5&Cu&=g#ay|soryVuZu
z=)O*N+g&&|5n_KOKB1`)Ap9uelvxc9+m4d;x3U>`arojhdi~9`y*ea1`;(1Lg%2OT
zDY$dpR@Xq+msyhjLSx>^TVSH&a{1wu4kSB9yz-5qkj7jzGVfbgR_*S-xdo`~^}X_O
z#MT|p!*$>60};FB4|yZiW6j39m%wuEG1VPctyFe$`<0hNN~vBPO>%Xx8>URO^HZ)b
z_V~Lq{zn3fhlD*2Y<4*}`^S--Q6t3%ixDRr3A%)LKHOEsxP$_^;Lr-g>fp@I!t>UH
zAsX}?h%OB|{#}}Ysq~89y>v*nd|iWs0p?7FRqx)i|0PMPy;wfMe4%GLTiAE>y9#@b
z)}P!#x?0_Ne>$||H}@aobCsPhE*3n`kwQnc%lUDn+aD!sd|<Y{GvxS@Axq<=pq_r;
zyK;%hS-04?2SDbsy42vRG%t{u?74eYDjMfdV-Owg)6+Q)M9g-yq_ED@I0j_rrzL^r
zb@?03$)=c#)F4eSt8<F~=4ivHzMhEZ6{(ft@PTWUP7Al2exkL1|Lhh2bUm=!O5q+y
zyz@U@Fw$7(6&yec_?7DWbHKT@{rTUhhe8r0QYDA4R}z-aT@i3vP{8S9xQy_=DaMoI
zQt!7Rr=Mx37NS+tWf+1n$HAkt8sy;*NrSFSQI_ODwClBY*`XK({dK1gt@DR^U--NI
zh7s7kSu9?Hk4@5glh!|n(E@r5d}-xiYa_;}9x->sqT25@fG&)C`VC|DcTa!`+fE&I
z)AyZelC5os>GdfW@Dn3vrG<0-Xa_27Y%b<>vjW`{GiP@{QJI&#ld<{3)4}#ugRjyB
zCb4OCIV0Rv1eDUsj+_n$B<C!5tt|gqYpR#zh@&9P8Es~~{~xmcDy+@0TNi~}3N2dP
zo!~CTCAdp*hvM$gLb2lR?#11qxVsmMdvLcDLeQOl-<<#Y_g*JC$U$=D%KJQHjC<VD
zTow@G?;a3R9?|+ziH=FY7bYvBtxHZNtK{qu5PQqiKS^D=e0e*9$P9giKXBk@POK{F
zXdVc!i6QKhC2A4-z}CAUlQHq>^5@;{_DTot%|TI@RWFPQD>>pP0qW#Wj0Sgv)3<)F
zl^N}6NuPSb{@&MSqq+0OWETV`*i(=sek;>3(_yTSOH%BYvIDr!)p#3WT{StdT<enF
zYD?!6gdQ~#ZU;67&34cLR#hXkPE;OOzMo;0Ju#&87fVwEB1;AKs}}AUVgxCr6*(&O
z&7M}8k(d?buG!w@c0G^c+*i3Gwt79D<3?m5n^mmg1%2CEwAN`u<+LhDpZdW|4m8<0
zG1O^3O(yX7a<d!>I&=~Xo+<ZT`ze21fFa^G@uSp9<V$~rA>uMyJRkBw{otL=JJse-
z<smd|=zD&q@a2@Gnmdg3VAi`+?MXC|K%`1|Q$lVC*%atjpQ*VOAm|j*CApOmz~`+%
z0ZEuE__p;6_Vvljc{{<Qc3g1iU-ntWY!YtVj>BPwXRVP(Dns34wPtBf13Nu&G=I5=
zigH<g_c)Nk6q~FTy^#A$MKs6$@f<z|%9Ob`N{=P?-)?C=B&1Lw!ES#}1ZH5Q>Jhw`
za*i5hBmdDGiTpvG`aL$qx7Bi*`so>;SGG~_@$ka{X&0tv4GyiYuY_ybm6>-s{HsU^
zuIkFj*LgUC#a?V@ZAq=UrXm7X;XW0kDoQU<GIWDZFoF8w<-o<CC8?enlI?I{2ajY)
zbVMF+8_QJ2R!*i?NO2qC8aniOu8VF`CK7964^plE*yuVq7UL$O{0P0exba#?R}(>D
ze7_Z_>jJBaq<r=0M&OYs4#9p0)SH88x|hjU*dU{U^^o_!T!8;j<Nh=2z)BSlMyo3r
zi{}CnCk8FE1$W4gA(tomQ0pHSbBqT&tO_kPhr}=)y9nG};h78gSc%6a$vCUGKHuzw
zt>5aY9Pv!QStSN_Bn9MGfpTy9zaKk|d8NA>l;&j2i6h-i8lo#~Nyts5-M#{N$al_T
zUjB&`*>nIIdmRUD`aOP%2dQb%-VSvj{>{Az3&MA7bWl3}^WH2c%ZB9^0_*TQ7cYkY
z_C8ku%Z(8fPo_pSBJm*4f<lXD!_mPaaBbC0%!(0BC7U`n3q!Bi3YEx392UChnNEwL
zK~%G4(knYU9IJ%;RVcREaq|r2By<esm>D66>cm2RX2k-<F)1uCfN^z^H4e8hC!Wn}
zeB6<7@yk9Hj5OBYEs{jE)!?Ah{>=B7foNIOzEGo#5aXS=T0t#d@$k|M{p@~iz=GxJ
z^Uyq$XGm<vg6;hbLTQ~s#zNA~A=Vmhz}!es(27-sHE!97qD5VQt9?+R_P5Jq^!iml
zR$s`vJ;TV<mtc3vhk&cSXt*{O>~obp_6CMFr3*UWI~hA)C|($PUV-z)B7k}0s{wWN
ztX=yZQ{1;3XaUIZDoRs1M*n9v7)@jl=d}#e#y-f3TpBTjr~H*6H#PPs^dhFxmf1{d
zZ)oe#0dK5$^{Ij>v>C)g4l~5nP#98sLeC+OzI~i-uyF@T)slC(qIZ=(Mh)@@e9>3<
zM9J_~-c2s47Aec<&y&ku;!mv9bpNC1e5~FHEY}PMOXZ77EiGNYP4*PQZ?K81tZlC~
zo1MSj3dQrUZG<8rcbbWk;}h7Wx0zog+>d7c6uB*0eAWa;WOai!a*>JP;%J>NS%;^_
ztXd<!L-6{nAAGfhaKk(z!v!ZK;^}ksw0-yvNDhUTA-#R^kv%^gSl4X4dZ@la;63~*
z%rWS}*Rx4X0mo+4DcTp9xz)od>eePl5&$zOhOIfA!=Z2_0e=Jjcg?|L=V%*S{;srt
za4?uD+y#NV25~e66UyH^^VaX<mD`WJcVj{yV7L|htKbvxYtOhv0t!05h~!5%wNj4h
zJE~E96L9dOxh0pI1_=;1BVAAX1}FW#H5oLFC~fzZ!~e6c%4<YVu@2$IHdNo4pgDX>
zE;v~<ZtIKTZ~4+2YzBd5vc~j;8Ul`yN+&oDnYNw*e2H0?9r?vD#Lro@kPzQ%aiYVn
z>m!<WuNqTZ-9R+oFNX1>9{oNrw!4kJINRq48N+Kyc|%1ShosM&->geR)A{RyB5#NQ
zrJl9eN#BkU=~|lsYMSFuMFed0=G9W8ZSb1a5nb4_e*I4^{YLqhsQwA?(!rEH5nZ1<
ze+ayYeVE}e&Dr+ufY%f)=3GBa*FnvZ*(>0xLI?bcahI++E8;2k1%7zxV!%2=T%(BF
zwAM_j(`1~M+l@W=m?4a#RGBi%U)!=p=2?TOfm&WovhmJWZ)x%9mCbl}yW3Roz(I>~
z)gfaJo%BR|v9d?zQY)Ie=ySJ1#p|A*{IQKuiL4(p`F`+sTm_q^*Q+7HBR&^|2{h^K
z%sKLfah%zR!!VUR4(DlUd*~q*m($&LEWu_mhi0?IZh=`3V(?t>O{?SiN&h!lEAAn#
za^pDTHuO>kyQ>{TzASYD2}T?FVkt2f2VwDPq+n|`ba%&xO0se1@mAn3>mnGOh>3{w
zPA_Z65xm!Jt+k<i+Uc^j%Wkn@RKh1aRSd1w`P50QKCaF<aBNlj>5?1NWC^kPn@3Z|
zbaa|$<plhh7MCacllXidNByBmN-M&&<f$`bxsgWQA(?5EVTbp^TS!Rj5=U;xSAS`^
zGe5_KRdrHPu|B>hw>9v5!c6AMvlKRt+h?Gpv<OR<*u&TgYpqOgP?jDWo-;Et&Ca$X
z*ttQF_978p&)4}jIu{cD*eCv!d$-6#q#5j7Iw6&&oOdHLqyrZcYb#x>-aHESx}7L1
zlgGkvo2W<uOSQ&@q5*Jas)g^fb8<j?d+<z5-)F*jV?`hSZTjbGFQL=?aIj2DPYu4R
zFumP@Lv;IT<ru+ym)qfT9$?<TfV}@v!oxy}zDrQh^1M$<RT9@w#94bFS>h*{{T@8q
zkAY7Zs}Q?*k4L#q7xn8r819xh3WsoL7}rVZCJzDLkD9FNZ;@;$KwQIV?^n|ojnciH
zp38NmGXW2@d=oYk{i-)FMDv&jwDLo2N{_Hu6x{4ittbl=godb}Pen(sJVIDqK8ryI
zrL-7RC$@uD=NIZ!P)RwxLe7YPM<N1I+MKVFq(7^({Yu1B1m|NRUW1>e0=`td6?yU#
zO>H<nlW?YJC{sEeOe;%1D(_kXuTm~tq_w}(_3=@PO4W&vyP9=c1^FxS$DENbTQTFV
z<r|s39mayHQ%W3p^I8F@OT=ExEQ8O+U9YHy!rCBMX}0gtFZnAI_K<=VGZL3q$l<<8
z3r~8V76S-eaEH;vGNmGV;;1idQ<wsJ<$5CIf2>}m4ckJoLh5EK$?c;KmTcpi{Lfu!
zFyiY%MZio*O1fAZwI$F^E`%NrDiBj!m_S0i0C!omT$f&fXMhmVE_e-cncnUgCwr2(
znyVv(4YT&D{XE^OS?zj7yoJKwnAPZ6;My{mje{8VOW^v;i9MJn0$#qsVmf}B5&Fur
z!}VRh?mBCkSXd`R<NPp(L`CWF9VWR9D;}b3>@B0|0s(%BY#CEtdDZC^LY8WUv(bD5
zTOXoJ<TDJLkSAC3`^>Mc1a^qu>mqp#FCz$zC_t?(0n7e@inew7j3A+AbR?Y1VGc-v
z4;MW*KUbZ-*dd;lo^~P$-Lnnk5eYuh-i4+Vt6|>Ah&W+I$(VUQ>nozxC}#yb&GvWI
zEMM`{AK7<8+DFS|-O1+O=TQNc{q%Wl!oK(L$c|b%n7@<y1zzG#!sLSm2>vSt)>M_S
zYVeOg?I&_`I1u+;<o9C|SK2*=0O)xvzf1b3H*wnD3&X_7vnc$*OkNylKoj>dliMzo
zQX=%H@8ccN*+PU{rq>xZ)bkD8qP0ZB_`(<jq<ln;-=;0^@R}II<(_FHkl9yx5-7ze
zLwafm-^QxEFQFdV4H^lfh5xs6`A;qSGXMr%0{nM-SKirI!V<Tf&F*48x$!E-=KmjQ
zU=PK3TyBvGk<kHF^MpZJU{z!-1ZEPeMnAC9=dYFrZ^3s?cn>)1c)rKYrmY=L@m5*s
zk3LiDE{epI{7Bh9Q3l^1r2r<<`Af&ZpCSKX(OX(({Unk$N+;en7|$=c!;tlx?p?Rf
zGcp8sjh#*BGaB*X&JPCPHQHq4c~XB!ZdV9+PqWqek%RbgCJ?iXjzLZqo?XqcaCnr7
z%vWQ)ixJo<OV+LEVt+yuQK99;#|zM8k(zj;=1$z<&XnpXIPAwjzjZ6!(XPR&+1B&(
z9W3RSdgxw-N>en1lcWyz*?h6a!v|%0q2d0%CSa75S;bC8ToXapYdoG*%=6`S$yUt8
zv|+>2(3E=DXcSi+S=__<Yt=Z-q3pTtOh~E0fI;Q+9CTxKA?Wh5D3F$Itzrv07_Ol%
zH>OD8NL#GQstJ#4j~pbw#8aX69_lKma8^-E{=!f6{EFQB`qDO-sw<EvTX5UNTrTi3
zITdEj;YtZmPG!4B;Zepz44qH!-yMbDvBG-lF&*JR%>-ba&AUv#fmufPg(gf65-dj7
zlBhtTFc<}40N-)=#E@>KovcegX#SZKk@FiPG+*Uu?hQ0EFo*loaNMV~J@1d+w`8p9
zi`(-a`N4bdXS$D8TPuuS^4&_h9@SKBE|&{b%hZi@$+X%XrLy{$w^z_DA}dncQ!ar=
zy}D>#wG;!#_*<*H8FBs%c1%W3fc4YKs`Jk-NiJKbc#ks@zdr(hZ9Qoru=hJe$ND6V
ze0bHp725H17Ak@;k)aEbc?!{!(+iP%id^~J6GH0_uarrI-GW&x{<A?s_MZV;dLqVC
zt+%-y&Z80wxYOe?>wNKi?-s=e2mCi=KO#|pY53rVqeYLvpOZ=R@82LD{i#8(SYznv
za+vf!^!_gg?0@?8D*%S)?F(+FmS=?e2Q;)mb<$Gu`ed=m|FbIc{wtO{{v-l57L=iQ
zb8c>lSVRLed)5r}OZMIB8QGOE$5D@iJqbK_*=@0nyyTG+NJZ$HoFrZJ@e--*uCnF*
zPz<3VG+g<Y0#jP`Phg!soGIBc+l~88iWDBH$14V8Vc$j5nd8$lZHveQnfj-|6D`y8
z(-uC1wpW}Y4XK#%arrT8IDRRLfoDV^guvKj)YA6uem_R^wQr$$;S46nyw|*{hS3i+
zmZzQD5>)xs>H~~e{d%VK4B;C6S#|P>v7!wHv@wDKLi1AZt`$8<kJGyyPQGmZu3_gb
zzRAkWa`%6zN+8K<s%C+xUure87dr{2TWZoQsWQwp@{*Z?USqfK7rgBS)GV{NSFzr$
zE~(&PpZH&{5aR~YCLG@!Bx4;lJw`3|_*|BkQPEEczQ%JM2`i~dYtJ-_QOqKBF9mmW
zb1*mSiAej2P+Lt87Vt5+okVgUXV8>MeYD(D^IZkEfShlQg8W%G<eI^>^_@#2tedcg
zn4{JhbYeX2oijC0p~Fvb*~NOc0^mcxqz?q`9YxB2)ribY%iG;Q^n~)~et06;K70fG
za1ZNl1Df1Fc6veHH>RoJUwRn%D@ruBv)ngty}@kte-*K}xBsN~^<<u5zCv^2P4xk-
z3%sAuKef~@#{rWwd^FM9C8-)jKFr=bYyEY+fgE@wtu|NeS6$7qh<~n6jQZD}+t<Cj
zVo)WKT=ajB>7d#9bTgl~;pZA#wHl<4%0kxjkTLJv8qW^xRu#R}hLCg7kRsq~3B(_&
z<Rtw@uPX=|lF%abFa{H;9pv46ZZCu(8nivaOR9CaLy-E7x`#4|<#)15dV#+@-Qft8
zJ)Tum>+gLXG2q05*E7K0_^6kwlAKcSCs9Pg6QXz;x0H^T$aW3&WpN6ylFLh~lv8cf
z*0re6D-u8@HBb{zPK#EC%4C?9?G~53Jw9@CSmM?6_Zfvn+x*4U_QuvfP>17@fI)g_
zv72$+d2abjQAxg`iKX~n<6uvQ|IMoJJUM4nHH;x*^h|jCS&ERA@*>#{TUF7!lV*YJ
z_~masupEq1|8vDix+rl_$MjTY#BZqTCx)UCQ`sjqH<z-26^V8VC*^#f+F!ShKtCk9
z+LVF{Q+YD#5xh9el##?J7<vhj-0EnqMBmC+D=7T_QhpHc=Q%#}K^pO|9s?s!=>1E=
z&-WQmV1+A6OcewRwmxMTuZg|qKnHtAH!{qY)jEzm1RZ<g+bxOm(-RDCG~HYq*piU{
zp)6w5ej}#+bmEp<lyVITTHM({0Ieh!e#cHW{KxNSK%u4DyC?f)oV8O7S~R;tXFXG9
zx)$&5caxdrBO^hF-A3g1UKA9)GZHXWIGCO05Hx%q(CYatlvSVyHRjlo_#A#Xb%k+t
zcQ;%lccRIIjL#Yki%TWQ?OUgc?N#|FkM+a)?L!#*gM-R>Zy57a?yAoEYB(9-bBQ(K
zm_FNeG*0Q$pQzW#YuF<N7Q5Dh2cqA5_0GKWBfA2NdfI8-3Uxl;^Z2~;9eVr&k_($1
z?M~n@qFKqMVkPOD5*vRkI^V{uV=nJh8Q<zm-5oO4_H_etUtb&b=!=H}OeKx?etptz
z+oJp6{~)w;pVE<zS*JpsbcB3`6;gF|U{<7-ADQ5aGwxjRXx#M(zi#=^c${v!qNBZ6
z_wy$0O?mfiRvXz`tf8f?vwkVsaIHlYO?)am%BHkrL22+_V&b4WmRIm$j=*d%qNRD(
z-Jyxv6-{95M6%{`-e*7c?(V;xm=nb>&*+UYDGE8u9~@4#zPv2hSB7{R$(cz|aKn;v
zqE@|1#$MEy$bP=W_bj)=y&Yu<K&8tLgRMeKL~%}OD}LHUIXT=}52&;Qea}=*p2k~M
zXJ#PLO|roX0kXw253c^Rtzx7^rUySi*0F1*!fE;E7O@i6e1&2$izJ@<Pnq)-2j&UA
zCR1}wK@e_gMUmBa7~=1e4+qXiYC(I=;<}c)BhdUyXgT$#3K%y%Mw@PcVOLi<6y@dw
zeXRh2dL;G^Y*;6j@^mtPrhV<Rm^=`XvlZ}Rx}X%szL#sA$0s+uW_y05V>rwzc~sLQ
zBzJlJD4sTnX;T0HmA63zdo2*<@t6=W-sPsyDTmv=aE%0wyo0rOWQhJBuE6nmVNXJP
zhxq}yVG#0-p_|^Joi4}xC1j2I+SJ4rx%*}E@HgfC2?E)jZ5<tEt4%il)8N(pr@^~J
zZ;1tb`HlGL)-h6I37v!3F12z@zUby(O7DO7mH&aE<9*<Lf_yL^=#!PU)s_%mdQb+*
z%p6)mKS|y9z{qPV3z<K$$~y_)!)>01qqgJ_?p`q^fF3ASVv|$kp79{0;6U^r2QN6d
zg}+p(EHoK4;knP`e>x%%YqT^<dz#~93(=UT(nx-tHFyYMFroQtVe!H0CAPvM{OWyn
z$1Mga_IR1VVr{hM?srP<uL1O>)|dUb@Yo1r#cbJGNjvJ}`oI0u4E1s<JWB@e^i9})
ze_IuhYttaU_xfwGuZG>INkr^;4l~Z9zNZew0lYy=jR(<U!~TXh^3&^bq=1b)kvlpM
zKSP31GsV1Hn{Mq(0K2o+o3Gie%*sIDFEY_bn<W2MZmmV1$jb&eNZXRqw~A^Nu8-D!
zu^ynE)-*C#n^w(3GVgM-pOWGdE>&a2!Y`daRXe67pI6)}Mdq}dveX;M%9B7<)PyUs
zCQOvv=&TeYe<;!gNu)hc?bJBeHrgJ;WsSkmo=3BB(d?eDvF9!FP+Mo)!%T`~n1~U(
za`!q9Mt$Aw+YNG&^n!Ib+lOy~=(zoJPeR*=0if&;OmKL+7lnfT-UPRLeP-$31lsya
zyBv(#ToF<jHRr7FVR4wM-9FH!76wNS8H$vDj13LAM3n(dStmq$zP7j1C^<t8Dx~bV
z2sg4lL)n`cpvV3H*8{)_KfNIoma?W5nY8|2fdoU={12HPw3Q#;Ny(5g<O_9^Ol{Y^
zWW;d#Es`|*@E%quhNcXz9KSL$B6i=5_cN<!O=3JqsC+WFh}n$uYRlz`^SiPk3l=CG
zUaE*7jA;KewZEe|7c4%4&*WuI#i)NQXw+F;H_S|77?u`@b~@q*4Xs!B^#z0%@P+>o
zQbVv=%`jk-`AX(lwLa|E<DQJNPC{G;f?SxH2d*ZHhJsyAf4B(im*4%t2DbSsmfMUo
z@?<8U;Cj;j0p38jq$K-2I2oj9FrY#YNX5j>lPZiV#OT=rKUdx+x|UCY&KKXl3c9|n
zQOCBgHbH-tdyWs2ccs-3qY>X74e7IItG3ntu{GE|D#iFMv!ZBxam#B5nPdaMz8=>h
z^2-1YZ!&|LXwx6>4c3I<<Hhc`K4Bg~x{J)Kl|`0sI+w<WG5BXbMJ+yI$O8LcmWk$G
zh@P_Mp({9gH@k+dw$aCp-8)Z$UbkdX`E_ZR%UF{4)QPr*jK7Gk+ek|&!jRh}7<~-0
zH1NNsM?Y!aFZplT{JoQ-hgMRGG)<}8Es+bUb6pU=@p@5R=NFv+o#8*T<!kQpI#awx
zH1nak^Cy4Ac4pu+7rC-wNPHvmT-Bhm?vpIo=x=ib#tgprz;C}Z@lH?t>3Kp3^d%zm
zj^^);%5vp5=WD$P^`#J}7j+WEV!e=(+s;`RKZFo^aKG4vwu7dcI?O;(4}m{s=JjHZ
ze;oo_J^x=Iy9=M?Z}`2bBo*gc$)ouH)UTg7?+>88-~gBx%!iQAeE$mB7umzUePVpf
zd&u|wEqw5=f@QX$p**NRDoTLM?y$*$BFM=BX2jJpa(lXjHlgE)CbLcV&uJM*=-f>r
z)2o9~M{W_9rQ+0Z&NuHannr6FZDHp_{y&}je<7O}Y<Qtm?``PXq5Zk4gNzBxVeU18
z6lgd5|6};u@BJt^@1O@V<EC&X7opMq%m+at(PyEj<x%JWnrf_3Gs$SgtFyRu$<ROI
zRC1`V;WM*Gyk?{TJyihK!klUieKrsu&|yvN4^k#rjWg;+1kA^>kd`K(mJ+-|>Z6W?
z{E$x9%2hS1NS26=dxr9t<~5hU9SDr$v4tzCf9UI$q{K}`i;0VUj#P~btam@f<mq-=
z`<~9bj&wk)+q_!p^n^glHy|sBEc7h=h_crYFN8h7R=>0*BiV0pBTpkXk~GS}?J@nz
zFOZt>0rwGh6AC|2axzNGtpn7eW&CyaCutA1rgzgoe99C}2{%T+_ukTLQs)^#$HLuF
zcp^ANPxr?yS$CLX1f|Iy3I$;Tx>X>cl4g)Gfy6iLt2SUjZsbEqHYTL>$jLENs`6Pb
z$2g>u30x~Wt@a?7cZ&-Z8;^A3Js(*f(jPR-ys_SIHHs#n1Cj$JCG!tqzt!!8@$_=G
zlvL=U30K}`{D!aoKC#>HYF&x%i6I$`hZytji++`8!3$)~o2aV`!9%hw&-$e$I+`)U
zgvyhDe%?!jz2U5AU%sL0%;z45G-%7}0~N*IC;e~o{xZ>*D@25x^0@f(GkNF<sz*Ry
zQb42@qb3gQ&`2zB4iP%S*{Cc`lYQjozA(;&%z2i0^&52tP3+<Fc_>jdIiw4{G0f6h
zlHEyul=;r8wi!A}+=s)B>z9(p`;fgAkk|>+HpjhCA@ML==QF(|JwC}+o*HKz2OOor
zAe;CJdK^;Ab%;P3P;eLY=YxF@njN>;TFH4zpmyS6XMf(4aRddyh>XuPxOLlw5_$i+
zbK8ruoN8iM+{<|XkQzwmFD){uCxTYH*=Y_D_%DFpde2M$8V92MTqKJr<olp8#ZFF6
z9!*~R)DQ{Moo3lVeDeH9j2tnW)cy}$wY$TYW6qy?EiI%R!4fs&t(&^r;K&F-A{6z>
zICcO%E=b#?w$A=9r%)S98OhNfc@6rvmEv0Ty5-bocB=sY`lr!!hC-w3?_XzhtK9Ls
zXE%Pjy>lmDBh_6Fb3LZp>>2TuIqJy|kBODqmrl9uXKB0ldALGqncnP=QRltd0ig)B
zGIr)W+L}aCSw#MHPnoN$wD@pi^Bo+NpWRscEoZpmy8pw&{z<Y#kIi$uu=X7}5&n(m
z0~Q#;>T8bJ3Xw%Zi>y(z&lAZDA=7a67Z=GX-!8$4*S%pH@-)Bcj#tK+yS$A!r4I;!
zvj2xu|N9U#jrOq<@kVM{2&Ow;S%<@=n-ZnvQcK;*`%W-kO6JO$-BU_@qoYUtXm}5+
zF?}|3oC!_h=Iml^ijYzgfAWst>0SIN7eD)Q!INaRa*uWJH`t7+*Oy$NQ$DZjU)O32
zl%|-!wblH9BEvV(KdE8XDL1AU&37nFa5scPYkSzHAII*QN`7-N8H>y#wLMpWH=cGc
zMKY3{Ns9uJll~+}nd`c~DdiJ-x`&I!neBJKiR>NGEn??<SK_=?q23qD5R>qVpg(~`
z<LYLKf8Bn)P1vra7pWg-R41bE^S)4Gpd#YnGAw*18l&y}NQ%eNv&Ax9q>QV<a+-Rl
zC<e8ZNx5_BVqP?KqGBQUXZ#;qb3QKh^XDXio*FmB3pyGG#m9#^=w`5t>9!)!UXb>S
zti!?f^!=Z)w;A+l)X=~C6WXn8Ymr`Aca45Zl`AGWGC){<QHHgU`X;n$Eh12hUv=Xb
z-}jdCumY%Sy^9IkCg;w6xgF&1o&Q-2U$h}fbVwG$g0#}3ALKge=cK3B{eghGpj%ln
zSIx)V9)vKETT?FjU|#MozqB&sm2y??Dv~u6{xZpi*u$xX$C=+AJ6P+!H}XO<6{)P_
zcmG)g*UE=Z{Fs^LAUGiozV%k-d<^?)6PO{<-Y=Pmp1C9PBAAjT60V98A}DcwLxZ`^
z{buOqP<hr1FomMzHIp|Z*OAvL)Ic8Bf+8~he{T{{sCC`A+@fDMP)bVvW^2-AjgQ4V
zSuL<}Im1$n)Lc6_h6f;**PhQ0Xv*Zk7==SN(AOhFA-}&md_d+&)e1$1WM)mc=F>&-
z;tJd^j>`W2@q<MsIdqgA5VM>6@=3>$>GSK6qkML5ZdJ+=;)++YFyY~%+1yrdnbrvw
z{T_H+##HUcn~3~Xr||19*4Ja_993YdYUu8A24LA`+v%%_`HW=*OVl^YthOTgANNc6
zuW^(V&Nq8<^VRy=g$ydWLSmr)5VSr;$9FQ@Knd97_`-Dr?I#IniunIuTfE!I@{j==
zXW6lF6;wYVzIGin2$SG<IeQV_VJXP4C}9yQ=epn$YXt9N$l$Q3YbX(lzQo*ec~7&e
zp*4pZ{)^UHtkV6XwBF-;jkkv9Fi@UTy3MnJ#=VA!jHWk%LIa51g`ryeg=xzjU)iR~
z`Qp-NI3mar5hC$LI|9x>SL<?{Pbczgp39v-6R!D4itB?|_uvyi^sEj0^7MJBsIPPg
z<W!89x;OqPFJVqFCyd``V)UOCb&ZrT(-J@l%ucm=I3k}tAZ{|{^2MInPhVb(4%*_A
zTV+pArTrs@(?iUz_z$?EmiTMj*b8pMC1?7YRw-FINzzEFuUW)lian3BSU5DO@7(X5
zW$z7T&l;S~Hx`44Y*Yg_+Mx=QFI>DAg+o{TrgGEqUVz!A_O59+ekx9)frIz`{xrv^
zga69~xMbj1k7EIFX5WE@SYHH7sy07Tv8(C}l+bmwpl0rm_-%4JAPRgctMJ%BX(6Yf
zpY_GbCGT48t*ixbT1}VTzDVrr(y_jh`|t+Wad4k0!2U~79sDLWM(%Su3qH_F29c*N
zbJB^Fc`H0sLMd@}7E=nS2^yB7u;HHyl3>)i8Mdd{$t_T{eZTFd!5lh`Hu$hXrvsUe
zsWe+z&K*rrf|?YU+<wgN3Sh;?i#laDX2;68F@n)jct=8ERzmHU8=H18Vb>rdF;{dl
zDBPd~+C82O!t%*ic`cf*J%tV>gQ)%HtbCz9GSge8gX8#@!2~Q?6zGAaXkyLq`<ynf
zIdrT8E@yDF)e|GP;#)ko9xu-lMMDApP?HIe{|ev8d(ArAC#M2`U4Bz(d%?1n*V;{g
zWGi%<0JT5eg=XDf1_NVs>d(n}F8P)g(>Z^><Fptp<1p^=ef(Ivu*UA&CAwZza_qzB
zwa=;-mp!o)t1-{>h$_}dbf>o)^ukIBmmg?Py9HDaH5*iytD&J^E>Dq3-RmmxG-^EM
zdr&dXDkCoji+Df3?JB7mOhC!EU;oXj!~y<BHOufyhAcn)-+Li1_rfAVcm~T`n3ab}
zn>Mbmg&t(2ZGTUK<Ic0`ZD_7REr4zlq>_MX-vi4SlYAzMNoHr@yM#-V8oS;Dx(Rs-
z6;=A&laa@Dv|O4_-FWbQzfWJyQUAln?50y7fiJ_?<}MX$z>kJp>)QbtuzSL>{Q7}v
zAcl}O5j5Chi2C;la043TAZzv3rx>W#>A~4&Vi<D{bp*D=0ac*BJ0lyVre<c_UB9ta
zXI1;Ri2k#-!S9;3-#qRS`!X6&snZ}?7E9s4ZMQgh{3lRczzMd%kA^Mti>)rfU2gD+
z0li1Pv*}%KJjA-A6&OD6AD%R5C~77LQoO<uVR5Dk(kybAgt4aS$-@F@r@_M;M2+ad
zO+#rg!P=)qh7J0}1bO0wT%2GkD&YlNbtX;l3rzF2+~P<M=s%VSOd>aIUR#t-19nhY
zZDyC-;sc$U@_1o>yD%~Vvqh_S+NFu!j14oYEQ9maS)byu>IIuxsJ@6}8$rkJ^YB1P
zA{y+-$wa2Uru6H(E^hdEa^2-_PVB7yM+Jk2gBH^)Y{X79^Dg+YNH;*7{_>*mo-g^;
zPSpNgNocWnDDe*=82@W3v<LDR4KQGU=sV~$BWw9glEAM{8{=1+SlkGA_vPl~oUz>5
ze#NuIs3c12+x-)oJ4s7;f^j={b>Fa)fTxz68k2dkhtJs(pQ6DpDe7OCUY6I^+I-xV
zA<=|3w3=ER<=^auoZkn%s3ng1y#Ly7ou)M^SxkWaP>HG3sTNwg9{UZH2Rdms$oZK>
zVD~Oid|`Pg3)ZHaG@Hp#U+j<8?3BJcbo@EtNmD;%MNyu_9VVB`O!Vya+~?M+WE|AU
z3y&ZGJ1c|%tWL@bDKB%a*rO84>|FW$&5dncOBWbsn$#p2>DIz3>2RsZs4~Sm9)Dpt
zR(m%$6(mXIZu~In`2iO@%^~D6)VH9=K!33>5Ocb&iCudI5~iX`z~|N_YY*62KSA>S
zT_>-57tlbm<m7UFwNZzEp^#jx6M;I0vr2vV2l|`cowNthgl&t1xCH$o5L5Xh5g5Ag
z=xjL<yPGEI&3!HiE3hTiTWE2(`no-l#gXi43=E%a$2+=8Yx^A@1QKv$%|7Lz2f}3~
zcCJ)>m0~8^k2FPawTrIQkIO9#JLEKIV=9ZUb$><kU0uY7kWb5()eODHW?R+PX3e7q
z_t5rKVzOCWKb=j<{LO3-r~cJksZAIq-L{?9X`UsJ1%?U8hFOq?8-mrQ_f)7Dm&!lu
z75lqR6WbVBPk6snV?B3k-k3jHcJeX*U+ba(=UkxE+i22H0VqhcO0w~=00((Rdj6>%
z<(d0;-smRN(QV;~_~DImHn@ywBjNS_{fjkCeR~t#hFF{iuu#gRco1r5CmoC!Pf{I(
zYPEd_=FE>kb|qHKbq2jkBDA4h_-uJt6USy_&39I#q>1(#>~Dq0@wa$w<pqfv7?9i?
zOhuuS2eT0?tie1#Iqa9~^2`v8LjYfTVVb|}w}N3=!<g)!{#U|Op!-e`RdUL;ycWYk
zH1F*V<om66Eu?f;^51D??b}@ojDB%4>dCjoK1U`u?)K?Bty_W-yka-nJY{IQ!eti<
zv{B3p5<d~)P%Fcz@Uq<VL>@Nqr7p_H(y%OB3F@`S=A@<Nmk*N!y`#?MDWx#yLg}l{
zFC-hT8bfp%$jz3gnRUM!Lp4p9%thbFljm)Cg>u38>HYaAjDu~AF%>^sw4HPs8dS0s
z67*hXdC(I;IQ*(Gi3c7&P=WB`V!z|cg=Pmey<vBoKd6~;fQ7s=dIRYNyOq;`kxAg?
zh);nH+~fHOccHK`#KmrHjgfuAN4Uz(zHs(GewdK4#IE=K6}D4<cP>6PtK@MDZr7_(
z6n`qW3@bMCrnLGRMJ;T}WicL1f>&}zx?Ogp^CU82tbq+kTqzPc&q&n%io)fx3G>hj
z(Acy5a5hwvl(t}cnOA&Krd=MbEV$|a5d9)zcLkY6VcZ?LwGS$PCqQL394#=i<@1-s
zw@Z8XEauT(Xhon^9UJw06*Z*FHPR3Dl_Z8+Fs{$){9|Vv2F?#6O_r$JTu1tB-n7Ub
z!r4KERuzGH`nVrs@8Shzeo=$UPYL|V$&m12j^{|<c>HVMM)ww7Hq-x|b^Dueru&v;
z<pk*Urpnb^ePt%Cc8fhgHb|@@n0cl<mfmO-nJ~f!je_<L1%ReHHnnCVuB%CPRRunZ
zdq{CuI7=GN$ua{>l3u{8CsheRI6Tl2K*2hs+slY<p@Mzu)Rl`qD9QKYc`tG~K!$(y
zTM4}3eS<7aSBqhme0OmOlNCzpO^EWs<kQbb#h?G?HmFh8?Eh-&4c=Pi2-N<F!i8CS
z=ma!1xq8bC-J(?Os<oOBNC(bI6_c+>IJM6pt$u*eWDSuJuN2ZqyMGgXTbXU|tM;SK
z=a2)yc#*o_@XPnr8~j!v+VBUy6T(8r0rz;H)n|MSIeJxja5q}q;Hu#W#~&yBZojCY
z3^}dN7Yd&`KOZRke9dQh79-a6@8PUiH%7M&i&tdJgfSJXfzu3`5SVyM)+0{ivSuCk
zWttjyta$f3z<BzMN?EnunB8^rJ@eLA$w@D&pjDi)IOYsa%8rn7g4emuCA^unI79Pp
z4SFs{dpn(l)F~OyBt6n{EoH+l!+isDYRN-9Q((welG~;%SX_Ve&(oYwnN9^gk0*Ml
z!z`T53AE(0Ej`<Bw$^yNqUFZOii;aj>m!VNgyMaNI6WX3dHBO_WV1}IJM)UNbEgH%
zg8GgL2Fizx$h1$yxs^_vtdAmvKQPmysq)Xa-=6-AkO1H?8KJ@!&!#<bk7QwWGAho`
z;~r;gxPji^LOS@kA8+oiuE^p{7l+W`fNG&&_dMox{C2*b8JGik;%C~0-!rgxv)I2U
zmZ7$|KC5Yz83g3bJV6KZRF0pZe^1VtCUuyT{Ywu1>$58Ul<_)pg@K5KBo>8tB0@F$
zBaO|N3U*_JT77B>2V5onlWXLm@)^eh;!`o54fi5{V869T`MUU(;E$Q2-sZ;^VKji~
z{wFy?WwB5*`0B+Xt$5GJyM1EiOa7RcM&7{W!m<Q$5{6vN{77C+cq$+)P5U^)!&GNj
zdW343mmrJ>d5fpO(r}t6=k*A+!}vV-MoDiqDvS+-4}h_xg0vq|#_eGnSp(>xs+3zC
zOg*SA#dymgra?WR3T?{!ePPs&l9!9cxqjC7GqJ4`7CXZT$MmbVx7<z~Y4pFQp!sLn
z^}qhlru0|P`DMdnQQ&Pp+MAd5y+0Oc6vb2`n^=LWuDc1N<lO^uD+-=G>TTlMwH8zi
zVSX}_jq=4&4I&z}*Ebs@R1CpJfP#_px->=Kn9RVv;@Ohm=Oppn&KOF9<}ZcIF$b#x
zW&}jqEq}U6k^D+)HwNN${iqnEIm#^C5@7{b!u<?i`foxiUbTra7K%(;L4|`$Gxy#e
z4psX87^U~ul3xGIegLQ<0JYRKJQj+?rNw$%>%r))7H^tf-Eb}W#@!ezT60aaXZ~}E
zj+nG3pa3f&5=T<3A;b@Oh;F6zJ!$i2${Jl+9nv6=*VE{dc67>c#VWxu9MatclDuiO
zZ694Z*bc4E^GAwY%DvpsCiQ*Zlhh*xR!YG_`DbK5qj#UP$HDSMw;NyoQb|vv&t-aJ
z<`udewBNk5BHFC@$WZp2*LC_P>g1?kg!L}-NvjvRfa+ISSv|x$8W+J%67$&bZ30!G
zd)*hdaN#>6v$(Dx+%uvW+`h=yZ-;Gh)pb95Be}NHW|*DEbGRVWS-#$zZGh#>dMc}}
zwPXmhu=C)_2IsDvX|TpFHN0%a?-Er~k`C&kUA1;);~$Vb5v1{GHEM+lCk++il!=t#
zQ*&Lx2-MW%n_Nt_a+1hgxPGy2cGQo#ArisN??1J)a|gTa#dybAT;+Qj9f{yQ2x#IS
z0L@mk6f+9T5-G6R-sl66khD^q{2rbNCw`HK`07+Co?+S~j9N+|<C}h!Gn0GY7V{Lt
zV4pfqg)6yotJ=OLJ7-Yo{Chd4C_P<zS#!CtyQaroNOsiu*6u~64~T;gs>H~vcrymm
z9G#0xtZo^JpwB2H9kw^H2uS#Ay@)J%$+Vg+fzIu_Z{1JL0y9!psn4q7(V!i|PaD7s
z-mSBZc>@3k(EkPQ#fZ!KXdCmF_I$PRhv_8J-*s|;1fd#x!z6XJI#W4jyh&YZ$HxJg
z%rAipo;WOe)JtbGagVXE*&9$e?DKkXsReEwZ}{l<(h3~0*NNmxQ?3_o@hVJ3gk)Eq
z>`lMcP$EW?@MBG+A|X%B6)U8%E;m}mh3bq=Z6ofn{EJm<xd+ek!)>9#eW)Q3rOQ(;
z=yC+C|14E<y7@!=YJA^A5@vAw4kpKRO{Yl&(N3oK&X3RE$~;(pyB?KdzSWJiLeJyq
zS#&B#KPg>TDKhVU*27wIGV1O#NNa<MnM~$ORI==Azn!FMWG{b(b2I$Yt<ik8i8Wvl
zn%}R2JAd+e>t9>%xoVmXn`Nsh6NswCtACd%0h(9A9@Fi4<ZIs*4FstMln~Q@%qemJ
z4lh6O&TIt7Wn85aF%0Ow*N6Ea8_!hLS~*q`8<X#!V*AsPy-Yj}ffxMZo{y}5Uw816
zTQ2@yvii7PA+_qIHnWOmadPs!9Y@R!ImDIrOm7isoM2Yp>I5c$LZfV08KmYe{8<NB
z-9zB&CH#JFI%M_Q8TfxHuY8v$MZDeh1yH#oaKq~v_xdgbnNg*d(F@g=_T}RPe`q+e
zx}hGymoW`DjdICx>=TgR<(U6fxKZ5xS_t{&9Ft|&DyO$cg@#rF>ownG3G!j?MYXKM
z6qUY>RZ59mj=+ZwWgpGZEFO@Zt-`H8kL9sip_bY7LlZtOyguE}0)tGAvHRp{o`>4!
zEB#Kn(7z?9OnT~0;J49FWhaMlHa$rYa?R6&4r@S(pHekL6Je_3>Yru9k-_HArGA!=
zGwezg^8)2u-4!oJJxZY-&e75S?J*}PwJynit#yIqxjlX#a>0vcYnW0)QpdMcJ7q&Z
zSnJ3)B`9O6pW4eCU*_|Ybe%BdYO*vbh#9-QdgPSFdK(<V^&0g<NoXAam_sEQuL1SX
z0;UHy)jJR5nTPqDF`VbXCbs3U;q9X#V#49Znfte8*IbxyaUoH}klZgt+D*}%<R0{i
z129mE4g-(IB`+746uH>PrJXTHR=!)T9d3WZfA<abY+d|VU?Q&a!Vf@-e>~LXE;KsB
z>;HsWw)rfUVcgk`8FwGv<YpMe|EOA``y#xQ)8$H}Wpq%`p-f~GwJH!f9(}SEGT$W#
z{;ms|BNvv3K7pyoUitm7v<`mR*HTEe5>oz|tx!^ER^Gi-Q^O$SL^;($Vv|2woa_7h
z1vp<*SgG~cnSKeu$de4sI+GoD?)S2OX=qa<jNKi&WZMdQbp=il0=rT?->E7L47wQ>
zEeU<%w$7b3!|)SE$GpG41~6)r{U|D;hQ$Q18iaG04e#Z9_*lpOIG}*7r!pJyV5+Vk
z(tzJXEhu+}cz!rGf6CBkpG3q8&-aJ-o#4r6MARq-FvXa-Xl~sYzp}1HX!hAeQkr#*
z>?{Fc&MCamZdQhYDVK~w7`%Him?!KR3WK{u8I3hgh!#SIP*F50Xe-SVX8Dk`u`q1#
z84`)so;N=NN}@$MBE-d|Nv}?I#YJ;btuVg?sI`@$(y4@27O2xGfVX4#;Jji)MHS!F
zP#c(7rvKs|x{;dzuNIMz0;d~LjB>E1lW_H$#Lh9Fgfa?$ST(wBNo93Y5-Mvor)W&R
z2sk6_$7T;)NAVIw$bSz1V_DAi_`Dj;Acp%?fz<)ovv6<0ll@e**wk-D<z9((tmr|x
zz*zayLZ|(ly%NWq3-y;o2UA6Y#SeAc6MTb?M?`(}ynz5NZ9-M4&CNbTf3bHtwgp||
z?HbgX7=#yV<rT;<Lm~6l$(*>c3qE<wQ!Zm_Xc4B(TuaLEmhTQ0EY^CAsm1xQ{;FTD
zGSuUDTgcn_zVH);M?)endf$bu$zQ6}1DGc2w*(pIK;g7RV>*w+vWl5ME=UAtcnmv+
z#cKS_kjx*kQz=XoiO4H$cHipuulPv>@X@J86WS>gCiFg)g~Uz~efgYboKBP-=xQv0
z-SM~wtlIR26O1){Q*2Qz&apnxilo@y5}NxJw*+}<)4(&Ivd}9jQU57NMyKtR?=Gi3
z5IYS)s?9j&@9)MK%20Q^kj&^Z+EXqVaZ8m@dG#+OITAXDX3yvOUS`^OpzZXVr1u(6
zc^svoe!Z`B!)saQ-GCJvyZ463uQWN0HEZ%ULeS(>uDqqMQi_A*k3tcp_Gald+S|EX
zZK%sajEM(?{@4>^)H#*n(ev$2A>7qNrgKf@{2v^TlKo4!oEQjvcDH-tREyVcD*HK5
z-l4&3`YC$CpX{*Jl_-pM?m3lNCvpYUk3u<o9s@eRjF__+FY7WkXmdwU=Be?#)8Ny8
znEETdL}s|)II}Cx!Ag&NZZ7kUfz#j<Y7EhJ+B_d={Kx{jVq;DhGu#q(O}+0g?+&A@
zrn~(SZ#0=N8|@whpRD5pQ=}qs`d}ftsulAs-kU|B&(2_&s)`DBcGksUO346JOdT(G
znoy~a*ry13?+a{)D%uL(45pa3&c$&K6^tOoVIE^JFv5a#!(ah&w`tlc#wImu1_mOm
zMXUBJ)imK4+70e<gs-$a=$QRw?nLWIC9Ea%YMtSHn`>!K91~u6fanCZ6wNwZMvd|x
zfUMZ<-@_8E9@lskM_OV@cS^sz{k=764N2)`<I3h1(~{?MsYUuSo~@jZmuc49-BB(#
zyCsS%Qgrf^{2T%ga<U5DL*_Cv4@+nAoF8`_vX{N!TG%#*r%?LQ<;+zK2ilpts*K!(
zaai?qO1291<09R(n&mL?1`I<Bb*@b_Hqk%U0K*Y-Q3h)q`t5%g*hMKqB*2eMCv5h)
zbVbsP+HE8xmi(e4qB1nfvIrMj?Q#BJUj`P90+CD%J6s$-oz9~ceN9c6%V;83^PLb?
zNM_c?+C!&<WDm9l9TKdt9cyCg%baA5en->$6n{+~LH_$8STAH}|4*7d+1G@xWea6m
zp{Gh)e|V*w`rBnGHPs)t*35z*5pk9nYkpYS+A@-Deh=RjjN2W*W<qvzS}kWljX_%x
znv0^dd9}7*C%_B9L-usY%(6^sXjNT&4Q(T?6H)t2No389@l5;I>}#h#gp3JJ>&t8A
z^YLb*H5>bd+2jEWY1;fK*Kkq->z^!bS2a#nlI|aWyj2=oQ|M(DIjVv}tBC`zW*3$;
z!;PW7^54l4pCg_w4u-P$>UZIHW(Ieav#$o2Zk4=)uAPR<begG<@#xc!`L@R$`D<m<
zS01ATZsQC^LRb8!@O%YWsXu<#l%rwsPqpgM+9{D)^=hb>p=k0uIFuvn;m7Y8V<2Z|
zutrLC6h@^&bU<v9KnaSztUMWHGj7S&=ZyuvzKrZDfOk(f9^>jzV)}q-PNv0M)JHbp
zN0?-5#M#bewQbs2iKNN>1{MJ~laO<ToX1S2SE`s9na8L)ZlWwbC&`IrjDL~lb)1HB
z$U3-Q2DR_G_Z0>5wXPI+y0A1uqO{qJ`v-=ip6g)an^)8G0)a#*y5+f(K@WnqAn$sY
z8+r<(w)kij#Sw2=ZQ{JgQ7<N+4u6Dq+Dx>~(QIkf_=B9B6X=wm^zrMZ%*KNUi2Mx2
zfm96SdAOK!<$P=E`^?b_JTtWaT&3e;xs)^5;p`?0$g<ccehJ){{LE$k%#z(f;8u8G
z0prJ<>DT!fg_^e}C!1&UdwhKSb~!SHlIB7u=2@~ZT1SaW#3PzUmNQ%DB7wuG5ubFz
zSFU`1e*mGIY6a$Wxg#qql@=GpqCKLlF1R;z9$tmE1^sAl=B5GwaNnP9^#ETU6h3S0
zhrkpyZ(#ly-fv;r=XgrE_p$F=zrQk8E_En5rMq*C)uGk)VvfqeDy7iBO8J(yF$Cv>
zEUBZ^Hu9o|c+nj9@04Tt%zFUkEe=a=37OalEdIi3cV{*#x|L_zisWVUiubN&z4D7%
zt!fwhs3k@u4hxl$_w5<(VsLHWmj83k1&(|QOUPTTcfU<UKV;mE6V0xN+c)|xewQAX
zubt<r2*XU05iJ#4O(XFbbCyI+9v08xmi0@T40opchT@cfJ!T#F3LomPHGb>1RL^6|
zg6hX18e~i)s;{CCla37phv^VDaY`s6+EJN$kge<PKNUU4WA?)_JD*Q-SAEn+YcHvw
zAILiC_I<_!3m<S%;iZ742WHJA<4`sEH#<ECH16nc*>I6475$!^_`i8^t`GJ`z{(-<
zp(75`_wgNRLXkNKShnnA@Rn{f`lSd3ErqlLCrNWctipq$IbXk#F7zG*2>&T3DMZCq
zq8%+jz)?U3&8VOfM%P1^-{=r<_h{!(w()z6K7O|f;3QVpjUe8ny~K3R*bV%kLe4Hh
z;_o&%3E~28mXS1V_hY;(qDDL%V>J9;ID84~4T7A(HE#_%z3_t!0nWRdX3=MV$P!`!
zs)<Y^u3P^l6s|~9KsP=thcXX9?Y?20@*;Uuyk?(^^guqcLfa_z6JS-z_9Q%W5&CFg
zWy_0!1PMv*&WY|QFkX7ee(D`V|M&7!9ulx@!GuXfFnt#l8iSVJpub?d&>hhuo3v2o
z&#Qv}#F3s0U|^Ig@q3BpR|VIZzo=cOa@nrS>L_x`^ofTguVR0@B|2jv&0r+GwYxM8
zGgVZ&AE6v5F7X=<=<+vHHgA&sq3$u`2(i=G5mXM0WfeAFL`_6*396MaN(|`1e+B35
z$Bd|iY;lI(lD<{oqHN}7%Z%%VK|FP)L&#Nn&B*9vh+9&ym1&^{2h8JSfMpMH5B1*&
z#?^Xb#19939}w?A>8x|nd!VjXN!{<zpRWlEfmlpX)VO&Md}XHSy~e;na~&s_O<-CV
zG0+GFNU<J&D$nS9eONg~Bq0w^Zef@-Q1h_cXl#ytE<%+VchV3*KSt%k=PEg)VJH&C
z-6t6um3{s5LT)=>8KCa1Tt6Z3AV9C4kD4pu6Pa9uvn;bxs9Y#u`YM^3dB}lo9#&+X
zGB2*wl^KlzK7bNa57UgFF$hn7|5A7UCE<z&XdL)*AG@ow(<m9bECJ7U((<J*(9`Fz
zzU;eHPLeAj=G<TXBd%BBuGdd@RqH`Bc_O=xPlom6V?NftcDxlaTO+Tr1w)aRwq2Ko
z9bxQi?3V9R#1?~oJNW<;{-$Kys`GBjdR?B3Ri^V=<W6=*{3t3c9avN+Tr>>*+Q%Md
zLa)`39tixiDzz-b4x|<*FZZsFg5|h9Q`@P?`;l+?N}Bo<+<$A=Ds>^1)&uw_<X*-W
z3lbbh$%9#^OfKwV)hu9;&t(d>Hiju~|3j%C9fFeh{)t-oZuc8)Wg}IAZ_*^pMyR!U
zJz7^>nH>>9NTcNYyU@`^K=d2=VwbgHMO>9cT20P8G&i1bbG}GoBp1%E+Z0EM;Kmoc
zXF?J3P5<8EqZmbwgdN#}1G+9ha|P|B8tct+{>HRXXPg$S#O?X?^0K9--2RuRmvisI
z_amm4jJh|B<#cKWqa#14jV7NbXMb9}Qyo+K-t%ZTZM*|h04g~@KY5fq$Vy>_g}}gz
zZhDn$Sk@t`u*{vL*c}f`Pu||3UwjgT&f|Q;#pi?RSCo3AZgVecN!p+G5e=N)zJ3o^
z^gQt2`h&sdAS{~MK{^tIsL~v|KYkVQ^i*<Wgf}yR;#eaRp24-<z#>0rU|otJh=6u*
z#RuX-u-3Axy{{B&WqG}K%XR2a!WOG-c~^NkYfL8Wi=;k7ufeEI1TKk|Xj8+oxFvV<
z_wDrma8ZMQ!5cD>95y=)kHeVaKjTsPYnZ~-1~JPCsT%950W9HC5<2)N*^dqQ-A$06
zpPF~~b}u5yx#_Fe&6@cyCsdM$4^`*Yp9`qV2P~Yn?r#uuaJtelo6F+Gvx1^nB60C?
zR*N20zJuS37;A1)Z#BzVOnZmokC(}=JuGF$q>DJbue<;W6F$HHi~YDEHtx)-o)Y`G
z{URY5GTooWd)#BsP+urMgxu-_;#6-f6fY8Ms4U?r|Lxm%f@|k;e+jS^rlL3M<U2w*
z?tzeFo$3qf^tl&B$Ubx=iH>`yJ3p_YX))bqx&1E}z{+IOr+C~s4_InCG^fLy-XUuM
ztf;N7jA7J#4O1Zp){7YRd}oQ3S}O5u`+wN_%BU#Ywp|6Jr4f*l?nXpFLSpC`x~03M
za{%e??(S}o?(Poh?(Tj2Jny@|wb%as&aYYby5hKwIx|f!DHpWXjvnBw(tqD=+&$yo
z5puSo;yWNN7>dW!{<5#G%Tho{(~>WJuiQm@yw;YEzTVl{$$tx_^|vGBpBj*~zuv_H
zp+AZjbbD(18$C-nC1ECC<gZLug2{T(`)y&!$#b~pV8k3qH}r+^!WMafbO<^Mx@bdG
z`@={5&jIMxIY@;+2BO!uV%N#_ImNh1Clk1mv&5PysbwqPwB~r|Y?X`8&;}M+mcO|A
zzLUoZAjdPEf_#V694&%H@PY!woAn%j1%NRoiO|hmLQj0Z9~}!XeDCI}0!3}eHMeq*
zrJ(Cb=-%#*r3;YLg2kf$7+)2m%h%E=(@158!V6E@4x!9q(2Wsr{`xWWW5QV6Y_UYI
zn&7$^K;LXoX)u@uDpOrq`M2r1&qvpsE&CP5GWF{#xFphAr4p;C6z@gfm98~-gor9;
zym7aIB}=x@Y`|8>vBG|&;KOIO@f3B*Lb|nD1vOZ35-odtqQ(~a$~>7_6!q{mc0q!i
zmG+yYYkT*E5b!J24kIBZ@WjXB6cfM;GgrfJj9Z`}i8Rb{qFfPCr3Y(ot!d?br$xN5
zI7BmxyHRfK2g;&CZ~)6A4%mlGvDYHFJ`iF#S5ByoLy9Fn8y>Syea(=|7aK8W>~6Ap
z;SL5(UPi29`YodP>TnBA#2afjxNWfN8ubR*sFW&wmR<aFU|+OrEO|D1iQ*M=OH+YT
z-@6o3KOFF*`p5ij-5B?;pH5q2V&&E+j+$;R3@KD*D9k;tT<4eiC|xgIj>PeIZ}fg=
zt0MC0b-AeQSRka*f`!hBiAYfXR>lATi3<uJC8d);Ue;$Z+;b-41dmqL$enz|>i%96
zk<m$IQ@+Z&Y^v_PNUYBNtwwxIijP;E`x&L!C*G@57BoV2$B8H3jSqhtEu8(@mvZXi
zQ;2*llIwywrVEwfAjl7dsm=CK=pE!^T2;DR1DN@LjE{%=)k2fm{njLjMW&;e0uchO
zYty}x=d6O_O-<+c`R9?ROO-*H9^&GNQ0~7f3^tVj&P9N~fX{TMkYZMzylbH{UW!e9
z@7{bk3OfK9U8|FFDPJl*D32k@u#QfNZ{2}3PwKpPJ6L&sSAY^)=a?BmO-+aw3o(~d
zk-Y139A0=@<jG=da^zbIq2ASJmkzg$^a8cgq6S0#*ZXQ5W^Ok|#+2svhfl0)3Vbqt
z_)LHq*mH;U?8F_h<$L}oASFkq9$M2k&YPEH#tcMW|86w>zk6;7%D4pq^amu+ZsLpS
zT?bRSuT-u<$i`qKPQyyA{37lOousXt$J2ElN$sLP#upIsebCx*;~kKj2Zun|Mx!9C
zogwmAj0EVd6^?k5@5nSj{nTxQD?ek<I~6pk3>rl9MJrK@Ksl9lZM6m#TXU(rQ(fK+
zNcBGy_h^;;v3~bHhJmh|W@2fPqN*0oE&3|k&EArAqss+}@;{8cqxd55HSf)8ZcCSL
zdJnx+agKKI{)Ym!<|44$hIEfk-Jy9Txrn%aDy$twZ2Q)ErgQCMX(~9E)&4nI)P2a1
zUhigtCuNuO7kH{@Nh|gus_5o4YE`H5RF`0pyaiiKWs1jax-`?yo!9-&$3mhm&`fa9
z7!#T@OL+GK>8Q__TtW&1&Ol<YQBKaRksUFfcXD?bDh)9j8=D`^iJZxF6Z&Dj+x7+R
zT=?{~OZd@<R<ArqTfB_6ZrzZ%2A4<R^;hS0eRj(!6%Asrocczykl_giU-a@&&g1jh
zI&>Y=0wbH_&>IQ%!(~asr!IArU>LcsI==*VTMY4dI)-1Uh1qfsYWEFdR?o9zs!Vkx
zP`kt*Ak@Rj{nqmSNLFi6RjE<|fJ0<plkBv4N=Z^))-QXySBFD~bil4)MfQql2H!XM
z!w0MD$>-JIHaM5dyX!l`>+2@PM5WA@i<~jk^4|^J_MHoY1WlgG)0&FV{0ZUg6dECC
zkHJ|N#?r^cgEyx~8^qOAu<GH6-Se<B_<zJB&LX(l;+oN=op8xgsA`K8KxC*~GJ|w|
zInTn47$4gh1wuBn3u10xx~Mc-Y~U6ZX{#T{&x7Y!^v)?$np}eKH4C~#RoJRT1h5CK
zm)ja32!kX%tuZum7k7WoBR>!WcVhmAWtZU&#*;=RK^9eb*BJKT)A+fqqaZTTG5V;z
zTGRU<4P21)Lqi5s!6U=<=BHPWB5#^#pZFN3^?wo_`y~7@4CMnt<4q;9rL7+BH70W(
zrIPVkz-m3>*s3B(_Cr&U^v;z9%zPO{)2GXL;p*0{p*1*^%9#8M3VkIskK=i*^;02S
z+uHQum@${z<9EuaMlsZ?-%Ef}_mA;|o+OZVFK~fj6mlQD(Nci*PI$d`>?Og<5zDll
zXdJxv*sCi_Tj2|5gSHc;`TR6m>)V((rJHx){YJUp!oVQ(X<1n4d#-46D&4yR9cs0h
z$!^*Y_m$N_flY^_hkwf#7eGHY%dY1g?Y<m}*1Irt`ye9PwEPoR@_O<CJW$*DGh4Rn
z@NR-I8^pbWYg*bz6E2>g7uiiOfuMk<4NO(oo5+Z*0c(s6RYO-PkLVENaO0=yLt@|s
z=@+e3!Bw>#0I+acc+yG8phC=2L$%B&nQg@+LBt=Kb(v1qhi@eW#eA{>ekb3Iw@{@N
zMV+B2_?rC?LA+2U$5?GpiF6-E3K1$Yr$D>enwy;?&b?X21r7W`{fil4K!Z@O=9&ZU
z=2ySk-0wc5`;F-5%*84ay}6aMLM1xH6#98ZO^R&IcBXS}ZcnV<rQUe!Vwxp^LX=V&
z+0%_Xt2^Vgu{V^tEyNg&p|g?j9z}z>dhx&rhH%X?)pAO6K?}LsxT9q;FZymZQEtIO
z?`fm`i`-ZTnw9jbb6eAsNfrTp&I?DeBn5-R68f?jDbd9*u<EFW`kePakB>gTV@+Xw
zjl9()?V7Pt(>^+JA9SU^P$kCzQFPZu8Hnc{staWx)gl1pa9r~rbBd@nb9MzFR8Uzy
z%XuZX7bC)}!PZXI8KRuyKR>~$p!836$<#vsQXa;HA>wuIpEQsVFHxkI43WC%j$6iM
zr(hw%fd$~GS$L&L&kpc<#Gii^)yrvlee)P$kZ>Ds-elI9p{!0e>_{B`s;J;@^%7aJ
zVf$|?X+gl#qei>_6edi&<!${$TlP<9($43G?UiW%22O@!4dp(n53x<VYql$?aDljn
z$v4mD(;0jdC33;Ub8qj4fgz+`k|YsM+~_(2=$vlM&vxmGMM_>aj~vFpi;tbdG=$Y4
zh7UCw`Q>~QlvgCO&r>+xJg;*ny6`*UMElZ^sl7$@Qm#rLAup~@-;4i<(1I5cvy`&G
zWm+7v*0%d5#au%rbwH~Oms#cJe!BisSL#>$5ArK{>l0ldbAqa&-U)G$9x2TTu>nRN
zY78@FMEwt)Hc77v?DAR6@qjNc4SOv48}a2Xy}rBs{wkh)p46x{Lz!0Cuj*2|WSQK-
zSTu~!n6RP9)2UG9f6J7&NTFoXzWG`CoK=0rW=h`QTdV^ys<r!X{IlU3fBoGQD5Gvz
zVeO9oo~ki;x?$bS&G@sX@Zpg}Ybyi|9rrkINlAPcd;^0tgt`Z41%fAu8Lzd2ei!=^
z%O*P*560}|Ag~BMk$lH>eZ!oCB+?WbBbIo-vyCsrl}FPRS%sB5L$5(Dr77<5`KEj+
z-1ye`#)y<>BB2OgX3~O*s2RSE1uC&9PJm(FXYmi@^VHU>*o0(OlgPdj)tyvCVlZp@
zr=JrTZ4`Iuqpz~%V&7^`=d{ep`09^-FyJLmx`;=w<C4$=r0&%lBl~Vs^N`hkKV=-u
zBG;F3=ey*`YKNTF9iF*F9Y&f956*5wcnRBNEf=_Zj@8OlRv3nZ$F;{-(q$5YFHFn&
zTi@#MEfMVt(w>&7YpEg&5I<n7W{eCpO7y{yb-hz%G2q-M{XGgWDTz$t^(3qWBVw%)
zZ7wL7tF08QxwL;bhjYoPR`ZakfA|SJDj^{$VFE;7A%MSm+?50#ns#vG8y#;Nb_;7&
zto9raOx7%D{Y=>FYQ9m&=*V7~{M0yYMzwx#M}duJi6n(;Xz#O6YbYFyJ6_-@%4_^)
zq&wtOZkEaAT!(Agk@~RTbq&kVwM>nPM))IPICtymnS1;xv>Wl=&SU#=kDQYOm~$@H
zVgm1>`r@{6l)IC#83oA%<#2A&f5y1Egc!qF0tj#Z@fY%}LPx`6bXcHG;kIQIL=I?}
zt&{*IgnbMH+$jZat95!q(vmhGJeogYe*wr7JnZ_MFJfmmXFaQ$kK-){YzgXNg2q2M
zH_EzMvf}f@xy54{DHa7`ngjk&i&T(-mG4r}_V9ClZwX`P_K{<=gOnd7X}Tzu#oHBV
zYFkiJAqGCxLD+Qq0l{wX26?O8Y>=xH5!X2uhTea3p_4I?*5|}={h|@x)^Wh3R{2IM
zCuTkEn>#FOeBs^hotVQ4(>UBW$WQ{g;J(81Z|Dk(YU}a@eWO5@oov=`#H;_nU1ylM
zGW7s?92lE`olrH>Wn0Tp)Ird{e3P^1-^KrS&rLMUGmp&r&){3Ti+%E>ul838xO->y
zPVDY09^{V!vRI}#w1NOWNL8f)yeDZh-wy_rR@8O8Phn!HQz_kLXh<mZ8_EQmXeqjc
zIYk0Yz@VlueHc8&?WsN)74#PG+>M+`h#LgCk|^u!Lsr#I_$jZUQ2UayW%baMZeFEA
zp~7mL>zJs;<LrrEI-JzU?tZ!b$Z7x`{(Tk!W8jo3<3faUh)E0NNLZsa3p9Z#M*$7&
zbMB1NM_e7fY?WUGAeS$AQpuW-hcg;d0#fVa<<B>&9w(~jg7+M}mAr^TUz}e*MXeBH
zzR0S8OdbPDcG-slb3J2YQc{}jB+>9jx%$;;@T9B$uI&X<(MQ-GA6nLwKIU~2BbOHd
zAYG`WuR=b0VM43&3H>Aor?OYFf1}Er*W}WpDb6x~nY@(s4R?E;w3zOZU;ypJvcr1O
z1u8RJoU>1)Jrg}nOz?~PCle1HvcXx;kRJ0C8hybwIKYQL`XUCL+IL28--dzWp7oB?
zHWR8Xab9?Bv9D{3vy_+eD)3hEKYJAQKZo|=gfq3aoju2{X;-;H<m@3H?#GB5f>WdR
z?{|(siH5*a#qnm)M8!EV_*NmA1*M=kIj<>|c&y4z8bI^>qa^&|7BNbfTY>VN4X0Kq
zJRTV3TECUPeu`|Sw%^;c$<`OOjDr1Aw&EEbN_Vo_80_L=kB+v1YN2US$p_B0t+VYK
zIsBq3K6M_qiMEH+-EU)mFeC7zfX~i!Kj50?`T&axHO@s5Pc@9z&>#7hSi*=abtF~1
ztM@&NMH7{s*`eV_yUw4xb(#q$v2ms|i#iy?BFR~@ncu}9hXR1<?E?5nu4pu6c@9#;
z$mw3d&MPul%5s)K!{!Q}Ip#Za<dm}EKi}w?F&(1&3vwENwTJqc!E0mDqkP!^j4Jhh
za!e(LJMLWWpE}X@dJvHGX#JMg0s_wTaDJRM#(788Pg6Bt%u$QqDy==I+`(MnK!wQa
z;|F!$330Fhj6iG#1<7y<pIh4XjJ>gP$WMHV>?e_KVp5;NCdM_8$W05tIipLXbGn|<
zA4V%o&?*;~%#yml+$D~1?<+=iF0+~Wb3yM20m!nv>%bbYGiB#PH&>*P0a|JZ_&^-q
zg3(rEyd~O2`-Iwl+uH=m+IWh@kS=l(R+@milk(J`m(D$7PJu~NWyZpG|Bhm~CS?F@
z!(lm<Km&)dF4csVNY_s4`5cHfLMfe2oeh-`Z5tnk0On8-@r%|nGO)=-rrzhcaFRIa
zOAqz;59Ifg7=vuND|@)tF)j=$kQB-WDIjSh;bEA&Ac+-%-Fxo}YDk)Gm2sHtbra?$
z2w+;`oi-<rC5Asu)#a)1nl$d<#zV(-t2%TP%Xd0gO)?$Jxz=Elk-6PNVMRpwZKi)E
zS|*+(Z+7B2`?t4)RHyG2dSn%K<NNP(E-lVE#RHQduBpEVZ?7wqv^EBow;G3uZzOcz
zPWyhO;{b3gKNY}7R4qjJ$yl#13lng=g%tMDkoCPM;()J{o<%ef61%rZYDOiimqSP<
z^w#FC#aVxHeI=Qf#+p>)ZdlYySdbkKO86x_5ga?qXLCfX5>knEm>l*ouY!FExEOuD
z(n5>jjgAN~7;uUkedJ(+GSi)_s*4$&ruvaIX_t3!2c4@%f&E3SNKp}2>^2r~GdScR
zJElH6+Z=tzGicCUa_pM%U~Lu=>V9~*mMc+Vu^;XT#$iH(C^{83NG;7K5N!olZ4VaB
z#3}#95kKL)0(wXq9O%N{&;HhAa#DNT8Epq3#b3%H0`Bh=6C%%1|1Ef9Rw%mgvhIrh
z8YU<7S@Z&MC89ie)|XJgT?8j}_6rmkj;4!G8ntfjZ9ZNnvqXQtTl}g;vdlsd12p5j
z#hoS0)&ON%Pju=+Ky^eq_5~5TQ+os1aycsK=UnJh)TP_^WKfxmWUgC>Mn*CfX(pjJ
z?n~8oZa;4kz8`%Tn&Au=CJGUaO>J<0-uO&*BG&IcxoJ?VO&TxU1aD##_-FgWG3*L#
z(?XSetTu1Mvu7c^ctCUxO5U2JcIorc=3}@Dk2?!_k(Z!_pRQ-mn+a)WzAMfB;0KZG
zsZ{@_8=lwOuUQ&d3EJQyg?wtNy0|E3kEgT9LL`jLV8lZ77mkN7iOvQ+ykf7X@~dpy
z)hopg)h}YX77^Q@DqjVYQCZqDI{U5JO6Bt0@Dk7G##F-<*NYW;Ls?%{xy?!oK2nF*
z=kTKAuShMcwh8%FqgI2k340!6+)@V0Wt5Jh|6n>&qF~r_iS#4Y3f8L4bl*#iWiO3e
zn!5*?U`x!MR_TRP9u29kpnAkOfS8Dd=NFl6n!}L-e`(EJ4OA~Qe%Z2L`oCi5c?~?)
zYh%z^aoyIO`wLUwuSe)`xK`9Nch7%72Fk71t2A*>={u8MuCC8E<C6nzyG`Ff4xMxw
zpM+)4Ypy%SuwUb!d3M1}pfjFcp78T8V|v5S60+S-(l&P~4|w%flz+-=Av!~TrB06=
zs>{DpmI;2a>QW<)Z{_UZRFi+`VEeA}n`IN!H8}|N4hg3_ewryD?Hlog+~s_&@6c|h
zNpM~SrJGbuWqbN&Uj}~%RLP6?o^S#ouYa!S`%pv?!LD+|UGS#GT8p6M$Y8&98Ol+k
zj7!5*OB_R%U0x@f9Bu2{)xDYe{7$!N)i$3cgoFc*_wkHchVHe8;~(VRd1Nbp_ShYZ
zN&*i*Hdd_j<<`%kd-178NFm5yrf>{~;-AZ&&s{^wEFBjmI1aj}!o6{RFIPB?xUtme
zb-j2y)Uq24A9eFs4FQ|o8ipMm;ESfZ71udzU@p+!LW7_+enVna$Hy9oMXqo=+DhE|
zP*E=qNnUg|1Ywq7hJX&$*I8Z(RA-`P7tbF{Kb9LE{O_NwAJ^$Aa<AfZ1l3Ujftv0*
zG`uq7%B&41E3MxG34#ftxBysix#^!9lkTsse5f<J8Qk2cA4GK#aZ&9QW_|s_wEAgS
z&eBr5Z>R{O`OO`hsLQvWs<2_yMmSiSp~@h?h^ZFUickNLj->0p^y-%u+qc>qKq};8
zW=-)y6WJ;ez=R8QFnniaf1K%Wr5<nDgYim%O!m>SV&(@Hy-Gyr>*;R@x-PLdK8|IE
z%khCVvZGL_%V+(aSnNk~P7l$0&z_f`w)F($H!}L12m&$v)^!4P)S9g|!*^yz<=6w(
zf1&j&TA<|jjOadfFyD&?Pabk=r!@&_l3)XL(Sd^E8Sy=5bhv-XJXHtPizH88g6IO;
zj%QJAQP5?48eDSok}=i{ZBp5qK_AhJEa;N>VhL!hU2VHfSmS!x2}tZRGBi4zyQ{<T
z!~H}dz$v4&Ka2VqF6<9MLwS(X+Pc+CN>~3^03A3#UHrSyBNpKpofk;|7wOTe`rg4}
ziSXGq&z)&&O_2T7Ra7Z`>C7QTC==oQ`JJ})&X($(^fh#0qMXUh4$d8{OEx%HoOR4=
zG`;c7kc|@Hnr3La+3OV5$sLUOK6ywkd3n~6mg=fwtRH~a;ZXoCMJ(%E^UMxqJE2?^
zPWSk|zOx?`l`H;d`!YUc%i-#b1-IS|x{aSWrdU>b-5Divost`Y$rOcv>y0!X&|&_S
zIG61@l`C~Y)Bs+-soF~kfuS}{Uc7G`GM3iWH-!p!ONiC#s9XxAoyR;!-?dvT<Y!xF
zj~%-nSEAuV`BT4PAgpt_N`A4%?EPz^2T)~|8U?<0{{&fub0g(B-&^yiukrs!-dsBJ
z`TE~gPV{$`-%)J9{)G97`d<|jmOfVz(e`LJL;pO=l!p45dXUjAEh!TyHM$@nMadN#
z8GMLHO1{OlVi^xen5nXrz-!Boi4cO%{@xwO65U5l-Q~;W6Qth|azxToG(4^BsCCgG
zsmXJlOx=aPzrVb0YyKkC`dE8fzOaas_RXai>xjHi0xpF34Jj`+tp>HkbE|}hKLuM{
z#f{X`7lQ+HilG?&ZWgd^G*y>@2%hD{l`tvZXm_tqu2^h|KhNrvvXrs}p0uW`rB+iQ
z-Uqg`jeVrLFb*~g*)s4~j+gVlSq1UCiQFmT)U*8EbZV`v24oM_N*OgEC5!xJ{%6Jf
z>ss>bhXHGm<B$aHe!4EXqVd>uIz`|dZ_{Ek3l;SP*0vg+<%O(3B7YJx)mr+04{#+v
z;4g^*a~j74zxxe|;xTgHSQ~JTD3z_a*|EhPVFi?KOT>Y`>q8C3^|zUPC+hM0Bb{KK
zMz2wYg`7VoKvEiYe@R0M;ToP(#Pp!p^5}n#cqB2(K6C#xcjMLP%CUK(wyvSvAnA;2
zA<3T!1~7BIsTne9JfgS+MU8l8pABqHQ2bqyF}0Uw=ZEcd-{XW$k_w|nK5NLG!#TgW
z0me_IfhY<VqX~L;4s&u$E17g2-{fd+=~vs2fN7tUng{Lr>VRZj2`a@x(|Yqk3U6}h
zP`R9?X93e6l}H2{+8xH{A=}SEN~6^puP*M!>E;TL-hwJ9$mZg(Fa=pzWBbKvXA7@E
zBA4xWyb+qGB|ihu%M-)mV;pJVk&yci$_tecFmFwwC$eQwKe*uaq?@v)R6xWwr4B4#
zIy|V=>*&|k?p{VuU)S!RE!HAeRNC}DbI-*xD#9KbS_l_n<o-N`Gv>wkR|bIw8W9*#
z`F90>0*AAg@<4fpS#B_nl{WLG$Spx=tHp!i9VI$sC&_F~gYDT*#p6Y4OJD;UMinm5
zRNJA$VV}0L@M3CuTA?IN<2`AaXu#*33gMg&&cGPka=Crp8tz_6yzC5Q98+Rij9%6%
zf2`u1!(^@mZ+M-BP^x<mwkpBYb5Xpo9Y&I_g?9^??tB7)(y2Ft2ay|#q;10UwWc^-
z(-om4WG=C<g>G>Vzg;gG$^2H@?9;F39PP9dz9v;F)XNa_7$BOS8*MIp_GgBxIbKmJ
zwTX(3Vllg*Cs=Y0sr_8}x#LgZuO8>2$IMq+$lhlu?ZEJUX8{hA)}h4$*`jEobZ~Cd
zsDtq4Z0I>aX3-<5_>~U%U=<e4-tP-i*q+WPmD)>qq=u-<=j>SOd3Qv^=F@apaQVoB
zNGO443kR^cOA~*?(sjYGrc^aV-Vz7(&5x^;glm;!S;^s@nDZ?AHRbl0`;ZmwE65ii
zQxO_n99BCI80I~Ja-IEmLdHQ>?~6C*@^8{Q?Sl7dZ$m%JnF*2paEv~|-Q=JYG-w#!
z$PG(f2$Wxat#53jf{6CsWt@wOi|M{#n14WiT)jzh!g`|QaF=%RwQW1xm@-@EQMElY
zR@h&>O}-l8{+H7JPqj-#JiME<Sg&*a`Q4Qg=bXtBZ1!NqlWUoTxO6A~5U*48Rnmy#
zkMhht^X1rl$DRFDXBQ*J>yjJZ3Rc6<FTB?9OlfVNI?=t>*0%hWS;EtnOeFEF7jdCT
zWj)$Pj}up)_^w_`^Cf`_6E}V*Yw|dzFtOsPdK`J_>dFc|NVP*nr^@9-iRMB6SO8*r
zsP9DZd~IW}VYdC!Y>9gWggM{`hpW!Cau2(3;Jih0y<Syz1j~;<<6hQzjV5CKK+hgA
zuN%x4e;Q864Nvh}e+7EW^KaQ==~BtQDI6DvIyY>n=Sc-I3xZOXpCNdN*o-AgkP9b9
z%~}=1w52+<d)Wkh^mou*P_o&E*OKM<@gH_3J2li)Y{`4&SKEF)N*a;p!w<$NksY(t
zgqvs)mEO<)^VvQr|1E>k@JdAOL?3s9Nm1e?F_u7MtLc^bjHZK~c`ss%1W|z@BJalt
z!W9SxI~rE;G`{-K#AvpC;*lirEwz#F%jEpl5<we-0sj7^lan@}CtKqrSJ|5!FAlG_
zKYw}t45Ck~#N}2x`r&RzxQwlrAM%a{WRRcJBGu7ia%pXy{V9@f*Z~(-4PM-E9~(SQ
zx*;zPMzq7<>uiW;Ibbj8O44_*JXnP2-Xg$ABBB*H3pHja)+-%w;2cJh&ej<$Rg7C@
zPPRs^^+z=cZDf8VD<qG0x1c-qCtaXTZDm_BJYD>4Zc%`y6E%A^u>vB^EKAe*Ht)Z<
zxqHAz7n1bPJ2MhlxvVDSy%1|Gbr(%AzVk1>md7{Xjpzt+F}$$py)7hke1m>#p&v7)
zd9pIAj@xs))2-~^Mx`14>jn6(ehhTQqaAqCE-Fr+A^!lT2+@bqV~gY@oBRSBOs(b&
z9W588m(<7&sc6y2psO^Q(Qptg^qU*Y{(f5qqd6e&`qSm?e{`c+{%jL;lP$f0KNzf^
z|Mdfj_Ae7R@By#U$G2;!_Z~pg)i2k>_>Bq1b&dSor8j=Xm$!|Uvlq<WmRh|#>f8B6
zXEo>dF7<Ju9%7cOPt_G(wmdK}4Zuk+8tf0OO>}gO0iNt0rjd1GcEVUlQOv%R^VJ!;
z#xd3AQzKG|#g9E?yVDt|j<p;y4a01|uEma|$dw%t;c5q4oDl)DP3`V;M~k)hM%>Cb
z3yU5Yen}`jS0*p>O(ysfeFr#u!Q!!Ww6)4E^y1M6M>_$!J^Rj8M;;MTQ6aK<8<Z^p
z3^{DqBA0O^Rai{<GHHIz!Bp^=00H;jFSYu8O-;@8NjP{kN-fPtqakK8hr@!juFdHM
zgEAzMI?q-UtNP7|8oOdMq?P#M@ogKi@#^zgd8%0@Q2AQN-ZW!ILTQ&ez1sU4|BsHl
zrTzCBcr9Z{#{=<O{8tkyIc1)IdyH?Crjm2x*+L5XlnjrSX`FX0N#0FWk#_X!SoL6Z
zIA(gRaK6R~GY$rxM!KM3^r!HlWhr)3W=sSL%6HQUBFb;f4i_r#ST!o#D+<!`!hnu)
zRZh)M)-wf~(4&!~1*-Lie)0`A=dgyZ!y*fa2ASFAzrvomz5`YXGf;O$t?jI?@E+n#
zqf|IYMUtg2My^E8lGR47kceb<P0}`9yni1LD%#foMD!q5?s{aZw-wXrLd-4Jl1>rA
zg%Sw$=|#^u(zcZ-g&pq}U;|tEX>EnVa-3x|?X|EQ5sQrwF#ZTNkV)nQN)*BJv_5hg
zs*L`K-?qtY8+cp~QDmFvnWRg!u)Xcogi(99v0CaXDG<KnaDFE|K`i#pDrZU*ixI4%
zGyh-ukfi%<?smVmkF&W-$zVg|m%ID6vBaRYTnJ;OgwPYd?pGllEgPeaKU@k<Duu=Q
zNcr$ovHF}QYq8qICgz1qiswvLZz4ftR3LBoRE3yQ0mEJCXZ>8MJa1p+QF|C79#_vm
z9n2mZy;5BrUGu0zhpkypn!xi0W=;##61@}8AciH|vbni*Wb(uIMxvC%%ykXaQK^_k
zcITE0r|t8k4`_TYJUKibp1~(3yG>lYl?Bt^UegXwctpFIm-vL>aJlt))Nm(<%oZX7
zb|<F@_)hp#Wh&27d&kbRUw=I9A#7-U*Kb&LcQyRzKZH}>-rM;>CY@Js8_-(t2)$t6
z?>FZsgva;bM7E@ur{E0)_d61;o+R)Yy{$7}9z1~agU!lwu&-RXHv97D$94dBpyT<{
zse3|aa(*$@<J0aPOM*T(!aJl1+^Ws-UEuKVI4*Ws`~;*r>81RDR@|4Apj@PQQhHG|
z!cVnd#X}dXWn?Im@{~-W3L6^T>ClPF(i-dpoGbz29?r?7Ia`;46t}32PhwN(nO?IN
zMnvkpuqzR-8?v&iata?EPc@;H2}-cord)+=rd{7XFrTi|yVK_SCxOR9Bdc1iNe|Z3
zq^Nfgr(y5O@J7#}ZA&8r;juG1Kb-lqT~z~)>w+X`uI}5<LmLue|9|%o&0p_g{@G>6
zSEwYfe<dV%j19lq_v_yl)!C3Ym8@YDK1FB(6Rmm~QyY9l>qVa)|7ddr*~sWgUL_X9
zkjG-`H>k&3b@VL*^Q#Mxt8ih!owVIV^wdniZ7L)wip;@?wJJiT$lS*a!c$}>MU=wl
zexHK@D*P;FGg43*ImnBEM7qOpfinr!AKmja2pOaYj3!luww1AsDOQkXSbGIRSOGUd
zy?m~(=!bM<q!{a?kgT=<1oT0V{iqwtf~pQ`;eDP{XU>XsPq4;bcp^oeLGJs)Ty&G4
zyMw9b-vTS_-R5j=5(jH2rpCJyR#z3Yewv;fM6Bl=vOnaa9|<1Gt>+`sT6}ZdA|Nw}
zIt34yur4>pI7*7ytUb8noxGhC8Cog~5sJswRaaUpz@}~ir;(sFF=D}uA&IA^;yd(e
z-Trnql(0m0vn|3)J4PFgHn@79x!XP#-RQ+l{Oh0Z{I80g0{S!SE&Js7{r{79tc)&a
zil^Sx=@AIKv~|iH(j#en)Qr-Of80;q?0VjW3kx?IORKa?rV2wCJN}5z(YaM<$!t8&
zpJ7>uPCPI$0Dpzl`S2NnuvXga;C*Yy6vL=4lIec2RDYJdGuCe<lw@mN+gH^yg!eVL
z-%>uWDN|@5inR84<Zv!z++pvJ4!enIf4@2gcH>@;?L?jhUF=J<{iQ?0SHvw0FXUaH
zjh^7Zn3`!O?F!o^X(!jaAVf3DJVBjP^LQnh(0hrSx!1I3I4c=qf_b{-;3B1BZ@i9|
zVatI-q=S<-M+}VhyI6KncLe<gN>avjUJoH-*@lFlbxOx-@W{zt57&nS;FGPRHkW(a
zqf&yiGp`J_-L0KBQ*vE$8T<k_s+y18E{W~_w`1qqFS!p-l};l(Pqr_d!Cn8=goxAb
z;s1d={DJ`F;e4+OZwU+fRF1}Jz5Ryk-VVb*{bkpVebOREdFRaLE0ih8*`FI|n0c5_
zB+{~eWQs*7?%V}4aqtEp2LhcORs{60>HZ_b#Y}+T5`W}uxc?*%$_Cs&n|UFC(FCaq
zsdJ&rXr|8pNQJ0{0RMyK4~u|3{dEg3ls0>QFYh1|sr~i7Oa_1EA-y=!4<FY#p`d%;
z^LKpj6nsZ3be4GdS(*H<W^Wz}vxTxiL4&NPGO><Ln7;J6bi!U%xw5U^YLykcB%3>|
zstsK_^&$Tl&(y0d3uB>)^9M^^A3W+p;k1PY4jF++dqr&~n4kbf$L8`vYWOx4mGo6i
zoM}o6_n1|UWP|~IDJx%C@73?Q#MHg(q2i+6+u@sloJ{NZqPa#5`jEFz408=f^b=*K
zJF#kyq%$9~24O-aYf8##(bUd-nci9bl~J7?{_^`<%T>YU1x^v~>}%2RA93;kcORH{
zoYcuOEeAO4W(^M#3bWcU<RUi|R2Nqvr>{e9>KKCTbL(|oEOwK!zSo(xRHQUC1N+l+
zUJ^M-MRQqSTn&)Wzvga&@n(h0$;>HWc!CPo`kul_g@dWDL%HWpnK6!Qk>=*@!BI6=
zwff?rlLm^@v?cJESRZ4L@dmlZ#PM@QF8EAgpyDwYbuPSX>KSick3gVbVIOtWrE$+6
zhTm$`iZ*}Q<CXy>J^=0mOqEMsz!&iLA3_`O&}qVGX*c^cwEr)8L(+bfO%1v7xK+m@
zi5`2e21c~GcVKNMdZP`4$Jx|`of8I(PCzyK(J5qWx4$av$CEca3L)kbP`EQUSD|Xb
zk{mRZcmFm~EQ1=$Wh{Zi)Hl2@(8-p}QEL5^ka55_Fp9Cs(8~gwM$Sl`{as%L>%%ZM
ztD~rRu(sL{kpMoAm%^$+n>x!?SY^2NRpxo0e(O=z%dhqq?0KBc!QB;>-<oxT^66+5
ziW2spuQbIGy`jHXn)*guXPGxvC9vAFy4;*FvJu;FruJK-%;scJ@hqven$je(OOSQR
z@x?P~Mv~OoA5EeYX+_K3@21BrzJ~O>5b@5pDFeDdtCl+q2<E%GGS$?jq9aiY(c0RD
z3*?cgx6Ol_sN>Ri_=^JR@14v-fy?D@l9v<zh*kVVe-m8J@VvxHWElwXcIU_5?(y#w
zQ&6P1+I1p+Q6}_Q0MqS)J$VoU1H>(w>g5*v;03}lh)G`Dc*yMK)MUj-vZ8atDFC`u
zi$=`<U>gQ6?16$*t--cIw1h<6CHbb;iVPA>!WsGOI(wpHf))5jFxRHIm%h}vC!LOV
z_^~$^KY_)>PxM9)hV2pEc)BUFfamu9(kRX{y5ATBoXM(-N)tVDo6JNTInf}5TT~<;
z8+j%kLz{{5#eVs2yE;DLXLyvHW3~e<12XQtWO*SGt#aXTlgs^@#Bz%rrhh$gZ{c&k
zeiK!0y;y6sLeb!W9vHBB#(7UQG*^n6lSi#YKgv|F>gI~3WD6)n9{mhP2IWYR+5?1f
zQ^I1R?pb#S-bOm6t+GEdXkUnInUsIuun=v3dU_Sh!nF>2z)*;X2Xe)uBvJ;G`-W1N
zIC8n5fE2K`ty<afT@lX2M}7##W}+<1=hEkN2GfR0=*wW5VdgbZlbpyA5Ukk8^n40s
zN(`sPuuuxHP&LVx4msy?{^j{d>Zs8`k%PrzVHUI<9k&}w(#UR_Vt~`aMUg^9S9jvM
z2)Ve_XB2o9M2${*csGY}s&`vi&1rps!KfwEp6c(tVdYzutD~Zg)5cyFZdn>SUS;EM
zj7^uc>olQC#=Fk@;iVQIPDZP{@os52?}2M(d^m$KZz)75^i?j2%yY-_&(=HgHrMbq
zuAFk?<Xy}mD%hvni}%CJ-W7)hXWkNqLEXDocE#Ds%{#_R({BvKWBVm>oYlTKED0nu
zA~3xspdNUk?b>`qXH$JAoO(B4Wjji%)S}r7-woHeX`k}H#T##NvN>?M=TjmD$c3@*
zJ+*Pr{@hVh8Ck#H**tHqG~WX1LwGJ-EMt!OY^5z+jfqjnqP=ORK8-*n1I<1|ndWc5
z->x?-J#|1VD{#aQg&4{ehdX9`GK!>7>~U^^ZC8iIAxpG1S|d0CKTJB6n>u1CN6aFs
ziA*H9V|D;mI2cQ=@l~uIU!N6?5v<?fqh7y$vJn`p%QY32aIonm9!pD0F_()&oX!g|
zEL0LS@~bu#g1@@5F)CC-`am6#`>dLVTyH)D#vO=U-|EY)#OL7fto@DcVKNhyE17%!
zkdhSCKxdPH`4tC3%t>B`9r*zc<wY>!^%1DM)VnHi4nR!qc@fWP%n0D$%Mnc4eC!Kv
zhMZl3ZpO4Tk)Y5ahqQ+U|Hd+77lQBMMnHc$e!&$el_u@r4*~%eV`=~d=4T>xjcmYg
z!Tv|()7g-<J>JO@+5ULv&6MGo^N-92ln7)GSEP|dHB^0M$^}{-(&b-(L5zfWl@c><
zsFqv-pIp~VJJTq?>6}=Gk@f@!Tt{nBl8iH=8KENZrF`|{%o1!vG3}9CZsg1hpdiT+
z+sQJ@=?gGCUHjd4TwE`_O?#7(A?B_kS<S!VXCK>3^9|u^-6QUYI|eiao9VEAQ6b&z
z3+D68;m-$CjO{2hhYHFSuE#00Dj#bU-tOcbCwJ0JBNz<E4$OL-U+hcJBdU-K@(&Fs
zVK@h3t<x%$WlV-mGZmV<!<z+?ehL+UDN=iH$^tcWxLa;>ILMHgS$N7v^%}^EJxKkB
z^wghiZdiZNdcSAn_tGuekJLO<+y+WC*b(H(p3|XSCKdEbD51D0EYeWt;cI>%^@;a@
zci*)yn9s7LS2x*+>2)-7sVu(B3p^zM()=B*(8(wPERku%1Z%D-g|y-^b3jCFwOuMe
z1`Q#tw5d*a(1xwVr+y~j2n58Som$<mKM0@*aN6_7q0yM=u@3kw&=JQM&H5iOqyc?v
zO2<*EuHfWijZ#)#$P>L@zCKD0nwW_%u$~G$#|8^x0}!w$g(uTkM%@X>4sso}bAeO~
zS=*TLfyeMMn1_k5#Tt!Gc?|2B=AX3WHEVULyRv=yv07j<k;$%D*S@XX6XdgKA6sfV
z^gZ`7L3+s}$dtC?b<6o~ST(TAsJsB1X;Bg4eSPM^Ogv3UQCoA-%||0RJ4gtpv;Gaq
zsx=QLo};zdM!tL4`g_k$gNVB`mKMVv-Gv=@72YoNF4t%%*6`$%u;f>*R=>i-VR~TY
z6Gpm?nr&DkpH89sCI(d-maid_b~PfdIrgF+1ft>C8e@~?YK+vWC%PZ>js;o3ZS6?H
zwCg%o@wtU_?%MH~iod=-+oFBR2Am~+W*Zr~(LhkeHSEYo*(nJzt-cFcp-{s&{h)Pd
zMOHp5U@7ui-s+{1d%Svt0x4d1qNN9s{BteJ<yGfafKG)?MU<Fr(NWCB)>^)Gt!>W$
zWHJqJrdS)NBT}G1Q9@aVulWOn&>@<@n(92vSX{vlhRZH*x2fpa5s9X2p}CUdmlSLV
z(2t9Jd-CiGa}TFu5R$P$*(g1hdP=b>@5=YAJ@eWz(L+>^iJ`;$vn2pOH_2kE?z5#v
z+d$Q?f<tXVN&gH0YL$3w91C&IbFQnU0DN?HQp|b?){aDw-9+^+9SN(FZ=JN%4bRbf
zm%nI3>doZWYa<g2FP0rW1=?3(kT4neyWdoK_@A@HPd)9Yr_%$sU+ka5=Z2D>pVE=2
zDt>Ge*f0hC8{C{;ad7_jp5GIOfmXRJq)4$?7npsH%VCaNQyXi<J^9AA5rN=ss1RUH
zVgISxAV=V}EYRLQ@kaR+$?6@l@s~GxE@1<yUiZ`lUgo6MuYf5Nm~-)btc?7T51z|~
z3y;=J3=;wbQ%GgR5NTwAoJzotd`sttJsOF`CaYf=MbTRO%WZ7(!=y7sS3%^ldpzmR
z4K5-9wu-(F?6%C9*z;S5=r&>fpt(^bR55a4Wo?`?^fGaI1w(e!`@@`Z`=763CBa($
zhdAZMGb#PzluDG$CpsmprMYe(A$kp>rj*cFEt0G)*`dt6&n5Aa;u?VA#6wiS7<S|%
zz{@Bt%KGnyr~045Xh$l{=BeY)4_62jR^?822GP0?H4a6urwFEEPv%Sl=72V4;o5DI
zApu)u#isWMy_1`~O^@)>zfx5`lXjt3N~wOWxke+U1ocN86{Oj@4fYwh910FvE$P+k
z&6FsTewB$WW#F#$Nle8flPjXV_$vd8o#G8H8svs929MG0&_aFC`Oxrqs>zYJX%&|6
zmFF04TH`Z$JZv*__Zu37PWL?V@V5mloX0#g&It@GzK{<@aevvrstq6$>Lw&KOnEn=
zo*N?@7rFGh+W20+g)|GS)H^o%p!wSw?ib+`y5nOSm@Uu~>qs|jE9REj_B|GYs*oT4
zY_ne3Qu5NBai`OmD4sOpW5^Rt_F=AsX@`3Yf;k94=e~$0c29lc=v~8b*FpusUxOqH
zX-qpG3NRs(D>5T@kyKrBPhwc)ZSnUPk7Bc@R0nsM*B%2aP&BB%wwj9apy`^?wOuax
zW-rkOrh+&Jt2HEO*Y`TFO!w&O8@5l?)?w>T@bK1<KQ9^5Jma>5NV`nB6HgTL_1v4g
z=}~v4^9>~CgeB(^hjG*@5n3XW#HTXp^61zfeeZ3jV)mM@j}TAyr(~P#XncV?uf-?%
z7k(H4*PtJH)xz)@zd|`<!CAq_@R)Utn8@|j-yl!zPqQT&$rExcD~G+uA?UaV;Y;-x
zKRERD>mIUt-;%uSP1V<nv!=+gk>y@e_%^I8IqErV5xkpBZjoz>fpVOLu{)f2Cmf78
zHWIs@AA$_BxcBJ1oeJ4V)*f4a?q#!MCHjWWnu2`D4Haitx3DVV1lJhcUT3j1o-ae3
zC=ePN%3?ZA{RDvUMl)pMP_nbszd(9zztg7rE7XGUCw+Bzf;c3a5^=NGUW{FeaoV!j
z%uY0A&-EbZqzqiqAdMHLmj*&Ao}7__hzI3k*xF2aZU5Xgc5{P(s@OqVMLRGDP%K7P
zg^x`N3UwrA%0|B3@#Xj9w%tQ67VQ3H52VC4bMx}5u7as<Qrmquad1gO`gt&x@Lwj;
z+HUT#KOas%J)D8QZFiy)^L~0U=aCO;b~~slcu`!+;P_Z)+z*aC_%Z>9I2lhf^YVN&
zz-F%A`X1y!gf1qfpS;G(w}-!P1jw*1%QESFBrJ^R@zPqluPe>YGKM^ApkF(%%%(=&
zM{!Df@{W-2wAPmgQgN-9%M-Q%d^lsXp)r@DBx4Q9EJ=>$qhR)&-O~#X+!hWMzzqfF
zjlQ!SU(7&`T{qS$*J{xfH_w0N9Wl?1src<bAxMAy#J@OZsrkF-$!juNd`z=Ekce8*
zawv=;Q<3)hvExPRk!4Eyj-){K(NL14!u-C9_L$@`{FtZrDS7gDloe)T<YJUI=E;@p
z<zOP6HNjT@14SoZBw=hR9nG^Qx8wWdcnd=tDluZoI&NRiO#OwZh;Rw}S^py**E{;K
z)32lb=@{4oU{iKh4&C!CO%QjNQ($O_L5{>gngF1H$E;H;me}FnyNhVP?p$Bo@cv%!
zhaFzyNH2B2kJ^qzOZ2Rw7&Np?ey}X(3wqy?$_xl%4f-bRiH$J62?rk_`N?f$FUM=o
zlm@l$4jpw{lznj7Vv;6vl;N(7=aAo<nq^Chb-aRs2KQPYI8LVvDBZdfoT+?pmq)6a
z-iZPokSlR3qvbE17avsbbdC}QZ^B)726$-7=Kg#)9xhoP9VMsSOdBw)cko@Gb6um+
z&5zCRc74#gv9cV1@7R)>z`@Y&VfR{gR`)S*Ll`N)xqPB=Yj&lN({2Yt$$TMP;>+YT
zOMnv|<{xA|vR@k4^}`z!$LX7Jx8DPX?oPpYjBMxCx`!SKjflIdI-^2<n}gO*FR>1<
zE;VM*CuJC;_WygrRID^AVf|xg8YVdM^<4GkY_NRjg>&RQQrPnm?iLlVYB`?C>EaGS
zI-M(c6BM?2veF#T<|3uQOQ&8&|M9aWJgHHwnHBOkXm>0|1<`FsLxUIlAE$3s!s<UP
zC8$3k&~ejY>6-qo68<zMth+kLUM48IJ+k6kn3P-pYOzxDe)OiSB+L$990~sqzURyW
zc~`B4jD7=N2(l5ElTckc9*o~GKj83C<;&SJv^q}6L#{W(OY)%TDqL)fii)z{nN~eo
zt%N&2pHneapcT@cvdRdT*5u9tCM{4air?gF8XU6u$E<(PR(S;6dgvJF-4Aj~cP|~c
zsQb^q*0g8Z+D>>|FwN?Et6*~_Vmq2t)#ZwO9++l~T&LCv-7sN9-!pP3setw;vq7jw
z^Oc*)Asv8ry5;y+8fhzMPA3NOBS4i$X$#DW<1nqZ{!s3_Ci3T7eCeqCH$(70ya*}O
zz-bf<t2KEPI7$Vz!?2^`)FQoSmA!3YH%FQCwwaOl_lHrS?KZ6IwdwLQXQzDm)@;4D
z`;pn^tk2i1a|mr^kzKf831)8}tfn=yy&tS>jjA2=?r@R_%daWSf9A{suLv3<QDZo)
z9ZIt<p%R}XTi@P?vOPr28<MAxY6#d?blj(v30~qjUX9%c^sE@8Iw!F***bl|e;&rP
zeLdfqbr)@!y+zTJE1}LMoEgSwjs^Q6>W`ERtIHeqY=Y#KHXiFYZN*s`IYYbi>pL5z
zmpE`0{kGR^HNihIVAXCql5^(FTYo(w3)VY)o0e|7J*GKJ%)PpJcv8fY-rlz9zx6sb
zAr^JMCGu*0%L#cQCF?q8FI<MSQN7MwJ#zpV%Vop)O^ZQwhqCIACgGXxl5MZh52zf!
z98Xs>-R>_kAlpMgJsXRz&`sbZA8w9Y6W`yL)$K1vp7fjWP|qnuDi{j{45Gc~l2GI3
zJz{(`Uf2JyQvc~t5c<(^^G?OlnD8)H;2O38!;T{Po2kR)|Hy|X*_Fh6GNv~{?ys}L
z<qIgmCf>9u9;Hbtct?J{*LE{W^U<$bKfIM;IF9*74;)P_d~#o;C1V-tgHD%KoJ1J)
zx)5j;Ff$Z2BK3v!iR5=Z*$o=rfc*|*)D2(8H~aO{SGjWGri16i68>O^)i$G4n~oqK
z=o&=eA551NRo=12vg3Fm$Im#Mb#6I58F$idq2iyEX?cRUeQZ<XETM7BfLp#vb*Kxi
zS+xPA_}uuhh0xGCRXcR{aw`WxI_b|ApA=QL_2cAPJ{Y{J(^StVpMVsI{QdoL(o<{K
zDR+v`Yv=`i$wD1s^UG@QsQo{hp91I9#f3Xbh3O1!LP7#HuIArK8ITtZfxyehB*OYQ
zvIy6|F4YsA5zq#_F$~l>>Sm@@In&~X;f5cIUyOr4&j0g~3A{=^S-<YPhrB1E;|Q${
z@(rxqfDSIXDh;#%V{F>!SOax2qcC3-jA!1yLwpl62!65&g6ufCkwHYAU*AdtWWdcJ
z13fvtq=#Td_17S5hN8MQi+NWpXp0tAS8@3!GcP_5lag|W$D@b(wPkF;pX6q36~U!>
zuG@pv3t@~}_nF=sg4SRzk)a~eZmdY>aXtTN7!pAJV{CT7fJY)acHug`-T|LJ2mC-3
zh9H0s9Hp1_!;ARM7JSLgoL1N`h5L`(;pydD0U81EPi}3H*E?RuhP>3ngS+l5@GW4A
zXukY}-eo>D?LOi!)y*WOX10MIv(){Mmyr|d9>xoj{)9xujg#lH#ny$ND{*-hdarMC
z&whdH)h9!&Gbm{^f%u=S5s`Y2{}_<$eZ1+3W9VdK-e<M`!(XD=%t5Dl!2ygJ#{fKI
zk0kM_G7lh;cKShN7LYD3jEH>*8Z#sRezE>LM67n4#Qfx{^1QP{0^$?m$+nY=y9f{a
z;8RpL$e%~(HDZO|lgZk*{+=^$=N&DHxwY*r=kZ)uZg&nhcSZ^go$3TRW#UYJz_=?9
zID=p29PgEr@9k9`2>L)bGgdP%kx>^!H*Y5bw0_6luP8{cIjJROy|Lh+UHDpc2Tq#F
zbZ8x?)j|dsrXM~{9;u5tOD|LTV|Tl{JhLZv`2J%Wy^RgY=1&hzpU0a5d139${88F)
zv-Hn5Wc4N9wpItesb}v(ajuj3uNUCCeaV#E&P>58@iN5jFpPK-pQH@fOeMH4uGb&)
zq}eJC`%Q3X%Y79yPYTNA3Cq_g^$xu-RcfE#grznI?Ut`HyDoh;3z&e|^)dN9P>V_y
zmN)!%zeP#DA?hkRhO;?Z#;)|Zts7x}2q!Yu(7vS`-S1;6)Lw=>VX330wWnnpp*cPh
zl~F9~T^7pLu_IrbcFh;Og}puSgsrWSt3bP^dOmP{%j0~LLQ`b1=onHvRrbI(>*t!$
zDDW-!4*ZB?6~)}ahHU0YsA~bNQGekn!c6Vvyzj4gJ!=pqt&VBc_Lv~{)xp^>-(C&G
z&A^AwPQtj!ZKeakP@>N%LAK{#Dosx1j=p)IP%(A~z8L0c!?|<Uij-@sXGsoVh(E}+
z@)l~h{y7-oR;j+UGUvWSy<DH9kjSttrer>xi@jDZ5k(hWqT{v_Bf$1NJ;I$xS0x00
znWoM2cBiPhq`)xtL4)4fJ;C%Zj;KOEdX0I&drb-NThxSXZuOY4;>r#uUhVtbjC{;g
zsU&56?%yn>kJ#F+)s<NM7)m?|N5Z(7yOZ6(2YI-n!F=>Oc-C~$usV#iPG?SO1V0@&
zt2A2h-4IEryQ<xWwc+Yc6Mq}eP?DmsiLT?tyZ_Qw_Ta`yyjGtDeRh;!T}HQ*#dc-%
zw?fU)hueX0b#!MbqwFog2o=PhUW{DKvgWDU^G%=Nyystw9Wbo%_E*WNzpKNESFA4s
zrE62Cc?`{-DtaH=q0_~4z$qx|F0t(=N?7{0L5-|`i#?N1*kC>6*wCHMH{N-?+#9()
z-N|*d1B0(N!}%t)^e2Y#tcy9JlKw{gOh(nLB{-qJhP8Xq!Ho}4G4x1BIHMt(;aya-
z^3h$MCOtJ%LgEI}w=awqo>Y1HQ1#Dwtk-AXagN$z=^3~6dDyY<C2a0CwS4?@oe14v
z8tITWZHx3B&t1@gIFcGYaQ!MandE*XLK`RMl9N}vK3Vdv&mT3N`K$ih@9||up^#JE
z5MPflY7ion_L$AuWLv<r`$nhY{QX!*p~T$e`!A$P&eH5%8EMCI1wTj<udfpce-)ZW
ztmmnQ3Nd`{x84b<5ZqS`UjItgbq}HFn&~YXr)hY-!y|ysvRPkbh~;QbOm4S$VveDe
zBj_9J>u`9WkG1_jY<*==U0b(o0widV0Kwhe-ARBzaCdiimk``tf`;G(cXxMp0vmU?
zjqSH`&iUlt`-&=d6$O7-bIsACyGIZ5>BN>k`^9MducSeU#n<azF<ISg-Ym&B%enLO
zpUCO;R;G$xoJ@jl_&jeBMl+rr_CV<KSaVUStnemZQsl9PIBpP0jX1l<{WyPpxfPeI
z>50{~kV^V~fg%Fn*hGY$M>;Tb%?pR!n5w#aP=q@wPIMwjxlafUR`S_?r3)SBAw6b{
z`e~{z^E}sp_XV<l;M5qx)W3Pdr*bP<g|TVrpKHbxu`Q71R&)eYpiRBjQ7eS*0i@|6
z6(SI=H?gEXBq2%op4u7=TcaICtoO8U9M7enhifL2yzQ5Hhbq|jUi`sD?9s5{bR<@q
zmOzO8C7i*+rWcICRG`0}GHZPdD>W~%JZ~{iX29i0424xHdSCs{{fbj#M^hv9I;i1>
zTAVRyfT`wXXvMFMh)yGW;Q|ZwEZeo*irGwPn0*BB^w+4#_6(0`U-Rw3EB#R8lX+Gv
zlga#R=HoB<=4%?<<X!1VN_nSS2@$PIIn3FCZjpNN&$KcR4xnNZ3&3pJ^^w?@XRA|8
zTM?$=vhV%y34x<muY9%DAEi{;s$8r{FGia$^;fu;JPJ?>lu)R2BgkTzQ1>@i-aIlQ
z+BLYc96~MRl!Fa*z2E;EYW@K`mk6&5!Lr<LSNM3pYs+Yo<utR@3buzE9g=m<4-@UP
zN{v_F)D5&0<T{gAMD_@c&`LV+vjz-w-|#Vy3#eaT9o!n>gniN_k*d2)_H<8(S3OSl
zsl1MkB`NqkI~I<?I2x+?evU0nWcWBtyIq0&_HzSCUrjUboP4FQ$b73?h-#6>4b^Da
zT4YSNAV~Cyp<q(HuSkSInNctL*>9x(hQ*`X9)n3!Sw7e)gE1C#ZMiie(uP`VAN83u
zq1NkRlVV&46XS%OVtFMiN@$rwzI1(9p2$S%KatDt-}K=|ao~P2w5!6CDXmJu<_OnU
z!^jJ`MmEeLGqi(SG$i3PtO%=8i!?lYp>?hN3gNnPU2iX=ukUo+d%XO<GnU?TH=6l)
zdYUDmq&>zZ&aB0`hWXr5KU1`y<Ot^NZyf8*QTi5+V|AQWaAWE4hED|;n|gb95AXDJ
zEMK4upYOm{Q;k~h{!UwTH*Bj49CCN`^Ct`(KDXxw({Uv+B76xH@gL9F<<7T)4(H5K
zXh?6QHMT4bY!e!ZcyO3bhS;R9!#WSkyzaRpVX%k}qYsm4&GF67KuVS8$ZFv<Iv4$t
z;mM|zL|{5#%Wa;6&Z&S@&!;nc1rs?JoxF?1jxT(*7Meu1If%{=088>Snfjqob~-1I
zQYkn>Wwj?%d(?K$XJXKluHm%bxPzMIbcx&%2?3KnG-AOv>}a&SN}$#_4bj(fE;XjX
zV6O}Rgs492LNXEwt8^wi<~z!^@0UFnYymhNoC+2`7Rf5N!}-m?1&DwreTl*_SC4Vc
z>(o<jL9JD%_Z2)tJz#^?M2m5EPt~HM;T!gkHbOn+N7m0&n}Y{_jU6@}kk{zJ8d8be
zJK+h-<7@dx-!@4>xuyKYxLAQGATpKRynFd^Jg}W#>&S5T{MZn;?=cLeY`yRjs|h=`
z%gj|!7Spx2Pxgc1?kaIci+l2d%NpPHw+5LX(|fP3NzaHUo1Dy!m+QWaZ7Af58vwI_
z9s_9AiKv5b;DG5}EPAN#!f@WmFz->wyV;kOGc3cd-h1nuTJ_E5MT!lp1K%D+qF?Pm
z55p#+P!q`uqM{B`5}oolp^LzW>**8(=KrwWY0l2AeU#gEkIe{GW)c%vGVXtmQFdK-
z>9%z_L7!hpB*pXDS?69l;e2YecoL1M;Y^@2l%{LZWj_!bLZHe1t#7h@|H6nakwHWB
zBC`v2Bfa+OmALeG%J(2m&XNG($Ey9=W*RkLqob7XH^>7MI$yKH!{?j};q9>nu-_nJ
zDqk^~yJmbQ-yGJmO=f1LEvffw+{B}5BG~oqA9D_|&=ZG2PmJ8YxZ1CB6yl*Kd;5gM
z^b4&>qweSZn`aOkmOdVK)W>QgF1jSFS*+8iGm?ycB>%O2ks|GvFM@(fR{FFCoc>c8
zi37_OYks}ogsEhc6Z+2*_L#tpBlD1<N0y{k93}d5r$^wBi5akruN#|g_BJWs4^aQ|
zEepHdK``6&jRm_wri`R7;Rl*+9>KssLJI0(STo13u&Z^y)P8tlNxncbee4{2-LGV{
zNlcHwA0y~`99T2Hwk`7>#C7O)cmz~Z`~soX%SS2|^zNClNktMJya@Rn059SM>Vi1U
zZv@QEqj5Nm1-tc+=c}Ja^(gQ$N8uc*d|k2UTiEo>EWBGs-QO(4dZgBQw=FJ=4C4dL
z2pYz%I=x9-&x=uEQ>JxWz6`JlZZtX|CJ+hh&@{|YTx@P$?l7DE6z8pO@~-7kO?L4#
z56$N&>$YbbiQF$MXMGoJY;09m6<ZnPG%+K(1S|R?Kgehk>F6FR&Rz`M9Zb0QidS9A
zk@hUZ4Ex{vy&}}pyz;m9sojO%1D+kl5N!2*+&JApCK?3lr2n!_w>vMj@svtY5k%Ph
z%8w8gF3oyvVf=Dyu}DDTI#kZ{W#9Kj@WG2p(Twh}nevyS{!<ia-eu+1F366Y@4bq7
zmy)P_mA8q;GaVu|SnL4$bDI5r^Ve6PTxxl0L5vIG$ADYd9oXTRYwsU>0gA6*4i0~P
zq(U~7eWI@_{}QSmaj4R+;F&GKbomzQ?rdG6FB}tW##X|>!-FFPd2;$~A6QBb;|LG+
zDR_heQN$D#DjV1Tt?WDd(_J3HF*~8>w-LWTWP3{CW}_0rz`Yjz`qxK*xW4ngxps?N
z)Z_l_0z6yD1iQJP+{^r{1GPZUYQ?j4c^mAxgtu)QO_NOmkc?jH6bmT?1xGeN%eg@G
zPs<JFf#U0Pi3AkDrtP_oHN!wYGx0`rW&Z~^@ar0HYlVDKo~VFIvUfUW;~z_<=fa9r
z%A#+@B7$9JA7Fue_q=mNUGWV>r<(0UxsnqSs-`z29@)SpwiylbX)#@iaj9fKi8)ZB
zmtS4{Tn1Sdx-*Eai<w@BKJM={>1M`Y%;Wy8@>^sXvSL?a=#t2UyHH&042kZDHu#am
zMTN}zO-DPFkGH|10MlN@`$XcMXyptH2KZ;?vVg?+Q$*K%`=+KG=i{51Oe*`)-=?Od
z&Qr9FL8E(j4oQ^<<@zk=Br=#@I!EN0dpGF~`j+<61WcR;0gbMuaz~UmmDKJChjS(6
zrkbApblfl#{i~RM`<QkmzHHNr;c#}cqKWlFcSK$KczW%OP!{8aJrB+wqE%EZtYs)4
z#S8-?m5b;EZ=wtiFs8k*yT)dwwTHf89pF^n#&pNmbt{xNpBpZyICVY|Q$-TYe!*8-
zf9|mjg&6Tsfh?}TN;}X8tq;9+!}(U&TLr|@tfD!0+i}#&pXHPDZZA-|cUW-`8UmZk
ztG}a{EMdFqn3)Q}soCo8zT6%!!gkyclm{>rMVZL@O`wE_<q=Cd$U2l~ce8^&aVxav
zd4u(KUf+EmZ}d5%tT>W>jQX0?v^GE&Z;>|<F6{2kjgOB%yZdzbtHS>A)&^tx96V;m
z5lcx1bBX=u;29x`@4kmR3lLr`VCL@UxzzfCog(~A3radt`W*S}UV-hJ2_)<w{P^|q
z)z0f-=tNj*lx=#H(&ZL}Dm)_(AC&@k9PP07Z;E|VZF0rQRHsb|sKHifq5>L??(pnk
zpM?#K&|kiYiC3a-NxY1hd1Oj1z-Px`6KQB@+#eYu;D0j(ogZBB-XW90JzqM(mMhNx
zR3J#j#zX8V-3cUK@>L6)2>Qik&}s%pl!8{ODhB>(lH1I?^?3Unf(_zVA)n-l!9(!l
z0~N_0ylfz7u(GZO=J=qUn!d>e#v(SiKnZ5dZU~9MHfKQiev^3YcYTjDiVrP@$tGT}
ziE|L}c71+wfBk0Vj%AhP6>JRRKmqf#f_D*>rLT@gt{<R&GR43sZ6<7`B!Q82;}BVF
zY@vuIOnZf;ovMYWeu~v@j?CK~uDZKR&~M&J-<L6Uua)T=D}Qp=qtzg~6;qt8UN;cp
z5O*gUiMGlF1^eo(7%i+1aEFD_tW<HXEkPmYhWya)g5q7^YPs=?y@lBwM<Cq{p)~st
z-~^p9nPWo60ZHxPLniF&9D@2TPiV)0%YsQeb}@7;=XHyZ-X>Zl?h0cS`itUN{CPNL
z)ns8tuPykaJ<{$FoNMGH6?G$yMHHkhM9whP=WA$T{XJ5s#@lS_H#ti3ifq$OLf;HA
z9nQ?yaI;j&5})|IAM~CK4~F(kty+19QP!5Zur3sR+L<vqAsSygKf9KGzk}dR335_v
zv~u6c0f+lMMsB%4KG=x6NwYf@UGmKn$4<cU{N6n9U}%P5mHOWWWs?(CJ|~<HaK;4=
zc<aN6WDn)m_A`-*JbYplKVNk}y;|&QLp!i|XY>yIQO$RR@#X6AhA&+W<oQj00h3YV
z?by7Y2nl*m*^#QXcJzL}!Rq>EpET+wA-qwpVSFvs0%0N~OcycUBQWIat3K#H=o8pK
zkb;J2`ii)LA0^?=R@^ZD{G}*RQnEB`2`<5qlIX!T`4kd^YCczjiA?mt7oWx^fl5x?
z+4<Vf7*pm8+_nx(C{l{sLNq-U%(dvo&<ItLfey3<)3m$aNP9B&<rZq&X})aWD-vNi
zRH%PG4&hn{-TiZvLN+xX*nC-L`7w~&=KSi+s25EH{l0BMTW*{v<z$tbnubXU6lbv~
z-i$@~GIJE09L2LxIY7%o=!%y2Els{81%kh38}M18wihcZ7^#VWPCLf(=U$4-<|Om&
z1KNLu+CePd0aKWGo-oPE%XfRv%RcL<7Y4c+tK@(}KC(eT6Wd82eETuACCSx<TikKw
z4&@9FYcmY=Tua>}`4ELSjfX}!pOKMQ?9fLq>GwDlU{Ye8_N;VLVF{}R9X$8jPj=(5
ztfi6_uV_kEMQbx56<3WI;&g{xR>?n$u9wVMJY`x?&pE}1@vGr@cfrfe8c}FvDg(RL
z;!f|$F2KvR7vE1Jkw%5Ig?aF{xnyGL_m6yNGM9IU7l1D`W)|1(`*)3u**-YarM|#J
zb|*D*gZKtvLxWWM^aI8R$XOB|w&f8WAw^m%H_V!^dCm-G`S>mtNlA@Ggq)9&y4LFB
zN^~5s$V;Fbt0LH_QrP_|UkjE#m6u|r&Um1C`HeVe<xpj+Km%GfG@-)J)q$=~1~;Na
z&~MnAI5VYhcsGdcyyxznOSyvV?bu_#jc+>>+w&XQMPwkEQdCE6jS(YB`dvf0*UKf(
zTaBBpnmxIuj-H3_xdU=-S59!TXaD`kw<Tew*``6*>^Mxi?HjjRG0vTnf~FqRSgg2>
z&Ij);7s|`f!UJGYe%h?m!=M94WgGQ1Y0nNS^Pw!<HU~_`RwmkI$Pe3dnIOB$UJH1L
zUlUX;r{8|_4|~yjbMdd60h|ATibNs~l32Yg6erHHuJARWV%_|jA6P;au^5tHxuG^&
z)n%Mb=vQeM*~0l!oAE-T7CPfP{6Pl`;~I{rDJ%4{gC{0tw3L3e^-;+B_cPbaq6aE$
zbk;7d)-o<xA9yLMf3eF)UwIobpZi1N<dehsB95<@VIWHI^&z<i)5$FXkJIaChSg6d
zWo>k(InQ_RlKN5NV?mTiCsazG=t?+W$S|jI$#TBm9?>dXf7q0$BO=4~P>-6#)dM;h
z10k0WyJep@UxJ!7x@@oY%U&!Rwt6<XM4)lXtq<ZYUxx064*Zx5QZt1X72-;($h4Sr
zF{Awq1_CJqy>e~K4mHozjvJ1*j=Hv}m{U;SI>pbV&p|#!9nEjTbjaGI<#eZ*C|~AX
zKFMXbb#lkx-0Kvae>ggWD;O-0^SwXn<+;W_={0Io$V^j5JBaY)049zqR#5Oo?xti{
zuXo>dd^BuCu~eyuXAqQ@w2gP8?vMj;YA=0IOyDHWp=}U=Q%!s=`_*voeCv%iOgKb)
zgW}GQ1AbQN&ru3i08GWbG~v+P^Q^<5?{g~T;gpE8ofb#-Z~~|!Fce~CU||;`>f@OV
z6x<<<>%=`1mJA@@x0$0PY=L7(;^a}>HpKUSA6<`#PRP&P7m{(e{y8;(FFmk%yiweT
zUBhD0MMqj2=6yU2scjO13d<bDCpdB&8A5Jn9X6;yls*(7&W|XOh(^d#^lKK%wf8;J
zL`affuhG@1kUPwLktX^`&%*Y$4gJlS2^<UzOy|mq)>sMPfZ6Pi!fP7UWDP+b#QW2y
z*_EPeNW!&ZLG9UPO9+H`ZnpjIX`8D1-dxh@^-h)x?OzuhaQ0(S&>DRqyrZVnz9!%G
zGrt-LA+f2GPgi6O43UI;R^J-ic`rpDN(Shh^PesraoO1)JlIqk5hdIaZVt#>k`BD#
z)9Q3Mmj9zyUf=J3_nWf&Mn_evLJOgY-<A1z73^{?x_$c$%j^W#dHQj{1JWqZgmktb
zbMU(}HUB(XeGp>G5(zBUu~G|mI#;9avA~Xq>cXrfe@-4f?+n#zwyNV2*uumw5I3^@
zDI7dwoO^9LcTsaxVPjZNc}FDJUsZ#3pJx603W(Hc9%*qhX=Ju58@4Y+QTnY_X-P)U
zZ$>NAg;8e+odHM~SiZBz1Lh@yF)wvkPbUN2Mdk|mAe>~sZ0?Bte5=Z{K^osEv<H5A
z3H{B=7u*Gvx%OGB2IkRmEc#P*Bje^vlP^a*Z_^>4q)M<?9#wlcLUyC>+oV|>W)q3q
zE|TOLqvH0(;ic4<H<i}0lX*$n2G(uq`OCz^Q-o9_+w4{^jT8x>*&m*0uKb;KeQsZ|
zo%@_l(WMrlvz}@XNJ@z5sPPfM-f70O)p0wSS4mxdhoSa+diAYM+Ou%ne-))a&ujp-
z(Fww-U9c3P`nW7B)^ICunTlmxj`6wxVPF=doc`+<-M4@MczU3%a7O{^-W}nODTY8{
zU6D@jWS-<Rz40-d7aONQEjWogd9wesAAf#b2nGrr>SC`-2t1~P3(Ox+oSbpV6E@M!
z4dA2H!`&DU1&=8#2xx~m#&PyhKeU+U6Ljw=ZCxIY;o%E@|F(2-RArcJSiwp<SB)xn
zeYC6_xfdWIm)j#xM|dQ2INu*mt>>+lo2Gm5vT>o5YDHWr`HcA2)nhblKhMk>PunAj
zv98?Dda@uhhFtzNUqYyfSVX*^Zc#3(pC2AacBb+2g@-_hjSanc!49<e9g#c56O(E}
zXS)K@=B;B@pMdTUW5Rv%2pYP=s=SHmt#;AstD;Ot>8+(nARjg+YR^_)P*~t)@zOJG
z6&-wnxff^naKcJ7Nx-&NdaklGt%<?@hL1_RDLRMqn0sax#TWNj2QA=s?E5A554()w
zO@=5x8($|AZulIw*)?9Rr~KRdrKq5kxzA`R02kDyED9&B`@TQM`8l>VHlb04qDS%>
zoCp?9TKs^u8|9e)<?xzMM09n=ZQUC#y7@LYCn}wn4;AGO(M6{?Sx26!zw(F|LM=yr
zwlex`b8hpoQTp;>1OF5BzahzA)ymr!X18~uyDRtczLx>FPBiZ>W5Nz*?0yQB+hG$X
zjl-(Ip;f6+Hb;RB1ZzJFbeF_te{KaccLh69;tAr)qhm&A<bC>OJR6*-g1WTJEro3v
z?gB~QjNnbZW=3M+dwK84mmg0*(qO5D6n*pdKEC)05_UM_&)1aoX<8O8$U0k2(oD>T
z@NiroOzCiQNh-+AUxkYgJ;=O{s>Kuw&O1sohY5^%g1^-3l~7gvplZGP8V_Rc8%b)F
z>w6MpuIGip#kpItg!YTTH@z3qA=$PJ+~qO6939DhEvGkL_xJFl$KTUdNPe?00+$0z
z5+{!h3|vMX%x^TrdzT1<XS7`dXro+G%PprmdM}|MR_#=-;JZf+)34uxe>M>*<<8*y
zY2z?bBRHXR4O$y-4L|;REIhkEz^SO(z8eeGY}TZ@h#Kqs94h^Md>OUkEh>8mf7g|g
z|D_mAb+a{h#C49888I2u%@SPt?sN&>2`W>eKv`1Kaz=b+$_^sn{e*3L`+Ad?c3!d^
zsZj7G;#$rAyC7aoN(JUgB|Ff;>Xgc_rmj(`lqs!3&UXI5pQr>HBH7@T23FMqq&sB)
znpCJ9%M6PKJ=<*dm%2W)jncmNE+0SiNUu$Prg~BIYGhsAp4x1`zq{l>s=GmBS!{6R
z^fmvj4$Yw0abxr-13>hrXgMiheTDX)+QbL&q~a2S@=?$xkIGx2achAB^s7{v4PGJh
z>V)I7XiJ2XgdG=>evaK`m+b~s6+G>US6^T5pRvFiXu?GGel8%Y(Hy~-87HdP_dkk`
z4hK#@4ebVNEfDZ|!OXX#e;QmX^Omo0g&ra-fKdF8>tYZQE-x>uZx&l<YSw(+<Gx)a
zqE(-FspyUW{=WI*;$o3bnO=wBc$UDIMxf|)dXWknH{QWS+mgXJ@<Hi9w-^q5C+H5z
zA8Y9uN%TGWzOoQ`{Ga>Ce>Ljs3BDG7n|mUj#r$t1hlDN1hAtv=op2R3VEOWF!p^dr
z1Or_Ol@O|8^Rv(=Vm3BJw!MKK0b|C7Zv_5n*asoYbH6SZ(uVBkuB+8Ek*Fvt;Xj&F
z_RtV@3_TpMsmeYK5ll)4CnOxtuFVcad!Z0^^)ulEAb2`d>4p}wX;`;qH;Q|x*y7o)
z%6CN<zF0yEW*V78>89p9HOF<^rm>L0S{(TybRds50e24Atz~1(Eq$F4_HHFeZ>Q$N
zF+Y6tR$|O%((f^ht2|@wH#ssDu`|n(xMo|y*J6Pw4B<Fm;H}c^=A?s_qW?f)-hgXD
zzZQq{41THB(r}T>czb87^~-=)lT@}l=%3+DORS!O;=x)a(;;v!c;}583*O#MB-Y{=
zAe(jB@l4C9&o)?aseeIt*uEB8T_@6X?(792km?I+I65Z$51L6>U(@>4xGwjY&%y5;
ztX3PScik0Fr=NgzI-6Ttaj~&q5wRHUxxan;wgE8u>p3J4c>g=`uY>ky?+DgD!)@#S
zi~nV3p5j28>nLHQ%^YGH#EjXm0|+z{>MjL-dHzY?AKt>?*h>-tl-cZ%m1w&Fxr~_m
z3~Bf;B5T7gPIe;q{>cKLWj~dq{LHd`pA45fJeQPMzAXoYD6xl!e=E0AJSFkIj>?Wy
zvt#Bkg?|Coy)4+S^$q&BZSYzIA_)gwrU`BlaC^Qd2!=bfD?QQb!>CI;hX{G@@;^km
z^=Ce!G=0;!oQA8l=*MuX))O*au;(Mf=cn|>CcxSa06nWDfj;!`c?)*fvRJ?B=C^U7
z?g5i}RA`&X3MR|v44t5T7AkqhZjz~K%Rs`y4r>ZKqTsYV%U-+xrn!P2$oLuK;`qhL
zN*bH<6$9QvkD6YxNE{QwuO;*je;RH04t=xdl#oYD`Cz5vm<TmH!@yi2<V;jwc%6e^
zOkmuq{z2DcUlAE;oa{Rf25L`m*dpD$q>6@&4}_50GrR&e;!RT~Ju+`Y(hIL%m2U=|
z=@w1NPYQAHR>N<d>u@eKdrAJmXA_B`Xvg=n<JVnkIY@2@1i{j?INAO+AG-Lo%XtU<
zwrKVK<pnTIEylA{3`Q?Ckl3qEMTyqw*Lz%HV_n6b@>XfZd=w|p^H~7q1xI`#l^2^_
z%qH9<AA;&s@2~!XDmxt0SrhE*xiUDX&lfXP>j&W-t<I3106$WWVKwK%p%M&iKchyM
zwfbi*5MJ^9dz%6;a#|mi?~hCi=qjaUmyjtJ(7~7g+y4kpgvi}_-ybdd_R>dS7|8w2
zsZJ@BP3E_TcE*uOaMO%0FXM7p&JIQ}#*PIFmkR?&u*nc-5KeSsi*&NH`1SZ2{{Sq(
zZwv0y2Xza<6VVkKjBB^<FEa;H7+8P4mgUxSo<#sAH>B$<1~e10HQYwm+U|IG6bK|V
z`L04SrbM}3HoX1ZK{}g|^ng6D&{M`v*#7wdU^NS6{ea}g;H_Ee-r?vRc1S+#&*u#J
zYIl~h5CqGJ5i+y!+>XHWdSJWtGN;wxKvTJ>xzTnWNn6Vy=3R4t(iVDebShxr#Wsj6
zPQ$^$RKbJ9y^T+|fSjJK$qD-XK{T-9hFO6~g%;W#kmF-+N)nW6evq9Ip)C67pVlq&
z=zW#>8~lKu?zVC0{=><P`DlA7I29Yb0&i=Q%RtY3br#W|iM;M5n7Y%(_`vJFU@%qh
z+op|Q)%i~NC?yV;8yD!QUqY3e=OsB#YglVg>#5NgXAFC-fB#bAqla_-7<8#u0=~N+
zVoT+FFt{oYXzhp@Q+HpA6zPI`BRCjy@(2;Ri*xlJrMzkH3om-*tyZI1?||Q68}a$)
z%&uE!Y2SGqD(^=D8->xLB2}-Y2OcT7RNvbz5Qj(YBIF<<rScYDWVN|$Ro5>B{=5<&
z9s`vcynK6+OEtmV#eglB-634_ux=oK#%R7&#CVpivnn<4Kj6}71OS%~wC`Zx9hh2t
zK2DfCXN?qQ1d}}%$S<gHI*D@N0lL`eF)V0!Y-Dx)cjlvky3m!jsz*G}e|Pk#P!Gf#
ztYd@?r@yy#Wd7n$&tGD2|7Ll<@&YVrZV@pI5|V$VjDdZm)3$CBz3O|bw)vm#Q=3ow
zseoB7EE{3ksykrQ|2E)Bi}GK>)(>xD@1t@CBFa<APqT0RTNVDpmEx#l7TfR1iei_G
z0eIOUA>JIf?L$yF=E4txpATAko@Gxfdfi(_AQmnI0ZI?qJr<3FUyh$65jtba(uVd~
zFBiuM&+3L+n9bwyI&yTb`a@&Rw+;<U+uXmuwpvG}*Qg2WsK?=enbD!jHIHL5MmHzS
z7WmkMRrWO@a-%OACMxD!;)XUU@fb1E)VR^Zn;ucAK-7duDOjl<Vt}jUeWd{8Ts;me
z?<Vd6S`tQOE5-8@b8E--qSDR_fRI3_*vP>YRt})%M>8%#HTN2<Syr9{T3%EaOR1P7
zm4TzVqXql#oQMivR4*Q36k~|m>3c4V9sO?SIWA&^ScZ&vg$R?1Zx|4-L?E_8-+omt
zfkSDnyOiib+`ilhShUnbBTEATmfnKW0OChD4sxjus6tZ$uAWxNA5AvKLOrEbqBfu}
zwrA0RTDd4zk-ydV>Vxf!k^6Dw{7p!*I_}}vmX`VQo>qD*UZl1L_}8U&klpNtV3F$g
zz6-wSA$vYMe16&@jPkH|{hg--F3|(ZFL_5l83dtov1n)C$@s+;;|;4^&7QRRocZBT
ztoadb)iZbJ71n9vM9-~jXx~iHNiSjrc)IpKS=dsmbbl&rD;zI0tjy<Q)NSj!kkVvE
zAH$p9dO2$G!*0L8t9QK2t)xo&gYtoi6XPq|TVab#uypf!px4ud2~Gi0VTRCtov-`q
zQZ3Yii6w5(mt>Ke8#VdCUi+hFOuQ5_PYZ0=LU$JQPY<KPRSGPUPm_@M|J9!Z9Mc;Z
z&a(&Ti=z}rr7x+8y=kdY%ryPKg%tilN|S-F;y~1laIc~No7QYptADm(DcxcYP_qR8
z$R78zq(Q{Z-i+~0t5}!2f~V>ZyQJ6|(N2Ta&Fko7^6%I<7%6ORN8msS`e+9WT|%6s
zGvxgOGXOi=xl#x){x*ROv^Rv5&cj)VmsO2&>_nh43_CL>-F{R$x4b;vovNvp8I)Aa
zjmX1{SL9Ypy?bl0C@jBCXgN=KJMB;hZRZiHwO;FkhRGw{GKVYJ;#Uusy3D6S<E(23
zAZ1>V9ynON<A<Z7%SN#d`DK)+P)VGF5OllNV|{I9S#rh53gWA$u>thYR%>&&L_+<u
z(P$0I4;Ecy&vO8G_DSEin|H^d?s~%Z$)f5DNRTageXVNjPMr`GGfW!yg-sVQaI%c{
zD`^j$3e_vbf|Tc(cBWM#9gtCDBr^FOnG0H^cYAB4%e#|brEF+yYBIpzbjf897whqx
z51XM74TXh56+nqdi3iWHK~A2I;HY@lnjpeF(YOmrBQcUIw+5XsW>ecuN^8ug4Zazv
zWen>D0v`SFZojzBHtwFM+sHmhUkQJorntX{ssa`&<0GHWy>a@c-FK9BZm`XI&}<$t
zaeoupOMyxZP78U=b$o;4AqOq!cPbFrm28_aI+)h+4D$v@c0L;o&nv1L9X`f-;(0Zv
zP8EW=tprjhR)O5C2_fkf3rSfqtej=j62TdBzm{-7No2^s0vEf=1!ix%8+Z!t<p~Y}
z!Ln4PeqAej{~u<Q?R$qQwD?c6LSNomEGG*QS)wONPTn?$0<Q5|xdUkSe*~fd{@APw
zsCwUI_lrxO3ZpU!UVU>SwmvWY2cdsdOb&o?kPs1p!$yS~zZi!DlLdMy*BHu>H<QP|
z&p2))q0x)G#Qf*mfoeOcHyn}9usN7VFqL%mQfCAZo;oajGT!l63D4y=(=Fs?ICwS5
z%i*e(h4mYR#mjnDi}dmp2~|Nag<N0WzFc$;$syu^X*BW6Otv22oU;ODZPIg@)z+rV
zSNlIZm(Q3PY(+mqfpet9=wskIQ5bOPmfVg`eIo78(ZXz`gRGMw7|O-!x?A@Xx~*T2
zYBfT2URr`3l1||N+@DRY*}|NTlSN2(+e=Q7Zt&e{4Bae0BDXmV^_l9xO^)fADS=3P
zF(Bw>tG;x)o;QI&hSr~(^JO?{g=KU(U}8XdMRh=SV1hu?A~TyNS>`s&Uxx1uE2&im
z97Y#Ld14Ysw^e#i%A{s1a7#gp%jT;I=;i)4y=1e^4RQy!r32k}@!<9d?^MwIQ?#kN
zT97A(s~pK}VD=UmQDZ9DD9mQXUpb46+{JkxF$uo-S!VtG^C?4N`cp}-@+%9uEW@8(
zm7lGR9GdKdv0@IO-lPs^UcPvwgOg)MFm|@15-4D=<-=ph3YD1M<M99vt*7Npi+Yzs
z>v!>;R-hkpz~i)Dp^GkN5JbV}3DeEB7IR!-#ZgE|xKqAepG$XR86s2fjemD~rk|6C
zH1K3(H5M(UQ*BC?-Ql)Jbf2Z9qM8^paa&y+MI~FHz5To=ch$MLj$I|8Zs&7{q86;x
zAx56Bk&27~SK?i!LHAWI31xGEO^g>FC;H6aIuXQ;Hl<Z<2L%jsnE35KwzjqN<GRKF
zA2K!B!S9#q6E52-Stl~7FBM73)Sr4SB0`(<Ed1zy3us3kI+Eb8mm;<!uormAhFj6^
zTFZYa)KMmfAfBZZ|HFowO91E+gCOv$Ke7GGt;eiWZ1IOzaQJAU+3AuoLkbScH!ZJK
z@b{Z9&58qKSfMb8Fo>qlcFp}JUt&px!Ty&4uXjSM?kWuI&-Gcl(M!>pk72({rTNds
zDsugsO|7>fZhsPK%1ZQOFCX)1X7qNJ1#t1k$ISp5L7ul_`b%fsECrXUg{CgdH@~u%
z*$3R}N^(=BZBAZ#*ztLMkdNw5yQ85)#Gp`Gm=;Z{kJV+we=4z?>P^=RU>lr~-Kxc&
zYM$pZE<Ko<n1B|S{6IuJdgO}-xG6eY64K$u+~Tcq38jg96NCB)wgA0V_as_a&w#pZ
z9cI@7#s#*!iP9u&YV$<W=@-X>q+zcoas2>9pihJv?NRNp)s2G_9aogNmoi}Ix&?6a
zvk$AW@t#2_ozYU6r=|hpCz^wUVcm_=L)OpPEi6XktWRFxv}tdhl?G97<5$Be3_Xx>
zBBHb3BrIVBF}S6o&4*p%)vA|g6Fwop;fq+5TWKDq8e@DR%oFf$f>H7(&Q~uyK26ax
z8*r<}=5^kT+-D+-{(gJr|1E>Nz}yR&podyQD6Ubp-dWOJns5YOmqKRoVph~h$|){F
zjhNFmYQN%}G1Ya#1cZ}KNjwcb@k-qZ*Y0hLL%~l^qQ{)ciyBBbmtmG7QcZh_@5q4r
z^)^m_3yi+brR@uqxH_VJ+Tfku6sg$#Q-xkfKcGp2e{*{u;tTaFm&fnz>Az+=s*-xC
z!GiLbQ0WzJ-h=r`*Xh0NaSo^bE%y^CzEV&++nfxI@AaCHwX}aCCF*GEM2_w)HDx>q
zhK*}~5Y};J8{Em{D9AYk*Kn*m=bl&*<E-nCIvi?~-+QA3{kc!|Kw!H4p}(-;BLBi@
za6xaaJvhNNyt&*}7;4ysYqj}wx$GBn9Ue8~!FJrO`9iAzO=Q@%j-o2535BBSu&hRy
zfEu(n6e>~5kCTfZw|8lEmOpPB>i*>AJgw&Pl_;0dlYnR10VXm(U*KC=lECfh2j9v{
z(;mxO^QCWz0w$O<Ep8gzUc30?nF1jsA45<X^L|*K-;A9+Hh?h98FsbxR*LuMb-~ZA
zzr)ga|AH*BF{s3>+Jc(aj&qf0%-JT<JC8wn4fRju6sB6^oYo;jeE(F4zdi(6pni|E
zRqYIZ5d4x7tW&w>^hlL;(Lo^iZ>m|GL17~Oa0ucmX6~v2Lnz>ja5kx~z(QXL$MS{^
z<*2HXfGsh>jlHhHj;-qYF3SEKMTvBhC0;SXa~%}Hgs`<d#%I~f==N*Mz6LjHvoDe(
z8V|c~(BA2uBLWm@+?NO<AIOtBP2==}Up1qi?buFfhAe&<A$zllMD+{s%Qbe~?=*(K
z;&4o#MHT-u0d@g3#Mtt(6g4sT@S)eqF}{}N+iU|zyLi{PHy0ft78&*wc9G0nmY;>_
zHJe@+P82#qt%axRGmGZ7vN&cI*M)p`#}%@EH6VpvUZL9&yeD1IO}Wh=-WncpF0tpz
z;Dn3Y+5xDmWeN3?*i(^cS>l7Cs-^2F3~BbG9mSfG2b}XMaU=N9K)^EgQKw;@MF20q
zc$T1-e@9HVYIB9#-fb!-v6{S6Lm*ALNARfl_D@VHtQB^vz<7%AUw!rfmE~~0Uu?aV
zkkE77nsgiNJ*FX@A>H(<UQWZv(kJo<I6Vo^@8pabEsOKenA-+>aKPkFx7G0@eD=Hy
zo}of0*Wp}w&TgyVW5`;sq=l<4ug?nxSeUoMS?&m7`L1cNjnm(M-I-<?59Jo&|3_!v
z>lnGKKU$C@+V7+TfmOO}l0oI@f*3?#H^0_l_cuaz5d1ihG7&|`I|3*(F0({PI6+x?
zh8cB-`2<r2l9@$_0N?;Q16-8@mfDy*_1>u+M1s;iPdrzs4dlol66p1}^gm8;y)))s
z)(gIQiq`lL_8f9|j<NDnf$ozhVm8CSl#a=$KX_Rbz{|q=3rjlt_>&sf#;4m&=71!k
zSinMTf?LN-u|pAV7jAd5lzt0)g0f6hD^yDo8}MY$enU-fsu4<da=jcT`;He!c&BJ7
zs8GTUv`Q`}pfTabD8-pLmv9a3>*Loc_beFD`Bo`(x&92bU`#jC2ZF!gC3kx;{kr}3
zgZX?VvJk0k3iD`7soUbs9a^JIL_a8D&w0EnRQtCpI&-6&nh2WAh%bDk?yEC-c%1PS
zw0`M6*^iJ2wQ?uYxtuKo6Bbh#vaPaIf=FJDk}VLRrq*02$AU<qnOptj%B~7KJm0`V
z<lo>hroTKyW@o)17$7Dai61LcI408SM=qN17&7Dcr(b;OvEi@ufe1Z85>f_!Xg$ua
zo?<ag^BCCqZt?<pSmt{(8L_SAb<3Nh!?G3Th8rlp?gS3;_vu1wh?pG@tOJi1AhLq`
zy|8TosIJ%qkaKAQ<Llr;Cf&|{ui*IjoR&H4dtP<($OLp6tPO(q5m9`%2ayf$Df{#-
z81;uRZ=AQy>II!P{UqeCq2hi3HU!E(<~~DmO}}+GA(V#E`Lj|t4fPfv5JuGW7n*3k
zQ4F|@2r9D_kRfH9?^TF90o&sVchpgLG%>Fy?VH<|oJ~zjXv5x$Hna`nunIaA_p|37
z`v|=Tu^^;9IHRqfmbX5~fN`MVQO>4{pA=OunIWG|cZ<s!<S>p)H7WLH6PSvtM!(KE
zihsVi9#zP@r#)<bsNz$oZUWqc1LK7Lb!RXRcpghIKe`&;NpkgXxEPtQ%&a8irwH+0
zb^nS)y7yo+;qJRQ^Tb+@j~2Sy58lA(qSkB2-yspWn#`?bU({Td8nm~XtTu`8j?wcJ
zfBQtia=_j9y8f0EOqeop&Y9t?0{QsdRf9D_Of2Qj)Tvo}#_zrUl}1eQpN4M`8897~
zgrMqB!Cb!j7mL0+t;^s^(#cPZqAo-E-wkA}7g{O?8$3iCbr9F|WDJ#yhw%>g5sm&!
z)nPbQNRVgsYD}B_2&M4J3Oglsx37EoiIRM??7jE+y?!2k!H2rXGqT021qJOk4?I62
zr6MyHm14@9j*kPb4BQ@hfSlirCZpL-F^ACs&}IF41F}ZX^O+g$RBAj%Fpc$Nc)&~t
z;r?VZ4pp5A7Q|lxuoI5lrS_fcq3z}Zw09ejzW_0~omRQ`89{#XlH}H{4Fnu;ghq$D
zP#R+&$!!^8kOB{0od6@rY8Cot1B3P*S5=#&&@9^1dM>^`17_p}SH)^^jRth}XME(e
zsY4fMzpDryT%OTX=zRLFBurI5w0<-t<2Ge!=rJ9Ymv*ZWn<CZI2RAcdlPDb;E_C~3
z+`EYD&IX-@zZ<HwN9?ZIW4cd$Et{}m$fFtNoV@l>d4z(EGp_1i?3{Vsq{s0RlffU-
zg<cpMZmQkHNQ;`eQa_ujx86F`d*QTqF_7(5wKu7d!R_Y%s9f(%Q){hAl(}5j8{57n
z-{wexK|T1snw03M`s|^sMg4nX<X8)`{=~WjkN6nwtjO6o0rPPRVUgAq-jmBQ#xe?i
zuA6uD{k_i>vh`|v3f-`P=TjivlgruShmM<*==m|+pZfmaxb3gWc?bJ%+%oc#kf<6U
zbLk8BNBb6~a=*iTLNL6dmsWCA*q-Jq_)=-LyF}l(JY=uEpbgpAXCY+cTqr=}M1w|;
zM+QX%<;U2jZ$QS}ORWIijT7i-^5ww0ow9TtitRu^K|~6WZNlVywj@p=d!uK~=QGzf
z4T3L)_p;pyGe9d@UWjA)hyz#~kA%=9j9*;M!RV#p%QGp95KzkECq?3TU0sR;7S;@a
zh0>yo1uZ%x$}&d(@m|MyC*E+PtCef_ffF%GrM06Hr~--TDmg7jKfAlP<QB!D?T`sv
zqLKA1F|)2RiojdiOrM!Ag%8?8;fxcFnEZKh&|HxL5^pckK*$X0XPd*Wo=9JdSBUO;
zJ^T|~Hq<ogUu@v^9~C@BX#RUEyZ*n#i);I`0ZnB{2`{Aks}G9^<UKM5Y%K~kiag;D
zeqoQVm%`(h!k1OXjZ7|BN52v)F=orO7xd|dVTZy#-zd%|$o!n|E^>4Mt@wef=l~<7
z2w)lZ|Cf>SQlHS>ofxMW1OT@GMUbC!t<t?sYA+ErAxNrIo*x+;^I(p8Hha9A`-ODx
zS!JXM?7v$I1)bks1y@xm)L&d5FCj`JBJ@{NK!NXnA?O)^;=$(AwFiE?a=NregaT3l
zQAuwB)D^RJeT{VgXinU-<D^YpfpfZ28!&K7V(CzH?_n(3tEpUZ8P(-0C>i}%zjI!N
zM$Ec8AfDJy1|!G@^)3T<V^s#UpB1Mg-d*0YW>ydEsroOaud@>M-sP~O(ff|oZMe7s
z-4r?7LLIrp-RVE<dCaG}I9H-O@61-(AYoS}{7JdvcaK!Jwe7X-iF;RKSwbdvbklZl
zE;=S`RLV6h$IauTzNqD^7*8iFH(03<R$8rT1fMY5M!lbV6qhJ{2B#4pfiy>|5_{m>
z#B{{y8gtiYVKHfc?OP0MbZ1H0BR5l{QeV-`1w59CDb*x)Kq1I~=Nd6=Z#7lo0$i<l
zzyOyAo#AvCO%KoVu9r5o-){R8xGJk7S|q{YYW!W+$G_-c9v_I+B++EN6L3z>r~dp?
zBtjD3lEVlRK^dfXZ(#2YQqgOal5>J>Md5Sb=fO#YyXV1u9uECumi_yrYzSNlua<KM
zmZST+X#9r;_GyPYW1DcMB0*Xp1<c!(l9KG1HSj^<Nt0U}9}MGABLdaQ!QaG0VKo#4
zXWdb8PyCi49=XVk2^b+qUjVzX{Sd2p{R;xZcWhgo2E#4KPfYK&UeYkwy67QbYEx}#
za(MHV@k^$ugXYb9i-WD0bF|NI0WifR$jTlI@_zQwH?}744=8_*%Zfv`L(wDer@CzI
z9dzbt%0B6NYKBY`=67j|25Y6%eqhjQkY5P_zsrtvc6F}c#fiRDdKr-elL%R_R=pmX
zXWRtW{Oz+GJmRzbhg^Rk6E^04uKJk;$TI(fcGQuD&eZ=s#ILsWx&Oai3$ZDv?pYW3
z0zRIb0Mod7EPo}>eeF8<g>iduesoXH-dp%xs}JpL!kh$Oj-im401agPiy(E|hjuSw
zATx%Vkebk3NM7)V_h&FOmk<i8p5Op{M3(!bcF|GEO3=^vVxT%t06!%DCFo9h<+fd6
zG*#&2w|R>0meSbb5zvxOw-A;9;(tbMu;!V_-6b2F9kgC@*d-GsV{!5f9e-rfX$}nA
zVsJP;=5ap0MmHS`+Z8(3Cd7(Sp;O&bJfvW^o<(i*yzA{0fmXOVqJ3&ktbMp0)4~_>
z#rLyb4E^W_31BOnIOnmND-YUbNXwO5*EN6qs3nU!f7d{$aaMi5y7ICb-GTBF<u{?%
zIhM`^ZdCcDDwf&2o+GouxzG25sYJE-8`K*dFF7_E$%i|$iEN|2+swsQ3&h*AwU~hz
zA}7Ab>!T=xf}V4ix)Ob0ZAL<=m`US}A2X}=m&~n>F>fGIB4#TlFY3?4Pc(r*L0}hC
z2eL2_=%|oq9%p?rIy?%cm0~lra&f1p>20_0*5~nNu8}a!{<Bv`@EYG8vBy#Nsw_{k
zwapP%(ZoWL%F#SEz$1;d_k<uD39tL}3Y~3Kq2oodnv579WeXnV*=5KJJ)BGyh8Ct&
z=yne~;j8-$q}_amMLID`&YgHFV*AHR3mlYiC{f}V(hdBx$XKC3uJ}6I`EJdaI*(a&
z*7X0<rTtg^h3_Bi>6fYk@s$3i0Mq%99zUsS$2j(%?1&0jJh?EiNB{J^L>H0i`47hF
zG!x%=n6_WlV48L5=QzwAk^+<=!%foI*xJZPtvW@HK(F!lXemx?+kQV5ps<3^QBi-m
z|8SqICt_t~-ROytOo{`Er*=_w!uE@cd!A%-+5IM_CdR(n8{d1Dh$eh(CuD~WOwKz)
zc4oh%a*K!r8~GH-WsXZKm5eb>Vi^rYiM(c1gEMaP3Nez1?gn!5sbn)|!FdLOY+h7e
z*C#7@L5NuMI1samjK0VNlEs(3TkC9HN=nLAu92Yx`4VS~GB5=lT?G1usdV~~q-&by
z!E`?8=qxHaldo%hS@t{9LP>uA^YHc+HkirSRK5PT-g@r+YY|`b*~U0k$V6M=*{G`w
znEdKQtMt)_z`O04=R!`+$ZC7TI7YCVBR<CMp6GedeyVDUTB{&Z_d*gJbLnU6`_fmY
z6`6N(<W9fvx3SPTJ)#Kue+grk#Q(Ao{@h^`9}}MVAaM1A%9!<^NBh@D6SV)YyYioQ
z!g^DwFh?**fDm`ee2T6f=X=91i>xr4P)A_4F@mft=L=P2@%cXt!2Kwk@oIB;g6jVv
zHI%4#7o!}+A3q5Bn$BAA0)N8H4g;)yVA@QkO$0qbc^i<E^c-A*f44nWez%VB@w&aR
z%>l8oST}kyn)WhBTqHEI;w!=Z0+vjx85xy%T%9S5=5d#6H-D?3O&mCGVgcd=aXb14
zO^)WvLsl@{`284*sH#+(r7hUa<%xn2{krA`l*7zy3Ck-}*<Xu=V@#R9=)ppQPmK8#
znxtK%gAvI_B=?McC^a=Z=hnow8qa>W^*q$!ZEPNwI1`ciu_W;vq(*A*vG<|zxgKLw
ztyDS?V8_YSRXHTgQ+18aLC4Hsd}8By3rwSJE)J<IE<Rhg4{D&9uarkEyA!Rr+4$DD
zaL_0~N&Cpn3GPC!zt1r1>-)uByB}`Tz`)9=9Wf2W@F?b6V!C^H{ZO{(5kOv>bVLH@
z;`p!J*>?Ba#k)@}xY9P)or&20O>RmiZfE_=3jo&&3#gu6D3kjiL^nFb|8@jo{SK@z
zo_Tv!7Nn;piI;%jSLhszu0!RG0r$5Rj96eX;*xAfVmQVT=JM*Q%cdBQlkoRo&i-^I
zV!f9s1o0fwO!*WGR`MCW(2tOpOSWEpd-J3rQ0#f0?-^9=%}lPN=JO3za5=jIL$9{u
z4{JL4vicLy=C`&^W4?fF-76)Ej-=|2zj?9*>nwmJJV-s(O;7D|d1W??@a{=WNY}pn
z<;$1Y_&8|)hmc!c^ngQ7>mPW?@!|V4@F~YIKiqO(YRLXFBJzHRXg{Q12G>Cp1Mh5p
zOshz>4Hp5Ewzq}CP6pU`T=tpKYPl4iD!1-2C$s3OAI_;zE>lGPbL!_xqdU(B_fEr2
z7zcWAtJ~RXP+ihH<YS>hNbfO#$3zn-qxx8u*1sRFbwR_j2)F}F1+g#=XqE2~B9a+J
zYtGV3z=MfnrIeZxd9~TRFTFN4vyCJOALs`k3-j|0Ku=<+TxRXKQ?htCvM$=<l+F$=
z;vw0*p?Q4`cI%s+FVJx3h?{{Q=bK%RzQvxSdf%vcOwG%I9y=hqoq<&!8r?yp*h4n)
zbDtv9nLkiSC8F*nRZP_u&s3S+5x(4=rE@8f3J&xA>d&i*KtCopF&}sj^)^U&HeUYt
z*^KM1SS(jH^g7E({$<C_%yzv!YB-5*&kcLfPjsV(1h8dIyycMj1m}$p<C?h^H2ylX
zuPb?xL1e^k&JXk2!8BLl-$QDDUPo^jWIg^!1+?5x5^kO+p9<e*BF$g?n+FjxhnHfv
zrcnfKC&E0aIar52-6K7yiCE;`|G5A0hm!|jvj1u-Y*r>uq!S_V`uiTsCR*D99+IjW
z3H=j3ky5Salfs-Xw|5jSrL;?lx^xM0@F?9MrTX*TWBWe><_UMN2Vz?b{0SPMl7on-
zRf3iJkJk7$mTJtnv*W)}L?k=Tan6_c>R>8phm~d?>+5=izdzkGZgfC*P!*pw%&X?P
zG6JYjt*Gavwa7+}rTDmc<nu1i>*MH+gDzmM{9FIifbuHTbD`nEX8unZ#1CT<2?bF}
z4^+KWom1mskf`F3zVY-ZyN*ZIc?ki*@l3vuuq`TwYAh>h9=q<|3#n%uN<!N#^)^NB
z-Pk2wGH2f56o}MVM`@|FZ(h-_@MWFpf*bRtk|ae)3eybe^e{Rp?ALwcIv0cxO~F9(
z#H7Sdy7+82D!tXu{1dfw2ny!2T&Uh_2|kIiv#R^smSxRss=e{df!+)By8Cfkx=tgr
z@B~C6qt*MW^UIq=>OGn*j|)*+N|{vhei7+odL;boi0LOBoP+aQW6K~>X;}xC<XDh7
zL{V_!5&U!<G;@yZhQ1}Z>_*sbHz`_;vhR1r3$^YC5g-)s3^eEoHs}pSD^inlu$0if
zi#n!)aU}aQUl#RCloJh#7j++Z5tDEHcthZ!{y4%Sw_vKUqUmBr?LqbHX8dM+J3-#x
z{Z!TvVA>qmJo5u0#ga6<*m`9?2&!|;CL7LS${{1(7&k@xD2(Xd*9mg`r$`XPT=GKg
zwGC_zN#>N(EzBU*JtMW0*W6@>k`+QgFE5KF=++X5pWgnBOTDR~q=UXl@q1=pq;Xmy
zK!}BE&8M{aFBM(zRajSlkOZLe3xv{zSy;hegppcA#hrYhPYa=xN#AEzAfD8SD6vzW
zPquf+{KgdS-F0RZ$isz&QmR1_I}gBV?y4x784IftY?_3cz?M|c^aDFnZ|^r#a?hcP
zVEi<DtU>$XZ*T)KnVtJ+xn0q@Ch838{3y~6vRWgfqo4LinFB+_&lwdngp$sz%W-B*
zFTn+}h~4{D2-VFcnWOFBXGG8M)<e!^q7H#PKNPjf1WuoG-)!!DFF60zZ(r^5U}dwh
zW+_|T_TSjyDL#%dSY5FkH`;@s4rhyV*_9J0Wtg4!<r|xueIseik%0f7yvQFp>wTaw
zp(yc%FsXE|j93Jhhn?tHILF258sAGn<O>BpHL^L;zuH~kmCV?A?aE$qWoTB_F7>xf
z57@Vx%BDm_PDMKRML~`;Yz-!tmIKT*iaD^2V6E%PAf%dk4q&NgVL)E%;-gl7!9c-4
zY2qLxJY5T&MHSOoFjhf`>|j}Kwq1ECJCAb4is&-$Xev?!S(QmiiApGrlk8-)R%GP5
z!+yGWqS^XR*W=-C^j$iL&g_i;jaXw?0{!{X4}WT<`pEgGo<V!`evOS|91xsbLAfbQ
z7>%x_uAb~_+Gg||>ohK(>&(gmB2cX7zC%(S*QzPLfHvui7+D-Oj#sg5sjN^4`Fl%}
ze%aqo8mu<E;Q1>-PnFau$4S-*7?6NSK~^wI6Nr}y1Pa4~*~6lr)X9Zplj^!6_qmKp
zBKHdgeD6-yVy4LYLkIFHZI-{PmL(;DcKgReVp33u8r06R;9`&@il-?kM)Cw-ghF2Y
z3tO-(=S%&8-pbw!SXCVai$bVBH_XB9y-)Q0?6u{Ug(iyYQ)$L=!EwuZYn<-cR1K_Y
zFpC&Jc&e0Vy+a#kyhkEZQe4p4zB&b*adjcM^OCd2P$nU}Wwj1kgVUqlMwow3I{(u%
z$<8s&v)GSOymMgz%lH3L_7zZ3Zr%G*0s_({sfd(>BAo)#f=Eh-gmia`fRZ90UD7!q
z-ALzvbT^FB4MWWQ-!ZS>z4!Zm|K(b<-X(M1IkC^)&wlnk`*?eslZ@r7V&&!K)lAsm
z55w!KLnMCh6jMD*!NoUQ7`XzmzGic+(~$xNf&_Mxv<m>wzH1<5$YOjNd@GBwN9(&+
z+okH3rYK6IUF+~>TwTE{nZu!ax4>B3iowUi_mBn12KgF$FPB?$FQ7F2Oy2uP{1kBZ
zL}8)}(W*3=^LXpEJthjdCPouZ$?=!8+BI(R=K>25y?cfWWSDLe`Oh=Aa9#_teW_jJ
zVSE(<obZ$K4#O>OIYN5Q3%$;)ll|ZV#q*x+Ac#@B`t4msrsMIfk={5C>zC4jTD6{-
zwI^TLHqxklcx+}`Zez(6S9eU!pg9csh>A4q&*w!k7+CWYoNUi7pD;)DWztN1ZGyLo
z#ht~A>f>V=!0rRcFwrAhV#eXb(LXFXY$d(W&1bg3gY-R#8n5d@nxw$+@Gu_`@McTV
zb4ggjkTR9RxaFM64!5||N*)3OAVh~WXTS~Z95B^A^FGG%7sG*3yY>ls526++AB1MW
z&sN7TA47#sgyt<T?<{pBZ+O1=er#6<#<*}08Czj7-jOBO;z9i4V1NH*Q<IOd*h_7C
z1l=GJc<!k&EdQz*{U%-|6Lz>=ixPTPa7+uhIhoN~C5dQJ8b5*vNrpV~C&|Q!TW?CO
z@zDV@y9hrVgH^d6yI4YR(dm79r2wWuHObA9sDQ_^>60zVj2ea2xmOs}&(eL++G$$+
zY}0ttybu%@u;=Te(6@{AURq!U{43?a?pO0r?nz&DT>AUc6;c56U!Y?E7j=@-wX#0z
zFr_i!Ct}-X<6t(=9KDgv;k>oZ1=RLKvGCCg>U<H=-k36%68D@7AMk<HYDHq0TEmQ=
zSC?Jv_*g2u%X{}CWo@NfnS;gzjA=#8tRkvyNSW~9MIosea(CZTL1w4pB(e+sQ}u@J
z7#G9pnI;z+sp~YKz*Kc5R@2;{2k<i_X+^~XJx&*H-cg8q1(ZeKpFT>C$eXX>uQNEx
zAH8u;#v{4|)345XyXSK?pS_AWBnhyVrZ53MUZaETescr_WUY1F3cfDlBzoz7)V6Hy
zG?{xNip^SiVWF)HI6L82A0|3<T%l3#F6{E**J;zKh+y#a+?wJ>1hr5w32UZ##x=?~
zW=iFt3`qc`($E0>yqBl^3kk7>47%NwRl<>yV!v|I7oL{mdOG)Nd%r{o<KjsFK4<z$
z)cfIrTZe<=@g}yKX(@?0^|hcmpjMgr+5rQvaN5;5VdC{}b0Zz47gt83p!9Ryia!lx
z^IQBbPbeirhjYh2=Jb3!?U;ccA=Olj;&iuE+f7=Fw6?e~Gc%ZQQT$XorgC4=6Qhle
z{jL^4XR;c-JEpYRJWVJOWO2>U%EuZa^jKcL#I_sRadSioc}XDizA<mnbKax9Nu^XT
zk9+jIX%Abk##Px;n~#r4-g?p~!f^W|4`_a2-e+KWq<i?{w01ZLys*F8TkW=-J?2>u
zj9+KdqT!5M$o(PJ<U&P8FCX<2i+XXhVIdmXy|)#rO`tYKDxV?!@$*&oPbWT~q7TWQ
z?Rj~|cA{(Pc~E$iI1P`*x$6(42ys0Bg6BDZ;i2?Qk*m>OA<#nvPV78%84VZlg89V5
z2uQZ+KCKIk6@Jf65@(w|$g&9!568eF^zOQ4^*E8ceQ&w*fd0<PgPXg2K+JjuA{xI&
ziz<wV1aUQeb%Ea^m8X-(&Fx`B62rVA-s@AU0765>z+EK0K-jl(8boQ6Oi$?nl>UKM
zatl-aPnBpF50$=A7Wl$R+#}9HJ>*rfSVx(56;oz{T1tw{*Zm_|u5^poXOF+Ug|~Rq
z6HRmv_y-_iq*d{rpqc|WTdU8GGtXH5lfk&n$mOL|!~ekxa>E=`4xwwCeO_N-psw*S
z$df0xzO5{jc!TyPv$4%gO&Nrje6uA0tN!50aC#g!9a(!YQ}kVO1Ir7tfcrBYtUg~<
z0*x~Bg@w-d_sGbtVYeg$3z8e|T_Gb6XR7EX9yNr3)szE`%$-9>*Z}v6PdbWm`b9#T
zeEbljS|P9XY;8Mxfnl`r{E9=aAt<=}JUo#~nr7JLt2Vn{V_QCPLCJkP`fzHY>|yN6
z_tq)JUuJW70N*EfM4gzFl!ug;pMni4RCn}YBhat;9T!#!Y=CPf{)VcBp<DGXd;}Za
z4^f!Xusyii2i};VJe;H08O@}O-T7?)6<(Jy!ivSRV#JGM<lg$8os-J|<MWi~JuBZ7
z7yWI45XZOc8IWUlbR=Eox0hxy$1ur`TT4zTSx@HQ|KcNZUM=_1oMVXW#=r;TVf2FE
zxW%8WKXo4pc@^7oX26Hy0;;#Z_Orm@;X3CHX8A25X4~*(?a%2kD#lDF<Jr$|0zbBb
zgDQL#n7~l9Q*?c>rC<4U<_p3;Ee=vH^D(@KT}-7L%I_BASe?hFRQt%33Y3-u0n|?q
z2YFwfy$%lTI9&KTle*buOTk?%HSKIwhU1&`Zcz=Sll6$4Z8|S{Y|-ef_{CZwaMZWa
ziDJ6Ss&kCBTbnhU3bSA;E_k7@B%bVagPrk-XnNlxfQ}uHa$mziwVrLjWZAL%P}*3j
z{JF?0qTWF8Sso9v$MNRF&hUhdBMKhtH1pAX_lQSSn137%TPUN`7w)C(Jz<9&R2EEM
z@7;W*=lyO7m+A5?@G;*_NbuLfK7%|o{A|u`f#MXWLf=jbDlSfth+E|euep{y%4Lw_
zwhB5^j_q#TWWRE}Apx7b58qGJC8YkfeIQ6exCd%NZ*Q!>LKhf6;LlO%Wy>W+AO@dk
zhIc1cYcTG-actOi0P?@0Zvwo}1rbSK{0PDL`!h>fGw&iZq;7Ep_1JDUW`)Fl>i+a5
zLMdGgtqr6Bmazk>Y~3SX?g%A~Aeb$8%1|(_j$%+SJbl7BQ|A&v(@^U^3naOG)@Coa
zo~gzRu_NKV7%-R<an`+i6$cgrhc9<VESl}S$~=-;(m?T{dTo#`UW2R?|IzqKwXH#g
zo0w^!;gkwFxdK`SaNTT+;f2KJWP_BIu+G&2U|+mRv*p1TaM=V5&eYlkQR7A#EkIX%
ziXYC2tUpq`Z6%Eyg63v1T1&jv7t%gH+DHP=4h|_0P!4n$F=bvC$5@d?Z6O6W16NVm
zPHR3e50KvBs~J+bzut;hQc}`f?{+Y9^~HRn(DByviGly{EI`Mn5d!8$$rlf~hRjdL
z)Fs_P)bw);Y7lBUtk{XIGEV-laU5C~I@>QTH{$~1gC)1W=P;~VXNWY{{_@>zG?8Vw
ztsnRBxIKXeMqfyGzm2Bk+lb#MU1bkOnL;<Dc6DFJd#X+Bq;olu%s4p%WHB3=Ml?^k
zywo7S%GWQ^${$ATrsVx7h2i^Qpq0nL=)pmh+vo~{_C`RuO_%LMB#!&sJxYS5MrsxB
z7Uz)U!ydDcE40EI?3CjNxqTJ|vB`KvK}SMv7q<g(qwv0t6-$P^M~<}KcwfZ9Lv{0v
z+rs(5Rb2AWmjSn>obo3Q&SJEo3bZAmlw1UUpMo=<r;2!7Cp9H4)T~L+b8&D9XozYM
z+me^_eoFrtxXb*RPj>d&(rxnbR^sFE;|uhcwZ{SiS4Zc@r~$hIx#n_tej+uzlW2Z?
zS@nrgBQ0%gk1Ue4IU*0%?P)5}eDWH%rU9haAZaOo7a8Sc8<$3dGZ(+}nT%(0WD1fZ
zF51K47^*1#M`D2)o{CUmT%jFsk_cMF6VrQ*8D5aZ^Q)}F3xuMzTs*hl#>c+_EF&E&
z>*N0VxHvKmP0cmQvhF)#&H8{rO8Rt{p3m?#h9k~qFqP>mJ>cR3p(_SPE!dwQ@4xvg
zWj+w-`>dM$dWuLFgww(%u)5nW)oL4V_<^XP*wtPmm!a2*+E-J-KpPYA`PfjU{v*gq
zJFw>GU~S^+F7{8g#;B$p{aQ6hZv_$XMN2P95Wt_i(JrfA6`T=_iC%n}ef8RNULiYE
zR0heo;28XDAbI!YqXVF57|v&Fg+L0S>!#qORjAeQ*kiNV+8DJa-R!1UAyF&@gamGI
zOv27qz+ZV38vLa^Bsrgh!Q^_qU<1^vNVBSK2I_#eTIwQ-q3%jJ!d>UM6*^H<3f#F|
zsNhLn(>FGCJ(1gb>Ezx~ecFO388H{ZPDofX3cCxtcY{mn-0LJ_zo9JI4a$>=NFxp&
z3O^5*9VU(!!?K4vI%5K_0s`rH8RPVJ#+%r0eV}{tQ9>qpIYr3fdWKYZ$EsYcQo149
z?b{t0GG%u#;?tMTlQg1P)zuhMB2onb?w5F;e$Q^>n6*%+fh0#EwId@ag2PIZqw(@s
zEBSb~+8c#rnRR#I=EIr@rdFIA_od7IBNqi{i0^$}-@e9x$Q`H7qq(eR<hiWwVHrVV
zP!2!P04yHj^c_4T0_O!-F|-*t#{ap>Mz7)-Kv%wkU+(xYDrfgGRQd)2nI1d{#$|p(
zaSmemJ%X*dw=qWLE%2nwuwUEqUOWPM{s_*!65W=h{c1~KNjecM`J?Ds&GZ-ClI@4~
zx1Kf-9X*eN3h1lEtas!3MqB?r(dnoCvbl9uMmo7TcO#)t)^CpoI0R}n-9+3Nxqa)b
z)DKH3YiYd2rItV1_6z@__})!ApDve(=S={a+aRKA{JdEQU^c;d-FM(gd=A%_JHtBz
zeAsu1$xQgMhsCpjLa2s@up#L{i(y$)Z!Kz4TmcIK;k9u+pRo2`lJpm4*|M^4^EW&o
zU@G|p{Wq8A#|P(`w`<kOLA)~7pI<(@n0lKo&cXSdKbVx~bAEYy$9#<u4@2y|EbcLh
z`$>_JX7Ey@PJLl*+R9QR9V1A;<bbYF#n}n>0c$q8qLi!ItGENchfl-;jYwh8a~f>&
zLyFYIusYXCL&x<IrE>LAxXVs-w{KD$DHW2DkrCIt=_`3DOhUrsf!J9A7o!mj;`!C*
zyjM=P=h|NZG@?EnX&arP5{Ei-RCo-som_XJ@eC@|OSm+cp^l4>4#4E8d1U11dB+ZQ
z>lV{_kva~C6j2jn)|-Z``RdR+u_{2_yG{vx;OY`-Y3U~vAEhFwCq5eYb|wXP+yVmc
zH~``aZ+VH=m4)%Z*4)>DXTWab&V?Y-wm%V#>*~{}Zp7LCUzNk`xsYc)YZsT2LuMp%
z8WrOBJWYx#PsGCnpjs%Idid+zJ^pvnrg(vqvP8<Yf2lZTkB-d1;3$6PlwMq5x%SGO
z8{ppKCvjV15@TzcQ0j0$eYv>|jh(a4(yJN5Y<LgUuBHjbjo)Y3!u7iyHGDw})Ltri
zUTQ19&RFbDa+|r;m}d@{tA!O_G%aqH#?4eto9kEMcNo38cTAsGmuAG>pCK0>SbLt0
zJ+7J4;9@=h>2`?BM3p*iCBV{1f5BEMRFP4kj+*KWXE{39YyA)q;^rfn22}gHu32Ht
zdK`ffA?ll&)_16ryi;Pk=Q#qJ7hQJkg8_AZS|;tTg;Vn}3zIzxB_mfL@d0eEWSS4H
zp-KPRWkQu%__KSTbCpj7&?e;rrPRg*E@0=Hyi<I^sK9|CKtT;xRy>6t9>&@_JK_`X
zdvDL-aEh6{mUYSBaM)`>3#H(E*v7}qkKEedU+sgRj*qdcehMBrzrk&pW|Equ-5Z|h
z<UF)gC(pOrpEw?OgDf<+%WT^vh-TVoZ1Y*}nUiYRUKgy|ef5TXd0g8$kL3g&7Ad>v
z){c3|rCx(Oj!*UjGQ!?-d4PCiDh6M;PHZkk77rgN-41G!626$&Gg=k8K5nD8Y-c%T
zgId^!X3x&mD&|YY1CE}bXv(*m{AkR<!^3kosGCBJmK1TqTDzL@Il&JSA)NB{({?!n
z_{b&W3vB}V9<4bIKSdBGZZJ9$(iASDO5;<6D+A@IgM4~vg@sqGd1_lFzjwcyzzYpv
zAMQ`I9Q-`Wwn&gQ==jx}h*|CFv9dvXIKDT(JZM$m^)Cj{Z&xDgTYw(YDqsOwJ`)5e
zTkVh72}MmW#qI#Aq-dg=^0QoPb<>@8IsL_QULqE?r0{?Yqdnoty||jMWPJQ%yG;UP
zZ@t`hZr*H-Pl#`sZN+&M7Hr-x5^r<gZgB~1bFKotK)o2ylRRymGH|3zd=mWNA!AvV
zSx^{utPbV<ce8DnQ4&KMKrrUha4*W1k7Yzlz_o+R$e8#5pd^ApM>)Cl+*yuUNEL{u
zu#mOg`!4u+dlt)K7=Ot%ncrn;nag?#s1ixEvRt|>Gg<1)qFyNW-1YX*&``86z*886
zFOVx6SUxW-7hz9o^6gd|yJZ)S-{6VjIhZbvP6YeF(dkbGuXWse`U;G#Oc5Kos99xI
z6nOr|h@qN4k<v8WuL^yt%7#6>H+t=OYZ6`0MkErWQ)RK`R_l}PUiO+~z+SDcJJ|;a
zY}Zwx-{7$ZP|me|w@FhFyzWG#P&=>vBytTSg`2peJ@k=)>XyZ9HkV-i4Bw>EHcw(I
zQ{!diA^J<??BK=;fvMKf{zrMKwn<hPFa2LFTav%HXKQ{1C=SuF^}MsKXRDvh`%A5a
zOn6F8>XBQmFx2I)Hdj!mEoRg^4Z<?~JMHq=#|PaPn-jsrEF6{-99$Top`oW`ukzdx
zX^4J`AwUwXFl*xj{+n-&P*dtXU)~twdV4J9O1H|@hP7N5Hn7L*_#>M9qo$UxZ*$q9
zsTpMKx}vV&B1pVnd~tC4lm6l#=gekXuwt%WJIV}Kw)|#d(l3cv5$U&xh~dI~d=G_?
zVnxc0lkNlKwO&(R;SW@E!i4L!>(1%9&&`4O(LwCMy{>%R6d|9UODZ16TUQ+^Zg&{^
zA)_^iH%9R&)kqwp1G$Zen5OIeMRWLdCkLm_ckOb*rwyER4<7+xMCYw3b12J^A#+FQ
z4Hklm%G)+qH&S(AJphMq=y^*W6ctrHu^=CiV~`Pex)jn<?R|N5;&f<j!3-Cj0iB3l
zSDVS_KE0p~J566|m1v&NmW_o2J>Pph)lQ?#(qf%Dz`RDy57)?<B3wk{-uJ3yaF9Q`
zfrSz1IW9#NaLJwP(HR$TI|%F9C)n|$$XmRr_9+p)9QeQ`I8Y%L7A(Mbdzq7%L$LQY
zLaUA>7P-#x>&GFkY<W^n50dqfBD%AIE;C_a;W!cXvK`BwL>}vf52(Fh9sYDV(tsVE
zTE`CU_?<!#EfJIX9SiRu|KOoy&oH1OZXa<}W@DL$$I<qi-7bu1c0GkHXr=WQe<+oZ
z0_&a01a{4OJj;fa17%po#YrQSz3$?maa=FEshNzFQ_yEGG|@z82ub`r&dWAD3(1}V
z83Cf7K7E&9=YJKaWqq{3A?NdR;+y*VfKdZClSI#SpJVWL3ec@Y;B9xg#5A=!H*JaS
zDahJpzLM(k4d!Vor!f-&>IYoX!gnJ8h6F`u|KVT)ZQtXR-i&PT(KB8Od+r_5Fn+;m
z>ebNc-T?QZ>N6KJLB(7KnXf_7;nH8<M5hqt`6EvLOaW}o3dTy2r>)l~m6>O(xi84Z
zb0hAs;M5g|ya^E%8@j)3Q_(YuPf`6kigZP<WvB^xGR&41h>Qrsgipzy^JcqqT3={r
zC{#*7_gA^L$45rw$T;mor30tWASC7Hv$7zs{w}rilP!o;8Z5Oz04XBB7A^wAl!?A_
zGmvXDkPGYO-6Bg(v8S>5PEv$~)38ysU`CuawUA(q@|Z{XxzjtQT7(1@YVB%1Y_Bwo
z?F{g5hu~qz`IgY%xrr(-1{;`3j7NP}VsgjDdhS3_*e)*KW58|Jq47k(1uSGbTYHpq
zks_im>SOe*pTeeAsj;yU$GWMc<cVUxw<i3&2cBZD75+(km?oiE?_!^AYqDW1M#Si6
zE`U_iMLQ@uCf*p^@?2_Pgb1Jf^TIJz*$ChJ>z7|zBW}z#8qqEoMTq2A!JzM9$N7AC
zpDD3<@*j$HZfuPEWJtkwUpHPX^(wzV51*LkTK-JFK9}Lqu<r%GJn87OgrAF;J~_#J
zakdIS&)Roc-B51=uIY2y$pP*t>DQPmxAt+oXyV(L-Xy#_%%$?g)FXa>fyhYIK>YFy
z(57dz{!PO^uuJcT$4NbJ4p1SOEJ+No<AK;C5bpegDQmNVMp?IGRxQ0$%kCHl>>=m2
z(%|4=cWj%c%S!9oui5YW#Aoq$<F5EG3QG2e<MB3#f>S617y|3&s!zv5oLVDncrVf2
z(2+4{eXk%sME#MquqxLXIK7wLS)F9_4LiqDKg_Xtu1-6Y8yH#%^DQj4uai9zei6GY
zCxj<@1ql_;tQ&c?uF|}9w!%#Z&Y(Q{mHB+lNx<1geP$$4C+H%5zHjhUiLmbz61e6X
zRqyerAdku2!h@MpTF|oid2UU263ni0@6$b_@m3$#2`<_Vd9X<MC;3?H8j1Au^x*I}
zW2@?+%CTUbK5}P@H;Uo$)08Or++KCv6YW7XKDkO6Yyt(RPKOisilAo}m*)*|v!nIV
zscM6aQr6BZ)ijr<owci#S`VE2$c)~M54S6JKFzflD;%AMpT(*b<<}J<-MD{9pySog
zW#9L5EIoryqhQlEKAT=)oC4Hju<3BD>@#G;09O+(GQUbg&SN(=1`je40qQHNx*zoW
zn75m<gg%?5^xyEl4`Zl>LB_u(SQOuTZ|rrtD0Nt+4p+0dfA&$i15V@3n+<Bbkllfw
zdllsyA*kChYN06(vU`Tro?5_x*N=V&{zjDz6&7P(x2j-HE!<bJZAZLb7dNgk8*%e?
zUcv2z$jQmYM8{>0Tdc!!a$uR=;k5$Sk7cGsJgZfDg=NMWXnYK@Jsqj(q-Lr7X+ZVs
zvN}$qG3%c*r2JDs)84YOve6fsA4{JXI0%F@U+&Dq)T?HmI?SzHbV3IVAx%0SF6%rI
z7emCdroyMu-u&C^oR~%RomgmdK5#D`mszJHS!yyeMk(xNzbnWo`1%}4rb&}n$fmd7
zmM}hKVtE>R5{^ZaGft>?@<sEg*hw#G!)uq+2ZFP7QKX|D?)|VSO=OVfHHt1VI`wHP
z`r$FD4Fy#7aKVf<aWRR0D|Cnx<e|L1jLCoMV<J51J;56QUv&Z%prA=2WDFr_HwY%q
zV$mHrX8Sf?dJ`4xwjYBxOoT+lX}n;5O7n7bffmPye8n+zNi5R4<M+DSC@H)2D9z9h
zQI@XQ`SOZVmY7q<9rP&oXjrWk)3HNKNd!&6j(x3-$Y^M2vX38+zN42bx1QzC_|SAV
z3OWR;$xc-bjbI~%U~-Uw4bfjMm23Nj_F+GEwSS4VP!u#b^u;e@z>MFPMSiFE!LU5%
zV*$&vRg5p{hb(8rm%+TY-gf#&YooPC@mf02(S4~{7Uugsm6)d8Q`p9j^ZOa-)*7>p
z#ihodQOA6)4_z~U08$TITeYTNioJJ0c=_I2PQdoi{~b<23|9j;D8$vEW-6k@Mz@O1
zAuPM^qJqWdvUewSl-?kAOacv}k8ZlNP{VhMgv+TpNNt9Dh`xlh_tv!w+AIrcwQ*?t
zyaV4CP}i-pUF*F;%{Npz&XA|!<+sy#4kjaV8C6w(0fegXZu<HLp^{-kZkjp!H-AMn
zgb`&NKK|M<ZB6bJLNZtP)?!Zk0Fodwe(?sS4e8sI+EUx;FV(-%jBIG>0xfm6vp``h
z!&%RDFWrw<$cTNc)}oR>*eErL+&x@Bd*b?i5%ENunciLwO2icM9Mmd|w-vw%fn}sl
za!MWZJdaVD#{_E9X6lKE8GTbN-#&K5{DL6z!#vN{`t;RxV4On7&8hVzBt+zZT<i)X
zU2S!nzn9U*O(d-qpIsebSDa83*C2LTccgB_iOgr?v~m<^irvy#PyGJ);klWaB-0uC
zrT?j@rzI04$feO$DoQtHr3ds8V#-qJ5r@MmQznZpBCEWeQ2|{UZ1O7h)kRtFV_Ye8
zH_E_rn<Z+_SyY-R$DR8Aj@!cAj<La}VBDbhO&P3;qias$^L=W~?Z^Ea!q4{7P1tg&
zE{;VMo3c}8-KV)DvJ8K!SzjGA;mcb(KHhsW>PJsgJ+!>swJpcjpRxn9)ch0~JFskI
zF^z*&mDb)p^W%P5=sNs(eSW=kn%W|dXoTX&hQKo;@MEEF5KBC<<gVnpJo}p907<$t
zS**(Qaq(C>1`-c_UiGF)bw@8sQS(SciS5S<@Cx{ax#*?uQ<><F3whVNi|Fsvv%y@d
znAv?2mFBfTHR2~T-t&`wd)7af><W?`#>H8ro;f6COY+&0w|DFv2Kfq!8b(yWWDF7j
zQZr5ve&4CQy!qk;VxhVJAS0BtqCwgs^;v;4($)$c4c}O~i*V9_BCK~K`--rMN4i(i
z7g2bc4<60yh$I@#HRxh{gx<(9NW>cT(?aE^9<8{XH13D*!I%R%wjRn5i%tg<;jg5w
zk%6qw=;kZ^)>Ptfz>aDjfxz0D@V9ivKO7V`qd8i}xb%k^m`t>xFZt&zwa!)P4uzBU
zJ%o90`EzZ~3PLN(Y;;e~e4du?(j0`HU9H3!5m{?UuSBvMU{GBt`?)k%ko#$OkArO1
zv&!3n87gj=)4ZUmuox~Tmu+Fg7ro;5o~c>C`n}ssUd1ES;0PiXN7>xqp5Rb|{#v8=
z+jZ~igjnre8L$4ZsIKXvFWmAZ1s{mb$9maBd+W9c((5LE3E<)@yQNOWHg-lKuKLVy
zEz0Xo|I@1JZDps56{EX};6D`CE|~X;TeZsk44+s`7SJWrbwQ^PU@R`DT?VuKQ2rQd
zwtMU6i4)zm(DgT_?}_$*t~9WHdNC-O*oSq7t<_&V%Tvl`YLXtI8gJLi_9Q|@m~Ko!
z)GRQi;`uezXkW;@TZ4HND>XixMSs!0NgM{Y9h>OmzVK+}oLhJ)C(gP0nazx}Jmid1
zdWiiCsq+SDL=&`{1fJ?PZY(i?VGQmsw`7|jLYszqzB|s0gFbI=_kVCWN<0U(;K?B3
zwGLG?EGPJ;iD|DV5urObo>m{%8>~4YrQ_c?3v4*3ZSNjnI%*gXJv;ttSE}nab>Tz(
zdr{`-cY8M2zogP+k_TMhnp9^RXk67HPAdn#;fn)EpsByVG;zH$t*K&o3t_8HfDg1W
z*6F~Q$HUZ^k?Y4|(CQ-rX++v#?yk`VxAT;r5({vTA=5i3;`dZR%v#%~JU_xiV@K^;
z`oloF;}iMLz|adW7lEM{&<-Jnjs@myLyBng%YfEQ*{WOVj>EfEoN9X--DbmQr-^wg
zj%ycMo4n+eG7M*;BtEMC-ab{Bn^SelF7QTYlInO|;jGl~x+$x3qt<W3!@=Sp3xDq?
z%87aVuI!VZU?QU(bqVcs7i#yMc-a<fK;*vfcMpu2lYOr<NWRaFm^aiwod?eMlNp9#
zBezcWP^m6NI8(82cn7pVnwBGJa@uLHB#~pc>C^tOeYOsiU0&<LR2OakXgm!f=??p7
zQp%jD<yOCgI^aj^u3bRe8DbjujWXQ7+soq*It8)~h%d+ieFhH&1+`47($aB?6Q(3B
z<j;BdxGFBUbWB#%j$r*%QMPr>!AZnt&{Uu8sevC>aO>Qvm@2xk|H)Y9l;{16niC?&
zue|rT{Yo*dsS}$ur0-n~<`qx63C2U8p+XOCdX~PK2SCd<tlm-YS6&CTPVp8GjMqRq
zFIH!yQpN{&7nAI|EZfP0IIo*hx}@4o2-^h$&0kFMmfZA6&K5cHWdUjN9vI_9P)=&c
z;!d(7bk+wnKxcix6^XH7^NSN{XY^98Ed}U_JL_5V^&{Gk44Qrlo-Zq^^+j*kna!9_
zSZy0n#BBevyE0eZD5figGGhtU@A+{!hHw-fu>xF4wV-nS+pzseZof@-%?KV(aMQ`L
zSQ>?6dO)0T!L^DOjh8*SHhmKXhe0wC_c4MOPp|`@bKL&1C-BhgIj-Lpy7V10E#}&<
zEntJkQ+q5ZvNDF`<RhfK0wLu^%MV{@`if5VJZNVPbr?O|5b-I3&7Sp23FoVE`PijV
z4dTB3WuM>ME$Hc8$eq;NMzA#Xq@)&=Y#UVe-^KNg&YqF7IOtCA)+=z@nerTRGj%b7
z(t{FRyqAY}PXOz-<qv!4Nh?f1Ls2Wdns*>>aMo0AtY)RlbJ%Z*rfoT2za2$PW4`uM
zO51H?el0UkU7EeSQ!2lBjs1E#${Et{5#7Fi7j77zqA55WGoB@UQh9b*J%PX<WWXxu
zI|0V~*(HXO@HD1`LWGK%_j@G}y7#efO8v6T^l0R@o8NsB0B#ahZ!Y#mmkD%EtR8OU
zSb&oq1ZzAZb}U-ABAEv`54yKYadS^@@5@BcpgO<)=Ez(@w#J{Z9XDp4w_bCaUCt>u
zK>>iql8;Nl{(cqtQL?L5k~g7?h(9&iVRITfJ2~*FdUfl*S^UFk0N#ks^6tXDX+TtU
z2ch+B+Ypx*_AM$^E4`LPh!tMw1D3+YM4=wb?+$((9On60F2%%>rP`{58$6$z=1XgG
z#qcWd!QV^n8z&PPe>+|#DqaGCn@lM+4i|3;y)y@AFLr_UN!FNhA3up&u~wVunj3*m
z?Ip!4o}pdc7~<jS0kJ}FL(srjNu<2&c<6{l-HAZ)nA6Bi4(YE4gS<)t8xk4jodKw!
z3s(##ff+R_%{NzWRjkzb<XI(FhGX(469RXbCY2DQ|AB%XDP1@G$B$axLHg^9bvzm>
zPDRYevy+M0S)bkk#+KHO!9iBsPHE0OHG|wP6->-euky(Fk)%>eIB*M_o`a7AX~-qr
z&xO%tdGb{EL)ZMH_4AKAVf=k^aXD=kgbkI#IV`TDi-tu__f%r&0Xz?Z<&17Q?{^Q^
z?D79l#u|EQ`Gpc2HX$m;dGX^%@%)OjOPh(~RKoTeRSJy?SwV84X9~eSqE$MC7GG)H
z$fG;b1n-dred5LLxT!#hm=HgAyR=QS_$BIUnJnWhNnM^*(l@oswXPcs3(<JMI+fG_
zZm*@cZ(fQM>bv*1Uu-u)ZlGP;z|RK3FSuklWjOUxp!@QawAa5)uiwFI6wbN0bhI6I
zj{SnyDfm4}i7%2@25{&ne;YJ|8wnR&O)~q5x0__{8({yC^NCy6fwE25TQZynwZ4NG
zI9{!DZnH}(km$2H3rR2bZTx)oR2e7n<E}fulsW1Q>+aL_TMc_K0wbPup>yWd$*P$K
zxk?*+Kj7-dJd5#S5+<B>0BuA8no25Bv>64N&%>T5g=~9jc;4sSWtMx5`=HYX(#l8X
z3(3AhU5<N(zUSvEEx6*@4BaJyHm@}#a&4v?6l=uu)2o&m8410e<Qic6Y3S^m`<e=j
z`mFku$&Q16>=%p&>7r0@{}i5j;`ouzebwaRkATBrbmH9pxM7n1^vh-01B1Sp$BaGM
zzf1(G;LQ??f$Otem(JtGCg0sgbH&XBFAdR!<7pdpNy)%F3Q4&FJN3#LQW2G+d5p-&
z7`%uPX(2|08A1^K@mBl9@M+hcI~;e1lMs&+=^o*rTOVrM2)ewNen=kfYyY-~OABtR
z;x0|LRl0i6+DN@exZz_T$$JIgj&jYj+0!Tveu4ukAT4s47&dd>0T)P&EPF1<lSoXR
zePzq<Sk#A>8V~nP(44JsSF%&pb#Kcxr-vfJYNv*PRxV3PO^*<UH|OrW{v<LqBSljC
zQnkV3p=ahgKM&UanaoM}lK44*Q$@^CUq`fc5^fhkE}EgA#34aBFn;CK!qdUm5<hIC
z;en3e^8spNUeD7KewXb+V8FG<6gO_<PfTdg)6;+b5W==^<~Bi&j4?v0&p}y)jKqWW
z{JDu)OBA8D2lPt>&DY>vmbkA0@$+%eu=siN4BSgUn5a*h2rt@ebW<W+;a-@$dVgcj
zsP%i1T5Y9BC2R0{g3STF1!pu@9(QV}#$n_XZrumywYHR>Qs#GufbkhT{^K^Nf8&;}
zM)5Hnu)&IYwKJGFeYS?QcE4cVxN&dhsJ_%ATLD&UyMOKqI5?n153m`{cX>LM_O$$x
zq7BA{xOku5u}IDq=If5>WKDv843rAD{BSqEUqeT~8>p-i5dl)~8!OT($jy}wZxRI5
zh#SJf*26X%&@kFAm?Xa@4uvJ~zu$Iu8Y*B-8&kpm1rSo32U1}1lsuHY)O*WIF`hmj
z&z(*DSOFZTv^NgGevSKua3j$DhigpUo6naX-<f;9)PIoR8g(k>X>iN>&eE|NWtZEo
zDYLH8np1M|8+ACtX7acY?}l&@^q3{1F`T~KMBck)*IBrlXDxI)>{EfH@3EN2O>aCA
zKA$+`Ht{6`hql(X$rB)m01D%F9bFX}Hp0E=MpnmXdYOY-+;FVvwd=<>?J(O(<ua3A
zUj9Il*#DB&x0wmB0a6E|&0)wdScZwJIKUM_VGcazPpTf~t5tZewEh9%0mN%e>1s4e
zQERcMQ+&bf;a%BM8&q>I!(#3+%?0D*nuxKrC~>{{T;DnW)A+dzkzvmI@I@zVkdZ6Z
zHzi{$Mag^GP0{03&kdqDGpR&otF+ODXoy)LY;d>n{)JDz;)yWM^+2PwD#8qKx4+Qt
zryOgQY_B9<ve&SOfUosHm^U1FFw2LPTCR2`RM-ZxfaJWkt&YCh*&XOgJl~-vwC%fC
z)3{ZUpPxS`9LuJYp9;T#P_tkOh+XmPKy2C##AqXj>0)A1jW=)ZO-Tq*os?6YbY%Y&
zIHYqd?X$tGp1EWz700uqVJ--mud?_saGFjv#dIxdc1K=ir(3aZxyr1Am!reaxm|<d
z2SD9*rJTTX$?G}jt1tGra6IyEo^#W=*s)gCg>h#zmX3FBGz!q<Iygi}&AT1k9q2C!
z4RTZF`pfoC!Yy<ZuD3Xouo@WP+^?-pozLpVRJQW+de@W8xVsd>Za$PbMQ@)Bs<aqa
z2nYz+KRF4)Mnb(y+pN!vMTr`$ujmKe85l2Lc_F#_a{4OE2~->HD(VeVyQGK?lh?~m
z%2!eYP4P0l28F-S01hVIUTHeHiB=1Au?9Cq`KjzX0APLhT94)L&~H6X22l;fLo@Zx
z^>I8H)Eljk>1k;4YR@v8z)lYpU@i~`T^^rDZ9&K4aD*f(hN2|V@QOSbfZFYBV3iyB
z#gCJ_f$QA%txN%^bp+2eTNGKNIKr_U*M)1Jrc717uy7QGoCrW{8B!AGPj)$id_Mod
z`)ER1dFB*HKa8h1tvcDT!Xl<BUYJgAM5z=uHY<AX6M7v_njYy^LsC2rdsslCfme`8
z5RBfV>!|G|v7&pCh5-tgs!Goyl}h>c93C8*&JgQ^GUTm(DE{UsWS}(6I$YF8-J6Tu
zAm!dcu?~%DP)G%=<B95f!1GH*U(3~+8IkO?=TpbAvMeM8B?1EZ!u2lwmjCfXIIS=q
z&6^0EunL!PX!^eGo5q^FCT2N{{iM8u5mqqZPA$>}Ic=<LsJK_Rb!}`nZMf84PF3H~
zH90fLrZG7y9g-I*mztv6ii~HBfAcP$QIU&qb&<|E!4*EBs%f(3r0gA;D^MI#wXrI*
zKI+>HyF4c*AoyV38T%!Qx;K_xT}6e+8i1ep2>7`o`UKkp^ERu@(2UOj7c=o^PB99I
zWw|<99?nm%alycEo*mk-c4?Xdta4J~*F^$XXBUA~#i0};!0W;Y8yqYTyQ|l^??kQF
z2<F3h*#wR+JJmA%Ks`m*n{ThXRPPA3(}x^1ji=C?=$fwZ*-Qw4w3058p3Zyp=h^@i
zzPlAH(dj+BF!^g{DS0#WB+&F~Ct7TviF}UEAmDf=sUzT@ONU=O(*V;}!*g}r*!@kd
zlm`8-xda{k?U@=`+54u|#php!vO=z1ySB1i-TNA#^fknMKiR(L#U^S|_z52Jhu$&C
zId;YgY^pq3fBDn!yQ(bI3C%MZYQC~a{oxj881Jrtf1`ycxsg5DeHle+l;eOOt+w$Q
z=l7Di<qP{a&P3w#71SqrK+~ljbxYZEn=MJu1?6CJ599tX<`5A9sqy_8qUKggD|wGJ
zA%Fx%2ej@@2ntIwuFLjhdUak^(%bP%$C=S>nS{xFV>HQ7vUI17u~C-!kv!$2&9b%q
z8o;77-$2k}9d>C%+yv4Z(P`1dOx*eSst<qC!-aHsLcE&2CU?se{AQlKo9QHU+=X{B
z%!>B@L|x5d$bU2ZV(Ua`KTVP7f+w1!?!FPL>t_4GEOzS2eZL~Ds(c?f4Cnn>0jO~(
z!uya8xCOEOQS%(4FDH9h&eSBQ4>pt^^&7I<;+B;O{BZfr9wPVB=oIHlpRZ}rTZpKR
z-98@!p>e6^|HN83_^*FZx^rdUDP69up92AgYJ!Yq+E}<4NoS?TlJCIp7td+B{)QnX
z_oNcYm}TPOIdP<cI05Wu2_mk84c^{l#XHC$%d(NJa-q+o`YE@A^2mF!oh)wl_MBU@
zoOS)*Jl1Xop^Sdr9~(KX29;lj5Cl^M1k%HH%uDlNOTGV5pZ70)jygkYJ4y^#L}l1Y
zAZwXS0T!=GfJ5JPm;L3>QZ%E5&&mpQJ$r&gLkRUbv_k||RLFi$rZyUoFF-^}s=l)f
zgTeS6zr008K^brw_WbT8|N7$ZaRX0x@02)iX@*fw?ppeWKU<_MYuy9r1`7lSui5E8
z^dm|Ih@gf>T7A7Xm4I{f@=kX|bTrYd8zh;8P3V#MAMg7ORe>kJGW<AUuiCZ#MD-my
z@u2IYJoBCo+=)vK7_<vF5kxXZ!tC!~m=IqW)3O<Df*pa}V`^Y$8%J{<JChmjMv(ux
z>xds9VSA#ZQ}p&N-qw6mQ%aj2Y@bx<a8R68`}I1V{$qr`IjDZZ{F@jGxpr}o3Lj~E
zL%0zWp7oT~y_u2Z=kopZv3^n!SCApJZ)0Mzy1Kfqiqtqj6o5F_24!WB6d*9OfIK=>
zaZF478hlBNuhE+k;7DdcVd3U0cfM>1y^qgjGW0^N0z{4fn3X?9V}^v>-rf#`ws^ue
z;Hh56@swrlK$?j{st?Q!D_2NWqFDey0+8g+u&2MAK#T_&qL!A0x;~dqz-8Eq;c!K1
z>A_|{l*Cer%^QEb{U??6eFA2f-lnG7U<Ypib7ft#O5(B|Y(?kW>;^cX+k(5NjK3yA
zw3p~~Ez!ntbK;(k^Q18=p|9reAOF2ze|ge>h?wi$v?sTJ&-MR)@+$+xcxKY=ugv%V
zZiR$(4Uo>4=mEEW@fE+9?BAb~l7Y<&28X!(kD>f21{zNQolYXiNQC)cCiZX3yReBM
zESz=ge}xYIT#3Im_$EfpFA-X9;bFS;zmMWCbN;iVrVIc-UUslF`~TI7gb2VNXt>uN
ziTx7c|3e>n1%P!<PA1*?6(#)Vq5S#Ik~CoI@47x=B={y`8R{{NKXU<o$=pxd_zwdF
z#??FtXr%4UdRLzRV}oz6^F;&pj_D_VvB>{RFMkamtN^UJDWtvbw_5)1?JaLT(a^}{
zwwlTX%qVvr{-|i0F+kWB8`JohNBkPtf{CA^FjN8s{klr3u-8mI1NZO0{anl6OK8fn
z+bIH50<Odu>F1d*tghy5Z)?j1s<)Wdk<S4Qf8obp-Nq}4Xja$r${HC_Y1r$aU}j_-
zT#gVa25i|>&dU)*_~G|g|9DEo0Lq1^yScd)PdRQ)sS+|Mr~qj+?nmp!QJPP2P&LDW
z)zbu*W@rEIoi&NHqT;yF$xO!9bX96ePDu%ev8idD+m0pmzyA&@SaPA4y=er<A-HiN
zZEanAcC_Kbowy8`1AQcf{#c-Wg8I9c9vGH0)k=WbumbK7<hC4qdsBQ<Ny!|DhX3xz
zA5TDBg4%QUp%DDy;K=;G{fe~L<<U5I?oCzEv{gWSidU%|%tL>*UXb)dLb=*yS2fWc
zAw}&bQD!>eyqq%m;NRX3G)&49b#kf@6B8>3Fp4`(JdaHZ0U_alqjEhiU`cNyJR2(P
z;9oG0cHt74(-b^Hs3@oey4B!*Gz6H3hNhPzf9{Sy)`9}#u2!YI$mLhXjnTHwPAMQ(
zz7C*F9CI0BJTo^yKUlbRCf)e0!m5>;@3ABqIadc#16dZIU*prwRJ(A8{Tk&z2-Ypz
zlUbLIQFXndlQG~x8VeoWvOM=E55TJZgZ`Og<#^tO=XJhSb;KMQ9i3^K;?_~eBK7s(
zx5lo$UvTg>r!RvKgK%kz)uUqdyq}%Sd4OWa-!9P&+9PNR`v@Vf-<H!hs(Ev=jLWDa
z)LqLoPld;??AdQ2_|GS>*ut#K{ODtzS(P0^0OkN`jj6z;V*`d3^~6Z1g@`p^p~C6;
zy$1SNe#)#dweCmeVK=yXNRBcI{*akJCdxIVd8=x^7(gjyx}QOH+x(a3^YG&iW+f9;
zyj;LvqkN3?7wP7~2qn8GwlP*zP_>$;oU=Y@n(EX_J$d`T36R3Vm(Lj;eKz4u3?fgJ
zm5C`RJZc~B0u-%=2Ext=^Ah3w`h{i>v=p~B@=MrB6G+_;NaHgH3cl>`?d{F=DB1t#
zy!_tV^(wUV^p7wxF^^>DYn?2JiHW=SVgc{O6$EVK`jd&jw#C9iT)N)iabjKa3IRUY
znl9|IVpBf+yI}>~b_-C08V#_|8hZFlUODt(ccLJn)@uRXpYyt46#6xK-vBhpup3$I
zO&1?^F&Pw73Lk}#JgZ*$%K-nj4bCxnO}saNGlc*nvaY(_a56W1G_GgUB&$q^s)@*{
z(wsobsr|ivDTJ<oUmX({r(qNH>C-0w^5V&5xo<xyVfgpyIs8sd$l&N!%>%M?ae1i(
zSU`4wB20><5e@#WlJ#|vI!PhG0mQu8+FIS4!1-)0gBH{cl9YG<Wp+6ISc<E~5a6vr
zbJk<0!5z99Yo(?IToB(2*jw#h!R?XX2(4Kh8Q2^(YK0m~HZ};@do(1+?v_imdG(hJ
z{O1(;y=Q&$^y!D~**XAb7E-t6q<ZJ4UaT;%QoNp`(<cO!=;VqhptFC8gT=MGy7d|-
zKv50M0E^k$itQQ*HLG^@U_wB)%I}H%&!>qe2(T{~%Z5nOIRre9ybmA5fJ9EEhY$Tc
zJv}WojPhTJ_9A?cUV*id-#$pQKQcJ$sh5|R(`t-5GaH)%&zx(<RHfC%r>7+>zdQA}
zr(BjNDk_=4zEcKDj*Vx^L^T6$W+H$~!FSN#ng50a8tMzlE_3&?)prKd{hhF6x_^!H
zZ?F8ORd?-y8F7OmWuE@$2!FrxAD#&O7tPTAzgp<B1BwSE@=E1D7wlhG67dIM&y@ey
zFZ}IoAcUNb(}*$uU9bO_{sMiQ`a{wE!<heQ?mLc%H*etHB>uxV{qrRM{tBQWt^U^K
z|JJB^1L52p1--rYzk2jX_ks95_y5-fEF1xr^WaQe8ts1w%<r)vo&XA#;?Mv2AI*2&
z5%P9<GxL`g{G;FhHc0Hi44i-a+aK!!ylp}(4JhXm6B7dvPahw7<pL=||K!;$sQCfD
zdah!su=OlD#cyN#pTlV`M)rZ9&pXX~oqV0L0I6zgClv#RPsepJz`2O+2w<Y45YrI#
zmv(97Xn=k`oQK2IcN$Jo-UGxn0Mf|e(DQR1qr8b<$sd1Tp*beoey;H^t+*nN#Em+2
z&b83B#25fneDX*4FwH`wjU<%<S{@u1%;TOYDz>kgRXDk7)t+7da}s>jF(hy|3yq^X
z-5)(0sCc_UND!U+&c2ui;_#QRI0lE0eKGE>|4<0g?0SXH8MCkOsm42L;J4n^ctY!L
ziOwoVVe;vnQsx>7BQx_;ps>+uHU8yFCpwKMZKo(7kPD-4gHWjFr<D^^#+H1!m?3KA
z6KBeHKP4Jq0mx4xhEI`_XBrYsreW!!*hL8(>66C;;lJRKB4%iJ*OSh@WuXdzh#dwa
zr^q`qjo{^T<Dj1%zJk0>+~sOi-4Su!c}HQQR{h#t&2nkwZ23I(<Anu%DqdSrAiY}m
z8T&N=)tRyZz8+`QQnK!CyfW9UAs(9D4+P38zk1|>x+Hb{)V>8{hiR8D+R;fq^m=wq
zYAXX2<EKO}2q6nCFLT}HoAkKmd7VD&jSlEACIGT4Bm5W4r0RE?C>C8$JCxs^q1}I%
z!+BGqNV6hemub6hN4R0P1*>11-x)BKDR_W7Vu+p$l0q9ATVvU-;41cV4hfC6*%;g3
zi_9;er;!;Y7})zsLW0qgaJ$jL95jsVcSq9e1=YP{E~U#`YOlIK=sZ{fJLhD*(9uq^
zh>fJu%5!3o3bMPZ082n3`b#{0!OAXeSEFKImXg8#06!+V$l1UkMZQ9?7F3>fPQ+sz
zWszERRZ6x<%cC5769_l_*slv0X!%%JstrZpqwdhp*EYu2E_nJ#EExbk1gIbF&ZPkw
z5I&bj7~`#V`2ZlA)(D898F3}yYfpZiXJ}cH4O2mmqmP8B7zDhFmd9nii?V3G=Li7d
zlm+eO=DEJRs?o6&(9uK8OD;&@r^?=&8){#MyQE`GjLG366-5#56&8=UyEV{lC^kgh
zpd`{Wx2*AjpgK;6U!k}vGw1<ls`qdXiy@d&3*r>{S&z{|E?!A0^=+}a@de}9;0ZE=
zxY0ZXA|VnH!^oS*eT|EVia^l85h=VPYF0h(lExW)hNT=}6Z2zkcVh%x{6yf;0r%6z
zpbe`u(}9$!8iyQDO3p94z)qe<X!qlwYGYMfZ^`2928+3^$)0ib*&uDhi`C!$m8~^$
z77!aZQ*Yc`T;PWFcx=j|p`E5qchoN^Pp=KAua&=$l=0G%iFMqM0onPz`V<Y}k9#-p
z+53w~8}7F_lYSU4aKW+NYAiL+J|%oX{}E)hTZFL->wdWd%hH9qv-wn%Rv+#;w8y23
zet++XzQCx2!#@D(EdDb|0Es;Wr4#G;YxTj|^S7z{YGJBR(&lUE-+1;MBO?e!;7HcS
zC?^w>%z)rQ=+=fJ8op>V0JJDsKXD_suM%HNJ}M0kS0Qs+A1sJP7O(A+5_WnwWuCJ9
z(8JcNJ6!~k4_P51^O@!g?qN{inYfc`%s!9x$X*XKT~mU~vOc}!<VkUF@RDIxP+<Ot
z+^$?3&~fAXP`2upeJ$A77;3@B6M^zO17|mZcYzGK%FZVszpvCbIEhEmi;&RDL1%X*
zm?&0_*|ATSm^g8G0eZZhrFK)toRAE{fnh4v>foS(N_`($;LqrqgRr2tHu1bCi)w<~
zY`qjZ<KEm{Wj~-`p_tE2Ow4x|0@b;mRaqwo;!rqy3oHx+oGKT>$|Z<DdZ@_ow%jpl
zjctHyZjEM8Qz(=$d3h<|V6mPs@$bnCBT&)L#k0(8j2yMH64Gh+X~KndWY<PPru6hL
z05C2|rFjlz{8(mmRnVeuMU0R4dS79CpXSe;WMMnnrj_dBn0OvuR7nXlPm^qQh*b`Y
z=g_8hzt@W(H?+HH6&FDg??&Awu{k8XJ_swpJwFI<0aARvC}^0<YwKy+n18&@2~xan
zR?v5-lp}}c)_{i97k5t_@W_i920?lvt7{3%7_Pu!iC#e8X!r7hD?!HK#-{-yG28j3
z5O1&1Vxyob%cgrsrrdoSI9Llt-dbru;P3`Pn$MsJuu-HEQ)!sPsa0W;4p6sFx})OA
zV5-l<m2RtxNtDK~g6h`Cc_#FzrplMr@2N*@IUL35`F|NWG(F?AOKb1&%OD<|Du<)1
zr*H~#sR|hyStX9IS_*yG%^fX1gP!aZs3r4HR5q&o(l#XpqE1ky%aU|J*#!6#dx)#o
zsl~c;`H2FmCOS1$&qhvBaPJkh$7boB36w-&JK7*ZfmM`DJ#MmouU0B#pOy8xDvF*T
zUn74Y3LLA-FXjv@t+_~aAg?#q?F<6qdA75JkAZA4d+&vdFTp@0QwP$CGF|y~H(iXb
zDVTcVS?`^Q3Ma{0;hPVkOiYqgDD0M5e#w+&Eu-;dV7D<|Hx9>SNCz2;u83X917WuB
z$<<E?rISqwmr(Z7*VYJ@GJp!1l`XQw?-R>foW7e=%j9u-!MMC=m%hm`Ei|+q5zn-u
zjX%;CRo2$>*7f_3x=X3b?}pK{ZMhdFe$dtPhGt2InY*rub{5)(PcD92M_|O-y?kSp
zMK3zu`zMr`1r&JCwq9jQ^GdVYkZb>ivRtsDW24O#!7jzKcbj~YK3sLQKe0}<IuE*k
zzude9bsUE7)Hq0nhwwnZx7l^}8#3xVn)Nd(HV=K)s7;mep}s*g?7hej!+UfV5a~aX
z?^V|!wR~eouzeQ0J{%WFzH?|9y4e;Yx#N9Hmw^3EY+LW0@Yqhd>l+#-$s18N*y<x+
z^rzQKRW>isvK5l^(XdEHrU$zs>5c&RG!=e*0?4v)jX}8A>$XZ0fvAQgAeFN<lnyxn
zgXPOLJxTjDAXKopz9QHpCN`Q(>ov`!TsI~i*vMiG{$)Lmj7naS=h{?y80gHoKJByd
z2qmWgG|4cY53%M-5T3NC_p6zJJb$4%#2vTEYzrcMzvX1=Ipet-CmRmv*QAuay&Lb7
zEH76!1!c(y>vX@e^iPU=3OEg-p;RllQYU$S%cVo$b2Dx~qi2au_0=Qw!fwj1-;qji
zYdxDNJwl@_Xb*Vx$_x&>Jh~=eGbc!Y=Z<C*0V268K?QJD9`|&22(rGd<kqtT71?c7
zo(z7fEpggNuqXf<!;_kf_h+=P;^7`7{He)q@YB4f<j?e~nbwA?9@7El*L?n&i_3Ro
zU(o1Zhn8C*Qr6<e4>I-n0g46>kprZdE4K8v3T>4pC(TJo#tsEOc99If-&UT9eq20+
zicL|dW~G^^caek(AvH)zM0Wj=C~$M{?Ew-*9BU0F2K}0or30{j<Qc?cps+_D6@m>O
z`5hiQ0z~QvOF~3M1OTB&vgAlvND}V@r!goHzKEnY_m;O`SI%&Zs{R}?-VCcmvipf6
zTwxOBCxFs7@}r{QJQ8=SQEL54cxC~{@wd{}`ufAxJ7K@r0i}@Y=x>}o<@+I&W_-x;
z9=)o%2H<PsHfN5H`r?f?dZZ+M(hFIQC8;hw97oZ-+!K`TZz%8Z?Ob2wqb^4WE?#U-
zdh842m))i&z%8P{?H#un+M|=|--v4Zn!)F@X|o~tz9+e<86d-o0MfT&F0dnl1_T_I
zJ(x$*0(1%paL$v)sqSCfB#KLoEO@b#;No$124!}fcv4|Ly58nws-SE75^_4C6B~3a
z3WSMAiiahD<flQv6y~dap7Y`_F^Y}UL#x_TpEFWtW^A6X#QahTJ&576N{Cmi-2Xh7
zVl~>cC;p;80y$g&ol1H0VCEtc6?Y2d($7^C=(WA@Zpy)8<IuG&<_D$$2e{YO9CcGJ
z{9jroj`#Z*=VS}jwCAQqq=KekX0KN^EMOfv6pa?0-n_MC4u%hOMLxHG^~2o?f{9-s
zvSO2s?rgwT5t(59kf7k%bH}@&wcmgs)#^ZldF%l)sh(%4^X7z-jS+%$LeNLqlbBin
zlK>(BCIiU+zqUcd$Od}yr7%L(F|*|~ht12Ksai{YJXRTUu5Bb>4)cR>842xjpqk;9
zqZ@60O9&IBr5@9svHH5V;Hnx^Su~%U-^1*0I2Yz0LuNc*$Ho#I5<`6>KlsC~jK&6j
z?cCz6c;avy${%tJ5WuRP1xHri|BtS(4vV^rx|J4C5D-y1R8l~xp+Qm*=|)nzTN*?<
z6#=D@?(Poh?rt1f8fK{b8yH^wzI*S(JoESm@5{{ZoU`}ZYpuP{54y0}O)}BG*o^vs
zfH|!qQ4<*rwXA`c^oj}^#<k1M0}0hhV&d_RHNJJDssyEdC|*-K*0q}4B3SNa%E9kU
z2>Om!ps1aj)TKnOJy|ryWa)SID*8rJ?Rl=PV`6%H$<^5s`dS`28xkx9u_tSHp3O~9
zmpd`wn{s(|bxF7ZB#v6~hX7cyMkN;7F+Q%0hD9u6Wo0$I?f(MU{NpgP#S{x4C6E3V
zIxtzGAk)~bE%-LbvmU<>@@Vfa<7uO<j~ISuL|q!uR-ig~9gC)q!Y;TSvaF=N0H@$*
zWV(YzBX#N|)i-aZxE%9k+wWW{5zdsn-+uM7kIXi%yKQ&o?e|^gKj}bCK)QzAinK(r
zH?Rcue(%?%D0A-p1q=rfc?Z?0lT|TiUpX;_bJ!6~I4Fyok~v*NR!teS^vqA#$_JB`
z4jrM~7x9<xh+z#I3DE^)`=<GD6j9!%mlfxN+||`}2$U=)-BB-MIBnBi9GaB?fVdBn
z+0H3`dQxMqY|$&8%CF~p?94S3OgV1cd3bxwJsUYO&Sl&crrE!D_j8kGIff`~`6qxY
z*)~14tlovbYHoRtvOS#^!c}^7(wFmDV|KN2j~=^WA6^y+#DlC#_bb`A>FifITjb?O
zso2M$=u@0g%!ZpdK4N(eZq-sMS`bHvo^yfrBt5H7secQK#@MCCy?s5p_wr!qN^?_S
zUGgTUL4fNuz(`7i2}@}ecbIZLt_$Rs<AIh$BbsiMsMctYqfaYaJCPmX*i3*J6b34!
zc8m?gCVsV_v*s0FX11@w)Pvh!+i_Dl>$)bnc_yOTTRQ$G`yGE<kpj$Z`8NMgp|sz*
zR`EYm*qkR{VMZUH2Sx2_r_0T~G{oMk$*h~Y|K#H(kv`__e{=_tlmPTKB7^LBTqBWq
zTCjV_*Qxg27R(nc(8vdS;{8QtIBcC0HZpjR@~lS7q*utx%;)Rh&Bi-IL019b-|_q(
zzay5S{S!ez5LXo?xt}CI=Bt`p_RrdBO$4Z?e@DTyzw!mDJ+{IbyL;P4bf4m%Q`3BR
zzkk7aK1jwURc?MeD<0pXXMc6$F~0tEEh8KVC^j@wc&`u;uSwx0G3#LC9S>;!s0;$S
z7H{0id~*7-SrQy$ReX>(E0GX~p5yKtjNxSqldP+%_Oc9c{Z(t_)?=|ByI^n3$nv5c
znx}s)RKG1b8%YtTSsLV3tpj9#oPMeM7RTI7if@|wC`V7%j+uA%W0~Fp$_7z$uGRnO
zW7wVDk&7hGVm6s!;nQsS8e#ta?EA!#gqEt&$lj-Xl2Jy6(!Mnj!%;UA(K4<-Ux2pW
zkC;Fkf#8DPoRH+1RT&!0m3-O(@jgHTi(&spOCze67m)=Z*-beQs|>IG8dePeTA8uz
zA&n6yH#e_e(o|obn)1|%Ui&nb_oJxUDa5y-N4xnRBXB-%nmIw;z|+i>sbROYUsawA
z6Q<8K)yC7e9GmXhDP_$I3bMuD`pQSJ-i)c%3Dj6olX>}0?)zF`HW^Khx?RpeuVWoe
zsth<9@5Ao3cc-Dd5LCy+s@adr+%hY}#4OQ#Jtq82LcLnrTAsbC++i>bEk`kStW2+Z
zW2>F5C7H7{?Xnu7?~m*aV3Wl|i=+5_D$64J4KU3<1#-mfV=|8kt;`nHX_EqJgJ30<
z26m}$Y{gGt3*r?;2kRkoUa2db`y!n|pVO$Zk!&0Gz(-@IW{=zUU429tn0ju~AC1)h
zk6juNnT)XGqzP+=BTlu?P75KiQd|zHGaUu#(rX2r9QTf!8TNXD&W%5$N|Y1_V}C|Q
z%T$goXAo<)c?!H03&W3}b`Yr0*yT%?N9L>jxSQfWjz#)w?QZ;SeZTKHM`1EuU?}w>
z?*5c2Z9?pQVr+R&SUTo^fbH+sD%6Lj#7*|dEix;+IOaC<|M8im6}gt*3uAyq%ykf^
zk@iMBV6kD_et%#gt;NrOpYWxyyL)}xJ%RiJ#*m)R&6{jm=DWsLZ8SimNz3VmB+Q-4
zoI(if6z!8!D<E<oTg>Ydul=p@r`d^S**)rCWk|7XfWez76mHxd6^c2&JLY`r365YI
z9BL`O6j7La8EDUB`?o!lb)(R@4;w{9tkkrDlevc<q%!E_w2zHZ4D_;GMaIV*uRKa|
z=W}=8V20wB*=SH3RkDfSdI8MLep43L;2E$|*080>4j=OMS;WnPYO4NhiC+Em7;I2`
zeZ^I5n%HFngn3j5t>vf+3JP@czc$YV7~}!xP7ZToJd9SRu^Vfr<C4=cOLj4`wBI4M
z2~}U#`x~qj6&1TQO9T}oBT?IGiO}d@cGWw+S9(Z5&&#TSFCrr9=Hh6akxw$kdYVFx
z_w=P8F5c5g-A_-UC$zrlxL|+VR!)~+Rm<%;WjdP_ib^B^{O$mT<Z-+lEJwGrHcorp
zo~s~R^rO9;Xh723*`C*FkuiY-G_$#~sKvv9LP$jFkIi>&`A%M3Y-tr+(i?7V-{@NA
zDs@f?RUXkubhY!u{j+U8%@Y-P&02m`0^oPTk%1*(NrG4Qe3R8uo~ZXrPWIn}1zG`9
z(hc+ORf22PbGO37(QEIAf4G1?AF-QK;z>EjZ8cwn#CSNoP)|Y#!Jpw&G}22E|2P{b
z@i5Vsx-sQnvn_4MjEnn(ebQSLzaG^^L@G*FK9=EWxv~;EOL%F&_oSg5VdJCf3VIOG
zK~xk$@N4A=4`r9rdQb+~XDRu2fbLLPN7?s4P3d(iLqDLfw`XNhAG(!W#mAYH_i9)y
zUQx3n_J%11!Kx=CvZzNLlgin<8AI>&w^%Y0KXE&j+Wnv<zRk(n`d)ggVv%(YDLMY|
zgsjB4?-aQWPsr12<I9xx_e^HjbQ5~Mw+6rI`Tn8cNdzc?SbB-edbZCx>D*7kE!Jfq
z5`0fuN-=C=;-D)&_k1xfPwXLS*_2{y5bBBiW<H5Zdc6LI>fLA`D<XfSKAEQ15e%%8
zxQ}8!B=6!)b6(6pEOa=};gQSt&fyr<Ow<6W6mftI&fv_Mz%Z<#$C!9BFcum2>o}c$
zNR^W7g2?u$$8M4+qlJ}IZ1=Q7eXZ14y4!F)<ViG;bBwA%g?a}{_&fX`R+I;kxJ@Tl
ziY&TTa*HO9(9yf3-@L>auHW_{9HFi(*FI7)HL$R*_a2wVXgGb$k~qls76It6UjBUe
z8ECR5=VNb&=%~9FVc|y;)$BYZ$s6pBQl&z$m8Zoe?NR9B{=G#(U}sc-)q)n2ks^OS
z*1dTO5G7>v;y_G*)s`V4Y0m(+@C*rA{rtC#{ZByH(c?B!onq$D4>x-ROXf*wfc+{z
zjns)L)A6==Pet<mOk3#Yav(+kZgDCK`oBpg;Lc8C>wt2z9_wFSt0!UdBcAUx!=EIm
z<6!H5FdF@go{#P40}NPf-GgsqKvI66l)ClZX1|B*0ZG`)?&usng>MzgIfMI#xr{@f
zr^l-_|A1b-B|t@+`k&-WW6@7J5;YifB5J{BtM2Q~amWJ!LI&{JcQb)q#?#f?*C$GD
zuqTmb$*1j_Dij@B{ZXcYt3ns_XCmvs^bdMP_AWG_qHW1bY*szZ7$^m%gtkWz4$+`O
z-5c2I?5mlFXSe{-0x#d6^Bz>6<c9_fHXYif%#bi)$jm2bs2o&#i+jOUfFwb$Kjma{
zor`*Hrf}|kOzvgTW@LTwK&GBRMJNU-vB^L6q5vSu_>EbN6*U7-pTmd7Nig-c4-`2@
z76=5r{{8tx1mfzk1(oAtm2=D=S00pM8Qlk3UEnGbfL$Mwf>O;n)!!%fxj$9YdB4pa
z+Xm0nqmCVo56UdI4d*l#<xAq`(LFZrg{~f<#_d|)Crt-pz{W2eU`M4GSn&b}cOm)I
zDs#=heCHLKFDL0uG42qm`tyiCs@yJdDc#HlzjSj#a%H~!m}`+%=?xoS%jW%D3r%q0
zvUiwu801Zk2v_V~m<=7)*F*L$7^ziEenF!jTMPq3v9jovkZ*&c(>@cvUL-D<hBixK
z`UE8u&c5plIXlSr4`{wfSsKBG?LpY@Sw#r^p0ZSQPo0^7!dZ8VsR%aqGk9KZw%a7m
z)SUhFu2RKu|Cb~{5oD<7N009%yN#TSzpW+Le89x1#d-qSpYA<Va4&6M3%*E~R&I^s
z920MxQ3&i<WBOK>)-&O}@QVV2aN;0UhdbngxVwRaCm4-DgOG4exF!iOixZN~30)tZ
zm~4+N^|Ckken4`%Ob`2gUepl#N}SzP*$c2&WUE(LIOLJS<h1879{`?iaaKCc-@+UM
zF<#S;&W^OChjfUIX8-T5&c0VzTix>+@zV~EA&OQ+9;>_#X`c|!W1l;fEGU)tCKmJ<
z5BInZRYdywIz|^*=EQtW*sRN;#`aj8GujEscX=6iY`wHg0~^F}Ie2RI5*9Dcp<+#5
zEmKt)9d~Oko0<aqx+6ToIoIYx!ypA)Bc;?8j}=W<|77g&KcOVk($ae^vAh_!JI-lT
z*Jpc8NU`WpU$oR_d7o;(`}Xd9<-7?Jai08RKci(DQ{ixy@gKfEkwji)W97LQKE0CB
z`RQH^v(Bfbr6s%W7Zo(j1NHQYl>uLRDF<``nQ>eftIScMw6gYfatWW(AkM+g*qoaS
zeX60q<HH0w%3DG4(ZkH_+#%lr_RQkqA)49$JAMBSYptt6qfeDgGJX5W_NcgR%yz;4
zK#bCrFu}?fz7c`;GBI1#1w0ZSd-u$2p8WfhE()slGaa^+;9mLpR$)woBq{mSr6CEQ
zP}QuTx-4P$Nk=5Tk|}<#k%GOfLM;}nCYkpDmKnw`U%Tu3{2uk7g(@qU^c&81?TuIR
zL9UDN8so2`<wLs2>2H7h1ROB&GFt7PmGc)nj=<FsW<l0TKkt%T835%GHw*RLDuq{S
zS}d!hQ+`o}Cp$1GXa8$V>L`Sy{P)k$5}v-CkbG0^T8m3()dN5U(sNx;zWP4y2^WI=
zqMY1ks_Zk4fB#{lBg4VX?D0$C{RZ?PS`>f%T57+e!IEgUT_D+9GfwM&FuvK=)wXG{
z<SE^P_%C{Ux=?k{l0PdlRnyQ)Bdjg_xy{gkUAlQSx$b#y-{?#<K>c1!qhy{{HXSJ$
zb3-cTBGBmfp#gd=K-Oob2M)Qc;}-dK+Arh=VuygMk*gxqcAZRwvqs^ovbJU-@qp5g
z^%S0Ovc*I-cNG*BvuweAblFu1)fROLfN6#kwd*<oYyc@H^jQ0eSDafKh6mlOV#4L(
z#b}YO-_&|{{P1!0B{Z|TI%a&c3e<kS0BabmjN&JEuWUn%(L)S?31dQ%0oY>QVJ0o(
zMDHPr8}F&`xqoKp?>D3~cECA*_GzZl5-!b=at97z+EDKSUOW3D5b%pMVHN1M*MXnB
z=SS^;k5-39KI{g69Ll;Ev$8gDh^9x@zv-7pmmK*wrzq8Q4MT?9O{Vj_^}ECbZxI**
za^KsepIey@9>?a}a0y_X<&62y>m<<yK>HqH8`qHU2PV-f1H-XpfxfRoSaNy?NE^iW
zn_nTKcbY~Pw`tJxt^GU>tYocS3<{xU60_d_r3&!%RamZXz1pJAn$gInU?Wimc5VDy
z#nfBdXX5g>3jZX@MqgL?#dR4$T8ZIF=ieTt)NdmpJ>+9fEgIcAJ&{ip%TRfS&oM(N
zm1E2ilF#_IA$fkll2Ge?^riKq3qAEY1QG@wvlUKw-QTS-Y%3=CA9O$<nWkSU7TlS*
zKYbmmZvt^iqu2duyN@Rn2hf!XEVU^1o?e#^2M!NUe=)fyk88T*X^!gZ>K=AdCr%V~
zjQ~dKK6xOCfzdnKV@1jJx&@_FN}J1yI+ArvJY*k4OI9(sK8p5%KtdAdH8z`uEKaLJ
z4$#2ZOP;O@`))<AI94k6bKnp{Dxy}^ivE@l=xj!HVENw=qEv%3wpahp{|`bqsD4M}
zLj7M5LI9F1w|!9z<R%VZdK0_C_c@YvGf!n?Ws|W;xYl-c4t^(9hXJgsB5ZQeOsOGz
z&>h4%2Eb)w$>+>Y=hCwjDKYQbm?;a+k7p&XB@^O&+|?w5<!kd-GUi!oOus)#$4K{b
zKL^eZh)(KVWar-Ch-II5BuJx7rIOn{(R~#qJ+aTCoNP&WL($s3^+BfIPvD)Zag<t}
zmlk0X5?vvVhrwavg)1Kb_4q$zp&b^wVPiZmKXS5i$9Z4f>7w;cP=5gIz>Ce|F~g3j
zvgb8sWKHmlce<K5Saqc1SFv^5Pk-=?$-fjfoMMJ*7wE}YFQ2x^0c1@N5Bx`uvH|19
z&g4sV5r9{H1mji7_*@?H!AXSdo?S9*MS<9{<4oeR|Darq;6mf2&{SOH=9Dd50GMQj
ztNl6kLX7KqV|h&4k%ywF@?6O(&tepR62&M-AM0RTs~us`3(wO5p1tJ4LWL^zld4<B
zz|{BNFiXu_IAh>2=5qe58jz3LkXSHL3T5tKq5!e^b~J-_CrGBcLgsADJg<)L$4fPx
zi>Z&#vpz7D<5sRlsUHHp)8jeAfT?t3p<z0EUjZ=S^{K-4=-YO*SPAm|bvu3hZ-lX%
zgEg_Q&yG05IAu=9S}I8Ek5u;6v2D3=#=kFzaL)N7k%k~`)=y<p=>sr(nEG%$jtG@Q
zal@c<+m-N1)=(}(irH*6N{%Db86%o$Cz3!iJY+PQBN8>tChzZ)1#h3k^1a8g{atCU
zr8Hsn;gFzfO3vWnvq~+0E|J41n4^O1a!G(Y=^wT*XsI#u%3=sRjqN`N0LJ=r{oPjx
z?L(QN&*VrX(-O`<0v#qqaWDJ2v&F3e3RDWnIJS5UvAir^kSA3!$N5eb8HtIB?yMVf
z5XhNLNRX+k#_|dw(|>~{`d(tK{GpT4+`&3PYyIm5P__B(^JnUYD#+=ic4M<!`?Y^y
zyAm-97*!Yl$8d^|=EM^<72@2Eq|E=r3&2v&Lcb690t9oMLC@Bg)+qK@xhUp?`u1GD
z5QN|_K+(qe*T9Fi02kHhV4Y-S>V@@T`&GeT5+GhxuA}BbCe`jq!_|A(q1^ZZiW23u
zTJTh>Feu|ByoY#~6?Q(=L`%^B;|6p?S9Z9Q!kmxqWyN#0JNX6KK7W!B>sY&>)0Zps
zZ{wV|Gq|NIdl_mn>a3`wL<kxEzYzyO7@Qrh{|vWi>W!K&*1IJEld_PrvqLvykERt=
zRrSqY+05}ANxe18?p3$o{*@MqK~f?o^2>m9`dM+sr!#wfUvq<r<F$L!xs^?UM8I1Y
z_cfKYaK7BQ*R?!#+=;5F_?H~0I*&kf@@tG#*31I>3Q{$7@NjVt0S}jx*fW}Ez!wez
zik%pr_;@a07=y*o%xV>?XlywGq0{68mUcV7I8Kg!DFi@1IK&1DblC7R&x1wdeJ^u$
zM>-F%W=etZEtjA0AD8dxiOKrGNXwl;Pu0Yo6)Xl@kn2frBE&Gx!BIN-|H0qa?4fue
zY!I=1Hxgx~ikYO0d0Fb8EOPG!*HyB}Ar<tC;}KPRu$500%I~921oOxrd|fb>(l!y+
zMmk#O;;5+jgX8X~o9k{9TJ`T3)0qm}a=^rYYq*?K{SPCY`gdGz43$M}Qbr%3fTNLN
zqTqKeC``)%<TRL;w&E&t<@#|&ys93m>%HkSQX@*|AT0go?W6vr5kFg-Z#bko&W2ZB
zyg(bI%T8XVUVGic8Of%)hh-$6t|tQwki{X+J;3hj_89OzQh$$T{?R1cxwS1p(03)?
zvUAc|cu_Xzz0aSQKNF`;w$|b!lSgg~G6Bp-r;EF?=2G)2E^ao9+P~%&zHbArY(u6(
zqd`ikAFtBp(7e69L!QCRv;FQVu0YS5S}?GHYCXy2N+NhL9zm*CZLevY%}!+-`^Oj^
zZQ6gGn6*OW+@?9l_{6LjM7)J1QBtIvBq;pta{1EYGVT(Ad+W!7=J=$YHke@X&i31o
zql>6su{GF)^h-30OGM1{uvq^Qz0U9>>cytFX|;NNq(kMlE=Qy+nZ}NdV}Fqi4`rd@
zHak9;3^;wE+e!MyEj-0q4!Y7_FZJ4hIgnGa261I^tj1q2q2moxB|{?^fO7m62&7V?
zoxXhvx5@tKSf_=`PS)y-<pT8QZ~q|Z&BEDF+&r!5Rp_|%^ANeWD*op2Sa5y-Dsr!0
z#Y@n9b(u<0d0dt6Gx*~GsYx_J%C%&9hz1TmXho6)S{A>)(s!$wjReExe?V{4hZ6=H
zj(e**CeuZI_OC6!cZFq9n@&@=2W6?O0|Wdr?f^SPfXB#sFKJ6VpuC{Ndej91BM<;@
z2cV+a76<^0D$o)$y6}$gY>VZ+J7jg|#T|9PSb1y7{gWdAPy93=$HU(R25<ezM(&;~
z+9K(m8$C9L0>UV;iL5XHW*03Z2yH&MUr9EZ)5YKBGUsF`F=?})VeSm~7YoB4+zK!^
zn<W_zy<hsrrqEEldHQzUo+D1imglUTy-Hz+a|zGcOe+cDN&n-^Ol7en;>m*FJE&k9
z4Dgy%LTn1%F0)w4(EsRHHvzY7(>KUZdf`$P1U1kKcVbeS$G6LIo?!4?91c6Ld=cB&
zPQ7)F1-pjB29^#~!Z<n-Qv6%u;h59;P^Qe7NJbsieH%P}IeanIzj;UiJjzcW1p(dk
z0rovFHxxhMva+bD<8ar8o%6|$X(J328-pMEk+`ShqtflIt@C;BlW@u)AP@joEN%VQ
z>1kEan2a=UD%r<4{_baS37~uQ)Nz})<z>8>&8Vb})p+*!beN)=xzL^^%K!b}ZkV{B
z#*-Q><MDCz_hh|&kMj!M&HMD_T~QH)7J&JZV`5}(3ewBTexyUNM$vD;BlQ{vDvg~|
z;o4+c6w1VR%1VM+MV&y)YL3U%oK%%l`EB#{-BOSjQ~g9zaj{Lw+~!=Q(s|4G*K^i>
zaU^7kqN3TX-5R*%>ivBiB!qH??-jsM%X9m8#LZf^i1@@-w=<$F{gH^q&6FTt^h6Ve
z;XZ(NyFSH?YpUT_97@4~rv*k;>&s~-Y71Nf@NC$58cYX3yG{a)>PwBk_2)NK0-W4z
zpq5ahJ~C5floI>b07!cdYl7|PztWJ5Vi;7ntV?TtZOgt6^oDK%4F&eo?vu=bZ&(aC
zhR2*yQ3ajn!5UI7<F(Vn^YhZnZCKF-S~>C>OZ&2A*-mwX^9;MmKvhe4MfdiW5`0fr
ztn4~bWm~{*Ij6It%-7y!XWJ<T71=Uxi?;=OaH9IL`^!Dyj=;9|^K)0Wtzs*x#FZsx
z0>8S|CFa$Yc7`SVy+9b%$$+)h%b!W2e)4_1_mek%d^Gh{Qc{UjbnWkp1Fgv{W9gU@
zqBG5fjNKf+?8pWtp~c5FJrnx%v;Se|#_pxYrWYtvC4;rJ9H0(qr<^uAnt^`PeL0K`
z5=59Id1!dJOuGs34-Ipg@EtD$z%s8<LiQ<2)^lke!>13bA|gf?jDbx&A?D`nX!;~F
zBkDFvdlDhj(dU%vt~x6!C=jZ9Kn;<O{BkH~3aL%aihowBtS2XD)6l)b7|GCb)H49N
z%NZKo7czaUc57FhBSj;2n&r?k=-wOO^S`nG<$d{FmHhDEWbY<!X;1oK$(Oh2%`Hy2
zgW!mvpr}B5)^a}^+SIf8o0m%<;G~%DJ9qCSJB-z9`~mBK?1}q0?u7C5#cXO-w{1vX
zsVPy?OTF*)hdgf}AUO9|YUS0d?qGE0V0#>XL;stbe#uRJY97x$4searV}#IU<^#?3
z4U^7{KIXeHCuC-wU)*haYiSEWiwZ^Kwskb^&%Y2&z<Z37y!YcrdEe$&1F*@C<nOJv
zylFI=tuD#$2yNX4o;nX+d571zJ$SVWHo#bVY#j_5V71mhX?s7_;8Xx?Q~`DU^swUp
z#@yAyHsPFYZf$ug=a~flr9ySixe()%r10=ijg=)z0+uS~S<ZhkCowJkc%UK?U>t(i
z@Z-OOGU!j?HT)bq`I4^v7NHrBqh_sLeTVIV?(_dSgh+{>>>C+p_I)jb8m`UV5rrfF
z5Z?FBD`f>>xPxzB%T<*h<;vr@4>>>k)x|PDu7XVMPx7BMKR-48LTN&IiO-n-LsB38
zjl5L(Th$0jmi3XIk<=^EG+I-_40~b0ZVid<Zz7Vd4~4%-J+ekXLBWxg5rla3an7Jr
zk}sh2)t&c&huay!367A26G%L`9_*A@p(SnZF*1n0(Xd;xG7#=3?78f<3=B@<O(K=U
zhVX><9_$*dOQl}Kr4FX;+9%wNF}Eyrrz3iY_*4M#skRzGlo_g>$Vb(oqf0inogU~;
zJ0Htve+%n7e(%ctt3uwDJFQF|#$f<Lq6}UVsSNf`GTNI?)k>?qVvw#vzJtH(q;!;f
zQZixWW86<OAw<8H9<q;Fo!Goo7-aB67Wag|roIU%Z9IASTMK7eJ4ddG<5jWL^zcxm
zUv1Q<km-Y8Cb=&5-rY|*($jTJ2wPIqDa33(WHW8h+Iw;dkQgbFfNY-<Dfc~ka}eRq
z`{2pq=urOCZfXQva^U493Tquok=+aGrFamb>hBqbZNf;hy-b1H3g~i8Jw0#Fm1o8z
zz$MT>Jkor`z;L7uq@yqWXxj&0`^DC)#IG9fi6~20J>^qe))HxWMS96t-fKM?ycF~E
z`(t8^y91@8da3l)lPW|a)05)YfARY78u>StEy<;lb`QAfYuK?mL(gIvO21d?BJ)_k
z9p)c@W1%ZRjQ#i5$q|+*Wt`UuUma(df~hNlKQ|PV1|i-4LO2pYQ+F;hezzh*<PQ*0
z@cmYnYDva@?}WG~l2+r#P&U1M<&`GO$^0^C6|^cL!B>!pR6Oc6ADSMBVBK8Qm_gS3
zx>jl;7I_H~N%k@Rv*W>hN9lX}cE6Y7!yGM2dd<tTkVX(T?$|dzwG+7enZC+K>d=;$
zfWcE)<@=j6doRHN|G>!Ng(JISpMwc6;?CzUzjAstq9eZ}?{Gl$BE?H&aqZiUmd9Kd
zXuD+9pbtNvK(Te9+7e28b5^uPj2?amJ=}8-W<1@SWl1r%7Jkk6iKEqYihIx!x!P`B
z<6>Fp0t|gF`B9TnfDcL?0pbGzBVR|*HmHtd=Ree}^OG^I5G;L^OK)qCP|yDqTeoiT
zraQ>Lv1mbC%EK>=2tzXG-Bjn+#=RComj%O@%deaiHj&zapKUMaBzK4yu-7|eumksl
zzWj*G+p|3%9vT?!?KJ|fO_>_-k4L$y`7O7~Y%N%4?ZfPkf;C8By^rP~3u^#H;62$7
z{**U+jFyCkPx<o}D32`m(K2R*ppuk5HpC6SfJU-FiEAobyIo_onoX1XJ38<ID%Kj(
z3B|i@qPs5A*JlYhP6CK0Zx#f$tTh@m8Ka6>wV9Rrk8AsJSKT1)X}D=wLs5lpsG!Eq
zGBT|w>-XGRF$Za!*MB?->-%S{IFKFg>Zto{Qkvvqk#kusNy(X6yqx{xtREb$a>O&l
z1D3)bp&ue1cfL4^iC=vc?5D$XZ#fxD1}ziGX4ET!KVI1B&C&S^<~x---CZ~5z}h(Y
z(Yq%N#7SSnE?k7qZ69<xv5d<ye0(?+@I;$uYLO;kYj^u%jc8nFA^76F5&W?%Bof7D
zE&<UeO@u`<;z~b5Nazb-`I?Zi#kBp%3nC`z>4gk^W0HD;+R980Y%JJ0Wp*a1{CG^+
zm%v5c8#_565E+4l>GS#$ON~YSF?5oX!AZHx!fbAkxP^=q%Qgt>Ys`-~0hnwQ*ynFa
ziGjMRtV<p{6Kw=(1K~)(;d1DgJOmU?11Vz1uw5bE_ik4EUa%|(f71JuK>h7BiydJk
z^~MP$ZhkUosTiY(U!nrxF{sfiVl*z7y{-|V{MHKq?KH`6U6Np^%5T-Jsg(N${Q!k3
zs5%jRLBj=jI9jN(DfpvqQNwzEbuJC9;I|SMy1<}cmsnP|pS;(kH!AT=`c>V|AU|9|
ze5GMX@L;%bAPfERREW1$-U+Q8>*>$psRH@@_d#Ak1hK7gN%il(eN$?7(qIUbe9Av&
z{J{fs&w8O><uZ7+UuF2}i6nPM{BWN!cmx4&7@;XlYQ~8;vLXjeJ#W*r<ny{*ixd`|
zffqSHu<s(~d0#>wy_o(H%r>a1%b#1LraN?svlK>pBq448_A$RCy;i4?b5`$O;~rr4
z03vjjpDXFe70=sVP|#K!=1-5+!D&KT%k1DeL$geztnMfZmtMj>`l$CWUh1CjZt#d3
zWa3^)b8Ym|?(;ybTnG3vu}8%C$WC2zFLvH*UN%+9gRPRM3QCT58&khMj5ywms*Cnv
z+w7Nck8j3pAtO{<MjcJS&fZ?wbM1vt?q^kvMJ;<T3pw(A0o0KJsie1VNqjP78~w6F
z8fPHKa~{Uj_uqa~zpo~LW|I3){IG^OF0AMIC(HQik{<5Rr?77ClkQ-Mp-n^QM8zYt
zjvDe7o%R<`liXzZ4;y+@`Q*s5!nVc93bdHSMhXIATu3j!VSlP6(FdauW($rF592KF
z<kR-)f2KDZy-ZD})8T|aRN%P|pdxo!ep^0eQfsjI@DT(>MJW@MRLkf(+<H}c)^1E^
zdqi|^8i7|m!)EOD-^2Y>n;CI++=6n?4{r+6d6P>E_f;Rj571+ngXW%8X?!#RZj@)p
z9#{m!ffxTVN0*Rf>fDIdsMaFHS<d@Qn7AkG@mOwvV3S&*0&c7<PHt~OUjk+@qeX0$
ztEMRS_4y-?L3l#1!)vTlCd5RfQFz*D?4db*_9>Tm{F`B+&&dMrw4=5A)vr0h{~u;G
zV4<;qw;FvB=UnQM81-MiJl`Joxk;|i8uA@EXcR4&y@8>Jc4VhNcv5@XU<VF3XyMM{
zyekI+ef=%LLaLQ<@=trEp+^EAK1ctvFnh3t*_#)dJS2Z!@&JL1Y@`(2^itw*D|w*A
zO>ujh?o`#VLSaAqM{hh&hY&{%FDx>*3tdcZ8<RaZAp?@@?xy8bUdd0bC)gg*8@V;b
zdx-bP!{O=mVJ0#6!&zs#d&&rsls6e2mekoPt@1LFY5k;4OvngGLJ}@7%|~LQ3fQ0L
zKSq!y9#&wrE9QQ9JgZi(J9cwaZr-(J5TBupTiIyEr1z}7pj!N6syD!;^7wzP5J@z_
z=^KWhuTn?zE}#?5hL!Z6BN>_l)1l3?9}|F64$Dm!i-jX%K!W-LrCD2UqROh&P@se=
zW4b~q7+R-O>v}621s;lu1h|!MPGfr4Y5bT749F|htKY_jh?{F)Tu(TYmGX2J;5G6k
zO}&f?{@R?gL}Q~-ct=x52Ur^v)Mu1*$AJ^=nti+!W?#Bi3pKEE>@zp6>e{WFfuGlB
zN!Ou$f5>a4!%m~O;d!rg-*!kLw!Nb@vvG7(yeSGI0h8uuZ*L*`kybhYXH5|S7^19#
zd+zl`h<Jy{uJ1>w^$2R`rs^Pl%+WU8H@8Skyzvk{u0h5>rJaS9MCv--9=X{AI|tJV
z2M6?sICp*it_PWleHku+a)(pw-SV&#zvYPO7;AMF3)439-pahmx^u5?r|57ehWE?h
z(~Pv_oY-kMh`micNY@bI6Tvgnm^Bs`P;RxoLAs>@)IikM_p=<3rEUO7huy-^56aZ5
zn6&uB;60yda(d31pJuT*IF)wmxXvd#AprS94gx(E6cj8hW1D?18@x7FnDtZ}HmKeZ
zx1wm9J6QwYXXIS2h$0{9ta-RMfZyaa#h2eCcuVa5c&v5ZkJ_==J@;LAE0W!Mt<dGE
zqgR00i;($oj-d;C_c}cTLn|;s3Wt<5w8alw=&pZBMTMo3>I<jA>lB@(i|>ZY+1#n8
zEveLA{P&q9jou}mCkxw`>ntEDwWH-xDzGT4p8iRhw!tjNV(G8QO-I6P^?|>!Cl>*&
zArRc@x12H;z&V9><PBB53POdQ(&NFQW1I#Sbp2<mLpN#qb;ohHH}X9pn2y&<c6HiL
z=p&WD^91mlzcT1gF#j&0f`J|`Q1c9M8#=gb-XGr^_|CvENs)KFWuz0jj=)#Q>>+@G
zl#S~ioSD)*l4!5_=KvPK6{L>``JUD#@fjI-WQ75nSg&TRy@!Rsit;FUO>0s`l^ncH
zE2E;u+UYf2WlPNmS;5<yt|(};Cv!W*dGtbUB(+!#+6_<mDj`)qoV6LjB4`I#T$dlV
z!ASh$uA1R!P6`dYqnpy!u2#U05ttevM*YSr{4gi0KoM~-pq!7`*=1H&8eb<H8J}eL
zc41#^5l7X=(n9P>dS4zF;jC#7dIcF&8AV;c0odVYw~MADRb@ds_?(qu5*5Bvfy2oj
zj{P?OZCI>3Ohx|TGdD1XB|a40cD$LG{Tez~k^%X#hRnJ4(X&;2D|bVtz-~HW!j@UI
z#$Q!o!lJzGxxZWchvHbA;Au^0gCY^lm6Uw}TR~x_N%|#v<$GCLchhZRX@+A<Xm+=<
zrjy7$-yFA72xM`g$-8*(1HFXAgVn8Iplh#43W*1;Nk@i%vNY^I#otKv8@EzkH9*e=
zJLN9~;HUsiYAMU9p@Xf9Xc)hoPY9+!Qxg5Wx2AhhzuuxP0CtT$l~J`u^K0M^?NT2-
zh_I~E-fO;W6-Gp*3XhKuY4SemcOv7my9>1JqUHple7S>ODh&goi6^t#b(-ZYPfUNN
z)h*T7mo<+BP-&yGC?p47eMM?V*oL98n<v^k;Du*wXtiONu%_j7ST5QXn{5EGXt;#Z
zKY#mO2CT-y!lG&(8B=Zcq<B0RXWH0!7DY3Auy=!r>2EMbsKNooehsPZvDGcf<>g_f
zeI{Zd#cJeog9FmGDBf4A>mJ7GpuR}-0Y2k@aIzH3YwOe&6zwXT+tZSUDfhFj{N1nI
zwAlxzz%U}OiaF<;cOE1pB-WSb5Qi2ol4l6GKywT5^*GC)YuPvp(#|N08bdk9y74d0
zCSqJ0CI$bNQ%1CiY2)2{!(PHtdyCQ#@{Z{u)lpcv*VF`*%_#QyR}0?Qu$f0is;h+i
zKR4Ic{jD9xAFgvr1SEtW&#<?2l=luy+J#TgSxpm7F>Oa*KQM*O#@<bR@fk;zWpMC4
zY_a(ya-nLuVmf->ntMB!)-mj~)L6cCw$iF)Uqb%w00i7h4zJ>`v7p+E&kpW3@zkFT
zrsvOIyL%(Y2ncjwP=D~tajONI(^aZ$^zkH?lwnatfFB6f)*o171;kYl+1n()OdgkG
zyd~V6+$#P^@%IUa-?0L{L^3>J45GOywg1!?)&j%qPRAQA6*KNp^wx%3)aQk@hCf)b
zn4w`88q7iawk)R9z?Wa3Uv}~%|AIH6pbDD}GV-)>uW2v2Mw%0dL&=h!0P0|_<@?Z<
zv9Mp|YMnHht*{#sY-T#1BeIF@^a6#bLh@pJap1n@FpqG<yABBkh9Lo7ov*j$v-jny
z2s}4VC_%qNCD|Db2qbPx^PlcdF)cli-^GEvaHvGrwm-_r;re2cOkE}wwgU(Z6&NrE
z@}@E;Luvc)h;m`sJEjkMMyE3`--N%QfzXE!jAs{w@7)nEgAD;eS#;f{Qgy?d()=Xa
z{$zKt<pJXSAqfNmDXg%pQH&RP`4aT-L6ULIzMv(|mYH3-vXV4JQPHxZyMKrtcb^%r
zg6#Llg^x+RGT51ZBSOHP#uAKLf?WmetnPi4Yet8MCmfS+TCQvN<n2El>pN?JG&`cj
zBH-_Xm!Q5ObU_S;15dKdFju}^i1?KNB{{)WCg_R~*Ii=x-YwK9c4y!oV@JL`2lFpq
zoXi)<SIK1w+t}Dlvj&y{$NoDpu;@Q#UI01-iaTyl<~>@|&O2h<r=eDPmR^WgeQqhG
znywd^PM~+N9{_|PYzWkhd#zP|c=#~N|8f*D(fuyw9=pHZc!P&W96yu(!6@%*At8zn
zp{6-hjLZ*wWMy@a94<%}UR85-TZK&W^X4r>Vk{+1s3^(y2i;;@_hcjP02@H?Ec2>J
z`iwc#E%p4X=T}K(;ki)EM~(Chxp^A-N?9DtG&r(Favmh8)6KTc<c)7hP})7xj09}R
z(L#}?5a1PShZY*}-G4=5j;BI8tLku$lZzG4-zW{$-#y$M$F8cXvgf)0MlR|ASI%u?
zqPj=|p!>#^Wh7Ox;Qv_z+0uX9ZkOWf%i7ql;skQkGFvUJz=sLZaUT1Fi4uVTL<T-6
zP&vqv9i>gN)PY+jF{7>5RD5qKf-*tkE?^)GQjT&#zWmrkksb~KfnhjTjGDNnFmfmg
znknN!%m$j|CgW%#gncf_a(Q!xc>kG}DAG&9Og=<_dq6sBdH~#P#7Fj|;hAyZzO<x}
zGDk&vwWR35#WiP;+B>&!GGS#WYnMOho{c%`v!Q#}eozZ_JLp*C?5EpcO$Ofv)j%9r
z=yIL%>8}sQU}`<rMfQ#^qjNd-5zyV%wl-bBpyo_O&#>lTXU7^ES{Mmz|6<topPZ+J
zu*p=>VPrwlbX#W7j(&-UA;?jgTDmm$vsHMJWA6{lKKcyf@s<T^q_D2Zx5k}|02<mC
zc<-~*5WI_=9qgl6kt#m{lbLq+_6F65CQA%q`(RtIK`vluM4@P^UI<mYahF1aybonb
zgPidSWvNttigx>(g9M4uXDs3?&PV#=`9bb|FuEbxQ1-RyndS@CL^$6-hk9?*n<7Ph
zzQU{%m^l4U++{-6ZQ%CW{B?_Abgf+0xy~j7ku}{R`scuet9*Gqb;cQsUppmq0E|}x
zW@uF^eKyBX^nr5^hGLkuJ?MvCK2B)zYlLV2fOqrP<lBcR#FCvjo8?8t0)5l+Nrk1%
zM(P{WboRY5Y$^c#Ah=nrtOG+Pa_=*o#4tbdvjPsI&zCk0Mmq$(7e9^Bt}&)3m-Y;o
z6$KTVVQ3=mECnVJQKpTJcCyS&H}9*dgXF~cmamR1X_pSK+CNi^HvZH@bGAQv1_I9?
zQS(s~*en(ZZalx?gltJVIDdr6H^fK&rpEu>!MTFwQ&u$y%`&UxL7%CzMRRa)pkZXp
zA8y3vKYvhA56$Kw>Qe$T;Nxe&L4UGpZW=2Z{Y|+M5?@EdnDc;WA<xeDEQyaRL{?fY
z=L!HOPuZQskwK&&KurT^4~nDke($ONmcXu>iTXb@7Rm~NEqd3`gyTHB8!$X-uXaV+
z$8`XA!Yx<KyPL9si+`pf%l7qUW+v=fkuIY8QS`+fuBKYXN<#@9*HDBEKVCh4-Tr`U
z!;{!<!9wKOlH<16s=4vp+bV|_sEnF3Chkh6$_jr`ib_S%J1f$SfvH|?x4*SAjMPfn
zKPK6lkMj?_MJcZh&g!IdYc;UFfbmpMhSQ~TY}<Nz?g1frZ8V3V3Wfqan_yjl+B8xP
zN&4IoMbFiy=rwiBeN-<s)8~zk!~aslYT=DVBP`;bU3jLuYvv)}I22#Vsl$>6F!8No
zYix#&&3^LLhXPfq@)n6fe5!P5?Z=N>8owiXE+*YBH!hde@hR;}TH>P%VpDAFw-RrR
zaT?w+`Y^gMVVaIfP6|~|spvNAwPcU>+*8c_^u**U2Wx8OHmI9y+&PlPFaaNU^`r5T
zO>ID|HYyH4@pghSiFWIszGW8KSaVKd;Zg0@%wJ#u!viyia~c~30L<Lq>fF5pz@1Vy
zV7<$ZNSMDuhcP7<>t9I>PodTMHBJupiL3P+{6&18=XW(~G($;b(PMml_*cEjE!!WN
zW(Eae78j@b+^=p2?inqzTOWFqqx^pN=+LgUvc`17ngoo(O+Cxd@?s8I>7V2B;&syb
zls)mZn}Gar$F9}keGHQ?)hjYa&JIszE|yuP_Anj?^D&mcm8J1}=<_zR)4IdD<69N{
zi47o=RBCF3BF)DQ^hxI*2oIRKa8|~dc?aNRW@ZY6k@8wyK##b1&w%-6yIUpwF)Cq`
zB*1Yle-0ey@_HmuIge}Hl1QMwsh!dIK}(t|LdxoS)j%TVMb*Ik(??MGlT4Smx62=5
z1nn2JU!%SQYd*T5kIW&V)H*hm?ccTZI3v0QX3>pj=>ZcQK>)4~8hZg;kZ~g+I6pit
zmq86S-s4h5_zQ+gQN4O9mn&UGyN*q7Bm}l>^AfJtnvzXz+YmXAGdkwq6dWY{llL;S
zmkfLV81BPzz7bgD+I5e#EqIaO=}b<*3nGmNu<UDN_n`tbN~%hFl~~Tn=bC%K1qd%K
z=W4m?N(&1cBk7d~%1o8<s2g{Hy<Ex;TVHl-s$8a92E;XHUS?vNPFqw<k+DAep=tYB
z(^*p$a^Ev6Mla_Mu-70DHYSE-A@y-;TSK6`KF_(28R3hMcU2cZZ&^V@qvq(!Xr80`
zAtJIWx}b)7H`ixe&IK|_`&9TH5#Pcrq$7DU1Q-uzlmZrdBZ%{r3L6-2e*;)71q<@c
zgF*x7MH}(jOE2PAy-2ZUpPu9;{9++ISaGTgGVRDvtB3LIhifpud=nTmvEY#DkUEz6
zGoNPF$NPTu(xZLS7EjPUv|`Kc*;Q*gaX%i7Di;AG>bNdqxg8nH&8H2DFAUmY%mQp?
z=TvCj;0W;zZIb)Un|Pbt><e9=MOI3|hNRQ4gCPl({lP2Kkbm=buR&P^xTW7u&T4O8
z5h%hJhOu=h=JJ+^u*mp`-@JKazu^qrvI<NWwJcR3f-4tMmf)?sIzb9Y$;VNrs+0nx
zUr&1eAvW9gdBj5lh2GSElWMVULj36CQ}$i@ncpWQZ(JOF7m<yEpcl0Y-s?>R-Z=^L
zNoZo&p`^sTAZM$?h2US_DVpfc_$NYK7zYx!m_-h2=lQgDGcquFv&eY(x!GLp6DKD*
zgKf)th^S>72(PU%AGZFuu=U3^(#v~JNkPy=w8Yly@pOPFFh8noiljXDZz!9J^DJcB
z@H&-`aY%ik+|)Z)tvFR?vnr9RQuI1kGFe6&dbs|b*M;4n=erm1q|3yK_9-c?7(gn(
z5mu#Bp{ur;*0JC=8!-3VRHyB07$<Yl(t68vrRnh-C6qdp8rBI_j-}DpDbm)29Fn-4
z80DL9Vew)~zBBn&=P(YWGXhmRfO(RMckWOB6E37E!edqaKX))}pMMBunn6`%OS|hH
zjl2^iU}snQ`0h!8#|36Qw_|-l@yD02Gq*+s0F}@nRCF3;5el{jh4Vfc0<lw?c%aOT
zuWF!ke~ulB*2_&Xg>>Wv%?#OwbzmD|A5FHDXX;g|sXd1(OCq7_9hvnj9f+5dl>9>X
z^&@kC3=E<&j&weukED@`EZ11eAr3~Ms_H6LY%IaO#TZ_XsKZa)Si2-rok|2Hqnq`P
zHXap}yr7rK#-d_7@SXVFL<JRCpKhq;1n!x6E6{_SS|;^B^vRVRw82>Eq|!=_m_*km
z^c>~8lkQINn$&Y0He+giM$-G_DeH4gNeA&9UT)Rz%X?WV1A{-kn>Ka4{wBTw&AU$x
zR~IWHICUM(a~Z-9jvvx6ym(uFcB_T(AA~;Ib+V)5Qu)qO_COx<CGZw3d&6{j4sJI3
zjF_#oCO*6%{ttnC&($KhRzf)BH!P`b-`?N*<kXI(z>@uBv}bKJ7(LcU2LJiaZ_olL
ziKLX8q=U~Ihgz>6D6Re<8c4bjAnT;tGTonw__~h)YA7N2rH`*$rRU3~x3*C4?C~(0
zKL1ZUV!tGEhuBpZZ76mqFq;KfEOECHj>!*!{eJ5Bj=Q~cDV6p$a$9p#oCzvhmcYGb
z{}3JxeJQ>_!~E8d$CLU%m{hlvM>v}oH$J}qd^JI2g=NsT!j0<DUEO;!T={iK0u&Y`
zVjZy3j&)?)ynut{YC4xN7sDd)c{zSLeo*&1{{Sjv($W~o)_uv=?f(#kefj*^-ead1
z@POX|a~!VN^M4qQxrGmN>A<x>@*&7QJ)XcwPmj^T_Z$6XSF296&=6OXP@+(>02s{%
zD;oA81Eh)AuSF|Udo%_c_HR%{@!o~nPYSsHQDl^}Gs|rAF6s16?x)g_VyIU~|4P%f
zfFjQS09Gb$B=4;yhvP6ZaS9#U-q0Rt&t_;Rzg2D3G32T;>0_^twcvUQw&)n7kLG1(
z!;VC+Z7@uG2R=deg&IJw1NnxZ_ra@X{imAYw~~jM>pS+(x6<4jm^7DaOsStn@U7WR
z|64i-LW-|;cuIW0VdyMO-M)EwjZs@2Idj&}Mki%~&S}Yzow&>0h19az;KatrZ!T-r
zt+=x?x$Ruvzu`zNRKV#CZK(NtPL?)Y=Fspl*RX*@guKVR`pnR(35g%FBYw_<Wc<KZ
zftOv^-hqV!8Q)8R^--IxS-aQV{8fE;f+gG26Vo4#;D==36|B$Zl}1SQNO&Q#Suj7-
zak6E>&PSv30aU2`bZpLrVnPAunGhcvI*cY<>%$rXxb<peKwv@}9`gHsD&N85_=b3e
zPMKS{e#>y=a;c6m5sm>0>;VYdzm8WKf42hfWx%#@=@a+SAMTKXcR$6ZIJe^*`nFr9
ze6T*8+aO05_XoI#=1HK;cXYT^iHf@Mt&}}9n@IpZjWR#V^Patdfa4*2K6{Q$7UYEJ
zyWwH($qT0oFd+%3=SRfc_%lyknO6=LJLVYTxD{z`B@o%OEHfVXZe$e(7K)gR?QSca
zv}oWxUywO0l2`Chx_X?o?YeI(W%vz{KOFacZ|XMgA7X&*w^Z7t(h5T7etPr|4tVbT
z@K@`CjDTmlW5O`C;3B#HlXMPRDRcSzS-x;p20#$jeAtqVnxgyCXKW#G3FP_4D8ftY
zUw1z)eNXp@Al>tPKX37TTFW8$?~=cbY+S81ooINW2rQWX({nfTw`Db31n{o``x3}(
zcadT2LvZ)fyax2-lgY>f+<&+}+)8Z3*W?;=X*s&+H)_WjpG3q}nzfe~O>iJC(UfT+
zo(6TOPCB^(h6%xsPv0r>5`+bbuZXXDzPE#u(ANd_!_cgbUVE*9L5wPQ3JRGa+=eHE
z&5cRI!oM_eB&}>petoB998K%AnrGUX&bG~P@~D)iKQSI9M7}O`lXx@HOD%apg}nGE
zV9DD=z0X%GtSNCT|H6uB$b5>DsSey~uz?w~gd{TF=lO%x)EFQ_6&3dY(wAp|6cCh!
z_^?lTQuznOJ;*m*DJa8IP0w>mSLrGz)hl8=#;5RCMT8ertP0cB{y4D=s17PRTVy6C
zo=5Xsmt`>11W-?Y=92mDMw+3!#OfaEp+;C;T~$_IGiq`qvc({<`}2d88_R3ol$?-f
zdJRv`dfzV-O>B*Q*7a~obV|%R8FAV0L|<=&1ix<G?i}iqZaZx@4_C~AMCSU#*KO((
z3}Ah#e!HoVANbnDJ!_&GCqMKtN6!!XiK<WW^b?#xBL8t>X9HfnXSi0YYCqPrUfh2i
zJX*r5zkyDeu6h<J@hx3Koe_j(+v53>>YhBa<X@7~i~YHk*`_oNeN?gWJ+<nGq&dnJ
z)<J6J&Ts`IveN0K$$#&ioX^n(pB#rD_~^==k}Q;L3LB6xU@4=a6^xXKoww_HYP@Xy
z?O~+kK?D;>=P0o0#in?%9XwEl!zQnHE<Eo@Mo)gJs!}YZwmq4Z!udKqM6^AW<aGax
zPt)CY{bBSw>-BHwFXDBXAHwK(W}4zBTvn7+*hOzLM6SBpx`3OXrM;SmUSs8J1(hz^
z8PUi^G2nEw{~5ta`9vz}7fON<|2A+*V98L*K;#DGvxE%;@B`?!@^&^v<lGR{l;g6&
zq7HNq^$EGJz-egbZ*l7d;4GA(QESckDwGT-0)D02ZJCAoB_82|(bEz|Hi8>A5(CZ|
z^aTKfbDc4-B~s6RAqWI%y`7PAFS=ytJhhpa)9*Z~kUiSFyOSM`E(C&r(LfqOOKhRd
zl3nx*=Y5(FqZR${lk&Tg`_>$;mcN)O_n9t#ya@cmH*|W}ny7ihyw146HMC;Sw)W=|
z*5%1C&P?_D|DPX$>nv1wUdQ4sv0VSuoudNe=cwAccb=nW5}8V!h2I5Y84y(=Yc0TK
zwjkVkyvPcRpyMo!nny?kOzfd=8VpK{Q%iCCmUU${C?Q2u)L5R=;(i9-S|$IpXu**F
zSAk#K+!K&rFK6midHEqxn2(sY=A)!^XKy3E2`u89WUi_Y7RPt0;Jw(!pHGR!|CI5u
zyQ0YK2|UaI&*L$I*F-AuQ-hVj7e_bgS9D<GH^)nODgQ`YKc%swoy^2tr}pdL==DKZ
zpyO_|Bc+ttmoh)J=V_@OnfAVxUBxHGg$8xsc!dXtaj|*XW_BYBg>O?MkRY!i<0fb7
zFRcfgdsf#K`wLBK`rGtb1;czAJHkX|RB&TKnKa2V>iA_{Wy7<Rij~W#8F18J!DPg1
zY3P;N3|_%rL3851V0eWdk(0{{t!;QcPDh9fBCXRvHFBtt{}b5dkv1#a=|v2}GP>cO
z8b8`4<(?AavU?Ql*^2u5rC7?ybaShiD*#=`*L8VWSmNR*1dDR>5ke(^6#WmO4`?8c
z=YLy1D+mMpVdGAeM&)j{LCf~=U(mLL$)9!=J%#tN;6}aV3upWGxBwx@Mq<HbEM!CL
zik<z){7*|mM0CBm2Z#~&voZfD<XDI2c<<ngJjnj|>zY=U#`Xh9aiR_h=tpD}Jk=`M
zJ8s?Y_d_BwN_KuaeuAG6Ey@L1qlQTaK%X_>-N*`q-YuOuprE3A6#dGkRl;16?o?E?
zs|Sz8%objL>n7Ko(LXr-&fTN!n~iD^c&(#vBzD3^CjwKY!k`9FV1N2Wjrn==>r+za
z$;2ccZhCdxWoqFHWe7pupa_qS_N#Y0`-%<QL#3Rsf(!D(s0Q^m;iB?z89yd=u`ie;
zKeCL3#&jMGSw7hxl!lop{$q-9m>0o9J83zm79L;I#|I07{ThPS4Tfm~1DcXO?*Lx1
z-XDk}iF3gl9}ctkgbQ9_#2!nyq}9RH>7zA)P9N8&+Y>o^wrwI}DH#nO(^4yCR=u|g
zH;Pu8W^lSF*1<Kie<UN8hY5$A?4lq4UtA4#Jb(;P_WPaBm`>C@6S>X6Y7?;^_MYB>
zGlpx!8Q2M~`)`oZ&{{e>1B_tjFG&hE=0U2kJ7uU*S{EnMTGcE#R-WmudaA@3dzO5P
z;-BSIYthj&uD#k)@+jBVz2-gH3q7A>!HXpKeMl3SDlx3FT<hJEYm?&MWZ{h+aCPyN
zEg2sW!zRkji5zxj&>bBeVSO)0nj{2|{hKps_FM}nkQNC+_2_JSIe#}IxfWt4_sOEk
zj!2Q$jV$oCK7>u68}X*ZW-v~wdC|td^7CLw^RzC<vcb`Qm!LlayWNkKDZSh4n4%X?
z_Bs>+1bb0gBKBE%`qI)QuCcp~h2>|rOjgf|4odNbIb?@Fb0NTOB4ejN6BEf2+@Z!i
zCo{GujGpt?#})}TWDP-wP+k@Xz3T4Vg979IKj~*ler5O-?Pl5t)Oe`OI(M$$;Bo|O
zD$+WbaWSK19Wa}lS7BHjkBKf)0^@k`VXC&n_LL8BrKic%OpY?s-6m^jXs!QUVoa>D
zxP2!ZMh5eT_su{BmnLpetp~UW`-26n#ALTdQD|mAp6L&8%WCkj`21N=TmRm*Qoxb2
zyBRawaqBL$>nouesLM$EUGh9CP01|Gw;oEu(Le5gL6K<k$7fk6&<l1D>~xW`fFQj$
zejl)vOd$57>6nfn)x0V@h<I6VMf72mnBdEJxIhH@-N}XKmX=Ko4mYUX`RQ_)6k`|x
zEXOon0HJK=Vt{4IG&1LBc#Ia^aigI>6)7>_fWJ#idc=R)tH3ngL?Jb)$aq()FT)K+
zn68%%F%vX4rgc8v3J44g<Wd$E5up$;q*pj1IEEGE|A(xvj*6;nzgDCqq(QorZY88a
z>F)0C?vn13Zjh4h?(PQZ7(lukhGD)zeV*U*u5Zo!3A5&$bI*P6YwvySOA3pRGwLd!
zCM)_5d`<tTqBC&$fKd_?4KWl$>`FSDr2klrj(0zI7Z6glkQo69|9T<vRE*Mf6#iD|
zt`k#U8y{jdaB+7t3`_p?P%Lrz66z2Q%zd{<>zT$$uXOL*xQPk?*}R8hw^<8_j6@Fb
zu^@)9tNfy(qO^4ZbI%w2^%-^_+rIn9xM;E)wc~2NW9nGTWE1>u!RY@E0VX8h0EH$F
z+s9hf<bSP;fqDow`SLsC8R3xl^4X%}``xSo$e0+5N6=qP0XAN1!Vu=;K_TQs02#fJ
z+L<(bBlWMu^|s;7_WOd3s|t}Xd;{J6u*K;Fs#F0-4Fzri-p+3%))K0N9S<y{+!$SZ
zzR$ek)Xu;ymhGEJeQf((NKO`W|AS|JK(Z*?*>Kb%jGLmPxMtq+dQs`0y=7G|K^121
z=NE@_&yEyKOvOSy2m{mqYb0q3!ad%zrqV-jIg=Sg692HjtkK}`ug4*;lo&$<3G+jb
zU2&?yTtAP}7;peMpb|A}2SKaZ$BrAyLhpIkouJumRKlL-{*?BfB@BU>H-a?4rQpi%
zGRc*Dkf8^z#eW3^Dh6ymE<O&tDD5RWS92cuq@Iayys`i5CI9uv-RJ;hi}D1~)7Y$`
z+r--+t0QT)E*P4iNQlz{xEZkf-hU(JTJm|pE%GVZ(r1zRKTk&qs=;QptKl#j@N?|o
zeyBmS(pAq#;g|vV{xejK@mM!RK#?spy4)G-fy`|n@4>DSgKS6z&!dwHyn`*SZ5GE3
zOGC%e!AlByfhGm_d##aZY=spNmUkKV4**<W`v(9n6=f66{`7w}P-FLzM+JKOulaGh
zhmMgkSTJX@Md=`u$Aih~K(R{aA2TGl%-uQo)=vaCa@cgXK3v%7Bgb*{jR-Io;=w}t
z@#<3tk$`o#*wjST9}i^i)gj*Up3$k1-q=IU=6uVt`sUPn^yJW0$rd;qaP&xrCX2#x
zK>}gOAoLi&{g<NObOOPZ>>#|$#UKoXdH(H&VEg%g`&M+L{WWCHUeDT^_G#+{gmoo3
z67BYXhD7RVkg7mE%@zM=lC~_uZGv1X_UM(9zXn)aQmS(ndP#0bk??oEfWU8`SvJu}
zag=gR+MXo7ev5YeKW50~$*q=-Vq&*#G|Tn&Hfd5wvNcls<AsE#jEqP`OUq8{vZK!e
zGOnXf!iD+&J+7FSz89P5z2H*~DT!wcQ;})qZ8)ju|B>{s5u*Chq)9iXh2EbQ`K{(q
z8Hb}j+)zT-A(D1s`){Y0FZPV@w@rIb;QH3TPD#|P*uWTm19;D;`(4#WTfK}51dUwn
zCVLocZ0tztp!MIiivRPzFJL#^Qu&uqV4EIVgO!0sibGG^dxQr(bNZ7f>q3yy`k8M$
zNWy^#X98aCl|Wlg`EksUr-b3TH?fSsp=HumkwZv}Q3MGOqgPp3nVh0xz_v-2&`E!!
zKJ_Z_wy$!)*cAO&s6#4e=&Sn_NU>KN+VGE1kP(vQPNKuZQAY`zG@ZAKY55!P;MM84
zRP_JTu~2bMnB5sclAlTZbuNDv=pV8x1huU&fg=1D5ahJu$o^58V;)d1KVw&aV}z7U
zzq=Ql7*%3+#X7U0p`f5pF)$PgC9k0@{FVE=e?x)TXM`;(+y7LTJSqt7=_xxCdH=F!
zvO~qi;1qlLKVEYy;f<C2k_Cc{ydu7bg!40(Dyy2Cx8L>kN@gJoT=1^p=DcS(;fH#=
z=H;&5rLo{th0LnCg-K#52syJN#qFD?Ir=jVGC3wG4$Su2Lw`nk>axV*<2wA)4J{wZ
zt^c$&;CHuuGs67`1zCN_Qv5Chwyw6M#TWVye3j+%Mcggdn=imTe%wxsJ-e+VWkEM(
zg@Pnf2tp<e$-SpTEpj?qZNZ77l(DX>uOxe3ZIi#PZ7f88dLC+N_&;?Lbz>KCKi`Lc
z-l;N>h6OGSSvu^$xF$^}IilPX5<yyJwOr3h|4`~;v#<-1_}rUXI<17i-ho$?Qd(#n
zRE)}+o|njTO_=)Py0h^L@;NQ;3^J`=e^HI_S_&Wy{cU{6j|E!&E{05J_+-{bcRq<I
z-Ov~G3ypNix8b+pL~nVRcsuSM%{wMewE)zETnc1=jhKO!9o(YX*YMAnO10J1E(5{i
z`XCT8Vl5kzC0{oG_UD)9+r#<%G~Wb${>xLiKRkZUoy<)$x_+1BUi}p;9tXS4|Aqjr
z_SIhxePhF$yv(+UQ<QwOe^`pN0Fx%85$w?{z7Am>iWjQImpEA_yFJqdEyfy(b@=&f
zurXQh8@_SpY4!X^OO_p<@E<9xj%k-@d0+gSvPd~0oBoSe&X0CFn4nfY5H17$=VnKQ
z@++H@lDQa~|KnyzZtkp4ec}78IYWj`Ik{IuI*gxhKE>o`XJ2U1PE|-N*|_SdCdN|P
zA6w>m81o4Tb%V9?waVpxP$fT9-D}T!AQghda4#R@A4nnv`>nl9p6}&kMRRaeu88i&
z7oE6$My`Jy*OTq9dwrU4jfCXu*Z)WX=5es!6$ypScaW>1$@^~tD~5;ifFtA2a`T_}
z*CO<V)+dfIM2MMQ$F}<b>$jPH(jB_GriR(I^2rVd2Pe47V4z;xKo#TppJtys2{4G!
z|D`LMaz8yWS$yPls+|<O`!21PGrrI{A&-D^5yRqY%l!|f^z0+7%L*2m9w(<cul=v9
zGC5x56C~^!lFguWG(+%WLiu*Q)qyg?+xsP_<H5UvW)^?kVn>$4i)VVuQ~~fj`T2Q?
za<Wof^S8Ul;P?;}8oYs+)P55e&t!$<aK+5A<etfH<+c{JhiKvWj_TR4sN=z`FYui$
z80jRb3ncj0Tnizp3^{Ed8^Sp--Gu++prkk)s5ZEGcLqtX-F%)&Hp3wzB9hMMqn#OP
zTmt!}Eg?h7t(3De2I`8hms}oln#!a%EVkPu5zY>?mh*XQ1p(lLUeA3?6CsA}@1Zc;
zjoH@l%)SJ0&#zj4nhj^8_q=Q8_||LXn4pq62kxqc^($p9Z^)cC*T?gLDVR=^7tPAR
z(O0h?H}*dGwNN{-8=Vw*p0385QPyHCy*>`$Byk}8(dvaS@>KIOf!zjj`Tv3og#Unx
zLw2_p!V&wqgoOv0C8zk$eD!DdOuPSiD%Ea8Ywcce_`KqsWoiF<D*xPp_auga7}Y&<
zylbA9Ql*+r^6khbBhS|l^ZWX({#Yt&dWn{3O3%7S<sK-vEM$ylGfTJcQ{xD?6&whj
z{jPh`K~v8(HY3k7n8%n6{;^eP$DyP?xAO;TNKU6I+9Ysxv=TU!c$gY3zvAsc*-<wu
zPhhLAn~vz$`xrHvYnKceGl#{ivUJo$W*9KS^Nq@#^CuBea{M>_Z{8Wr3L|SH-Kp0m
zebGEebZC#?lac`{8BH-@aa^STk_}%uNn8ja&-zWMog0LclEBIPVn-_;=e%vS1c{T3
z?im@8usQq|7nh?*W3*(?77i&i+n&jg(6#+5Kl#Rri0Lcq@+>ZHxCIA%TI;R+6vG`h
z`is?qz~ynfWu7I<<GAJGAqS1R?-uN5`zRu;Pmp+HUib^=k8(EXsvKUN#>b1f`&rAS
zMSJ;<zr;<;1Aa!i8N>cLn(J`*&K$=2P`61nh>V2#LcpiOk;u`1OqFrhhtoA98z;nt
z`|GnJnt%f{i)xV3v()_e-Khw{L<t64l+%810Ps4yq1lT!1xs2t68+E+nLhYAqg)y>
z_zhuLUfw(0hIA>2Hf75A%#eI{;e1E;5`Qbd^!<37RV6>Gq>ap6Me1Qs=cPOm%>Pdp
z%<c9>ectCtQ6O<v`(++Hm;{oXjf-!tbLfWVK0B<~1iKJVqV)Lfh*Fxb>w1^(qe7g|
zdz_*KD9^7YllKsWPeSe?3if7X$WKQfDEuu#h(dZ~{-H&PfuCrZD0_3qST2|is6Qmy
z6WbXx6#TaKL!>bNODEma8_9G|H?_Zf2q>jVa;eE^*YGn5EiGaio0XueD+db;ORw|m
z<y?g2yo=}Z_+K9ey{wO<F)0);?+@zxq9Jmo&UO#P*<P}LYcA1jFH{Z%TP*w=!g)uw
z-ynf`)e`ToVRHZUjFLF8dq}P(CF|jfmx9Y5PNRb3fBN%iH-5VF#;1?<g-WfPJ*L}D
zP3`TU+EhzhNoq1fWpWhkujh6sblZLHHUy4D{m)K3$fMwQ`jqrNBWr~P{i&Vp`3a5;
z!V~bgW1^v<>DM<xMrOWMJ6zO+U;%~aVT%~8CB9j`WHg`Ev6yBTiK6eC!k2T0y`nwp
zWtOmG2nKz(zB>OgHg6Zc(u)-v9Ad-<mB8?VP_Z|y{Iv7xCH|@u;K`90Wb2wjdjK1b
zwFAKU632SR+mx>sLT5Yi3v<R`wj--dh_+DwRVX0O>o2=;l)KEtC7E%kKMgY4{h#${
zxg6tn`RP5xy1%jp50o`4)-vk&`v(vN59z0FM;k8r`q$X_@gQL_S8sVtTszXx;y)h~
z0-w(9@OmATT5M*wWRJCSR2KOIPRhdkAIl=%_4G{TK~c-_tRj22l1mC`OZaRBzTJE7
zUm6YmZ7?sqMs9%>SZ?Muxb~x?lY>QSjp3K|s0FpJ8$v^=*nFTc9Lpm*3mB8soZ?~x
z<7?QmzkOY@GP3Lt0r**-lqJq~nF<D6pDW6fdOu1p1<7+v>qG2|A*Vi(gRCYBe^!us
z%5w#R>nfpA__yZ*kI#rt@&@n6YPQ(faEW6p%WogRQn33g_TO?RKcp2?w8q8#>MZUw
zdc$7q|1=7XrF@H-i$Bu8ADF<)i%IrcH?2N6;s02RuQ~9&73*KCM}+k4{1X`hA$F)B
z8ui3qR7%0IscxSw)1q~BxH%R>Q)p6(zNOF@#TUtgKBR+x;z;&%9HLZ{;@D3~8_4r5
zMe|tR-Eo95s7%PTZv@QI!j4RX4RilK27U&}K0%qprT5kOA6B8zNW`J_X;9U%nu(o%
zh|UkPP1eD$v?+x8jbf&YaR+=D(3eY&^X;xlNj)G8juOlw?U`J+Ab^c>N7b@C!dv~A
zlWw!_iG824qy&g;SDnD|XmQnFD`wO?RCEYWaG><=!I^*<<oX_ctiQH^d&Y_{r2DLn
zfOn@Mh+vn<Lm%)x(G@XoLH|09^=fbC<gbwR8>PsO7T`qhQAer2QNziYc~4;N7Q7BO
zaeg37>Y*2r8|(~rMSxJ@_!vzb4&*mCn8mNT<aQJ~fsJo&VC0+LP{^-mMhXsMuL=BX
zR&cFrGBSy9yZ)5ya`XckT@en^T$@apv3~MV3#7(_J9N3Jk9iU6Hy)|`_;{Jb(b@o(
zt)VlG&wcek+YIP>31Tv;{O~~4+%+nC0%jL0Qf=I#tXiLNMd>M~j?*>MlRIq3jrKxZ
zh`V461qZq!&v8Aqi%Ac39~{Z*XB9s5RDXucAVE98xekA@GM}heSx1j)yZ!pZ>Hpe9
z5_j->g#Jh}`S$_ii4dNrik;K>NcOV6cs>76jf<AI-c37<nTeElKbU=M0L*J~M#>Gk
zX9Vy%^A04H19kIm97>MxE?7_ij?+Y#zX~>{o8GOoQ`5J*y#0vRmM*NjQE9e!$oVFN
zt|cYjYxZvOr{C@Awl8DlsZY4>ybYCcJ0>hX+&rGe<<-88oHd@a*TcIohP?3G?;^zP
zqYph4S9@FK=Q?RiPigsqqLIr`C>@kji^U!-RqJW&<r@Rn6C08jys^I?f648ASZG+i
z%9?Q5y0VGt)9QTUG0D3Z3;bmN=3YQ=?6w2e#Z;B#H)j1R_S`ldx-$MB4@Y#tJyJKv
ztvAIqm$N^#WLD?#1q;D?8?HzP5DG?Zsq7B#z>Tt`T(;QQV7i-qay19cL;AYQcKt2K
z+mmy5KLGDM1`@mT6Jm1p-YL`R!hWVz87OcQ(d7I%Y&RyYbQM<pZrlMGx`(g~3(3UF
zB08HjZR5w4wLPdi5l)a~Ez<{--(Qytsap&aTfag_gm#vNi3eD7MO-W==-14TpZDtf
z%xF2O$;tasnV2YI;)H_;IgHRBBCmJ+EEpQCRS?J4!M~W4PompwiVl{l$TJCYdb22_
zKBJ6`S2yI+Q_h@u9(OW1I4m}3=;$YXueL7@4<FbSiQbyC%RMW6ZE$_FwH#;W<OkqP
zOKT`OAMfRmjG4<sH&j|_4kRb`iW<I1``%AGU*6Z*7k;X7`b(W(%<O&oK;QSqFBb-s
zQur@U?Yy>^H0w*SUg!1INiwY<k-i;w@us~AP{K{^towe+fd<lfM|g+MZX??98lRKX
zrK)Q|&8NRIbfWvaLXwA9#%{@L#We;I$zFWXB`~D>66Q13#Y;}-`gY2!xDL<AHXTq-
z|1AYf8_Fql_=%g<G2=>6Yhc73AxEnNR`VjKUO%EvSG7enMUp`&PP|S9KGOVr-@Ix|
zk8{H-s0QoRuHBs1jcVF|1-8~(V7BwuAgy<Qy&ka7CDj+$!QTQ~#`~KBm0JeYj~19Z
z%jc}Zz#$Z?_=^3!`8VAMo9TBtl`gC6u}SWwYGH<pW3V$g6}@Ns(E#5!g7sveiaLY|
zW8H96q$IuUQQPuNSN%DaN+)!v&$ShD;~5+g6W%)PAJTIRQHXTzVJsiu`DRmZmfo|v
z<B+Yv7xkC`a1KUomo8|m6SfNLM&!8ON;kSL9iHlGo=Gr$@W>Xsk@GHt=34WDDPnFh
zuN8m~N^Aczh-9*ucagf}<K1Dl71%mBA`5(aOpgilr`{E6x&0#qj!l~xG*SH(0@>z=
z%PECH0Nl3a{sXOt2`lHrRpX@At33JGbJzP_-_Nd$J5T<AS!cch19;^L*&G!dTD9dF
zc1ndR+hyyjMaUL3P;B<JG-}Jzu2Xkh()u0B<|MRB<uzq*d+FJ24?af!e%*vkfNmq6
z*PE~TZgAFV)?aSZ9?Uf#mVVwIUaf&ReD;3)K>iE9=@`^&mi`!ge3V=3O&DppYfPa>
zbty&#&(@vCN$IT-ohy6_)V7Lq!in4gw2eUCu5+V+t;6vQ1o$mM_(^j~RLx^Mrp^zj
z#gy|Qi(@}N-2Pnr_Pu5S1R3fx!ifZl!RhumpuOF8J`{d^<9QFog3(^#yWZ!)R=^9S
zrfXUFD7Nn8q}w?;lziVkzqc~ClX=XP0Q9(YXLc_0!84H-PXYL;e94U-S#X~5QD}-j
zE+Pi{@cQS%T|#!4BnHt?Jd3kk4#dBU=`XHoEZK+2x7EST-ydp)<6#<YN_u@2p+)j?
zhC<fqmf_otWgiN-0lHzh+=7$V&=db+LlMc%7Vil5jc<2*g{$;};I+LpY_-pFGzOoS
ze+f?h3uFLlIhtR!@VNSm52W)=nV$!d3h|NH&lmgN-foYK1<^srI~`18p*a3XiPBi0
zP|{&amzA63WA7x-Hq*#gUTY6*BL-pFrOnTqm9V{)_7Gp7sB72dKmn&DUn<Wuh>uB$
z@!x%eB;~}?%>F<Q+F~mjBNA<mfi@J+K2yi#U9tNz(|lIwiMoAEHS=@Lr>4BHki36Y
zOCbiB`$3O=R;z|S{v=nZ<}i1>o|(I%4Q^Y3@X{!6YYX)1A*7Y&A1{ENDz1ZUYmT@F
zS7{n6@;KWP{nf!pp3}WBE;{43fyXfQ(NIjivNG|hWM+tpr$|oEE8`_CN~3;}6$&N7
zw>uvQN(x{*r@J+0w4HRedD};5Y0h^hrtWC85rQ;M_=tQ`e~evZFF)B_U#X2EsaR<6
zAK+h)srGlQ9M{k<t0=p|i10lLen=MiEfY_IxMW%A|6y2vj-~G+-0Ib{Skb4=EkOsA
zDq=N^;Fq%TQNGleAOGpOTlJo6YPktO2gD|%*IScQr5fx>hKJt*3`P@^I1lide_S4y
zdmU{ju=l|g$<n=L!xwJU?$-K!xu&SdP8D$sI=(i!?YQ0~l_VJCes#2EEV*hES8x3O
zr2htGH*|4>NTd2y<dA4$5^ivr#cs&*TywM+0lST+(D+cBKCXlIDiS3MnBw|9tMv!x
zilll8XCh8qK_&~25hO}zm|aspz;`?+;#BAq9g3H)&>H1cqZoN5XXCNke|yM~6*<HI
zv8K#GEae;+jUA+VuNU4j!t^28S9aAX2%EBCzSaPW`7+N{Sk2Z%&t<tiH)b#O^Fk|d
zSDDbZ>s)hMuf{+D0>0KWNAm^S2{3jq)wAPO%jktCZ4YLIJYEr}W;dR^0iE4>WObR{
zy*<A+KR!g9Fx0(Fn}Fo}7<0^pkDyhaA^RWpEMf$oubb4R03MKVgAfb9_3}Q9`=HHe
z1Mx?>oUJQ;nvWUBx%}xzvU={fT^;xJf`>JJ;Tcc(G>OdUTraDWeprdNNZJ~7(Ft{0
z;%wUZ87Uu+y%Vm96JQ;XJ6$++t)jHrmm`(6Ie?-Z(&e+0L1(u$Y*ckL70v693)y04
zRj40)Cve(s3L`0EG6y`LhQaqO*SdI7Sq!LFT7Rq3PK(kG#86Vw!PQ2&N53(CN+jSa
z2F0q5a}@LeIf-#lc%mPG+vB!-{?fRsul`&EcD1ay`?d)8Ze;CkdWX9d-N3-qOC4S|
zL}^X|C;OIH`b;Y>@~|r=p2E6wAF1DtFT7!XYaNY><f3W$5cc)l)!=&!ZWJ7({n&#d
z=4sU$^PO1A#vd9y^<z}bynLmdNBe?Ep#TE+c?9K$6#VljD680eZpKSgHqhrZEZ|TF
zyf4{YigH@=eh2-M%@0XHf%f<+7ZykJE;+&74kMn_5nv<Cu4cVAM$TcXtJR>5i))3d
zDFCFA7%pIU&@<<OUBG?$y0p}T%}pfahDACcaH_-m)`eiL4iB_Sp}QlYlWTFihI~=n
ztP#3ENNZ=hOM`YDUpOxSuOIOSXUFm7>5xxn-aRdTt4`g0LBCa3Iam~zEKs*Gh0z;%
zq@1nQW|sW{k>v`<*=Fp^u-(2^!0>ZBK&#|Z^uLu~XUsFFAew3=tNgpnPSpMhKj5Gs
z+_{+3Qvd#77Uns0>1sKe`yCVkiWM7C0Fuo*=-7z4Q-ULazka`>?Nz=x_9zi*{h8$o
zXy_Amc$trv7zQf)1wfbkv2bqB)J$pj2|IJB6b%XIZqxpbQPbGggTrI90k6vzvi!L8
z*+Ckk#Xl!z;ABc43D~?c@OYl2AFt@!??*26xn5<0F{W=$e?niG`2%WO14eC|yKJ>T
z?B@D6!ECw&gc^;yf`}p8F>&wgZRC`Hb&1g3O&LzC9TmEu@u$PgwzK7*jD6L$UA7tc
ztHcK~tS=8_a{Gw~B$gwR(!*8G@3pC81UY_(HsH*R706*?_<w%+GWF?=4+Z=x9K445
z;b>sVM6rQ3Bp5}~(eu%v0~P7E#MLKXw{dYvw^2CChm`k;ucL2kg3qkR#fQ}fn+W%S
z;T|%sm6-bN30#$tJX1PCBH@_P^4h~y(PKR-5>7q_(rq@|+w$03nW)=f_tD+hoz+sL
z>(3`!mj@{h9Op^_{j~$n-VC1^mc2cyh9fZh-#m!g4|})bSHwS<IC+u|FXd0RR%MSz
z#zIOkkaqXvGt3eu6ymB1Yl?tX7ND9&ZzY6GtdsLu6TToTVyphMpiDlj25Vh#A{Gbs
zwNB)Yq{j^gQr4N{wZ0f{hs7mrhjxJ0p-=RTm<O}-4vVDgkyUbvlj+&5z)=2&yQi6T
zfZ&ZBv+c>$d<RAccoi->>{fcPpM5t$s`B`-JkqTtxP!zn2ngVVfE1Jt)ePLI8!MI@
zs~@{Bvj(hyK+wuQ`WH`9`aRUB--#8p(B;2Lp5JKD6bZsivIc#N=)m~XdZF}E+y98!
zrZu>>OhW1U=F4xNB1D1}RBrGBd*gHjdC@JE$5>g8U!X8q9vHaty8ugt7b?yWL9)H>
zN9va*Rdl46YNC4-SC?1iGp7=QOe<MgQ_$aF+G`1l18v@>Po!)GS7}pzSUKfg@e3gC
z?vr%xKHolIaxEsZR&0YfzV+u*+56~h1W7Uu?%B^S2Tn~N&Nh=gCVlP}@SMGY1&~!h
z;axHgIR{FSuWovFeocg4(}GMkAwS=U@t>>s`6z-u?b=t#$&n{HeW#%sOTteg>XDy>
zN0Ay32)PhbPJXm_KAqiL&^VvnPYlF*RjJlq{-$`Q&l34roU!odYJX>MhzW4r^d#7N
z9O;z@@3@=dmIqsBc%q%2M)!Hh$uq4?uC<-bVK{7<FXs+%f4gKPg@lM3;)tnN3tp2v
zeH^M7`Ze1?Uy=iE|Mdj{+d9DykV*I}Fkwci61`rGwfYg&2rEJLNddE+o5qTq+~uUd
zKk3+tkdVxz$Da3m&KNU`m)zWM%Ma!ZAr4zC<pAxEj5KYtj^^y@npvXfYyY(#F%xok
z`(5h7jj&s>kVRB<6lCImmOr@E9MM3KAIo&KCC0U&-ZHTRj!^b(XROqwth^&O-y;xr
z+gM!V1Iigv=jd(+ze?Jg6is2<<BKuAxRA1(=nw(R=%;gyM%Rf}%0dMZ7k!|}rBJ(d
z?ymkrP1fd12Ztlj`XR@C$(1A`yf<-fAp;Z<sy6h>vabvpq%@o}4obTYuFzY|7ipPn
z=sz=K+=>OoLz)qH80Bp#fW;sAJ~gm+^D3qmBaV3gW%KyMA*8WpxUN-(AUV1{TC2b?
z&3l6I-3Z@|ZuvrchF<`AqLhM6R1q^EVHV8dK0J74rZL_@{2jCV$>~`Ps;to>H?Yll
zpRB(18cKRVg3+3?qZaC^Lx<*Awdm8R?Y$)|*;DB%h0}$^2-;aX=ZAE_NpKAA2_I*d
ztzf6Rw=$k{si&6ZSmQloh}zjG7s4|A4;0X~4sTkKS&i-Lk58m`JI6Q{&^<*BflA9Q
zF}T*ZOTSWeIY@Pny9%}Y>RoUEj*kMW%S+_%J`#Z9$xqcfDie%GaV)A%XKsrtmIFo`
zO^<+%`L1QMH=hrz<eL;vGPzHkI^5dzJV>&B0w6;i(AmJ>in}uEHDTC2EVKkzi@+0q
z<H@pw+iti^d+AA?>d|&BncZr0&(9`6aI9Un`o@ivJDq*X3kOALHVKX#M%dVZeR?hu
z=)Tv%sPWhRdFRN94h;yR&B@VPW+@9<8UDGuj3M&~*C-Bqfx#<U%zaKIrCd(djvgn<
zt3nU?WR?=reG`S%v5r+Qw78sl^L$@1`Si+GwMy^o-0gg+rg(djBB!m&ss$r~7>;)x
zqIUV+8*TZW7bR*Uwjf3L!?j@2F2WpiE3?>we%aIMVLdzBF88+!L|f2LtE<EEuh9m?
z+(&@0{m+`Z?+wdlsCUnrE|VWpYUdohf`^Srczurat}3h1f!AZ9igYUvUb#T>m8>g?
zq$>JPZT+Cwr5%zr;yGOtsr&0|g$C{GMpV!7h;G_DHfmg8&BW&3J$IE>55W<$UN%PN
z6&^x^w(}-X>fl3`jMQB2%_B%c%uE!htRDZK;K|Hv?_kGRzCGRAgv=S6#8s|(M`ohq
zB#@{+F;X4cm?m6(0i}m-nu`&($E-n2>r5H*q3MCm9c|+{jop?3q>@Lx7=@_Tk+Y{C
z#K8bWB8udF<+Z`KBX>pk5y2QGnVo|d<wP}5r^7>~)%mA{PIr$AJbZI_YiX`1*CYLS
zR+v!NBF5%@yT6FR<eAwS0dase&qGfsTDVk&R@2wg-qk_gb~rb+U^#Cs_1<jj!pdWl
z^{(8>uJ7m0irVhDg_}P7KhIYO$2cO=*u$o(YFHnTF1lXGU<J8O1Vpw;_#e~KwvoZo
zWbihjv^}8or6=`dR~`rF>9B`($p{8jX1*!c_X(Er6dy&p>3oIq@!<*H*`AFRb{%BF
zYrRf~1uR8);)_Ua{vo$wzJOaC(+IX;u9WtI7tk(WLz!F7Y;igLz9+~jk7Coe7;n05
zM2N<S1$FU&ga?xp`+?bZ!O-z!I=Aaw&XM#L7w_w?d<RNq+dV_>$7z+y>J#5@-It<K
ztI(t&l03W*Ej%H<J%xh%Tqr+z9`@;P?XB_ZD?H>a(`>mPH<Qk;($=ZUyE&E~y136d
z^;WoR&-nHXm0ZwH&XXA{cT0p)flD>f7XH6f*4lluzzwPxR|OtfQKw)$KCg4QDrsu?
za|hMstMUf!RlB)K|4LGi4zEkR>S%C8M8IjEovOU6w1JQG@C&4K@3(E~LF*3O2pa8A
z@>a{{<1RF`cIQ`e??BCHO@;bGtElzX$6T&-+@8cZ140AiS|33e{jp~t3^=ygyY{$X
zzWDa&@Z05;?{V5hkP@bA+=tIUJr3A=6WsdV5;Avl!?i`t8M!bnp#8KvzlPF$FSUk+
z-!j)Ki}VNj4Vj+3pYS5`C)U;<ob*MgTn(ZVG62!==j0+l1KmFXjwHVQ&)#nXv7s}M
zVI?`yu_W(G>9eC~TwZ^}1#z}eNa9D**R$`38wEZFE_(>)6M_^Cg~cD5Sd(T~TU>g-
z((B>NH=HX>`YceXI@kEBjhS-5`x*g!rj3Iz`~VPW7sNYG#1|n0A;q*G>`^$Ks`~ql
zWOB|l>yrWWrgJ!Zq?QTp(OY{_OrE#nWak<(UZ^iyUiw>H2~aJY^)BG$!$bB`woj0<
zhuLHK@Khh1loN8#qWD-I1F&kXR5GU*u1ix;PmZRyX$vV`zcDb~u;hNY;p`QDEEHy1
z8O<-zC2f!2fyD3WB0%#l*g?an*o0`Q{Yh0y6>K;-?)e>`B!8yUac;nD>i+eknVQaz
zNO$?9-RofC*yyR~tlQi`T@LCE3sV~pPgb2lq<Mwrb*aV<$Tk^bHOTO-`V!t(*9YOP
z4r7)=Qux`=3Au&1%(?6xcwajksmzdIo}g_16tyus)e^6y^7T$G;0-yQfV)`fDeCZj
z2d*!N;|({qH!Vz1XbQksvOEq~8p$$z7Pbep9x=3MsaHOS$m6*dJ$0dpMg`pyS)rmc
zu0xNuDDp0|mC_&VVafsphl1HBK3URO)0F(M7zA5^;Sbmu6PBTsh+S@={C2MClohza
zn=?AT!wH(2YU*Dj#+Q(2WgbGAh#*p9uhC{DUUd(&s~WUH`uSuJpVd!@^$L8zg&lPs
z)L9vvKl{YVxf*ajbU0Me7Lj!q{S=O$gdkLd_s2fTX4WBH#2GOwITTIEmBR>CfWox@
zRP8`7!as=A03ft7ZY*3YA4hVDrYR{t3419D*A7afkj-nCfb|&rN@M!o{RUUwrU-&H
z7&X?>CMMkXX^{x<LXzoyx<X%pOK$N<&qiH#^dnNv@kY507&zy&<=t8cIzV@%o6dpe
zeq9H&lX}_gkGtSkuU06{8=ht_jYxkSP2O6dcHGW*Bz|&yj7#P?wb}bQBWF|hYd(-a
zqtK4ijQ9wMg7K;(fS2-1dBvqyPFCCwh>(zulhvIcLWa`Yb6?DGp5ra^S|<Y!C>B1s
ze{4otWke8=0Cp#Szx|44YoGNZ*q3*2CBJ2@I*cJ*P_ftE@cq+*u9hHb{Q40)v;4JY
z+pXLWmw{$f%ybEK|J?v)dtML;8GBpK>*-=GvCmtd-{Jt-+I@TXCi+HNp)z~6VZfyg
z-cE}yG}ff<mk0!)dyI3LV#@V_WV7&t2pm0)#VqhG=Qr=y{=0X&k{%B_`q_q`MZBfS
z+SO={z8y|^FcG=2Y9i%fM3p2>=nf3XgbJ8-pYK_uF&oYvxrfgXsvyr5O{@j-60L)=
zyzfs4r6x^3*YliLHRzpwY@MoKUAN-aXRF{BUtz!2CSL%@uX~Z$Q?)3&%Fx>qU&0q~
zU%zaoa-XBFVG9rXD81+LFh$7G==F=oX^iU(22xd2=%W&&pf~D3fk7l_fl0%?`iXz!
zt96^$*37@s_!1$cAi&+HM~NKj{zf=$E;Ay%b2nqqUiq=x$}%-kCx@rPWaq(0K@IG4
z_`quEWr8NHY=c|;rn35IsVwwfIMo^@=?c0B1|lWnadZS!mYE*wa=@v`9I4h}4mjIj
zJJA|Sy^Vi~Y}4aX(u%a+(Sdos9{hRmiOuKNF2w#x!L|>5RpbwF92l_T>9R#()Cx@R
z<?uWYZqf&3&S~VUI~}g+f@`7)I)mQJglXiBYgVJ9@ZQ6H@xGaWz&S!LCpHd|&+9me
zB8041gjPe|kKgQxTB5~$so0dU2i-gnFcSrh-_3AK2M3WE4WU;#>?CF|8}E-Wovz$P
zC`kFB%Z2{<LT}a;gsTINxL3kB`?aqdU9pRp0<gr-2$iY~FQ)nMNO=PP)T%N3k%RyZ
zaYrwxrDV-se<FRrV%aAkM(X*fRQgQl-qrdUE}fB~$WD6j=q3p4+IqweyQ5!R*{sQ^
z)e*&aa`(Drrj~Xn@GwBqAcLOHJf7O{BmQa!<I{sfglZe=V4ZnT%ypTvabvP-D~?<C
z0~Z8(84S05=-EHhMB_lSxdZN7*&L<RxLN^aoF0lQ+5}$(v_D$hAaS#ee%4&E<vt~Q
zYPaD#QC1>tIYS8O<OHDl!2uVkAg;^8y&#tYnCzD1NVP$vB)>rL#F`VoP^{%ti66S;
zTjubP*NUUbFr%-9VB}D@4pQeE3!JXgn9A(tS^Y@hlgX*N(rfggU}{3b&}66iD%=YS
zDPQ=1C)02mCra8cR;jRVD?Zt8a>~lee#@GjvvWQvkhm3#!sH3$ohoScX|iGD>jx~&
z?25C82}B;~J9BAlPl&vwnTW1kkCnnFC8d6SIH7HLobQk~bi$~7&D_FMbdz5(IB*M@
zmWsj=#m62MEtA>T_DaCgY99}`i;V@nIWPc3LYbI<NTmM=?p~a{x^iODm*?5rmut2t
zkm~{e;xMFnL$t~hI97u-YzulKF>0jP@0T;TXBmer?F`L8j33m^(-jWWfwW7wo}Z<n
z#|D)<5!B-SDrTMATR+#jw6#pjMl#vrfQu#{XHue|`NBpm*N%O$J!RBO+4PX^>4h;>
zt`{sfn4#fD7?Wl4hVJu^QzScD5<&SPgu3iB6`kCyZJ+s^L%DL!*oV3KnU)1gct<d@
z3nK#c?yr2qz9l9)S%T24Ei}+ZDmp`6#Jt@gAC=rsP2~ATl!|0?tXM$EokDB+CvseQ
z((V)iUwzBbtj+bDd7CFHBXvfFOCawJbq#nU(RMH=YK)}TNjres?V81iQUNGIgX(q&
ze^WtN@I^6BAAIhG^gTBo#ctLDgjnhUM!CP*Yb(0&zPVT=JW^kbjs-~4S?a2{5Jp*L
zk#EvP^NcwMclVpUx!4)0FuWbV=UHgi6crPHCjgw%kS(>mcC-tP4DCHpDtEO%Z536e
zRckkHeyTrdK}F_@kR(2x#Y{(34a}yGOW~_iZ_I8x)R`PI>0HXsdk5I7U*xD0A#pF!
zR_UXCGN2RmEC8&Qh=h26(I;}tf0Vs{$WG{?i)SgKeE?Gf2n8F)JNvTh#LqL<mI_#B
z_LEqdiq;T{j?Tg^nm;vJyu8c^tJ?x6b7p&jWHrKodA%~`he08EQY}(~Arp>O?JFqi
zDE0{gO}S8NS2BBh(;gZ^kMBG~`WSEl@dwRMXnrh8rL&Kc517HS9Vui}a@Hzqg#~q5
z`PwFq1dVvGfnh!^*S)8QpPrg_jzTHnmx%iW$A84UA(LSh%7My@8b=iQO5TaK0dEV<
z0hJ=~H4ebyugq2iWaCAns6ioRV~BN|&DS&f$j+D~)w*T<{8#&=XLQ_1nHbQ4A881M
zK`#$|x7ra+ndJay-%y8^-ee(=^%yykvaGDXS_(G~wn5+*lAujrtbfh5gHR-&_6hm(
zKvhk49Q$!xNxbw=hwro!lolUs36`RggH%vXX(I)!H8^4AgO*!d5e}onlqO!1uxLKE
zu%D%HzP)8g=kL<*w*^@s4y(;dINaqm;?(v>#;%7Yum_H8j!_vIk3aEVHN;RTS*OuL
zz6GY82Z8TtG=KN*yU}%Lyz{JYs=E@+YP#u8<2+6&cQ^EYQj#%mY4f01ju&8mrLsOe
zbzTyEf?7m+*?>yc^E#j+H3G>5rDijpLa9Y?IBeiVMDOmVjm)iBC_hZ~P&arQW-fvF
zq}4T4r{S~@MooRR*+{~`cv;z0yNJxue%CnT+n9*Wn(ok)!h+)oIxs?1VWb5Ochryi
zr<jDNKKc_pe1AtFOago{utf?{_>Lw<1`-lIy|Dj7B-x1y(}ajy-kdLH&^9Avui&v-
zD@;n=Nm1@wJ>=dRrtO&Xe{&~_o!)*cyEL_Km71$Db|C(w?e#!VgAz!_@z~D6hfmm&
zYyO&y79p^o*9c5%84rk8&QhE|9Y#jFS?*iPXU!au$?S3)VS{NZca;<Va%8Tp)f=3T
z7arO7xUqHC<SBJ!WV!{*vYE;fHSPT?A@uQ1WLep^vKDnf(wW-oB+Hj{Gue!xEyE)x
zp(xb(<&cFP^iwYUyZ4^mDMGZdm|2~uOgh!j3H@GKr>lHW7Y(Ozy3^b5^w$w4s;oFu
zvgIzP`MR;zOx_55oz8O~m8a@y8b{JIlb@k0Dr&S_LV;t8Ko1qIOT86x0Gct;Xv--M
zii=4@p&duvfnFc@Bjl4FUyP_@(&lD%-`$Phlj#YsvB%1m5}M@NAJ>1^$Dy7UT7ws=
zyj2vA_%3+kZYs+aj)WKpmKWKfL_4qqjs^=Ez`eg_y!ddig9hy(gK?}>W)sZNQ7Wgv
zEpc}T3uKSzEUV#C?f7z1yuO3;(UK4S9uYr022E~v7!4(6DTAAkHgff3J~HBtoIX#z
z&cexOrKNyFRbb0{#^t3X{W7TvgLr7F-~0r*Y@=$RWKp*q-vk+Bk4sfHTmKExZlLL`
zz~Jy%*>DXosnV(`PBb9*{Nv#bS26ohK_@RORu!`xN4o)o#A{keki`>%Mq!gVg4LI=
zD^-HBLF6k=mj{5YmbV2YrpHW({S=8vwC>3iEh3NLFALOx7pfXbeWnZ6jo<w6W073S
zosFoiiyVnAQ0O$SWpEPlDI4_X*I%h-EPYWu1mief;uhC6d6*GDCSK{=0hq?K^2enr
zCLfQUOi;;e0xHdU2CRixv3{|eaa&j<S&DhFmCeDf3O9b{^Fd>E5ygre)P)=zI2C5<
z%X!5vhbs)}3w8YO3ebxpkIa|2K9(I%Cmb-7Y4_pL$8trVJ&s~-JQ{D>?+M#^;bPMT
zE#Sl+Tq13(9#X<zgI>+cxYj%bgl+{8QW?opuCf-GMN3GLZXO{N7*a9rND)kqTBC<0
z6fyK-cZwDzeP-KSh;rL}nlGN0{J75|KrqmtuCii1{$)c!kVUy!i(jzy2(*p#Ap&2X
z*W^uNblAUJ-TK>?(m}xf7e0uW7%uCbo2UYHd4_tQBTm<3L7mmDTeWX4E)Jof2#Cj^
z-gEP|t9b@wA@?=wz&NqcfRgUBg~Q(rU!FytsgU=8II+rnQsRFBgYj^!4PMYd#tGN7
z=$*yqi_`1W@IH(#e<d@b&07RFWn@YUMGwdLICq!EvPxXjyuV^*(X^s{|2R8n%LAJ|
z(tAIcn;ZVwXF@21PRV@N&S7fBje{${@ZADw4_GTAhK%*`2yOOw1%K5?C&Dx6es0Ki
zYxi?feEc?Wh&`3hT8E;v`y6Bp#L$2P>v`VlGzRCg6R|*>zs2ESzzjt1Zg9;6PO;fY
zr-yii<9uL862}zrUk|GGioi@1eJ$3`c+v5R2wdlCi^yB>C20AJVI#WGj21DqXkW{n
zOe2^kF*c`_Q44EEiv{Kbt{J_H^)FX(QSts$Ae4}wMj@V3RilMRJcA;?fzMD;k*eF#
zM`Kt7z7RAunK&O(v|B*65&k7<+mo3WpBwNE#tlL5QyKzH)5}Ov+;V!PivyZtI7I4S
z`vCG%8~JFB$tJ=p$BVoc9ntgjp9NUt-^6Z{^0qPQ;6DWgk`AP3k#(2Z?S&a*(VDJA
zTZ;@Zea}hfU>_>@<lSJ^CEq{oy2(qaKkWN?!BLq6nT3^ec?a*@a9`XDh9$k=-1F`c
z7sDAY1iInj*RXEBm&q&>L;2L&?lxvAr*|qEQW5Hn1z~fJLL`m&HK=^86IHEaGugw*
z2wGI?@K}0O0%^T~#W)Rna_BQ(z$GYky&~Wi@6o%gwPOthzCYIhYN~)g63-M1`zDvT
zt=$Zau`8^5-hrY@+$VvjyI|v#tRLptlR$fM)&Y|*9Tl8C8L!RF-RglrRD-!5tZ9NU
zY-<J(LD1q-hDXp+gzRNq#XWDgr^ZUqPEl-ft9*K><Q9v{%d#a4$mJ-=!BMB_!MQ9(
z9MuxrRplL}LQ^d9>PlX`V*8d8sib6p;>(y0xx%>@cCAI!VjP=%_>@NC>R%XY4wzD@
zRm`e!u&~3ld{7MYn8`FgNv#E5umo}|^RdFHL^V1ek30S21u$5DK(ky?EZq9~i^=BT
zt7emEpVIp^d4;o$9p)o7F2$><;;px;T#upL&3=OHeU#Ka$Lc)%p-A&qcND50okqXi
zwULt>4Y5-MpFjPOZ@6UFAbx#rgSN-4{jPC}8+|BZy4qYlS}gQ6hlA~Va@HZcV>VCn
z;@QJIMR0=niri}Pb#wMsb^;TfqP!J%XcjNYf(a)B0rvD5x?Xjm<h_(9zm%B34=^3G
zFtq)${2m1Hp%D-3qV2JwCHb*4rCqqmM;6Y;V-mzJvX-qT#*!c-V2Zi$p6a0Z-TE<%
zGUR6+v*}|X;3X)lcVzLl8=X7ub`vQ_i>iM42}N$6-?qNWu}_PBhiS+`&gw%}0oUWB
zMjX3sw8fEzMwx$3vnr!cwBFn9)UhcRU`tgfOzz=GP>bLm6wjszm^jZsFygphI=L9>
zpwcdAUV8d6?7=1GCSx3TkT$pfRoWjc?AzUs`}35;_Ps;{?Qs@-k$iU%e~<!I&(!0T
zYJJ6Cc#@2DO%>`;UEK_K!Du082Q^FIP)wQMW?p)P#&^p|q}%0)8v}1p$1OL8=f+EE
zVh^%X(hr{lU^KjDc15kz$+MB6qKWy07pqfeS}bNV=K;T=9v{N<{wFu+qu0@pyPo^d
zs$1o>p2#XYFq_;4oz(HT4Mt!T`1FP#_~<QM63Ro5#{sGxa<n6gAvJ%8NNamY>qv;9
z%qaqOVn>0Zm3}zy<<G84Te-Sx;eAmRiFkmpCBj#+jDFF(Ga@b=$<976!Q$DMBc8fz
z#ku|051&rRokxi$4pM*G1hrqz=khZfw0fexoDc)`F`IUuItX$r1aZ;J1=XZ+WVI>2
z*I!=JTU7<usV{c{CX@zERQbuDT%?0Tdi#iGlDdjTe3Vbm)Y+kj;{kXhin|vF`<Gt9
zr!DJ7geC)C#~w#I>+UOzYo5K`geHAfQ!P4P9CIU0I%j+<?U9oQbA9JOZL38>m#}pl
zG&Zb0L4AUKDZKU#t-ak@_DfKz<uVj`s#++GjeMF28QIjpOj7Jms)Rz7b?qMtY+moP
z*w0_TtuY{W*6tXR96Ov%TudtIVK%Du@?D^kX!%qr$1{4mA}#ex>R@1K5bo_p8{x#d
zk1lcHoN!nqoAXd&EpjXX9>vV<>;W0}&Jt(HiuQ=Br%phNwT{ayZETm(VB47#85eEv
z=j}<9Jj)!<+_}Ypg7wJhuoKc+^M^5JiQV8G`wRmYR6XXli>Db%$ptS|-Xv%i{%G|b
z)v4T?A2-5TV!-^4*KzAs8e}LEZYA2^6y%wAZK)CiGIepZ@MnEB=5LsE%`DxXmxUr)
zNVr>|gM@N8@%;6TFfL^w0pjejhusfrdQ|X_1uiAuK%;X4E^Xe*Q~wu5^;QoGIY{gq
z6SWBdQJ($1n<kEW2VRG$LB<TY3S|XZh3m(Cki?k|1yvyPb;0=)RqpqAA-2#~3P#9O
zKRBmpxqY&_hF59*n@e^ASj9%sW~OF*Qi@Mx*Twu3VvTQGGqp}XnBmZ9)JtF}R=Z3Z
zaIf`q8*V46j{&nSe+B7#kqsR`c%Yt)#JAwN;Jv~RR@<ko%P-iJ-=m1-7dcRi+5+in
z=G2=e#RHs6A=px>bP8%!q%rraToK*tuZIu_v}-Of6d~?@nTAF9%{=Q=C+dy>OyqWr
zm<>FO{GlE_=wgN=zs$MEEB$UVh5ck`lP1espbCw;$BNvj5w6<-C%^fadHQqiaCYky
za9*bB^3*q^nVL$V!O7Vlt7ua!FERgVy7d;XCJ^oRBNmzDJtGdh!k3FfX8R8aNk4ul
zvz~xxg-atE-QmZMD7u^HCAWvA0>s7wF&EzeC3~7)0K9*G2V!DRq%LXlZbr3P_!wQ&
zs4ZkKpCOAE2c){k^uLG&9ElXu_PqCg2zhnV0XUG8F&12|^Ey>WeN4JV+fCaE7_^xF
zmQt6IeudkmQMfhgoA5nUyTeMz2`hp7UTE&282DmNLCe70f`l~zjRnQ+HMO}|S=44d
ze{5|>Nzn;h25i|E9^vk-K=N!Js0#hfD)%be%J8sTaYT+qw5BASsLp7iY)D!kiJ<20
zc*azj279f={F{5j6gclvlUA3g7m(sIh8HRw={8O%Bz#xMRSR``AvuAqd(t^EHKA=<
zo_j0Hp*6rijv$ygplAO$t<ozhhNjjL3$c-IC14Y4w0*?0_e)UE2gD&;uiY|#0RWl#
zIi}M+lCs#KExA}*Srk@@i**v^qw#BV8+yo5IFrm5D?(YZbVr?08cdgAnMN8nxiddB
zm_du85OkTIM{*VEIG>G_6!Vn_orM7HXR~zr#Ps`Ck81+n8R$B8@dQrWKIQnb^7iP0
zJPLB%glsaE={y=uY>&=HdDfj>KpkSjpsg0=52~K?>uaselv4vwVikGE$6=>kuPX|M
z)o^2vIIDzBc)Y1AN)N%&^p+$o&9jA(a8{OLOTQ(5Elk)wkr(oK-*BGT2e$aIQLF_-
zY66hvd-J#{mg~Pr%!CcngYBfMGeb6NN$#k~!{zk0s~L`SWW+6SOZmu}BUhWU=FC;b
zB4TS14iBq{&%1S6zxbS5H9ko$%s8Xi2i2*)@6yLROqreSKjr;Ok5fV8NeZQ<t(AXg
zLM>xdY_jw~+vlLJVe#u@QIkD@f5GZ%?P%noOs5@*?fS$QK?tGn(8hI?mhDOK!JVXz
zEtr#4u3lDvt=&~VC@$tD<l1SrO!mba5U9r#2Xe*v5B0T?OFR~HDl=n@*?727)<gND
zrMjcf`e815DmpOeM&i-WlfD5$3>yq|bySt8l6D?r2+i=9X$lr5mx+Xsknf(Zh?leO
zKx!?!aT@JhXkPc68vcEjR#O~QqE*%o8{nph*Z%X;q;ol9zq)O4^~)Ddj0iMt9zI{R
z(pd*THdkgic+PHa0SJ)|8;c{2#N)rYLY<wDIUH%0kpwJ8lfcxhK~?T=l!$BA?DOD{
zHA^dots8Jbj&d^2Yl7oY{rnQD$2r}sX~v*oc*kpX6k(5~z`@*-iAS*q%xL}!$TsvY
z25``T8&>Z{4Jyd_(xmL>x!*ScI_x@2rHpCb4rs4VI4nFQF6(Hh$kZL6QQ>a&`c`Lk
z*ovy0Asl+w-{uoZ7*BCs978czyhXLZIVO<F?N(#xrNfoui<H#==zaXaFU}(PXbk8*
z5wfI<->IR^Pa9+8KFv4b+Y~2hT@83u7Em$MOGu?v6cut>QaB9x>JfDZV%s7q_Ei)>
z(Z2mx2;USUg;%GH2KgXIqi8R9iwnjs^1(Q-;HN)pKR7|w<tP?y3a_5nx87@qtU;Q$
z?vB!LADf*^7?pg}WbQ$5C6CpcdP+w0dOw;a;lLstj1!F+z+F8QqK|w$T~{F-g#YV}
zR}Yyu1kh>+gLS-AIklTs@sN*MB2ELMZR|R|=*qKD-+Z<R)1k&K|MCg(ir_F%KfkyD
zcYXRHZ^M{x_|=zpQiyzJUSxT7hYhHl%@5x}chM}}MV666L$(*dt-`y82wm69Qd?VL
z>D1A__;#!m+V}?<vA)n%L^Qi&CR51-^QCOOwA~e*+7827-n;8`H}D$30cW0Zpt7;X
zN*!8&cbM8ckc*ZP!uEuBwt3vjHdWh;yFh0ALWZKkLUiA!_VKed4k&_`QT!BW8-|UN
zP&ecy!9g7XV>Ez|90>rzgU(t^G0Dqr2ov^52d}ws_$mvxC%FX*cbxD&O{&ozFIpwk
zQ)Zf8+tpEf2pL(5ppq1=ftezi5y$TRp`PrI5cSAGT9eA+j!i@QqJ|Vz1pBp#)79>s
z(DIOWA2=)vcE3wMWF{IV=`q-%W~;OhiTUpdD`)YIBh{-piw$Cy67j_zF{4+*F^6rx
z1d$tYta?6(ii!2-up++Ai40o}a5Dwa<<-O_x-d^{uo}UiXmLcrY5}{whxNDTC~-Y~
zktnkei;gS0;u?NI4sve-sN2lG%~Ss*Avi%{Hvc-cD&pNq_LF-hQ!432Dn%l?;2RM=
zh3mM(^#x?>O!62jG~I1_TmbC{>BJ8-qtl7?Osu|EN(#wXj0GfJW=vNMq%X-qx4KD^
z0IL))6QtVW3Z<l7V&0%tP8rX;tJgdI?*V`nL?v-}pt9*5k8@<0(c-l2=`uz_STSEk
zk#p_&C``m^RkW9+_vophx&834fxP8(ajv<_Qcx^F9dW`@QS{nDU{Y>0iTb>WtrU^H
z<&93hD|S|L2(`y6fvueHONFz19V7Eh3nIcn3b-`~kE2Jr;pacc>Qb{rs%)4hJ-Ehs
zQQuURxl{*#&QRCaFzaBI!0lOWq6|;E3agE@N*G?49DwXN0VWFC&+7mN;)NFi40rrL
zs@^&(%J+@>RYJORq@=q;y1TnOr9q^-I|b?PmTpG6q&tW1j-la<-{1S5b<V#omm~0S
z-Pg1CXYc!b8%{w~RC#WNcSsc5ix-zG4)HGGt&c%@x=}ffh?!*RUMyhPiNK1kn;kg|
zC+J%I*Nk$`1`s(n6zuW7&lB}oE3A*(F`2<gjovE$LN{cyT9oc@I~Vj+k1OjB9YoQ5
z!gW@Xe7_(z{k;`?W1&%^Pmi<yHHppTAl)xLhx_w*_LvxSPB#=+7a0mf%9C#BvY88F
z8$MyWkBqexO&Z0Da6d@eFrLELRQ-8>Nw?7rd-IZ8sHohr$0cuRf~7)Y%EZjRt*LNp
zXW)bmJMn4KCU4DTf@;s0NzWO<FO^?#mcco(`|DIT7U1!*@TKlvocQw`le;r12hnQF
zu7PIeNaU6^*hz~~C#G3&Ph$yLm!odBvX7&Vak$deg-U%VtQQ7GWuxe8YL1FbD8rPR
zGHm5~CmJuRewfb_Y_h!SGc>`_c7=Q-OV`_8ovD1$1S?cId)N0!X}5Po8%OIqX2)D{
zy6+paPT;&eGOOhoSR%N$6bF7evoF<6t_RzK4-Cfxc9;l>!(|dN3QZY&M=Mk2LJhWh
z$IKY`9jLExUZ1Q%7i=qV>WgW|pFJLLQD0Jd$e;2P3@rZq?JHC6O4l?yNa2PuiV~jd
zYH>0u96rR_*27zfeY^k~MNvDZy28Iex^2=pZGUC-P$^_=?#cu%P5w{x9<G6RV$*la
z2B7#g8Tw-OxQD8Jypxv;QZ4c;Mn`m<40Wwn)eax}iF@_qD1f9{{#8PMt5ACR?+jg^
zCpx_LP2);kg!y7T_5oc6?Q*Q8Id^D92IQ)@Z+iubFta5;*V?|58@&Y$XY;kJbb_qR
zTfqeoGrH2;df=FHaLDBO1%VqVr96h;_!o{-*H!tN{jHC$gxLM=(m}n>2DX&PR=F&V
z)s_ScCvH6H-)qrLDGVo@{#V(b8A|bYF88M@cAppVas~Z_&$!&2(%LYe9+Bu`clHy5
zmsm8Vi^UeJ+dCo_(PZ{T6~o3dmEwnhxL5DNFaUOaqqW-iX5+(*0D5@;&tnT6F0bD&
zf0UPwb3M#1@T|9U8h(Off0tXR4n6KV^IxfR8sdQhL4Bn4jE=CZ^mdgPC>UkVQ&UyL
zz_>w7u5npx%8W^t94I%m|NG4<xIUDfifqmQ?(=JlU0=$OlFsUig={`Pt*^BT4bEXx
z>EwVP8VU(~Wz?RITyC@$>ztM{%P_LFtZc!vp?T|H^u@-2P>qq8@!Mgiijd~&*X$|Q
zVC?emB(%hXXZ!)xsW8O?8loGi=&RjDvCU~$UW9>JA%EtZGBF68b5Y8x8rY9L5O+7T
zbanX*Briu6?L3-}ajj@oA+I;)uu$owee2v^u^35Z=sFR@g1ng5I$m%6UEYyaDq7kp
zG4Sje5%w8&R^maPcl=O%S}6TH5jV2qBlq7m2nubw6j@juL&+%VN!^LFIo7o(8!>dM
zU2}Ov!O)DOp?P;pErwrO%i|1E$_iznW_x|wHM*$*iYZh39MeL5kT2mtW5jX+)jzNJ
z72DZPl`X58g59P{wqREL(_GlD%&4`W+;f0wl=`3c_@w}-(AfFksc~Wi1-dUoUzZvz
z@g1`9=G-D4G#1<a+g-eJRi~Mp&0&PD^S{!LXed7C#d4W*)8;D>KxNGJcrONT@n|Un
zXF0ykxK>2MSMq9?16NLA0cX&F#H}K4V@uN$+eIqJcaWs<%?isJntz|L)42ok<lg?f
zz3{_f0}50TqSUcdg$3kN1$>_G=FL|>`TiY@PB0v#e@2{yWb_~OePRSNLub#_OKuxQ
zBSlwUwXc67Cw7crJScr}+8O?6hfkhRRfeqBfTTfS<)fMnuMMYVTaK(X4^{IS`#@S_
zJUDsl9G*IcHzqBt8AzkpXE_g(z_x~{70R~u4K0+7@Xp!t!1CdZw)4{Vhz-hsMx&x;
z+hVIQQ|q^1Bl*Svzf?HJ?6l=D%x+RKo*Gebg?+uus8a{7Vo$k!rZ$O!7picJOb{Dk
ziaT?z*;r$KtM@)UBJS70O`L8Ttv*#kzwg^trv`7-ab7i&)+pMt2cq+CnkK_%|NJcw
zT%g>2;TGUCm7AMw1!F~uy5QNXFaM)p7ys2&PqCnH|HNm-&#=MU11UG9^n2qoPNzXe
zS`+LgvMqZLFSie^M1nzWfuwfXa|{X<23^KfM$~jSa<t!r&O_KGILVYoR$nYYuqgD&
z8l{nPE8#=^dB9Ym8X+9jRbK@j4pWvYIz(%(;y1KLn}A0HnXTBI231OH?s@A?s{-P@
zvtyLMC{1ql+*aZu0InF8#U|)m*CQ#@bODsB3*4eVF1maK>1<ib?+QLpVGnjvo7Ejv
zAPa%^;km1symH&R?rkCe!DN9VWa^hcjJM0OI}fi+21CZV_VuibLp~j-Q;Qgbg}g%H
zlE^_RGO*R#nZr?PU_$k>*^wc}8>5w4dkt03lSr~p$=wP$CAz*ZeyKI4zbD=HE2m1}
zqVN;E6m&h&9^V)TEqJd+uEun(Znm#E*Vf3Ytmzh797z3}Y>XPZEKiyRI|`?*l0^l~
z=Ze~C3`KKEX12o{>vvg}g8548wjd9{THZ2gnQ~Lz;6s6W#HuEp`e65uEp*bVAsTml
zTpBX4i}|xp$S_3lNf^yM5#1fn^vfbiX(XMIhg%Gd$)fJ+kX+{t6tI<IQB~J1hLr3x
z%$rx2#Ie+96L?-9(!>fKx$vWU$wMWslr5pFR1RTRU!`tKnqA0P(;;D-oC$UN$DUzY
zv=SF`<-0X(=rng*zYx6Jk)w7^iu?bXxY8qRFI6w`8^b=f54H+tnaSr0#}x99lgb;S
z$VxR{s~Ezbo=hus-(rR=r?Z~~B)z}*_m~3WcKSL}<(<zrT({DJi}*@&b~9?_M8)ju
zQWcIvA$^@^GTF7WG?YN4Go#24>7#koJMtVCg6Xtkkn0ie8paY*<Pli9u7u)RKjaQ{
zT?&H7fYj)UjNXc3IcuJ5`Yjb#)c26Fo$I3?f1WkywM;B(&K6#mL!;Xs#qD6fP>wBi
z<RxQkAa`PAF7)<lYxlE?b-opuuL;GEEy+>h2KSJ2l;|;5)*7b;WJ+Y+VCVs~QF<4e
z_c7=o47S(9zAw_tk5w5@et&x|p2Y4nSRi$}+LoBBtyQd<_(_dsYHDh1-an7jfNJj$
z39(roa3P25T8r=<!4gL<NlfM>2#NXD=9koB7=_F@N$~rsE|<@n|KOwlkDChV+k(3_
z`(FOWAAp=KuQzPG@S;fs>jiIYH^4jUMGqJrHn6f@^%YOWeM_Xmh7909xAG>M^se*l
z%5yzxUFD=f(-#Yd4~V2Jo<)DkS1&)u3Q;eo`W3rbZ!8cao0t*llr8<O16Rp{RyF*s
zD|<fi!kxK*EE(Px3ZUqmeO#;%Hf!dls+xrU5Kj;>SCR&;I*j!c?+ZrH0isFnrZmur
z^KL!mH#u;(>M9?pQT16c#_0$!uMy{xyshQ=fqty|;ZA~l%uhUn6f79(8y-F>IWG_G
zVz{_;0zH57sMkOiX*W&u)Q0%fHLQP<iEu3P!FGZJq`#%G@odyJX1Q5Q-KNs0o8j$f
zok&9teOcWFA>K3cEyqtsEQ~<LoLdB-2$9p%hMmslXR>Xs59Q~?8Vx94()d_n-Hc52
zR>Cl5SI}7%Xli%1S-y>N!GI!-{cA5~u~1!cj>Lf!m}zWlIUnr!8A<zOo|0E?JAR-9
zcUhQor?6H|Op+oD0v=Nm(PsPrb=Ay?TQl$ZJ8~=LyVSr5jA3i|ETcw5l0u{3Ch3Z4
z0iUCB_Zms*Vu#ki$PQshlXVJuDIy&c{Aa##IhPmw+^P~uSUP34WKl&aR4Le*^-;)D
zAuPdAQ5<Cz=2_sHQCcO=o&x*K_zS}zRTV?Q?Q|yExF;GjT{NO6W=Wm~$!(MI{>dx&
zDdD*pw~C_HkrC<Zx}kD(SFbo#MYIZg66^Uw?1XNAT2*1p=}MW>D!D$ovsC!-3iH-R
zQ<-^C49l+-%>fpHn8ty7)d)Vf5!I9oJ=gGT3Qp?smNT-uW|t0?Kg0NnST8DED3r~s
zc;Fsy6=u!aiy|V1y>Ggdl#~Fe=5g48tHtE{6oS0Pv~d=?{Kt(1!<Uigc{fJuYqj?9
z;Sr~qx!>tcjCZLKb$~Niz_M2C$9Vxo0-tV6KM%dQ06M5jT!X>rLvHX@NIuT*Z=YXl
z`jDhHhsas{UOX5K+SSJsAysx{EG(Bpe8U3&11|ty{++8L5ez7?e;y2at(Ag(c+-xS
zV7xqG)7@0b6A-AbAm$uDv0u;~`aKwjzw{+Yk2-iM!`^%mPoR`Z($F`!HUsz!Ve{JL
zW2@ygkZz53;~?@`GC2HDS_E(&DGsQ4RSsp-*F2+EHTo7~#t)@Z9+$%1J$dE4#N`Mz
zalX;Zf?wYYhg;2Ce{yvJ`eU^(t&g-3t;=Pw$>EKIjEIYS-;xHm<&;~7AH3H5Lx_TH
zZ6Z|7X~sx7%yz~0L%UK2a~2F0Qw)Q~@l&0zc7^&Fm)oI$EzT`6QvHcmFxYg5+e*`M
zQ<ZtnVO_%7ALYfJ6m?krX}4L%m8tUbC+JtJfS>=)z59UJN{cp{lmmZnO~Y0d7CE$m
z_}esY=&h@N5Op83%hlUVF+71#4(9^vl+=#(PJUr(e4>w9Vw8T8819Q>+>k#HF)ktM
zz}g4W6c)8G7iV!CNgFKb`HTgr@3M57#5{pQU=mEcZ`9Gp+S9YB14#Hb`bJyKrklq{
z^hsN&@r1p}L)~t#GOzCT>Ld>9yuB>7@&KmnJ3p%Hk1JKW)co%p4gs)5NhO(DBP3+a
z<B2BZb4Tz32H8LE(aMEpS659{Gc`?L7oD&05BK6X`FagIJ4=Z^an!tPJgew$Tgfl`
zDvN~KI!1$56A&AoULG+44$(mKzfq72qF*Q|sJ8Cc;lxS4twA9uN03O`{k*f)n@1+d
z_C^?WI{F)bQB04|pw?+)0Qs}5NMWmo`K#-utHITqX9>49BYs{lL=11#`nxJybC}g(
z=|kL6Zi~BA0gcxOO}RyZ{)cx2@6zr-)IlZN03WZm<u_WR)E6TY-|vLo$l-AAiYDvG
zkaIWTJ&{cxE7rr1?;PnpMHEr0Va0yRwh`_Y3Eu+YEr0cZbjh8oVLdN4@A}u*T9Nh}
zSqgmX%($B9yq{Xk3jV4;e`4I6GMyXpNARUWqr2X;4oWD>Itxfvh0Da^R{8V(_WEN+
zHbH`minYZ1_EB|bhU<K$?`-E;YpJ<FsJE8bvN;kGlwYGA+vX8GM{%nK#1UAzu{{4`
ze99Zk+Z(RTHL`s#-jP?7=}ol1xuV+t_?Ka|z%<rUw0%my?ef@~yZLP|<^Zjxt~>07
z8GEl%*Lr^PVA5IMS1GE_HNKn-l{IFLx@<z_^o~$rG)RKI|H$A}?V<KCe)zZ9-_q`%
zFDhMrcuj7q&*J7TIi1iJbn_N>IpX^Lza#|Gz4T(k+z#JPJ|6;bH3_t2Rmq@Q<e|%l
zF}S)U?UQIf*lgrNJc=Mtl%I#k!Cx(qT|3G%+@1X=pR4xQq&w<(q-+0mWKt}|#4WG)
z{m_Eqad98E4#nb=P;gMpf~zEW2Q<{{A6nTusJ^4Rk{juX#GP=QmsqFiJa)ln<OX_V
z-f=rVR3lu&8d-=|Ate&ev569eol%^9ozH~+`UhhQ^CRzzwVhoI9&-q^)XetzIjWYH
z79-=84Mg|)02x95ZSSbw)n0xCx$r-i!_f&+hmYND%|F>je;3tj`xCi_`L6d68@$uA
zLrzeO#=C`gISB<bsy?EfF=4S$rrfZhHE{7te+)qbAhTtmFM8OiRc_dcNJ?Ie&og6=
zVNIy0I{})^M4?`gnea0BAer<pP?L##%JnpNxbP<)nu9(};v9{HO$&m{^r&Zu)3N40
zNwG{yT;OtZqv{G80N{`a5yM>|(<BFNy;C$posNXENIWChNb{*c-&Kyn_bo)g|2431
z!&W)p(zmtUUVNGjC;4=CEz+F3UV&Eo?iOhC_Tb0#9Gh_g@_Bz#e0J|GrTBZz7uu>M
zgM34|&cHIkdyvyE@AHh?Lq#PW6Kk=<dA1Tr+h<Y7Lenw?OXCvqp?H;4gH7KtErsEi
zi4V}{9mA~oY;fZw=!q=lh)CBgO9AO{ccgkjkUt}Nt2xTNRe>q)P^Hu|?A5%{!FR%F
zP}!4r+{SWgss%3!|FikG`@gSn-{?vE##>I{Upw1n&(<i!MM{VltI2O&f6Qjx@6KWW
z<-w1k?2!Tx&*Ibm)vY4QO^%@Lz%F8)KQST;A%5q<P@@p%uxzMq8oX+lE5u~W`oSM>
z2<#DN7RrkC8upp|`_lQG9FVBzZnDU^C7gNOTJ#DTE70V9fT9k+7bZHNrt$txOKj#B
zuP~@}9#f;Sqnj-3+yHi$)GxG5In_Kjlyb_}Yup3wTZCYLM-X*S(TGfcH%^Wj-w#fa
zS<T`WJ>swbmj!sdS9udu#yHZs^gHM|oBcfORP<2$Cirw5S$%0lS|md`KtBZHS81fM
z;7ClVy`)x;U1~=EfzR4?545p=Vh*XP+4fbIyf6~<e~X~r!ozTL_$obpq7ZY*AUse|
zCvX~86V@$aLP702n9n%t{QX_lty-lc|8UCjBl(8)jbFwJ#&mR)Fj;c>_8!#sMwg{I
zBM%@BUcD*vej&nNa%cL704F7mlzS`busd#pPNzI4ckF5q=AMnu^aaX=k<d`HMsR;)
zA+F}Sh&Y2lu&Bq)HR;Q*j;LLxsQ8RnZyOvE1ZT%O+ws~~XG`izr4X@`95rzt8Gr+Q
zu}o|`RX=g@HiJb5=^JzI+JGh?GVT=7T<PvXQTC-j!mbpp3KzWU=kRUD6T5+VVA5e&
zR&rskQE&t3`3blObyZ_=ANl)C0ISVqPsZit>*QJ3y{Sn}B*!7kAarOn0^(H0socT+
zO=<j6%hp$Y-tCbJwxo?qP6k7bRU?6JoVWVhLCA_9Z)6>#S-tvEK*~PE@2wZGEq7*p
z=D~P6(2g_glw&;VRR1cxvr=;3Y<-HS>JK4a<SE%=p8(6$7TialvCM=m<5eZb5PC9R
z7A*zYGMN3c-@9e5a?o>QC|#Jp2r7d8zJ~z@Q4*^>2afmNhcF(_|0^Y{)}@_$-9tS*
zKHk1NSscz4dUwB%bRYgt6kv<~K;377TX&)FzkN=iO-NQ0%gV*pKS!L5nt=Cj?pR3%
zz#E+m!iO-^an%$~dCnmp?ox*4Fw++eHp>?ve>G0zlT}Yn*!qB;X!lcGBX-zH)uNU8
zM5ild4zHb8vjmyCn)WVqCaaxk=|iEag>%0PryEnwC$cgWo%2{k+`D96WujoJRS2|{
zUw~|-6vF9cLP!6jVsEN_U<qi0`dswf@A=ye5U8a@{KsrYSlj>@@XI&6=e%Xv<_0sb
z0J+qLuhMDP4I&E%I)?u_cfo+~rg|w;>1+MaTe9TWF;O45mx)Eay|F#;7fWmh^BcFP
zJlADZ#~+O09C&nq!-R-PPmAS|>ZQ{eN(1IV<M&^4^PfDQwoFb5G(oM#AFJHmgykA!
zZO%ab17^dx{xEBueKiW!R()?~;UBsSLe?S!)<267ePw_1#)?$TVQZ8TWB5{#SyQiD
zf}I&?J(t+7cqHmiUB6oHa%x5f$7Y&QMNwL`=?@qDn?m}_n6=r@V@e82lfj%1Yd*x|
zs?%=2%fh(kD_95sceHntsqmJDexINHAeJYFp1bLyM&K0Ks<as&)LyK~E_Nd{HH&9O
z9{L>P4cvg5R?#eaZKg8KYzP+CmM`^@A*jyUgNUm<9|FC-p)Og}mA}&|7k<(W4c!Y+
zDHaHcbRqx!0D`L$MKFVoNmrLjnENG1djna)bkNbz9I|_LBiYuN#c?Y(#Lcn<5QX9u
z-cLdxs<MtgYC2I!)O@Dp2!F`aq*#JHpvFWG@R5on@x`K?l;cH?A}y~vC_`V2Y#dn|
z#=xM9`T9U@jSt2lJ+2|pHodE$n67XVh6F#g#Tcx3ZV(8~CaO2@MKl$i7#hQ-A3qXf
z&IwI@Uv2CbDZN++B>3^;nzh6E!muZP*hve<f;)UVQ@JBlEzFI*0gb=g9mCKes2-3K
zSK|P|?rPS4M3CG2aGrf_nK%o59}P4`X3@)*VmDM1ec?NO@lKQeX1i*|MILJ1;MnDP
z^`RS+j0mH`hU<*->Hc~I(!c7B>l$19#r4q_ZZvISeCoaI_m1}01{8v?v|T)J&FSOS
zCC*s1Gv0~9isVToiDS}dv4h_C*QD6YdMLgT(%KvPX3uc$6MxJcw8CO<%>C6JrzjQP
z1*v`h7?L%FBAPdsWJ~2|-{|j-h=nUPt{yOlco-iD^=Ust)c7g)BocW>-SQ@&S8r<K
zZCRR5vq=#jbd-UNP|L(cXlkbLw=sBXjrRy?_df<-yfG7N7Sm#(Xq7C58cx)YYq?H%
z$)BA`OaU@#-GhGVSW$=?N57i_Zf?TAZgwpp)aFRA5E_)8skG-T_)Ph=y1(LM(`m*P
z|EgIX1%$yyTMPKFOt9CMu6sQIoDi3=B6sqYmUPQaEF*p!R}#Jwj=|l#kh1(ZRIW0h
z=ZBVw?&=a~b_Bh-poDpk{d3){{`qot{CEH1Nc^M{Xen<C`(;+TYy{qa4R#>nW1~pG
z+z`0tg+!D;WVX;N>~V|R3~=mv_&6XGfcf!^83Ev6{Mo!U7L!fC)KNO@=d<X}vjQ~a
zbg*5e3~4gbeE8GofaNhs7+Kllt#dph*q|vQ<S#-g9f?d$6wK93n_`WI!$t_8{QCyj
zfJUfSFP{UgvrE+({vu29saUQkTuFwUsz3^b3aZtW{*SeN|NhhXFa@ANlQ<yRUW7aF
z$*DpMJj8uN<n`+G%q3c>h38!{J&6Hn!_b6h8Uc}x{)E2<hJu8-iGvcR{8@-`VZw}S
zvXHe(u*um}GSuxarDYU)MyGzJ)oDr3OBKs0xKueEh7~-bq+GrvXT=EJ!y3;g{;bMz
zO&QfbSwo^k!X9RM0~;ds_4eh{X#D<g{EHdsAnZ<gp^T;BJpp^1za>hXq4uM%iPjju
z!MhMOrODIAhbkVos7KMF#>aKuo-egEGhW&rG~yJ9Y}>l`ic_=G3bUYtiw)NAn^zIf
zf0+`usG6EoDlWXYqNJs5%~gpB47>_ERknV-`o4jQK6%ofKB0aWfr~!2dvPv8Zm|VQ
z|H)j)=k#;5Eu79FDJH&yOm&JK8CjX0ATd%WBxH<(D@ST??3u+y8^0%qmEIK$(iul5
zA?;uE<P$)U#6X(No_%2hc?>ejjDcf!5|?<M-o<GI<U8=-H*kM~yu(@wn%p+(-#mQ*
z_6xMsw0s@q=J5@ut|FNo_zzR^)jeICE=IIjELPv>bb_hJ{iquD?A}hMa6H8WkrNe4
zvWjPLXiM!4N|fK~k9c6YP-hePTEt&vgS~@}7m)ee+!0`-lO^|bxyshpRAlCb%rirl
z7Tj2P6Gt<}?~G>4W^J)vS&&T*z?(;}^i|gPCY7)Svb=jEuVYL6AVZ2+nYvbY0c=Ag
zBEWQHBE+9BwA*_EC`z72)7ZUMl0<Neg$_#eQ+txm(}DId+1Bu7f6}RijZL+?;Ydr^
zT|9xDR3tkKIDY$5Kfs&U%O~GE7S!$J8aW>nT)seHKgcoz3Q8(sg9~6Y4|y1e9Y?~s
zk#ze_-NYsqsjKt(H8guYmXVF@^VX^(Vlizs`9d72MQbq>AWpU=pC0g$txGiAOzB)N
z5}Np@PLtOH(;?df^X|A+S>1BSVo?3+>u)o=xDhzI>$KrAfx9M<FXm`Er|<=ctFKIa
zfMebJZIic|;k*xCTm93-Qut^SyStMTXY-71yjGlJDCoP$bmFj6oSFG(2<Alr$$H3S
zQn$y|1J`roMSr#38ynpf3CMT|)F*r+6>8eibwl_IL)?(9qq+DEla;Hm>kIjFg{{16
ze4+Zl8$C=fInF$12JWW=)Y_pM|IjA8=~li|e;jN~h>`P6^vk{k|67K?Oy_Rb43whV
z%e@gX<1%#M?yfmDvms$!Jp;LFEkt7pBNp&7N0Vy9`|tJV3p5NjfQcade;krWKYx{O
z3B06h5wrL8i`@62M^Nu%3L)CZkHly~a{T;2e>g*SKZ&^}B)ffNL_q#fzvs6p2=jDv
zI1HuR@#yo&9aUbA5g8+!Uct%gRL3W)S(NM_b%#OZibp_={O2d%qE440d1-#1^46Y`
z<{(<)-(O%`dR?NHM|8URd^Ls99UdSEj~Iwlu4?}jD{Q|y@fX;_I%y*0ENJ(HqII3S
zA|)kZyc6=ajgQ@ob2S;^=Ah=(>rdXWUbQHEo5xZ9c$<9c8{r!{uIAq}jwM2)zJ?$C
z%VSe@Q;~4F{!<DAK|XNx6E-bl3MJl)&_gW&xHhBGG*kVcF|)R6FmrWm<zdfb--lm7
zw9iRUl>5~|ub4YXk|X6bnCHTuxubQnE)|a;O(kXYHz6y{0ZLMQ1bIRzJo=;c#7B|P
zLd$g3%fkzp2Fw*KlmbxK&~P;$Iy%+MH}Qe-ZgxNWY4NkEzCePO%~pp7ZAS}7?^fM;
zrF$nB+R<3>dStd%$Uz%Z)8DHWH8gB4bZjRf*7?TZhl4U<yAuZ;EPBF@=tWKu*PSDm
zTlf2(>_fcM>pvy2_|pm+fH;greqWvJ{S<l=F(IjBx*{JZ5gVZgH}0&j4n5&Yxl2Q{
ze9b+(Zc!u)M-7*FtWK`?J$KZT+C=s&OD>$qzF!fabzc$pHqVbk5G7)lK=J^G13x={
zPp)ynJwe%_LhHX$=NrMZUP}A^Z$6K|HwHdt+aL;^mS}8H`}Bs`dvE$0KB$Pf`W>=c
z1FI`FOt>w}kg0nNSGkUON14CFe20hnD)5T#yTwMJz^bRmfr^<d;9}zp!1`31-Rokv
zM;FTatyMyih>m0tO%j3?*p|@F>4%S_DEvg0IueEMEZONd6xoYO%pdgZZ4DB|(U^6T
zPPZE)qh>|1E`4+OCTnAgnX-od?#a!d=4mSqgRkm27J;0qNFd3jyBm@^NrqoHVx5I2
zZ^I#nXFtzXBR-sjF$gi+90!EO^|6L;#yjF9Rs|dK3NzjI;mFJk;8#M3Y!9>X%qiOK
zX{<`VOtThDuxYKO(Dq@djMPgkByE*l3};erHq7ZzJYj6r4Mq5?WX<T<s6@cvZ}Om6
znU8wT6&Zk*zfIeBVgiz+gL52EV>qPg^iT`G7D)Cn9f2=t--y6frOCOyA0fKH-4{%v
z^3)NC{^9eBJKDVU7=@SpTIEgDTBu`)t*PxT55H8D>tq!X(K-V$@i8d!?}1Rf!8z9t
zoM-n%Ao((ILL9x}xTKCT*8RSnwbf^i05DK1Jfj!*dfw~3B8$to$vp@Sncp*}RM;93
z=3Y}Sbc$zqsvW=BZQ)Cuk10Cohnm?4%hVzHDQz@-Y?^{OJPfM)Dgt^;DqF6jz;m&q
z&3aQ*4jw!|&WLLEHr*nsnE&RJWeDYCa-#1%XG#0ev`>hGQ9z%dDBTw11Gl&>vB=L$
zG`5BMO*=wh%B+>f=FG20d5Av>)poS{h&-S=tTw_2`Cyxa2!%WZ|JcR<fn4M@RUG})
zm7!Xw8l^6qGWtIj;Q@gDZ~ftw%im%)X*do!Tu+;DHP)niW&XVUgJz8p@*edC1*Vrd
zE8w8f#N$={OP8wk+&qH)r`NC~Dvg;}(&X-BpdT2X+UM)$t{Yl^S1gWwa=8sDc^_@5
zOe~cKrz8c|i~s_}N~jm4j}{KIrOqI1;ZcUbH0Kh>@jAi#GzM+FaZgnek%%SujdNHQ
z6={`i#k`*snsLmzNhP$|Rm058xg|2xpR!nx^*ivN&C6A*4Z?PM?xI`$_ZU`CiNhxf
zOKkZ^d=iqk#M!beS9Kz(TLvIIk~tC7)}(2jPwr(fZAwDoxUqeS0i-nKX8T#k3+Ho`
z8W-Yr<X#ymONCvOdJ!4k_bY!MRDM*Bs$kf_o-~MX(su7Inu1Bm)?|pP4;+v-`;MtS
zpG5h)9Ec^SM`z&gxP!^E;vOz-M{m>*UzzQdSDa4EH<{P2_)|@_j%*j-@*D;IYtkK4
zoKuyN)X4c>xYB0l2vWQu(vhx}`%yq5*LYqgU;GMp0SGvR56WKhe=Yg;Gs5Ji!#ya_
z!Lazr>Ox5To|9m}x!T1!M67ECf|9mmti#nBnN?jd!K`=FuTZkBQOLs5_XyBZ0+67j
z0~M6TslTtz7_J-Z)HDu0P&BovVAhi+o|H5?#;&!E*f9CXD)#1jOFJt)FbNcQF?Q4-
zC;^eQSeIPrE2$H<c*T^oq)zX}daYtW%>;sMrUZ}8qvdda<p8!Z;gd2W+5j)Y1~4pJ
zCS*BU?%?xUXi~3WK3i<o#j!hGB#;a%;!Xv3pL%cAON&E9`X|3Q8k=uv<*mEVg20Qt
z#%F)KegS=*HpMlwjF6DLvZkQUo7bGyZY8DsQZtjCGG#OX+YD4m3R0w0{jOBQ0}wOR
zXe5y{PHx{R($t)L;6B2{MNGL)cZXx=Y`RPC(r`Cq37=Gg8zZ|NMJEhO;?S>#o0ldx
z{NB;f_y9=me3Ru=)}TJ1d&s_lH6W{{QbBSmWE6_rg6?^9TuN|HN@vfDrVXFn8B|YS
ze7bfu39rxB7JH{xYC-aYs1e31ER^#pJu54ukPmMsAqHKoSW%BVZUa6zNJ8foJ8_9C
zXqmh~tHtOSdnn)5qcp%Ba-Yu*&5ITqj!b|8?@JW;b)PQwgq*zZE;V6`D_=2sL#KPg
z+J_p<FeFlYv@NOGDUY*sn-P_vxt<{PHCIyqn||GN;?f1Hg!+5ZWXJ221`C6}XgSeV
zx790m;*untwLtiLpK7#CRD6HG5(BLLpI<`a`e)(Qk`S0G`50&nVSv~OH5@?hTp4Zv
z6oYji8xP;7Ta1sHrCUG#0RJ*!WO>Cer&(r=9xw+bKElBMSs4#$%-2$SU~l@$Gc7;M
zox^%<K_c0Gs#xJ_4_6vVT|9V~LX)Vosw8e(nm*i@17dISbIbvSk0ZedmEYpOp?zKg
zYt_5RsfMcqvi)uXM0!kZ-X!!<f))z|$?6ydd`2em(E5I#T^LEsi;=Oqku)PdBH^kX
z3w60xSa)j0+Ye}_ClNugkP}_C*O=2Gpnq(25-0d%4L!MbLEHLe60*qMn})I(+72Ck
zTei<?&6R$pY}+4`w|D%~mnGffOJx`6-1a#=2l#j5Yi(}ZU82EdJKh!Ui6#PYS{dQ0
z8tG*4<sNi#qdx9E{V5PK{d!kd^{Oi>E9uwP*PB>-PW~Ymgn@Q}|3su{Yi0L>yW^km
z-2{D1931Rnl})E(q{Scee@I+2i$MGxxcKl~H-Vy}qHWz_LpwcgiAZ#*lkN%)`Z>F4
z;=OT3l>$?X-I2oJ-6D!k5A?wBueF;`(;wD&T*U2fXsk`cQ2YF(*0~HqQ#F$wJcwx%
zwur})w^mTn0~Tt?frvo<c3ucW`vIcRVa{;uL7f{222(O2TC`r}gd6J)$#yb<fP&PJ
zh)(=70;0q9%x^iz+njFfN960u>I*!y-a5ZdQ9zvC&tpRGBZ`yv26=#@LrZ#}@NEXa
zk%vxxz2UhLPw;%-sZNEuHK57iYlV)E?V1s4m2;mU08rBwoX4B`HMe`^VfXFwRTr!p
z1gd>_r6X9Y{dED1;Zrhf*gMw;67m}yTg+1TR5_pgI9c!jFrdrJa+L*tX40vpO5n`-
zPEQzk8XTY}%(732KC{JQ(fZMsy%<^QOgh_c@ErO<>z7-y;0dURbnUd`&zD*<0d!wK
zvTeaIug9$|+b!eZ#K4c_F>PE95l~);8i|ls!JP|o>fh%N)`>9-r6vLl=hmg_@pj5l
zlPRP_mk^+`eSwvWk!7X?R&G@$6ctwA_XXPpjac$etez~^p9mAM!Cdl|bc#dkJ=bX$
zRc9mh^+t^1*;JRJ!tN|aa;!}H`EWM;Um^UEXf%?#w_$fW$o@943qCH-A?}PD*c{h$
z-uJ_<l#=!Nxw70lF(Y~#=*QoEeTZOrfh7Jr##jqz>WL1RYHVdI*WnG<h?O(;yPqi2
z>L&CyWGt&4dMoaQx6Gt8=SDH>zP)7C=kkiYDIEFQn=~kF1{QQz>HjW$Gni(!sA4}+
z{MuZjIO?RROVdZ|GvBNL*3tOq)whUjj#juvbkP?j(Vv{g@aE)zN(${vAz>^Fm|W}}
zvy(aIzzwDq+OHN^_iY4104VXWXRC@J{il;o$k{4aSwum-wUlGQeOUfQBc)QG<m_o8
zs9Zvh6<XBB29f6MTf&9@TDy}l$KuW+a60eqLz7s3&)bdDuoYLO6igZuG`nWlSQ!e&
zWit~AO_Ddm=jD9Yiz5~)Q_Y&S0;SyTjjhuy6!_~*>V^4T>T;1YZFbB>MSCmzNc6?l
zS7nY9{-pBm5u{~4ywf=Jci4#MFI9Wj8lMXmhnr%!|7Di<#73P`VVkx$%+>*bb3JU;
zciv4_4T^}7XO<e{z;?!LW7WccjH~GqMSoXcLs@03n+;s1P}{YZms#Q}Ii+xIC}J-~
z;<=r-VxYv1L%|A(fSO9i<QEW6D~(!6AvnZJYSGiCT;42aZ2&itEvvIu(jKm-<lurv
zzl!i#+TrWevywg!QyJkbKL!R9c_?p%oJ*(u9m`rm@oVzE*}56#3h%f)JLRPXnG@s5
z<hJo2_0WR663>`c+^BK3JN|Y-kA1&13pR<0HKt;>;)gc>9sK7A)6N5U3k`vJ&W$>d
z{+#Vz?}Cls>x-$Z+OLTvZ>qR6L#lRzS`8y~))>|9(G)@~p`o3@BLnyiE)ouA^_ZyC
z@V-NoGV=|6`(Z$}+We>Q3`Z{Z5QpxxvW8W`n^5g!75LfvtD>~~X$Ckd?sY%kbCDHL
z68(m2o3Z-BgRb53^oi^oF?=zrgP_l)t)1r2BrE2*g`)TuQqqr-l9GA;8gjzIA6A;w
z;7xY>uWneKAp?*H2@=N}3%faTxng~?!&RkLO+T8#r$jm$({}pK_3>TH7{>{6ldh0I
zA%C|kinrFd2}}z!BSZt!;hoa-4bP8o2E;x~{Ma?HOJfnH=<~HUnsF2Vbi!jz{NoO}
z|IXuL#+o4!Ldk##;b+@pIcpZ_d%Wx1sv45$75y2Ym**S13rWO{`#iMxf0eJY7GztH
z7hh3P^_#5rmXnRtc-DSAe0+w2;L2)!l;^K{9^?2Quf+^2IKIVwbU4vDR#!NYOF=H>
z_3LK7@=d=OTXr$(cDjJ&3Zqdp*vPPnwxNVol~!a>^6_yhS0Zu#7F(heA|m72j-Rve
zXeCq0B>yG1_JYUA;iE`egOO+$fi%QmOtKg4cx=6Sf&J9#h%xKGzgWnT*)?5`TfD4Z
zR98w8dT-bfNy`V03ZXxaIbNhw%Si0mprqLJNzQ8|V?G<NG>a7bn#`=K%~;L99RK_R
zXaIj=fy!>=DAOi}y}P|+;lkaT?%krNi<PI_wA+OSze8v0nuwtB02DntpKC)|Ih)WB
zM32?z#G}YXb-&vn)(H89EwyrkJxg`>ovIuo{Rl#DHD~!C35KYvv7KlAUFev*b<%B#
zyGVTP20Pl4Z>7h~caOEv<QW=VDg^Ifxy32}pHrocEZ+;w1p&0<eNq;G#gEl;a{F3(
zuQD&<I1Uc>!sA<?9wJqZ6j30-ntLXiUm)Rep<h%Q3(VxrZQ#U&EsmkoIjudem4U)u
z9>aI<b83jm>2`(ddxhxGPM?fe<+v{TW;1Gz*`%;Z{r1Poo#QD+m48Sv%MfiK>V#E7
z2G{D;DI@P??0w(ET!LSGgE!Q_gf4l@FB=c)r;Wl)Y;C`iGyL*?Si%`niY5VuX;ex@
zjDDL7`yMxsilkw=!Uf>l^4Ztjak+c>y+@>^gsm!kfCDL{KcI-<Lgej60*c6%{Ts%Q
zA2=tWGls<L)$(g*lSG(8k+*l3=RC#E=%kfDPxN=^$7&bePw>!wd!K2qwU=>62*-}4
zY4`qCW<^&ywBF^Uyv8rzT1GPnM4UgoEW3VseK^PQk#3Tu=(87H$5jN``ObQXwJ?hw
z#z`18s%h5QjbnpllgpF#{lwM0@4tU>d*RbnM=zR)NY-2_!X@wb^?AlrBCDl&RBq8A
z`Eh;`_5yp<QT5%YpT7;7Ig*Xg<Z%-W7+|+a37cZ4i9mBS+Ai;Bx?9e_u+wfLh*{%p
z2b|~9($aF*xl@8iRMSRZ-Lg-kN1N<9nw4Z^`#(-vb9cK_RxDUYZ<ulQ{5HmjeW37D
z{L?c&>m>SXWv=QM3gM8QxJY3VB)#%;+vxWN|J0Mp!!bla(~Sfh3U$OY%4J_Y<cRVZ
zq3U1Mjb2I0hJMVMbq4>wHnAk*m}!?$l<jXQJ0TlU>&|IM)`oYX=d;wjdlRpSS1qvB
zlKg?Qv1-$vYrwNVwg7_z>0VJ6B7eTRFxz4z%|f*bF`{pJ6DqB%EpBNJ?Ft&s{A9%8
z|KgW(lWSFD=!M`I&$n&pO5ey@kv<~;aqEP^!p0{H)lT0F4p8?*VcOEt@OMXkQhRfc
zcKwf3()~v&=?2_(Ri;g%4UQl59zAk3(f`?~sDJ6mk&Fx7HNXSb1P1q9VSsQhABG37
z60;r<S`<k{Hp77E&R59`tBQ#;dsB1WBvtLaJ!R@VzsO`qF;WJFX!{Z?GV}q8CyF!W
z(nXxA24j__%97EoKR!iMjyhrSqvd}$fSe%a#D;>N0kK$cU#>HRKXBEk;{X}fV`ePp
zm$@O$CedY1RX<})g76Zzs*+o!intFhN|dw6Nr<EYOu3M4Y(}UU5fQ8%dlp?1Rd~v6
z$Ht5HPSp1OK;<NtyY>0qxz7;t5yua+y+&B>sTWE<{z+Y9d(SuSOSx9hxjhX_@pjcp
zzsqAJ+;Da3P%QNbr2dXyy_F<@fBT$hmaSF)RP{n_+sT+|n*xv3BEdWP%+2cV-9BhL
zK{{MSGsZG3Wc$qW*ejA5a*poAgiJlMc*9i$ML#pqF;=J#w)=Idv^$YDJ&>g|^E^X2
zDN7V9xc($fT1(hhU|DIVj5;XgNI!O_bNrc*7I8PEWM(4#TUvS6{Gw%{dg@ORRC|s7
z7!#VEq<L!%%hbF^LvB_^lz>)^8MylcC0lldLFypNrBy0w>-#m%3WLEolhVZ;>QBB#
zMHl}_uSY~R_w&AuO2ppq5{2e23D>`eYs5`szZSUt$Vy|u%O%qjTh8Rh8s%gimEO?*
zmjz%yaeu@71et*2RhtWi^tV4;IiE_;URW%E;AZ@}&N|7LZ{0qVMvyU4VQTg+$il{f
zeovtAe4&}bMeuYlq9PXj;{X!7eeZguFTI$YxysILz$eL-YS=U7HG#5Wcw&PJFA<>P
z^RsmD{`9pk^k7M|*<z|ZGjjiflI_;bB#&YGUnD%&6An!!70A(+MiLb@f0SVM^!QmV
zDaq;V){St5S1b0mq`f&pdUnv{vklkLC!%bh&m^6}J_n@ACCF0!l#&4>rMWdE;vW!i
zap!imGex5FD&@le;#kQi3HtZ8<tbF<=od3qH#sH%;qAO%;|7Vs%9aziG45G$rQ8^L
z%7fqO5;r)OtF6}Y&YmYkw+bvc<M__kZ9U%iqWIX@vSuC6%U0{%!yc?SgpLf9BoNPy
zgyN>MqX+!hA-oG0DzL9&{LAkA*SS$(sfrfHZ;M|^Y<VSqjXeVUR|X!8ZiBOBwtxGl
zmOnz2bT<fV<-BMSKu|Gn8?HZG?dmr$bkA%uOX%Gh?K|<rR%zdtyoZR(y1;$zUN^bf
zHHUNWKqzJ#UjA*?`}6k$)lRGA!JU%>m8JT|T<Jf5%euL42>k@GiEoJdwe1!>VR-K{
z;j%j)(}1+)Bizm2JYa8{jquTf{OnnhZ)e;Z*7}i=SKPi%L_w6JVb%jucTYsq-52u7
zHQIa%K218MGtP#*%{YSo2oM-^p{dbN5^TLb`!c+Uzq%v+s@nIO1dbjX@~v~6y*N6@
zyS|%hT~9#1FXu4gxO{HXd>)>R18sYzqQB~M3v-!#gJsCj<S9J+Wpn`VqXlz%divXH
zlDJ7wDh5ALqf*mX%gDp(e@v6@Ki9BUf+c#D`G4%>n18TOd)m$%{wL3)(9Zo96!Bej
zrG;;hRLCNaAiv<u8*Vz$A{<7^hPaT(!@z8-D=*9>`32eav8Qt>tyULT)GsEuhNp4m
zHbb*yR`zhz=V?XXtcsoJxb|q*od!d|gvm9(8=F?Q(t3x_=V*Ow)e$TKXcOJ=8Ecvg
zvt6U>gUO{j^G~A9RVHS3Yt_tNN_sIqPq#ZYPfxLNQGa+Lj=1=*pr*u`ha-;q+)Q!E
zzxRQViaEeRP@vfB`?Ja-^Jea}?<LnhkF<&7(sxJo2AkqK|K9ZEEwD5s60?&U&(1fP
z%Awfm|MGTe)Gh>L;2&XM+P5rdW}X*4EVNRP&>eJ{Fc9w|w6^wIEQi76$A2R>^yfpY
zl~Q1w;OjRE3!>4My~gC12iL9Ufvf2%j|`WczOz3joNQU?&XLTzLXO*5EGw%de9TBf
zQ!44xZA61j4S}e|TdI+@MRV%<-VgZCwV?jUOYc_jF0Gm|c>#9uqn%C@Tl->p?6m|E
z66j-7%p@d}SNFPY0TZ@v=~h0+r1$BJp`tI}k4T-BDX`(bP<$pM3%OK@)_{}F8M;=e
z-{d~!gxg6VhIVwoFD6)LWc!;z7oFDv#F<0fOXPER$ZEqY|I+8`5I%CICr2zbw}?mU
zm@vB1U1<W}g%x+h*2G)CK3qiRHQyqx)*w5aDn-411&o#p7&>yf6mXZ3X&U#<WA~b?
z94B5(Tgi~uiHdD2CAy_uuYAz1J~cuNs6-<%194h_q>JNJo89WFBHcJE3Ie)wPMwu+
z-nP8AFh6x5Mg0g_4EyM-(l1;AQ`-$0YzFG+hc@asbGz5G@e>)?o-XH4*n?w3_eMY=
z0Yt&D2%@0+mYPHiJ96@TJ?SrigZ}y->tZqQxtD~t#(BZC{tMyV-EnzY<XTgSuk=9X
zGC!vGOd~w7MO?hRCvElv#JZbTgTz6}_rvo{@2GTl@ZPT&3m{a*yc%Dc>D<rfpcOof
z#8LJ75YN7eGQ(N{HCQ=jLA1)mEnK#y;@j=q<gFreZazpDq~yW>eIJ(GV!aD9hB{Fz
zWrj}`QSG`Rs+bWO2^TAuk1dKsRt4bZ!0cdz2$P7oBiLW>{OfBanyPZ;dKI-CNM=%I
z###Cml1jY*K0$cDOzqK`3k={bI@S#Z;eBu1&c<C~gy(0b36gP>&<Yfl5>+BDe{8I$
zu+SxLwgzAVFKBkS5&~4A_11=N_7I7L#;Xs)yCYTB=nf1;54mqPElE)H)p1uZ;MW$A
zJ3nm#TRpeFWU2+fNXqwAm!uH=CHP@_!NT6*pCr08(PZXKDlR>u5YoPHAQ^d6#_OkA
zruVgjp1cs{(S;=QxpDvEbW7pV=#{=fu5SCOgSlcD=b+okUkTFa)m<63HJH8*seIW!
z`|6AoL!TDyEEVMt@_yA-DPpa*XYH8z13#(;dhMSWxbjq#T-U-(B)Zh*8`CLe6Mx<M
zMS^c^zf0Sw&E-Iby_?oheX(3vcV5oLzgv9N3Bu+17?d{oo15nj)=9)7ONCd@ZjkJ3
zboH*&i?lksp%HHgOb#;DobgthEH~oC$R;1$TfO-toxgqO*hK)J$6$jyMR_Cm=dnW(
z6L@(bQRdUYKjCO3%0<!`6jW3dJU_Pk5jgBuIQ~D&7}>vN4C+pxqAOM@WaLyVA{NVm
zt?P4UzH9Y0sL*x;?*CKnM;J&LDg1IkA%7<b!6AEZ=93(7@vX95lnki4eWQ)BuqXxx
z>#s8I4x{Ru8U4Y*4QuZl@q;hasERj_2RF5`A3nun5mWOk&fwXNVmwd`QE$%0v`Hrw
zn{oZu@fJIlIQg5S?Emd}qpD+dYx*i8m&wNC*6CyXKMR%}zSHA7UQTqD4x%m{IbKL_
z2sScM<T2|KgKr^j;L%tKn}cmIz<L$GC-EOS7C@fI2!((!lJ-30**}0V=_+TLX0nF<
z_FUWiaODNbX$8Zf6_^dh*<@>!DSuQ^Q2`AhZ2X^Wr|3Vo7pdn(SM9%kuLO9@lH^2E
zOCIgLh}0DQ_4Ie?^btE`{_}TeA=69WzzZZFqbXhjxOeAwdjB<$9~D%1jiB+Q7H167
z7rUslMZtQD6Yy7R#s02===AuoV%dT%lAo}9U*<dizBumjj}xNBg>yJ^#={Y#@e^Z@
zdiw}m-X+8?^mS!FUtT)Vcu&y&51g_%IjFC7b=Y2mXL!@{h-0+23mBO_xxS}-@kMy|
zdq@GHJwBIZ-3Zd@LIY}4?*>8UxfG#dXkV%O`D2Bl_3&;VEC7&K`bV$bldH^r0#NDq
z2|-HV_bx{*3en&k_zUWBWDH!pyQcW4P64US%rBSVGrfOzi$gN-k`UEFn$h8UcSQTQ
zanIMIc!GaXaY)_`^IsLvuS3gT@Y&&gnIN;a9cxh&>}O4`eIx3JMTEDEc%bIy^9K;z
zgW#Y0kKJfKTY{G_<5GW@ymWoKug|qe6z@hYw-L?`DxS8$_92?JOuY7j!?2VDbFXcm
zdb<AiDgQkHkRRyBz}7=NZg!V$kK+IF8++PS4xnf;SK9%Xs<D8IGCiQjzu-0ge~<1k
z15^?6L7g(BXJ+MWm8MTcce1cU%lLk9<gdh#immZ{iNEFg-W?A_L4Id}VlLQW^#@A~
zJddQ{I@QxZl1PT2y;_NWVPQc}q1Os4FF%^x<+@M#uLO?e-}8|P6ZBsziAR5DxnT+X
zpleMZrP%{UT9zkF*d%?f78AnYy?Zy>rSB%+88ksqpA*#!k9T){wJM<qd4}arYM|rJ
zpamWtKls{oFt$P?aBs;7s35gQ_y7B~yNLNSuAqnR3x+yrN~>o43<v9Lh1Tup&SciR
z*`($XSFPH_;zwl9#7B|xi{u*4&d-1y(mS^$0TYi#Op<~Az_S<*hU_>5Mz@7}mpX#j
zXy-X**4o)a9fi`7h?rl16W!ae7)H51zWZ5s<@g9jBtPEQw;{KAPjngoUgBa5`jX2F
zy6;K4AN-Hj{3o`*(?BX&Qc<V`IbJC3G_%nK(!Xh;X~^M7*66nhQ^)UCAzWSPw`sT2
z`(^L+?F84Nbqm2)(RB-rg2nv&H}9q1e7>B+x#x8E6bMAArIM0o_N-KCb9r1pb1`)y
zq72Xy5KMbyUlkgw7uTOU&ZovhenUv0d58(oj^g|Hi9Pr)J*1z9=(`0Uo*r0F$iWUU
zOw2m8VU|A=(6O7E-?-2;mjBos9%(Oe8)6CoIGd^DE3s7z*skV>wRYllVnwyDA9k)%
ziMAm)-?_tt2c~E2z+PHtenN>%ieo`~<YNCMg2!i9XbIV5t8sG2>_L-v2@iaxgu1hE
z2FV*H@`TI{;FaU5kI?2bDUhU$P-ChQwAS&AOP^|hcOQceYX9wEmyD9b-G=mko>>U!
z`S>p>&Xu;aP^m=r;DCIy`+nMmpr2y#^!~mjbeRbg=_DhkuS|>Eu%?<P!5!nPN9CG<
z3vaVB$mQ~>NMiOgqn>vCbjjcH7etxWm}NoK=;GGov2XQ7H09bENEi|0c-~k3SNu6H
z3B}kt6P_Nqi<Ni+IW9Q|-5ZmxzUL=vldd3l4`s)7>&nbC9Y-yGZ-O)WBPV2%m%g@o
z+$&*zbhLix*TP2*C;@&EG&K;YsGuu6)<2qLSqp$buK(r4wu7oFGIF{ni*N_3YcNKA
z=?sB!P5)OY(yH~A=s8+uKm!7IW<VkI;Rjx)B;L#1TRe5c&hcRvsO=q22J*tQBsx59
z7x`1Fz1>XD)8Xhz&0hP1#Jt43x?X<zzlG2!KQ=q+siL+-`;?;96vkB;eR*^qb6>ht
z9MQRsW!G>2$noA(8TON3dFc8{LU|-1<Vn1oh5{Yto3}5<uSO!fQm$UE4uDJvE;BIj
z!XKc}0|Mp|l#v&E+_`PV(zkIl*`tpQT};X6o*6Ds9dJ^^KfrwX@bTZ@JeM<G2m9(K
zbj_OxPjh4jc|W$-iU6msyL5C`@bL~F=F*nqM?2hWy+FJvZI`Y_J2zE#n~1goO9uG<
z{UldXrFmFr{bC2v;jg)l%00ajaV!yZbgGM+#$)@gI%8mZhkEvS-&<z#xysZQn-fj(
z+WXrqGmp*xhpw-VimHwJmC!MeQIU`u5dmqD&H+(Skdkf?>FzEOP$Wc3x;v$NknZm8
z?#|(!fq~cWyZ5g7=UwY{hI8h5&fdS;&w$&h?o7v;*VkBUsxI{<M^e|Ev6G3R@so-Q
z)32~B9oaNt)jRL0I-+_m*Pkp<zO-r8jfBzp{Z-{mh7C8!7oSJ2z_ZV%InOqzRoNRT
zcB-+tzAeClMP?Z=fSbR8uH<{8&V<N8xH_s?VyaRo9$QwWS9QP>%yO$}_5Gitfha20
zQ|l9kjc`BoN%OdgNk}dD*wv{U+6lF2hnpS~=<kXgJtT#N;d9GT&Ujjw)DIMfs6NC=
zYOa1g^Xx1uM|X3B;-%!+GsbVB4C5tQ-sZ3p;0H7sKAJA_Y7oh>?z|^D(u($9zzE}z
zj-{joMyf40hiVM;rHjW}H4JBtfzt?3Ig8&a5hGSg|51wKlPpsPr9)j-tC?o&{Oy{X
zR7Ylsb~huEOR=ZweWPj4qfn9^jYNSZv!Fv2gKNg`J1QFbiis{~2QImB;@-$`zg$p@
zmndFS9GkZi7J^<ZDVrB2cVt%thCN87JMY-}io+anFqTtN?mS&?av1UM#(5QVIh%mj
z;Q?gtP7HX_O?q!iHQbt-dSX<dWnx4)k+GpE`-dz@%fZloDROCv>YdOly6L%@%uQ{r
zmb{kiRPvZGm$YYTUY#fwH;8yaqbi_%nvR}*)^qiFb~rpYD7d+$Bik?_9h1xcxOz+|
ztYS>ZwWugrtD{5?w!VFS);ilUPwu|GAtHw3eov_t`}^b#(5VX>FR0VCu-~<O)p*yh
z$U1B$z$uDKnTQfPffapjr8eA@NOHMzBflV)*dm<TxBrDnzgdv%03Pi(A1*OjrNL}B
zg&IxdKLftt6D+o4=LRNbQdL2?{=BR%md6W*=F&VmeJZLzEWDJkZL^y37=moc3F^2u
zGLmrnA=wWzXGwReK<6Y*o#5Svjs{*g&iKm?zzx3e6lQRnc-upBW}lpVru!d0Se$<?
zcyf=5%8&BpGu1il25&JXYsZ*4jbpif6%4oQ$4rK7bEYerNn;o{EsJHRH2(>oQ2CDq
zm`ugQ;LVHYFiY6wQ!!>KYg0=h6?0K-v0bzgPZ$<~im{eLh2ff>{@hfTNjT@YwShed
za_V--RyodxH8TAOQVZJRbBM_CDc%EO#xlkOg5g$AfKzIWRk67~t;BrDcBE7M9g#tl
znp$7y(9SKW5f!06x;3xGHp1<45FUkcAHrBu58d=j(rUy4KI_<pr_lO5=Pc7%@1WtF
zWt(uKQ{0F4g-$ejx?9$-)6d^O;{LW>WnZw<ub&}}SW)ah%SvaKEoABahIpoH(d=A^
zyCI#BU*KW`p3M$xp}mbc&H2O6KmG2S)mw;rG+ix9WH@5&OwNBC$Wnb4^_^)xfNOoL
zN8hvZ)f<}7W!w2ZgQfkf290=6cD)L+2|E2d6#oz?GZfyQ#<1*ueqkBTX|*O=Qf$s=
zF=($U_DeoXoBr8;{hAQ>UfGG&9zJ)PBVD^Rvcf%Z;U##KGG3&r53wCJb)OOrP7yOO
zAW-?EA*Ng&uRtQ9Pv}#8*1p5MJRQ|`Ua3E*BtvI#5Ohzh_~xc3l=>bkecZ*D`~cGS
zFeim)b;4T8uVZwQFgk+u&7T>2OM~2|zN;-6cyU`ZSg%-Kn4b~PG^?YG?QR}E+%Hi&
z<*xMbbDTqghfzPMLcu&lF?GR=L0hcqSg8^oy4ifbtVjyWjU131I(Tcv8$93?G8*_Z
zCb#v&%olj<UwrsvG$C%0;Eyv18%sQhm>lI4#rY|5C)(}<Zi_v5%;elazVRN%jKE&b
z(Br7cF{k%3ASFt5A=g(S>x@~1gl56;KnoP;`QcGR2h_tA`M`Ffu=ctRs2mBc^gC($
ziV0XrTJ`5(Hqw<mziJMa=r26<wV{n8PU2lVvg}zLhSj&-_qUSKy?@aQjlVeJ2O&R7
zAqCCdzJ+-~c_E`^^n_kQO5D$?eS3Jr%yWA?pMzuA^sl<#^Csva*UWEs5_!B_EIN#(
zZ#PdN=MCqPjt_zUTddC&h-=)Fl-y0&PUs(=zAhoL7NS$I6SUnQNPS%I<C{)SrW&);
z4qZ8<Cq02;b-l?i6JUagAG&)u#i}miKCxS9(k^t|1CGS2v>4qUW%Z<{8ihsX3Dbct
z9xtE@ke;MuW}cO}dy(`tqkmFqlx;6WtCuJ3fC^eFCTxGXepDamu)CJNF8-2RQ-5b#
za9>*B^;lY2zhmZQ`>&}Qf2bi<^p=YLvd;o<Q@e>89`$};>dU8^K$k<C`z?(*U*y|{
zA7GWeXFt*+<|IEg=ri<hzIV}>E;eP=+3kd6r_|}V4IhsNvDHlPg__Zh+X@($mzwhc
zMaDzFAw0Ao_Jiq<0nV)7cIIE(w+@UNa7-xw`>V95`h{yfr*n!#0@xyXl*J+01>PCj
zZQ~g<hTcESPlXQ|^ZMo@RJMP5>Pfe#vQ2_&Q8+I4oXf~^m9;*svmZulG0QIT_U$r0
zp}|l&o&Jzhoru-^pS)<oSe1&h-ZJO?v948H;)Z!6-OZOfPXFz>1gJ?6eDjATs9;mm
z{I%d%`U`rcziu{g{h>F+HJ5itdmDfszD1O)sf5_;=}cKRZJQ{i?62Ngu58=d41(xu
z8Mrn<(|=1uo?W6ZH^&j--57apo6WsGDo%&m#_S{kuuQX!es*cuGnx-U?(QBHXW5mn
z-88a|8w`Sl`4Aunv1YD?q&fSsp&Ypf1T&~^Nuts^LP8_TtvA~VwG`8<*R+W3>(0>^
zj|otfW2Bi-{#%1dsXMI>7H*YCNyEUS+^b621BzazJmk9C`<Wkrf-G}ucyy_IM<{SN
zi+!YLVL;`01w!Y4aK2+90~><rWc$txKEDCZQ+VWD1^EWry}|tkyadMI^Mn`<Cwfbd
zcsm~)?0;NpHQipB?h_qSK2g=uwky5+>zDUE`=i4bpVoV}S3k10^eSP+5<j(nqUi?n
z{rcr#B{LBPdg^(|2kjI)_MCyTEkQVQoJY^L%nY5+L>N1^4ni{XrBs}p)}l=cda!$N
z*JZj_*XG$`{gR`ctOCo+Y{KM~#}ZP86>NLMx%CMNsN=SH95%tzXpX#)$UN;O$ReZ(
z%)tXysC394chE>?=UG}v?e6p6-q|xq&iK{BJK1|@b)-lPI0udgINtp}V$T@~0AGhy
zu{yS({$n=%lIC=M?{qRHyQr+n6xdN~9enjWY!mSvTtES!DoiF=PU^HVjGYft5glcR
zswRry#nrx8X)n!4O%WGWi_vRtxo___ZTt~N-9Y3lx9nhL%tJUk7^Pi<c3Ih~2(5X+
zUHr2BHPKF7X{R7!DWK~H<zIGP1J8$Lz9nzo&N18+m`yLA{@vvg{t2WvTFgTc31W<D
zw4s`5YU`!jvBxJ;SMo7s2Chm+n~3drM$FD+d4a>Mwsoaj3!QQ+%ZVu9vzhemi07^Z
zM#X&54%^?e%D07lAW)S7C3By4uASC^Qx*uKk5BR-ilL`{P`N?+M{{9s>YElj62ol|
zz2#qO{T(FJ2mzq~Wl7gvtXer5*@M?-`<znc7iHfi{!_p&GhiC>;)7_vZ?9B53J=PH
zUTdBhQxD!rimJKcoa4^Y4Tmst7iR#6B6@I~pU6X<5qRw9rjg;jzYX>7j{N-0g)uQR
zzp|`~&9(!`gp_kyin~h|>{goCNxg$6ThpZNzpQc<W`C_e+58Lr1~z|s$1!Jx<X9)`
z@%wXL;MHk*R-<U_m?_3X$JjvO<{<TxF^MSmJspZ`Fjh)h+OO<v<9Mt=0~Sp!t=vya
zEnV9pw|T2>A~-}nYod*F;_LZYwyalk&UE6lwlv(s^@Hr(S5FEK`6d9C&nPttYO2YK
zyrovZ-Dpuzl2*zn|M|evk)M1bQ|00Oq(TptO5}*jwv!~*%#_9i&SL<6;?qYU;OBf{
z$s*cur^>O(Y`Ad&`O-tdC`N~-M|<CW3Jg?akv)2U%uy_--g{|Cr{~1u2ELJA`w7J(
zmyYk@l{1h-K}S@eN2_aDi@-L)r)cLowhj&_YZBbsn|2F_v*jdSL=!E!wlp5{ZoJBW
z6<@!8O{&q(a{`p*Xg!0s?TE_k@)l9B>RDfrXt=!^*kH&q<C`!FV=#R~pz7ble^qQy
z^s#;yIi$<rIP2r-a%LyzZv4p4{~P}3kyLP~?%<%|yA%)Z<Cw&?I<~7i%>j1(f~(SB
z88rXglb9vEIxmmb(4jgIRg-Q^dk_!bE|}P15mk8Phv{WC13Exe8oG+gW{^xayqz%G
zdo^h_LtTa$yNu?|WLkPwcLO9m-E6>FB=pr=^M&?;pNYRwf%ovn;bq;gcxrsY_qIcI
zuDDt?iy*c0hO;f6c}AOTrXITo2+qqD16m|GLpdWks(L?<c2m=HZx9-hM`ykPdn3We
zJ^V6$Wl$2*+|skWs3S-!z_tJFls_`P*3w|5nB!`t5v_5%`kCQji=;2so7wp^dnu_F
zvBl|$Rz=CIMFLh1g`-ANe(Xonvc|LmCyFB3ZHr8P2jL234|xv#-n$$=<-yN3^x^@?
z2HNr9@EN~^gnN{o-Vc5ivkAhi6ccCM4gp;iZ3=M}m8>Xk`{m?lC%vzq#!4+q52t@{
zaz0xHXe$k(4zZNSGSJUbYmcDJ+zq)I9S_RF>)ti)ns<ud+1S)uF5+}@Rh`6bJGUa^
z`39b@+1}a;NhH{vi>eWD!iMH%zj~BWvAiDk8;_{9iD7lkQTJ>0I;`w|o~-D7Bv`nC
zKI~VOG}N)U*^fH&1!ZdsKiI#^33tPVx~|sfqiAJlgTR}J%(4w07@ERgvTLSs16?$v
zim}5X^~;wlTJA8o6!mL^Kc?+YIdBuD2jlhx$`ntZ^4jnJE*srKpbj-M)W<aiU1P1o
z;z`AI?b5<<JK=<5i=EhauFr0^;8HSiq69o6n~pxDfOi3PmB~^~dW(nq;JbOh<|Sb-
zlwcXyN*)UjO7t-`_4Rneb)He)!95@+jK#+%zKJ%CAZI{Du8Oo&Q-$jpMxL+=Mqx=y
zC1m~mOBC%=AaO4@cujV6Hq$o`db0X1(%!d7!`<Yh{oyIIb*Bp&;Xj`sZK<{wwy|3T
z!{*#BjVm{~3Enn}r}@?MI%e;-lGTu=LW!AXJQmOKVtg34z152HH%3I8HhFk)gFd(I
zsRZ3msC@Fo@ECZDlpZ~|wRu-i;B!8l?r8(zIRtHWiGw;pd5dU{Zy;TQyInQ>k~j<c
z9aE$NPRz`9IUno6y73>{nMIL!KtSVI)xO{sMeRp(QOc)@KLfwSA><bqOu5&f{?j88
zBFYi{S@ugr;HCdu&9K+&<%j~d5$n!idN!u`s)*~lQp-D9x`N_Edgdc!^^TMQ5!>mw
z?XPFDG*D3N$u{r)H4Nl4Q@EW^?3$bV3Ygs%E3oX4$1fKW5@OQnjOj1K(?%@f<TFHf
za3k=Kl8sBSZX1eH?%3@9GD05&U<Cs=YEz-BdhoQ9{dRegX^6_nD)rOmY4ocsK#iQU
zF?~=wAq9+1C79Xi5+-1rl$RdA*sg;$=e0a(!v(R*GPS~3l|NwM7BfkI_5t?z;mtD)
zj!rs`oD(;{0z-+U(1pX378wO9dMFeSV3zJ#Q|s<q)$l+dVrP3tKP>B?M9oLg+nXjk
zV&XMu{h{ExL<K;fJw`j`$881iz!N@%hH|!Ty@!egW%&k1(|0)2hBqoU<^%>l=02Qs
z^P$1#ut;Gs9ycHnX1IwcoB;sm$y~kv4=`Yyuo|*0wK}{ppT%e$S+WxRyR1DBnsmxA
zm)#?J6c1VHbjmnB*JE1hlTJJCYR&E@$B_6GXsNB-FD3PY;-xtA+?+=RpYu1_^9dDq
zT&on^IH%8ih}Xp;<$GZEN<sRT?TN)jf=YxG<P^egk&_p0FGj3HWW{a&MkU>rRxxeW
zk#acYRkqZGB*zSf>w#bTiU0%HH$ndbHKimdm%>>9@nO&ZfCO0qHr<bIpJKR9)p+Xn
zB#HR&Z0+`lt;Bi~e<4CtLg7a-cv(Tl`rEmgLT$7R?ZZVtp0JGU%cG}4D7H3jHTu^W
zh{-~&tJzSY6s`}1qu18bzX%#+M*J=5U%DF*P(aGurxPePal;*Iyk=OV^n)kU6@h1g
z3KT<o$3XS8)T-jd)$|UMfw#0+9+m_|(?~3TQ5N7tMJ*57UTvH)j4O3rzLiE0fg*Gv
zVQ#5WxDdyHveH4>^h2p&o)_1>zQ$Yx|AuP-NsdMX=6&MP-P&z#`~4SBpH6-jt<gsO
z)nXz*I#SpQ#@@)5YK$OkcoL9>pVrd8G3F&ZDRtePV$2^7DmuM6pli0qtC$V;_aags
zgDgtd)ivL{V>-zBd<$i-l4-g&eB8vKI97NQ{#~PC^RM$V)>EEue#ad$ouFM5N0Uc7
zAjb^&JV9pVM9JOJO)}wL3YPxdCc`aXJ@+@6rh^AnXMerEE$YhKBh@^_^~nWQu}%<5
zPxtP%Z#CKSKjy3qk#4J_`|;3b^-lwnet(95%v%8gy}mSQoID(J#5>p!=*4UN=g&WU
zp(_V?+T}T_tlV2C5@|pjMNgF)3<l2>PpJY$NaVEa%)A_o&{ZxlH~6@tJq()QoGVT3
zRmreTJz~op<f!_YTFW@gI!)xx9jcaACj5TG#RopWncafO<mG<_k%?(&zAv==tk}>$
z*#8g;9N`_`KSW+US#or|-xDMd)l;J-X=7EmLZ*>9826R(X0uXh2iwd~5^df3-PlYU
zE7408=;C|sR6(DIiDv|*R|2a?T1LsKFK4eDw!%A^f2TL|u51y9Qq92W(T?SW<3Z&9
z5jzTE)5wzXGAL1xg_f^@u(>DXxw#jP{Y#(5gwITAnuq9melM!8Va%FfLOINPiN58^
z6XUn2A&I^SQHpmy%w5e%s#uG#47EQe^SZm-B{USi5;-rf5+x%j$cfJ~{k>rp`aDDB
z$<H&|!u#qY4+h3qYUSx%VPWXsWT%5{Ecc6R`L%sLw=-`JZ!W45u0Ob>?^z)~3qJn1
zqVIX!EE*}*ug?e@C<NR9LTX=qNVGBFhjR!ZUcsFqbU?&<0Q{Hv5X*6=!m+rxTr8}0
zeELS@$-K#1_miTJ{f)X$I}6eKvsJSt;suyX8?&AEsEOE2c;WCvvypL>c1%DTeU8|{
z-<?|T9W@VGRk0~bHI@jKZhx84*2*bj1n+Ne-UIXo+rx$$s29l#4;@>E#$D@v{hf9G
zTh)jDjzRt}<APqX4sf+0kpaGS`(VrCV78;L*njZ_Bx{1aY|d+S+&aEfvri!)Co?%>
zQ&UIAf*1Mi7i&D8<XPD4gxp%a)BsG{zGeg|U|)y;o<4cXR!c)+4yrGOX>w~nt<e0r
zEy$Mpnbyt=Ap-H1Emvg>DelhoU!#VxT4|LBCScU}@81iFi5d3}QfC#vM$i*`JU(g!
z-pKz6lwDeM*>in)g679XgC%-RU!ZT&#8@G=jqE#dOQV4YYmAJOJyPqw0IQg9<(YOn
z)L8`qz)Q{1@i4<CI_j*EeI-{XOyqBcgCdE!d7I>FTXT1PF`MAbGnLcy@{<Vr)8Er(
zUM-Y#)&g{wB%Qxzn;WR3S?PW!Jc@~#spdpk1eZV}D>Sg+b@~;MU_9F^@#k0=rK@L)
zJ>v<5X-C-2bFplYZq$IzpYlXhgmV)EmB1hF3ipTqsW@pkJw*#U8TcfqA{&UCHOV)g
z`IEuy?(KYT*WqD>z4bB02}f<CEyeQ!ovvqNb}N~N6SuQ~uYflO4e^BE-K@D3`DZ>1
z_~6HI(s9DD+nXs772v`N;awmzIf!#u<B`(5@%vKGm&333%6Tb`gM-h4c{pZsbIj=;
zV`S*kXcl8+cxq(>c#Hq^XAmob#pVNu$I-dlf6U&SOS`i-^{&YKKNc#u7t(?t20?==
zL5$)wNMTsT`+sP?$+dgKk}}r{eWo6jmHba$b@kZl>M)h2oZQ@x%4JpsF@3+$fs3qq
z4?x<UU60-JM`RWH?fa{C9w9hmwyEXLYjY9)ii(QH)0Li|2?q#7790=R$T_AjNn$sY
zEvm2^YPAeET{_M#*zC#nK5f3z?S++{G}{K+cP+4=lRc@owOmVUlvNE(E@E%#t?bSj
zGxKc=-Q0Y6XA*pfwgq$y&&AmE15nw7qv{B}va!|t^h}eQHYg-InmK^uJrYHjN93d@
zt3){eaVY2ZTcDrc;MP8GyHs(aI^sZ!D1i7I1DfAZ9O$Y!;Cay#SN9K$$WGSWk;BL%
zca0bH!yk#IQ!M-daoBXlB;rU|$tn}wDBtmAzF$v;Sg}2Th*{&Wi<{;PxaI|?mzUPF
zwLlu8vcHqwV$*8v?<?=)ZYkRMQg!jCQMg~~yLa&eI_Mj4cu!6Ix%G!4n$r)`(6A%v
z;jbh@-nr`;En!}F=g@DZC6$ln9F%h!OP+rph!XH<mc7R9F8UAHWRx?di$p#GkYd>Q
z5>k{3%nZk1V+S=n#`DkK_R^>yGA^@>$*Lfaxt~=_dHrU9EBxZM+!G74nBb6>4z*Pt
z_I^q>-j_{F-&8`<Nr%pJvdXCOwio~tMJ;>9w}EQy>ie9RVN|zMOr=mdwS=tx$+7iD
z+RdW3xsJC-JR|n^N%p62z&UW`K9A68cVPfr25=|c`Yb0mq78dERRi&_^nFergSI`G
z!lk}1!0&jxCtWMN<D325z}Oz^WQSRa{ZH&mpa)Y^S1-Joss&{JV`@o2<`+DxMFpfF
z1$jszELTqTTqD)L_t3*b&w$T$g+e>It9q6~IIvVL#MGk_WjyR)v=K{cUpu$VZe^SW
zno)Zog4(vGavy>y-Js*n-5E;2A$bE4Q-<Jbul9ByG`DUwJ%wP;db9h`_-DUHQs>Tx
z2!KqoH+>d#H#A`8aG!Jlqi=B196zEBzLQ?^W3%#T2ZKihs|tz~s=6AIZp#%pzu<Iu
zV_ja_tiJn8QGm+otEFsD3KgA<%DZnh<M)V#K5qO8DR{{_r+<x5sw{64QjVneS&yyg
zbRw~ZCH`PDE#<gG=Fb#n-#qd&O1ThayZx=SuQQ5*Y-vnvEQQ@JAh`O6G;-n#6)#`R
zNdVl5Hqhx-VV`cffwmGk6R_j)!pb3Ark$Am0t6y<s-Xz*;Nm<IZ5)hg-Lv;L2Cqyg
zi>X4pm+uv*LGFG0^A!8u$Jk3wlAnsnwbK>SOnF-0m~F7IC$;bBPuto%I-`#i510x^
z-s0-jO6HBmPvt+lmMKqBT{9^=mPSUg>yzyU4eoZd7c%7`@k7p5b*#2-e{1Y2Vdv;0
zkyB32A31rqk1s^18yu6>9Gjy%bqZTU)_?=^nN?H!ex3#Kwa$=^0I@f%SuP|4>nw#|
zc0F~aIoVLZhLPPZHkj|B<!m*~_XYU%c_*|Pz00+y-?h!YnR(BrgD6yf{Ve1KRf^!f
zSIh^8&9UWNnZCYOOje(p36H=Plf_|!JH*->#^*6Z1lyI4kTQ>z=hcY150w1BqYIB5
zxlAKXk3IL-Ae9r_iIwl7<{QnQnVI)-Q#Et{%E!rnWh_z>XFR%iI3FR8yZq%P46AS~
zGc%JJr|OmJo=hdXQ|qO~1t7@;=NM;;3=BahwOL^ze~6>}W<^rl=rm!pgL!dW^-9$F
zwX9z?hLCr@aN_K|t~4todReJ^k6RY3>d1Gmy;}CVRop;dMJIO9uV7i@_)J*ux0>$#
z^<l6_9yz%H$lMHrpyQ8E-%n5gjUj#gmHgakd)vU~P=aVYf9NIKQy!I?SFdOc9WIQ1
z2#W!s{n3z6*dqg7dj`zwB?W+NydNM|&`yNU1Yp@y&m6r>5A#w+Zr~QI7`*kz#-{xq
zuj**6*0IlrBSX?uSs@Mb5j&f0&eO%?Z{z^0twHuh4hRE4O>E@#@d!CZEn~7<evhGO
zQIY$o@~A3*U7D#aX;P&!`FCPTN5g@ZrmlukMEqa>Y?_~9up5Q%)k)NFP_;<ZkX)r`
zz}2xNTCYX2gJU})CGMH!{X!#Cx&GWUuki9d^^Gq_@sB$j!c~h;tLJcylHu+x-zZ3<
zB5l8d07Nx5a1K>2QSC^=<-By_Q|Pj1uJEi2C$Q1hA8#lOO0rXp%IZ3-c>DnAbk8kt
zc65%$WS(J_Pwd9fIeb7p1i`y{YoGv3;e~WIh;~9p;X;*GL}{<49rbqA61^d2UQfd8
z9Ub1v{ieEi&qn}UJjHQy%(Be>39tTjE#_rJmzl8j^IaV_Ea%i&mU2%@kPDOkZJ2qt
z2*O}cZuHv42xj>|^tq>M(x?YqIVxobQowQq98+r^tNpheuA+RvBQ9Cowq#2tWH%78
zwq)-S68xc5i3hgN<<7V%aZT%QxE`aA!>9Etervav?p~|@n)2^BdwQgQem45PJhzfA
z%!u36liboFRhRe1cZA-5M(&9_S+?A^mkoIoAnsHRbL6Z<CLS**x@$$vY|ty}rk0<c
z(u5Br(vM~)42`szR%b;i>~8B=Zoz}NW@8olh|l#8>`qK;VCBXFHyuQG2~Hbk5l|Y~
z5$^!iR~(wHR|en%d;5aTKBD=<gy>`;S;6Lxkbkoh95(^nEV|7iOPDD!nK4I|oyAaT
zYAltL1!1mYkkf$roHg8OAV;zJ-NX)efj$5Oo&FnO^D?c2{Z0>lf5LT0YlgI)QW;i1
zAQO<c?$;D~>u!dMCM~IAG-yB0ZLM3i;ha3I&{Pwz^8I!LU=l=tq2EE%t~O7JSL+@u
zx(cuGiAWd7)J>#?37Hp6TnhYLO%kjJCvyJ?{D);mkdz4@?GgL=8!sCYha%}Sl<$4W
zWIYIBhqN96@dG(fn%`-H;(4*Cw6vB=2n*@E@78<fF!UgtEBV$06>+28B)>uorS*)y
zZ!Mv4m+Ey9g+$D<9^qY5Z$!7^l|)x*J`pl@P|+*L9v+ds={HDoKkX&EJm*Kr`vhf}
zi`m~~aX0h5lQrvi9}2XM+4M%7c%H{;EaGvfEmjrYytlU(zJX`!Ypjz}kMmfLBU{)h
z_f6_Ot0ye`8w;5?cLYi-{yYny|LsscOF#}`cd`^$@>O9R`r1VPfrY`jj%c99J$tQ}
zQ~2QbOaLDbRpS$lB`%y9WRZ$=2XH@yY4VkG9>ADYAzt{?Ty#3z0sXqh@WxQ0mVqH%
zwD)8=#udz3-qrcu&Y{UjoBlNC{*`T&;CPXA8Jj!!(OsbzFJfZ??;;h#HOm+~vCwVf
ziW(rgqMG~0@skE^_yG1sip;xSd{0Ne5BK{ATiOdVD)4S0!SxOF451hU<AU{K?Fw@T
zv+)OYz8jj-=6{f8XJ;w`%sndNyinD{^H2pa7PA(N@axo=-TI86a&5;188y`$Oc^z9
z01+ZeMH8!py=}SoAL*Wb|5<bW#2-nzF&0@IehVZ)-O#KO(4M&*Lg0eWzFsh>^rfm0
z^Qf+(a~2B!{R<O(t@8cI`+P2>3C_d5`r4U$r?xZT5C?LlvmmrdX4H-ZY-;T~Cv%5a
zx#78a9wiD`y5OtLK#Krs^e!z(1Dt2iVYCjvKRbL!asO}6e0$Bw{3O-?Hg0lk&U17J
z)J2+(N7ZewI%SXi!1?f1r?i+Im9i{bd_7>yM*9`1`P@K-@y47&!}sc<lQ?*=FFADZ
zTyRm6QWvX{#b5}Bb_!j2RWvhQC`)U;deZ%itQn%$8zrHI!)OM2brmCK0*o<brEHW{
z8m`S|>IXpi#a;h|B*lll#Uu-Q^~|`1!>#g81e+sOZB7I^W;duoG0quJzW^VX8HeWF
ze5^!J$@6(uc8m0cid6k8_V%)tp&S-gyc_$x9_51XkFK9f;FcQxv`lR0DHq5C5K=ug
z%0yE=WtaD4HF&ihfwaN_wsgsX?-OXm*)2^asQquZ<5#kAS5n*D{NZfvc(V;pYg*DL
z+{*=skwi2PZ6o`OZkSahdkUeKo-pF+&5@`PxKFS%3Q)8Qb=wtZnOAhG$nLbF%@-r%
zlGM_Mf7`fZw4RRsp+9DOSgSI=<WYaGpz_s_l%I5UMpVg*N>=>VxtN|pOi_z_snI;n
z%d3}pu(VG7BE7aP<5}@T1wYZnh3@}0(gBb~&hhgAC?laCX9=TQ@5ydHHE!`0`}Diu
z4k3?&oOXn%M|s7Ty8eW220#x*gMddD%1?u|#CaaFa)=+iRkY^akRtNzxGGXNT!SBc
zQ!W0O2uk<P7WE;{2;AaVFvBl8YDZGmHKGUh@$E2c;H1pQJs&@eRDV*y(b063!O<}c
zx~Z__x#+2o2~mhN2~VpW*Sr?MM`m?sb#;)Pu+nbE(vf+|*OlUK-dRGMrCho3BNwqA
zA`bZFw{rZAhh@*r6CNva>^ZW2^t#MS_NqBf1}AtrfKci9)Q)$T0BItmO{?c{Yld09
z8hvkdsE2G=fD}=wF0`1i*yt-cTsvwi(R#fC(Ko06+^?C(2rnAWVfSOSjcA(rK&%~y
z7AXQ6KP;EpG4t~I|FYPl<K^z6kj6ZH*?mDxXSw|b&FFvdi?VTfV)i^t=fCh-N7(zw
zgkATtASU}V_2}&E<lu4^1?39vCuACd-V(0_T7Z^&X^yb=xE2KmFV(qowq7gWNXp9N
zT2;mCc(yk{fyZIDt3#_)ERL&W;<|^TMuMPnS_oW;{@W{_`EEpk(p(>NeHpm+>x=zq
zl=|}#Vb)dEnL(O47ti-M{uQ3X*1_rL_8(yO>h4a5|33kIL99|1&8XBBLiQiGS-V5X
zY4#>s**ZXvs%`9Xb$2(@ONOqsHNAEhuV#(A-i@-hbqq742jk4|SWoGya`!*|*Vl5W
zTuJ*HEP66mq~tZ{Ep-{c1lq_#)5Cr!N4iv%0%L$$t-l3=;mfkHZdWwW?8e5#ofm4)
z5cimbkXg_}W;GRop9A<*jx`vc9(=m>Q9Cri@m5*I>URxu0Zy*g68b|lq|0dc&bjlm
zGaGb+x}vhMw=w&*{Yl#7!|=qNo$A@oYwtvu4Oa2lcprxT_}@_4eu~VoV)lNG+NJ89
zlfTs@7W~4}wS;KZ#c5n2UoG%oZxw;v=qqw1jmz^I2%|B!CHO>Vo;^z+_EqkG{!+8~
z@!4Lsv*G7aX<6k&F!Z!DOh5x0aXSG(QCC6}uREc}enru+m7Pbflwhj9U{R>tniDcc
zFjV0`Mu6NI3L+pYBCa?BT!;C_?BLN?AODbMl!XZ((`NgaXDuHh8R?SxY;w;%!q;GY
z&aZ(Nkk~?7baA(sY|I){u#_w(?RZl8`TKEgebnsBgcHXM8^MCtas6!gJUdel>b$wj
zTuJr`Mk5@{jjhl(b;EA_^9;UqfyWk{c#diIdBJw>T|3!lkE9eZTxZbr^!5A7tc@KH
z#;kpQf7a4ko)0(nxVS}81T(!=1ify%_ZPE{OmP`?^kexHiHUbM2L}g2JS7(Kop=dj
z(ALt>{Afoq#0s4>W@QU^?5IvE-w9JpUfbG^v$uVI9}tc-50uRqP<Iq=8&Gdx+`d3?
z-$0xw)?IG&=i6T^k3N@lts*_{$N{nsl;ZQ<jvhXwblU^wV&H=O3cRZ<76LU)e%Jm{
zz-0|%=Xo+DG9+J<0cd2Fk5Gq(hK}2eoq8IF8Qa<T=jP_pPFfJnj8~f+Shb3>=%TZj
zxP@%9)0Q7w@yv0~tZAMp#Bp|B%M4epON={5j{1Q5SU^d;zv|q$^Y^$fc5*D|Jb@v}
zeO+CcCju30Vq+fn^_3iS&fl2Vm0Z@ox)y+d@Hqh1H0kkR@2oG&It0^rG=Y9Sa`l>?
z=Gp?MF_^ubDDG09C=dL>WRMa)J4Mphzm0pH=R^39#osXvhCjwOHJzLKidC{JNld|m
z+cX$5%pVlb_;T(pr)RUg+OH=t?j<lT?_vl1U&N-2CIN1bo9)WhR?Q+b%Ywgxzm5R?
z2kIu7ZjT-uFC0&=@Yg@Nh^AG4>tPFDakdos;Eg#yi+7*<z&We)rgP|&gPgH(Wf{%c
zYkLvYQTpKZPfK*khmU}S=2mCwjQZIuCC6^vEjK=1hjI_i&&5t=R`|5u-^UUnPHXC)
zWaaF2dwYJ7R}aT79I<paoKIk&$aFXF-juZZ8Uu@iy$0bgRt0(%>C_~d0hsTTu_hP9
z3Sa)2sylB3R!VBBM3K>$l(ckT3d2U_NmjidWMa!F;U=Ppar_24V1qI5MLSp7C??t%
zH7~hkaUhYYH%H<}4j(puKkf;Jcd?0n1Md{!07)2NJ^1;F7L|?<-~VN`vq-%*y{$}+
zN3237kf!%NrERNVhvNfn&G%8$TaW#cZLSbWHrfQ8Y8I^@&pNWTvlXvBO2_0eM{?UP
zuMAm%RA-po%-4c~?ah<}dp`|CJl*sqa(|x`vjM_pu?sOx$4GX=!ejO;=X-t9AfSbR
za`$#GcDYCd+n=uYrB-WcHJ;#>mX?)Wx~d`sE0uG(xZ_Vq_De|i6{P3YR*Ghav#I#u
zBEN<ZfbZWP-Ru@x<I-st4Zl>Vi}}(O$7`|G)zfHbm_%PbozjYg9`R^Rzaou(a@5)=
zcp1FR`HDF^VUm|7B<ypgPV3rs&)%BA$jg+T+9wqdHm;1#UN6&+A;=FSzmnpB%VK(N
zvJ<JAVj)|=&cmJx`<Mj2^$2}5ig!#V#kgWmhLm%;Q1t9%;(UF7UmXEZUHhN%GDIX?
zDb!qfKSFqZ6Ou)%+H=g)j_|RvdUTwSqZ~4wd7}Zbo^N14n|T8Xx4UJq2cOz2NEA9R
zVqfXL(x!h{YA39H5R>Km_)9%|AXIyK`yhB(R=Gk<PJNL0PFp5n^NChMRq|4h<`x*R
z<N!H9aQ9DYP|*lrRvquLN}Z1G{KJkFG?co@I{Yh5ihgMsvo_pu<St+`_bzvgnEBd!
z_;E(qC!0*qLB1&1QeVEPe{CQ$5gG?Qo`Wy?Snn?$g;)U`3z;uh6$@$Q(|wFC8eHWE
zg}-oicJ4bfJZx@n4{p=q9Fp8(M{xaM^3J6nQ&j&&&RIFOSV_DfSTFkeg)CbRh*j%{
zoJKe4H^N6=S2gV7p3c0%1d>Sz$88rJndh8x`bPj3S@arus`EdC=n6B(CRbU2>rnd6
zIoDe>Bj18uzuVXTLX0_}Ra*iWfK!4Qda8y;7#y7B`u%$_bLD^DcaCZEnA_u%lN&9=
zKh*z24{7MWgzQP*w^nxf6GI0Q+iriHRzAznqqDMnh8h|gD)jE1&gZQm?dVL$ll{<+
zHQNi3%Z0;{K?!MfE2*n>WF0q*g-qdGf9eg?WIi<`t3Pn-HqjZ3*a}DK&2I(T@z~@P
z`ht3C>c#`EuBEeUtWLIiRKi{_-NPwIYwrL>60p5o2o|ek%@<NFOq2W<F7@ZXx}U#i
zodahvUFgqUomFriyO}<TtO*^3+t-ZqY3H}9@|&E2NBzVAX=@a`$*vKDjw1Y^_;@*e
zuS*cZoU7R!NEyayokeC)dmFK8nS3#lxHU*;8`O|Yp^`xD5z(?eNaNS`>WC)54{<q&
z(w#33^m~rDJI*J){;ISknhBo-<*b=~u6T2WNjui^8FXed8t2%LpclOpm=}34QOgRv
z(*mC;1#TB217EkhO0N~HFBjEureQxSs8`>>^7GJBlIkG0M%6L8X#lTS3i>mJ8j@ty
zTSdQl?=<d+hzv+70lEH)!$D~)sey^(k}Ephhu_i14npS$<h8XevD;rwjdnUtmEi2p
zQcmxR7qr#^3KZ~!qR2k$O#LzS2$4LmV5e5IP_wRx5ND;1NlhKvSL`31VIOC`oM^yr
zKtU8%#C}Y7jwrOq<2g0Uet3-R`?AfBiWRgMtzS1Qxch8c0)KCNqSfNYks#dPc@}2g
z1lw_ncV2`XYd<*~0e(r#ac8Q*L+s!(c)nZ`NFDhS$+^Rr^on8dV5?YaX1_vU6|S}n
z5Vj08_uAcp9RQ@le=IPPdG$H0sp0A5FuID}64{t}2e%LYlLrrCxb0iRKq?n+Q+u?S
zUuR-F0;aNqQ+Yzl&02JP2s?>jS0xnqbR81cJyl3I(T@T4%1ES_U7c3~AHbZO)3vn7
zkB^OqSMrVt`idJGW^!?HNu?{;HpY){ZE4S3djE(*a|}8b6y%u<JNg?--@o#sY~5;>
z{b5PTGBNht@g0SGiocP3+~Fi~fv&gl$0<>O7Qvi#gXJg><*x&RS=c7SG5cvv1d}iz
z!cm;^2tY23&Z)z==Z1YygF~?r@-y2!bBFD=DX+WPEh{#BRcCjuqEn>&Nn+7i6j3(S
zXZLGSFE;MGm>!&(y7Hgd5r!a)zyXZ4cy=s$3ygY#7Vv;Q+h2qSwov1MY*qHAV4zho
z^}3beWcQVJ^IxUi1oA%(FWV(6To8D6ftvP;W2*a+j3aya?|)ot-@Ny3J8U)BTj)fV
z6i8j-g8WIE^_`&KgCfmMPy?pJw29sB)g_Yk9vg+b|2>S8Pff(8wHzz|jI#Mg4=O(C
z@Z<xb20z`b0Yg*7qIuHN()2!y0nX>^*N68tz7zO71h8EC!^OZllMCb8rUGsH<UlzK
zh1;Q3<Aly{-<q8z`9%%drC;!XI{_L4Jq>l3ocUtBI|+r)s7YPw+L^)mrNiI%k-D~g
zNacap(Jn+yb@;8opA?OQO$Fa6SA<sSCc*DZjF{^x<#*a#qlnho;qM`|6_t|kFTDjS
zQ{d>=TWeA%R5cAK;%GnUsKZTnh8la+;9-7d+hs%XaKpp=#P)6KDKU7<rcs#~d`7*%
zDH5){h$Q<PGA#t`579Zoctu2fWHRJpHCuw8oCgI3=`ME04wny=;C~V<@?SwvdlPKb
z$Xq-+B==wcYxoc?h>0#?yoOQ(n0g@sr%xg-x7`yPfM|$u#6@0nj+b!bXuSp$bD81Z
z#lJ#NptaK5V=TH27dk+go4#w&SlSNO`R7v@w@<!mpax^6$N%7&W3f;<<|z*8wL#s{
z3_i(ZDkic-k!?y@c3q?l5N5t2;<%S=f2%~OM+bx`B8WZrz@>`vLGgNjN5RgD-VA}{
zWeP5N(~##b$^mJP)0v&Lj0_bORr49X*bfB8$!BmCmBk3TqWVtC<-6IG65D456dn7v
zg<?KnW*#IOTjafGIq5tQxI8rXR={a)V=b{ebjF*jB`q5{U)H*_%_$k;ydBF`)}Q&M
z!2ne&=w-SGVk8U&fcGZGM2B{!9$u2&ty~Qgs)gMjZI6!RnKLDk<7GJ;A5_I059jfY
zj$eu{j;nazH9c<>XcV}3^P)G$&n=nIZj~1d4HRcw(xn_9ohqXc!2EJz5P`udI#a#=
zZSGRUrO%{_JV4W=SG5cHerTVI?@eeNw~Z?G&B<G4=ilKE#M2y;4`If}ZoY#dC1P8m
zZoKIx$vhd1Za%<yvhIB9ev!Vt>@dZ#n5PyALex&?iF!_le;ds&Is6}Y`M4G5%>P=r
z&T8)oh<eK7#{U!vAF#Rd_W5g+ZhI6Q$97+**u}5LS7NZR7NzP;E~dx^BCuZ;Ys*$x
zPt5JfZ0fY%TQ%3Qf$5@dGA!UB0xXi~K+8b`jNr@(Xac<KUG{wVh>rXIzVu&G7UAT6
z_ok*{|1iD5LCemk%G=`~Wxi^>Eyu}9ZB%lcU$(9<IX>ie@|!5zR$x|5<Vo%xz<Kij
zlDskTC*u>J#N*?d4%;8idlA@9D5(?ih~T;csqIT#1SF9HJ}*KPDx7WB>xdKr3y@sC
z9Zz`S&oyPNqQ_cfsxv{7cE0Cp2Q?OZHy1dF$A%_!<Vvfzz9-Ubx(J?i69Z!^fX6^Y
z3p_>^^)@fK<r1kE9G7o@y@;mYUdc+5Kt|Ima1*)G&9h3n9g+;4hoZO44ITk5@C64X
z0Thf$oU<waNY$0Rbmx`MjLWgHReC>~(j^KnYbomIdOq=c9X!No)EagSO|I`TKbV<3
z-+WA8X14XJZVw3SHa#b>s7ggs-QWR!;=*wqIWM-d&2ib$fCtAy3pFwckM4@3SA+ID
z^^1I2HtfyW^Tw!cU`rABJcr;dH>ZoZmIk6@r#ahfLe9JF75DL3M%q_WS5KbNM#>rs
zy%U$$N>T^v9s~Z~ay8b5HUIen9?#wm+2yx!#CTCu=y1V`#13yJ%FkV=!nqAt+_kk8
z0ej;2$C%c{6Clmvy*p4{g8P4^SsGWLIc{d3UE9cMsmP(li&C-p__{HJg1txT7pGCi
zLfQVJQ;~}^xFVeeVEiH$(*eFmLRMbbf6dAxUgIyTK)5f7xvzkqO8L7nI%3Py-vuP{
z1i6e^_jSRb%Na3leBk<vwYdo4r}r|;=D?V;zKw0t2PrQOg!)Be3vBh9u^ndw8d!1K
zAW8l;r?V2qZ!;A7%rL+WG<v75HoTAsG!U||?xR@BcDn6_Q|)@5umYpQ7s>N;$wrK$
z7grTfhOCp)YwBu5jJw=Q2a*VfKNH;IH_V)s*dTd%;bKvfU+`&Ou0dJxx8P1U54%mz
zbLC`aPB0MAIBl$Keut;t=P;C`k0j^k&g%uRK6gs77Gm5}<{30hlPd5xv<uZ@R1#EK
zyoL$<I-022?QebqIo=q;3ziIj!UF@G2XnTHWOuuJ82SGd%Wj~gKPuK}*2+{?H3H^<
zFGk1^3d)r)ZQ_MaqlyuNoFL0NU2*+`KNzpi{o%g05$p5D6N>g#kB;x?z_3I-yww>1
zo+)BCfJ%@79OUCrBJ1ola-9tmW4Lb_?T1v-1>NXRX8DgW2oD)=Jvgt0S<b+=DxGGb
zh?gkpBJ_B}>G*sN1X+fDmn1SGyv&BJ2c$%P63#=0%#9s)>veAK?-X{R=<OT4VeMLT
z>Jr)bOu>8l2v%#n6G;Y`+QsG$4x1ZiLQWZ-Q<HCPr|uHZI(iL}K(5Btfe{Jd?r`^|
zeD5Ae;~#9aH@ey_-qo^K=nK7ao^H7AlbFnvi%FZsG#3wE_Rj=Ix=}t^dBhm^YjYFY
z&Gr#XxLZ?wl0+M~l6n9aOH&Pky8eCcLL3JKqE=R>n$xo+q?<6~W07-ZAWfw;^P!6r
zIrNj#>VEhM)lns%R!sk$;Iq1`JklK?O+^Cu3Z*&3D!`$(R)I1Ev=JN^B-<$Z>#y`H
zBs__OPNV!k+otX>5oPDKrNjHO#@12|Pb(aAqG0V4r6y)PwSNJc0Kh!NaCCbXT6X4m
zDg~6~Ud?Gmg%pQ*guk}@UI8Pw^3C_}J#2x=F#HqUEG|cfL6D#0=BPP@Yn|Z0$MHj#
zF|gZGlF<^gX+Lg7B)1`(DCxwXWopMMBeJ$}BVVkjrPR^HsqL-(&t=wyWw>Lcm#+)T
zww{RgV|*|^Ihy|89x5ZE@!z+O^B~87L8-0_|7N>F*7h<!TFWjqzoq?~F)Nus2J!mN
zp%%Ox2w*pM@d4MO6N)KyRwNLTIG0A(k!Jt@o)Jst&ZrtO*5_+(`!A^BqW_RhMmiBY
zL@=FhBGukKF;Q3TQ=C3?F1DhigzX2{heuS|Y+4a}Q<QF&iUz9ncb-sW7VTog0*)d_
zb0?Zi+z4j87b!S8D>=k169l@j2GXWMjM>8B4-;z3jtf4`H>O^Jp&V3{7MvO~`jNC=
zC36|x&Y{ZV0kn|3G}bbnoRzoSeV>BtWlFSz?6r9jM@Vd$m~CzLQoL0tZXOk&zMTD%
zCxICCdmx4H@eO<?xe%vqe*^(hRHO#mrpOc!Ud_DyguB2yX?I-XlA|P+s$m5G&xVS*
z+a+kM)Y|?q5cT8Tca%>r$u!_ad2npDmJ=P@&l_FQVw%W?PlQ)hjh0lc=~FQ>fAw_x
zbgyP_BzB>AZt^);(RQ%4#>(sdi^7<}uhRrxeo7>e7u+^)3KwMAw$m?Xx?F-aV~jR4
zkyKd0u4|`bDPlrJgP4_dst>L^Zfy67%V85T4Lak60NE`Z%vfig1NhXtB~eElbN6cv
zyp+$QKbIWm(o${+i4G@9)31*ltDSK^5e!qD@&M9ho4Td^4gbWzoHje5`C8{W<^vMS
zrZpT_GI+Y?`)3nZxTmTcokkub1R3>ar85%;S2z*mq0+3?T|)@oaIC|fk??$r9hDi5
zOS;?00@D&yQImZArTG=354nY(cjsbjo$tJO3m+}0fY1kGI*9fa7+V5>QhY3C44526
zw!4AWS+slHI^@aPvU(mR7iSr1PZkTz3OUoIMDZyQh66K^2v-fbclD?vKr)YOLc`sH
z)tT!ys(<#wMBox|17l5mRc6<1w4u#*#@EE>&Z^P(aCd+&U5y2h-p-)qP6$^q%}Vk|
z65NL(x*__+ms48}1p?Z%5=9L0Z8IJxRICOMH=>I-UO9+9N6Z!O9n4+CmC*q{8XLJg
zwK;65fJ?BjS@~^L$<ohDFN83il<pUj-748N{itEGo0X?H4u9ypZ&pXX&5aOdO}Sov
z^Amu9x)3pbUI>OdFJy!tGLU($VtQa;Hfvdeu%C;a1rvTNVWAodscno`4CiY807fdC
zSt>9o5anCV5;d6G<A*JYD*Ff~-htutTa+;Sv$ts@THQRLfkC668a9N{0X`Qo6ClE(
z#<R-@?y8CIsTq_Vc;Vc@M+)=0Jd}kdk~AARl3mY(MX^|2vJ_m1dli#-u*ikpUstyx
z4%e%hL^{#=TUQ;`v|73%5FXMwZ@r~&WX%b$I7+-9uDJcvxuN45!St=mG(=30<x_an
zxy~N$f7W!jEls_p3Tg+}jyJK>p?aj_{!r<Z2Z6;-8>@aRp4H!#a$?sxfZz(>kxRp)
zyEu9!+d8N5WJm7)a#um3!6EV0)A(R<7d_bUNFZRs0zjmfoHlqamz7$WU?w6xQ7I~%
zbc4^e{`1`hyeN4tUHWI1@B%-}or^i%i>)}H_wpil$@viibKGz|)82k8NQ`^gaf91d
zkWp;R6Eir2g6moN+>|=qXQ_jMsL3He0{EPxc_P8#Z}=DIh(=dne<Rb_u@%P+2O%2+
zflKbb<yjX)-Q~354WHx|$W6N5;_B|qw^8uoy;O6-p^=L*ovF7pVv_>{q;c}`ZN&W|
z$!#m_q?<3yNi)S<ss$x+ybx3mGd17|{sWst0VBW|_yg)igY19%dUq6i7#<m_Dst&_
zi<0zN!2Ggq-;>mgWiMti>4spRF6NULcBN^{))=m_dB~m_t^>zGpg5b{CBHzlM8x3e
zA*cAXwGkgH<Y6o8AsLace_R~4a`tV@2Rhy@){}(6;cmS1Pg)B_mSi)01v;=Fsined
z`O!J?17wv-B%bJjjx;K#)oSPDUy+jokNX5n4u_jE6~q*Xbhnx&YA~gTa0j(SciFr_
zM_#Fou1Ky(rM3-GbeIOreV@{SFdThSU394y0*tn!a3jW-kKK}vYeHn#dNcf5t6%;z
z;??G3os^vZl(H$OV}%C#HdBR6Y=UtGyP<F!j=9$#rg8pwUrhb{@$^09W~8-@v6y;|
z-DX$)PdZz3r}V<!9-%qi;p~x29|sSFDswkgws!HdRiQwiVj}YqqXHnD%1S}8gphyb
z?w0vKc>N`JT7NU}i*c#LjndEYnRI>%3AM6KUZ(db1gjp^nfy1{g^>;^<M`n0P1v}I
z8L#NQzD0v_d36`?UJ8EjyPhjNbHnfdlf7pBJdbe`zLgcwl5+I;`Z+LBM{S^@4q_#L
zZFwX*MomL2f`)~qTwIqtkpGXR{f6<8n}w{y4NY($xA^`fizXQ3^UI%I)vg#8n34Po
zG$9*OGG__j^kvfjY2=YAxs636lhx?Q`)ldBxw!(g4yRiUdxIK&B_=Jp_=rmwAXWSR
zsrbU#)|cUrwl2)xhBE~xL?=v`y@$Fz#rPuLSEu#@XtDXpm|R<?{whi(HP?Rjie@e_
z&4cXh?Sp&eI~MXPR`%26<#&uveT;~aOV?%*bJ#>tuv2$aAZmsHKkd)Wo0-4gmW-$V
zx*X@4XPe@MXu{0@tT6w=xdQM&%Q$`JO0lrs`K7K5IO6FdQGHamu?b9SQM-0P&y|2A
zy;1IwWqNgACMGiPBO2v0a;<ZF);X8y7L%(CHkgi%%B7v50ZB_a>Z_zBBl~62QZW4h
zOa@A`(6x&qjlN8Zr6`Kjgk>Wey?5E7G5zA`of#hud#IXofCeLcO8pVyP!{A4m)!vS
zt`5eE(N~N<%%};Wk%%6Y(<it5h2+RJmAFxN`wq{S+7jf++g|AgW9MCkc2I!S1(7PG
zYYuAw;9(2*B~}krF7+5^_)avl_4=s5Qvq+hSYYT9?e>LwEb5QAKeEpT;`7Jc9z!9R
z>#;@ax1n#*9SAS}{{?7^;c-2g7*23Dcom{iba|qR0)A)GSM(rFEqoR$*zO}Ar=9!%
zhL9|XQNt;qF%{>B$?7KHaFMB@p<ossq>61&{y85uZnG4}qG;+DhN^dWG$Y9c`<=Y&
zWA4vnpAf1PHIa>)LM;5K>pDXw&*4{B+%6JRSSSOWq0nzV;W%*njh<fGuV247Sf6Gf
z`d}w2su6(xk#6?yie0~lcOion<8pzXdMlaAQzScsN8z3aQ!`2}D3@BOKm0Ebxxo2g
z=t%>>F?FX~(ut0wukvNxAO1~Zw?Zc*4!+$!(BaDbFJA`bpX^-jQ#dJ`;N*FT%$I3i
z=gSV-scRlHbWi;6$VlHx%iakm%N0@H=13nmUj1ceZmyX(gde-Px!DQ)B#p)Xkha^M
zt0hvzy;iIPoS`(l^rLL8?0Iu+-+&Nkd{*gScy7n(q@&x7t0O7ah^n{79xVc=QgGOw
zAD+OnyPAKwc|d9}^ViA0RZ0gM25M^;)kJ-}7I|PWQQY?E1(jQ!3RmFiX(xuO<-kM?
zKXu#tgv()VbzPaOX1`UMZm&BYA^z`*BZlma4riS6^^{8h2lP+Tt0-{Zg{<LL6=^-5
zDyw3f_|T_?9;QSsli~8{qFCj-xQyV`Xd1qd9R0KaA50I3V>&cY)wT?&X#N4yD!?V2
z{3i(qi$Oa-6$@Q&H@oicwC@<kW~lA+a`PGJ+Pel@c4X7^#eXhQ>=%4pGp+ek0>=4w
z8n<W>3_Du*Xjpned_Aqb?w?Q)6NJpAHL%K7vI_8_l@HZ*59$9&r!Ff+=n<D}z(sMf
zs@v~iCCc+Ip{=rnO7RzvZP8QyS!%>g88b8Ua(`wxpv>blA0~#VbSJzI7_E7ra0lq*
zRdu(E4H=Y>cJ*ZW`8Hpet>z16)!uT69oda1%*^z?<>Ds-YXra_ib%s$`c?Q^)lqw1
zRx%z$t$F$YS6ErGh@HB+T1s9{C&FUV@qp^^PpA>Y3%P75zy0l(Zk|uLT9I<D&PNbW
z#qKL<#{%G>yjvoXX`O+A{UQlUs|M<Y1yrL>wK{zj&Wp*R-`&DCG3?`43V-TQ#ZNm1
zTx^_pQ^Jd51Q03bL6x%^O{@Xk5lqJH@Jps`^`|vpo;SwwPB}1fj}LWnh!6NL+0=pa
zH&Zww4=A;f?k<WrHHpa6(iY^E%k!6}CK?YgI@--Ta<Hz8N<N1xuIzfJH@H<k1U{}p
z70E6N$lMBN99d@7Na;Fk#<E}DYrL2*s#4iDLR9{~CUamP(pD_4zznP`=5=S(4zu-z
z!3+`;x3_o9gP1b<M*cId&&u<sRHv4gXO^TGPi7_TEzAu>$=y9ut^SQwO};vt`ch}l
z#8}t5a(bm=bDbORp@(6WK6VW^Dt&KUd88H3qU^ZWcRc9^R!Ylcx&_>a3+F?VvQNeT
z<<xF13=Mo;iV_BuVwwLR!mb0H>h}G=i8rg1tuivQLMYi(Mv(|5S!GtTN48GAl}#ck
zE2|-7WMr0IGFpi2y;t`5-_PN2-s<=MFIU%ft{mU*_j#Use(ukGf4^_mb98-VdjcyC
z4G&y%_x$odTC?|Mm|kfwxSsLu0B?2rr$f9d>s(Y==9eY4=lLCMnp4ZO3Q<&RrD^_9
z*DO;byb)=iYm}H?=2?wgchi~b(z3laRg-q~FQ&P>X(Lh))WMjFj)_tVse>eSGLBBG
z=e4!tc4@t=pOtyI(OPv_L$QWvOBmaDSD?&p-Xd(BzPkF$$0KdSwiby;QO4AA`$jX(
zTHfA5dCSULz@UdyUd%I(Ir?o^<qt6t)9K=_IV%PTCD9zM#q$E#?lsQ-x78G6hgUU(
zasAq(!5^}oMXqDvZrMZgeM6I9lP=gZ6{6O?wW6l3zihoxpn2aCouBDE>Z8KsPQv?f
zu++7;{3ZEQ+vYESO|Snr*7>bv)O5-9b6(V2B7T*T2R^tCeuyvb?U0PR5KE`uY!NVZ
zy`U4v1yKEa{uzsz-&u|G!#8i-DjrHLN|~FRQ;@3>?II5i3riXmSQ|ujnPE#d&mG=Y
zgXckFVmp-S%Khsc=ptR+7D*RnnlskA-#sHcH~RhdEg#ib?wHwnc6Vg(kdyiPFg26#
zggtIaFLKU1@7ovfV2A6xNu=wKv6h&gL%TL~VM!A6BAC)${QG>J^;|H=Ak`jo?Q6I?
zciRW<OA9~F76dR)Ee5v1jCCW)kco@sw-_i)B_FyK>rG;UKV9|DHBDZxZfEy7JWw%%
z63m<4N9NNVcbP?4ZMyuA4Ei5wy|EK(b|d~q3sYX}u_Ct<pcs5|eU-jmG|Y-P?o!H!
zwIMJ?L<q^0-hzNk8OH#i4bgtRds2JP&u`^ICE{!S+uc{<9{{%i*T4O}Yx2XP-X948
z&$n6}N(-H0j($U6K)dtemMye9u?8x}!l7N}{^+;X>{b}`kBtgaq9Y;|$fy#jx$Fj{
zyU4@*CmKF!-Hj}7n&a7sXK%<V?#fwwa;#vvFub9O723Cb^HbtRi5e^n2&gzuf5{sw
zZ0?W!@PP00>jwSiLjR7dNL&vn6kGUoOlC1_PP_Hdo*Ps3sY|%!`YbD}^<w(n+=K$f
zlA@WP^=o4?Fwzmvw)V$q#~(v;eNHak4VBn2m`!$0;#2e5>@^DCguK&s>t*jQ=JNd2
zs5b_-B)kLZ!3TL&GJ<WTei-%qbzNJmIR8qEz$hKBOcT*?-HXr*i_H@|q<q(qbr>J9
z4y)gFALHAdsLVfeeQ(2yb4))Mj-cAA=+9v2;PBek`L^{Ebbn%a``^B({opXdyb&vy
z>6<nELtDaP9U=R}@tIZ8a@JthDZO0V1UEM~`?>K;&d$zn*W<cdH=3z8f(8w%y-YD-
zMt2)@jNY9tmacecbpJ2O>bq?CcCb#<(`n@9*M-dPn2B?1V7txkD+zd1xDmUO?Qi;f
zh-wTQ@07jW8-4x&Ifc^~0j3HFU6(K`CDP%9Qq~Mo=Uqn#8JDIt*lJ1EfcFN~n$vN&
z*MkSqlI=w!$wakxqJBH$ht|#e-QQ%jKwSr0^>)zI$l)gxkC4^Y4$r#Q6q}Ut=0!>K
zv8^a3&5_uW&2{GkcEHOJ<I4sa`sA1>y&sWMBRE9+F9Jq|_Q=)YuZ*cy*1AP&40ef^
zQWowV>yub>s(Xft)EZ)9O!+@ZHx1waDBbk2f*sd4Pgr+1;L}~JR%gY>qdA{#jIb<r
z(VXPZNjoGw*Qstc%L39&ac{Tg`?G%Soq3u)#{PCuWz2wr*{;k9W+quz^QHDtY><^K
z8f2Eb;+EeCKQ9)O4#&1-Gd?i0y@WG*FOL`V`E;z$SG~0ml6LE7>ERmfE!TYBC30Q!
z>7-^~XC+8%fw^dN98x)=99m_IAD6DKwR0Ipl4p}BpZn+b^YmABPG)r>+o*+Hv+(J?
zFgi(9fGPQXr=f6dIChRIMCH|4=WaI#(u+u2E6RNIICvsmD`~vp<ps>Y=G1Jrzc_t(
zNK@Ca*I~VU{h-EZqXiE$!`41oX1&qU1!d^u&mM8Ywd2=hj_Iz2vFOy!`xP(MI`-e2
z1*h*Ye|Bw*n~5NAWbg1Adu@V3e`UpI#nxAHxdz^<cifJGY|rT8w{bfdv07+7E;jK*
z7|6U|+7v?oV!1jb<tNWvclX=ZyPuq=&Wc(*VyjKV&Pc5VUg>%TyV?k%01i%Fs7&9f
zZq%~BI@%;N_IP1ybfP&MShx)`+;_o+yIDiby+v7ypQEz+9MFW+E8PzH*aQeUC0X;G
z)&uzN{!X&0o`l?efjy<_0!5i$$i^g$ZYA9FqJu+vMI{T(uG7U{cbGjS6YDr13@yf)
zdW4Ex@}i7;DKV6u7<)_%gL%O%aAbJ9`yluB2scmj?)r$W8y>-)3u7d+6cbT*@T=Qu
zL4Ckl3)gU+$m^am-i^tq)i&=GZ`!X~xDSmqn-sa&E$U4cp8QK%W$dmG;FIff&wjob
zVqE4}KCUxOA)~n*|LU2WyQky^FmJ*GE!A)`$nW+W6Y9R9=XY*8m(crwKASy0EcF}c
z6-*3p{-RPZ_>|I)d<|`?!_Kyu#AsA*ODr;Ymdjg>o64pwzRb5alC4CcE$E`&N}0=7
zWZ;+xpmMnMq;-t96P3p!iv*CFE9-hPSI_WW@Onyk-#@=Q`S!Qc7gp>0yNd2S_ONf$
zJ-KytM{1%HBapiYiY^yY2dtyWuN^LyXK*}Poc)2Di+cX>?J|CxR*8GvfN}$lsrEpA
zg3Y01TzYA+U29vB_=QESZGFeJ>S*kV82b;2zq0TQ<##C$Uni;&+N<sUJ`<L_)*an-
ztb~s5Q|=Mnwi&u!(oGE6pG4E5Igy5LyIrpsnuI#MTH6|VL+@;;?^e$(_gMbEx9?Ql
zsp_3ucdb;$R8l*I9#tm~KXjLslTwlT!?6%eN}9vlpVHryGYZ<na$lcjU(nI*dB<+w
zi1oU``b8dwde_SJ><#VMQ#0ON`*~{KVMc$*-P1gJ*`mgI_JSFmO)5vjQW9^rLV?JX
z^m4>_z(Re2Msh8oF^{zd<}vBg=Oh2j-opNra1x|!OW78FU23ka$jV=VQKleukLzrj
zjbnpVNQ-}B6v<8&*&BahBB(1H4RfU9Xv_L<Cy~`P?>)<wIAcB$y~TQJv{!1%vUR!Z
zyG)7<t+l%?`jM(09_ny5qASahHz&P$Z8mbZ;8f=WG3Q&0C)Iryr@aryDbg)nC{P)?
zEaGtWbRVx60U<Hmh{cnTHu}29OU5IkU5gQpBeC<<;bM_)(JQH?{W8hCm1z#XF9d(q
ztCDT86St|_g`R8+0io@w_9foc-Cg_L-3%I1dfp07QH&1RUh_&gj{UXJF%~x8j^6y4
z)0vDD!~8J}bg%vH51w1C?xc1W#`yLyk<!W>YTZ9-%jCqIvjhDMj{$q;Ts{@!sJmsu
z^4DSI(|#g~X92SQ6iaMa5+OF>)|{31$Wq$fMbrH2;Ag4<U-4za??VGbb8bu9s^&YE
zPH?s_cPw`-v@e`L<;g4}Id!UfsL=+7wVBoiT($U4UB-7-Aydq(wfyp83ElNFA@}Nu
z`oVu*M3?t23}JtCPBpZJSn7J>er5)zFlN10egOs21*wmvR^~2$UJ{;lV6d6xCSKn<
zIzpjz`jlm<+On6cdIs%}5B&IPom!rIk#F1D7jDRUlPop3*H%O{NEa3j{)j*AH+a8!
zR%_~ecHnhW_k8wi>Bi^d?lSWTx-wF&?<}cC`|VZn!>!@MPh+aS7q}|MorxpOnD;YJ
zT5j?+`erxgu_V?p@VfKKrJrn$qf6DL-3d!ICG)S$V_n9!HEe6BxcPL5Ok`{u{M<d|
zDiiVjx=ppSM^{C(s$&vO+~~__o7pG=*7Z|7&eG-6vGVG&%JkCmJH0mo+pp2JJ=cDI
z(ejo086HOu)09I+Q`XXM;#wgava6w=%ap@3>>))z+d0@TakV~WF{!8FweOe*TW-<&
zj}{}$D?0^dhI4syMFa|{X^N&I^@_#{pZy#q)n8qTZE%woXkW7SaC1CkKSnVYSWB8C
zog+<M@$=Hi+{A=oTVZr--vI+Anxe%YE&0!Vy<|?g@Ve+POg@dKs=3D0Ht%>l;-i)!
zR&YT>bZO{idXoE_S3fVl>9KoMaXrx9AS|hiPXN39YM%%gw8*tqA81cBIE>NdkYCm~
zVmY8IqvSsMSFM?chdPzkczaxfj7Ra^TK;W)6=qCsQ!C>u*6!Bs)3f`=R;3T^6_<EW
z^mAhQYpBIgxrJUJ-+*f!>xif2^1{T5!Rm#&ihEoCmMhS)Osg4PNy!*;8gkG{ZFVSi
zD25*vwv?@^M<n@M4GNb)v3OJcj4bn#PmCogy_r}T(f2AUoOxRjO|>%jDyQfB8xF?c
zl{K{o<(fI#o)a)-H19b&l(JOSVblC`bZ7=szi(~X#Z08enu0A>y)V6Z_j@d9%6=s^
zF)FK-kn6GCw)FJ<P*-P^YSINH6Yb06Cqu)AFzY{WE6zqgZz$<dP4!BWw)6ZM7=7J~
zr{}|HN3!9}$jofp(tiEiPM@2_AI>H39?^Q~wUvi6qy2fOTG<nyl0~#6?6UGww|X*r
znC07kJ)1Xb_kFpG`An8l=y2Yp<`OT@sl0|RpSP**OJbKo)6+UVFUTb*ZOcjGJug#|
zCzjXtvAlN9nHi>=g2Hz<IJDPnGYCwwT<9(wy)joY#TTcPuytlUrOpy}z6ErZ4-IoZ
zHfL4rDtfeQPqw8%C2M$E_oRq}k(p3WlV_HCf)dTr$^iLg=GD=kfkbnnGWjyP%(~1n
z&%<YC-?k~vZmTMI>$<vdaB#=U^vc!7fePB{{KrYnGanxtecPp2I^(~GO#PIplGvn~
zjPjyMN6KP_z?@M1&}wc`ch}O%DU-a}yR(XKi)8fRqPooLt=r_C-CI^#T?LaYf9TDM
z{k*%~*NTEZC@I?^ZGM@e%Zs+Qto$Q0Qqh>WPP0DdMDkM%5kj&(F)3SdKiZCrPGid>
zXJ+I3uK#G*p4ah=-ujZQff?JHP7+f`%9c}?=p7#D$#I|Y2`tKad6rxMO_pB0NwNvf
ztS}Ml3eSha?<?R9J+oS=<kcgKNk#e&XJ6(o&)jt5Wwbk$>2W2TqRRcm=<&$|i$Rh_
zQT|8hZcd1B`iPyGSrQ*_Fi!~x&8AqKmhKDHvR$y8d9jbTb7oP?-HqAxOM_99#>c|B
zl>Ol>V&Pm1zAJ;Sk%3bsizPqzEf<mMGV5V~d^;ALetzUNY}^{%+E5^rq<y-rLMZ@w
z2jZQT<QPA$JFY*IXX9EwW;lBL4K%a6?Dc#mAmHvi@Il!?{|(maV=aG|L%Wrg=gx{2
zNjtFojfE<m!{x7qyLMy}$2Lw5CJ8>(FVN#-mN7l$mZ2Ev{$=XxQk3W6)DU@Y-&yz>
zx&c43XAHJU&o#Fh)(BR(xkL*)Ufb(3(<E=L7|Nhh_{P@!i@EUF=US36X%QFul=*=o
zGU-Ul<@?Xy%B;*+MMgKsj7@rPiw=}Au=Ws@A$?&c^PFI@Ru$|M1sGICPiCL)v68lx
zLcV}2iv#XI8uJ4EGx9vxge}G8kH*jwK^G7+j6CvBbssS=YfMT3uNu@wz;K4V!R$*c
zeSx)uN8mrhIOUdSuN##(Y}i}I=lkqu^<_t!Xz4W5>HcbB1!|nsv%*C86um>y^8;Hc
z<iDtRi!?u7<rQ2|9a1{<cO=eD5L2hf<~`U}Q7v<br$ynl0t*>olvAg-XqTUGtZmQx
z`edE#{;x?|ndS}rN@4twMD=9%;AEP3)jG?p?F?Kwzxz3fxDv^q#=2`bFKbLwSOmRA
zz>BDkGo!9fdwOg9`<gSGX8+!M689BK3~h#);_JPIW8o53Mw#;+zS=^&-?6WY?aoFG
zOjB22pvcFh(Sgw!wjg1GA9>dUNvQU_@kj4Ie6`df`v+}bgx%?yZth(mBF2woz3EC`
z7xm>n{8==lq1PMedwJYoS5K1hqi`n_n4sHgv(@FDjHLhi)aF}ho&cuCyiK|_{II#f
z4hM3qm_hhrPIz^8l+j<o|NSGUS&q6HGl2l@+JZcRyW)4&L=nDooXF++n)7>7PvR?E
zPjam(onwdeTzg}oyQ>msSGm70^UC;L`G(vFa85ye1RVP0*?FsjUz5lZiS;rspMe>C
z2cn=L&ZUufDvIL%H10Iv$QlhWPorFT#O?coK2;wnn*Vq{&v9f$WXx$@S8)%qwu{un
zODnTQsuGLM4TUqW8*^-o*kvrx!K)o2U^q*D#9{EuO^5U&R;~HG%^&#l4vgkc)x=im
zC#}l{T6DL6@#0t!GYkV%UOd$MSyhHH`NHe>d&s;g*CPnr=3y^uPw+ItR_RnUyh89+
zJd@#cf4zJ|8d{|+i%=r{=p*wKZyl?G!y`d{j1nSfZg7eG`gx<PVU{WTzdYAqr|1|i
z)U&Ubx2cty9oA~zu^)Hn__vf5n$q##eb}%ffRUSCReIhy!twi+AoMgb#6)m2%Xj~f
z)I-fE^Z5f_E&i3su%go0!r21W_&S(fzOleu{_*~}!1CO_sK>#kut=klJKN8^(0s1T
z8N`hDwT<t~b6JlS%$AqY8AQwoS3*5aF<i*R=hbx;q$$XfavZ-`U2Y8n1rr!UOLj;c
zc)+Jad-jE<;MI*wz=uXojDQZ0LvM9>eTrV5`_+b{_`ls)<q5*mk`tv)qq%+R2R$b5
zN0(Z8F6x{>UsWqb=a&uLMVDLioex~d<g$R5Qe_7+2%A1ke&VG{`f)>ZcUGEUY9b{R
z)Azm|ZN5#XPfsB^F3bOv|H$ShtleDp9*LXK(8}9w%(}+zmmkU|lD*W<*gyE?`F)XU
zc+FgS6=!tB0Jm2lQWSxpGiBD^z29G($6YG9L!?Zd&}^C0acAypsz>P0aeUE;F(y>-
z-%SHgsxly{&GS`=h(mb43&$@fnr?7$ZoI>0z9+0=m*U=CDvw3{sZNVNX*avSUnN#I
z*S0A-hMX8)LYL~fJd`u@-CSb^p3U^7PIgVHjO)eAbg-7;RM~QXsYy6!>MfeH{39i<
zCw6PWNj?-ZQ9`1z_&Bc%PmwB*qa*lP>$zeVVf&uj`57zD-EF8)@d`-<XrZ8GOHPQw
z9dZWB2EFS)Rt}F|CeUWJZn3E-`;?%p5nIsd1Fo#usy6t*KVZBt!?cx^hB^P(kGH>_
zn2KJNOtviZp)slo=CJ9$r)Zj8*pK%RT)Qy*dT*#DjWZS<;grjv(?ge4a3Eyk5!OEh
zeI#gGnaS>AUY%1^A{92P@pBycZen-Ukn@!y0k2kyA2gjO{<vNrD{lj3`LQz+jws~D
zEK{MB&iGu8vBDjYjQEe^d2P#<?4AdCG@rflBnZIME0!@rg+L}o_l4<`Re2ESq`|Rr
z>117N4L|Yl*Qa@vN&r*lPJelJhHa$d@e!+l_o2LF=AQ%Df({cLX|X=uV*go%dt(3n
zrI+W^Sj{JQa)BbH>#-O?PgNtFth|ta|BJIppTt+&x5;=_FndZiSI92mDFAgwaTs5y
zxbtLpe@m{NpklnlZ=yOwDmmYM|B8&NSnorl&QLA=s@77u(-#E@L0qX|g6T=IPP(Ew
z`4*=xf7kk8J+$Q)Jqd*X<v0n(*rY!5%ICS+qSZ)|T3Bb<jV)VeDvHza9-;e5@xaq)
zCXha*s(h_f{d?7WHzKBWX93@C6vhZV+BKvovv5Xoy6(8^H?IO><Fw+NL}Z=+P@O*U
z_^1%Cu8>58KiWVaMV2nufIIit6w5&iAD)Bhz(RtqGsBm`>P&bI+}_+}OvFxl3CC;f
zx2^K5dOm~&8qe=gPUtN?sC;wtJ6PMq^*0*(SE`{Z75eEj*oNl0Dpe14@?HHmWvfVa
zJ?47^fpZhW5+JiOcOKkI)!$FA!%lXG%bo)}WfUxVc)XG=FrJYqqok)vn^$jhmv6I-
zn<ytpufWUAZZR*#>~yIo{M8t;%o2feQGzklSBmwe69|BYZ2s9T#t&x+v<%!GTA=kT
z&0*+Ut#*sY;)qQdjaGL7#b0m->Zh{b;O-yd_kT#({~^&M{cMV^Sh*hqRiwBB&uL+I
zT*?ndE{Kf)CYJkpQ*kC5+kMkED}(76cI?VmHV<IkQu>^3|Egu+@A7jn!9J+Eun2mG
zxb)uG*E^(`5HPU60&{S)cK3Z6Zs-YPVg|Aq*bV-x-(H3vJwuANJy#&;Ff@9oy}a=H
zMSB$nM6?Wg9~m{Co_?L>I2s)%v=MZ$s2LM@&Xq97n$<>ek8rLUOt$-DPZ!?_%C>(X
zVsf9Z7oBaf({L0Yh`-+%ZWi@ckwt*06$T7Egl@{OoRhZG94a?tPm&p*71X^MN?k0q
zG+siaoCvK?Y4OX`s+(R>mX?fN)qMX4CpTcL)+eewzWFXkxK>h_=j@A+GHsH`Uu1$s
z&5eLg&;CpcP$jq-^__b{ZJI5zX$)9g*0XS8Ik|5%B@Bl*Pa6&;Q8wNFXol&>*mKDj
zZa2Nm`qz5UvpDao%dzQlnkb{wuyuzq!K6&$nYg3!(SU|~o2x^<zdp*pjl3A^EARqy
z-x~_g9UNL2ZH59I69gbt_f91oU0bfPFU(rF#=vcZ??5(p0(W_qTZB=rqq(9JrBoWM
zFZ2ssUK$ih-(sa9U6auu0RpK1n8f$sLdgw(`mYa|W7Kmn9)U_r*G68+7Y5|JEDfl7
z@{~p$!i6M_!duF$U3-n!^80@;2G3UPobxND0I8S*^Zd}z_5)vbDb%Ar_^dy!+Auth
zPo(Z${6G2ElVBga(gVX5Aj-NjU+bC3E`;m?%*&qP<l4G+PAo%&)FmcAf=GU|#{YgU
z4gu>LvA$Adf<Psfbq|get$JvJP<3ym-m<<w*U|o1R&qes{I6P#_dl3-RE_B-%K78(
z$OJuCmd<CGsvw!V-H`JDCmpW)V(2XeUJWbZ{n>w9aSK5>cMLteG(%qVzxQHucV5vE
zI736+soLh4qqm-^J$a{puqpJvo*Mu5LfR~;J{{m!qjHueKg5|gq)Y%|&15pi>vx^~
zA1jzS#v%Za)$*8Ug;-w%RaZq|oBVz1aul4p9yQUz*P-965^k4Zt(BUX2!T5~udkkc
zd&*iljDNX>K7>z4r1#@vI{!UKYQCiHqQif1yuaMsilBW2M&*9q5HA4A5h}#adI{aJ
zc$7fQfc*N(d|$<J3aJps?jdl1$Q;A?ZZY9MUg+NfbA<>+sN465_(2{g|MYYuvTZis
zuRNt$vP2jGrW#&3X_&*OlO-3Y;BN@A%$2V?*_MnMX0-tjCz#Yc^ZIA=Krnt(D9e!i
z`9;G9EVha6H5~Z&uSbcwO!W7Tz4XWL?pW64q~Sg1XP%O~|L^U)H8Ug4u?B7CJWC&V
zOK)(0ir|C|hrT6WxfCpO3f{8YYIH~f&pre(6EtU8_@2DKzpOFM$mo0Kl{Bz2PB3MN
zE!^b00^Yguh)(araRG3<ExSU@9!-)0aV|#(Fh00bidGhuFhSh#dm&wP^udA;4_-QE
zKh`QWV$GiMj>F)ZLx0`AlDo=Icp^+4COawD=e3JXo$19qOv+(I9?<6Y6Xrk3htm1>
z`M&5t*!4tz!=Hm+&p-l48K5?ct~c&2+X~K+#K3K$%m!R*BIEL{9pYRM4ro)}W%kI)
zxV5>{TL=U%7T-W)P}JEW-=!8PoMYYhr~6IHwFfKndxOE`<Y<PVdY`^i$8E_oLHl8i
z>Bo_V1atU5I%37-OUtJuA-;Y0Aul2_TDe@?o^+#WIXoK*`9XEGjLf6M<`1%#8}wcI
zfcm6k>~8)ii-f7(38$fmLPQbEmWPye;~gb&o~tY2yPk0&Nr;if-uRz}6umu2_TXQ<
z6cK+;`^~u@V;+6en_>@H)n3QOOZN`gy;>A;8gHjRbb03=*J-PJ5yFW2G^2`;1InR1
zG$1m*@T|Q#H>Tl^L3&t2LS73IWgG+qY4FOI%(%l>zdY*)+`!R4{xW`8ZjV6%X<6{|
zTZm(z)J&t(KtdR4fy{P1x!cg9AZ>!aC|65KcvT1Y(boWgnASvuYrplu;sKLZra7wR
zlHpYyeKM=g-!)jY_aD~gKddeC6;Co{S4hztP3pnHX8!d+>0{f}v0gmsS~o27`;hw-
zg%?K{0p)?4Byg{oP>sy$($b9}4s?_!SOm=OD3^1F(~O2_oc1<W`fT3}ED$E;2M4^7
z5c%HGc!qaX>a1cqzR*XKx>FwvR?=VPeP)7-{7XrZ>M*d5&vgY>K>QL;sRDs>*N^LM
z{ysFkrYGXxp)<m$5>95(^KWU)yrOaF>K87o`;uY{PEfzkp@Bsj*7vD9wHPe8VP()h
z?!O=`^fz`C-Sg`WS(pif8PWqpJH2j>X+0LP*v%PXeGFkCJX$Fc6ayawRf&R+<;1gC
z&L)E`7qsatl{|Y-1g;{a!;<RT5&T3u>zel4kKc<{m&a=THJ2BD8bLa6uX-nb=+5Fv
z1fvlW;DsYcFWfK|831N_uRuUG-sxl4mA?zLgf7;3s+W5{%kqafr0VyN?4bO}j4{6Q
z#n}5x{7(B>i>0{<WaWA&?wo?Z50WQvhDJfZv*{0JA;QlwZ}fmr5#@Ay!Hm_$KB4Bp
zTd_Qc^_Me>U`~0}A2EwsUu<!k`Y8W3+3}YLfg@nEu3)@g3OR~;=K!E1kT!4_by5>!
z-cdS1MA!w9w#t=JbA!bT*IIG}gQG>zF;+@K=(5UiU)p`46vWD~@zo`pCX+Ce))U@d
zn>GYPNj29a;K^{L)yYU{6=E^(1_RH6(#}tUcn^vdikKY2U=sP2xeA#?`y*>wsD?Fg
zW&5q)b3ZMe9{8-WFa!voACuIZ#?HE3=mNu$5YDYSEe{#2`jq+7J&pW+eeA@D%BD${
zO(8)xHKQb<8R0W>VCil7Oy+;{q|%(@$yV;X5Sxa^3V{=xL!g$g>}e8q;i9i#VjPDX
z?*%u&-gJG@Q2qGajRVSp7+p#=@-XE%ZiEfsuIO_r2J*c$NC87;QlXV5XC$lH=cU2+
zO|WF$@F-nq2l6U67r9clMQgqo2%zq#eS^`mPe@T>V^hTu7N4p8`s&y7MIP=U+!~2$
zi=XMylBuI$$?r10P~vT%*a2&3YN`0yU=SVrBthNIsV)Ie`LU-#O?)asR1lK*2eA`i
z|BgHmFgWckTSL;~+!u*RT;&k<ONZ$cc`SDpE{=Z4b>ooxpDVlwKE<Z-(%l4$Rpe(-
zjG^7O8!y#g+y9XGVu(`r6s`PB)i@>!NNMLiMJDd5x@nb|O2RCwHm1){RRaZ$%1gVe
z?i+zEMK5kqOgfkB^HLSNUBxS8Tz8s1kMrU778zMig4MCr<)ZUB)*XH6Dx}@PPmiif
z{rs}-Z8nYpebtHpFk^l$g75Bmz$@)~Dzkj0(5CVT1rgw+w-9Cn*dZH5Mwk%nWV$jW
z7>@p7|FY3XG0m2Fd?|$sgU<j8LLM3>x-jqo6!zB39#6>D#GUH`CQy2Ti|5=+dI%Sk
z2QIH;C3F^lUEo}|QNy?4@Cu64dGQW!o^KrHi<JU@c`u{5eWev~)337P%=XHD;FIWR
zPM7*5pFyD;0(%h#euS2C9pw@-fXD9*yhr?-Zf{B^Y>65x!EBqS=k!p6K*AV=2jv;H
z(mHr((6~Ir?Gbza^*0+{XyZfBxppibF<HsRow}X=APX2c1zfxvbOtUVbXAmat!|Eg
zAR8Be4-oT-0<K~>KiR{*1>WE%T_Zm9vYJGbef{1TFE#SML-;;=)23|f9fleM2P*&|
zLGZcW=+z~&y&iKNUvRf%;DDj&JMS35VbtecA49ov9Ju7$b+P+#U-|Pw)zS%sO{={E
z7$p)3&lI@)L>jmbY#>RL-3^9s(fPtl-CQ9UBDNc}TnW-2L0SR*z2J9#)ShF+6(Roo
z{Fq=GgksVZ&6NqL-|Pp7@6@A1&y8<=mRMgIi$6%ovM3({6U-tV&ar>q>X{CfPk@$R
zm!5`4voY`i<9{baa0{pLx7)f#^Tq-al*2sugZ3T2Q|i#5r<MM80GH1PDZsp3=_uOk
zKOD5C$J%5@>S3`yFs5kC=)(?N@eWV-@Xk-RmpFfMG*1xPJ{}62ovx0Ol0d-z<+eii
zUP3O)b$eLs2^r-`*Yksb{v!;^ZLZ_)6`dT2WjbSh`M+0uWBk=h{rPTi^sa1>2U=x1
zHqQbM8VGbf9t$$SVEp4Rra)`kV*Rb-|1J5O&wf!AJ(TySu;I_JGsp~lm)apUU3dPC
z{T0oma~3kLHliC6_Wyiv3r7bfDFYNjN+=Y7r#%;M@f?Ie5CB5Jho(Jdx~=h?!+*-%
z94pAxZEMQ|28EK34qc&y6iB7{I|^w4=e=Eb|NoI%q*U*AJ`}Yv%!D_u0QRI!|M3bo
znirA<<FIc#6f3A=U1vwH=<q_d;M8LgJ}B|1ohZ1nQC8Y?o4+VqqZY=PCg&D~v?zLA
z++_%<r-z$h<YSlLv-4@4g(3r@XpxnFp#_K%(;2wgw78BHT+xy2kGW-taA`PrI>W$6
zID80>nMH#LNpsaLL2P=uvkV7`;OPvD6%mt5{}a*Gr74ZcB0vecFx!|ab?CgoK(aIw
zQJ2tk5?i(vO5swA=Bu~<-fsBzCefJ`-P~itXEpHj+t{ZVj>i!C=LEr*&$X>CxAjAE
z0mTohaLOprKb_26pzPwzh!N~0HxsJ!NrJ(PKwo50M)#g9PS(p4*)4L7evint6V0~j
z+S0b&mB&WU+Tq<cBwkjXumQTiGAYVYcf*<I>IuQ*s%>~**^qfRh4zu`wtx$-ua-eI
zMMD=nGA*NoV>fnZw0Xdq5&vyD+LB9T?!#&B`?MJqJ)ru;e>3VZehzh#_*!!oAHd(V
zRkc5|jT3X=?TioP&M$02jGM~9;*UCL7-SZkYSs0h!<`Y%1@qWSu;x5n69i}1_7*T6
zh7inN(J6u)!l=}1m;9FuQ#^g+BEf(oktp}y6ZXjggu|$_)SF{hur%IRHcs^fTOHI@
z9?E!nyyq?aaQzM1&PBAH_4NBcKaYYs2}A^bdQy9x+A;SuAI=Unq<Xo0J%Fc(92`N^
z`JiKb35)3_AIib`90<LjzBSI^4~iDO3xP@yws9s`K~zNtzPo*GI}f1k+y*bPR<S{K
z7TB)NpyNBdVV)OjFsU7W;*iUUvZ9*_>TRB@%OTyUh5>f+eL=hlL`Zngg*{T14}^u6
zL$jf3RRP7;j01Z*P(lX~3IKF-5$E%rY;+yIo<$;vLW*6piOzUWt)w1f)mA778Dq%K
zH}(UR5Ulc`aOwBk+>n*q>pGk^@t!v|{=BZ;$HRi4$_nTSf_mUq>281B#>aPiiBCm}
z9~O5Q45184P!~4ylRNg?HG_%>AnEgk?Mx&>wh&&J$l_^Z46Y2bcNfIDY+Jdv<3Fo@
z`-cZg%K%$3Lq&&so<)9Bsy438qKZJ6hmJGZY=M$O06gD&%JT{0L|ID`U|1(lglxy#
zKaK;|IE;x>zON?IXtKMCV{hc>Z&GA%;Gkit7t~#j!~!N`)9KBr@bS@Mn_ZRM7+7K{
zWaR3$Zw3+{cP<R2);`z8hU#uKu!K}L0V}*E_{eRf^I3eM7uZ{N|5`I7Jre_`gs@3H
zK98x#ZGK24tK%IYDWWq}eZ;D^L;%g5ZkgAbge)qOv{D1$-03{kT#%|^j?t-+YxT$P
z#?w2NXd*cIU4RLAfz?Yqh%);6Y^yf6-g`LTkK4PTCj`yeR)MG*d2gx3eG%q+GAZu0
zk-MY7e}KSszcs)G3POt=qdMy2^G&REQsy7^vB&=$g<lTFrFiC4_8TdR_(qFXS3>4g
zB}OuxCaX9zL<BIu#+kXiSWgXctI|yfHv*yd+-KP-^nnZ`;CjOWs;Biuo~!xRLF{m9
zlSjceNc*Ml#XBCZa~LtZUO5Q3G!`;7ucqy|Rr8Bo#S~!2>KYc`*ut#{F)hb_l}DL$
zd`+RUy{NUBP26$#A?10nk?2LTXP{b8O?E*Jt?QL4tZT;m<G-rdbA}kG!Itm4yO$CT
zLVbTaP;&A_gk6o;+%LC<6qln8j4(a)mMNV&(&lDIzE|{MoFNT4rhvu?>X~fX`+q`1
z(o)Nlk?*g4!-dU8E$n7c(2gOe1;KfHi|r7e;G}D@%v)Q49pom7a2b3q2x;xG%BgOw
z0l=Lf2>EGJw;`Q2pR-Bl61qhoAjlMM?_<9iZ+n5o`NmJcJG@cu15Vix=o*6HI*^~U
z6SqQ?RF*afQhJ)bK7?&r7>waARHyGOrBXdLkOY;eG>_#4L+_)UzufmPjMOftzR2sb
z;2*g29`Nglqwoaqfwudm!^c^Y2=}W-?FEK~4}xrxI%%=C)h!ny8FHE9`p}7XQ>GB|
z&iCUdZYyrCrr3-~I!|*jyJ&d?MXdHC-w&UL!V~aE$PeIc4R+&Y%Quou%38#TI1i{v
zc`5hK^)D7{f%-4<JfU=3hKfJr^Dc7WxJeVurP^<7RXVs^4^J>(WfOpn=5>1213+YR
z^YIZpySBDBTp}14fnHws<haH{-g=P0l_hGAv(AkMC|>71PgvmzdMgO9%*{PD#v<Ov
zdFy3$ft#&|@Eqvoj^m7ztTY)^bNBbwL_R)taQ&dB!WBl>HeU7tGG?!~Gy^~tVN=e4
zDy|`w!G}LcTpx|Nw2SQS?-$(sAIvqH6Z;1tiZVn_o8xwvfrr>&Luz#LbMMD%->N(g
zkIx{|kEB-DbNT0$ILLR@59sC`ES!oQGd}<2-M&9q-i?D}*+T>(up96lU8Vj|Uaf$S
zkB^ucDH*P3>yzg}jqhw<0qmC^$poNL3H~pk;6+4a1DQtUtIFH|naIgzkw}ggZ=JXw
zy~NHjJTuZ<fl9>|177QbTWN=$-iaXp=Qo}DSC>jocqBHX7qz4JfC?S7-G?ji{id*J
zZ-Z0_K#uh#p>D~&k7D;BuD>amB2=Dx;mR{R!PHH&pD?|nZ0e<K>Z6?PPcO*IUpUuE
z?yrCTSFOMuCU#eiZxstx?4|k;CZI~<0P#-!bB(9hP8{`H|09eZvZ8aw8`Zn-#NZos
zUWHr*RTUQ9`sBYK`#AxsS;HxDSU5ra5(tri+136|T=W}{d{GCdLKwdutMef156Au|
zcgsPi;voKFh`K%g$GyOvapfTl09}e;V~s#BqREmif4rIaZk}p#A*T~32yO|W%&F`=
zO^o=7$@cio#KgYp?-!_0zY%->_90#UV$b_fY+3Hr#iMU)!QPai@=s_k@uPLYw1T76
zKDYv#7zk#(6r$<R3KF1R^pZ~2w8pG*|9Sn*TT`QKH`a!4`W)w}hdTi(S8cXL9suer
z^-qr9JvOFw>%Sg!^Fli-|Ar@OI;gbjJ@H9dz<(E;LcIT6%&i?h9Lf($0jQ3WXCT4J
z4sJg;0@GhU+=hfR#d@!Aawsj|h0=WIsZTQvU<6R$npPIe_W!Ldh#QL`#05%!XBP?=
zV(Em-%aRat-QnFby*ywGUIu9cV5TK~kDde95^+t4Qk(D=^UL$BZGo&b$#6@qFjU@X
zpheH|kX5TBf~2HeroBMO`lxQ{f;$eHMIqogHP8So5nU4HY&ZKsV)l40bul9;(|+h5
zgy*+U5JHYaRdt*n3+(=S;~D*S<gD(iN{jUr11M4s+b~mjR1t}c>^V}feeW;%(hC`%
zAO{Uic{BKeUflk22)G?(;avziUt1Gqi3Slw0O{aH{s$_Q{_6#v$3-kcID8CG3||Ej
z0KD!C%s`o*{&lP>FECl$4@qCT3Ye`d|3<S;6;m&Bh!tqfSnP4P+CnhTjUvRG+jK!D
zXT@$9EOpW$jI)>1=Ox~<aSI)eL(4+$G7YeV5Zd6{PGm&fuQgpl4<BTKw$q}tRC&Xw
z-jMx9M$RlP$zQZI86p|1-4Wc>aqUk+I-FY!$wwUUJh-<3h`S$NiOUS}z>lf&VXVi|
z=7VZ=(p}YYygxRMD+U51&E37WC(v1QEP?dO0cS8nkWUcLG2#Rm2wtrPt`!djjno%L
zp&|x}eX^Y4Csbb~4)Fuk3btv$hZ~*wts!DAhioV;_Mmlpan5}2<0)Om1-vJ~ox;qi
zc7xRgt+&-yD3erVc0UBOS@O)bH(_ZQw#VQMPLs#6;|Xo02qtgz4e!G2STv<0&}}*a
zlCI6N$KSd&;{o+KIYPuU2GKn+kOzq71YlJPP^?g9=Z{1=lWY~z7KZBucoi`qvKnM*
z4jKTIe;EcLM3J40#$9gnY>KtWB0Lrb1h0IG4`GD&Hv2(&^_|v{ixfycD!IW#h#igy
z#@R|(E6Za-6v{ZkTnvq4g_G}jyRw=sKO}0F{sCi)X(U26bsMvr^))~q62=g@%wzl(
zGqvj}actNh4AxyMeZ>V&q*(%qMgdgM8BYJ8PIw^nQ3dDiFTcy-%=va8M*5NQ=dhTp
zu9bj{pKD?>(e5xf$EDp?P566p6uTEwwKB}$(raTb-~K(PlV!;u{GM&a41u%~$sp7f
z{K3zQX56H~=3Uf?3uAcc8_9-}!iP^}5pI)J6V<}C@%fDq?$b-ViT-95WS@{Af)l)h
zW)~`;?FA~*?Qbb}h(rE?;eSPp5Iy=Ja-_Wsq=qnD`S{6y5>(Odb6tYn=aBsC4gTg7
z5y9M;U19nzS4iI-DV*?kf>K5tIM4a*?5^mKaB2!4V(JJ1WS}#Pc6U7L8p?B)=V55!
zmCuPbgU_EndQw;ya#&NB;CidG+0rm{KMbURnM5%%;yByfj>VSBD_P(P7!J{w(Y!}B
zyc`q)zAD+_o0MAkCg;sp3cRhjLRfh1`4?dAh$BP>hi~K5jzx`&pz72#!cLWR!t@-q
z*>B5>5Hzs|Ra@s=e0wt_D8!4N0>=dq$n6f!0lYMWn<fn_+xx(5^2$|TyL=U#Y(128
zoq9sGVj&@#7|y75dC#`z@6)(gWD9{9UmetKPLQ@aws|akjxH~lZGHSXS9!fLVPku^
zq%i_KOuJzqRtiIxkHBuJU5(c<ht0pqfgXMmfyt7!@=$-DDX>w*KEgRDu6!8mXwGZ*
zTwQFdkf+n5hL`-lP1$-sAq+)~RjT+P2nQR0E_%Qj;EvS!^u)OQc<R4ThJ}s@Rdm2y
zG_cL*F3te+Q3{n=bx-M>7UO}xUG7(u@d1lJV#lzb$TqcT#U;R$j}M)1BNZHpYi!8Y
z#ig0-B7fs^-u5~WZIPLJW#ENbYC;6?f)Y#3RDCJO(Ux$^Vu&+ofR8K?UOfG)o_6~c
zlL_PpIhH-^7WPI-?5QZqDgg}Ol7ff^h}iK^k7P?JnE9U5^FU{k>}VER&A1pmh_mf6
zR9-=R2bd;v;_(DjMNw_jId-4QrZk3g)&Z;fp_Hqm(nVC-1P4=8nSVnbmvi6_#W=JW
zXbCX9KtR<S2+lQvz&L&RUj7x~CldDPWDeoIk9z9nftrR{LTRRASCQNNB(iMRe!Mws
z@vTb5(l)~!4~dVNBLXW~nA0EYODcX^`jy1ua7NtqvDP!G`URa73W5JVlN(UpLnSUS
zv(do~0OEDN8A@5N>>)YtehT_8xn~wS8suh2e|$F8OEWA(5Y?Atog=lH@tgiGrkki0
z>V!8L+93?2K?Sh+8*&?(u>PStq&U0qXYzTVASs;&i{q1J(cG(<g0r-fTHr320+wNh
z&2BaYzY~w0g@#Js-6o4WcAM0hi|sdRG`?Za<qVSiM+?Y7A*E3FVg%B#_hiq9KK)NZ
z>oN7DI=c7{fNPuMAcaYT+FNE*v~zDbqGR<u@eBFg+-@3{?x<oIs?!P8N7a+V@O^)<
zMclKL53m(x@c-H1|0nJH@rfifNSnKW^irk4c;ydH8F&6WD;YPEQ_45(H_1A$7cJ_I
zvhz@_$r8&LK%YGm+y|kMf?#z-z~m-Hg$-DRJP7L4`4fY`oGSjt5T+Z%0Gafj+mW0_
zk@d_pL*^(0&S<Vc6w6<$di$|ep%@|_Ws=@ya=Cm!@690qJRkHaeZn)un;vxi4ndIw
zRi*TZI|Xc(1om-|VnqnaGK2-s;FkOee6-yKIcSgo(3ar}#|1>&f}@9Q<~}an!9C07
zhQt2=P6{y{B88puegi^}h`a|rDCDfK6(a!PSs%GKBFfrSFw#L)SSiWoheJadi17(L
zv+%1B!f}}pp!S2MZcdq#-ftkZ7PRZF{{7P2JI|*ZpG1U`WQBFGG5%jq`0I953Sp*d
zBm`iRWJky|uwM%~05PVo1H5R*3U%SFJHp0c#r^ehaqu#QJP!e=>9@EHB_m$Mng8An
zGNJ(BnV&Cob5`mGcjbLLQrzT{!ZBP;-7W2gp{Q9!A0a=GCF#X%jrTft-<ts5j3NYv
zB>%l12?Qty1$io4nx8U)>c~V`(GrJ<RqNiWRRch>4}4T=#BKXDh<XGAH$hcM<<@OG
z{Gej;2`^CkbxYtFD+pH>8j2KzjMgaVu!BM7a1C1vxt{;T4K{@sBS{UW@9r)B|Fi(;
z0fPnz1oFp^g$Ni(L)GQsz!M+TC_*SC2fBZU{GWsWd3;DCF{(sJKviXy?Gbuj4+KL*
z&4LJh+nTP=9Rm+=ddH@nhpw(C=yiLDw5&~^i=Q;%RNU=#G3rr)ZV7IQEU3IOc`n&R
z`@0{0B846~4A~CbMj&EutXNir6rnfe0gQWq>>0T5bEv|PK{#D-`=6rwIVB8gAAt6Q
z@N@$}x!Ii6G9flHXu89Qmdmn=aKg5SAMhA*m;;o+Z=mS>renp&byEjydLOGNymc-P
z!XfPtf3-iD!r^dPs5KY@@XIYR3zgtI6wHNY{*Bl$fATbJZ|nBlfAIdUaz6%KuOh^G
zoy)Ursw9gL-TFeLWuhT&_$!7Ek2xN?*+F8pvNVr+BT}!G?7$vsG}<A%78Cp(q9uv&
zT$!J!J5G@x*8<y*YRt4IteW@-vu;)l_{M=<CS;jEQ69nwoGnIx;a3L#9>(&$oq56h
zS>4s|T$gX;^^ARsVY_3EDgwAW9zY1$3aqHP@*{=!_?~;+-#Da=55-!fYXqW^Lf|bT
z1ZpX?q8WpJkM(FP0R{NXu1ZP_l<eeC*BDrD?p<wAbGHGk-t^@D&Bj8U=)lhO5CE&v
z@Q`JT=*Ve^H@u)uLb(q;=FT-)3^+z};=yHrLNx4wA*0s(70Us_t{vk>*WZw8&;nH{
zZgL>c`axJM(&pU#8$gP>9JKTt&Z0I6?j}^%f%xtbo3;|RI_#qa=X<8A_flT=4@Vt@
zi@y&OU`G62<G~B~A>eJcAW5fB;M6nyfmc(b{`_{yOPb$K1Vc&Pm}_VD<;7XvR=8gX
z*PR`VyBl=vR7ez_nxk*yKpF<7F36&GLTSw(if)El@jxo_y-h}^aGR7yNYbTa)2LPM
zz;xd#KE6YYy%mgU$c?+(_t0+ro4hhW4xc-UJF#}?8wJNN3V_Pz!Z#4FJcBCjT6@oU
z0y1wbiyW_jh%7pi`MDEyyD3SSMLC;F&NLb=el5I(C*^aR?Kl`I=#a(tL%p@pfjG)N
zp5Ku@N47NmRK_2*PjQQ(S`ccid|9J{P3qMo;hsJfVrH)4N38TSuYTo`1~@ZQtQ7>~
zp4%mS-r&VmqF{<8l!|6<FIZU^Is;Vc<uN%jJmFmzdO{`Hi{Nw5KtzPN&`M)fTVqkx
zokT7{8N0_KV8@z73>oOhaN*~krY+m{?nL$z3ENgSx=sEVW|(p=Jwg5M`awBzuwVG)
zSYzslpqo^NTJqp^zz%3#ns>9Sg%CgqeO2!jPt&LlG$<NEWN-!yLhP`9GkRc<IdB8U
zEXv~kqd)R0pv>{<jtw5r$-$VoIUj)7xBx0bTRQ_JMo(11cnF1wPG132cMrUDcWuLG
zRAV4qOhfH93c)12h$aCcf<HKBtJU3(hcK+P$e@FM?@tI@hpK#Xu46kDWm6zl`4D>V
zK!+r-Eq5bIFrsVt<D;S^iZAd|&L+NtQ2DzI*w}Y=NL*e&14f1;Czs1Fn{zQ5@!QXh
zNB$Vq0Xh@gymYMsup5B&=1Eh%^Z<%?uF#g(33aKeFLi4)2ZsG%mrDL-e_Wys9{=rq
z4g-5O_x0!|T}(a=7hZw35Mxk_=ebn<NXieX9zT<y1#*uNtHM#7Bp@$<d`p=C>b&Fk
z0EF+UeG5uX#OzEbw#6$FxN-%Zdh-dYUkHDDzYEt*cS+e1OTKr5NGoj2Om<&iG(Zyv
z!WZP5fod5C#QEDTNxitsL$(mM$>s}AF(~rLhmME&+xv8W-HxzgB0Alj6w-4SeY{Tc
zMzA-r22q#!Gf9afk5|P*3@=(%U~fCxjJ<Sm6lB898r+K`X|k&;%UvmXqe^+JN4GIj
zBT_LCTipViGuW^Oac!N^bSFli?>T%m<UtP<?ED}#--nkVVZH36TDxo8qW<@7JpfBn
zLzh<7@m<_z8O0F(bQN!<SbA6cHeKupc9WrV8WM5w=6s8JxR;;l$43c~??jGX0k*?|
z^RHN->o6=^fpRlQepN)E$mpQSuKJ$Ix_C=HOjGS|$}j`d%m=i;oFWy8D4ZEu_xF{@
zFd(cnZv5lW@0B12^vZ>2P+g%l<uG2~x8hcKgk}i=*RdzqKmxHx;a{fW^ym)>z$9Pa
zO>V*iA22_N6aw}#tj>C_(g8Q@0Yye6(=P@>wmMvx^Lv)XW7c~w8>gZD^#0EE4P+5O
z;X>KLuv5X1jQT^&lHk6;?%j`)5?$-JL>yER`gjI#(@(9NyY8T3Ky!@Ri!s~`92B1r
zZ||;9=z3F>Y`ZTcGhz}xn(e-pZ|<L_5bF*?MS7o1PybDU{_UDOE4N{whEZ1PO{JmE
zD01yQIWP*DiMvm%?;tTmWH(vqur#@Fp&q+suwOatrup#j@~uTXHU;!ojR28-BMXkW
z6BsThPfo$B68%9>dtDYLaRNh)vKLnB++z6uPL%ioG)pu6^{Hw!HpB)7ihOJK|8s|8
zz}>q5K~g@M0r1T|H!ZC~w|TtQ?ie6nMNmCCk4<-wqeEVJ0{Xae&M$oV7xMPYN-%Ca
zaeW~iUtx5nG}IwNC<twFbb|GVe_RZlgia`O@MNuQlstdmWC#_R2@yDd^CUMOjP%>}
ze&0m=|6%bZ0)?YloswC-AN&@@Du@?FYqlSUy>AHfe;qhHixD9zN^F!p6QMd?0Ydo0
z&-*OPJa14$G}J+zQDjpOhk*!#vCs%KPsVD8l!i+Pyqmm2(xqQ4&$__i7EhgA9L%8{
z3*Lk~QWH8rA>+|S%|FHZ#O$}NQEdU$w2T6!XaOQ>^Nkx>gOO%V0V^^}Yz^c7<xKGG
z5CiA<Nlg7H#vQ;j)Dv&4n@{lG00R&k;;wWd+XPHXMF3_d>5gN2EugUzA_2d1CBLZa
zFBSu6#KzoT5Dy9F_V1kPs}0O<bxDdTIeP2wvi7&Pe6#G5sT#>J0CPUQ==}xIPnrYa
zK;)|vg);1HEKwfTXJS$TPn!LzFoW9+{8QZAC)snv`4DKNS$DjftGQlCRl($X?Of@B
z$|Y=TO^bEOb{pt5l!M6A2-q1D(90++6C8$mL5QgzBk(xx+d4>s7Z3fip304J5<3C0
z^^sYfudN5()>2uJ81_2ea2?i5Ur`75|K4r%t8r}%=%;*blvEdDor8^}IS_W?g(f$4
zTmiy{fkn1E;ST;)g2J(=;a&hX!x<u;;s0vwi6<2<LSc-E$QtpYBOq=>M2u%3Y^~h_
zXRyNNf`rZL<=tbHDAdw~B6-k-&$wY7kW;qMqHN0$C~t_>1!Wk8Q&Ey)D?>#qyxI33
zF#fhI*>X};1Ow*q1vO%b--|>CpT!xXNfJQy^z+xgh}Ih94ziMNRXxeyi5SiZbpw*2
z7X+&NcZjy%$QKSexN%{O6b-6MzNMr>zya`u4DxnJK=*tnR1;c0n-1fVIB-W3qekR1
z-M<R~p}z<$og3Y?HS>5I(WhUrzt&(UfVS_p&dkrJeStqfn;Cs7eDQJHQsRk)Pn(B^
z(cO9C!Gj#2UudX5hXj>9gD(u~U38y9wr)Hq?t?5X2|I<w{-hhwqHPJS52z>WCe+Zz
zJW&X`jmhaEtqB`8*Q+#!%@3CpcV=p)QkC3=as^}Or?r;QFm?Xz!UTR@LY*l1W;UA2
z0%4!JeuTKgRYc8c{Qbm<sgsJ~lwtukMIBl<SGO|=v7^Znm!XWvp6U)XuLIwl?6J(`
zT$DdvuMkNRehhQk&=7|!c3eQfKcJs48UPmTWz8A9%AJoA3%AoE3}lG-w1iqY5gX}@
zLlR!o`-X&HAp)ivb=9p$C->>5L;HdTs&t<?Cp`B#dW9N_!)$AZMk)`v*^`4;<6jYB
zi@1BD07Oo-cmojx(pl(MJ$C7Ovec&LXbWu48d03==ma)jEWp{?(QzBM-u#e$ygh@L
z!A+e!8;}7M$4CMJNf26Za=<2_3QW_zATAURu*895H<=#qzHVI=W)Id9nIv95xpd8|
z>BcYcgOgga{MD7kGoU1B5Dn=^sClwFQ#Sm8<{cwShzMyYgFUJhO<e*}5+5q3p3qk=
zoMi-r3+|sF`mTN(lA7)uFcN#2+pnntliax;u**4kW$EWcTNsy727vOTkt;r%l|b1i
zWQd*wW#))aKo53j3m2>T^Antwp6_<*4L2iwvk*_uKR*V9EID`$PS>-z03U>k11rgg
zZLd2rp0WX`UNW(z=Jqt8Bl_c-vTsIiY~A5Sr|-0H_Yunovghxypa53jpgxAC&w$#7
zeu_#^C-Xmld~@4zB5AO19JuU?uvd;WZ{4%#Eus01DB~RzCCmjuJ2RtAwyeQ_t{{b$
zR%OEmtY`p{fYWZ44ax}F15Em6qRcX$CY%06&NeUtk`854zac0e^K1&o#v{n{9K6=}
zDhDc^*MCkl!923ZM=Bs@R~wfaHwn|Dg~K%ufsr34y6Uk^`0hg3{TAk~h)Qel1c<i)
zbS3%kZg^mCpIl+^rmIn-C#bsr8s5w8q@+<bNd;Ldj~{>LdIvtfCu%1n3Z^O4hq97h
zswXu0&5xlrh3!XLc0!H8%Q)c{1mzK^ixjFv0$>Uj%lg_osS!$g=*lPd;CDIp&Bd6l
zUA>R3_-lfuHz|)GuuK5(i4MKYP&Wi7q5{R^cHFja?1I=n1Oq@)H`-@A53fO+JSNm}
zGTZV8g0zrFFNGS`azv#bCHf0BTQJB<HTt!SEJ3I4J;qOy7VFZFZ(baV323qi6rWMV
zI?BQK3soejivTv!mz0J_4oZJ^k{SahdBre5o4#nNM?fa5e6-BkDt|&@dhjbR@Xhu>
zH*x?>F_2aY<p~{m?u!+2B{i@LB%MPpZG=#0XGSH+@3$z#`(psn_8zdC_5XHebCF!u
zP}Z)120e8lWohHRQZDv8Ob>AUTPT8>37|4(frYA}V^AK5MO`pmgln!1Vee3l_&CyS
z5_XoL>ZNtnuyyngM}a$wMXQTNCu#;HMswH@p9=-dps^fDlhjx~q45Kdnd;pX^)OKf
zamxU@M6M>`M;^p6X;2{L(n@Jz$~47moy+Ar2>q5f6ZSi{K_c4R#lHsH;3A7kJq)lp
zDquWRWH2~fX?jEr8{N;J@shGJ-`Nh8{E37M-5r|?V9u-sge$1&=O#g|al)(aMg27(
zt5FLH5l>$ZB@GO<vN*J~I{bqq<m}fb&e5>vGza7fOP6S>8qVF!+|A;3V{0tEtgH(4
zDZ;XSiSqJlrTe1kcm33`Ic9UJ+t+6&Nl?42v9;{porKhPMc*!D-`piHpGLU+&?Ry(
zGdUuOJLZnvnET9ym+#YSniglJ7w^dba2b<^1pr|n;yvo=EQ1&kH4G)!`_XP@m_bAs
z_&-$o>4MU}YS}!rHeOgsjQ>@8y6+Z4DLgJdo*mvT4%1i%lTyp27l-dN-#(#zh@i^I
z{k~~-cdf5LA${(n4t*HZGO5vl0J$8%N?@c~4cBhw+t@mA5835ANp2d#=kCi+a^D2t
zKw&a)#h9J*?Q#c&r`P?{;?EyZp5JT{k@X-(1OXc~e(K%+Ks9)Q_%RRaeb|@~PqlXn
z6ilmUS~QDroPnH|i<j3IKynS%AZ-k)yw0?`j%gNlQ|H2I7A$Gz#P8yFHoiLfNCc}#
z#ovf8-)Y``c-B~<9gXusz*-uFrlpW@sC!(9z}t%AY$Et&kvxRi?^lf`_XUH&kVKUf
zSS|3t@BBJ1q>>8DPeG(~rN>JzSD}K^cap}C@&p<Cc58q<{I^m_zFI1I4CbeJt(Kdm
zBrDu<E-jZ^nE>S^<V=#H5M_iUoKN3eZDok;<iK}CaYs>=#PF-*iRmQx<qz3>R%09&
z09iB5Q|-`3>W>=c9c*l!N<Eh<JT<RTYVE3wDMl@ME3IGqqCJ_=6sG9qVKatP+*q`(
z`_>K99JfXehX2aZcXGL4pjsh6`K;<=1(Y^EsA7%~cPACgtV{&<w-tG!Uab`1J>+x;
zE2FZ;{DpuPtcANp@Ce4+w|&~M9e@1r%nN0_Xe5`I0E~w>0Ye<m`Z)(DAZm4R6;c;h
znm<IHczMS@V)np*Vc;#j2$7m?5k|%3V_avpuip`4L0d<j{qdjun=S=s1@Qj|=7jn|
zhdO`WaQYWk6dR)4;%r_{6K;E)Sz{RZRYjCe{oCCAaxEXN7??*f0W;bX|3WoVM8o!_
zQ|Q4xj1@@Zlr8>^abYyxg35_dBM{UcPA9l63_R@u3svLfGJ*38U)f?9Qd}aE|3J~%
zyvSdeWJ;%@<m_0mxX)ctEyKjb1gTH*qXre+(;TC!BiFht@q+0;*F>sxu_ea_jc?FY
zfyl}qOfQXQvJbA7cg+_B$)es)02h+~z9RHLFV43E9VGxLJyNgVBS#DP?6c0VD!2+m
zCG(b#*7F%!!PuPhe_c!M9|DMIKR6S#z-+M!(Vh?uA2qou+<IC-Uh;)}F0}WA83cgP
zORw~;E=9|a#Qb^OAXe!x*GOLisIR1?go-TBZy*9o9OQ;g<oq=2PZ9Gs;h~bRf4$+z
z*oVA|D1_Ku!13`xbS-GD3HHCJzNx>zpY7b!VlW?_sC|+%+D#}5QVGOsn7{4_%|ho3
zTn=;MwjZ-59$exWK$J@%HbN5#4H{a(<2sGJZit2c5&6C+7ja+zeUmUtyAUwzY}5jG
zqC|C?6~eGg79uc_1t+4)9RNF~HFozw<zvhe9DbOm>;4H5V$^*K$C2X%(x#DNKlp_k
z<_A!L>-ceprsY>c%)gh4>D-A1+;suQ9suL!4x`~+$THXuH>QdDXkkscvAV3V?*44~
zXD=vGog7*6_wV1^@(UVQ23bYYYn>IV-lT`%iSWkypgM)d3qhtdr(?1N#RES?3`G6s
zV?9+^0dnR7?W8!O(rrQI1pR?spnxkiClO=|tg)sn`IgO!AWJPF%=kFJ;uQ)}c`54S
zgNm5*ISh_2uyNNd1YfyPuQV4Q-^bi;V3%OKz8%WbXt7`6qN^kCyA7p_sxt*u_X=_q
zAi#ewQ!uW#YR#*3P<{inqJD`8L~5;n9Rb;KT3wzkN)f)|1uY!fZ@(oU{Ov3;7F0RY
zpOm|x8|#!xI)qa4lcoG?gH*o!1Df_oE!)(ejQy!TN=r;@qauZsfzW^&74ID?XIR$+
zctFy#7h%iLJTw;6qOR$UM%|G!if5`RYyCdGr&xp(@qeL_KkAVQG@~GvfkoG_<ytJE
z!x_V1In>f}_Adfvs7QkzON)6<goT7j@*H+HezEC8V}giaoKr*&5cD-T5UMQ*Io2_L
zJ^axc>++++jS3=&&{vapU!x=Va|bzgFVv`z>p1q~QTQ&zIkq2d5d)%YwD4zOLKwk3
zj_@j8;*BH1yr)Ge7wWdF5*q<F)1E3$H~h{roXu>b76lZ8bzX0<Hx6hVJ*w?O*k8%2
zkU+%3?L1_v&<>G*05DmPm_>nPl}Cu##RAp#m)lX#6)Gm^Lyw#dycP3ouWc_<rH#kI
zeyFs=bT7yMaBSv7#W_t5m{H-5D$vgQPv+3^1l7U<C!J@FQx9PP^Dl>i1$H$J(8q{X
zaUczcW<`19o*PQcCojS#*?d+xlEC^JGzTHK3D58X2Ask9@~Rl&8n8N*glAyD&@@9w
zr~~ajeHq&q$Ylx!wk&_~>gt7y5JA_&B2PpaM4O||Dc~7QpwcGZlr<tw)W6jaB;cAi
zDAofl;LCYL^D8M0x~Uup&@?iXqdH7Ixy!U0m=LXRPX1WSpm<J#?sFV{w11UI9~w>?
z0fuF};`iY&B1Pvc3?@w#snKj&l*)at0G}-fNKd^Drfu^YLZ4Xm=3Z=ku@oNO5QviA
z@NUPDa@2JSg`mIJG7+^nBtRi;o!V3<m|$Xz(BXzgczz=&r!j>+daZJQqlYF&osj=c
ze=PmstD09+pv=mN8kUr_OK@`o@Xi(HDgc3?4Ej9o^TYz5CJnN^%P=m;gN{oIeXY9y
zR5QQr7?EaHF(Su3(6OBVmWdKi09EBDePuUCA<6C|MFUVFgbm9$6sW{k%BV$!cPFd{
zU;ID><v9av9plIhq;E_>(cz1(W5oV8CYb0@$yuRi1Cs_90eY?n90#e2vmj*jfdF~2
zI6@4xBygH&(;5^iV#h2T&`xJ}Kp@9%HpKwr-@_`H(`w)`Z!%wHRn_pZ0&-X8<;7V6
zL&gMT&1BDk8#FQD?5+$cM%8v;tyx9DtIq_19*+braS|_1jo~2-PR!;_%@;cCaOKJs
z_NOIywh4ElWxNMHmgca7@&1Cqz#S$U@&u>;jopoBS3q137i|4j`?471Yw%D`^=JSV
z&%|on*?r?EOM!z8vl~QYOhjv-`S9p>)&&a%p$p7F24kqqu?*v5v{&N!)=uO97<=z{
zs^9;ATqm+BJ85vTw^Aw%PDxf~_D)nNp%M)`N=3>hl2P`Ij6!JGMHz_(nWbSRl<|FB
z=bYy;dc8lt-#@S0dDZE8uIF_<uE)GT?ho5NAe*b)_AHyPgl+eA7W0Ok1LnXgV5GYn
z6lz12I?@N?Gp-4+F89DXqC6WUP?wg@6hU2;)9H;kbT^{@0Mx{%{}%RWBTDR*r0h{e
zzaV-6-pHq{N*t@Ypyv^MNF_J~`W^TK<R?|2#LPZ>m}3<y>~<f&OK3Fyg$P>Dj1E`v
z@xef;59pG`1=9vs#cM$lI9Me~Vft57&=z}7Xvst<PIc3#Tgc+{9!~DWo!DZ`nMHr9
zqk}_b(Vo+f%nu_R4d^28G^0&(GJ{SD?})4U`W*%I;%`qv*XjiH=*H$h8sW?-2he0U
z^0S~Ld{vHjBPqA%2v;w+KwxD4mc?DAgR?;-Z_?Vx0Yxe5dr<O4{IZo^e^1V4UTrd2
zRO;j`AV5P?S88&MI+r|T!V$iLso1>*OHUQ(LLjvjSA?-okVL{qcq4Q>qkh!7wVj^B
zH>~!lsj0<>r7ZGp1J#JoFB?9)t`w?%`lax0QZrSv6MT^`r8p_~5HEq=BbUQ<{(CIe
z^LU+RT#ed~8s8sx0`9uCi$TEr@LqxLF3L0F_~9Ai_j=aM7p{Mwr=L>}ZW9*(<PK!I
zz5vpFFa&AtXPPkgVZ;h#bY?J~3A-~3&xf)oQ5qsZnB@?6puW)aYb*g0S&#0~i0k99
z#W*01k(o`6X8!Xjc%2O>+N)T5vLvYJH}LEDLT=dL*h+3}+x<*nF`EWU+tc-}fr8^r
zitc@Ji|31bRp;A7pA8Qnad`|d&-FHFgWlOE-(|Y5V}NZ$nhrdXt1GH#I~{5pB7tYX
zy&i_Uy&q1`Wp4_lK@7VG*gr(u5%&y-qFR82BxiLvW<Npf5L_`;K9rXu+Cb!I^s+en
zw>mentzEYd-|_{Tj_`(yh{_5Y(Dc84yB#=xxeK=??g1?-kE6SpmNGN_7b!$5<BBwD
ze5wpboW}4ruR_EbNJ}fAlEh~PCZy1>3})IPS`u*uRxa!$|KkAljj7RM7T@M02cYRG
z<=N;XY)LvV;rV_Z<1Lo<I5gYwJAm~&<U62wiU)uQsd-}u{<X$bcSOe+hhQntdY$Zd
z4!FT=!1I*Jura_=2`T)bjSOfV+q*arZXHpb$upmXKu<70=LDR^@Y9reI6#UVQSU(K
zc{=12ObiEjMHeN3n1H^t+BecdJ*J6PSo=(VkOTv~BqAaJuzGas*}2dnbLfKfd_}nE
z7(_1+!4*PO$9g`zP&`My4676_F<}}kOb~+LV7#Uhp2W&9xv2%U0cZvG^gM!6>GOZx
zGV03kL;2-Uzh7V2;dx%&7Wr1sS8NNU{)3b_u;fg8ANv!O=@5}2qJ26E4htogNd0m;
z3oH50mqqL%IulVNHZcDlhT=j;TJvBS(}nKMrk?<j(2e8|NI4YlGggp)u0tx0r#}5M
zfeSJi>YBWfQH}^uhf1}O$WQYQ)BHa=iNTFT^<0Q@OQ*)w5QF1)08UPt-}>$$ywR>U
zbma1GU=dH=4p6lFplwlK2zYpIF$Vn7=rB3c;WceOdYmzOTn-77@w);0z`?tT+Cdvv
zAN`Lvo7+DEo*AbmVBW|&{tAIFQGw?$Ag)ZVi;~N86P6%ZmU!AU)%o5@Jy)=os<4XL
zJVK}u-A<5YU*xd}L$@Y)eom7jSO0b2q1_K9{Vd$tgdl4hA<pk{dH`%-Y*~dGi5y88
zB}9%Dn8wR(0G=W~RALsGjB<ot@Et&duE+5vV{RX<Z_t6~d1C@)LLnrBFicf}CU`mD
zuXi@;+x~fNSrHO_zF2Z}aBcqd@Ap+uE#jYaAbpRR0m;s~KPzh@#&I<@&ZI6bV6-#P
zfYRUL`e=F723(STa0ykX_{aoqc{&D&Sc7BWwRHmw<5o}lT<nD1t-x5)Xb9j9NO~3v
zLn{!U`;?@*1IdNh0b8qF*vpi~NCYfG1vJ`(p85(%qY6-etH7*f`5t>acdK*`5#iXw
zXN%<~izZdq^lEMYe1)l@cke$YiTIsCP^3%_aEV)wzYAo74|%AD=k`pUoMb9hA#9G0
zpeH)8ZwIJFH?+TMWx4^?K^24tkTy<;3Z9qMEq}`_7_jdykRdsEAHF3ooTPQPXpE!y
zU(YPAlGdsCM*&b|>NrdTI;jF6Quo2RmJ5INjI?Xk8pv<)4R>Tbei^1KnxM=9+(|tJ
z4T!Z7Cix~<y)$c%oSExMBo1)h-eRvO$oJ*lg=R>hP3fjWV&m}GvfMR`b=alyt<39A
zzuq$p-AsvnJc!f~N)+#Rl`)c3HQ@{j4W;eBv0lH9cm_*0WaZ`MO>IHIA6G2-1>u|#
zu46o<84@hpmMvc$Xf(#~PcB2$jfjY0NBdr#l0{Ur$YHk;(NjbV>Is5-xT{@KmszP3
zp^AD!xXAu5x^Kh_TySC>UBFbNS}_O*f{2!^>VhfV{Bu_lvJeOsi+Y|kTS~uW6<0!D
z10`gYzur5BPP<wg1ZVQ`Nxk{j2=*V&W(|ot>{~1F+8Ab;^8m;xE$8?utXW|0O1zUL
z)l7g|F1pAb5i~+<d7{lWi~LPFNw|Yox0R1j#U1_XwgvSE<=N*j{t5+)rD1~qqzmIj
zN#p{FiHY)#R*=Mk!!8iD;3Oq(CM6)f<2A13I$v^|ZzG(i4VbLdu>fcfyDjHvc(XTm
z4$OcZIH7_L%B;L79X4<kN}>`-hK7C3$yz}cC3XeYZkXG4_;({}oe0j0Dra16_w`2Y
zx$PEX(fusqP<IgYn{Yk{F5Rx&I%x?v-Z4_7#LEP#;khP9<b10pCj}Y3t0-!$S4l+H
z(*;r&Z%^<0Hvde>THr!JIdq95T)BL<s+Nz2)-^IWGNBY?vaWD_$UG_!;3Hw%4vzm}
z$hf6$K`l8bD=#2cF()3FQ&27BjK(%Vttk4BA@KO3tkG+-uv?-P*PY;4;BLvuYy<^~
zsIA0R!6K20nV@W1E~t{P$o#jXb!9?X4i~7TRYJP6bZmO9)1wUl*j6Xk^m`Ifg#2M^
zJZ8a>py-U^YVPYH?o`y2<D;fvunQmqQ7mE<?r-YC<I16Hfd>^k!3R@Z3H8Z$P+gY2
zt(RcKimGpjk_dGGBYzMu1yv}m2lafZt0OI>NQw^fgIP4&4Aka}XNoitnCTx*alE(_
z0JWNp2x>jxcWC{5AtPOI#i`FM-H>IN+;IxJ9`8YIhs1IT8QmD8d<L`axY!FLYG~JQ
z0r~DTF?Au*yPZbfn^PUveBN?mEr;qS{O<1$o|ncv$jjRT4IA7LVh5mXc35Fx<U%Kr
zKx9BoT)iThClws#Quw%3eJq$J4nP*tXu8i(@Iec1_M6vq_K1KV>JZJ0d46;5UB%5h
z5o0Av^>|L!S%rJ~eOxRZrDmPL59C-MQV2kTGItnl@&<GJ8IGe`(TVa?62t+{+XxMF
zx>2n+mf;T~Y=W6SC(}-OI8whnr!laf(DyD!9idRVSAP{c^;2YWeIV8BJs5*sD)>}?
zmo`ng6LgC|Fz$$;2IuTn(B75^MG_QUPT~$<AZ>FS)e}H)E&2n7AdG1jB~%<qW>v$)
zjE57_+)p*xe9LFK5V)$AZ`&A9WG3jiLDdlvB_rOa^0~+Tv-}Z)H53MP_4L@GilUha
z{=L!R*o(hwE{pHOEUm2(L@{pT`(<%KQ-o$Pxu}^_>n0GdRUu~`{vc@%Pk`BZOt$fX
zF2tU=%g_svr1ukmQNv&c>~=kIr<@+~J90B-c-4%1NXgecVE0Ao{AhOQ*17?zKC3z`
z6f2@&J3jXk)ukj-HWXNie?hZqprfwVa|t_EWs!yPlhWO51$l}4tUo{VL44GRqEagm
zSYls5p_F-X)pOoY;Gg4QxRV2z&)un4C{ZW>^BQH$Kkm5(Vl(wlM2ztTfTR%fpuEr;
zF!rA|76K6M|LW<hP9cCo=mIE%YT|jw=cB$<9!)87;vKRQUf>t>Hdv~u<AFqAZHS0=
zy;u$e@)a5|R8dGsDB=G&PFZ1UM5>QQ(SToXE19z$+ik2W;=)E6C-l3+3mAdNT2F~1
zQ-HE?{HyeKV20%%iNxO_F?9e5QNlAJdR~-nyAz*?#TGr0$kxB#=)x$dxI7=uF}U)P
zdAFZL*7eWNuU*EzHFtK@t*#C%uMQIE1`DGhGyugF1ZL$?7ss&Dz!44GG5AT%@&!_-
zAu36IYf#%fxR8}_<hn2Rs?T8psA5O?V5>omLHFK-UI(DH56yhgh^=BoA4EIGDf%5P
z>b=?a@$^Bh>i`6+L=3a2<Se>SS<ZhtaXEuvS((X`b7#bB04{$Ys;WGyqF71-KZ_PU
z%bOObN{5&%pvV6f_(gr3?|_V?3ka#N1KDwws!!-#<!15CslG}gw}5t~13sysECBk$
z9dtpCK|ib<0x(nqy0!l7hn%$Vjbvw|+8_4+<#8W8ZGm|w0_JQz(f$GeCAW(_pk3((
z_*eZ?PaW!vTuC@FYzG#~Qu$jZ)ZGKwr73<WVwK8Op#FjeK&`yh`lf)J)x&#4>4TlB
zV|k6bM-W#UBtUuzH@(FAaU@U72zg>fr(8KEk_oP|A0S{^w|1M;RC@wgd(@?d5Hs@s
zhc>k(kP16ct*9&sU2&)n!vc7W@_TG?7y)II<pSA`Blx=vKH@{~J7%_KJ?@J-8D<0;
z3{6@Au8|U*Izv(oh%*Ovg#XipMDAPVkj$x_eHuklz^DTbb*OyiOu_h9X^17<sB8(y
zA}`;brfNhX^{Mqfnvt_$rw9rRSS88t91syy3PqR)vj|m_6Pg98khojr^7uBvjSg9&
zN;J6!3}AzkY7O-{e(IfIY<+7&M5+)r2d1i-1H>%8WClX^o$4AIH){9JAAUsodE&q-
z)lUIxycKsGCCv&X;fMkTZrPI2j5@twW74AEEm*3k77;N_!Q>IW3M&}+!~0w}T87bJ
zu`kvlx)E525b?I2gd2=fhyfqR;ed@8=~+4#(SsY>)#vYls5lX{vNUS0)qjO@<Urgl
zp==j#yx`~IZWGQh35xvM20<4%>S3mCQ0qk%(E<O%Bq{^E&I$DNo9%lSJYAXyQ%$y+
z<{b-qcnqosS*_j<n5n~b7=%F3uY)M2^N)eshK69|z>*A4;?qfT1z{XG4VtgR=b{)`
z%_b(nfR70LoAwLT=oZL)nfMDj{1XT0KQ4!Gm>m@ONr$_D@TmmQyJ7D)GxFtZ4h4aV
z3N9q&7%;i;BXnL}>vSEQ-$ZegWav4)eXbmW>CiY=I`A{O0e_(p8Wxr)3@$)%cWc?q
znO^B*WNdEW2Dmvq?aOtkf&w3Oj*-GItO9UMD4kl~!PEs5gb<h+xaAXMAtaOmoMwoq
zm{{tY3lyl7ffzW4;%8nh)HKKx-i9_i^)6J~hI)S#*7<@x7&c9aN${q#n$m|pQgR0(
ze#0*HB(f%NL-qA1yMyKt*e}7+2{|KvbjTFOrlmso5EQ-cKawr_w+JK0ntRw%L7Q(w
zseXS*7L>ZKl4T}Yp6V3XWPc}r;{}4jNrW~jyc;AUqkGZd3*c(85=~6|XQyF1(BUBf
za_aKMRToiP*khlPVh7BIqcbGzr(FreY^cU0uIXy?0a3&MQx`!&lUbZf0576k)n`La
zEuvd0yAeXp?duc3kkAF1?~Pkc8v~%P?cf9Z#p}yqjtE2ZrE<+}A34Ff*3A5_n=1E!
zCmi#yFhOF{!^=-*LPZTVHR)djMR@~NA8bF85DiOk9pLNrj}hXk=};vEah%)y{@zAv
zZHID#pZUH?>W)GE=wg66IQ&Ni<mKzBmX`YA-8m9pJF<pRagb_31;3Po$@Oowe`=8~
zGLXkZ8Q1py3qc(&9O`gIsSboa8u<$9d|i##dl;UTZX<t+`hw6KM6wm2&eRWqGvor?
zG~ADGjV!vpOK|Un153(ep{l03^7%yHYAi9a^)b1ALx=$oQC-MQ_7H*{Ga}+g$m0pf
zn3u|(fuODnH8end)bMw3AN3SPU5Py_+E8^n5mDwM7{Yy&9*|DBuvjgX7f|_lP)j)k
zl%gm#55)NTP>lC@FYLCBEQK5(DjvDIHtZ<}h&@esQji45fI9gf2(VuG!BxXf+Fp9@
z@j<)9$LJL&_8f1|6H7$IhTCBXo!CN4NAzY@)!MHe6%?$*fSMEDfwE)1C#?((gK)*!
zSQMw6rv{A5YO^~VN&8AX!3yZE74A+#kPoF41H6thLIb)V-1nYSzPVI|ZmZLYNO{2%
z?)zp7Jq@SULDAa<POJjl>AO!~@Gn}46gR@5n!B~xS#=nL?L_k`5i>CY%1}$ONEL7)
zZMs32n@#ip(F=O%0^lTe9Cg7a(|{w(^w8aa`t;`O1XtWJWEbAguY~Of@&pIL8_ElX
zfndOh{7wW*OiaQm1Zm-NSn?$93pi0+(;75{P0VPV2BMI!5lD5t+1S{KhMx&c<^@wj
zEkc|uE%bh_M}srgyS<B-P1?N3(Wy=#Ku{qt$>TnX$ofV&y2zLe@al+ZZ19$HSx1Bd
zw=IlAB11TKB4+K9R27}M!=_#|@&|$@6Cw%;UP`}}ucqb;3=Hz&XmoTvl|j+DoF$;r
zygO{@*r9O1%q{JY8g3ytK4p0U=}VNM&K7Z{`=5?CTzQm22CbtlCRlbO5>!>tH&G16
zual1~>D$)tw9ul1Si2#~5bXh8+FGFVK=c^Bn!CNJ_Y~7MR5RnS*BYuLsCy|PTx;ep
z08t$IgXOT3qGq3IXYvWqA%Lic<z<e;ok5))SUH=jSmu8qI1zhjAAuQ2ERBp_efAln
zh&A=brtF35WpLfm0XHf90Fo{UInj+6(0F#s^ZZ04dVnD4H?tr4-|J2@w}xDF4%DXi
zjhq?Hf#`n%jk_@NSjLRKBekN(*hAtl!LV6>PzNWFM~k{GI!Mf1Bf-Bz8nA?B`?>;f
zQw>mL!Bk`sfeS%_ZLFI(s#S;(OYVzRxLCUR&-b>6-QeTCs%o_T$FsoJ#GoIEnrcwI
zm^yhLx|LAxOIW6(4JKX0SxETTOAu*@(Kv*&MU$|`o&dsGF$=PLpru#JV_*GOU;twm
zjk4+hJP55pQ5Cf1+fk!)oeCr=tKa>i9Wj_+z3^Z}ohrznp%l*rDK)46`3r<FrYXcv
z1x$MhZ&3>uB3n>z-Ad|EL+^Mg<a5a~;AG$>8En+}03quzq$UvN`uYDAZ(pOrWlvzm
z)JIgkqMri~CNJFrh4Q+lrho1#mh_&#I6~Ed)Gfw7h%0r12YwT*Xi|~`^o)StzcwuA
ziab^JXM8Rl*)D@dQe*$*KRz74Uyd@Mh?(k@tTl}J0tche#jTKI!V3o&Dg=by2v8&=
zN%B+5kH99E2>!=|=KaQln>ZC-4_h>wS*{LE$A$*ww_EFfy5VM3;xaNMK`%$3COQOz
zLs>LLN*fv)il>Cp{ClBbu_%i;|LP$*F3=^7*Maz~S(15`<aNLUkT7pUUbLva93`s#
zP=7L11<qQNpQ|hbuHqLf9L@a22ku54lp-kqs5n`pjk>|wm_-*gGrl`u4f~7KDsn0(
zV}YHb(rdcas{)YVC_~hz3~;32rcN}cyxAO$rJjP~YDTEs0dfgu1l8UZ24lX@KM7Gh
zX-tf;R`hBP;RDwpenz8r&`7*ZP{P5b>50w%HeTld4)ueeI&y`6u4%eXCu&p&s9NUK
z?vPiAIp!V&eFl$;Qc21U;Q8Qa1Y8Ubo`37<qyo2;t%s7VyyH|)h&JkprL!pOn!h2W
zEWoU~HAFPXz$Qv-eA;FL_-;Uu_giUx{!3a^Tqm+!Q>!GfayRK;wQ(C4I$0oE2?a>~
z>AwRw%uY{eLp)l>z=&yF>N^>T0(|h=h{~{&0yt`%;CQA!3)3W<lo2Rdna_L=OQ3!m
zK8PM8E?K$^jhNDt6AqYA<sJJsv*>4F4N$Q+kbjTvRfIV&sE%ios7rQwCc@EmU|4V+
zs;`to85Es*s$`9TLa@q3cYO!p0=0$T0=U95+AK803DBJfB?ck4h6X#WQMFYgJuOfZ
z^nKe0?$r4ksdE|!t(u}I-Rs6un!#rwjmfyTS@)kB2c8Y^c{c&vkhBp))?swWv&bn<
z1TQ{}5m5!9!O%#deuoh&1_if#shi5A(h00(0TA^hG3@~79LPChM5Ff_`~x#ji7W4_
zxN#r2GUT(N7bN0v+4S%9&I8h9)8&J1$k4d453R#Kgc}c<;vX%84pt;tC^T48bBNRn
zi!LK^iX9bgCd>G?@u7qj^v#my^;yxW{ZmAHx1ZSas}RbZTjlCN`Eg{!BqfqMK<5i!
z2j~=VR5VbEtE3C=D|~vHz6xM9W0{hu=raWkM(1z#b5P?p{QLQ32fFoS=!NKeI@Nz5
zaVfS$gvtW2)4_*-rp<(u?vSVwnf8Ko?Kys*pTSni8W8CuHWI`WxcN1NYP^tCa^Lz)
zo$7i)s|uQdi^kgMmm=~RG@q^QgEv)qV~P(F^{5Jnrla)dp%DsTZSLnrY#|SKuXaQ(
z_mhjyII3gH)kN-cqDJF-YpDC)g2H_jsz*HOWT!sVcE2@*KcwMWBOsDT3RE%DbnHII
zMbQQy961R^*k%Wq7j+$>38M5UwcMTn@X4myB-9vte98e}z(v2#nSh8o<VOpD2^e*h
z0nJX`jdNG1aaxKeX&f;NkeH%eyWWisn62#sH77$aQHg&Xl{6D;a5oW#040_jdJ(Z4
za3Yz%@|7f?LoUdY^xyM>8;B(wBo=y*xnUS|-S;x)g0~$wavnwz<<$G#A`(~<dFNXP
zAW_{UXPtehyH8Rq==Pr`LMz_B0N&ktcEZX1*%W~6UJ)Jg_P(=$?KL~l2Y5hn{*B-o
zj(A0jB%kY9LkTDJZN*!$gJR%}N8L4RQQ5z~7D@>pj{)F=ElqSm$q)UgFcA`25wSDB
zXZ{X)rGH0LbHSZ6IKb&Zt|)RLl+qz_Zg%{GL>$VZqk%;IC+39@1*c70sV1D0SU$d6
z{0vkPU7pf$a}vGTsW+IcIfypgs-FZrfW{Oxf(LD>5(+o~J8~Qim9Nmc?@87+2ce+P
z4JA(l;Rk@T`2xNqs$cc5f$}Tj%R%6)FuRSLF#ic=K9Z_?;EX3Rkz7&8&{lUm6Z|K!
zd_1unkX4*0oeN1k4~e7B<aRmH<FL6tPs-8+bh;M?5GJ>F45X;eym6xLE$cgTyuJHQ
z1g#v4^+GhIj3UvjAS+AvQ|~b8gM)r2j8+!&!f8Rbx5Jie71X_$i3LTS4JQdsFsf^S
z5=VxmK0mlgp&m4;c*7y=-B2sUI|z*r|NT)I0Guinp!inTe0{c{Ak4s;VowCJw=K}H
z_3Mm(!iFqJXZP|Apn2I2J7UJLA6AA+b=I*$%g?lFMx8+Ydjnb&d(EL}Q?%)5ee}px
z`MG>d`8GPh+oRr6dd>v~1s^Cz71YRJ7MCZ0Jc+Co$9MwbZbS8htH~)njIe_F*N^wx
zP;>HTWhm^}40lza1o+%WE^>KHR-w_tXzEVt^_tW^sB6`^o~9Co(*YbQE3VMUg*I0R
zg&ZRS&t_qCVY0Uzppwvt@Gmfpv|z6W1|eJYa1i!)2RdOyyZ_gw<YaR+j&wUf{yeDq
zs5uZ6${O!NNumwcyhe~x;bY(#?E;1If3KYz!3bfTzz(icP`7>qxyOvE2oh@_vH%Z|
z1#olZ+K(^Bp)>*F5$;_ABdXRT21l@GE8A8`@=(uw@rTY>92)yk&oTjZN10OCC4`5U
zz|MuZ1_k*_L*YU&9~;`V4<Qc)fyZr<bx1gl|Jw3E6&z6IS)mrk)ED&YPwPOHV6%t5
zFE-Z{MpnZ^-Os1-V?VQaRU{!#Q|-)nSnmgE2LZXi2xEvk3b-;IsKi=CKk(DiuLnE+
z?}JB5l9de%6{&Gnk^*KsM8yy4{Qqj$>j@yS3PGPJnv%;j4F+dFpjt+gS5wQa|5<4N
zM^N-j%b{CyJtzq+%@4qrY7EX{ugtOi*gg=YV7m*I&87c$kO&(dXG{kaPeml6h+aOp
zN1FimW@ATj)u2qc8Dv8CA(!`m9S&v`fa8-0y=c;CqGtmCtu=Duh(bbAB?6-&(BXuq
z_}Sdyngfw7R6?X(FE3z1v~Gx5p?M3+Du9)@HA_D06~y#1T7I)>!Ew^8N4+I>Bux%D
z062P&@FV|6)el~fK++J<Qm<`b1gSlQf<v*+vEJG=b&{n8{M0Hw&xu2*XAAvJ*`Jnd
zRM0kfobi~Xk_b-$*dg&Q(_oJ4@4tjBOlTtjG{3-<AcrETL|5KZz9H_&Gs+oBa82uk
zrjDqzBDhMMpqYxlNr#-=f9D1Y{k_w%1?YL!ltaXxSD!)KqOJ^+=W3D*OVd7NwgBKA
z)9bwp3@I+4bGaSb2GVRu)VAQWg_-uEDGQLMOY$??Mry8L72SnwM`5x!KO6?i@_{<#
zQqJ_F`W@9$LsO#-h!_rWr}PG}tD&;QD!@A`FR~co7O44U_!+az3#dsVkY=w{1dzv~
zlsJ$@#Wl8WKJ=Lx%1s-hD5w;!^Uy|F30|4S1Oo-SABMnA!Fx1SE<xMq1povL1BZz)
zViSjS<=H79G0keHdqo}G2#9lJR$^y<CF>V}3;+D9bL59j@$+x1&TU-HwTECYijk5#
zv??V{oa^@16Q+o_D}O(bZrdPkZxY_RgN(P1V@2UpS5}%Yfye}7{z|6ky}=2LFuFbE
z9C=}2Ktyp;jcAx5Nxf2o>}dLS7pM~Ti{}B;ET$+V?a}n^pSx*m(g{<6{D-@%lKI`B
zh6IBnV|u@knIVWT5lra&TZa9Yz<Qxjb^FXjH-a+GtU3VDm@a57@q>fsp-LNpjZrXw
zMU7V=c$b##=q7*@0a8oAPgN-usU+$pgz%)4@d5-By&`2oR@?Y^pWHcFEJjLm+M*7k
za5tq@z>0e?4lpx_aB}qT?(@I>dsdT&5~3%;$Y0Oj4`UGrQ@Hq_ZDo_1Q((Jfq?flW
z_B%a<P!}hfni*DW4s8+>D2fH0mOa20PI8QGmEji+-Sz@^YpGEu0FQi8r4T&eu+bva
z(IWl0hux35d}Tpr(DEvc69tn)Db4)n*hCkN1?fpG2i~+%G(+>?*8ouCw|9BC#uy5b
z=UJq9fNaZ^(X)dM+e1V|AwXSs$`VBP2O1`IABJ0SMZdD6{@kQADqP$p=ar;7G(xc?
zC$^WO1R``rsVck#`iUa_GpO_hds0A?Zb)WnOyPweZ0SA@berO8SYop$IuqQ$9QWa_
zci_rO*vmCh569Lb4r=SeFsw<6YZOI@DA8{cAVc<<=Z+9sbD@HLv**c;S1Q!%6$Q*|
z{S5&*@7^f}H);c^xfPYY(YUjET4a<Uaq(wE{MFx8xs>;Wa_%FbXzzlf(USV6o7(b(
z<1IBcO8HH3ac5|l2vr<DHtK?gIwK5~pH}w{vc%&7e6n4J8LCoRq$~&cP%^>UkI_Y5
zk4Q;AwFpt8=<;yp3otG&D4I3;TMj>(Bn&+TCnpnr&b=LFclCn^BY!hdb01&_LQM~$
z5$=Ex7dLIoBdr{mXy8^>DWnGzo-zuGJU;{Ni~TaYLu^FIt*By;isoAA0#{8;HhA$O
zcEf?S$#>mM1q0S^T6il!9!=QLR1MPH><W8u`%$wl%n1((OCh(X5K9okDdN3|d(u%1
z3n08@3C_11u;#oqV$P(_R0Q<{t{a$Kh5Z1X^blaekkyDtbA=3-KeTXpa6B#_xALHV
zhx7p-OA9$(#2%)t%nR@x#E2QI9z>RMthX0%AmZY$Va5aEnX9kU(9r0HjsXAQ;9x(k
ze#((Rz`u(Vq4g#Kq>_Cn01W{Fl7paY2EOw#a>zb$Mxp*9M87UycMghlh-*<YFSv#5
zdZt;x1<emJ3#1{|a-z!DJ#fa+{KB5---vDpYE~(-5nJYx(aY#4I3X>eov3?T5m@Io
z-!>M+lq1C?#=WBrzFOGx#1DW5f6Wp_R=9!;I_0>ekl~j4zTm%#ACiM{D9{-U$s5G-
zG%4`F^=3~*Yz@mQYQRAFjFp9--`8V4+hDNB2tjTc=#fzh8q#R=T_CtB<^W!C0ZY>F
zDDO)45Dp-yXR{lF(j^M!<PFT{gp6l*VG=3{B6K^l7nyh4FH+CT15Z*8$Z)_+RhBf?
zAS}|SxGXT=46(t@{OFj;oC=^s$j6J}y-sY~0`Pqvwuyk*i2%gsP}{WS_`8j$r3guB
zOwj4MMmYU^0yJ`UX+vmM7~TdsnJyR_brY;hR$9ci|LG3vJO@Y0!F?&X1cY8DWZ9rc
zl?`!>p>qz~`Is7r(8@{;p?o;x$ku}G^MlDpMN?i8;o(PiJaneMEcy^=L4sBTw9w|-
zdbjHND3VekC=xJx(?Eng0nWi;5OK+rC{AkdqT@o!odWHv-tbxq78tu5w53X&p$D~m
zK_@CBAQ~cPTDg8l=xiH+lTiQX2kJLvQRn&PJN$DLxc}u}$w@o_52Fe)9Fa^YHH*;o
zudY6%BFP9^pp-3FFL*ahp;ML6#8qnkUlYs!yHlepThvOmY@708p<d{$I%rw{`Slg`
z-X-BkWeP+H*yn9RO<O=`&z9#NM5_A#&kvnzNdzYTFe1&)wNXu^0*po#mChKCI&fu+
zHrY_u_FoUt*^H}6b*}m9d?x@#6v@@Y8$~hN8DLkE=mn^ZpH}RT7p|YW@8l0e8+I%z
zG#CLa&poi475m~1DNi9-qDz$igPm;H|7#hr%fvKJG$-aT3|iBYnny8X)H~8g4M5&4
zgA;mxsULgvqKid2BqohcL|+c5PkDo8WaAJ&{2{J%tP7Apgat`FSYfd8zodkE{AM!F
zh|~wrYe7-g7}kFJ#ZRiko8k(qG0-~xe?EnFM41HpErEuY)JIrNT!Y6I^N9QqF@vsC
z@jXV?YM~)Ol!f!1R{-A#CSK6}Pce$CrUjOuwNSG{6Rz}ou`FfvzECV_TP*{to6NU@
z8oiD(4*A50>L!@a)oUaqmgILs#xea~%K_{kP(>X3V5@5lEi@>rnsiSN=((UXA+n-c
z)=tM7W)`E0X6S5v4du&e2z@(t&g>=)(Vr2pLlr193T3Bh!(3>H-vn-XL{47U|2TWd
zfQdOHwq(d>-8St47x@OH>x_CE`=s&2gm_u|j1RE>`(Cp%y<RXH+LFK}?n8DGshD5d
zv^{8U26CRFC4`bIrjh@Dl`|ZBmw$da7azFg32H@FQ|5u*b5v@D<|(C05=}>RK+|qC
z=^70Z1f+()5uTTH$kiv&o<8wv<iJB|3<W`eGJOHn9=**%0cx7O5d#e0PV95rNs$Qr
zWn4|SdN9g+El=X`T_CLv>T!Uf8Ac?+;o4f0J9+|@9K^%VjH-5`A-8SYAJ1zb@zB3X
zPFKKm0KjW@#Hbwr)D!d6&$DGWpvjcKKLa(nqai1#>0I5i<up}@Tf~N-gSrA<0<lN>
zLp3j9>Q%qAMbQhm9wv7;Wl*cGBu&S&3QhL>4)!MX3e>}HffJhojcZq%ZiI6Yp)^+k
zF{z$QJy86RhGnoH1os@2z+B?z5_IeZAN;&hacT(Ruj|bK1hWd9xpB3-e%?jcgrmSs
z#GVKqL~)N`*}+1ZF1*^oIV(68z{cbR9{`Qv>TOQusAK{K0B+QO=@Fdp*dF{SxzO6&
zNkNPngh$?EjV=s}8BJYhLJqm-dMSXPVghKP?nATqFrZ9&qj$*5I{~Lb*prM&swPWQ
z>#S0oaVY^0HMPa*ac}r`Ai_-(^Rc6T1APE53{}KNk{<Qs4MW3a@4?e1j;hv(s{_KO
zWJ9b07A=l+1P~B`E86Or9!^6XWK&i6+Cs^J4y5%k`~vQpk||x&IhSRIQIPNWqk*(^
zy{Cx&4bf_hWaH<`L*&H}Uor|NKNCuRenLU|$*)F)km(y?BBzm<pefsY!_TxR$OAeM
zQM#&r$G42=@okVQTP0(;4oZ1{4I2jMk~~ptsx#|bL4=#TDAk{i`+AN?%rX{H&M@uA
z%bz|3kfjC_ZC7z~*BuyZmp)=$NZVEbm|_p%*D(7SMOO9Al!yXBYDO}|U9i+FZ7Bp2
z6c!ZDf3SSk>2!c=UR85``JDg&B31zgGOuN{h~hxKlZD%7*4GPNczy(wLFkSB;DK=A
zhUXC_US{DZQ(AU)Og(dk6Gl{nQK)FpjF|EvM<y|^^%heM`Xs((8-zWEjOE7szWi3p
zwJ5omZ<j+}pD4<)RLn4;9HTs4aCaISya+^2@@0yM6tF3oh051S+{KROes!U^zy1aC
zTM-O5AdPKNG|F1-M;dp9sLXkeLgE#%q1g(hts@J~Tv3*o$%Rqr)O(zVhC50Kk)C<?
z)11dChaa*CP=7+{cTOYj`v<f)VquMH=+KRmSi2pPyxma1sUIKdAy*DOz)O?PF3<pU
zY?>K<(&r20I9#Rw_AAcHgUfFAF(`vtl6$ZM2dvJ6Kns$k`YJG*sRYc_iQ8?NRE0zS
z06^CH>6<Vw)S7K_^sa%wBn8bXCk*ME$}du&$rKpSkop~Gf35@i*6AHiV9XAK=FMqP
z_}inzD2o|Nh6_rapv!KeATqfIK$GjCPExYqej+!8YZaEWIdgGd&StLV()v9(S0{?&
zd7v)jOoe~cv_kiaoD-B$5m5<qi9;{uI58O^GK#3^)!Qyi3_X)FpGyN!-Zhc=7c`Ut
z^eHf^5U`hsc+Bw6Ibn(*mI8S|v=O4q5yuZf=#I*mh>S6%yhWOd{jJWWf;W_rt?Pz_
zN6oyd0A+iL%jf53DCsi1lD#tdK@}TdPq+Uz30rU)L)t{(y&QD;X)vx*H(k7m`lHwt
zIG$gAl<x`m=AUbM_@uog<j|md<f(rcc>Iw<fpPpUT=L9Qr>M~!$(BJ<<<4N1pINeG
zebX-+YITfaYk~SDs_T>ac8O1>zuYerxq8490ue}hhvwfNA$^5-4{(D<obNMKF@KX}
zHz&$)_h+YUBrgtb#xHmg<Rz+eZI3f}IN@@?B?l4!%FwkYe#R+;a#UO>pSwxK2s=5F
zY!H?(yu3IzeS71#SJ@Jezc+{6tErzAy;K{z*tOR$EdO8>;LrH*>cuDD2ep2%YM#xq
z`nU;`{R>K{{dOVwF{cl3Hr}y$;ibTrt}vTc@`-nSrAqeFK{3Dfe=^njS@7L6@ZC?`
z=?WcVv4flQu3OU2nDt8AkfGHCXp08H+!VQXflTB1=#qAi&sP2N7z|Z8nb%#r+NE7h
zxA=GE^>>_X`4@US&veex&+rb7Un--Y+TA(0#EWI(e0j&x8D7z6W%Py_DO|-r`ug`O
za-W?1i))J6DWBBvZMI>Yr{D%#@)`g3{16WczA#Q0e!eB4G@q)g(K_NxX$?J#w(g&C
z`PAw-_P0&*jeW=0mfQm6g^Rp{JJomGPl?5XTgvRM-*c91u50TJcdk?{aoge){i;OI
zPNnm{_H5assQ|C3%D>B{?l+YxR%;Q|$VbEmI5ml`S4PxO42hSYvh{~wt%qm6@r6bd
zxBkqod|bsM5#HFcumsSt!FhUYZI?Im+vL*ol{{y<>UVx~6&SjU>rS#7Dr+m|7!vjS
zHQX&@5~M88zJzTLtM1_~YfP>LUh<}A^}S@i=hhdNeVq-t(zW;8X*Q&M^T{oC$1^$L
z_*r${9i!R9s;WvW?_ZfBiQE5zhH)9)zUQ;4KjqIIIVZb$L)E$t&zOrJHhyd|h|M}z
z+%%e`KEmXbUvl)ABu7a5eIFVe1O5^X&Xb=Y8y#!TkVd5Z0qIW}@H)G3K7TcAwJN2}
zK65Nn5D&^cYuHdD%C&@+l|V!N58j$~*@{1!va&PIse{d_{i+TQA1~R{b7h(kpG~*c
zMf)c&ALOwS%ssR3m7({;4e)#L_)BI>f<J#*NBJT0pV8p<NNqe{v$E~r=>v|3t5U{X
z&psRF*HhV7KV-G;3%&o+E*h6h+X?IND_QBUFMYA3i@;Dt)R30n(BYNsx$X1^+06}M
zpPsb5be{TDJ=l;n$Z|x+WB+eXWyiZo2lY>>X<cU;7#H8NlC*FlfdkJY{W@#)m@*CV
z5~r$~T97Y?h`PJGJK=4tx&XEyqIh74m(Hq8sXPO(Kh=K|>?fWQPOKc;Klk?NQwl%v
z`Lzbsheyhy!lAX$qcJ5&-P=Z~To*4=mMm#{zU&8I@=gD+BcI0(?WO!nMRq(-eSq{b
zgFQ6FP7YCN>2R1QUkg4?0)VoD!@`8l3=A=!9<s1Ey+S#Y8BrZe&;5FRC&Pks51Nj#
z+jNAzm^kf~A$|sXCgoJ1w8ACf8L?!kR#EhrpTx5saMsg5w>uXRSVd)KLlR!biU`@g
zk{sk|SGBqDo{+NR%s$eh@Cmf4SD91mG!L^8;G)53O=+m(1YM^8u{i*0QwI_GS5(Hb
ze(KNdHN<Z|pKaUwv#Fi$&Cr7c|1|=K4tZV4pkJY~aDN^!ABc2SkMY<)+b!0;)n=cC
zn#SQInLmE%gd@K<*3_=I@o3P|*4uw3{J}<HLR>ZyNd|o9Nu1#G$@g6G3JiEwP8|3~
zA&~X1hQvzW>zD@*HkuSV-9^1CqU~w|RB`m;5|!S)A(h$nM9bQ@hi|=5&K_$W(34pp
zuQbFps~}ZsW%vAQheiF4qxG@9*<QN~uR1>eI%U!6Y|k8-(`oSh{8d%o8&01GTarA<
zLV6{Erp9@LN1Oc?90TZKKy2<4dJ%a-Gipl^BooxYBS<e^6B>Zspeq>!Q$u18*Sf~`
zj;!7Y*-c+_0`ucOFB14$xZYm=6>HXoaER1DdVT!;Tb&)JCpc^wwlgHyO<RmG{qkMe
zo_`!&Rsl>7JjKyvwOpMBx0fj8{O6bbV<?l^!Z6LG6<=S+U+(GYQH6Zt(&G4LUhIwz
zDZy0~k`*}|k$Hq;S(2))<7mIEVupeN^(@Idm~w;&$#x~EmL83Gm}2x>M0>xgv9X-d
zrLLvwan7@6WJLrjhGI7C=QmTK)n5xLhx9t}f&^J}QHvApel$2a@dqb66dnWDM`B)H
z^RE(U4TylKFLnR77d6-^5p`H^Ke4^`U2I2^T>jza{f~Ek?m*YP*=lO^e_ZsGDq^dZ
zqS*My@f&s7)fPq30(Z7;>l;_`^m23e+IU_q?Ymf!={K3X_o6gh_v(w5SjOKZ?=)=&
zoZ6D}UsJZ@7-+o_!KUDb4SVXNUG6$qw0M1OOq7MTEy6J-${hf?#qGt5%*IKYnjK63
z_A7`3u6)1}-LvAp*uj#e6S7iJ%`peAy3o2BVa+s|21PH+SNh%$X|E8pW~6eD{ctz4
zpj`T;_kQZLleLZx7ib;*>|H<hgA{Ur7Z?RMtLfJT=!TCu{=Pu--n|KlduxjE(F~NM
z^EQktnetC~*&!}u_lJ88V-8R7>RG4P1rb*%iWFOxzE)?^H}#&;2CuX7?q3`C%O2Bt
zQp;7VyGKOFl2=%4v|m-t<Y3+Tt6PItcWnLAu6CJx8Vp-O4}1UcYviS69oo4>PU3@9
z;Rk*I6K>3W53F6%FxfW{Fq0D&A5&S$rwzAJ6&KldUP$6>tYPo<tJj3=ny1-U{Z_%t
z5sUaS#Yx_M{fFhg97hxOT(%LDIlKD}aj9maLHYxStR)rneuN`#!M43C^Gc4Uh=|?a
zp<#U6^ho3i%4<)VT(aHnRl1`Y{=j>GC*ZHyIXDEAm6elD3^XMNpxVZ<b`2$}L`AK_
z?tSw5x7~8SH!NlN2ZZuXb3gsD`X*gJTi;*Ap-@CK{oLilMupXUFX+1j#_4~ueBkp?
z=$yRXFVH<`K%Bnt=+JZf$CG{crOOJSs4HAi@oHi^s3SIioZ*_rojQNIS=o}K*FztD
zStU1Jba|9=00bdWo9VP`t|7q)V}lt4#V;x<Y5?1!$>IfR@g9`6xAW^;jVU2~4_#v}
zlVEmdVBhPkUL)B}_WP^v&@9|Ie5d%(%gtrgp<;H@e9Ko49j&QmxVC(WCSJ<1q=)On
zF^`7V#=^q+D$6MQh1++@_7Tlv9eobiFC|oq41&V!ovnQ^0W3&cVRE0w?ZZk`r95OA
z=j`SCvgV!LPU~mo$%46K+`@FmM`fLjw#I(j9MgGNnlKyZwNc?Z%L|pR`;j!{t$L*B
z?8T7x_jx%{#cpyWM7qWEciVnh57(u2WgHz!@}=BA%l7x#?s9N%U=;|;6e#i1eHuei
zcRJ~y%Ui@1MO0{rI-IL_xj%sB%a?gy0kIGP%G39c%vwyiE?qi{dl>OOaJO<*-rb$b
zC&GKK7>OJAHGUk?nkt^m)qd1zQ?PE;w72J;GUqaL{_e}ePbaJ|kR1ZAfTf#{@<ssj
z!-n^TJF~iS?GJAL+MX|6rrjQ{w&J7v*7G&pxyfI|;TrS>N3-^RWM=w}X}~g8Fvv2}
zOT!4TMD_{jpb*-+)o^5VRP)X>L`$`(L0?@kkLvrR<iS1PW0F_!+HLAr&}?AR=RdBg
z$0_eiJxqPU2X=KQgYTSIcXjm)DeNdE476q+D^QVC+5^799ZA!j?Qfgy1r8^gxKHaG
z+i`+oiH{*m9HpTK0*())EuH>tX5UfzWSyet2g4sQ#G@Jt>C&5zyks4lZNsj>5?b+9
zqB4pCVxxuXYHeaCs;;~n7uuw-bSoxy)YB)FA0>{AY*dR>)KIUDGXAtVFX+WYfkMRd
zCD>^dqtnV#JPN19x(YCtVSpYp;s8dHAV_oRK)c*6)>ZLev3I?65>p3mI{TjA50-TZ
zzT%TA<ZB#W-4P*l&ojRG18aB4#EXfeQygT!4Rhlr9M@I9{3&DdcpAKs0kztcdiGV7
zJ#?i{G7p_3FWbQi<hs3VSu?DDuQ5WH1j9TS^&>};@&H<TMd;WLOl~^w5L~`8s#yd&
z?mT83K*!Gg<c-etVj1p?;-SZr5iAp5aKBjoTx*MH_GhW2A0|#zPZR$x9j9;IUOpZ4
zS72P_XTVbzuU+Qt=0_ymE5XL;b=@<Y_?_$hgpKU#mdn7IjNYgt2*H_z(m_vm0yG!K
zA0WJ^e1T?3EOxKm%!|XV>Ru!yUzEPQbnbwtpM(O#<xEGIx}*kq*1*?(my>PECBh}9
zz~$U=+_xO44k&4p<HdCzmF?@Z{hll5^{l*KDEAEUkca+R*|(9qCO(+c@Lk+b|A&71
z`GHGt7w`(Ke2?L#cFuH)&wn+^4}3zWWM>XhZs*>yd}c60NPXZt=|I{hG*lNZmH3*S
zYNd+~>faI+DLG{Pw$b5`MUqeKnI*e^*?s)N@{7K5>F_0Y)m=TF<<o)vOJeBVc{@EW
zp%3d~>5s@nN}l)20u_Hux`t6Ve(avlTe5{<wFMEQYokk)$E1j4#|RsldwTdnl^vN+
zXhtiHY_eZ7b8zFf$bv<-4u6Rt$Jo<zHcZ)ZFegmmgl)za=4>m?qxX#|!k)qiR?T5=
z<SEL=Vn#ALpWrdcqZV0apkAwGZzuHPh`P;G?L+AZ9QACl>w^D9TU+;B&~7F~PY-yR
zb}Ow_^FO)v9BbZclFB9GQT$LXE&6KD+{J=V-UgUY9L4=+C?&!F+<|A$nj3`&yx!V5
zO~~+~4O<w*V3)!D*3!tfi-Ho{-i*Y=$0Ya<NFF*gupPU;47Tt_xXA(Cv)BU3XSa{m
zPeWUEz}wu+@!FVb#G@}#8ye3=Epk>s9YOWtQC{Rr<3OaI#a!M<rHycg`cxWieTZQJ
zD0`HToH?ndr%V_84$YE5s&EsE!#iqUByg7~-rh}>{kd;nCYZk69c;wvNU>2>^b{NA
zjcgSDaUj%>=e|wL%3w(KO+7f4&8YC3aW8W`WzUUW>^WU}pYVA`H8Ny)8#w2yeUGo#
zy?Hd@80AqZYUsrE?wndADie}eM7cBTpnePpu@wKAeLep1lAe{9XRP^H{oz2miNn@h
zcwSbP>+~;{b`hlsy99aJS$wbkcHe%!ro&GAO>g-4@>fFg8*2N{a;?J_G$R3y_^jnx
zhpVu4?ws4t{YpV48s724^w|61b?972SZeR(|2FAN3OEp!^!eJ-eEH_J3pN1h7Y7t7
zJ_K{{((p3H9GGUXB~#)|V%jP^Pga_;7z_ME0LB24L>)Q^duv2uL!tdW`)YlyJtn}!
zD2{n%;70;uPP)*hnafe>G)FO0hpmo=7{?|m!x9oY*_4oQz=~aNL%~hK%EA><zu&M@
zzVjMX!+jBSi<XEEK?ht7v}oaBJYG1|OlkoD*i-Z^i@7{p#}0dD`4fogENf~%CjVf4
z^I)Ie{#T(7Hg-2|t*=UJXdY4dxmBU++6>ZCx0zY_d;`G05Eu@cI3`?9R>wbCo}Tfl
zepzaf>w4bw)8^)oP~E#ik{S0r>wyrE{E7_w&`lUl(>)5ZV?u-iUJcZG1R#^0b$JgY
z+iYPJ9y>2DZ|>GD>oCdFh<o(wN`iECzMO+`Km|MZiCOtarW%geo>9J*rqNlswk!x<
zP4)Qq$;Q+skH*BhN!QOl{rat6MO(G$Rz_?M<6-9AOz{FeP0dgcM$%v~Q&;vw+ZPuU
z5&?h#M(BjYM1{zhd<f@rp>u3a7F#ma0ZUWCOQjP=m{!{R6go~SM&ACVHx_+eZ_#D)
zHtcM*s_UH^JzJEQZC6(`Qo(Jc(Ov1q)LziEQH7Y7aVXg6^;1w5IR_3nu&`0?3vDU{
zJbTU^vR?%OPjWc`loDZ!?j5<l`TdcXw>Lm9Jac_zlQ_1SJk-nggkxEKSoPbED+m7C
zi!Qu;gcgS5?WfJ1ic()EK1kAR2<qy(ohHwiIPnB-qy*%Lj43zr^Y~bjH$|iC$iq2_
z_gCt|{}VD96OxlPTwQa&PJU}n6@cOJ5wNmW)9LM-u%B`+p!$6_JCrX9*DwxVS6$C0
z|J=(Zj&I{iQo;w*m1K7>ySsd}fiJWsr0ndZT)x*VgGuR>Yj)k|@i*l&7ccw{b^%{c
zmg4K3Dae-Ij>W@M#t>a7mTuQa;fAxw#v?CR!=U7tuT}Gc`6Xmm4DKdQiyRO-&ZbX)
z6^IQ=`mBP>TCY#G0~FC=CB*%vTgEbD7%nB@l3LxQT^M|g&q?<$SHiLu>A;W<-Bp?E
zkMC^A3(9GfEb+{bjo`pc{tw2$<woa!NAtjwQu4|nZ0H3N4A%VPX>D)gmFR0~YLGx0
z{#m!~I%dV1^X9DBFQLyIHmw>qQ&CY_%lfGjM!72NK*e{qnz*#=ab6#k7Zc8p@88p?
z9`#u(u4LB%y*@=jA-m6@+D-^=DwVRPoEvKoGR>28NgfdCmEfRa83MzgEP+o&bFDm-
z=sci$<js9491DM5Guj>*49oa%(X^O<O}Kh(O243F#)sEh-*P2?M%DNKam}5$&HP&N
z3!gtr*V2V|{R5x(*o^bU5MN_P+t!FaA0OX!il1}`T4y=R6zL`h10vgY3rdWyf+IW^
zESvW5ijM*3_~oAl%qz&63@?Zrmzp}Q`)mZfF)egb3otS=y7u{>bNN;Tg_jy=HuX+?
zoQK&PZ(~di<U1P&^nx0gYTWxIW2;x+8<wt97ryZQRHCKC6OTBM9|iFbeb?p3G{c)#
zcXgj1Tpu<w&1@k(oZlKFQhc^be`xL2R>jp>v!5G<We-0UB&Upy*nhx#=I?7FpTGk+
zU6rFq!{f{mpe8sf%Aws4^g#_Yt_zBbi<g_JjZyAB{vBpw-@bH{Dhu&d8})utlZSLZ
z*##w8{J1rA(C!+$Yx;Ch9nby%C+X3?{PfDr2lEsmWwau=#QQfK=uJDD;2COcfS*A7
zG~;-H8TP4~t~tXQNEdEE@#|?`B6O8L4rgtJF_8F(I5=^Vabr!P=6%ORmYd^)GH1C{
zUd4Z+{556(J@L+UQn7WwGJYzxro<zn?#mnQ$vrDl_nu+o{8@IO&(?nnXv(1Kw`?tE
zcGN19mQS-p5ge!b25E*C<T!~7IQLYu`TF_>-Q#a9b3NEQ_O(d`KpZBK_vEqImNmtV
z*rwrD9m7A(?5Cw$EGT~{@=)8sZzEf#jI5HD(f?;^_)_wAlrCMfr8=BLq}IYS`Qwwn
zGWWvO6Zz;)Dn>oA4TXeJGw9C8f+7#jY2}xF_)JTFyNMvspPr{Tr7(eFa^8vqsC&et
zN89Bu{>ii71m(UAn6x>0T<d8Lwi$#ls?XZ1#WyKqzs57Kwn2%A{XON@#tQ_-&fYLT
z|HCa}S!*_*a8<zBum=xb=P=}l_w~#wugd&!)bTe(KcsAWZAm`#y{6zZ1(r^20oCQL
ziN*EBTO@YvGKX5ZB$Te13WS>=c|#Hr2J0nb5q~HlU?vx1&RVej-84>Y28%MXL-a5c
zYeIK_K*8+iuzanL#n<2%t0t7VQkCQwLL;uJ)lQXZsF@swtA`Rs1#9Z}1yNq8vjOzO
z<Y-I06Z`=rKY~L-gm&-#=+-0GYAz5N8QBO)Z1vE}4cPk=#xMoZ&d9%p7cBZL<azUZ
zIO-H~%(aK4Ta{*yT-aR7(*X(nIsKldxYD0zH0<S_v#)!(+d-E~X)Aub)v@#PukoKR
zzyAylJ0Ai*`EE6y7L#Ys^u4g;288b5mWO6-JppO`K-TSN214^rc33hoF#%)JhFt5X
zxKyiHv2vxZt|nFTp5at6rO@JT9AE%(qrfFog25tNJY|3v?mO7_<nJSmv2dAy@R{j5
zlaBeRttJfXq;CfDFwY#JxYhOd_N}ByI+1{qeLC6eNaX@?)^`t&25epX0TGIEU`DSq
z0%vbR7i{FLnr>Dc{@2c?@x7;sHb954FOct}F-}A)6v%AXCX|@J+WX}Z7k7uC=AE}I
zGM4F@Wz#n1{!z1hnoqv>dgrrm9TvglK>=*A_h3{bKjja|iljW7j%hriI@;f&V29S~
zM>_mwUPq0SDlNt`>%@j1Kc-Rm<9ZN9%aJso4)<KZE4V^9Sw8*L;1Wd2Rh=b@Fc*g(
zkk4noj%1p)ge;|Z%A49bt$6#7*XJ1xJFQA*%pGVQ+@G?J@+$hVaIspBYSdtYL3u}E
zs@d1xq1>zC`b_<e)w8v;4zaPraf<J*v4C@4?ry!yn2zL}FM((?j+qsl^Pk<HcMh%#
zK0o-1TTUt3W$l)f$C4K3)=!HtR5g4T60p^NBSi7q{l+&@ev|bX$ga;&MKU#cL0SYs
zlX``d?<$NK*srnFYijAduxO=wB;&$Wr@L-g32qy|-P0GYZuzOfb^N20ZDWGQ<wq^?
zZF$jCt05y)m^s|Ezf%3-(4TnUJ?RulLiMIww=mU7d9scGZlLaZ=%iU<!uL?O(kt(d
ztNC~63-tF1F8XOO#pX~dI20@kt^T2Mxe!qc_zs`AoF0^ismTivBh(+2T{+KIAdQ>i
zk2&P)*Sr}zs>UN{gaXp7C?LJ{{~eH82uW=F^q&E#9f;Q6$$=)f1)>#q6$?n;w_P2p
z4u15d!ST!YXl#*lmdslXN-d9G8au~+-mR4ILb;G3Nw{$p<yk4H))IE5d*gy<FAhQ>
zbyD$(?<JOGoVT}^_JD(3*u!OhVsQ}4`Wn(W_XQT^1@zOwk!6nz$*FU@#ke#=&Gy@L
z%Q-EkMM@JggPokBA399?Y?q@<FtTh<+l)=I+p54ExV(Qe5B;KtPw-Pi(1e4{5y8=c
zi8EU@_SN5>YBiOO?N*v~UN7R$VsUAL05^@5FS?rsyd;#K5@?{`#W!?;RlR}!VZ`$A
z-O8DeCPM{|P?>D8cia@ez}p;!)n(;d7CH(kn#Aa-%elu+#Ah-!sg#OoKG$%3s2=@q
z=!w$@_jV14?X}F`6tP{>z6K`jlJ9NH2FM4TNzoIA|7GpO$&?Ive*l^tSD@9PbJs{w
zpxo}}q&c5d%f3bR`m2x^6VdG%?EQ;srLc)KT;!;87V3~IAO}XOL(-gB(Sn7$1E9pS
zifqNuD>4v>Kx|eGuAD8<Ew};%Q_a#c2Jy4enC%UtT&{9HTG@x=>V`iRGR4_e)Ct-(
zo3XD-3eU=MdTV!l!i2kA$?EO38G@JfE{O53y-@l||AEB=OH`qZRr#0)esl3d+g`VM
z^#~V5FZZ>3*HdIRt^mBK^dA!Cljuhk*^oL1+D2`liDx#usc&O$Y8n;W9O!^K_IO6l
zh?u};y2l0LEZ=gpy6%-NPY~8IS|rg>b>SdBPrPYY>0P{HkcVGLy2vNC`4;w+ON!t?
z_G=qmE<%sZ7X(gH2Q_u|;8?Gjw!~$CMG6D1gl4^0+odSyK%0h@9n+xf$O6R$<+!ar
z=N9G@LAem5m=o+i3VYB6QOw3)Q;1M8f%gK-Rd1EOd<;3-4mzDcBNPPWJhx+VpjRjt
z*ws(8>0_BGLit)OhCI>AxYS79_S@hS`Ou4VSDpHa9=@1qtF_;XVjqa?lvU%cp{TiT
zJ;_lga6o%uHjMbG)koYBxu^QI7+F|Um@zkaNEEwH@3Rj#w*p3B2V1HD|IrDVmA{r&
zF8S}|+5{&gBfpBA_xmNJyHi;%qUVZez@YHEC6tHmHG(_z`pNG+lreB--R?Q{XrGEo
z9gMm;V0Z5KcQN3U*Gs<jhyM-=7)c6tBX@O;qUepctJ)Q!6<mY1tT=1ow_fp|!zGl5
zK$*Y*uHo7DPL$gfv$FdX<%<l5kP@BfYB~*nP;O!kMchW<SC<F3MXVnBXKZ02PB}HB
zl>`>xC}boJ#l)Ioh1hKr;fkF&M*ZtKlcB&I8qpdQI;oVHkb$Kxznzk&X0GCe;7%X2
zIB9wVF1b6Ws<w7SPmgWWG(-jtZrkNF)={f5osxl(PmFS1$4rh0tl^Ca)x9Akc|huB
zElPPUR&&CEkryx9_7)UUD|)NbTPKpLDPhBv7|`kQXJ6T)h`&?U8feHzpdBRgCpRyz
z^yK7ZW3HV>K<|U4lm}nAh2`@dhaT|W5tME(HBY;K&g8@a|L&u^zOejecnV&@M&8vd
zhD*ATS3uZ>=lz0vq0|*WKD=@BD80$UK6<E(3$w3%zb~sKcT<_yvZsR`*EH7F_|vIf
z9bX>A&i#@9v}N9FJ<4IE<RPPz)J=f@$Gg*M#VSa-xVYqA{Oi`VA1FQdxVpL~fcQVX
zL8zlhbh!v*VQ@7P+Y{C|W%UNgZW3!s?ruYo{Cp!!tYz@$ixtT(8Oi_gdT8hr<YR<o
zJw{qw-=7fFHmGdoyQNk)05>th?9=8g;l65rmb*~vNS8!FKz1<>%fqg)-orrwlzshr
zDTr;8aL|+=Vr1zdwC@b?UXl-EHPx;2irTnQgWrsgmT~c79Jr5l%#^{c(v29EG<&PL
zY~}HTKl=M=4P<m4&DunUys4uU`1PlKft!f@D&Bf<KjKO<wlKjTq^u#zIlfihyVpBf
z{%~~gCrdZbe9+=?&p~Uy7)2^rA5guzKWF{g1!@>$SBJ~xN}qh;HD-6qxt!=(zMTr=
z!MuKbgLT_oy3X3awhjU~Rq1nS=~9iTt-B*Sp5unbU$<#EMeJ671@?ke<1a7W9N$Ix
zHK2!wf>nw9QIxa6spEx&lp)u3$h4mtX!}|1TRL!0&xHIsX*yVmU&7^2b#LRAk_@H>
zg6?98z71S?6ZcOK#`KOHN;mAUBQCvsh?xLpRl@d7ySv#dti1m~NFO_M2MyS$_Tc^!
zcou<%5xT1mLKB3eX*bLmuLd@pr73dySP3yroiN>dv34pG3JmHJvK!nA1uF+L(z?X3
zggGWfL?$|*oIm9;@%e=G-{osgr-Glm`P>p=9ZO3Uy*`$};M`HF6?LR*!NQMrlr%2L
zrs+po9Y=tcSp9qV?<c0DRL2)UT1ggQzq*@W<Y0N6FIbpn`dnbm48PcYLsjG3e|+s?
zjeo;`R?{yC&&cqru_~ZRQV8af_|>6t^ZluYGofBn-8(!LS3Kxw*!=J?<ZevoayO}-
zFQ{6>T?SI;<A&c+o#IMfIke3RNlR;?8wbN(vZi3@N-*?{_~bF(BwHjXxKYX+Yo9*g
zH4~ElW46?HOoQc{;_$;6J5gS!>=S3o-BAkP_u2=RM|>l@%13D$mhNJSo*pGUJi^>N
z^f^SjEaqhU>1#j)_F-i3#T~-+ZJ-Ql{pqFwo`&qMvS``4_#c=^p?ZQiW7Q#akfKa4
z@9s20>C?IhnFye3zsA4h1m@S=zznXZL+gkL<Z{`zE*2!1Kru62m8f~=ta*L%5%!&)
zQ?m)4)2AFVq@%xlhon!JU^c;uJ>R0PQg&17bQ`xiBjpLK!6;tou}gk~meEUhi+8s;
zx4HA4Ugl0p_{z%Qj#=N#^&HExBC6ktZ>ld^9y})C`S$3Ar%K#$im&-rvp`*A!*Jl=
zCF4q1iH?|dp4YhObhcpR)FDG4Q?+|vFcZ#`n*Q^Gf5*VuGM~S90J>)edqaZmm6s{T
zFQKeTXD7ODw{{-IQaa>Yg)_AN9lRcLuD1<mwRAhwT%Q?+Ke27o*2uk+Z}xDxE?!c}
zmSwNGOcr-t^CsTw9Yr+5flxU_rIw{)yjoGA0x~5J-uIsUHr$;Z6k$2wFmbl4xU6a%
zE`AVPe2VX@zu@BkMJt(+i(9r}%L(NSSQ&Bbu7d2rsH=P|xBnc}U8I*gS-&*)?yjky
zd6T#arV8E=6tUu2cXK(>u&2jL_{Lkey!_Mc)yDR#EEQ=@wKyoQW?T^j5b4q;)O;32
zLmrsJ=_Kcc3%&3Z^BuOLHIcAiNm0%eb9t$ZIa<L&x<7U{DrI)ZNh`12<}iaTr)(SX
zkFTnS$9lEM4m0)I`iGPVFNa!QtXBus^02Ghdcj-ZPg6bLr)(>{!a{ZV(6v>Q60x&z
zyE@yo2l>1UtM2H3Th`JSoY*P2ey8-QNuLtk`=@%}<@}x4El2RwZg^R0n(LZfK(*qx
z{rIT%2xJi6zGS`1LZHtE0Uh&md-b?`+S_+#RJr-yn31+^*H)rz845#@L7Y;7X04AS
zg=H7J^x{7~d9Qlp2;Xd~sGN%g)Fz29jb9_vgX->hs-lR_2b+9UQxy|Ea*Rh90_T-W
z)$-|v%=jOsykIl4tai|r;iRZJd$Q_;EEk)=+v;aNKGFt}n*<D#G{5z*7$mIZg_Ma?
zV2SrzEM@YNs%pjAKxwM47NcEYHyJFcmyF1<%n1U&il^_bP5X2KnG_;>!96^~GE?S`
znSoih<q6x*&uHx%dm0Q31bdr<%LmIM>Ta#C2oB5FE^yi{YIVV7-|rt1O%5D9JR5+M
zC*<cxGqu4_5$frND-u^qY%my5=UNpGxhlTuEQ@PG!bzzmf1OUgYJGH~ZRGOlA4U3q
zldabrK3aeM`@7+g?;lJg@9(!bC{{MG>)7XK*|smPw-|W({rY7odtt)Bz|tv&(-@MV
zuBW_0Cr8UdHShTCySX7h^)@G;SI(|6{gRa4v3Cyp;~q_1ku=p8J=?Kfh>dN`Tx(#e
zB*$aa`fFovS5&IGRBU2ygK3&fp?#BT(HY0oiVJ77QwDmME&W{DrO~15a%lR`(h;Z-
zXmpfDL|S?rKESv0x2eZ)#T}B0dmQ2Xr%uM};P3D9eYl)+XS}|)9n^kcGin7J`U*k;
zZZkTO<^Hgtw;FP0%+8Z)spzK6NP~qav)V9lZGorRoLl}~Xs4}q(2S5$a>~-+=m!@2
z^!sXglwM89hwi$yP9CNjdKurWaUT0>wPZ&r(9Y#an*Gjv88sFOo%@Hf??tG4-`;fp
zq4_y`fx4L&94EVdW0sG{e{VN9u6nh&{Ls|#Hdmb@H;a3112xWt8X>}!2_fa9($X^q
z-ww-^m6!=@|HL0mUUSrIIN-Owm8`k<az5Ug=~m&+=`lmKqK8L*z&wntcPfQ<KTW*T
zoINXYLE>{dU)RSmNJO&KR8JV{9Xyr)?xgmWv%5Y`dK@0@6a3m(@#sR#+5U%CuglxY
za%-IH@BY2Dv%U6Zq>JnH*sGFKg_x2LBF`rSvOQ$YH6r@!fag%SDfwhrVTW|!uhZkh
zExUNKq#M%nCrY3x&r8*|=g!FVc$J0CIs32M?LQb#yR0b7KDVYJe{%A>kcFMLE(fI`
z*4r|;@;inpWpEcD{H6~V!>|Cd(c9F&UtX~DS5nA$r?;iQUj7RD2Czt(4QYEGX5#*E
z`-}7ZEEZS(+VV|L<KD{6&ke+mui%1qC&ilSaWBaM4OgpeayR}TRp%MiWY)d?8Fh55
zj0KP;qf$nyfOG*BP$3klp+^J)geDyVfw6&f0RgFz-iZ(rdNPW12rUp3nn;s`A|(M5
z65d;9eAfH_zz0|?R!Hu1&OUoz*Y6teADsH@@G%FI1%*$$HW{}l5aX!6dXaX;)ZiJz
z00o7~78~?a$M{^8jTV3*eitC~HjI}Li6NP4isS1t*?B4(j%g%G4R#PTs<MdM9LOq;
zJg_jeWMsP^=TI^v$aZ+meicoLSP)@QAX}O1cx~ov#iBK6%4r5(nE{|xEA*K!@=h(y
z>~Q-q6Sp+EJAiZ+A$5it=!+EaiuIz^`tA@<rIN{7Zs>@`1$J4-?&M^?tF=TlxZ##j
zgokxK13uwi@<YW<<tq5}zYU|}`-TA%y!PAo4db7aPk@`bHNRpVf$_QH*OjOLy@eQw
zx9O+!|C622@4sZ*oSR#X@x|Dv8aY%Rs_+q157O`Ln*z_x19Af@Q5tAe^HPki-gCIB
z;2M;ybwFnGMmsuQNf01i<*Tqn)vSwy`z)X?t7!*22je|Bgx16f07j(MH<S<VS#5ww
zfFV;l{E8#4SJfZCHBj=t`6rOnxqdy&u03k0w-Pm7BGg`2K5c(YgKrHV<=m(~Dbg-I
zeu9{41V6TI`VlTad9>8Ikud#a?6X5f9;~M=77N&uO}z1lCa(-GEx|v9)E&%c<p9m<
z8PJx+rWzrq$)d!tZ#>a*!+mpJ0O+*emsZME@wH(EV>}gX=_jS^tR8}Rs{fK%?$H<i
zl(MeugOmK2GP~jxxjk$Z`tALhOR|a0T@~az5V(73*+r)*;%43Qa@%S2JGfNu>KV9v
zPUH-@by=Oj`r7k&QNfIlRh4&&lxt<TC-X7+vdq7A@OOnQeAD)h=MSOI7l|9f|83zF
z-)ATlJjFMEoI++#0Ks7oz@E9L9d#-hV7Ky%{&%4M{>zj8m9~_$7UN2Ho%Zv9j;*cC
zS(H_ho@>%~*;+ZGvwd3F{Lz!VcXc4zpG@B9hKo3rBK#3hq)Cj|&g53U^jRZOOoDJ^
zn?Y8<vl|{kd7`%(9oEfCdlxXIpdP0rFYExZ2#T_9ygr<=aw~k}0_5b$q~2GTD^K-2
zMu^CdG#kJkGnacF(@yt1ws)qQr=0@gNT=$G?>7e_x`KcK^(6`WW~+>ildX&pDhgQ8
zJ*ko1MS24B3mL(%vBm=qz8gIllnSR4Q_3cjVDSB~Xipay>5g#}=(U((mBJxk4I*8|
zZ~ply;XHTH5d%YuS{VY3QvvOQ@6H;}r-8RMQCMqqbKhaEfLCf~_?Gf%tDZS|d2X4Z
zsk^q>8_gwiXO}AXIC+_fV&?@1$Nxn8i%I#1KC}O@_)og{y}x?0Uh*yiIQ*ada^X8&
z9H4Do+nn%Wnu-8HI<EoV`jLmAGWmbYqWu3=4*WN(y1ZlJH2+ZhrGC5Wx9#U?Nx&N+
zxA`SbSs|{|x{nF|Rg1G57k{bEMqc5d)WTo@yNx`Uln9;*Oant;<<<++p{2Jv6Kp*|
zFI!k^9`*A4ifSZ0KP?*Li#)XyomO2Sq1we0_Vhk??%Shrz|X+X(tb}i`^^l0JPK4Y
zShpa<Cv(vzgyHoV_;mysaHAXCjo-axUS(o?=Z>u&h$d~AU3VdW^SW`A0+Npq7Rb=~
z<w@+3GZ}f}3OtjZ{|$g{{ci?e@x3uCAzu8wpYY4CmACZtVkQGg_I1z!{$sBJO?5Xg
z0-gt0$j5%Rw)V|`%0d9k@n)HUL7jN#o_kt(I`a4@SDz7A5k7SWyfkxmzv_c|m+P--
z-Hi1}pNyFEBywV|2FtSPAVQHAztfScMvn%vXT;?_#!D%;VbCUSNoHSl19k)V>W?Wi
z<0#u}EkSG%IXT|mmURZE{}gZQpM+XbMNFkZF2C@r)8nbLJQOM}bn9_n|KnfDg7}(%
z>vREp6r;}XByYsEt9SRo6D?pAw(jw)P*mn@tCUE0@aTCAN3s&UK4xQ+X;r#5ssQyI
z=9FSRP+nr7^zF3D8rZpc__QM68P{N+;nUoQt)GDcri`-JfL1Ip1^gEJ-r_$gWST_u
zv!8qGuf{-iAA0V~k8ar${gXF0yRUj@=m6|)4;P?QmL`*Z_-9ehJo00pbqe7Q!WdrM
zE!V+nnj*xD=HFfzI?pxUGuhPX+-TdD$7h6i!lIMzEwFd+$%-BfV1jeoYw<cJzC@p?
zXW%IgfGCUYLh=LojQsDiPwrSyqRY#bVR{M`D*&qhN*oX3nlNHJ*K>;bj1kIW3@j{W
z+xB%QlX}O%%5)f1A+0+t)l5XvGF&=a2_dT<^i{8D957khi3Lv%8`)rB^MOx40reor
zdyT8kxd(j4rD8#L?u$+D+uO+_ZDAg&2m8(33?Vc&k`r1Bz!*xiXv?DuBgG&|JtN%R
zN~glR^l)n2F|JD)E$am0sj+eVLE2`gKm$nUBN8ETZ?!cnxE6TASf&r&c4ZyA^z|=b
zWNJvz082Z%;n@3I*Y-X?J8)<8dd%RE7$NmwoE1QYGi}-;NY;Ds^DH8u4uCdtZ?ms`
zbFBrUx_vrO!^ufDJ%U{`bWVAtZ!eQ%Xlxn=0hx}T*I>H!jQ}5mQRvpdpe~|@s`T)m
zG$4pV5t#itWpFk|jq$x<>j-J!ow#a?4M))<Eu!B)ySlK}Qiu_COIR9Con<zOw3j<o
z9|w;&tB@iW`#m$2yi&*CY;SAY$p2mUpeBx&S7lp!p|TU%0tW}Q5ieR}s~!kMj{CZ{
zK_yh3z7GA>w9yGNn0sC;v__b%L5#C4zC6>HGGWh9T+KHZ0QvS`Q_Yz4%i1-Y@D<i%
zkKvie-l@Q$(OX%Niz2^s1ow#}1z(ZTT4{kahTc@OqaUq01GZt%0f6%ceg6X6L&hR4
ztC^NW(=JH3iKl>e1s>vWsIWY~dQMNjxwk%&-!PgBQryJP`V0hJj`u1h8Y}uchlfk9
zs<F=b$Hw+LFZB9gHu7@WZGsx`5v6hUKv;gqB}Jv2iLc?pKo~S-J$I*k;LY2$Cj?)e
z1bc1)BKvm7F<RR&{D5CW?Jg)lUQDDi0tNjXFbzon#3yY=GHvy>V8ogyD?TrzXeTXJ
zX-?&_W@onCqlDZ_%hg&suOh($vP?!-^K=2d#-}INdpfr)!Fa~=waFlwLVW35j+V0K
z$+E$m(~?_N56G-o`rhJz@|MeJ$zhvQ@}0j0dc_YajkAXK_mJu<s7KkcSKGoG&{Bab
z6{aRLzDjZmV+SwWIg_xPox9e0J$g+$(_9#<QsBh!9?o-Y-7EZMA%wLG!mz^3&Cgpm
zZyNz=QY9rBjTP$<=tBAVN~;29D|{RP$6IP@2{0*IWUYE&(JCja@(Fn*<Xyl1-Bk3P
z=TP;}_7y)cFs<n<Q~fS%`vZRHyXX+mVUPiPh5OvS_=}g_-QDw#X~+G$1AO!US0Ym2
zWFxJv%l;GSH^#lx7yB+A`_?R+qvA_RIn|&|137hJGk77bhEx)`$B=ucR5)_q;!>FB
zvwkFGL)aMTtRSS{du@NE6+5(DANAvv=~otMpAzBHEK8}v_ToA8axZHm^kvnEeTeg7
ze|EjB`_$MDdhiTT=%U8XrAbhT8Q&^eqc2hI97A0R#N?cub(b$6ljcW_xh$<p^y9NN
z*>|Pos5$<x%W1nre4~ANT44?#tHS`<c#$qMnuu^)<+PU=PwU90*`pYp<efP?(b8>N
zL%$_~hT0iFCwbEzf*!Pp1X_Wo3(ifN*WJmGkzPmv@q^wpR~fMvh~<6&Z6McXb9}f^
zTq2-(F0l_oD@)0}IYz5QCHBtu&=vrnHTHwN1wL$lBZN<RP1c*`?>KvwPkvqlBtPjw
z$(VyNT6ShgcqwIXwCju*K#qQvE>*hU=8kOUpj~%f0bqd%LZP_l<hA5vSf?;6kmxEr
z8o54_+Q9^Pl8J2iV*db8Q{qO@OU3doy>~OUX7Vns+WCy3ERcCOu3<m;BEEbS-lSF{
zd(Im>hqXFT1%=eHuS$_?GC)*VIkDVn`b;oyw>`@!muc@1iO4X>X2eJ*_uBITImX?o
z#?vCa_QZcmH$Im~W+HGdeZMYK6sIiH2D0zn&ydLo33LF8+ViWQ>gzRaMwS|Vtz~=-
zMM>p4D3I*whO`#{r6l=4s#p@jK5g|{gpjJAX%tuF%&Su*h`zp&TWxKDN(r+(ORo-$
z3OYj~e=Yj*p~^aRa|LZH8BbbcCTt<f3ZCTVO_SuWzGic_$H37poAXH;YUQktb4kI;
zzH<~1n<=43R!rpiafz4oJYI)}PL)h8nLjECHTe1$By#oQy=l0l2`B(a>`ajHowSK+
zb}-f$GmHZ$S!P0^>uo*&b4hbNC9z+Ldk`?rm7e1B!%WcLQqcZO#=vVAza(wPinmEL
z+W|^BFi1WJSA}gEP&pIN2Cd>l{p(y9I;5qSb#-PGHXFD&EY{*D5j&X8kLQ(PJA7!G
zMJ}5%!Ky-M0DIwB<?chld!0*vP7TQsRJ6o_&^}23tI~fWk6UIW-JVXz74%x{>;ZYE
zBdZwj&Vi~wyq6nn_h&~3KBSfF^ri5${Ut4pqELTQ$m)u#XT!!`x4laa(l)_}=_+AT
zQ<J0Wf0uRjnO^F)wI5yU)cw~lf2SWix`qrM^EKAYtWJ=`G=&+}P+6pP{K4f4Ho&(_
z?qcP~j*gZFC%=DIIG^KhZo7*Zn%L-zSi~FmJB*kxE-d%>{4n_)>N^Jx(Vo}-uXb0F
z1lqNP*raDawu76;FK~0~1q9T;PEXHTV*=o}ULb&4gS&m{X9rXJkm}wsmto4C335r&
zq?BQMWOY8;KkMPSA=ml6`-_A?qbUEJ-0r3U>PZ%$FbTMSebm((Rx`jfv)ixqVkpvT
zBM-vvg>LqEmnalV+wH!PzP$4On0gPOMePIH9VJ54^2y`p)qsM0`jHwaE$hAKeD?ti
z4YF3I7w2c0=y^fu=Er?2Zld|7h%4eiVR3n`3l+3`4k9X1CL*rbmzuR_AZpFmP?|@T
zsj#P0n+R{Ao-JAVMgtB7R1ayhfgxambD2+dQ?1$o>e5I{jokOWTVgcz*TKGPA~r7X
z<rtol@zoq%dUO9<7mzE*R1Cm^(NpE*Pw{$w5@a9Fz7Ak?Qz}uwlw}^$;)EdM`4@H<
z@THuM1I`R&KGW5EoSAH`g%!j7l?{?{cLrTqo@^=xpnhQ7UWpC>G!6$e01+eDYJKx_
zCL|rv1B1-XTJ2<536J9WNx(uXX4}Byjw}x{%oCK{T3o#$l}d%5-sTIh>Uo^e^Wg=;
z9*WIY^^cRTr7oBHj6>Dy#nwWp&jvNa&^LD<R;7o`wU(3L0};h-Uj5cu4cL2Z&vBH{
z`5W{?tj~eTB}Bk8C8yjvv{)d2mpyGfp&sX%llx$7wzf+NcKluEfXy-PN}|dGL|=3H
z5&5+Kw<%{T{S_Pp0xxL{cTHr@u8^dzUF($}osz4D*!k){2YgjFG)Z8cN^spA>KXqI
zqFGP5;L+L%y;R{Xv24;|y`+*qh^sfXL~)Qj#{il!JanzeE`v7b?ODNb8;7E69}1?f
z^#TA`^?GX#vE)mezEV#gEB`}8Awx(D2fWxflzmaLuL@e!;{h-%Nb%(O_Hwo-!lEPP
zZUP~YPrKK>Cxv4pSTA<pW|c@+NEG&vsukC1VeD2=6{werzP#tKwbUoCCa{fj?Wag;
z0gd-7U)Dxi+_+0zc>;j(_%oi*#dbDj89os8y?0O9-#s9Z?g9A{<3C|$=95xx*0O8*
zd^`gg?iVLL+q>dmpXvGn-G;K7d!1!ZQsq?WbaY*c%ZR?j{1l{7?K>U9RevYse0B~i
zQU4BNZ()^ks;SfU=v*6s1&YJK_xA>w{*}W$d5tBJC2PhArWCiLlXo1?>G`&uqp@nZ
z0z;&IaGF=pRDAre@(*tnwO_VC9GJ3abAxe@^>VPC=ayKN74fk`Ku+GD+R+YkW1hj9
zfe(VtTm$B6>40rk5vyb@z$ZEI!Q#V>|Baym5P?X$MUrIkkGAv0pHg261f=EzFq6`9
zas&bE^I_=E<=uCPdh*qOvMS5e=bzeAF5*kmJvx>TTigOJK$8ATRKSAX#nE*Luf6x8
zujrE6c3iv=>q&bCGD&ON;^>*y-q}WE1XI@m?%vY6`>00x+GpTF9ZpdssNT!g@5e4H
zgto4K^P$G=h$B~I+N129Y&m)w6>Fb13YDG_vJ{zBiqw<;a7!Nb0G;MnrZ3DtKW@Fb
zn9b}DWV7z|Dhy6u$$+diu4FL_wqEf1E$l5dP|gf?U;OLvQM&~{Yl|-?ew!>TBVZ+@
z*79E3KQ3OW{DnZ4aAE(Od}u)3pxE{bx*<#KY9_{BG`y&2vPjgj+z>&off}3kH8Se}
ziogEdB*JT#z7zHBM^DG5kRZmdI};GLu(1sP3fTbwf$fy3C`k%>!yE}`ZfDm!%hOVv
zE7o0{!$pm6){@h2mf=Dmjo*w$i?Y2Y&H&dFzB4`Q1hx@9sVxlT(iv|&zuGY7naNag
z{BXKpKR3gb=o)3FC&jF5f|(uH0;+`9XM}9r0xed__=7igLDtWulRY=S$O1_W2DVxc
zAevemI>FJ5a-9@$<fIg;twcP_rs2(-#q{SoiV8W1(Nv!45l%i>Htke!sh)gJ7N2)v
zyRXFFEz#o12-X6YQLfqOJrxX$(A8`w3JS7%M~V;EyC^1P%ZIV)Lh{i5V-&5*jT@nU
zIZ&v9r5TI;O*<1fsF1Lp<Ksp3to8LWvYI$VwJV)S^iC(I*oCF$JxG%@E2(N2yPW`V
zXxdWK-#wf%&jfhc)#;wDhAOM1b^T2BhK3tH5jc<c7TdC&9)+=!t$UY*Q)UHGnPXs4
zF`M(=7EfPodQi-T*Qu`Dz!jtvV+?S(yeL(#IzR1A_(XO`oo3B%bSd|iBB{sm@TXl5
zGmda+I4)^O^AQ@$8_OOg=-;T5v8Z~qB#H9OAex8WAsw)10_9zeo`O=|QC-<ATHmpX
zH|$CO23XI<&f1cibtZkr{kY|LZ3%(ZDG=N!MOp#jH!|Kl<hi$<$rNP0<IfO|XuQW^
zlB8?}&5k$rvbLAC07Z!a@HsHBtXw}^T<Zg1kInTuICkkJciQi?)*FF=+k9ij@jK$d
zzIv-mllcD5m-=;PgJ(b))SaseSLy4p#bM>?PuGCQq$B;hkidlL$Osdrkj&RWrIK*Z
z&n&5#U!YB<(ZbpNYh!7lmSro5jurA<_Qu~?qQ28N*?A>3Nci2onFIY|*TKGe*L?i`
z!WCP=Jq`<rvC2&^nTfOTw82Q|wKP};oE~5%qJ80TEk4bgl-HwQ=l2b0`*vl}2y>^3
zf@0GIIU{;|@O0B?`bCt)1EC#Rj3DFl`@Z70pWm(KRdL=HnOtR|v+34m_UCC^@2uJ3
z%Y7?xxBi{5>FW0q`#)0uHT9i5^>ae^p!)kZBPsjhhg{0)x9=0*YuB!o{0U&-UhnUR
zALk;uYCrt1!SjT2w!42?(z=PsJAd}HF<^X<P8U5<O|O*+{xI15%O$3t-Xi-pB#)rV
zS5ta9`Tg*El2-2m+0}o3_#U0FmdNXa!be0qcJg$w$AoRZqHV+9ih{~ywRvSPy_NnA
z)~d<O$XF-C%2Nse@+Qh<@lO)7Qug<IiFTq*P0STrv(o7@QCqfNc8K9j<jxw>dFZXn
z=791&3>>#*6zSEZufRqBZTT=)RV|{YGV)1<U{zm-&3W@_v}jqGNKS@>{;PE9gyS#<
zgN81w9!=3FmJAjsKJadQuX3`7uTM&hf!$dREAG$E_>0$}EG3LkbFel6mS~S^6DO0I
z=W5|sI;5j?V$qT_RS_QJD`S(@J*LJQ#ZVENr(U|B1|)*x$fsd5FkK1(VfsmvO|WHo
zx^4~*ub;R(F=N>`;X{i_U__9J+wh7bSYC%uFj^@U#Hz1XmsZdxv&<Z49iyF9g4v@g
z5wWp}w&&7tqnyzF5RG9YV^2R%q>YlHsEtim-1i}Zz~_2{ikG+xrHXttxuSf8W`hcc
zXPFAhsv_;zqJKNde5oIxkSOkf08VxBok{da`u)Olm)F}G+O%(~gnZ*|Vy*2g6eclj
z2*dUO0=q%PzR3Xv<O^K|hAgj3^C$(Y!ooB9JknGY-N%VVFDg;Q>ez-oeMt-9%%`&b
zNcMdgGMtEk?=4KYe0t~BzS%M^L8JR#I;cp}*nZI3zN@@dwh|)mAYf6GSLJ9YHdY(a
z$tjTa@-Ti_-^HV7mb|xX7IO<`V8x1ywPz|S%|JTT7E0}ij1MO@>TA6xA;x0Xv#(B-
zreFXKh?4fx?6z=?_fD14w}<l#%G2e?l>@bgT=R&mr+HQ*<X2W-tY+x69hlzWIBkD6
zT3b3XNDE_!^1wF@7e~_MF88HxS9CqsiAltg20uk?y%{oQ(iI5b0`A(D1a!Yzz!qtA
zy?MBX5^pCbnWSOFBLrLzYg~n8dZqoT={{o9u3`@QO6}2L%J&CxEwNjIeF{6G;$*$Q
zK1K|%;`JL9@dw#EnMN`!8rpMtM_hxi27g(Pv^G15Upa2b+@lvw3<{=rug*={+zSh3
zRz-%n5dUJO%rg5Qb&^d6U*ZfZ^f&7dOkQe;Rmv@Xir~iEvFNuzDUTP{q8U2UZD-(B
zW{B~Y%^W1RIUlu&IvwUDNX0*%T{;ta<MVFS+iNZLBt<>~o!Nv(H8PNGVM#qh`I4SH
z9)aT@KT6iumO>u{GJ>WL#=v17kpcNW<1V4qG0n|L>+rxe$ntV<@3&dlq6CTK6`pq^
zuA#HGKzxDW8!~-85wFGhL%PX<uY5K6UUPSU=9F~I2Ao|v$0_FA($!F|T;;ITt<Mq0
zgp=QBaaDReGQ2VFWnS<$m_7JuIyM6Gwsq)Re?QIukkQRIyo{oYxR|@zKCNswj1&IZ
z>U<(`0qAy|kh}I>V0G)Cw!fKmrUPe6p)m>*|LG5nWZlmGY>i}{`<HgHf|fo_+erJU
zJCPAd#DBa_*z6q)F`pFPc2-GyS0L`0AT;ql<p`904#>8Z#qH`H<#K>#SKl$dUMF_`
z&5kmn<Q<*DI|LiWe`+zLj-kb#m62VE8Lf=x*EO095wmWB_KZ#Xf|hc()<Ki?gyZ&C
z#wIFKv^#W(oDn>RKjQEf5kayibx$@0yMe~Y%3Vw#YXy8MV)v?a`O;#xY??_S#iP&J
zMwxw*XhBJWKbm;xRU#I=P`+VhzKs?UuRNlr;j1sYQYLFkw?=GlsW-64z~`kQJV^HL
zSZbTRCnTk$B&Z==4+ZmQET|`;AQ5|pVg|bpro;U+#;4yuLn(=1o3xNwW6nc=B7&}^
zouU)e4~xic=-b`|Qp8Z(>-dk1rs;KJKVps5U{kW{6UiD%zoqQltq3R&W1=!Gs4vmR
zew%@tBcKM`baO=%Y6HF;<8@d%#%qyP$#-tOWgTDVw?N8@VPEv{OPzie&&GT&OO$VA
z-57RV9BNRL_g@s3_o5t~z<)h{etceuw3U&Tn*ub7wDmxvslQu%_`1D3CDo(P<am%c
z#&CvR8KOo!;mo^!_#6jbKL2jQ2U29u6}&o}nq@ihD9&jHy`zOFxfF)Xyitx8o$a_J
zT(zp~;Dbo-<hBvohI7@6v{2eABsjF5$3B!vWgm{#@`$eJHXJuYjS*vbV4(w#mxt%%
z8|I8LT_FR|0n->SiTC|%`aAJURA0LWTu*)kd}PSacH~vwq*C{Fv$r05Ex`wFpua2^
zyeC7FE+;o7D%loqa5ml_)5h?f7^umh;w!YKBk`J&S!#HG9~fC{j<e`x5~G_)A9eMB
zlm0l*jMIOhl*qe8+u9lwmp&fA40#-X9KHvCl&t}!I6ciu?%9W{3;R!ac`d9qut0G<
zo#Yd%^cfuc?e;R-3UtEU_J9NK5%Aegd-?lZeC{F(61MsRXpi4&qDm;X7di-UX`UAk
zq&BW2rxT5xT1V$n6f=7_;3_*Ppy#93pI0V6Rr5}C$|&oCxUps^+VQR)iBmdlyo|P_
z#^h6Rtm!a4V3H1Vw?ssSrM%9&@3KykxHi%X>|Vn1%OtN>{y`|aujveQZm{Wq!?11I
z-yZc{CEpsT$GVS3*bc+}tc5{-nHqW0c8wFWsy&>lhe@Rbx<aQ+FlkWr?(z<I+Pg(+
zDYP(0-X%AE$bCv){%GIdznVC#`rwa|Q4Q#y+Y37v8imeqT|qNkZ9M9lHN#eh^7}B)
zHR-Jzv(fF7o&Jgf75Y~!6&_R3e~9-~H&p0cW`~aj9>_bSs``6p;5nGW2~K%l^q9t7
zdVbLAm3qFv69$dfzSJu3pklvV>k7ji9Lt)S9bw8l-Oj0|w8M4u?ds<Y5~H%E4t{zC
z{wR|N`Xm+cDNlcFJU5R64HA8zPjZ2KSwTU?2Y}!PtB2*3*4-825PoU__I|pk2}b{p
zr?waNDHy4=7`?>4^<4X9L7~_~@Gnm^9CZ)AyvvwyF`Wr<ik7}Cb$Qm$3oA_<iQwK_
zKCZk>k%7fjb;KQm^`!wDLSG*@C5f4l+MWwN<tIHA60Z?jhV#0|20hf^GeWwr87+?k
z8_+7NL|h>JAxL86JQLn1b$Z#ZE^BIW4BXJOyTLpqys?qdU7&kd3hlNWCNPmNdsT_(
z!Y4Q5Q6-!&?k?r858O%Q$I#1qM?o0!E)jZ=()wkqe8DaaRrerUpA*Vlw#5|P*4Wr^
zvi(x1;rsE;Ur+~}jVU^KU6&FK#S%1STZhcl8(IcJTY(%zaHi@ObYio7Em0v+KqJ)i
zmFN>f#PeEw7Tpz7D#u5#S0w~qzvQHZL)YQbEGlv1=j<%5^lS|ZxJKRZNk(Vq?G`}w
z(LMxA#*wpK^7Mtz>6CMxZG5bXBVj2|^V}*P#p-)kBo2glneRbY1A`q3@#NiLO(Rj`
z3cf$aRJjCIN<{b6CbDw-Mgzys8?E#_&U@6R7QH10Kh%rCj;ceP5O-@X$$-k4euk~f
z%A5=5%_nlYp6RP##B-am)1<~Z_f`uq(nN750Exj8nouf9XXh0D#=TzuW0}~K_4_g|
zx~y0!`n0V#h{r#1cx)JjSVlqTK9c4VSm5m{(@fXjAnTIwx`jW1Wl8fatf}5mEZO7F
z?*CF*YUNy2JlrEiDAyVa`I?W}T#sCGk1h|}bxN9YYc{W_otX}raM@%`1gl9++1u~P
z7G^S^ou_vRjx}A_C2w0Np1uUgsQ>Dbe9+Sgl;lpCZ=7@VR6;0vp#M(Ce-40|@)q9X
zP6F_QCvSQFKg5zfrJQrc{>MrAKMNoLK$wJF@t5BwrxTSNKYPb`+MD`we>S*%?d{-7
zVahDleN{l6aTI8sBJSSsXs7et)=c4r^bd2Elka)giySI1&X$-4m^)>_EKQz_vq$zk
zp6g8Z8#QEnyn*OUU7)xEvs3Go7n6~El?;VTqm`tE^vrq!+Cm3%<%V`0n#=8ghmUNY
zy=OG=(62`{0b#@FYWn`@_q@6c<C7T7hWR5k*Sp_8bD<Ejm6PiF?-U4oct+<*DLkOm
zAK$Ti#~b!l0a@<cD2!<X`Zwu7|0dKIaHd!{)jeR+!|tg!YDdTEs`8~?l9aHZR~{3#
zt6qK7n%9+>`+B$#+*{`Yfgv^M#A=(FEvP@vdwKb=^X@mAXS5ot^?>tWd4kmFSukh4
zeWdCX`Q7`hzZdMahEk7YbSuvY`cB{*Jjt^5EreiqeC<&MSMS&EquZjf1I*UUn2k5F
zoej!)<A|K899BJLP!-!q3+vPzZDSP!qu^V8=U#D>zJAxQj1-M+ZuTKz>J^zH^dlm3
z@_xAbUN#Q6G&AHqP%VWCoM%VvrBV24FuyV*r+e9BK*N6N>b(ezNIQZSC<V;4Lifv_
z#?M)@Zw7qD$U_Z|Eq~g0q<RoOXOJ_E>CTRrGnTuJ=?%U4MwQQYeT8~$Ekt3ubu@6Y
z&kSi27S=py0}Aj?-MT2pIIrkoVPg{+^YWcIs-+UY<sa>Ts<I|xHLlcOC|q^C5tZe&
zIdg!s=WOzBWxkOBreZ=n+)f&;G46TH%u-xod~1xZCCG_AIkgqPBgE!XKiatO80|lk
ztyr1im~O=O{JuS>d$i3Y3D1~#x7K37b0u|%S95p30AK5Qf6qHDVdIMzzr(QZ@IFuf
z&x3O-0f78~3nz0<eJ}NH9se#Z2H5kiv1fn!Ac~jp^SS`Y>&N3?0fck@r>eCoEz?fP
z`IhELzE0b`1(=_)$By|38RhH_i15k0msiLsbOPpbdPWX1nVC$#B#9=2fk$5kf|vnK
z>8}E}zaIAr*0BDwvu?#q$hz@y2GDoNn=ed%;~%L(Cu_;qZ*Dss)lU_p_64?0!x(RY
z*v4__FTT^}Q)6;4)w)C>TYa79`bNGpK3E^wI?;NJ+whQlST>x8F*y^lJKK0l7}uYE
zD=klRNWLY-H`9D)RVABaXJzG?a*8xn2?sB>%HrO5f!qXceC78lZ|F=IQ`}f78vJJy
z8&tmiy!$<6#T};C;n)szWcY!!(l^_ymFe2sHwk^cMmEsx(ep9%=Tc#;T*UT~Gc>=@
znuSQZMXiXablRX6$@S}YJsvX!@Z8bTF@(Gocl*T$T1lGE1_iuXXq*A{sI7bs8x)cD
zczzm;bcD89)<neb?bdXoTU+gjyWmv_o}(^(bbc|yN5B$#hItBNjn$bT5^beG!D`JF
z#fiYo|7vyBNtrT~J*zoTL;O<(+hXQH%pC@{+*(?a@vUS=CCr;#H<(`bwJ#97VEzJ6
zrd!~&I96h+sxxs}Q^xk^J;D`R2+V$czB)JtP@xG%_uC~3?sZ&#LQwUtgqK_YIfi)u
zN!Wbvw7~c%w4OGxBLVQ;LB;khVtxcwDg}JgyKPyS#Yj^TQKR)ZkGrHw%0Q3{3T7iC
zRdVlZAyo7`X^iL}KGWeS7pUgWf?C&E4KmBLVQz#xW+u3yXw(#Uj2dw;mxRA>MEF%1
z(HXcO(mbdp1BR_b?tTxknYx_Jd2;N_{BmhnchgdZ@0x1jU@ogXUZ7Mqjk+KHz@AOX
zhFj<6qDqwqb;c3fuE{|uuP;<f?}QUnVaEfvLL4%eM1TXL@LupcDb;n&M1I=r0HS*o
z(U-mPxi64W5&?_((z5O#1H6}7(w5a$z8USCrBtTCTs(XY7;d)|nUz|eA(b|-L4fHJ
zFs-R{?ub2<u|}+(C?09q;kH4b+89?*UCGLWj=1fx-mQH*f%|#rjE~=`0Kv^hCIFr1
zwlt#uf5Ks6JRlraoa?v#v50d1b*bA!2)&`S|1mOa&{4nqBqdo${sg^pt0Cq~M{evp
zkL5idxnp_@`Qn~v)mOA6aG4rTk(y)1mk##V_NN>8Dm`3YekKV0I^w!%Rl;2!N!GZZ
zb)$B06B;DrlL)Z(XXb0W#(VIb;z8(=cME;IZ$2-K;<1GLb|>xmQsXkZvbBa?nAiWh
zNc1oqt=eLtJJF_!qLeo7@=7PK038XpXdsZXRrL66#H<57n^WANxY3J%$g88(oe-8q
zqEYj&htDX`EQlwTIMw3L<<2UIEzC+U<FJ$PV^U<iON+qav<>>g8z2IrJ6{{$O~^f_
zI{<BP>bkwR)?!L-wm{ksPhmsRG{p^PAL%|TJbF2qyQ2nu4J_o*szh}~Gn&<-WjRon
zGI9)Vzs@S#!i)|MG$fvS1{>E7&dz*Ms^V;!-p~c4Rt5u2SAE+@1=iLuY~QbkYsR68
zfXL)Uj(=k4heBm`C&|~NDyei+(hQ60QfYxxjdAG^3&h=~P=BNa%YrCoyqMSNsvGzX
zo5Vk8lIbK-QRp9<qzO+2W<lXE22Px;EEIKY5;z7u78WKQ-(}f_=&<7Q`BGDQ@#{dy
zw}q@wGvI>$R|cPglH2h`nAzZOu%KZrcF&~@%bI<CCVe!+nv}|7HT6@ib9ZDF*MRsI
zjaAN^x7qP|HdD7}D!A*ur>7*q1=vC7eL0S7IRVK3MdHRC99a*4$p6<u7mcL;6B+>-
zFyFnY6(awdbN~Yw0N2qgHOl?LU+qc|Qnz$K&@GU?&H$b0hY9SP;s0+8IDGyQ*ZoKP
znlmQ=!rkQWUm>QEUtgaJZdvAqF=VgeRXvO~E*lJ-D)nI+TpD+7lx#JK+FHA)7^URB
zZ^U7*3tBg<MVdnSK+uE-5Tk=tpIks&CxDTysLtlR*u1%VPb1U)zG%v9KtqLFO9L;!
z4>InXImo+)!sG1Kbq?<A5Mb6NCTXZhxmF|f43211C`7&zukNIAcHUlFeiET6)=p*x
z!>H9_J&&g|?8y-i1=ZJ>Nwq5uZT(%hT*BK_?~1-hZ1!mGEoo^bk5y_Ut@p{a+yd3t
zCl(K6+oNLqJ$<naj9;DDt*ay`YmYJ$u9~qyDVyu3&}we_><=Vbpj+Ga{c6g{Jk<)g
zCiQ6jmk&TGmsH*kUh;s>3gUC;URjfIf`!a>0Uyr{<(WwnN7Ed&q^tyg#>(_L-r$(N
zI_&heKxFLk3Ju(N7qGiUZesT<))a7)*^;nSWm0{?hSl)a6Ilcip4cmBX2Q<t^XH|7
z$EZSCmW!(8YHO06@R0Mt8XdQMI-kjs`X`tR3kX@DjeOts0|z@n0igX{7ZeZYP`=9#
zfaZ7tW3o(%J{hpyCS)fA3Y@_cg#&;D4x><7wU7+jM-6<J{iO4}@nngpP^B|S=}0R;
z)ItFeMxR$H^6cQe=byggGsd&E$^SET{ogBCIQORpALMtlDZuL#I-{}Py5;=->2E2o
zU(b6aul%G5YySR&v16yPCtWamO^**_|DitS<WrAO9*UVd$qXW9A$oGb9bs_rNMSf2
z(;SgmA&Pa*+};-HTk+1zP~8eZ@XAaA(mkm$>Ra(cr_;p@Z%U_DKvQiWsfhJBW}2fH
zV)XSR#?mM3nz#J{brfoXoYt9?l;(;we6=+QMtc&IrEPHvBJG%WBJHX&{>$E~BF5fK
zDKQ+TJ+*YhAF({bb)!_W#_dX9`sq|!jmO&9W&m`_q%U?L3!?|8H=6=dGx?^?UMSY3
z8A`D8xgSMy){ENmChN07EK)t3dfZ7H<6=s8IXNRI(w<NY=>FD%32a+q)5Z=SE&9SS
z3;D_0tF3U;Sk}Vn1okayW2|$}A5GMXlNy~)2)@Q*%BPqZxpl2h2F>Xoj25^Pd{E~A
zi6VN2A8=TFfE!`J=1QS;L7D!EHV%_#q-jfZ1mhv#mD%XjWdsix&1mVb7p0VxOhff_
z0$RQEs7qXZ@qDNjW%frEZ;4UOG1>LD;1^{J)&oju+e0{Z$2s8{8`cblq?F;BM>d}g
zWL$`zlufPJvBeaZa;49<qfqbgm7HxlsRJl5d3u5LZ)pMLHASyyS=ZL88;<HXcxR}-
zUNXdQ1Wxj!qJgdPyia&_lH8?`QDo~i$QaPH$Y2e*%A^2){i~cqSXiXr8YEkJVy?ba
zWo0?Sa(L^uA)OZc<<%BF05=!{LpC76KcDV+A=;ki`Z8|e``)M)W~)WNS@vZmmK$P|
zt}i48#CSFFBr(>c)AF@>J&wDXm<%{Ck2v}bC{i;~L2D2IM}6$fYnWW%Lhm6jtV-|>
z4GdY8)Q;}FFC;BY*z!+lU?;I*OFQ+{AiyZ3Q|BjJ`=3!L)nK_jj6djSkMjTM;@_qm
z0^-q906#S)F%f|M#K)g<C)L=Cg64k`9=-uKlz$Ozt7=tU4d-~-smjrHT!$Qve^tPS
zKfE7h2JvLnJcO!*$xoUY(_i1aw7fVLdULgCaDF!Dk45)P^O+>MaqWSNrBGAm=Vx~S
zv&Eh9>d}mI7TNm0=z|2Rq_v8%s}pfW#>g99N6*s9S&CON^MCp%VRB~QIj-vmVR6K?
zJe{vMCG&i|b{1R55$1OWRvYjxtS`W)$m(m`nf?qqi$&0jJI4F^Gle}4N}wjz7zt~&
z0Jn1X{+Bpg&zNA~Kg)8|c^(XsE|(2q(Lto%;q#qqBhVRthHdMv3?Q<yK~YgDSdLa$
z@FlQCrq<*AlQNBmuc@6mn#p5QRB-eT>#0dmLW-3{nvs*s_$PUGiJ?hpNl&?36yvVK
z>&bU%SKK`_58WO(eO2OyqP*E*9bN6&9XMguK_>V~SaRQ#=8~|&T4>%9OLK59Z>|ZF
z2ejRMdMYeBC`<IGRecF8s-%px!nG)~-vnosCOC`P?39!`1FRRznicQx&5^Yg6}4^Q
zwe8<^LuoXT=qrEWuo6$ZqQ<e;VNQ<OtNvs4Q|Syq_;F6p(B3^qxYv$bMMNx5x)gco
zFAaFvqt23S4e!OKj>+kvg5_PoT!kRRr%Jo8mTP_I8!HE~^FErbv9bJ3Y(|62f<A&!
zv09QH)m`0ihP0RnN7Q4$l2~HWvY@$F*3xAUgKBbDs}QDcm`Ps%_<2FqFWa~2H9+cV
zPoL~TzFy)x*Xc$xqX!!lf6X_|Z(6Xg@7cg+to`pMkf%NT{pHF@SPmT}>`pp{DSxz8
zio{tq2Q0gh$nJ1^xxs|$%))ANY15Oe23JWEncGuAxNgTvh1L8>?M&N60#l<9Bj~fb
z#mM1G6jZp^>8jR1Y<r)h3U)*3TkJ`wT1V=_{N<Xd%wobdy4u#xH0v&F;Y`}qQ`ybp
zCNKJ2bIyiWn615Lfbq`$?|=J#`JO?Vs;{ivd~!I=-g&KHw)2dbsLp*m^T?&Z`+Yx+
zi+`W~HFow5|Br8Fcy|E{tOSr#ss+@HYso+=EOMkg)kgH$PyRs8wO5(8S9C>L^)g3a
zSoT%DUL~v~Vi6B+$=-ZWsl2S40YrLEypzvaXq&Ee5})U={O4crj=Lk<R<S4No0dB7
zH+YvuK*~__wRRrhY87n2vQG8a!lr6P)+d;8o7iTmaZq=?du4nFKbJ;8W<L`779m%`
z3z16Q4@VT2xb%6a#GH@*r0BGLr@2GyMZl;tvF!RB&WkS^^K1E?7%h&R?KB)TC_LAh
zD}JzNTsNf)EAird(|!?@b4t=Ov!R%`zAb;X;Kcmv(a-POR{cyXQ-J{VL<YI2$X`T3
zL06#AVJO{8$}Q|mOr@@nKo~YXF&p<K#U_yOPJU8ofTw}IJ$gCwW@uA5bNNg)OB5ug
zRP}mn-);L!vsCp<BiSgk%IR_oD8F*sXKFEY@K?l(+*Y1Ji_xI$iLOb@Cj9n@u=78E
z{{uZNpF2u@1_^guhUVEW4uIUO(opWBeHlJ!0zTnpEF=>amI}zwvM8VYt8(}FMxC4_
z9c<a^@P32A%VUc)l<_Mf@hcPz8;(ox!*FwmQ0XLEhe#hVPsCxu*050IV&XHKp05WS
zS+GO62|g+(&{EhM1FJj5x2Zno&FB{h?mZ6soVa+_AMCM|RPF>?Rres2$$bh#z^;ke
z?zI(fjfZ;=zLC?<i%{#ax=+edrDP$p1WH50uzvQSTI{xeSkenLO0s0mZ7py3dSbfY
zLy-SH%v`q|phcc&SO*dl>WYfyd7D=16fj}NfisS9NKUSXZfBC#Aa&L^nw9}CrYQ`D
zUj}lel6~Vn#tl-RKS$eSnHsV7fG6nH%a^4ItB`<UJ{!kLZt$rTT8(;_JA34v)1l0!
zXT@bjN!Zc1L$@;1&aGM<4(I=`WATUERp3#k{GI~+Q2`qneg-n!)B0r4>v5%Kw3wej
z<+lOBFPxoy<nHtv$Rxjfr!9s$udv_!jl#pJIr%vs&(nk3dyB{Qw_aRROHG^2f&;ma
z2l2mwl9Lev4nkvA-4FZlWo}86zFoHG&W+eh9@RG#s^P|6)I>#wHW>)iJDdu3(w56v
zopVPBV@_lu%RaO%^fkPs22WrS@(6gmL1qDNzIG~)Q_PCCJfW`^XEWut*dIJkCn>|6
zT9@)pE|iznB>(;QnbWBJx;1+Uq_jG!F!uK`9${e_YiniwkVeOjL}pZ4a8<{2(AJ*T
zfi;TzH6M0+Th4D#V?zuz3=75OsF(mViM^Pcj5iMP#aNO9C|!%QacFL)!RLUy+fm0v
zGBdhOqzU)h9{Qx!K3ty;lLaI(t#`xN)SgtjeDGSN3upfTw)*wA5ak;ewiF|AC<l9~
zo*pL5pD7X406=jL$Io4pW_Rez<-iRq0+VJU)Q5|{-sjV#6EWxhcxW}-%?vT<%I9r2
zg#8uevES<Om|j3ueLLvdUUEt*Zq@+ojb53l`y+p2UnvP9yQaXFuU`luegcp-gs3YY
zItOXGhU)?+y#!9`WG$8^2t=mHIXRGGb)}bHug+DXqA)gR1XU!cY1vL)d6K}_RUx%9
z|GI^}s>lyG+Kc7dT5eP|BWtT+9KXCYMh(x*JYQ@qOz(A@m{Zh>+&;t|m^0Q&Y@<I)
zaS`iytgKY^z?}up$K!@t|E}?kxWIc*m5YWT=&9pflBoF9SFbJ&Ftdf%JDUo-O7We<
z@M$};NCe`AzK3c)2)$>8L7{e?v(9@f)_9gj)Sy13yO{pz@n!<craw~jU|As{i(MZm
zdrt6HTiDDE*&DeE14ykB)rQ)#a(B&PFd4STbk-2F*sdN#9dLHS_>^VLY@e7ed;ys4
z)?o{NpdCr|V9Rbp<+f`tUWC;L<%KS)nUIb>OX4sMwnD2N7CcxX)>^y<kOP_(=em#a
zU?7Y@$$+d3fI%7Fb(^5WQIiif9N3EGizM!LS^lqc=+>HxHD|}WeBkS2Z#;;yx>g6_
zpw$U~O_db9Am2>#RZM4?I)=J#>r)Kt_#Cy>w5*Zs5>9B0DCdrw3HL{PQnphCRUYdA
zw?dl0iH}}`rS>fnS^FP4Q$(egcCdNKl2vJkRuk{Gz~Uavtpj5Koryn~8eZ9RU*G%O
zzoDF=)#?&n+!CgD7mH(JC{Qu3w3pXOHDYdZj=RGD9ZH`S4qa&9>(3W^@ZV+jU#!V3
z-usYRz24r+`h&!nn}3$aF!gmjf3Gk8+u_}29=Z3iQx#4OXiTGi$8=jYZhmVA0~6xo
zmXHb8i={T0FMzr^r4-4rV?k<b9RW?8a?d|cO8K;h@$-u84{R}Y%P4ePl0<sQO0DAN
z4xWbEBI4g+gff`%dzwI+n6LKjbO*aa-q+P599wR5wT+_g-9%5Rs=6S%-aalp7*&36
z{ydwz3Vt@%edD!FXxGLkq1?mxeLpMCcByCR&J+q`Z>(OOee+;nxPEMPf8QX^v8tiU
z;aYYY#v0lm(zmN-Xc4|mt?U_T@NwDzISqe4TAhFVd{lRzY?6;JceA1gE=|DO>XE0q
zx;jShY_7xojVOjcn;Wa#OD7>Gq}ubfy8xrhOtSLy(8-Hk332k*3Pyv+;tbNWxN&cI
z+O31~d9f}O4X*wl>k?ZhKgk8vJ(<4rw}bh+&NbC)%}<k>wDHyS;X;T)Gs@cEHm<Vg
zO;;gvC@wUsIddsJbnCA2g_k!J$H03wqLr3jqY;ny7M|753Gh=Y$#X1%I>cisJ?lJ7
z0jZFfcXa^4>5~;mN~jL;T3zp!8wJmVLmWc#Z+RGTpRU|89{@zaTU%W=xY0Ph=^K(!
z9ueGc(i!CDKvd9l#Lm>O)&IDY>w17G#^G$CO5j@2raV@3(eUGR=$w)2`{A39)?lmq
za=M;12v-81;?Xdav2;gKti2BoIGOi1f^FOY`#{W1&hD7li;wA}cdI`!DIPQzW<c93
zTiJpZ-A9l3Atuu!f*2(}TePu%Dy3D!wrS1WKc%hfL%s0mWbr}4`*H}_pv!Tv%>tzQ
zuDGf2i|_m!RW;4g!9k>!4MAn@#ww)%8-1zSj(A`udBk=i$Co_<ev24&{>{eX(|CA|
z(hOiz^&UR)(H=7ux?~GX`k5!*mP{Pxvb%DUN2Cd6we4}`74C{@Z*LgvKwPX2StO>!
z)A{R9q}%WH4Zn0eUX?bA{8^{{aP<Ct^0yPQ0B_8p-1mroiFcajZ6+`0`x3b2H2Z{8
zJhs}91jD)+$a?lE@?guaB=9<16i*eij!Zx@<!!4U`*sOcCg@5U-}Ftmx%;`;5NAik
z#}grGWs#2RJ%mRp6Ri39*T7!m!gsPPT^4x-HYRA>0ih~dD^C%vE@GT}3=MIAOUa^O
zrk{e?^d<Kf68ZV#kZa}mbKee!Zs@kX)8!s2mn|5Gi?TT%=CSyXD{}Rft+%xd?6<R+
zawlb{yE&7Q2gQ_F1Xju7GoxPt+4uA}Bwtc~tjori7{BFWSDTd*-o>wyzmyvatpkAZ
zPtf6KHun15+T1*A<C>`0afb-irk*!I()D=JJ40@B9D{QD9JfoegyY<UZBi857|?)d
zJ5`}tQ>2uq-K^y9A5=Lw*L^Iy=pEIuA;<Lqk}PFwSu<I|zCUnSaXNx+0RA|#V_Laf
zeq{)dP8@bpfXhu@5a@U;dn4>jKk}efPtI;@sUp99kpTz=pZU!tfL!J}g)u;_RbA-C
z(@+t<K{P|^+t|!e|AygZaqj{~!8ZH1h_Dn`wu1QW#r~#V>fR1PeIok;b}H00XMawZ
zM3FjVwnQbCOUC=(mF;iM*K>q=5YF1NzWyf4{Bdp}gfyTPf0M5Jj8U^I{_*U7#ZhWE
z=~}C4nO3qX<Fe_SgUjiq121fJA6oq^onf+6EMab1`MAQ6dmTtBCxITkvKb!o^RTlw
zjm(8Sdi30K_fz9$uDe`LgyG1^xSBe*{#?JOg`poO$ACkL|M7p7U)dol;NrIzAXtEN
zMXjw<s~lu)V)bM_CFOarWdHTIvE8MHw(hOaft+&7F9tBvs+kc%uK^&zF8EE-`|+H>
zdi$-xhDC?>H-gE{4SqN7WhNFDhu!7Y2CX*S2sRgh50mwshU=RTfc&eomU==$GlX0`
zY$uj_NoPTXo@E|5Q-Qh=9A_80@EOWwZwtW_F<VXioT{H%fZ8t2ks1ul?6)r{yr#*A
z4^OKSwR3GHpxVD8tHa{Waw8L0goSctvXyC)QCZ%Vmc{yhn|p~jI$U%AeDFZ<igjoL
z{-5h%u&U>u8okx?!?OOz>ySF~SCo_K|0RlPUU_TzP~W!ENP_=_-Z`foebd$j-V>T1
zZoe_&8H$^oPRoFdpTv6OHH`8QK~)23vqabJJFi_jznp9<lW^edFU0|ffX=ZT^V$!Q
z95azf)<hS;sknz>$l}A_IYXzNO>P1A9<aQNemWx)ut&Wr28mCYoFeUO!iPsBteZ?_
zi4Q9CH8{J9HnR>W2WlXkSpp2saZHm-q4U}%FYE>q%nnA(WZ#)h{KlJ8gDL+){bT}B
z{Wvr|*%Z3TVKMbRVPTs+BGupezp1t#_!zEKXYwSxh^dU9Q*N9LFsru<wlC7rrIQlb
zK=d4DuF7)5y3V<<_Tp`XgV1`0L-TF#X7@K^AH>=oZfXN+g~9mt*1g4K8@bmur-vTI
zmn}L3F_iWuW;W{!ZeW&mgOq!VGtSig?%N)v;n#lJ=VI5nvuCd)X=c4e)K+v>DRCDv
zAK8-e8&Np5TnceNoLWvR3JV?u|H2<TEAP_BqtO+y(b3NnH%nalPC52&a>?|SD!hlR
z9(Xp}e(WV4j~kihu%i{90h2+WFB$<~*MSrub6}P7QE*tolfA7iXNG6ntL{913q(y5
zVmobr8W5gCP?++>)fMt_c?7deZ0RM_FJh|M9^f048ZOqq2fTC8ZqW=Yp1WBLAwIHk
zUMMskFyXy{Uby(F!>zb;<MT6^ZhiQiT+3RsKStFFRuNmSxi=r?6;TaA;gw7hLdPm>
zb%YMPdckkUM0g5jic9a=3am#SuyW%CcA9+&R@t_&y!DA-D|rQFAi}3caS^N?)-c5*
zIjFI74A7Wp$r4L=bEMLeG)GRZW=Lg<q>6L-r1$3``u)E&`YwTNC~nmc#vaR*84>9K
zeR&+fT!(gkatL*KIvHeZkb3h`gkhHQIxKXh_DJPiKt|46-bym^UcbF<uevm5y-Kxk
zXkdw>#Yq-vn3OkF3tD-b8<wk1lPsm2Pu)`_Cx;!g_@&KlIHBN#j-hhcV%grj`+(*^
z?q7?mips9WqW3VsQr_Ohj0JWC2r}%VZJNEw3c>wbPn(xaUoc#AxeKPu=v3~9{Z|tc
zjb$eZ$*#yTu&k12Ow5woPGw0)^LQdsp$4+pWXL|(F~bQ;unQReM{MU+6szo>FncIO
z>w9t5JJ|$QZh5p--rt_iuyCRd&pEC-d<vhyM|#KQNn-{FjSb({5R6B7?&|4cBD8WY
z3drbf44}TkVH$PHKb6x%b-*zxEwFL?{}hV<NFMRq4A%Fam61;WoQKz=)IZuF!90l>
zCn~z1YWcdggg!9a%1(RKNGk?{yaxykPoE0U6V%ZOx;jy*KE$%e)i3X2DlK1JzmQOl
ztdvC~BO_YlFXlV=jt5+lwQn8<u01x@{fuxD_`s?nusX6wT^GJ%DS1wj`TF&WhiRj$
zCZV}mZIwGRYHWl!C~&fl2>a8J1=f8O$LsS7FerBwyfxqkV2kszEkPFg?#r@<o5@VS
z&Dm&F%G#8Lum0$qtFfS0;B;k(#jg|yXVW3|&8lHuFZBIWIh$Xz@ab==gpmNisi67<
zgIZ|_Ni9C{=|j75Dut3tnUuSr$3KY-g;te4;%5(38|G4Jx_<un;l<qL4b$Ja+oRYH
z%|WYG+;6=o=F=iMMB>@a?NKKY5uLU(70V%GzI9TMOduzekghZut$ys?TC+!B>cvwV
z<sXDCZbCm1C(Ej;hC8RMPL;;0@v~~BY~JMzZY`l|O;45LhUzsO5EglMUVp)7@BBZ;
z{sJnh?`;@{r33^)LP`XrLqR~gML?vbq!AGiB$RFt0qK&GR!S+QyQL+R?vRk~j&IKl
zGve=mKkNP0uvh~#=bXLw)%|NrYbVP;^|qXNk)Wco^?hzjR-R7#*&Kg<RnN{iON*bw
zeOOt&%tCU_-&_vpK~Xgw(XXr*V##jtdut-pdlP55YLm~)B_3=?uCG&l&LRN8!xN$z
zl?@VqYfXNOoRb4k$Wr(idtg&g>7)p___`wZyl|@-NJ;fHZ55Lyik8%$A$>j!WsZ@N
z4FtnVoSTh=!T2vpdIepB$Jjh&PiJLD1|q!BL!K}tn>ocR#AikW<3NRlo_$Oumv<>!
zSTpgi2xafQQHNq)pgaIeAB+Bt`5=G2S#@#G-AYHlt1Jx~!O97~Wn~j>3HG`9eC~af
z`H2Kg+}T$Bg*qeN#>ENbKWOH1oq#Xkp!f07qSMi&7hibThWAi>&sNSRR}E_r#m$tQ
zH?|Y5YacpDRo{+pmuKZs#J<oL`dNtbYir!ajf-fX%T{@&bSd|x5=+Ski=Lj8;f=dp
zJd2#ndZ0Qezwx`bmNs!bSKiICu6rVhKzmsr%A*OM`_w<mz&8Wjios1lsT)X%%GKIU
z#0A{ReRj(ChPnn<2j$k0h*Wb%u-EDEtc2X^Qm<=6)t;{Dj$5OPTj9=*%-CAy($Y=)
zC#Rcl-13I%7F~CD29{Wsc9$~T*uQg~=^h_%ag=TUY>HI5+AN*(@pvYrK!=-E-LX`w
zhVS5nu0&DKoV(Q6BUSE3U#ivExI6ATADlu#kpb`t$jA-<I~HbCMc@5-P=}U^3Hey+
z+cA7%+ip3j_OF#<qI%{t<^_>cr`^kKRz9VB%b(u^CX(EGMY~wGw@u%Ap1zILIFoJe
z4#@745q~<u2Ou-#W}9JJH||<}Fh_jQt~Xv({xS04VXvvm*1R4~gOB&HV5O-leGz{o
z^XZcsUH77l9J%Y89kHC6^s*#I){!{QGbZf^yA8NCrf9bVi1RErXDzK=3vIA}?5SEW
zrA2QRMNY^DjA!pW4XJo4q8hU^_>8jHL9p+Uhv{so!0NjazN8)M2p(BApNfhU#<b_=
z&hwtktIof}P#u4V;r<3*x)+Q-6&k%@K<@UNEO)cBrNykABorql&lgX%+UHGWjxu?B
zO`jGRRhcU_zF0X{od3ozGWSD^2Goyb^$KlwM-DswO7W`Z|M*n-yC3irJRWy4i$hx=
z)_bBUz+v2b#TuyiIohwjYsjBb4{Q%+ie?9m^#{IaofvjobWyTC_9p&7)EMjDf44b9
zA%q|@MENmMd$GB5QD1nksiU~h@p!4sv$|HE;WMrL2iF@k+<zoLWpLh8$`S(_H@~?>
zu}yA?jqV4%h;0L2)hC9|*0%Ci_d<X-{9e-XiwRO|NxvjP4Tdhktys0>7uB&8H>3<6
zKXYCDTr_;leJ7^G?rop#yrp}En!L!^Aq(|N!l!ie65z9(OLBDH?HDp*C=8-31`>#L
z;QFsu3*W7N==o~4dG2=JRK?Cwo4fOJ-}@Fygs%3OgPg66%a?>IXWZ5wJjSifluKDC
zS*pPkZq6SX8{ZDOW)2@7rm3lERXJs4s!nvc*VVfy9x?|Iw|U{}0>WEvlS>#m+jC#j
z!oJ1l_%2)1HjBk~Zp)Q26Famlpp%PV_R+i0QpS<Ir+bcfFTtYfqRC@GV>IJ2<^LCZ
zwHe@kaU|?3LbnSZSKE*2FAOoDl40R)4-kyIIlBbtz5Gz0$+pxR4T8EQI{mpP@+>&Y
zdNxa0w%!m8DhhbETg&FTU7siN<(|%_Rlj+!i;=4KJwJn_v)Z+H>3Sb=CLf<(JF(uf
zN&e4B%^#~@vZxK02e)YXojk_!q8_7BnR*MVE^O7Dt%S<Ox%D>X`NvOBcZSAUkg*c$
zxflg_d*}F7<<QTw)sRQoF1dC47pmX|{kfvgj8ml9^*s7&lv$pT`ekg4!b8p8sHTtA
z3u@oh)+!6+1TOnrZ^2&O?ZDj<?yXcFx2(~*P3f*}s_iY8BYFMC4c3{at6^q&K`%GI
zD45Ok2FMb{5jKoHvuBS_5!&&;$92{h-4CX$mkb}P&nQ}a8C(dLx!gL{(0oH}mIHjz
z;1kKJ-j76p{&l`M30b?VpMsU`cGd1g9UmtU&2LbiUQ0`$^ikKFn~iK+=wW;DOli)t
zI+hs)eYaRY_siov8QqL3ivg?POx!zKb_KHWB<8k<nU0_R@yDI>o@n*8D`F%cs(a(^
zCTiXJA?NI#OZ<MJS<)Oh-_HcU4NMG^tIxORKaab;8Nkoh9O(B`t?}M(X;l@Q0i7&b
zPGxtY^<N5l6JNiCoet|(O+LuMHfHpRGcab4KRL>rNe_<Yw@YvSVWkTOvL3wdQ#A(N
z=!9PWjMIyl`%#hp-GH|*hQ%6UZXG`<1jJ#kM_M==&rVKjd!jOb1Z>W-OzS$u+H`-*
zM7Kj+&~Zg7;GplZ?3b0zO=jRTG60=Z!s+=K=;^h4+iaPb8Va>d9g2#Jq<>Ltm8S_l
z6~62||59e<!&xQQ-j(#S_Wr8KPke1aWgF!z?yaVoZ%`0X1*yR9DLPr$18BB2Wkx>|
zKCUwL&MoZs3V&fd-e}$`q&6%{OAmfOpHw4O5p8Q@wzL$jsi8b;Zkw2S-S+B$lIk`Q
zrHEq1K-p^fmZodSoQcaFSGVJRZ#Aq*;@c!FxPn%WG8J3a3|9>2e5959PGzLM?xQ?c
z^WSm4%YfsV4$ogf$^tGucJ^~BI`7QLM$9;-bBlCHD^RIZcKjCJQ{O5<(|SZBqw_=-
z8_&>C-`QGE8ltDiYwqq*0j&JLel=_jZbYxHf{TMrIU{@@J$Ul9xv8P{{BgAvfJ(P=
zXQc1_>pt-oJ@l7*tfrfGau(BaMOto@hT2pj+9{%I(_SzJH;4gs9tHaH67z5o6{eyL
z#ce?^zjP4lGI)BRSd}Yq`(^O=npb`P`M$VpcQu$|4~E9KI9<QTK32=kE4@iDr0|%4
z#*qxsz^*h1Xr~s|j3gLveHGeAvU?X9?K_xbo7X$I9{+jbQBv~6a5ar4!YGBanyYBQ
zSwYLhgmp)gT=>{#sC1)_d)Y6Y9FLI^Cn~#AafY=Zqv_E@MRlnUlLhdNdME?&YR$|R
z^l$6+T!PH3tms1TrPPmyrQGit2T#HY&jOY%Kj?yA2=BT#$YgbPk~JG^+^^~4)D$T5
z`(R#q`K5PAQJSbuLQ?7-K^f*R0$hT|7}2Twp5{Yn=7>h%8Mj5*N0|tI!dz7>X!7yA
zx9lfeORz~f9;{Yw$*bddSHl2gB><|f6T-DhWp1cjb!ke+?N;p{#o;Jd%}G|tH+|iJ
zKqDUaH;Rsj%R^yW%5dCP-{0c7=-RktlvX<@wcf<ScMc8&ZlC6yfHp*6?mPBbqHSHx
zy!ECyZOK=K25X;7(sA66dd<xKjw!XrPbTS12)_Ib+St_UA&7cZakw2jK*xD6_DVxW
zwKTrN!zrrsg{^U2&GRIZi^EbWg}qW5)0vL-%TKKaI##$E14m)}%p4)6v5GGIos5Is
zm68{)`!!sv7&W^vW<{SH$5xnxwp#mO_ysJkk$M@wKR@e9Wr%;fcZwb|H?r0lRRw7j
z^oe+Wc6!p3d|zBVt+e#Xn>TMN85F5uU09PNLEMY(u}Z$eh_j+eP{*U8v9_o|cuW>f
zQ=Yrm2fsdz2D1Ps?(nblWd*3Hrj@MeFUS$JbgZyR&{dHoa-z0nzc(?2Ux8FY2oB{A
zE|wQmnwiM!w}#U}Mn&z=e3A-@3K2fUfHe_Ab<u=}&B!`rPnD?agAWTa>;6SC42NJ*
z$tDIcnWX*pWtSgZN}F})_RjKDx*R`$j*ZdMID3GjNWsS=TBhs4yL?U&4sn1)#_b;X
zLGqN=kvzl0@%@5>qrmuXtsuM3c)swcs5un;x3C#x9=4v_*W*&ND<2&_l{G9%r9Zt*
z?MD5$D~b0mE-!8i%5S92OO76bq7+oPq3Do5OS({$(<|v^qKPz*KiRL2e&cZ7B`Kk-
zDLFY3aFxh>A8Bvy(Af>8TH3!<n**U*XA=?<$Qnhhk$>di%+Jp=l$DizRP#Ge_pcp5
zJ@>@2FY_teOD833>OG@6pPqy`UtA(ym;KUgtrBs6f4fIdQx3P#4-V#;4;DUYa5!x{
zQjo{aU#&_mjW(R!;lgp*?R|oY`4ejf(c}@y9s-}6Ay=M3;qz5yGFV_W6pYRqS~T{0
z^TrU&rUw%TE!4GK&c}9JUPDm1rk)TNd-477X_m@c_Y7bG6nI12YwNE-H$ogR%TWjz
zz@5UZJ#JG2eFxIR^}o_SmOlgZV<qiRIp~Ix3?JwtntUCgY!Pyh4s>s$;>pFuYSYux
zhM*g~8u~rMmTMq;%gbi5w^itJ2P16%R{5ZFr+n+3(X<`>(XF?saw9aSr}C@s=MXXh
zPD`~cU|gjaUCG_dDL!$C(Li^4>ZVJ|B6iMNJoYdYvFC-F!yiKKoJ7dT#ulBK$@n=t
zyO8W``_C`)T#ES!eN^b^kjWzCBNj8m@sIbPxkTXOJk;)enimvZivS7YnU2Wl%mQtr
z5-75Xod1&*A$kK-5qm%~^d1x|Ci$VA=M4-#_6`ha-FyJX0L#X5e78kOJ<qlr7Ng=U
zdEozO_l3~aD?+RVkKHVMb^jp41Jj7puT%ox3n7fOv2^IX##r$9V->rs0bG-0g8&-~
zm-!ZC*VT-7?N_y>Wn>D-oJ2)EKwlX@Cv)cW#uVd>iscpeIr2+L#N%R*10KA+NsOn;
zn5(N-|No=}Kp3^vf{Nt=q8#~09|Vm_nDL(0jqgA7jyXS>S}Y(pw(7keJp@Sa!FDB6
zb*gvK$~2PKOv}Ncl#rNMCWJ!BCCoPl3cX5gD%bDyOe}dnKL+{<Qh)ZV;Z{a9uKl`0
zhaob6w*Kvu4sXlrpzc7-jY=e(GRM`gbH{TDWdJ=FSDUg_*vj>52TG!|tE?U#_48Sz
zt6ijr(xwKf(IOhRV9hcCE1(Cl{K`Kj_*j4lCufi&BO^Ip4&EIqGK*IA1G7h3D%Pr(
zW*)tTW-3ma!v!iyfBfg8RE2I#ORNZo|ABe(O+d62(rUUh%vLbbIbgRC%L@OVXE9zf
zNv=8h3s_H8In-i{F=iU6;P=DBnprf5M-ykPpy`R;m|-C{Y_y-J@?WW{S%t4k%U-M*
zKAokIqqGa}(hf&A2-;?55lCwdDgg0$?_7KezYyz&%Y-sOV2-Qbp8baYgBul+&l@Z3
zTHcowOevz9Ur_K3rWnDUTPfAlrdC}M;oK|48+#=?PSmG0I|D>hZvu)2_}JNEh&!-G
z@{VSXC!sGmL)`QznFFGiB3x_m(>bm6t;g4~P<*<Edd`ve-vRa~{5lyEQVFuo<=NY-
z0^GxX>@VHj-D5@Et3hX<hH6PM$fSWw$hyp5Yw{)M?l9kz<Y7vAH2u-E@}6cr9Hakk
z-uMc6<}#Hvv3HE1F)*q!Pdk^xixGSv147<}YGlwHKC7ZykdjJAPcPYPrVO}TVc%qK
z>GTzK+<;Tq1ty%H9i#Y6%tIMPJdA~EMrkUA8YA}*Erk)z{A#*&Neea&y`O1dr^y|U
z{C4kL`goouh?ECF-D!qH=L!9K3>KIQBAzav0IWl{D9%n^6e`%-D<$D_uZfeh>=AXn
z8^K843wD>@{-wbEGFFd}odF?skr>K&%HQalGzUGyOY0Eyo;<9D<9n22EoWXxT7`hL
zq#Z;TfC@maE+q_jr*9V-4+Gu_{_zKpp(u#O>r#bQ8$`U1RDTiH_<$n1x6zdCo&PmB
zrXm<i?RDfoEXZlumCAr=yM=!s`Q!hzn*n~GZh8pQ#-s^C+!p$P6&gjwO7;BBM?ncc
zVl|mu-^dp$%x}Yc247eM9SELKX*dc6vj0dVW4R@!Q^hR#2HqHc0-{ir1oQM<Y{~9{
zj@8@OCo^1c`&h|UJAj&fH-`56fP>Wo?3;)+!=;cmouWEJc=26D<}W8V5;@<>JU$jx
zzv}A-M+;AG2sz4;ynvq+Z1+~xaIs=bI9ee8GhNlMrY7lKZvck-bgkf|AOh9@RRxZ=
zKs#mP);r5T*W+)A{!9-tb?BTtHYVSOs^pMgybwSbxBg^>A0PCSZXbj^35dWqrDa_#
z;a>Jk|1Tu0f<$%4jsEoC5oN8s`;!Z0y-pWb<*<EJLP+jhyimTUGCU`Ub#4fATX~0%
zd+|{YX`)6JeuR<BzYJ~=Iz2W@Bj0G_#e^2-!;cJTi+iZ7mFs7;B4;=L^n3~cj`gzF
z2eb3-AWJF?giF3Xj*UncH?mM?V*b+$HmYcp_CVk>!v{FpqYkTk7?kerr1SixP4KSv
zsprqJ>)~Nl-%BStev|oO2>h`USVo-8Ir?~1VD|PWhOUk=ygU5mGV*mwgOuIiYsJ-}
znP%nk3NJ%LdoIV$yDHT6Z-j(dkr6`gJ>VxDDE~h12ii~dm*M%F%;xXe!x(6}Hew?2
zj{F|Fa)zy2yTniyD_rZug}t<p0cD{acpi(oePtDOqZKLEaT_h&a0WfdbOS#37nvFJ
zxi|pnFcssT!GHEs-6kx569*<ysPi&xx1_|e;RDLZuK|ojx-IN~k<dV-T~urXOpa#Z
zvJ)^i4cf*uU6t|OrMPv1uxrjKqT9(6grLGIR-NiTdIBR<;dHMiLv!M;v*vY1##KfQ
z$Aln}I3X^z`)|n9{SX<G*Vb(Sumw*rINUBC!MdhrdAXA`B+B?D#EhN3DKRv4U-SRa
zR9qA$qqGvEz|_n<iAh|+(A!u<$31(Z@g%Qw5*`lMvp!bJQ5#V1BCV%ma!V#~PJID?
zO7aQB&+2aN1hp&IuX%!=lS-iVFFlwMy&H<hDGU9gB|%H;1XG!JMd8;s&iyjOg)>HZ
zzDg{Oe{jF<1l7#cxlFo1%-}K#E1}<R9!E&4=%tjCkT+fBl)UrtBdLYkQUUHi^Ry4%
zjaoH)KXxv;dRCPL1`of++8OwCc@I0Z@WrR>N}v9b#N2Cebl?8~P@>wd(GOZ5Ud!qU
z`VK#O^vHjEWkj9Ze3%vq3CV7y*%wA*CM<9}>|2>MCIcWqLoF@^fxm1D_O4WWvODDy
zJ_+ymBJb|91~tm%?N<jRRnIS>GoOIu*QS9_@(_1k=IJ;-#VZ6qlmxJe!G_%-2J`ll
zuc8!A3&0!rn2k4v<GH_~k<m0vu!vRw2S?OHrB>mbF0qsy4whNaCQZTkzpLQY(*N-x
z==Dw?u7(Y$YENFCsQEcGcBFC?TXr)wv#|y)i9YXR5}h+d7byt;=j(p}Ma%9twP7?;
zzvtpCqx!OdVh@II>%A9PEk!x+I5;c;7f8EYoCm27-hNh3gpblep!K8>^42_C*iZqn
z7V7(Wmv7mcTqDlNwy$W@Do$u=wA_Wh!1AuX>Jx|fufq8)(mX^AldEe&nZ>%O=YX!W
z9LQc086rP8+*x}5yCsZ*gj5`=*{t7#TVgltWDToQ>HGcOCB4gh!+YB3=~(Od*8)Qr
zWDZgkrK{z>bPO5Y(@85>M6%K76frF|ULy;9qK8*Rm{`;u#f6L)7(V-jCGKXv-9Shc
zJ1j+P46XwgcrJi3&vi-|N~3vxd)5pZ{#F@t^>%kFGO@BoB_~tqB~wX4QJWZ(kPr-r
zu6+D#%{skbnvzwBm7mP!A1Mn|OIgp_TK3lq-AVzh!e@dytkwH2#&#>5=EQ_Y(rSsi
z7#YPw%FpO%GqbuH6_U@vJ?RXw-^bY*3?L8kG(yFp6u#Zl(*q_TOlY0-F7$k&<Ka<1
zJ|?52xCCQ1qFR_2<f0h|yBaN1b&h7nO*pl}IsB-UesgdYhe<bW0MfzHQTtQ_<->to
zq*cx*xHqLC6n{U=-&Dw>aS&pmEu{3^eF?uqtuVmlW$S5W=puz_6(0lh0%vDIseR(1
z3U+~US5%qx{9!0Y3iM<PqY+Y+cvP9z{KLB%kvx&<9jWn+Lz-Zk=CFaGu&&mBEo#}V
z+BnPHNa=_<t>Xyi{q+Kn4HNBs@Op8Jj-f{P^^NiSH%2+nMQgXN0{;S{_Q=E?tZjHA
zo&t(%f};9;UsO4s)}v<8sAkoF?wjgmhL0XAjI2FvBa1D`oWd-6C;MnUqLfjM+MGpi
z7_N4lZ`X>=Q={PYPM1(gn}H&y(Xm*f`>F08SvGp`oM>9hsDy1(m;#L6DM&D3I13R-
z<ODzju(JunOhs{IPNR-c;2ZBTZ4?s|%Xi!|0we$SZsFaBX`<<Hfm}{PxX4iO_&8zc
zrul!I9#oT$Jxfako{yEU5IuWRxsM0>TW0C-b6YM5a+b*ZKkD^=zj&^G`h0x{AN+UC
zL<z`YBENko&3fcr>Fw`-q^%wQa<sIExi0jUxM73ORmey_JvwR&D3WpG!+oVWKV(;_
z@Sc7l>0BqHYY2H+rRJBkguMU6J?Pp+o2$ww=%U4CHMS!r<h(L2BxY)+@Z`9?yl`8N
z?`Ro`TQ{O@r#*2={m&|YYm4xU&G@I858k4c^eAJ0K%Dug7T{5P!oq<#5T;f89Sj7V
z<*->$<a1b0?hBJm64udmJ9thU5)NY~f;w>Pu>PIjHc(!1xG4@t4)gubEKg)^Yg>KW
zou$}%<5*YuRp+=4Sc#CA*tyeYB?7%a?@ZIc;M8!V>p-`vTt>C;gOS_NXV<3rKU-9;
z5;^Z9m$K@=KiG}vZG5QUZd1RHU|<zM;uXaC<3cff|4S{HvcRma{mQ#{U<ntuqokt)
zSnbdK^ER(R^9O2~uD}@ti*>qPl(5($79(Q1B%B~Cl_%$ua387llMQ{Hy1{gNlvTj)
zI2FxlrD!8OVoPBSyFB3|GovsQ$T4DBJy(-QKHHZg&Q6<6ls$ea1c?T!5UpPV87uSc
z%b;C$@I5LjQGb8`2UXmW=e|5C)WyX`I`b+1&zZuxD^VmPO5xK2{6A5X5c=D>-dQ_?
z#?1Vs=wqpehdj{#Xat2R)X|{o^m?g^%2Lf>0yJF$MTd+05OJ+GHCTzv;2%OL5DCz7
zt9CUH)*Enha9r0X>o&*iCmao)1VTs3)p~<QwE}b{R?*yVCKKJlxpM;ckG3?a72I!L
zotZC^sMHM-ltj4M5KO+58u$0_Gx{>`!!K5KLMXK6!<jzw*sL!5VgqnVDJlT?`9SBS
z4RpZHWZ7TKUWQHc?#n9Yy(myCCIk4eLD$vbFetSZR-7}1?!mRxx1IRJ-@jO`xKS-!
z`>##Gc8yn2gUy<_4{m1aUL(DQ5ca{qr!yZzoCK3#uRL?KO4?e&XwrZyD}sdG;4P@H
z9O2Nfy+j#(%LU42q|i95Wl}Q@lZr=@yKy*J(y~6utvjA`AcTGknkfJes}Q{e9oRKA
z`0?l-2YkC83Rl2GBYYE*p;#&}G1;fC7y0K`hK#A{mr5{E;3we9$}fT;V=ItQT@>Nc
zT~e5{EbhTu+t8+g1rj$yha7Bo;8*F^D`;@h7q$`xfeD019IB4N^s?j)w`7|=%+?kQ
z?dujsn~HuISi5x$tZxJ1s=+%&?B?6Bl>?8SqoZSZXlS>;&=Az)aS18x+`ango!>I;
zXFpR_-kob&uCY(eX6E<ritviG;P#`ng#4BQrf%5rL$^YDgqd4y?fOvf)q}P+1#Ur@
z6c=<ukbsLCU>`|B+o-|tbwFT}dySD1^f7cyYs#IMT>)n3u?{=o;eVqbmymFY85vLI
z@;AxUy<zHgFWcB(m8Y5sAhjih{l|YiF2ooY#O;z0dvO5m_brcMG~YL@2ySr>$8`tK
za{jPFZ`H?2Is?FD(hi?ScXH82MOZ)a{y<yJ_}E+xhbfF_qrfSoSjtMlT{@14@$}SZ
zfM+A)l#GGgN}s}t0_KB85C`l}=S?7Zm0Irz_Fu3=Ya6cowMraf!Si(OL@^QK69TYN
zWki0&XzE)EeB$eqiC)iD{#g@bVUzv9uc-k1nvwKfq|~&mUuv{MV(SpL%*riMc5P?I
zU+5Pt&;3fcpd-KRZ5OV=<SA^4d=XTzrxeNq@1>^yqhKE3zIw00)RSZOt_ooYvq3~_
zI}YO%g5JPY9!Hq(S93!Hm+v%r{*c89fA_U>l}AzXf1NcJY>MSaJQoqy?&uZgdGz6$
zXtEF_?2h~V@B-BVmm1M|LBKt)m=@OX4}&rfjvu6sH+n-17Zbr-yp*sPHu7(jO{4=z
zOs^XaLO7eZ4!*KerswUewzZi7-@`UUmL$2Y-d~5Gr^jcA)r$TPt=Ia&-Hd|`@t<$~
zWXSZoXmG(L#x)Q*dNWPd+V;zRO5XfDL}z(5x)gC#zK{@IOyEt#KcJ5}`iJ?4l8lIl
zPdg8na3=k)I}KG9eD}dmqEqVYeGG8lN#Nhx$wJav`7Z+E60$r#+|uj!a<&6zSq*`^
zJA<txB65fT_+%eIZC?;cEC1_`U*u{-4CA|md>LXkVVz?|LG4V42&<~9>SnL9LXmYB
zjI(zqBK#)N5%_OF_)N<g$FutH8e#9OusjzRX~1|g*~8tH6{4AqMlteC!CThEE1Zs<
z*KWO1IuT+gY(nraURA(JGB$tm!q7Je*%#+#cxwbh`_WoW-bojTHmWxTlA=InH+XnN
z!%Q~6(@-Zx6OWEZf0ok1@QE%WKIDXcU}Znq@RW^r2zl$X8!<Te6rX^zao@PaLT9Q#
zoQiV~xKj8aJ|l=LW#k(%<zKF*05RMI<{m&-H;L@|^XFs8jO_e;j%RB^PTL>}gS&OI
z&P(Smbvcx)S*45qS-sNQGH$SJVDwtg4`XS%gZ0J#Ohzxc4rVGbfL`dQ&)BOgRv$$M
z2Hq>VSyF+WZL`PT+H(5+H(vvUv)e^avE#rIv^EET+&ZhFP9J)%l}eLaphL}-DPWnr
z#2W{-vkXfEdC|E#2P;yRg1e1?0IONx9>z=TVM4TAN>!Je6do)d5TMT?%_IF2MuW1j
zNo&W+j{z2;Y&D}M8_APFJc*!@@rLmH)WR091)bBfj~(DX$O>GYcku3Q*G$B~F{PhJ
zz~zY`Z3nP#@~iYF^%vi;C>2x7)&BEasap^N7Exz}$!2ezjuR>@c5@`l`4Dm9=g0&0
z3e==jsgSMI^tqVtrH??t^N=yFP1Z*Fi2fCSTp6`-q?xZq3PF*fRBmO%W5EP7@LTLb
z+`g{CXf)7&tQ&Z~d4pWG@G;<l(EhjLs`YxT*Pv{}>D#3gXbk3-qPjZU3#<W<fO<MI
z?DwrMoUqGG#${%KfbL%`^5F8!t7eYEqY1z521dl=e*Q9yqsfVUXze2e{GcBJ8I{t0
zR~BqZg|7qk1K^gy2cA%x6BU4Y9$ePNvZ`!BELx>*6wI9;esRwR7JkBpN6XFlOO2(Z
z<Lt`!kB+v=iFLicn|*b-O7wpKtH5#cJ8^VNAZf^xW6tBnEVjteQ686miRG36zJr=a
z7EdMMQS#%HllYB?Ufc+|_^)f+;Y}@wP0xZl^a6f}l|7XIJYaiuY=q$c(?BQ#FoM(h
z%vr+wZuRMpPE-G7?;Z+RqfSe%`3@V8u)Kx}jKN%_y`w@o>pIHJ?8vf!ZM{<Ek<_=4
zD0v7m;$TIwfRW8E0fV`TlxKr4t_)}y=yG<lr{-L>oZoJ^GF&u0-)y;8H1jlt1rS2A
z9O}8g*59Apoy)=UTr{pwh6cu_{xU;{d{M*Ske<a+UE<cS$3xBXs_G-etrrJ5t@osr
zT2p|=>dC=EH~jge=bX&|sd@(zooNrFg%Ml`9wQT&Lqa3f*7Hd!C-~r7Gev~s)|{iP
zAU0IF7EL`bGJV`3O95=PF~_U;d1BMHB6gL#QZWo21ap`JSpO&nx^d{BtcBLobN-2|
zjiRb5bCuJM{;5{%Sz1NNL&@C`Ak2M0lFQu?7=1%eWf|hn#Gw_2dG1BEu(CCa9?{Fj
z=of(o0ooQ8xq%cy<vA5;LQp2bfSH;%Nc((NRdb0w%rpGMz5vDC*7K+Vw^)e{O`&-n
z$V-CdpAh<T2a2O(QtLtwkLIzNOLsEo@1r#=tHP<6YJ=$akpm!#1@+g+JAD*GND+@c
zh?ecz{Qezuje-yK0Ye7E)Yvp{nky=Xz8tUAsBdTxq9yZzv_RU=g*#u7J7W1C9-`Xr
z>Rh9S4PQpcQ_*CxNlSNUH45itN+OB~`k?4E8NJ$8e^IelU?AKI*Qp^ia=z$)9f3R^
z2>@P0nBSYv(Z|da#PIwA0wRKL@MxMlfjOs-Ra78;M)NVQ6{vDl7$&NtQ<Zv0o8fw2
za}g22GjUob;nX0_3{p6Yg|A0YOTQyV#qsQfi%3jxZ*&IK-Ih4oS;>=BoyJc&f6mMW
zIR)21t~yFWH-Q|&jWx~!^wd2HEox{K514_Qo3BavZ8FIP9l94yEGt&H9e4WFg{bg-
zAh{0v>2UW98s%s~M5>&C44_lT9{~w@?NhdKl@`hlRYV7GT<Xkr|4Dmru%*Dj^b18n
zcPnx*`D&q)+t?JWe;W7WEdJOPq}xFJkMQ_<^+C6HCY<O!)qt5Y@QBwA7Gq4#K@~pT
ze<uC#;n!hp7ne#G&{oR4@A?Xqhze5>*4T9FTSdQsD;5nDCjW_3Tte!)K3Ha567QJK
z>|j<7;cCI;Lv)J@#gd>MK~3wZ!~}>!Lmjf$5XvbAFC#c4FMXGD;bDx?ktCvL)Rj1v
z`iGR$sKvTVhye^4u#t=|g!a#}@`vZem_1iCv5$@9SRm9gwEr)_Dy8&?^7}v|MfZFR
zMtY-CLZUfGg=BNW)*EHchOW{OP6(+bACtWE3(=B&QVZhZ>))W)gx-zkdS+f;bYtTK
zjRK<}R8&-xJV`k@Ion@79Je8#!=on{?3|9F58m$Fb*u`QlJXhQ-?Z1cgnLqy+<`s^
zGRuGl`cU|{Q|%5Ihz&E?YLC{?SF84vq!rwIrf=1c*(*eY5cAsAJYPCWW^mU&5R++u
z$@^zJj)|g(qsoEA#<)z+hoyQDpfa<tfd2aIpb6%gme182b8jh3@FC_g0Ofh8C8sQ=
z<rgg%dv?x;^#AcYR^Fei$I1gKm(&4IqHa4%&?_WPUS}^>z4nKtIwF6Q<~&9ZT<>xS
z*Skj-gw(Bl<2~SL1bN`33OqCBn;)NY7pPH0%J@lpgdl`-c06f|43hlc{hoED29@z+
zw8d-HM~^P!3jbP~Ha8bV`;Jb>Vp;NP59WL|GmCpB0tE#4RXY#KR5$=bV-e&I-VL%K
zg!RrV^XSIGQ;^{1>})@8{a^mSA1^8=7esj7{C$1>Rp9eeYi-k9u)bW3+@HirC?m0B
z3HAz!rceq0y)uIelG--;jpgDa_oEO*dULd(AXP2^AH>M<J`>}K{6F2`5F^lT9?0|8
zATeg-+yn%d?YMaU&F4xk0?1dS7JLHZUND=dHUo86y}zYW!n-Mkfi51X#<zr@i+=sP
z#`AaG+`s?~0g*?o^T^{_@r!sh#GNRKfY9*86;8S6$MiUX6Vk|2MH2-;3@^?)U7+{_
z=X$Uk+{JUCvG>hp<Lf@xh-VOKQ&ioTA*)8BmisGeff8<y1vdbxjOnF-2urp{m$Uhw
zRZkKEY*!5r(=j-AKnkR>!SVqhyLXWTI;5V!bmKqXdmOBDi8k!U9_C8R_3G=peZPlQ
z(ptwiap%NmFzXI#p)dVwuuM*11bVWHp63$=9o%)5_yz*(z?g+dv25YB^#(FKgnjuE
z1Vfs85r0Cr{{duP!aNoVKw4=7X;lYprupv~Sv^Tk-Y0HtI~p&YSpNJCAHTP6%5D(h
z`4z1#hnSYwaCC-VT<ddD`Y6ygB?1BR)B!>6f6ogCYoUvj8}LqO{v}RH1J2*iiUDCf
zv%7fwmyo&PrBEvwFpu2BjlDgdgF>etK~7^~PiNK~?gA_MJk1rJY^jY~O7U8x_T<tY
z%G@3b-$4a!iB$<mUQn7|vs`SXAw#U)Q~7!_$*Kz4(D~3?9z0MV^d9i&mi4KIF~bMY
zM49!M@CY9JJ84W~Y3<%MK><cdc$2Ze>u1H|Rdke{W7V|WjK6h%`r~1R;cXj4wAP^j
zoZGmIjUSXyYK>8gEhku|rKMjl53)nG(xrheV*>_A&wyHubA0a7`0?xCcs|6fv!7Xe
z-ah104*@rvq$jn3RU5w~vjkfZua5q~<}8?ert<mo=bl4<pge#9Q`(@IZehFn4+Lt!
z&OeWuYc<gPH!&3TnpH||$8Tu5M?GFQ7vd>ZfJz+KzR__&50Q<S*KsoyOrAywPfWDH
z`+RtYMRa@pH4-!nI-cR&MB~dt(fr511-&+j5V+|e0W)Dz0WkzQ|Hs24e{FyJm%=QV
z_gXTvx!9NWe6;kH{6Z}KMV9?9{{F>zD-m!aqiZpDsbK(LCIWqyeSfzyBo!VWPzX7-
zq|3z_0@tx!Ni+KYKlfW#0rwN!X<>(kG>}jqq~_+zYu#K0rfSI|9w{l)4epnr7Gvd8
zh%)$parrN%fOj71z>z%P=Nml<rA-$x&~?x~a$&(_2#k;Hv$8Cmx~x(5I*m>78V&sH
z5iY&|xvd8-<Yo=%N!`$W0F7a!*!*=J8gUEAMasFnbG+3B8nba$GU%VD<o?f*mEiLp
zEf+kWto6jdaf5}OogJT&k}V^_?t@lPSl!>}In+Rq*gen>>btI$Le0<5H{AT)s`hE6
z_j3dxAt7k4LgVtl>+b)3p2=0HyURl~PoF*|8cT4T1BvHq{kjdwlrZ=}58^H8t%n5L
zu`aC@FlXCj*s2V<K3=fh0;uT;FI)j*?NG3T<SYN{EQZsKenbt5HSrZs8!oHkhMXr}
zLB8Sg5O88j{=v6UvYb3YE-EVGFzsbL0K@LGFFOxCk8T|<GW*`u)pf-F?tlFAu!jyy
zOjaMQMi?i#_J7O+6}B=Q92~>Xv28A7AbEh~qF_fDygZ^YL2rBo{F^uL)EsXH^tDj9
zZYx1D(V`{fc>h5OghDXM?1{xzskphdb*H?mqvJi3a(XFfx$swdOnO1oyGbzySW3<s
z@skpZ1dvQ4@p|*ObN~oXT7hg|kfLj{=1X<||8Z1bJe;&7z&v1ylNZ&ldtfk>!@7RP
zuENZPCB*=(y4eLrP_;5>KY#t29d}+UBGUnu5J;oL#U~+INb4c^AD<NL7KOgGAnpz_
z;hH{X=~g*O0~eb!wx{p~aNUL6-|+nY0chb75zluP`{byxad3!soZh6S(y|4|cVD;w
z_EjkI0+%Z2;o{ydv^V;!r>|&0mxYgyZ-pK8qKmZ_=;P!O=c0>^h}UU5GmwP+J}9fZ
zai>k`ljMu<uTGBk_NFs2|0*y4P{y!~K0xm9$jIkCW0j6upffFnS>Drbcp?adE+Dxe
zvZ$~#uGCJ>&X#2@okWghf93<}P|a_%Fgc4yK=GG^`9*us;JE;Q|BH-K7+njTZ4yFW
z_a^$YHE1v~F;}>-FZzx|AubocbDYA(P*GER1i*m%;L*Da5c^B1RNcDB=s4t|p`q8Z
z#=y}_=jm1rn=x+xRd_r+=)v(2{FUtgZ2j-4lM^`rrv*eiCpwallxzchYQ@U=VtQ5W
zX8=h1FK-ZyH~;Syd3fMn5j<S@w7EYMqV+%+h$hfbI|bMO+WAm(aFs_E#J9YIm+bz3
zV=V#!1Qn@n-4cUXc>!U7tlB}HybW&PyITl3g5dzsQxe@!NkuLm^MwO~vs2b<7}(hF
z#eH!!wgHTdAXY7w<b&wHd#GjwW?f#_rRo0~_WwDsS~1U+krEjdmG|Fubt>#4fy*bL
zrl!X0&gQENyOGxcp>t)PRJ05K`+L}b-?==R6gfKx8MPZRCt}yL8+yJ81lRt1w7S2h
zNNEVz;;Qv+63qXPEmGWC2jdspAO<MJkGi;MIuk<)GKCzc{TKwPFJFeYa|WIwSBQL{
zV5d8mnSa;wpLfG@5E+vXI4p#B((>_XfPCu<K6Z@K`+(J|XZ}K@bHdO2@4^1>JJC3Z
zA)g+957CkO{{YzkTsC<H2xjdR`6SpU4xzPxeNveX<bsJ3YL4BRq|E;x=L4?u$b(pe
zW>WDEFwyl@Ir+<%N5GD}4fx71g@WM5Tfyyaal6$Y=S~>}D*oqSV1JtEfh7$O#ErxL
zT*vSPuA8>fS2KoPi9$^qhafjbiqe%<UB@<3s<D;|@=Q9bndzPU<E+<rA=iZ9$M6cO
zM2)@nlnNT^smldP)NXD|cckoRwjir*zJgsjPvPgZ$`^98MhHcr{CR=PdaIbTaGNpY
z>JmJM<2td;^$yNaBSGxH8K{%JfW8E5#Wi88ugyY*B&-2MEg@3|48m}!KL?a#tg5mT
z09S%ThG<Xj6+hQG&Nb}dOrVx#p<P+4P%)_Ys^V^1Tp5}DmC2YF6ap<yFhh>MPAt>4
z)w!h^Pq|AT;YsZFr)UEY)Sx)z)poM6z2}r)o328_Cxt`eKOziO6Q4xDF1r%+s!|iN
zJ3W|hhsX{ps^m1IeLir_<YOc?Z)+xz1R4)f7BqTE-?n_MDreQwY0yE=?sQ2nhtL3$
z6I#24dKhwpfkr)Vx6)JK?lc33Tc}hbv$Z}hG(cIZJNcpAxg9;gx6F+y@Gto$%YkDt
zQ<f(?&ruinCnf5t$<I{je49Uy*D#_$aR?~)=ur1wvhU>kStey4=7hIF0JE1Txg94<
z8l(&-cqT}!zm?=&vHw`EDSQPptbC@qD5mSmW`u4+-&;e-Gb*Jf4XaxEQEVS)td>2X
z1-05~mIfd9P)+E<Gqd_zcP&FqL^f{lDjqu}&Gl{IlY?W-6&>5nlQ1<cyu|UVD}rgp
zeZH*j?(4ry<ev(~+ApHg(td&De2;@EgiLqlf$Cog7&r>Vms)!VO%|-*ZurH|`3_HN
zH#h1XdR2X>(fQ1lao(%F(-%dc?e{|7bxqbCZ`G-`(*uQrLpI>ySs8Ct=;ZsMcigD3
zn3gE=3KAuDkfRv1FVuC*&i(aTr@JNW^~T|jg=5#Kn%80F*T<|0UzIF##nJ!py2P9c
zFuJ)t*3&z$cN(R~ZJBC#ou8^Aqb}rZ^S*6brg0)kufrs32?^$@gOL6zB<Z7q9J2`)
zYw);n1@0xbHB8{Q4Tsn?l0js%J~YOl@?)|cc`mDdCqz(+-*&C@jn(aU&GjC)0~hQZ
zlXqM2hP0*yK)#mQui=8OD4Ge^44w3P=L4A}i_=58_J03JA1pGFlIo9i!X1>IlqDUM
z-_aG?HV6*$t;BU5CF;+%8S<)yYGzU_f6u1PY6hO(x-~6wx5HdbfA(_d4bhScBtvhb
z=sDkVor<*8=o8KAa@W_HmD_vPmHZ+E0Dv$?HVFaK!D1N0xec*Q7{h0so~b&SZSD#N
z6_}3{`yEC=v=J*}q+hC2<w<rQ-eU9rGBoiqy`uGLcHXH7O}HN|hO#51SV&|ydL@*m
z=JbmkkKU8(s~NNH&u4;oOgHSdiAnqO3cg(dZ9n6uMwCu-^WOXVCfdj1cAzWpQ@RP+
zJ+@Pc?|rQ!TI0rXIG;2&%RR(211Y^q&g@D({7PZnY)#WxW2~U&WSAzErvVz=zpOGr
zn>ghEJXR|4J2f#+Sc~sItKZevk1SG^{$k(&;NA0@?oKc<S-XN}YGy{q$|}!o@zP<|
zZ|GfLo0Xo+hPTZ^ciqafryc-YC&)xu^$E_wHh;z&mfLy7UwG#D9v*A63Dq_qn%#(n
z?2=C;_$J?N=bGId)u}7Fm5*_<b+U8<5BHX-YSunah>$(1eyNkTUzTXU{*ATvSl_L5
z<X~f^j^MGYw!6gQP>2vc9tdpg$nA&siO4KwuQ&;>4Xg5B<r~PFq2mQMu(oG8N==(z
zeuCzu`GheYe~~$J;zZ|rq!6A16<FPsK*n}lJKv>=QZkr+;e&IN86B6R@Nfu7=q_OS
zjvENLK)8ujvB#e;RzmUua)&tFI@ddgs(uNB8|l~7vQ4aoHV79oYfFfEy5Lb-qLOV^
zT>3MbQWm+pT=F{t!qO`+$WL?RtB20wUh!f&?>>x1LK%JTAWzy{RPr(kSEMKDx(zl~
zWN^*e#9)YH5wp9cW|coWMS5Rl2YN=A@5H0(i1Ie(J5oZYlT?llbENXE0rwEj2dYUY
zh=_E@CONmQfJX2qBdIM?U6;w$_Rz94&T?pI2vSfQ^3Gd~8u-1g#N=HecG68nwOKRD
zpHrM!n!`48-pH6LN;)1Y<-Wf4cB}qizF~t#D$T{&kDZiw-2QXpTXbjR*B0+gw{l9_
z62Au&)$FYHi*;hG(n5I5&UtlLB|Yx3q7n;w$YS1vz)8Q>^TCf2x(OrhgVwT_b@$WQ
zn>?nJrU&>uMwl%a$`Wi&S7_&Bz2-GeESqQd)xS5n6m7L*%?AFeaIg)g2M|0LWYK_u
zZ_R#EScZX(mvJ2^neCrKgiUSpfJxq3ZoBjwSi$yRRfQp0JcU7`5vo&8N%S?!S@opV
zZ|~aFP3*AAt#_lv03snyz!k|-ox%Z>)!^5YlE2)KHB66J{(`01?O16Ow0+<$w?4EB
zqd+>f;MeqlSo?+YlO+*5`8ye8Go8e<^D<W2+Dl|xp;d`3`(hN1Uu18X`%a2tz7uR0
zm+S<#F(|(@bsuFZs%S6znpaj?;_}rYCJ*l#8yxBWsjvlzB`=p;U9F2xK=Ay>2QIZ6
zSFo@mK=#)oef`PLwvFQu<ne+qzN@@!kIT0g<i8Po*Cli=%N&#u@Q!s?o1ggBAoW{{
zQtNuI&F;$syHgi)&F$r%V$_FrFSPMH?<!%-G;Y|tN+=6%)1g3^1TJ6_Uk>HmzI^F>
zt&_x(_r5el#BM#fe@mFcwvT`4bOq0RGNFr%G2y<GOjU0{(*6bmP6tXSN=X|^Gq?L4
zTdAElvmM)8hHd*3j=DdeH=dIcpeIN+G$I2UfR0X4LxTzAibeo^S;IpXbbckeqe%<?
zK*Iv$WTFn?flj-^a)$sM4rdf=l)p~`;?XT)(0AR`>H7dZ#g~(0{~1sSKHChi)8Q>p
zApSE}_34{hc*(nsC{FZ=*Xehx+A5!pm(B<d9?#%?7W!<I(eiG#eRhwDX~juobUeXT
z_X$(9{9vZ-(m;xrO~}oD#m3=rniSx!H&O5;5V;*45jm3D@S<mW!Z~%{9@nyKRjY5H
zRlM(UL5|8uiE&8q<Aby11bc0@`LQOpsO75nw-}X?@IV$wse#=u61ks7CLdyAs$YaR
z@w~8$Bl#p<6l+-!M;5^?`N=$rlvS2IlM#cF0~L*sG+%;n>ST0pN&{QM!NXWP!@YZ+
zZ+>)s?`V7`-XVAX@VT5<TH0=&k5k`m1!^RGD)`S`eIQXtt>*9EJOAHZ73BJpJM8Ru
zKYsk+cwlB`cD$44{vARJnYTDu!Iuc=yL+x%l(t>(d~0oFN-K)oBrRH6pKzB`y>y1s
z#Y22=f=!^KvkAY)4TtarS@J{Wl?PII?9B|##(8eCiap(LW>n{YYtEAp`NWo$H93AN
zf0Icp66v)kto#UP7<<RZyv6FtlP99QZCMFZy-6by*`o@=<dl*kBAq9%i8UWpE9DUp
z6@#Xy<i59h+~1@%^LllxY1d(DGiHUdi&*)JjO$Q1qGjF0rGsw~?DeiBWS7tK*p1-3
zyo3xUjWey5FR?l}JI}PU)im~gR=!Hb>?4Xys~?QYUtWLU>uti<l&RbFnOtRG!;w##
z!v7O?GfHzqZKtoYW3!^MQLdE$KE7gHIxjk!(_r0zEk;7s!b?eGo*ma197~^t_5K_4
zbu<h&dInvTa&14U=RRl&r=vf4GCn+fxn$vHrlm=kBR{aoBeM#^Ionh1HRyGOs9Jno
z3a_>XS_$@AVu)#3t7dR@ZhTpLI?L6wen@PdIgt=t9MDFVkk3l-i@r2=o9xH$haRYi
zpG*EzyjKu^Imv<UX}V!=ctu2NbW7?8q`^0n*&usElL~p05b?}8;>$qR7nb#cO}lVG
zLS%AWa$FPv%1*9($&5?yoGzyeM+G)g?N#6QyzFjqus1=oGi>%e{v*EX%TTQBw@dFo
zXX?eEFGcQ3a(#dG?Db6g?n|uAHnPZ~&Q2Wl=;(&wG9415br#z-<YOCg(p9H>T8S@b
z^>H)yGsn3a3&1J^{YA~5u`iz8_{C==bE9<J3IA&$d#MZSd-bX1p8KTgDy1~}+zR(f
zNl33@hF*gIc$@-LE5kkgW!T?+L$<8%fnicnU*_NluyNyNL2K2sbgyeS-^v>3M*qH^
zoVyvQebX{vt9c%)8MpfppF{Kzk3DM}x5ay;{qkXh(&&>(v<8+4@Dx!IycRaI2evb4
zLx=BtDJ8eTl+0g|*cn4hM5)hyaepUmV;!WlYw#tNw+eg<4s-%To3|P%sOC6V`A9AC
zl~uKoR*g;x+f^Q;*{S+i&7sD8hWqQ=;&rYsg6G2#ySRi&f(c}5VCp9ldic&@A9}EQ
z2IpGQ7)UmZw|eH^iU_^hW9r+lkxpl3Zps@_-_fur(Lt^~wFqirh4?iStW}b+mVKWv
zB?*vlUMVOomLb`BC30ZpVj_g{>U4o54rlk!YPDX%Fd<XfzI!|SHm_7{v5cnD24CRA
z;m@5zT<n*l?22`A*X~W5v)#!0DDljIi~u*A{oA-B&x$+Q{Oi;4KS}g2X;yt3_SJXX
zs;bH2O1QZ`*;%IZkJl$iW8f6dh~)eQ&LzwkB(SZH<cL`6f3_7vjbyIQtImdt>f^g`
ztmW7oFHr3~?WTkL`_s{?&y3fMc$}!-yp*&BlIf3*>#R#*Qi}U?armDqeqDQuW6kkh
z3(p#`t-IF949uO=Hb3~JpAP=up#lDb2!_qg5ebL6KIGVpAYYdFXZ7X%>sb$N_HLlB
zs$+!WHY=#tAhCBJDD9n)7pg3;6n65_Rr%hcvHTTijN(`GXP@cb^{6(J+=MB=HMJ0~
z1FjF47oo%mGRtNjBG@JD&0T)vn>thXn=sUQr9QQLApx|B5%x^?kt8JNU{f!RI&?ao
zTnXx*Ga9O!I%T!8{hBhB@`98iUt}oBNH6+i`a5?jWtOB_h3(W^xwuK|A6)eA&jJ#w
z(1*%`I21CJX!MtsUTpGliUDSIDTi2b$5P~=WM*!63FYf+iWgQfrn}E)44y><n*CT(
zDz#o!^a=<=3cG+=+qeJ(hrTRY7g-=!`GOQ=**;QlYmtdee&wr88gEWIbUz8Q62=~P
z;eB;)Vp)6Yidt^tR5N9?J(84wSs>2qnzuVw+o_dRB~@dGvf9dvnmCqsQkFY32-O)s
z-Yt2HDmzw^N15by_0yl5pSd0;CeeF+`&#^F;5~1l#yeHnvaU@wJ}dH=Rl(9Hj+go-
zek>P~Xbkc;Z1H@*%yXt?N~UOiSE^}g*)Zrs_KUBDgk9vL2<`>mba(>od9J&?<6lyj
zeJ3iVUfMm_V`XKN8s(vhY>J7k#i<{4VqN?V?|^h}_6maqk5)rNPfRq!nqZxH!qkeX
z@qPPS>GzR=xVqxmQP0V1fBNzG?Af?GrC0z_tl)?53bT_{UGKxacHPFU{4qEV;$x{W
zLp|;%GtCR#TRu@{AQwl@^;zmDy6pRTYT?zrgzV5)LNH<T!it@P>RhZq(8KeT24oqZ
zeS6*x9(M8PL{xO-dgqUJ6fwLMO?9Quo+^491t0NgalJ2{i<KEEEJbBU$>8dd>En%q
z`i+yX%w%d23<J3m!z(52$uG0-uyc0=bSBuardhKjn0bf9RN}_@EB}ctHzIuB-&WMw
zKR|5V)`o4yzUphDsWmc5l<?;8vc#nYP?0J=s?*)ye;Er4>y!>X7=<7S-ByIHzU<i&
ze-5kstFGFo9Wy?CXE`~U`+PUe@3c8Z!bHqYLt6_yL~y`=up<FXe1<s#1OEOG@?kOo
zgl$dLU+--FcETIU^<nXzn_8pZ@04qk4wDWfg9lc{DM)2T3s*{<`){j&-U7x6NzpXN
zRS6C}G!RK0L1VT3fNPPgtG%*5$r%Tat=i-@pPlT&udB~XmwH|v6ct4X2{Js0b?KB1
zMSUnLp1j%qt@MdrFQ(s7d=SYrPeZa|0!!9?bGwv4CgxE*jW`A}6&Gxu3yVxX1A=4V
zEYK0{0<<8H7Q*Yc^S0jG75y5_5}(LgZv@$oC@!i$XeL-*U#UIy#eS`cErKEIPteu9
zx@$LBgt`q0AduDPMJ4M@Y@VuTwiQ*sPTassRcnR3obfPMD>PVc<b7p=IP&B@gC`<7
zem?+DAtCeZw^rpnSqszg;o9!Re=^8>L#4&|r*)6^=TduCXKGXh5yF>>8|)+(*5`T3
z0Kq2Cx$6iBz}{?&CVyhlHhf9AoN%*uwpitP$^O9os82NLuP1p)DjmC3+==)|{hfmc
zA5+Rm8iZC*sNHX{pmv?QPnf5B$7}m_K{JMX)F+N*if1-pD^<4?pJ;lueV}(VovimJ
zIofEhdeM~G$;2$|izhA6$9_cny|!(<Kl8QzSID;+y0wo{I$u`jD5LaSgsoRQDy-dD
z_2py(lEK9|y?hoX8J2umzDZx$3z9ep8N}b(!+8dUIQWEs-s*V&6{!2d1MPM)F^yY$
z!REH5!OlioS`=NmyNaQputW`>jY(n|Eepp#CM3x;;^l9=A&(=z-Sqw{gztIWy2k51
zMg7ydtv5w>s0eE{f}Zrg$q89QyCv~dtUA_ju4Mp3QsP|g5Y+_@Z#Vs{?ub3Aja(Tm
z|HHt@gnm-r^@>U1S#vK<+(j&f6xNytO*Sjo3N9s1FKco`UdCU*q1w9uhm2!R>NC&5
zQJabhiDi|7>KP?vWiQaE_zKUs!zbza_TS(ehw{l0ETe;rynF)pE3p)-yk_)f$FA{(
zqBPe{c6#|7O8!ob1N-WkPm)^wnfAKQ{#Cw2BL*A4${U<WNOqj>BmTzSu44hma*DKB
z+t)c;AGok0xGh(YymIjhuL(Kh=3C!*p1R6TGs)W#Am3<XFy}fRxVm*c&dqx+f+<0b
zr6hG=iqJ7cEC+qIBL7v2*sC?|Ehoa&c!6FFH4UGK7x+`MB~X$7#4(P*VS`^v>j|m@
z@z^TYku|lnW?#w6%kw&{Q{5N;v)-5@j!fGv1QF@DmI`kJucRyo<nt)*N)C-owvCt1
z=~>a@8Z&pcpsf`1y1P_cRchQ`6FEHmSXQY$M@IG=6PE|Qpe0Tk5Je2Jr`Hg}tCZP~
zvixo<n)_o;S*<98VVYjQ4_<I6$e6M0cLk!{ZtP+!1(1$G^DI_-pS>%t6+nIW*SJsT
z7}|{xMz6YVZJm(KDn1ShV+ONcLtyF94X}9qp}Gv^_YguAMMeRk+ft;rKV;4Yr3W9R
zdr8Ah`%zEJ>VtFw9}v)c3VKjkSxw3)TT{?<B%qelKnxulZ9?({@~cWH;^ss}8syPN
z7IK4w-R~OEe{@}6-AAXK>^|9Re?HpTPxg%Zq!;Jb7jLE6YfW8k_kq3ej0`Y#FZLh4
z{$DN=Mll)~H-Ej`(eZ;T0DT`Oa2`F+mG>OCqP+P)>MrMN)K3izH-)usC`|P>F_{Rd
zrxna;JQ<<UpJD=-*-6fNU;LGK;FhPr@y6=r!-6Lnv9uRdEKy?szgH><RSX-$6VzbS
zaB~w)TdI$M(9~Dp$UyS*CcXvKC|2a*T&u)!3XGdKFBcXu1rr|2;E<!oqhBt>`8`jV
z{6*VQ$wVG*-jrUxx@X<2s7LJY#e|d2JJ|8PF5N^w&UKrRtSd3vY-}Nt1ux&w@0F0s
z9`$g4KX<9bny9OIfZzxZJH8!7XO)Rx23Qq|fM=8seth0Taq9!OgMs=cOks8Rr@SwU
zic%cf#R`in?9;OYnhP;4d<ZCMz#0)M7l|U+*i}Xg-N17??1-yf9jjnBAHG*u$aoey
z{>phNr)0fb95<GS)<^-FHd!3P8CA3i<F(0|@{?QMP-FSu#<sX=l4<s!n6$h;()~>6
zaUK{*JN6M<RojLtJNvfdxeTIK)EapE39Svc;`Yk2m8~ORc7mK9;qD(Fj?taWX$RAp
zm$+w4Js*zpzWAzkr5$Hqqr#o#OPKMTdFnAQ-q6adVniM>`l<eASdrsL=`r6J><a*@
zjRaYXt8ES&Ch)y^3ghpOOq}}AjUB%kZ)&=H|1L2l<wSGVx()b-@~z~REAz86eqWVR
zKJ=;>Q1$T~XIHw=oNEi0x_m1e4l?hFtCzZ3&wqS5Bgwbob6BeHketDBqH;O$%b%7m
z{}d)iD;Fy_lG(#h`{jFI%sShOEE*edX%2&UWM^L;g*$4OP~D6xW(v+&#Y3{fA~hm%
z<b*Kn85%1mJ=RXO-nQCkA-p#klvlW`X-U{GY{p|8WQb+(ynp!)&K#37P~CR2^rn{b
zYF0O`Kq{c2p<(T>Uj-RIHh94;3FrVZu^IuO#NRAZkuPlG;5EA))Zukv=)W^HSTh#9
z1h}Cgg<`yQW7Odj4t%dsIUC#lRMn(}pp~AUMK%LT^xAQ^cdhe3P=Pkm5g@Ak%h8a6
zRf$1Jhy|-tNr@geC3Bx#l6FO9p3kL;<jKg$2bXb(rck?&QFxDQrm<Lju@=jo{uXa{
z=RW?^Mi^Is{dNKhTl~0U{w)Mx67fw;OhmP{$w3lDh2ZnU?LUM0&#`Xb(J;u$kaZ--
zjG=}oJ9wYOnZBnq2I6z(ENW#(neq0$w*JwJFkUu^ey^^+TklTkU3hIk*4db9l(?m&
zu(h$j@#uXdi-mZDMKvkr4eprfPxcw{x^5L1U6q<tHTLVTzJJ}A()m5s<6km!Yc$WS
zPe!;Ll~<ToxG*Q)Sw!T^mcmqr5@^sK>uNax&jo;qt89(zU$WtFh(MGwNUi=Wo4-uq
ziYI941Y+Mv#$C*5M*J2EA}9Bv{(o$JbySqw_r8JB4bmW@ba$h)(jeU>4MTT#OM?uc
zq;$j3$j~X>F{E^NeP8bN-s|VLerwGivt9$<bIy+E+0WjGL6eiZ{4f((41c)?)3Iy>
z1=<tdwBkR0*l~xKKjjyFBG`}Z11<7PyU-UFOs%XFMPIeuF0F699cC5l!(mXWvQ*rD
zyWOEeEXgElQTS>qxy%gPqol5relKvi^86)Vt-|^i5f|2=S9-mHDsJGOXz*Mq#><Qn
zu@Pc2Vlp={8I@uMsJZz)tM>2kL_c9h`Tmda1Oyv^+u;2jb$#HTM?T0ySiJ9UtV{2q
z=t6qHxIF(CWK7aeUNgJERcXT-J^3*)agmVKzM@+F<K3#!R*g2kWj20bOs?&?TsaP0
zfwws%Zl%L3(h7wFZCs^*fGYUc_qVNWPfRUFZwnP4JKX4;rd;D*5?_LDKzI17SGHRp
zQDwp3KU<_}3O~u+o8|JM+g;K&moP_--N@^CL|<Ht;WqF9e)4R;?J5dH7qZLV6x}!@
z{bh(KomZfeg2x>VQQfD#(>+KW>OBDH$%-NR)3SITS_#4K8Fde2#6x!T7~=v;cIN}!
z7~Rk(IX5@=IXyiJ81?WpB%U1ja}=D1qiDF2wX`Fw3h`YY+6Eq;fZuv*KOGtPi7SgS
zLr5Cc*~pJ*(T<?&at#^5jEK^sTtn!jEi9^_ZI=mV4Fm1Xxz<fYy54)jF=&h(rv_k>
z)m`KhH=e8K_qTF%6Ez$9MRF4b$pQ&)FPtnHt?;g`Sk+A{$(@U{Yc@?@-~V7Xw}(JW
z>bI=_uZ7);04XJMGP@bp$Vj|yn)4jc6X6fEI=4+s)ObJ9D+K=HbBTwP@B1(2$VQ(*
z9*4ni_EGP6H<Al0$728|$8F_R0{RC9T?a=-;&#3XpjlCT6Rb3#<fXEwyZYns$EAM!
zwurJa$<eXwe?Vxju1@F))*Jo#OeOh6kOC)JG7h4I93JoSNXq0_u<ItiY?s4k_tZpK
zqYV7;OlFZP&Gs^jm7k_JeWYdFA9mE*4Y)fKBa%+gZz$A?gBYKf_zH-w%F4Zct5LIQ
z&I_R&78A(^Tm7Mc-93HSl$8KFaA6`H_=Cka-slY6T|`%EIBrsI(&r$-u}RJ)pOpHx
z-E<D*1fUb!Y=dkm(C01F!qCGBZ(T(o=J9yqSZ!jg0~-<u4#Cr9uOn{w@DCUQhTV3M
zQshb$Ru5J^%6Uf3@RnmnXE69j2D)R1GPmnevQ}y4-$CgAbqGsG0H4voPeJ<bAKPMn
z1Xp7@?FTd_IlJe8!Nj7Xq9|`)H^L!)Zq>sAaMAJMY5TA5%*`H#X`B;gqpWkwQnf$4
zH*?((J-F%6*8o7L+T_mVmBO#?Sr{UP;9p!E$sV-b3yE2bL+%nY5s@22hGyWQqRMm&
ze*CVpT|&zkr2V#4PmmIQs=D-VoJ3Y|8G+`b3R(2gPQ`m5r+w+jK?cr@gtC*}N{N33
zAcOZCbg(YN|A1&~{6iw%+w||@1Gb1>vj$}dxJE;E45$eJA0lSkSKjkxM)N(O_e)`j
z4i16gav?#L;aT)uS&Ab0y|9DI%&1rBIs6z8`1|cpaDBaa_U@2kseOL*8w!^0+ptLi
zLw@q#m`UXP%of_s9QHb*oQ!}CSe_h2;y%96Fgr0lObCd))9a>O{UzORz*HABK=70Z
zUbC`0?IFv>$MZ!tNsT#QA$R;4$QOq}Y-*zQo>}9cc=PbeUJtB6%Qs`_zjo6^4g#8D
zR=UW%<MWUn&#CHqKC`jeUtsUB2hK=Dgij3}xU`fiGZPv?|1?MnFiu9eipnF<42m{v
zFd=>^ul21XmqbTYKwPDNCuRxXieR{t3H{!FeLe9d^`PH_-_=$oBM!J6p7VGNe@>!Y
z>Zs(j>vx^~{*!IE`{nK{{$q3B{_*Ka)OT^vfmb#U?pV(Qjy5X2eG~eI^}xc!Ww~M8
zFp>#T9B{h-QXM~SfD^XORb`p{kNte*f-^TaXLj0An{!!-blI(*)U&dhPyOVA1~|7E
z%wNBdxV${rf%zX1^vHj`2O7}fNLXf*;acBD;(6z~%BW4-Fx@eLnPnhOC_6t>Vk{aF
ztyLeZYO4hl%Ivh_iG?CZu-E=YN1Va4sr@W&XQW0OoW1vCpP8}tpiO^rAA?W*HQ<kG
zjbK{ZaayGljiiBm8)zJ<)n7vYWd5|$0iq+xq@Ciwn3L>-(GEx>Q4f?|d4MhZ@%rmF
z(ner%vV3#n-U3B1)8Qpn#<3yBhQy~&CF=EUq3-R`&CRBas&KFLvx1<bT{6l!$=(U(
z9jkAFY@3o9eZ+~ZGw_(ft~gw`;e}W20&jA3&r_$}-%MJmiLqheccQL$8s2t)fwHF*
zKh@Nu{o*NDX-%l;#*E%TM`E`Vd>yf&|K-cR@7@5-8ef+;I{vd2z@zv)u>LZV5y0uk
z=6E*tQT|BWD)LA1(^E%O$$5+P5>S`W#iQW87TtxaogFL8RQzwG-(HlxO`V8X`->#}
z{)C4t!b0#PnAhL1?!&aeZy=KzeEJWNX-=o`Tq47R8^2i7H71UG3ebM|o&3~1Y9R%N
zxOPo$OK9GW4g4dP6Xuj3{yZ)Zx80paU`PL1lzQMQsU9GTNU)oMv^2M|Uc9fu^G_gF
zX6&-7VQ3?7d`fgh#fLh=g%=yqLwHg9c?rV@U*=;z`^>O+uuxOCqeGu~qA*E{i$44$
z&)1(*62m*MDk~C(=~Q1C^~4I+m=fpJ*3|r9aD(3>!jWkFxvBs<a^j3yB6tIwNp(z&
zoYcx@UY~G0#@<Gyi4MaVIF%&k!gsAk|1IiQ_=j`;_%-PtAoIZrehR_y*0srm1f_<@
zk-nxYQWW3BD#WeO_~td~O9(?sG6xoWZP7b=%%U`b>xq<<RJ&Yg5Fin4FjaQ+mXGB2
zQbG6}%=(lEh!YVef@9!(5m|BHiizoI*w&Gq8eeMg%bnkpSV}`$xPbUTC~U`ys&yP(
zSXVbcz{shb$-YLzjY5JR{Ou4#OlTv&{q8iBmUywSXj}mn)jh*s?tNG8SN!Km$LDyc
zMATp7^Z${D08Wbw!1oY+ktPfx?A+^`978vfMRD5}#>Ypni(xO#uFG$emoc`~=z0BP
zxYmnvb=0#Eg(G6ZU;ayFk5-y5ncW`T+b_Sj*6S?b{O>XD;`}$vnTsw^OK$DIV25&#
z$8F|QW``=>hTO@AgiA5$n9UW$Z;Aqzu@@YB;T7!{S-Vx_EsU{2>Wb{qDZu_~``sN=
z-FLp6qVc|Tgdd~m&r%0821Nhc6dyP@Ss?vLuUwA#&m;Zin%rJLfccRCd`LaF(_!>>
zK@!S*ljV}vS`P80!>wH`P=UzWKhCNt!E8`s1Fk#SGVGB=v71~8Glj^Eo0u57J5=$u
zPK9Ra@0%|2=!MzRbmrjKUWfKy>K`0uZCYZpo)N2@!~UwTh!N~97j<=qg$?P8j>tNO
z4^MWrO$@8pKa0}!Cu`vq2a|L?zmC3j4m{t8<I`=+7btboFy_zN7d;sFp1k$)=~A^f
zWUFW%knfU*up$2Nb#l|foW}WMYkcw_dS5@f#kxM^hyQ;5_N91uM7iAZ-#L}j<Q|R^
zd>8yi*Vkfn_<^Z^fk|vra5P_ASAZ#sLRFY=1%0Vk-%%`;YsK7UYa*YJ*9n{~Zc<;h
z_3r`;g?q)GCxod*JHSLo`2$)#n!U(W6{r|qE~?#^By?3~!T%oabHm=`cfRK8(M&9d
zaiW8$CKLI|V5MLwg)i!_3>%z+sG2neYR<#ShF<>x?3v8~(@$|YX9axepO2*<suYxR
zA-<`p*d4Msd}hz-JG{Z@&soXL&56qTq_trhTib{JwOkshscUl)#CT|2a;nhAa%S4U
z4%{2xa0a}X#Mr2b36<%#-vyQ?%Yh5oe@ZL|G1wwY1gmBq@0ksNf6|1>eUK}}m|nU+
zrZYC*()z<BVxeR-HK?L_NL0xxj^FX`GRw~ecIZRo{rA)F0|baqKYOJAw>p6m`4E{G
z<(VR|OF;7)8IBH;^2VcS_D{~~;JWiHOc0~cqoV0j$>3S-Y%7Ws4H_5*$ugLTZA_)i
zTRN1=?B1XW-9aoDdpa;#u*K@eemK*AGr>>u=SJf}KdnOp)AI(d)KLKi^u*6T`PXNs
zv-RetsQZ$|?~ZQR2@(?`ahGjElZHOOASvuWuCyL^J$f0<nY_wxR?yq=ExMieYlFJR
ztl<q=-_PfL@A?|PWs(NhQm+-Ir4`A?C7?+4Nsaq(RN*Ea{&%bZi18Zm56@#cOU3?h
z+<;a206DG_WoPTac&O3Cs<bX4AgDrg&x{IIK6ug-gudH*j%VWitcru5IMUU%;xZ;n
zqLu1rFyU+S&nNlsmR*)zSVIou@9dusrU_6@*0b!lM1K_vzp_y0fH?gv37|``y$rb0
zYroJNlp2)sPfi+JK$qnh>FxLarZ>7QYms2zY7}v|d+uK^h_lBe?Aok2c=ekf^Vd}R
z-?B06!Dv8Nelh>P+OJTc7ZMdfR3~GEsXz$*$@Q27aCkbBA*d2WGXe0=?qi>JJ>Kwr
zSNs&7^+?{WmtPwuTcRCaHEWYp`zV#1Q!+Up@(guLQ(h<ICO>M4f9#l7UfJn+-DaI#
zapmc02e^XOQ|OA&(!IFY3KTB;&J!v|u<~m<3%0c~*Fmo&kmu}-2e_aX!|uH8Jh&)y
zLqak@BSaxfc2hQ5&q05P2{+_o_H!b5qK0!vHzOz<FaDI)SerjrW~t?xU<WnG_cxjO
zH%bIv?-8%>119|Y^DFqRg#{fj8qU?@WLQE{PA<5#wA7366CL0h-TvY>XI|I0ZK5&~
z0kQWd-D4P1tzqruznJ7h?&EKd5fD&6_8R2toE7%^A6C8^AV-e^Mw4b~kwxtYM9dUu
z)J+5mofE$T?%6Iq#}cvrCo~ac`WFfqltyp9(n@H|Z;*!Q*0IIPO#&=Ngp~*J)y7Bm
z^ub?aX|R?%xOB{xy@T7<QyVb_+#_e)?rn)czE>8<q9Qsv1pkAJA3yBh|GeNynh(go
zVJ%U!tok-*H~W2DnPC^2nx^J?lWh%fJil^1q!4#)xYm}C^jm>nnxc!8S1RT8NnR7a
zo||=_w*N&gO=n$0(ybVLnoaY_Q-5g<@SpfzZMhlk3TC6N6)YP}J%hs+Hhqc(bwnO1
zH`7dLRFSuQ`6Tm}ol9OBj-;CLrl`>B75dT731PxZ7t^MxfO)#$CsxI26JG-s$$X0~
zTECv2!>l9}JS)UIt9L>%_;2M8e2U^45&m@FhYvG_U9O{IegpqUh*j9#-QBG+T=N0I
zl>pn!z=d~RW#;3wcEElyV?21tSgL5>hVZPT5gF??MKQJQgNOJDlympo<h5r*;<X8V
zMgI<DT%X6Q?xB*p5R8sO)jnib_=}h9oDsB21=6QktZyXj*k6u&d5M$Ke&}7(!227`
z*xoahJ+}+YT(-xLsZt`5xhXZ&-!Mh)ST4VH)f!awqC}5^A1JMrb{8B=$bDr$pq!>c
zLj*7PT2PDIbo)nL=cnvji4_&qzeCvbiZ84vf4X@QAWa2-J40FdH}Cs}>8q3{`(tRR
zbk-2)$&dmXU}`0-4uR6D@7)`K3&msZM_yTJqbd1H?UVWFDb_R7t=ar7oiF>kFeFvG
z0!7BK-<;01voHJCiJ{*RNObspImpSG<tqXK(!*+>q_tk!4@k~7s?Pt6*9Pi7W{QhM
zoUt$dCg2L<iBArkMaKFOC_=&dK$3KV=Q}c13$pARP5W|=((*WK<F=uN(S6z}yizvv
zmtHyl-V8zdh4txFUpf@1G5i-D9$Lks{tIhaR0nxNw}?P}$UMoFn}-JmwCf<CqH<jE
zrvRB2+XG#|2G9kmqhG^ec7oSLXr-=V`=)WKo~T3LMR9Cyv95J)n=HNC9i0vPpGxT)
zTjUKkG;@==CoM+L7)~GIkGlp@ei&I_B@%kNlHmd`qBRfSbdM6LB)2k5inogBPb4wU
zfjaf6vjilO;Z0g6mDJ18Gv0@qMb%--<7}SKm1F-|-%BS!GXhs4{G_?YJC*%KDCwY@
zu7)o>`fB$X9vmuEN`btrvP%BA+lD*%dIXK4p+td{V#wZ^lUt|oh#maNtzl)}HU>my
zVC8UIY&$|ccL=%5{-pIe4_1O+aHt4&Cw>D2F5=U_YgiN4U-Y|C#fKLBceIQn!z7V>
z_a&K%yOKH8THBTc$?n&1i5CxJ7s@*qZYi{$Qo<Z2G>C7v1gG!HAdC8|%5-|hR?(c)
z{DUfMux{@cUk@E}`6<u#GL{IL6`FL9iIRS-Io~AVayJ{ML=1bbT8hq{kOxh<6KZ)A
zA~7l7JK;s2L6{xWd5G}7E_VmX+FQ0q;9)RtmOJ@3`jRceHxG&$%KWFWXexbb32@2(
zZg1<!10Uq(*&FfuS?dl$RZWdSZ#>OZle=?LT{IyF()Ia%u)n|mbh&=pYdz|}02l_4
z5hQ3UJ`;F!9qmzr6f88{f^S`q0Tq*L4Q1FKH-N-UGPNed>jaE{K8t01@!~tvpvYDq
zAq$wFiB^G}`O&4b`dL2mokW0XSy~*1&6~UTAu#@W#w-q>+sV<3gj(-i<4Ba3&Da{R
z-Y*6aNX9R>Ii)|x$oLl7;=9)}=D)s;M0Is1p{vWDL}&aIiY*T98`8Y<9+xs0$KY5U
zqO*@S6q6MSf#pVYz@e)aG0}7_Kb3eL_Us3=jGh*!MkR5JcJUI`#CHFVqPxW6bC(C7
zSjaCe4&_$VfDKv(SvF7pZ2UpPkjUZF+HVm<Ox2ywT+Nl6xIP~;X2;HKPu`I^tq32x
zUMk<rR39{)s))E+?wxZxWly|x_4?8n1>`QV*3rR%I9x<$l<gofnkzhLr1sa4ncm8Y
z<!<lU4$z8M@Xtb(){z+<RIxA^C*xa@27Fn^^lu8I>#HnwoUgoZF)eolP#C1MLM7a_
zl6i;5YO_EkDwv2CHs#J<$hBY1{6|cx^R{E|p!%*>zcIAp%_HAG1I;5utso$=BYyQS
zf}ni21l+R`5*ZnIF+0v_v%nABX)^fQ&KT>pp5?=~4oDL|MOURt>C_9SLA#ftI2xgv
zB-qWYS<%SMi)kaUK2%gIYzIBs`fyZ!zILrJx`#s54$`4z0Y5%aI30@dzD1qZFn}Nm
zqiie-{<`IwTFI-O#4~_2?MC?uYe}Z}?rby2jn9pDU5;0yPSqcOcH@!!JH5KnB1Ne$
z{zr9bSY=-&F}H)bn>p#klCP0%F_odZmfji$7H+D%b22B!n%6FCt|CUyHERbki3OX1
zVKXzI7pDTQ&k#)@PApp2u7oj_`Uj)yuRAmkvzDw4i{Er|MsTA>YC0d9Mk60tItMB7
zND@vhj*3y)(o?9GU`FZdC#MwNc-v3E_)Eb_m#}#ycJb0uxU3UVR`!~cCTi3nH4%I4
z<sQ=(7t35U$O8B;LSZ+Zw`k;T1cRMAu*b;kPUJ!M$l&h_rI))MOGk89uFt|0aZI?Q
zKTn5@IRz25-?8Bcs`56=nV5^lnx~_sKj#v<>OH@irrmL4fQ+)5>g$K*%f<g_n`yp1
z<0@O>b9N?|U8fZqarw<SXeFAj|4$`B;(>9%d?`ouk9TG!egf>$OqH3bnb{`mcVGf1
zGP=OYXBb?+q-3k~gR1b^m8-siR=|%jF-tS^x7&fpNfCKQ8?}<|o}oJolbBhb52m(3
zgoN4gG26L}xz+v@^u@tT?GvYxo?<j{v|v~#k_opajQLx))0WgxnMsN`3fGE_FbdZX
zy^HV5{J^_Z<6E2heQ10`Ov(N%kg>zRYtwE7@Y~b3{Dcv<l^c4DVR9<HJnoucAwga4
z1!)!YHv=Syh~H(D=KTD!L|}M?rtz0tHCm$f7Ta$O@6U{$-6eZ{BxRfFz_e)>d)2M5
z@@Q<lM|MI5-@q*Nv?H&aTgH|~9b+fhm_@&=vf^XLD==(tHYu`5BLUx<n9`-5khm{F
zBD`!l<P~B3a&f@0xe2#F@&v+@hhNHUip0tE+KYW`jPBTb?Lz>)Li%yPzJ-;QzorAy
z77=5w!%7JWDi@hTr#hXx=l8zrZ%yGFM8j_7m$6%yVc=5GNvcq{5wT|`hU?0r{{9P(
zo|_^|@Tz?Sh5*T3^fJl@-aTKyfSGfU#RfSTR?#xX4UN>iz}9+7MHuu&H{5wRN>zSp
z?<-ZiW>z2@*gugC?0Z(~21lsoF|*PZ@2axy72IE4>!|UZq)CFp*0x~%`tc4Kxz^2^
zu9<1w*MB(vR37^A|Iw8Jf*<Oj(Wj!E|5D?Cyhq1<mxKRo%24~ck)a`lRDNey7m;i$
zp!<{tKWNa@-y1G%=0e&Ws~vA@Poug}I=x_fxmZ<okr8d#CJVv#Z@=p-mooKq?V%ZP
z+c21Li*)O@gDxst$P4EQWJ78xYxtl9A3O@UpBs9Zvi;gTt?r8F97<2Okv`7j@_0_y
zQt;)uC5xdd<Y>P|Vyk-c!%d<}q9nu#smQ9xF-?w__Zh+Ah9gz)fnLK`hM93hvWRHm
z<jO3aBL-9caQE|1<eG1SQeQ-iq9+Z?O8NKOj1RA<*{&w<*wZW!l09=KO5ID-ea^EW
zE35qc#Zc0lymFV8SFx{$7{UH1*)8UMg&-Hzp#QihzSIlsI?5(Fu6S69MCBS@)ZSS#
zXcXTc^l3cFsYF`pCy;*l|I$=}F0e^6X$X-A^^IUbJ8HvW&*h0;U+>yY(%oF*;#jNp
z_Ix;;7$DORQAru=OB=eapLxH_V%*CgZRP%0tfgqD>IGVO15uXsh0rN5e?zbriPrXF
z<B99;caVgURk|ss1qb$@%En)O4|2!B&kX+}AAJ0WIm3pxSEt(IYy11r5fLaZZf=BX
zV<^FJ(s2Mhp92XoNfT0O!6GcM+j0cxNaNojv&^lU2>nt)an4CWWSJ{TdXe9kJ?OSe
zl*XE`I-}!Fl5smtYgFtd#>v#K8JF&%n)rTgg73+UVHm%ASDDb@jOP-yY;rl>Tyveh
zLXB;fC?EZ63<?A2_0!59Rdvk}l<;#JKMeAqYi5?`-v-g#LYZ3b==Ww*^KFBqqPpOK
zd)rvmQ<Oq*gMm2LoUVwn!O||ygwkJ*xy!;4$B*iSu)4z&z(z%9<Uv8|g+FZ5s9#oo
zCZRXUAla2`rChQ3`Qj~H^dT2b-eN^R)90o`os<j=Xh;7w99sIa1ihzorhsNbc>gJ(
z3FN}v%TIIZwf~AL#PC^EbvTS)<P(MYo<vzW-u>M>iH(H_|BC*hm?i~V6n5iu_{ljR
zw!m3F2EG`s#%K4o=P11w@9#IM3qF_|N&n5lIWn%ykZr%7{U34%^?}?G7h*;JH|uA3
zVltF2W@$<PP*h)iXt}nWu6UN1_{B~I4p1U}^?pD!>(Zft4XIfZh1AGZ-v*Y1GN!KH
zIF-tONJC=L8mZie(!9iZ@yq?j@r~=$X>6|aR?%d05GY7VNBOkL)AMiuYY?<+0MTzF
z*z0lD+g{ekxy^GuHRI+F&u&#DP+HA*<KTnbDrt{~ka;VB{eDXg`Xvt3f9nixx(3tX
z-0^l#HBRExO2vd6&uOP2XQgzEQ2*-+qru^l!HbtH8n!{BB3{o<+Xxu3r>j2@blSvW
zfBv;;4!Ti!&jMbCWQNpc!UWrMiF!h8Pi;25cJ41)aMtW+yFQwv-+Fs0eaW~T_l|Uo
z?4D}WY?_c6>h-h#7=K6Rhg#L~58BcmxhXGmu2CjtQC=({1y>0YjowuhRd>&PMMwe%
z;#=RDPq~}&Ku^hqU=PW}>5FBWn}4(bd2GhlKr+GQk@@40PTKX^Gm>_Yxd}vYl_`!B
zdMrl*T;~)Np19n0R#Ewa8FH6|&f1jR!Z6_YonhL4CYF6HaZf^f(@V>g*B=pmC#cfq
znr->fnYWS%n^TFDYM$b6Kv!SSO8@WH0$)U4KWO{TkPQC=bZq~K4TSM0$IpN%7KXs2
z8qa1lq6c0$9WX<!&lz}>T@qT&B|~o~v@yS(FS!Xrsx<ou7&oWUm$0UEC{I~8@T`Wx
zL~hV05$<|4MLX>GA-byI=(N5U-*VsDQhm9d;W~(t9ecZ*d7HRiNj)?bx=f;*b=2(@
z!r8vgK}$-Cxx(ymC36r)43A8=O}aK2SFfO+Q6ik1h;KcjwBM$}4L$YIm77V@9-au>
zYLt%OnF$z$SCn^sEwlbcts%Ix)%%F!fJlglGfE?;zp6wx!TyoZz3byIgW#a1V5Z#W
zGQxDs6^CtNM_i_v&$p$MpkEnutEb|FbtH%0YL6}Wf(I|FYrd5Lb=rpJuKcQ9Zu^r*
zj9@FDESpDj;^I)7HcPnI+VFf_OgZQcuq*+_0Q?dm7oDxEt;e}tcWh_dyC3+%RzB?@
zC7pItP<tH9Oh=O535I}iERJexPpM?}4;P$Rz^QD4;Fn++v3q>&Ssk+Yb3JuSxSb+)
zp^ENOn|9Cql<opEJm>(vx4H?Wk^!&0GLuj@A-%J?Km%B%#3fjmEolja<|p1EX-Qel
zPe#hAEKq}h=~|4TT1+tzKrphLO&|aVWi&Ci(h!F$$)g>3YhrH9@8=ki%2slG+`02j
z2UJtO$2$1b#N2kW1O+TOiK5G=S(HLd=c;9m-8Z`!MGQSUyYa%_30CbCZ~HM*Fq*$S
z<##1+$*b;bl~)K_k|rV4(7^bm7fj7%(oe7zxP9ZfWX!i3n>0>}2V8WH&d-#@b*}s6
zt^Sn)xL1fAozUrcKys?=N73h|8z;<kSLKvr9bfB^R0`r$W%G2#y8=?9Y{EhAvy9%G
zC+b8%KDqHNj18Ox3fg{mlZWP6@?zxK>$&id!X)U3I6pe=C6@~AAn4RM4Bz@@KLFo4
z)~!V)&1%if6xwlYqM&`kCQEjQZXT{Qtd>AE1k}s|3cdL~vaDmg8Hy}<Ye%K-x)fEu
z>ZK}Dj<Y#%8U(oX{ZO*m4d?H;-;&G8dbOeXkz_V(Cf4)FElNipL`sOf-%YfH%LzYG
z!>E0fFOU)xF=QiQ?@3^p7WC?jD+4;tq*I$w-%Hk7PEwkwH>>C>&{FRF>Zo41jBt6X
z!%<^cP1z?5Y$NYz8II#z&Cb<2=m;Su(t)W$sgA@^Y!B*JFMg8Z@$$P*3vJmNz-?p*
zC?UX~V+wo`(vG}uE@BEK$mw3xrN6%Xo4in~)F@4%|AUKqyZ9iLb`G-i{Ri~60{w@I
z`I$!M&a>!3H$FV}>-eT8DB*Assc)%NA^GJ8cJ?DTM6S+*d<D){&#Ue)5}C@Sq)@oT
z;@_j<41%<Ah_*1i_3;e2E%^AV#S3b<$r=(FgsT(r2dOmHQ1WI4XEYRUj%&G3hStm<
zYdL4NdoL5W&)=JH!_&&*n+GT891ZLUqnK<gFHFD7PuiV)u|a-&izi>d(ItiEK^{R=
zc|Zdv-9(!BB9V8p<)^&A03DfF&)~ih(F&={W6WCmpuoVyIAy%Uv{C_o?a~|4yfdqW
z4H>T^61ai-9Ho2C3>t~6E~b^K^ZR<E2odl+%O_X{hTCphSsB5|8}D7)Z&=Xx=}oJX
zlXo2F$8;8i2P;DWnx0hDjV5bs8lIV>?>hTMv$Dk5U$_m-wjW;9uI8bbX}N=7KxzDf
zOlg9MvM{0Alr2E+UB<1<qZ$SS%s{r!mb>K}=J_s9_M0H7F<huw_2SG-kd18O%36^}
z^`nvou}_1{HjT`b#QKcOIo^0&y>zRkAuybGr_#DikuFklx2cmqLxfCLm%EfcyzkxP
zB#&_CrN!a8xNcQMkOZ97sK?{zY~z*L5j)qb5f)4}VobR3zE$t_xo7+kCTz`K+BMgG
z%1*t71N{#O8oNnc9ER#nf`WAPO-2!IEc-p*<>xrs$V37T%EPzm47iN}Fp6!sq;EKy
z6BjFZRXQ5Eg>=0*H`p>Y;^(PrBLV2Y-kZ1O${f_Ke~~(C$ls0ltEVOGCnqHj98`F=
z<C-%5JLs_YBM%fc@<+03Vb0q^s?@W($A~Q=t3b`o%&=6jx||g~xT+l#s))i>-j~}E
z&EFgC5qML~Bn)=WAK<$o3>Z>>F2QsgPbHLigmeTXOFHH(S`BgPa&NWP@yU_;Zx0DR
zu6lX*P&_;TnKA@587tM(|F+%h{Vo6bwC*kGnnZJ=m8jM&s~@G{7~`v*=4oV;!q0X3
zn6$^<ik*jZ=q#4~R%x=P*4&Xd3`}27mcB@Jni}p${UjJ_^PYR6rZMZ}T66~ag`V#A
z3oFjjU@}O=^t*LzU9qbGlJg*md8{id%^zX$|I-Ze{OuAfn5<LNrfJ;xwY9Z1K+H=3
zqj(;R_F_o3Rw4X?v|YOE_d*bgb>p3LUSTukRd4k>sOMRO&X7$YX?<kzB&(!Y&j8Sj
zh=GPpa{P-}cX*wtyi!{d&k6dZoyi)&O+%%7qJmz-SkM92rg?%Akvl5}2JQ-&075V)
zOf+{UfLj_8TU}En`ELI5$~?^z`woI6Xc%B9ws-;P76L^y20`xuc&^+>P=dajWU?o|
zWg+DN=Orvpsivu+o>kr`YLXG@kDa-%=BIE>>Jyk#ts<oM(&Uy;(*4}E=Ow7uF&VPz
zh+^y1V034^DP-7{fhSp_6LTo(b|%1mi8i^ZJ1pmZcScQHaaM2{#yuyJ3GkW0>tATW
zJ*sdR1q^n3w;@d%W@@vA#Ov%UBJNrUEa0|L5harzpU##~VD!NoqLA(Tip56lf7_Kr
zCVQZ`hM(ldXh(}^buHYyCrN}tJMx2#7Jzd^^Fn`6R%1<1*_eUvK0&zt%pNZQZh5Ud
zrmVW+qkWvQqH|th2WW8s>?6ulE33s+^LT&tg@N%I%jRE;MgMb{l%cq|B68~(&*7{o
z5hL>bLfY(T{N$dxcQLOAN?#>TOlSW$9Tyt|`nh1W?VNjOW$+KObCYd^n}Dy<+{M#s
z8kvyt_a+%})f{(Rv~b~2e?5whLBhTLb<(cNO&-?4?OV^o^$Syo4&`9{XadCDq<)of
ze>3QuD(YQXmcML6W(Mg(Qs`JOA%`xeL_`DIPEnEkxCd!)!M4N18GM`Hfu%I}PBF)A
z@p6d(<H>+hYFEM;TeRvrv?oi9x|f^wdX$0uvC=t>tjc=D?nAB6qr-91Ed=+|e8S6&
zriRgG!tHZmh{G?bNVd7D%Z%0^h0th``pe;g_r>;41KH*{==7`%OD9S=KdzG-&d&qc
zB;vVsPy=<`%(teAi5J9+x130dI`5<^moaEL$<ZpENr37raz!Mv%$E(Pa;7a+8z=?!
zCe~Vg1D#UIwWn7v(yEOz7#84S(a&(-iUA)jSW9<Q#M6wet}BpDae(c|3ak2LX6Bze
zgUx$gWR=FkdMSs`vPs^mIH$|5v79+?L$Ps9Wejf4@(TQ(dKX20W}Osq++V)SeQ{~?
zT6NOr5T9?~aH<Y|yNd3L=DGzv^QxVc!yPj?WVm;RbYn;*6S#?20%Q5>*U%P;sUjh$
zo8*Sw=FE!lB7dY~`l<8WxUu8Z^??4g_n@rG_nr?yIXQP=+X3Twlk|UC!Pk0w0Aav(
zj@f$iTY%8oj#yPyr3cK9lax}rgXy}b@;cITTF)WK$;n~owVp`o={<<h8a|oR^Us5y
zb0P@n$i!EDM4ZXS-y<ky%}=U}I;%5h0T)}RjL8~6sDVeBw>6wNR<k}$HvA<o+9Dgb
zRNCiGxxqn5!-+=*Uv~|Vdn*-V4iD`<8qsAkknerDxF7<NEO6x^!cZpPqfqr`pV<})
z#DoiYd%`8~^b+n8U{XCZ66vC(G1i$q#`A*|dAIo9*a*ewB9dz?Q+SIVMAY%3VO5Js
z93GFjlh5GhD8`B9k9FJ5i%(CuVWwE92(MDrgrwq;TMuXG8)wl3I+xTx$uGSxrC$U#
z{%u6_3JXexoj*=nevLgxLig>y(C12+wxdwT{5i4sfV+I0)b$TwnHd10Wk}E|SUp`w
zdDLeg&?d+_TumH+*w<<NzJCs&zKB9$4=*okj>EUM9O}z9PorxGjSvieQ1Vhrd_Rr_
z_2jdELg}(uk|O-E4AxFVKb=whqHX@gPBc}fF!S8-J8k{?oMO~`+;JBkI`sU6FpzLa
z1HgDhf|8o&HC0I&Tag%8<{GdKPZcRSYD*q_PbCFb(TW+p*nowZ7__l;bctxME3Y!a
z*_M*=1j$9w82=2wuu>MrSw|zwRnq6nD|y^9n}9Z4LZ85EJoiQ_N5|3vs#|$XeuO=H
ztN-po$`UTfwC-FP%op7?Qc;#aZ!3#{kraXUBl2-Guc!L72PxTFS}R|j*p`=@mo&E*
z8B3;8kH#*`n|9FEve^r4yGOD7Wt%2o3nn4{_1kmID%0UdTH_$b^1oEMDa$_i+`aG$
z?<r8IxCO*B5t8=e7?2J}r!EDYgGL3p<hh~Rl}}Viz_5`?v)*i3$T<~F(CGBc42agA
zq@u-DG;tLt^?qvcmC=<F2!uSsUt3%A^{U-^)yh?YVa5(w^6g3KVqjb?weIY-Ovu3z
z5RJ=gLs;`YfynZmJFMuQUR<WfivUc$Pbj$U>JpKhd-GPz(#lE^!+B>znR?;A{HAwi
z0#Sjh(N~}4*fD*Us5koMZFs}WnQBMUFx$+J1>PRDq2U{6H@kf~-SU?2u*pQ*m8Bes
z=LfmsDYmzR0Y0&qUpxdwBv;_^;xO0d$=YQn5_sxbFyN)!xExHiDFid_{%cQ;0M@Q8
z`NNV;Uz2j)vQ;|2$h-(yM!5d1Ox~m#K>1S|bOK0&xSJ@9{r@%LAn~tNRXFA4<=+bm
z6oYGwdaxRgdg;nW+-%KEZP`e%m{lGs$Up9Mk_nma4(p`~>9Q_#;;U}N`T)APlC13w
zb9RGp3GX7FwR))_*G<qfNf!`Jk)>GOn?MO|TT-XSY??T?Q=-LIDsw-yr26G2J1(1o
zjx^&C_l%D3uHpK|6dx`f*4;|#vD|eu{g#~$Ra7UNM((&zHa{5MD7hZ4CNX0xrF+5G
z0NJ0${`IMWJ=<B!F}+=jI>Rg~*n<I!Oy@B?Qw!s)Rf7v=@Z>ECyr33sG+tY~SsO6=
zGrn1I*_3B8q{&7g(2wknAv#p|=85>?(N}|{;lT6vqbaOAT89|2B$6>*Lte%tK^=0c
zwUZMwemm9o!Tw~`4Hgk$UD<C+e=w*;Y%=m=0tq_ADhp=_F^`A5JN0S@{(yu+C*C}2
zk`xcv0k+c53+GAsHTU)!jA0)Vu%dn!V<R0mca`V<5pUaD?O4WKY^w{{iL*V+%XjRO
zDtD<}M#0`)_d}2C)Jb8R+jDT;^ZN;N;j#W`0;D_e?1P|6Do4AZzCxQ;Prg8R@G}3|
zPBQ35ZdGzh4XHjYZN7IX4QhB@QJ#O|Pz$V{mwJM>&ss`dMN<kifNf3<GCOn9Lr}75
zfSfLn?C-sv@p)Hgdq2E|QdIwYlU35Lj;Te*qCL`~L4#ouX!Dxdj93^P_);GBb8IAx
zYNX<DzYS}8eFUNa<jyFtfr$h|4cAdu{^T!F@vLjp_kEk8lB8St)}_^`F2|COb$GK^
z8QTL(sl*Md3Z;^ELr?enR&v5Xhr-AnyH<y=)$-(>-F*8-had|wVAbABK=w!oON|;;
z!Q7#9+P&cFzeFPyI>jlSZn4x|J1?4^o}z|5T?-%*huSCO?`z|;IkWq5R15*slV?*W
zf%Tjf5UiE@2==fp<4HBbBhKuDqgw!UQ4~ByXT0+mp{x9aX-`}`W%)V(Ve`NN!k6z0
zOfKogy-?q3!gEMS;ho>odvBkFYOAEiA+AJ<?H2SEhe56inNvsQD2Jw{j9RNe4U`QM
z%63g4DQq&8TU~VfhA7eOjYW0l`FiCfvhi`2%>~hp_N~K_DuHKK^|LOg=Flu5ny3Lr
z;bfO>D$o&{64|0%GiGp$wqLA&Mz7(kQguM#KxTgaY}azfqof%nSTuCZ&+6#n5V_gU
z`IodKAeO;XfE;rW-)=J5wP?1wwBKCaFRilGTV&~zYxQdc7nKHp%5{*AmaVl>&%eG-
zTajX8UlG>jAHP@r*>}ghcHH4seHnLa%HyqYX4qTrBAvSQsi==>h4P<z4?sWRivsi{
z>MIYOFaH2sy(I+8#b%PqPXm6%#SDVCCrW?-!Fh>skS@lD;6bG9JkP%36o{=osg5#f
zZ^FH6&ymTlZ2}oI;ErSigWhtV^n4&TcV<NXaED7NSg%NP>l#{2%SA%Nbb}U|Czbuu
z;wrol=7<<v)rYEJ>luk}{XCcdmeKP1mywFofr+IJlN{}bCd&IuvI7({_a6Dd$w|GT
z?fdW*8r_~1obp)%lB7`V7LV}wmN|;^>&T2y8o4`dns+lDPP7cqQt#Py_Xs!{3Es;!
zGq~p0e96q3I?P;Vr<(f}n|{Ohgi6=G2cJoYJNm@Ckv`DfAiU8Pvva<5n=?;<@doPz
zdTi&}eVc{|sJHf=OnZ%T9tMAWZRR5-#NM8-{PfLkTZO>x?MLdMu9E?=kz~6hMxEWZ
zH)1Gw&c=fvpmb?$o5jlq-R1xOqTZ9wJvNFB?7-O93+>(U2;IqT)mP&XB(xT>WDI4#
zm1fwRc74lRfcS3W3(X<^#`HB_`|=gt7*HMQ=;>EkiI`5ko3pJ)a=wqH6f?{m+wOY5
z!aM;zjg=V(xA?rdt~gj+RyyJq+u;=MBH2rgWu?Da^O+N}Nak9o6D^wUqh+yaSpNKl
ze+=0>1A`7&3GmGG8|#h7o9r7#h77Y<nN3nw8=9JjBj2(VAyhTjQ2)cE@)%xg`x!&?
zb>`cH9T~C%@E}}8ZBFJ{)vG8J?>!#T#*0jK{j1+2OxAaHU_SZB25u)=QO!5M^vgs~
zsMwn2!+~pG==N&lmVV*RWO>Ma@`yS)4J!m=<V4mx@f2ASj|s(^76FZu)yQK8koVz$
zV|txNrhq`Mad(DW5vf1SMRBES^P^Lz$afwuv4}wi>N~VAtTcc~(;ardyMe)I+qA}1
zlF=Y`Ngj-2GjNR{LL5NkweO9i;7I?>?7m*R{`+(TY8PZyD6r6b?#(flv0i1f8-88|
zXRn&L_Hl)S&gfmthiO=;?b2e4RrD!s`{1>FqQh>Os_~B(JGUD$M{1>?2|}9IJg0ck
zl{UuZz0~Y*WbHHQ&372cTAjyGllmK~*Z8BXy&ebhEynLYmvh62u@9I9U+i-PCY?ck
z;tl=yLe>eIA8AA~vQz?V|J4Io=YO54Uc4Zk_>YYN`UznE3?Q3E7Dxvg7XA|EzSjE)
zx8Qy6`TETpa$4Hn@5`UcUh7Feba07ZxEpE+FZSNx3-5K+6m<<T2~mq2Vptmuc_K`{
zJyhI10=NTVl%$L6B!>?c6E8LlyGE{oYo=V^x$Y7zwoz$N1X&dDMrNqqM9<V&ZjJ8i
ztroc^HI;$hNO>O1QD1B#yY2vySghZMtZv73p5$}+C)|<y{KhHlCzWQNI8ekG-A*d+
zKvPvq?ea~=7qHj0*VXq44Noa`0`mdelOUdwcY4C7Q%{B~F_P_Lb#_$EE3Ld6hL7ge
z_$1}tP1_6CeG;pE;g*}8o+AhCnA0;zn~p5u`nGGReJpQbQ$;P0WWT3s_hJ)K%xHsv
zG#n^D`Bb*#5~l#`2|uLcT$GXlF}zp6`BF}&b-FiyVSHqCN-cSDYB|Mg+FqY4$UlLb
zU~pLC&pKf0jJGNNVg_1~HM7Bpg4$MbF<fr|see942Zj~kPS)<1e2G9@AXM1f%>Bd<
zxqhh#5I=X$R9*UBW%tH0$iCAn*Odoi4I(r4<uNYo1t#ZNC4{;u3BhZ@QN7y2ji(E%
z5W!GbK==N5zgCD-#{_QbUhIv@oh|)zA%tPz@c0K`+=@fHZx_*^!SLYIijxA2GEWu(
zh2Z=GZA@riaMbdt(>K_?crYw98KhR7goD@RsUaUBp@>Ufgfk%%<!^#6>I2v=sBhxq
zS~?6y+trU?BH{P}019OMOnfh0-bN#gsj95}jOH8f(oxnzY4GtOO|zhd|64~+W6;{&
z{x+af+pCFXVlFUz@`ocp^<qFH8*Z_=T@mqMeKs>2&m&v#vPwr89;)$Z9woLxsg07h
zK}uEwr^eh6ofWMYT=x_gT?<ar$jH9x6O-Pe@9UQ>iCxEaF}@V>Z=Zd__G`=nJRRgR
zj7`gi>#!f(V8tWaV6Yt~<7`E?c#g?>UwFe6BG=9ll6p>Vxx?{luX0?ByX=f~EPf~M
zNlkkL*A}yKH9(@(8`B&qsq~h%ec}&2AduVSuf@wN7Vqr5kGXDwmI%*zkQZUUYL&9V
zz0{1k8q)3Y*ScM1-m4i0oe<{>zJK6Cv~*o{7)k9HTbz~9dgZ8%C;sWtb$;ES@u#UJ
zJ($`X?sV~gOzm%j`zrO&Y<D_p09)HLI!TVQx~%ZW-}9%#fSL^|!|7&97l>PChQMB!
zg~FPLF|C!?KI!#UItzy^YdsM+xXQSHd>)hK-p+okYqQiU(YIQzKRF-T9-R0?Bhj?!
zNCp#Ho%EhWC~L?7avbS6RecoA-MnH!(4Q?-B6@-9d``_wZig+*Mt)&W#)8x{<c%;1
z`;r7MX2AP=_Ct|a=f1>}X)m3MWj#u03(^(3Pb8bPv|)_?mEWCgN5GS%(%)UD%PAmP
zzAA?yahq}JYD74@(LqS`K6LlX<Q?)NOB;{qbX8Y;2C2&#0x|jsxxD(kkIldwF5_Qa
zTq?ad*AXb9zuKpyE+$So<T{x>C3(YOg3^S+%y@aEm+M87SR)JwsBL|RS^5^nIqRxF
z1MwM$XGNOsf2L^MP#QA@P+MnLQrz)ykY>$mxwYR{?|3xE$534FmV5@1RzW1R2^E2`
zhAgJx0>*aGUX6$wj)dx+@F|`~cK1#pb{KTFpdBAUaQ>RytaH4%*Svo=k(OjlJX|qf
zX&JL*oNS=DSnW%?zq4IbPfZ3Pw(lE&uPB|r3+<%jFFSs8rlLZ(=Q`Ti$NIg%g|CL9
zJwPrx)F&s9y54ouR_Eg)wOAB^iTRpiOS%9-oCo`DfdJ!214#2hqZa?%Vx=REx7_pI
zXxu(QWGx7Wq0rq_(O!sH#|HDO_Y}{;m$Tg5+{o~fqr~rW2y`9DgSJ!Hw9Gm;xFNED
zoxthV4}wq=>bMW;fUNT<Gh+!e+cv0B!1a;)<Qx)tnEeTg*p1DE<U1WXadWR;+SRJ0
zbK>-}M+%eoxGJ2>b+^6O`IY!3HiICB4fB}8jQbz?MfZ}NEL=_GmnW}Fx0XE>O)25L
z2b&c_T0A4^6AF3@PnVyq-dylU?Rsd|kg#0y1iTSKQ&ZcjTKuRK@l<!te+xT*M8M{T
zq)mXSc@{Vl%mx#cWo1>(mf%R`XWV5@<mPg;PcD1TJ<+3=Xts{4x&mJXZ4(WpgSKT_
zQiU5#wucP2TjISiG+~3U;o!dvp>y<at+)Tt;iLBm5czRE`s|AM2cZinZ{Z*6)Va~o
zzsCD6>AcrQTkcQ1Esx!He%b}4-rWR)5K(rA3-g~8+f!J71jCA~cE_~L!;WK)GCK+O
z>bY71Z5%S)?#r$2-}R1&*<Bzcu|<D2r7_)pACt8{^Wozf4I^LMIQ?acGfR6%lZ36a
zrI6<U?Pg}Kxy20Ie%s)~Gl%hvYGL{shZqg|b<7t}-Bn9TtDVd15KWLOvhigtCtEHV
z8ndt+k9b}6;A;e1-jc51Wkdimi;S4x({YAr;Z&f8-#y8a!nF4^`RHQ1%+^V%*8W@b
zbDBQy=vV=v9>V~luUR&F(HEh3?j}c-!&<rwLuqFxd<ZE5a-N<5=)TG=<#%T#MX}OB
zb>gSF?w`fo(|V8Sg@p2P@k^(cz_US(LF0AMkcO(7Uk6`8Md}EFwYl(68mwwi8~;Uw
zF``0R8q8~<)HbNWEi5X0x^pE5vcJLZY6Y`<v9gE6M|Ze<f265=?25pWHudT1a;YiX
zw#ms}GbJT_ZMUtyx#rC>4UR#1#^`5Taa#w171#_#7l-XmU3hVKQAvUUce&@JZ_0(W
zg!-Gk57Ttf@2#C3j;kj<-`{z@C)hJ(igEK?Vs7LyD~f`Gse7)L2A^J(pHwwQfu7>5
zX}MQaSJm{4)Q%5qtikpvBcV5#guyJkEIH`u0->Q@T3_BCPgCwJex2*;;543OgvgDA
zj1KrN{yl0YA=bYcI?oKNmyDBbcx}jl<}tj7!bSfaQFRF=&(j*T+!V3}NX=S6L3}y<
zx^i{8lLRH~$5Ppz&2RUx8BI!{26{BltdbT7xl}UGvX3Z`GRmg$oh|m~=j>{W-wuAt
zkv7+z#j;{oArYCGdG?(~S@c7RS2_OjFUQO5M^!qe!n^s>RfwZk?Eh#1*e9KPq&?Kt
z%ShaXt|R$7`5c}vuX{e?+xW)2n?!Q?!d*yN>ntep4!3jgwjl~S7nn;(tmaH6OLnV6
zN-_6ncRu)>y0a!6&nP3^MlpT1YyJxJNfq%eU`&E8XVlbe0@V9&X`*(N+xsR5589mz
zyS~sM=F0pLZ@!`fKx@<g)$E@wJ})1$Q335(d!0rwqzy}~k5z$EK-9s2o>oblOR&rZ
zF!p9!Phvq4^^=@ZKJFf0KT&!UROx8Dugpf$0TNc5oUIXk)P^a?x3BYT+)Gj7)ADxZ
z?wYY&ZjK}-tO!rla0+6!)mK#m?z%fqdety5Aj^FwHlSIAEpD&Pp_s@ZzS&&ti71y|
zhy~3j+yY`G-H;%>=_k~WDw8YQIBS8Gf6q*k-(L?Z&*ZzhIquf3{}{B-JBB=58OOK{
z9%-pq@lBE_p`)L`g4Qi(<sTwAVxAZG^+`?QCK?>diuLD)z%pk$EI*K5Itt^|$N_aw
zT<pdUe5EssjKw6DY)q}`GwQu4S+=we(BVN@^=1OLf~xEJ8!j<`vYj=%5=l&&cYH*X
zN1Gi2lCgW0i;C-q61aUi!P)3?z`VediQugh7TxKeC7RDSaKSq|N1wJ*lTN=c<#IIV
z9JkEk9YD=BU8kGSyh$Wy@A!PgWPyq`>MJ*!>RHlck$Q5Kt9{1x{1jox?}>9eww1z|
z_zRU}WGFWUb8+M6M_F0yo(v-Z@3&%e4;QtdFL1>oVp}gRN!0<GEcL05T$i9e#dVdu
zzOkm+!<vnU=Y+DEX1XPe2(~b-h$56M+g!t260;A+sg96I)H-26=g-imcwJ7G_9i$M
zqk#9U%}2Kka{nosc4vyUW<1FtEOy74^KmQfKFGP-!LbpKDZMLZKSSbOgFE<F($DN*
zhod^s8Oq4S9om|!XU2*!3Lf`a4db;kAp7b*K9VCyBHK`9C5_m0o1G*To=>T?3;DYo
zH;yZIX1sH#^{4+nei-&~dCJ4~CmjGZfnfu^?nDGjs4yfG+IvYRFsz2jU7i)&0FvUp
zSI4Gxw|**VCd75r`7HTB2=HDtO<w`;bysY|^0ALV_WGvT;R^<Iz@_GV5Ckuz9{6;3
z5cIgscyGFy2Z|`czuR~nJBw;V^F%o9aAJZo&^q7STZdENZ0`+z?UHS)B8{P%^A=nU
zQxjR{YJsLa_otgR3Kuwut;Uahx~h>@Oq&uYNjLp`2;+7^qkF2J2kf~?T0v!umb1<{
zT_0#yryzg#kS!*Wo5pY4{n`#+w-ZCbVX@+G^uGn>9$OW7v91m264P}q>*BoAtZJYW
z{6Jdj6x5(~w;heU;tLeSg5X!y4p}Os3(T>N6}^*QwS&U^XLBubUkytjG{nmQFNpGz
z2J+?0%lS#SQExhQl8&Pj3RXB}O;QbG#nP(#V=-5?;)l0%o|NtgaQakmhr!cq&rLS8
zI}|H(zn_f2k4c1RfsCATe;8D$R%J*wLR*Nrw-pf`hmE<pUeqiUVn4{9(e+ele=-{1
z`UC%Z^cZ(g$ta=H)skz@jccRuTlc%YTLWxHutnlxSE)8~i>A$PUm1J7z4L>9yt!gf
z9lgBm^qjJ_O&Ei9UIt-gWCBohw53?x*a(_2^|@YxtA|k2K}-fEy2CmJA9-rOy+Xgf
zE5?M%XcvXRKBk^sCbejsOI!pDiMLrbIR^ndPa#yz2R$t>E;-thvL;FCsR8uTEyq?y
zp4ag=_$<foMF-uGUL^K<o?1F$p50!PhQt;Q3Svs$?B;>aNOG_{FDp&TbVFcv*Nj2o
zEkI)II9}?xo7!B)c4|*2Hvvn!Cr&iu<x~GgNk@>FnA@b2;+DZG^xs*<w|wBJoX`)h
z{w1s*zUZw!XwGlV-RK+V!7}SqyU#Rv!(m!ah27(If5>f@H#M~v;6OuJ^PQZ`x5H<!
zFQ*FDP@Mg#;|Svbk+o+thKvkelx4}@Rrw&~iL#;ufG1WUS_@4zB|jAdm5Bv>28I?G
zw3kKm$5!+1(9uYnD;C3?Jp38Fn<WMt%PhS7d(Z(N(BzM%s<7;EEfuaX0))8R6lV-p
z3!UhzTf0M|4=p;3vPyHh3QEzXHL109bBE}Ht>3QqlU5cNIuKX91RDH3M0e`I#QhiN
zrF>4Ank;i`N4m~MaN*w5Oaw?Qa}|e%4!z(HCXo0hbCx;aL_y)m&`;-eu06!qIhyBV
zWE{I&4|Ziyoa=5HW)~Gu@-VA5nKGUxj6>{a?<<{2<(0KG@$cScBT7nSt3N`frk#Hh
z(pge|?l=EhJT5u8JyA5PGj=t_V29%j%l#R$N;?|0U!eIrOq>MDg9#q{J}t;cOd9tI
zcVtGkyq*0W9_Rks{TBa&*Bmm`2{SVUpEQb4^wH4T`P&iRhmO5FEiV5#k2qs4Elp0Q
ziviXC5gII<GG?sxv-87df~f3wh|Xmw>_e~8or{{=O@LsbxVPM^EYNwIBdIMaAK!&I
z0-CMzX7Yh2ccM+$2Kl5dvk2RMeS+8NN{J2BI0k>v>hltcYW%!%3cl<XO4ke>_W#lK
z)?ra@-TSyANVjx{bc1vwAPs_aBPBI3bmvgg-5@314H5%_q#)8Y)F9n3<Zn3d`#JA9
z-|PDM7jEVP_VcW@?|ZF#t+l}<O4*P(FK3@J8^zrr?c@`sy+aJ5NXY0uQksyP14QgA
zy@v!4G@^#6oA{7?(5z6GhTG^&J@uaQ?0wnzeb{?A_r9+6Ydt?Q6%QbT7T7Fo=xAxy
z6yT^ia%XG()FwTP#zkdQ3W0WXFKUf@gzuG7QL`v}I#EjWBaCe8943x`_|p|({u<$J
zAPUZNT1mu+o^V=8^Hyxx-0%Nbe1=PnkP<5Cs3T}H`!aFpRMSA=MGZ0Z=;o-etW6qK
zWgBDm5ZN|kx(^yK8xTJFAvO{sN<R)jcuvq65%ac3&En)Pcg^M3{X_^C&^74<zeS+g
zBlxH?<!8fJgKgg$2i?!CwAO;#Jf#JL;#!*93fk8<^*m+kt$KzR3zx(4pX0NDkoyK!
z9^=xr(Jwzd0p1dvDL0eZm|vH_VAk{S0&M5wP!_uqJ^Dlanh$O_VV?0R3h;d9d+zML
zbP@n5kDlu+;kO6ai5YyYMT)^iYg4cB`c!92!{^U^@#jZWmoS<9Xmn#k-0a}pc0MTw
zWD%~y{A_Zqvb@Ek0rS2L)A8QPS_W;Owq8|}mUrat;t>Pak_A<Henz7}Cp64$QD%te
zktLcOH}?`SlGl*PsWpT6$9VLj2+1FdrWP{zE#JvPj{^c=?#IlvzIB5|4AxE~v;1n~
zb+66XtbyMRv6hjynyG1=Q@q@;z7PA+QACQx%@=_l4w^;XW8dR+q-zb|)4J)A>73U%
zrC5vM|4Jw3AZA<-!(Os(2K1z<kiD{9M(<GT-_s`*MJJ0JC+c)s7WBlc$$fp|=Gf`U
zy{nc%>y2G=*KK7Dm0UgHdv9!cV%p{X2ITC}rAQ@+Uxeh8A_8v1TzO`cKwM4zY`xiS
zS-#L#K-OB20j;A^?+^oeuV8E~_-F2ha16rlDdzfFw`{v!tnY<?FMA{}IlfaH7F*j0
zd6w59{GFw$%r^l4P-NJ0I=ju5Mqtl_m5RMF77;WJs6KY9$PYnCT0IV_#MAdAWoP;S
zV8Hj{xyK}$AP5$6y}`xw{l<}k8g=%C&ogznh((VPe(R24?Jf@wZ|ME*IhM6&+n!E6
z6Epdftv;FMli>!{Gr(=Oo<Vg3_b(g%dtLD@bqlMZ+BTYXzCk1{VmjbYQPFf+LNIal
zG1q_t*%mJNc`%bb2HsKF$+$OD1SBxjAkBFv%YVrCEGqA)`6Jb6mN%g|fZ3!s8^Qs_
z1KaMBBzq>GCx6V&yka)zpoIF>;+2e7ezs*!J8TEq;3E+E^$yzbZ=$z;zQN+$pc+cg
z;&%ZZjS~EPV_hv7C9f?f%)yZpX5ce>jE~uY6$XW#<(Kr5N30y+K*O{!t~f+j9r!<#
zUi>P0wO_Vg4M2rY`1%U8{<c&8yS&%IR6whu8a}(Fto1`Ce#-3|4^R@yLM!q3B4+cO
z0qKOh4WgC{J9TW}a?v<ITBxRj^F7y%YrHas$Jgk%eiBu|0N5YkxsdzI8lKhD*;^5x
z!eSEELKY6>%Lf?eHyS*!Ub<OZ8kuf}7KG{H`UgqIH9Z&c;5y|QWeLAsXxSL4(iay0
z0NBa301sNy1Zwa&XC{A7<2IJ(aUwSnV#GKDNcYbH8W{z}<V(iSdATWJ8ND>R6#j8}
zWL73bXUp07#U(f6cWu1c^rgN^LDqUu8B0=5<_B&^3JJ!2x6I6)^!R6QmWMBV_2)XS
zx;v;W=7{>vigOG2#Yva6+zpflCZ-lUdAh7PjLo^TZ{sn4audE?4I3Z>s;o5UYGIUC
zzs4Uap?^#JjvjPoX5K1HS${TB$O#$UjXU8NgAawJ40!eZLjT7&5x>7}PWM5c4o0a4
zxc2Au8JWa}3QurLOz*}qooKb(L1}@rkM@*S*keuODdX;xx8eK+300xgVh!YYQn}OH
zpGqnV@%fj66xOAeHDS$CIih<RIVqs{7pog9NRSZ@x8=LFX+N6sGP~}xyM8Yk0&o}l
z#MJr+eb{@-ZG5s5jNZYc(<WeTO^|wk7Yx#ODJ-n^KU;n>;8QwDbHQQEf13EyvZAu&
zKAO(z6ym`Wi-{0EgQ4(A!y1OQ)++9Szaf$z>mn^Tev8Aj1)``#7Wx}Z6}`oe44~VD
zNC={JD`f!1Tlk$(^DrGXb6hTK(cDnu^#D8WJ3GN!zZV1_v@h8oLpt)<^I}SIXEnfc
z<b>lZwGOK~X8W{A)NMPvpn=xTZK8+asWGB&ys1)2BH)z$q4814pHV_6^E9oZwXmj~
z=2r`HR$=bf)W^@N9qtGgzBidp<s*^0QX*3DDit3je(7_*3)grT8yg>Ua=bt?@6G?`
z3Uz64EcBCG*|kg^8Gko=z{F<2bW7c!yhK%yR8AjI6mt0?``!VK?;@rbG1;Bs23%`w
zp$A>>WWvl1lP*LT3<^|&v=Ut<KZ@tKAvNE$X$$f6cKK|a1U#O!+6gqvip5#Eyd3^a
znPS`EV#0V(5ZtutEbwyUMOfGGNp&ha-zBpVMPDKsaRv}sfWFP%Y%S$>n}}4>!3bX6
zfYCOFM`9!-5%wm$qdBajEMebwe@}qehXo}K4-W<}PWn>!sa*L3HBb5IfP|O`Sp#SI
z+D!v}x}9C`VBaghJp-8r{ieGSYoc25Xvd}RcNM|9pvV^O>sGM3k<p}QJ@H;vLQ>2;
z+(}3Ye;B}~`3EG79l3s`rXf#oe!JH=xxV=BFg-ORjc)M*qq*XR5`{xCj3oCC&!3T#
zmKIiQ|G6g|wf^F6x)-Yfp33MV#cx%>!krCplr{H`SLc{FHo@Lo@Xjyzmg@pm(YURc
z$la(o(=jV5>jhc`R=pu5n|@V2^c-%V8_Wbm+O@ZIChQ+^53%(qzGJrrEMc_<_S+PQ
z+AJ$1t+jqsNLf7U_}m%`W(*ViKKTm`(`c^Yy?>;x*8?TcL4?DBWME@PzzN~1fULl-
zFL&__c%Jeow!NTjxIo)F{QHa;v-EJ3B=$fCAuqWU7v;u7Fp(<k*lxjpxF4{u#L@>7
zP?$tn5JiieB!a=@4!UJFE`Q~JVIf##fo9+HW732NNWNeCVDr0iE!Jjc3^`xl?a~R<
z2hH&@a=M+6pR3D*k*9l(_;CS27QuUoRX}%bVQH-Uz#-m%Q*o7pg+7NK7MtBV=w&dJ
zh^igC<prUn(ggX-)bXzy!WauTY(!_xKCn$f4oC8+G*l0Ii^{isU6S+O3r!gj3`X0M
zP5N%d2lIE1`;V~`8WWArMM3xIV!W4kMu4RWix6kGtB*F*T-!nx2cjr%?1Y9FBA>8}
zkUy{7^I`uRKs1&xH)C71D$W*vPpeHV3xVM;$oel%G5I0TgIN7ars@6*W&d6OiihaZ
z(RcA_5Am1cHrgN|PhvDg9AY;ds&Wub1A=?`=m0<hiVKDTxwEO|uG6C-MV9ZMm)Y!Z
z$6DTX2E~XijOn8m-nf{h)PMBxh6=|{kcK5Q*4;HiOB*H|S3~hcYK)eTS%Y7j3$kVq
zS8t93>{TbBv1>!;X7zu3{O-v}-6-_e$TJEJ*J?BYkJI8S)s9)!$GJBWHajewg=hR?
zm_!e`K?b)oa2`rwRfB~8t8R9Gr`;2o&B*t(%Op>3sVCtmd`;zxx@ryVC+8Z^sliw6
zBw^bvhvBe@itY97=HfegV4j_X(xjhiHc|0FgQC6g5za^H_h=6y%n>GYxaTPXOZF@5
z;j3oCR7JLmy2PqN_1GR^g<kLqUMmFZ`Gmzb)+;{NSs2BYITyqg2RjipOS{EwVhbZJ
z;BLKH8AD*Ui-fQ3SaB1><Maz01tTIMZ^xZTl&1;Fg9G|#uc?*en}!D|us}#Dh$Foh
z)<2F#oGI&SR*v0)OWDVfg?{4Wc40NE>w7J)`T-+;qHo6mVJ$kH(nE1{+}PaxCV)-z
z!K12d?n)`KubUb=KMWX4t=Hf#R(mBy{%D~?hJH7>fONx?CVZ*m^-yeSY3BXYtrER^
zLaDg=$|3T8()sR<h`ZZ)aGb28b5rMijI{#sW&okuO3}$)x9g|Sqf7x1DN!Yp1W$ea
z@mY%c9R*7ziAUbqNnng4ra}n$zG<<?YoRSW4fK(WRxXp$uEdiuk6@%@f(5PWTLDgF
zFJGrbll6te!g?iNX_<n(wRnkkMNbA6eoQxmA4lMEx=a~)HM-SzteWmzL^afr`7x+j
zq9Dh>*vOpVBP)N8yjFI6@S6>DLH7A&@Nk&%=xY$K)i@w=Swe*M1x>gW&HAW`FAB#H
zX+5R)b}Ultqv#<Ehdp_efC6`Cj%Ew2)`?f+QrU3nqHXIjS{XxU<eL~AFM~kt4<_c!
zijQ%Q6#Z>(FFx*M-`+)fJNo;^!kQ-DseDf{%A-3hq3&w|>_1QQS1PD3zBXob$V<tH
z+Pm!r-$z$?l})FD1!y=vat4fH)zsaZT0yfI<BAWG)~g}=?f&n4m|K(hPw!u1QIxzW
z{{%aQTI$B#g+}y9c#_H3>3CuoN0WkFxV6ul9C#AwH<%;Pw?w1+)G9GcmQ50RD$@Q4
zc0Oe8$6=2*NmH=G(gs-|geo-uqLs<G&+))1xHhDq|AtKerJGv(2<AG121@vi9-K?u
z6|_7NqTleJzlXOA%ln>#S#}JF#z<6`QX?Vys-s&NL_nfJy^QIOU|vHt29O=C2goA8
zqOaM4SAICu!A;*oE{l`MWl+FU{?n9Zw(%XomvPGN(wHbAYfC?4y>4A8j9JI+4k2(I
zcWbS~$*U!irZns_J1eBTqlZA*vl&Gs=XpIRXN$%6i0panS>_!iBc@jxs|dfMB3EO_
z&o^fIcKI<BG(3(4&+*udIw1xTWE4crRd*S2BUlVy?asDn=bDb_oMA(6Tvf4XU$_H%
za~m70Zs$mB_^*LT0us)~T-=7X$CVLR+Ir*j5=`VC6|Jm>JDWc|kHe46YBVU&`995-
zbIGJB`uPto+kBRIO$4qwxzIYqDD79b8Wi&~=muSIPlQB5T!g?#hVdd`w<|cr)*(#T
zspWw=B{}T(%q@as*$n{jM<6Gnwps)k0qOwWt-B8_?=N`uX{)vL?=`|bx*Fw3;#+D{
z?UyQfW9wFPQq<R_md9R)e$)|lRj!RErD<df=~*nvO-bd9D^^=;M?XcbnXgOil}u^t
z{VyEQKHEKp2!2KpoLJ<lTh7@`Ydzm2@%@wqp`!9gx3%3~%oW<d_Z3162w#F6WaK+K
zD_Ng6HG#^D_Vec6G07v%xe2^9$?0HYbrAVgjLIPN=Ixuw3phJr&wb73QXchPy8hx@
zzzHd(;u7kZ&qeUa;j$&nU6bQ`=9SK?63$nO&$DJH<+CXFNxH7FEC{$0zSKC%$4`?~
zyW@Kjf=M`TcC|C&tHtrb7ejj(9zl^1O0;pnw>%hFcN1gKQ!K6nbVddm3UwgFTK@TE
z)@kWPg=C^Jxzv!0kjg#qs%!$<`MjOsGE6vgt3|x!4oY|a6`p>mkn-{DEaWAJUygZ!
zRK1@ap&MDNve-lFc>fq=>*o}SWysy_Qr`f8p5rZv@+rBiwPNdIW)Z$X|IEtqk!swD
zcyw_Zcm*C|gm@CV;aa&nK*RyB`7Y+V1JD3-B0O1tSrG8>UP4#k2$_IjyF91DNR{8b
z`lx@)DCuLdLze3UDyHAJz$lj6^=LqsAg5xqwO1?E2KchD9Yl#yl~YhMJ{(1KhS{B5
z-AzFr9*^F~D+6to6Kpiz)HmdC*UCivCss<zzJ|wsQSB~#PJi>|d}UO61_tvViLcYe
z3^@9jexd#}eVd<+J9L0HJJ^!ZPf&g8Cy<Mpc4uiWi0pc8=JwoNaNv5t0gi;f-)>`+
zzBU-!Oz#k6zyn{rrv)WN^>9SJH7&u+0{iW|PCKsXZ@*7X%89W~5Wwkl#Qicsl@tTH
zA3L&P0~)f<-*9m+y|&HJvgY(uff10lCr$xmM^&~<N~)_~d>zX1pfr|<!#@*GLcQlZ
z$5<2HXFbFzHI~tt;RRZCv;l>ruPPHFM@e1D180%~G+I?N&M<+4XL(||=xhDqi#ta!
z-C$uE@6D~GR}?`>-M+DmE!D@->D1$OP$s?u5#xeNyYG9imc*|)tL^VF|8OQ<H!07}
zI)0a+g?#zhEy$TpT-~K?s9S%*Z1WH`qWR|5x*E`-YM55=HL4omr1>5&-H$l`=TCNs
z3?t``K_;tU@#c8!@Pqg5vW>Ff>*V?kcvG%^Uv-J^Dl?P6)H}vTf_}hQLBV%Rjl0%c
zTBJgR@#ucQ!J3s-WseB22(pq)AHiEwc4Z81a4{-;##>N|vgN%cbaPG49FNc#6Mn5r
zc)jRDR}(8p;A5;^pGM+^zAw8n1L&;W0krzo(O&jIO+^-|eYf&s&NLCW;MESZxEMv~
zoK6jF6_an8yV%QuL;qJ5hvCSLnrBj!rCEZmAM1Mt9x@vpWOV)JK-G&~q;2cbidkZh
zQ(B^)$T-_^m22H~7}Ey0ABg{g6B1RV#g5+}<7gj_isjDri%>|)(rSHG<f`s^ft|kC
z`q+Q5_Oxb&@~e@|ga7#YJo7NR6b6(T2x4VXvt@Jb63vTBTGh+QfHPXo`i6Qz#Vc^Q
z*+N;8bgWhwQS1RHEt_=fqp0vgUrEpx;+AASHY!CC&9&CifxUIBf@x;_)F$%g=k4F!
zWv3`*P3A$-csgx2xrLM2Lf|sHAK#S}BKp1H=@<%p@b3%tbT!p+gV4Mi87V|7o2}0b
zIRSxsew<&qJRszk-S+eYM)GqfFi%gVYXo!=!`u5v?`r?R`ucg<X_hdGM_C3CZ^K!T
z(vd{Iu}rt_-oUn=k&%uWrP@)B?9UlzPj=^8%~^iQjmy<9xqB#Dvx#XWYJMlg3D#+}
zr+11OLzS^3rj6Iv@XKP8w;+3)#D+_JQE-ElPg3Qbob&ipU}!Af6{2LoNX^ZVE?^xj
z*pX^sHoL89^JgY;OUlX{?GU+YuqqH~$dgXAC2PUE*WxFO&7a@eLqtzFlFf)uG{eP4
zIC1(_R!p#)+4wbDZ28T0ELW{P7M&f@93R!QH!3+i(MLnw0w=T@fX?M>31o}|qY4)L
zU$6efn<YtbsJRVEBui-iml}#vSsPAa-I;G-Twh<WHSI-E&Jk`hM}FJDh)^sk*-J1!
zj7Y&!#!4<j(Ay6HG5|h(Z;svR87?B6goE`A{LB@ffS<klgWR4B&Y-=EE2(s-3&>)_
zs-&y;cTD%j@b^^$I2ff8eA1r@iHc$&O6GRXxdu?c*fX&o3=8<|6jC2~@WE9N#q=tc
z4#@*?+2l(SKpf~pW@PeUFx@ke^tI1*R(-vn1(-eY-1W@t$tL2~F1Dm=uhB))d1a3m
zWKFr>VLQ<Ax@NTCmc;14n#s-!ASe9XW73^amDHG!5rRJ5<OWZ0q~s;F_tu?R#pJ-q
zRY(Q5NMy2!X5b`vZ#m~O>eV}ft{q}vShU}8A=S?%1cQ6S3i3#wXd-ynwtC#^ri?LJ
zsUu%|8y78Au`@xDpg|mPieSBN({DJV-wU5a$c})0_{?1z9v}}9o{1WuSiXNPip0Gn
z6w1JGW`}iUgYCh~Ag8f;&%lb_ZG;JSnss8{EJ{(uoUSPw;fz!98XIt1-w=T^q)#`}
z@Kx|?$o`mh#>EF45t7F$dNV~!$jc%jQ}&yG53}QuBA@8m{XN<Ljw<uh@A$-o$Pql6
zOxJy^nWnn0DZG`g^Wn!xgDELFV@Sc#)oA1HzN^1y`bw~>CWG6K34)XW(zGSnn5pnb
zLq()%u~S|_bKpQIjhA%7!mG&-YAC}cY;u?4r3(?=V;)C0Q36&Tr$j)?kNvch-I&Jw
zm>#}6h@#;2Puy^D*x<9Fn3qrC5+`FE`QttB_I`5?>ra7x`h6Vs0*=$*FqOhk@~_K?
zRUn7Z>N0hPCN6fp<gZn5((Kpy?9&&XGERJ2XU7dLc<s$?<O5gPUPkHa=;ztigKiYF
z=h3g;*MNfFxT@@)ae_J(q6S6~3-;^J0sDd!F^m4FzjyUzGoAcTLMb1SXyms$H5XW;
zHw8L{{;w85f{_u~x!@MyaO44hU}Qz@AdYkP%rE+GDxE4yy$v=4ocYs^OQ;-O0m#54
zR(p#0oUnL{vmQjhDH)|2!iCREFJ<h~q!$|=^5l>pr`12|-4ZV?K5l?^GuwxTm1SJ+
zRJf$F_p_T=;p7~PZe-hQMl+jed(_rxN8jcX8DXYc*w$jW=){aI)OG7igE){%#z?b8
zdEJNf5%0N3-!;$9u9l3;R{^aIoMf!OQTZ>0K^+b89fl4XP8s^&>-nd6bFI#FPov~N
zWAvg*fA~NEpDqzrT57$8nMq5B5Qg)F+Y>D$${3;82hqp#z_1REsQ|`UIMh%C7=~qP
zH_JwCjB~eDkQ0TdWJm92X7Z>iOa}92D`Q;EN=-ZNRwrAiUK%WkwPE+}5pLslZ5i`0
z%m6QLUtV;!%cvOL=ghwL%vmC-E)s~Szb=NI^6|V%)1Ht?9-3607gVO7Et+K#ym-V|
zcqu0v%X;ZTjWt2cfnV|4?iazgRra9vSpS5Xgt4=WU!I1BLf9U$j;ybq1FL{k_w25t
zl_{5;<LDfqtTqh&H<;1nYL=Z7E3*$Le9TB9<*lsbN^|&^m!1?PKY+6{BbZI3Gn{;J
zthY!ahjKp<9nj2MX%RUbS@A>4>hSFQoCRQ(#@vSN_{Cz~W8=Lmx7@mbH_0umJErDc
zW$fgWUDX79>DPx8*bMZnZ@9CW{Qoq<EuV{15K-P@9X*>F|3{2_E!bLxtWQ_7a7ZI3
zHXzt+aD1r$)AFwFI<}L?j@?(Hwzz8@C$6hg75f-Zb(^f9OSj(8h?({`$;PbX=x}~S
zeDX^8d3(M(f=)AqWcgds1=j(zxz7{^Hd({&y`e|oI^QI~#5_NDh+Ri}3VT1o25zX5
zi*>6um|t>Yy80bPKb&M$4h&W*moGhkj|2WOQiNtMh+XX00KH5nd!#v^6RmYeqOkXm
zrSikBD4~-1P_^S@>hA~tP=WF%1#HU`rCY_Q?jHm0C$#b+mK$6N*+vz0HalXdq{822
zbBB|LGzA;n<CQC|=8KS*klDi8Q~eWeJ(I}8v%;-6dbh^Hn#=Zhh*gAD<uPT*if5zF
zg5gqCcy$n{*{t9`6^c%|A|Cat;$}1BORnAv*Bq7z+mUq?EYH9W8ntWnjc{Gld0*|R
zjQ`iDg`bRCn7oYWe@2~%Yr5W|ki=jHJKgf4<>loC$B+s1_)ke8meW43W6jP<K^0as
zf)srsWKLE*y%~3w<u+t2<2A%x?*Q6|E5CzY*HP)8@swpE&=3&N%ge7v43SjF&5>~A
zJ?IIIhYN77ul(`zBqWI{O$e@t3u<Eje2|%MyvI>{VuS5Q5ee2{u9QE%89?!tw-&TK
z0bZnS4vmlZHAG~20xnV3Yw|V_;FwdBUuOMGn{Z~T4by(nA480nwKt{|<*xX?w97Mg
zLn0t}C;ypv=EQ7w@R+=)zsAkTk6~$-eIrppk5|q*6Vx<!q*KNWC#N-^yTUY;=9>3V
zWwKtXmua)FJ`4yAz`Zk_*HER<e4@1VtR`xf3KLEKBY{VlPpbo2H3s<*u)95#VJZBI
za9IK`j&S~7%gy*G!ZYxMaU94IEB8|OS1vrfJ)NMH=P5&0;Slds{?sG1GE-B>zTM=C
zn+>6dw{Em<SmS0gB^!$@HH55Tq^SkiKPo6p^n?wn)gAd`c(8Q9_Jy2A5CJMNm_G-;
z<5B6s<4GQv|Aph;zyP3EVD*oNoZj6!tZQuiCH??KY!tXn9u>tawST~_()ANildkr(
zQp{nmM^T=P=ct@^+rnJ`B}2Agc;KA8C}?$|_2hzyBhQ|}RM|U{Zt?`h_nLSX56??M
z>+qwseV}%e%49|E8&cw(v*9q<MK@id&`UHgh}}RoM88*Evuo~*`S9+KH)N4`;&6a^
z>5&auV`c@yilge*bkynpK0hz?(*xt}a7Byn&B3iu&*=LaW%67!;2lfFS!L0dP$X}*
z)B%HySm(|-0K3{A?4@8J!T{PU+rhYkn>0U$%#;jd{<d)9-mDz1{p?>4khkI;=E;nX
zJ;uMH3NE@P>#WKx6T@K3Hr<*d*p3lSDaY3tsxtwcQRG092RLKV=baLW174gIASVEh
z*wQ6qg{tbN%H(Fozgq6S*t6;?urS5kCe+WqQ1i(dLW;o##{sn+?;oCJsX6CN*8}T&
z6&JsUQS>Esl*Yq}N7`FUTCo?=C0tdjr2J#I(i4?5S|7B|4ou*%g1K0@72ou;Qll;h
zlCI0MId><zXeicFa7^6Vsk`@K$t}TnVz}qXko5B?MnUDpAByo~7tZ#zI@-fYKI^|7
z?k^eeDHN>VQYrjjzJ?%F(%W0u(8vfE?@OoaD<B9R10ysuQy(cZqC^2vF+7k*r&S-6
z$io89)7X;Gdb11q;bKF;uZNL`!+E|@Q8Fg7k}twI8|}ia8C8!3p0h$3YMmYQg`H5Q
z(VxQ5d~#frWvakUPpN0t5<mvh(OjcgD;<7Kn8%d`#;r7y&{wC*m}+pQ_8*tIk8K90
zJPWeTs{ST3-=>dznhYiCzA#dT-9qQnmu}c0ec*cCN92~=^}B=)Uo5Q4H#oaBOO|!^
z?rNlA|4cRI^Wed0uBG9JsK~#UnFtWLxw&l?8kvndJqsGUIlo;abU!}$lzo=X!3m0z
zhAT(bPL{!{lrb#B)3&*A)l`ccL7rnp?k(ISmD&N%tqq!!t2!5t#eaS+V1NHd@pt>_
zgyC6(TtX_e_{=3lpw12L1Xo<7^e+ES-HQ4A+ff`*T<-dIfc!_FW%;_gpN)-`=11O=
zX}Gh_*B@!)I=t-c1ltX?gkHYlf<LD){JEPPA)a5Ssa8if7hc5r!UQW)%!bb9R}TFW
z@3zIpa`dTwG0ZTB8UKSA{6!e38K1PUm;ZRruOxd`SE(q3T_t0`Cg)wBQi!<cJFR}}
zk4Mo$MQAgEylk|4$t$zWf-8cgfEXb;KzqFDh1QH@n5H7U!|K#JPpnl*t!UOfdw+-1
z;V%a7+c5K+yB!X5F8(RzCiNu5F%coss%FbeUyF$ou}Vnr8FVxoj#jWS;u2(m+N(l&
zy7qV8lpwi&DreQPChS=IsDMuMUZi*zxUS}<#}M`|Wi@~^PtN>k$;zv_^waMqE*moh
zt=9T3!7zvDfV^(+grG-lqm(g7Z=3d4i(<W01BMvcf9ikgNGS2d)o5%yVekH4?j#F>
z&uKB8snk;xnVgyV4#yHxJG)!Q5Oq6R_2(#??qPfD3e^At(|MbyxA}^Ag04Y^kg4@`
zz|yc{^Mk5iyJStOLdqyIS|(gAA6BBi|7~Ze->pe#x_1UOTl=Hu^<tc{-YY=X>_%D&
z$3{bk>frV#&^tv50WBTnRZvgY8oS`_QM#y3=8RqBcv2iuLAXaZJv|ep&P4xltZ>Vl
zzJd#{0k0!*w2Obp^3Q42f1|>;Qf~ch;C~9|E5*RJKDVP43S?|bEpLZHxflTK)RN1L
zoZ@YMJszCu34HYfMzUs~XRS>Q{b@Q@@m{CF$@k6oWxBl$YU^`2>1OCahhKHQl%v;x
z!)=WP&3CfJ8qB50M<Cdw4)@1H9)8h|%`zsAgYQbg-79cyY`s)XA$ot$y0_sO<Sf+O
z9Z?%GdzZ;?WUsN>i#T=KcvpH}rYvhrm^F^Rl@p1*CI6vag5&TS4Pu=NomV4YPzBl5
zctzcQ_i2~7tnLWfw062C=YN=ZF5@#>o_NcTr1JM$!g)@6><{1Z;r(gxO632l5Tw_B
zSLoQxRFWJn<9~C=yEQF-ulGYsI6~6$HMM(~ovA}`^(LB*URvflEdoJqPcM3lGvhXM
zoP{J1-Tnetb<CYYA827^^-k^GnhKfwWl*HU@|Dp=)J;J5hk*#rlN7trXp@xcsqKk@
z-v$XVhLj_>@x|r>_s56V^E|AsSv~n5J+C_YEWTE2V9JPlghiEWedc)IaLl^Xq<k+Y
z%0yMZc3&HyVJ8}$kj9ka=<gu!@HS`LZ*%wOm1}@#tXB}&0PK0xF8t{o)<tl+eRug!
z18$YvzmJ}$Q!pDBej*7O;ACR`otsnlM-2@R2hY!wIIW<pnvVIg>D419F{rR~KmR!z
zA&H6#H(Ihg4d67-G=VqW1}D!0i_Zq97lY=5#>3Wd_n-pvoNu-v<cm&q|CsZ}b=Q~T
z8ss#H-)mLVDF=gH9Qm3shfq(H*q~(WWNd;xf6rH>K^=_aCtcN`9134LTE*)|gSsr0
zYjN@sfq^q>QhweL>u~S-z@n?BbL?~dIcP+<WZ-GRr!5&(Z3cd2ZYg(dt6Y_u$9=#K
zv2NGZ?Z#5wZ7c8|l`7y42c0czwbtCJhtKC9#}H1Ci&9`eV9VXs#{8)W<uBz<(u=;<
z=gJX2V^;ppH2fdnWL=gcy)Xnti{a^LYHzVG>hz6@LeteV(`zdbwM^d;Jt?@$zopAN
zwJ&ZA;L>RY{9TJ=7`)r;9AX*Gkn+bN(wqhX{!0J|v(1MC6^s+tM=?a4a`()d<>VRr
z`{X+!xpKo5Y?MB;y;xr@i*>dH)qCI{f2B$dk;1ur+|7bYt5DEkz+RUeW38N^T{=Nx
zl!6m7IQDf)Jp{ocMvoLe#oK>x;xEIM6oRW98dm?=)?X|6>4@JRZtU)E_G%Oj^mKLC
z;HMB54-bn~jjIqYQ{IzciD_94kZac}i5^b+u;5}xYezqPhL#GA!`D3KZ0n<e4B;ZW
z9gePQX~D}`ESA*am1?TmxgRZAPM#%-JH<CK=-Fx8+szhz^$a}}qW!xj#3&Nz4D+q@
z=WMw7KUB3bt&9;b%&}=HT#l6(pK|2H$@dU}br|jZ=mt2?ws>9AW|b@-PqsQ7<;*VE
zmI}J9aZnBRetRi>FSauB?OhJue|1*4?LNIaIrxiAM*hbGU}6Z`-?wjcSR@Pz3Q{z<
zxxS{RrhZmlUOup&!NG{I6&E;1P+ic(oX(nE`B|S5DiFDmbS{G|T(cw<ca1`&sNt{f
zzsd9?#G_T>GRFd5$m2^#@t?fQ$ZCoHOJ20YWwN-(+qk%olS{Z4zM={n%r{(}!M387
zRO`Pw6++lKEtK>f+$EM)0>~#0<=|F*Sfl_jtS`cFpF0(J`(KpJORs2oliAexO0QON
z&y>XnY>92|YU~+8FFs5)$}GWR;GeC#7w(hhu{l3vDDi+gChVKD*V}nM`#2N}=wDG<
z-<bF>?gnJBj79v*qo<xyX~lYqXrX^EGbtfV<Vl1lCE>$UX~|eV>;0(zvF(pfwheN#
z!2*KIp{*JTv+>)ih~;QO@)Mf1Cv$J~TBS_9z+>O@%R@P%d0(?(l#)|e*5t^ahcqXI
z$~f93FS+h?e$gO4(V&-vNx^8EGjDpX2DAF<4fQyQT(3hx(zCq<!v`>pI@_;{yV-1q
zQhCjRjw&%MgF#PfdcF-}^Qfz&;umUKLaJp3D(QolE3$h<p*qE#?*>zgbu6o>vhBiz
zwKHGu$7KGG89u%3puQ)30?4@{{e5gnA|~UtdS9@Jii*Z(WH5_3Sy_SD4VvR|VqHq!
z<KT4ky{YGyjnj@{ku`kvBt)ddJ%i^XRH^|eIie)k57!}_wRJ{De20@#&2vSTwCtKf
zQ}F2mp-d9Y646O{SE{-4cfWCR<+r=Mdo8aHhp)bChemAQ6NLRiKRRKp+9S(9d-kV6
zm;hA7Go;1z!5@9{zUMH!Q72OOF(zdy$MT#iQoW-3wa}Hnr;!L%xKgS4HF3@&Ddgvk
zmF(Qs4&FZxC^-JQJe3y+iF|?+q5p%U=aq(ScpKkiHt5swmo5($&Ft-!KSoG~F~%<v
zIKB-h<PtEi1{^^pVt<7byhrFy6Yng-=8W#35>-_0y|Hj9jV#|qyW)S1B(I_zpOZU|
zD}4C}Eu_6KJa_6e>V5OMKKv4JN`&W6{5i^1;-mG)wZ5r~A^)s_v)TJ&vvl`eEl6IM
z5?-LiZ8S=>=UCRwWo=kjp@Kqt0YChWXgx=mqzQNSv|SZVAf0gxec<Kn_CQ$FS?M_|
zb-ETLdxQoSblG)Xaoaxo=T3vlRp1`{!RMaY`|suaS16yUBq%R-e}m!oEgGP8X%M~5
z_r*~Mgf(!>PH4yMfL$VAskckhQ&OKu%Hm;bZTl{R;dq08T_yCNVxrWzn8wf6&Q2%h
zOE=qO;^mI)ntW6<^7Oz&t7rGg(e1_S7B|dk;3r>Ce!QjOxPEz<Cp>W_>_<e@h>FFr
zvE1y7LG)MpQ0MyN^qX?C^I1pV2_22+E-sQ1kuSEMFC(Mf_KfrsFLiX{t9{|1gY;T<
zDn@PG%7ats)<o9xt};2}%uUQEK<W^Z52d7uGF=7b`tejYusLg6!hf7r)CdGw&mti#
zmp7C9f76ct;87VqoKZn%>ySPHby_>W0ossDd91azD+MjbnF3y+{=C<`yM>yWqs>sf
z))7oqVB9}A5#~u`OfizDn*Owq-j72M@~ok9Oxgjxm8X(+b@TddRmAZ6tgHG@6hSSC
zNb_`F7^`$R{x|HwQ<W3naxC`IwdG<;TcMHe)TGA2qBahJE89=QyrNz>80-`l>A!(~
zHD|(p*W!KE;*i_l*EEE>{M}}<b5^#o|K3mb(|)cIrq2JL{n+9^xnMK5Y|Tu}ri?m)
zTKT%d2hl(N@pw$gx4i*S#|^@^J_w}Qp_K<fZIZG3rV%3Koc!*a^6s<uRPi@!UH-$I
zXXCHOkFw5w9lu*z@ku^%Bek;HIlgwbm3XNW@lf9Y%s=uZWv&T*U)Kc^>Fbk=OA9Dr
z^Ig?@cdwZ{>8G9T#~{LpTtR3gE!SRalCuKA;PkjhtDBvo|JR?&!VkyK5BnE|^Z5V5
znxOZupBzn#v5l3SDx$I?%E{=pG0nRC-H<TnqgX}y6*4f)02pcrA^wSeumG8l#LupB
zW7V0q-N^Om=0IMTt)7af<fFpV(G>rAuRaRDxiwYnT`ZyE?fgV`<jV%@v|S)T_Rs$6
zavb-5G?%y$9lirb=Gt2x!Nli@D7l;YTP{?s_=$lVK8L6G!~a1<mSJtBGy4A^;eXuc
z>)KO7tCuLg<bS;VFD9Uv1P8wHf>>tnmQTtiew^bvt3OzRPQy-%N7*#p^V-6=;PhoG
z3|(-szIeZ?$8Vj<W_OqA{++J3hQF;M6bYeU9_FA7Kjcqnj^ffz;sAbNta;-r4SO~|
zc6OyR?$*BZ-<JJu<l-4a(7iu)n6opqSW}p6SC~pQ1Wu;wQlkB2q+4L0k#N5(N4i}{
zyrY1_>6>)^I@PdwJKHb+HlhDtmw3{F1xkSE>14aZvd>v0Rb$`LOV&gpZ@pNidQX+e
zvw}@2rq){y-b5Q*c~*T3qjA6YelRwjX~ZRUy^nZXMz6@utHDBYlB`6RtJwD7fkio(
zl+^+=w%&q<TiMN~F7N_mcbWbww8=?QnRlw^J^xhxf@Mm;{6x*xzaE)im{YVBpGpXM
z7#kv$dl+}sB+==c5bLH?J#<-t4An1)O|g<)aIvla_kPl!T;h0ra{6C(79nYi2X_gf
z8E)CUnHZ%8pWWyuz*hf+=$_LR_VX*^%4*;?;oL4%|C66iQS5(XLjJ97x|fo(Q`u1i
z8g+tGUlZi2-XFZM<usw$nfOrOReb?F<2pXCe7ERUHs-csqLKdXnDVG5Hs`rFmaWz&
zKTY|kKP_mA9`Hw*d>py|m_y{R7;sGS|9-#Z-+=S#-4nT@{?}gM|B=jK%`Q%$bcugK
z@Kq}^FLKhO&SHT{G!kQrWYYbP&okzM)alnR=gq4lXUi+$2M>cd_Z=~7p+-AWf()gF
zy|f7IuFuFi<EMFXbkYeo@0o}Si>(6}VzR*Hk*xOmjlso*tkZmaQKMpj3-fG4@tWy+
zUD6iMW!_(!&gWNuT00FU!-@;k1bju$)NFQV*x+OI?!_^>LF^`&(!&jm^{|d}%E(xg
zaalJ^cBF500vu*zbE;nJ<h7<RNShC|znrXEmC&sn&9~Nil}UFOYEy!mlEUrpmQH4`
zj%D+NSPMS<dPXDfZhlqiTz?7wc2G)^LFBe(%3+Jgyv{nFNg``D{l?*FW>|y}@S62r
zSldRpF>dqo5KN9GI#J)}lvR8mHq`&|^O7G<fm-hAyz-f2WLed6eCh?^DwRk~la9>w
z>f-O3kQ6bIUUelUD;3g=^w5M3>>mq+)@KdnW#KoV8yk%dLf<`o@W2!~TuXZ^TGZvI
zD?U5}g=J}dUiKk=e)R=LANP3&8=DoqZNG-R%1PsFZ)~w;j?#t9*sV%Kt-EB+hG}>L
zYWs({1FBw~ZWnNT$NtH8$yVOwu^29t-)?+as<sYE>1yv$S9-B1b7Z=d3$E7^=P`dK
z$~K0CT;byRm`d-Tk>ZGYmG@N>iVNhAXyJF9#maqn7-+j-{oVAJM!x<gu1wQOgzwX*
zzJ|B1S6XvC^VmnmJ(sY9k9Dr39WP@<U(#Ou29g^K;Y}6{!rB;C3L(6Jr}}Cdm@^_4
z2N|feuJ2aPmO`aV9&EQJt)*i<QJ50K8U1B+I1su?eR4t(ic~FQe{D<OzqE#f`Dq?p
zsfyls$V(;+O09?a4&8ON28eU)<K`ACS7zpqgvM}HTE<=bo+fO`IL1=BJZuR`2yk@8
z_J<0PPketAGPOC<FLsN5FY#}ahHD~luJsDn=zpgF*OAF05wuhdxL?o`(`mJyOqe^l
zXA(_~fcP>hhg|JA#b@M#=(<&sP<!pg77He_34T`F>e`soOGNdCjXkb?FL2;FF))ic
zAFEp91|Venz@&_nI%-Fpy)S69b+x_C1#OqsgMhs8xxe1<gYyTnOd)HAJEB07_@pGT
ztd0DmfDGgnKG^N%6J@y_Q|iyiHr&wP7N=1V(U-A<@_CPlZ`@+cCushd@wux~g|z-0
zL35C=5k|0M^q5~o8xG2ncQ>6?ZX@RKSTt9QB7LQ3I!NLYmLx3<C&lN8y)-q>dDUFD
zOZn>lTMz^Wo3lkXNd}v&!(Vusi=NaMa49sWZ-G?+#+O0tNI-<~<0@pyVf8g7vZlqT
z?Ge)IjoYShrN^IRv1Y&AOc3`Yx0&uMs;#f9lAgNoF9$^Gfi5`21krh_*|KtQk>cXd
zOA)wm8>Cf|jy~+bfI`j+2iVOfKa+V+NQ>o=Cp!vDc8i0|!NdLPtol~_>R)LOd_}Bs
zJ1wyFYKN5UhYvXDP)1;4DpJ)mD3x69pf<d^;;i+J&885FEpwz`qdEN-dh^y2ab=7a
zlux^v@4WFMUnqsAej^^Y#_)Z~5ts>bFNRb!`k&qo#bw)vA|C`kf+iB+7ZgHs%QvoZ
z+dms1t5Ce~nqR<9&m+1T1H&GR;6HBX^2bZ!=}*s!VLROKCVejbj`2tISOJgD8l1I$
zRmZP+jVRvGhkHwX_)y}_IqSo&68m?=MGcQ`T%H24EnWGj;D0U$q9muo`>|S`jk3+m
zin7!Krmz=J$*R|LIVrfaaU}57j8>Xnmg;bF)T3|+XfS-=J-t9k7jVVP86LTkiGs}a
z3;Un&iXsP~5!8jKe4A&=3)mg`xPK1=E{S_dLUz}qAZ!#yKRpgvh^L!z{RBTW-gWx9
zNHnWb<U-f47(o@SSK-h8fF~rxPYATh@j{>NqudOWIr4kS^1WWcVO0n3q_8>zv#I;1
zdH4eNtm$O+QOX)JT<3>p_lBSf#BG{zMRfZYn&x6F`Zz$IU;wZrrev^*am9v8CoA3V
zVZ`%-U4`v%3ag7u0Efgmh11yHe2GMr9?r-50W(dnTP$H&HKVAXoyeljI*6}ua%LR2
zl%O_N!A*W0z2otbw6D%Cy!1)GAEu%|i~CHDf<3a}bu+RY;N?4(193qW)44oEkWo>w
z$A%ADdc^wiUXRim))_F8c~ode6Cp}182r1oA;jrzf5*xtLG%4zvZkhJUVx?!Z{v}+
zNi+G>En|y&32m1@u`^8^adjP3vZzFKf)zy8hyzdf#-Gz8Ag6vqwie`QHDD4Xh;9TC
zL^Ar=-{PG45qgl_K9U4OvB4=Tc7~%&dfnLE-Aj93W~;t%V?=HDtH*u0eL2KQ#Z0xi
zuoyfu5wI$@)XCNw5JzPmWO+uWJ=c}Spc<8jcl+!a^)rRHQd)ll?Gx&AKgF}=)nsGv
zJnuhgXEG^Vb2Lm6F4;MH?GZM@?pHQi19wU?ASfE=TklG&5I6?*5{|ck>secr)MS*V
zGyCvw67Y93?D?_NE6gI`E>KBimZ=gK=xaAshi#wf)z@hJ!e;vC)&VcPDQ_|dClq$n
zo5L%g@UWdyX~ga$e-5)H`7|$d+|5jObY3ot5-$kOnCF+P(F^G7`J(!5Ho7~mrm}nY
z_nZ7`Bbfp*{HsWkh!=KC&1Bx=cli3=H;T}-RZnL#om#J9oG`qyN|V_Yx=F97H4MgF
zDj>)}NVtZV+>X}<@{;fJz_=0)$3+`ER1$0JGh1GV#dQ@(@Mg8^wvDvyKy4M7CDG0o
zBC_Lvsb*M}j@jLxPG6zqEof|y0ZJrZ_Rp7mEj`*2v+oztS@q?EI5<3=WpQ54+zydx
zv$JAF5A{Qj7wb$8F=m?yO1!<`1(@=x<_81?*`ywzy<Z=gl30XXT5_hozrYtm_bBzv
z$_a@mjTH2fAYF7^k=~C^s~0z78?F1|K{Rb0T<PL1s-`lKZiKuw4k&!7FkF=S3zdkg
zD!T4ur_~C*8!p_<T<l?q6m~jRq%dy>yJW#fv57O$xi=!z9NZ3luV8(Y9R?3}?tPCY
zDtt(0KC(!6S+#a|$2zA_qucsGn77w6sbXqdXb>o<rFE#MFQ+G!t^Y6%zm}mV)$5d}
z%m`kKG9vX(MWJ768-s{8f;PujuJB+@7M@Xv+}~&WSX4rdJm1#edwM-(A~AS!k%Z&7
zNAGkr>A;dLqLP}ts?EWm{XsJX1md7}TN!yhlM*d`hEJC!@vSe!;m|AP;m#><H8_M-
z&Upr2NaEW?u8%Dnli37^qOfxTPE+-fYRL!7A6I=VUV@Bnpa#dN!#w-I0<RbS^4U6>
zalvyPyVXt_>(*tL&?V}RMd-J`|C0J<Ij?TnRO0_>&-pG->E<ncwW$A<V1ErnOiWBP
zHC6YQ@U04bXFEJV{S;>*0K<64CkU#EDRg?2QX(O~-{HkeXZYYPEMB?`%-9ER^fdR%
z)Eu7Xx{!&Tn#oq9IMHGWV2hg9B_dSdKBj@}F*(QFyn*tUzdYs(i()Oc$bNvbsDSAM
zX)U+X+0a<vMc%{BYohwP_hlgg%ZMZ9$%>Y*-%mIhN6~dp1cI8erumwVZ(y3@mUYCc
z<TCKCT(6K4Pgi-S;vFG?sIW4NahaAjt=lbM!P`%mGWhWFofK5ufnffb04X6r;WK!V
zb@lTLq<#sdULKdQjk0*aCQ7id$98uE_P&At8Hv{K3UsU(9%QmV;Z4CI$Q?(+BVJNJ
zo7Lz+b}>$!qVbpf=&MvhG(_O}nSt^ZBY5+j>qEjsE*O55NvJsC+k{_>&EK`0Cd>z}
zHZSZ(h27B%oK_N-^?@SSYrJYY1%@m{V#%ETk?;&Yhph+Iu@JB~xvAK`Yvk7r;deD0
zu*0^qU}lC3e9m!#RNIe<zHO7lv$9c=!ySdnE;d4M_#JT*2Q)1%k&u_@uK}`GKw16O
z#N!{F-=GQDIQ6!rwX?RLzF0^jOuJTW19$(^@@JMER+qsqr)tlz*v&GM_F<il)AY<R
zf!Ohml*_tRpeA?Y=V^i`H37JZ0@Gxht%L)m7(H6T<5s)N`io`U5i5m+tz_7V=PHzw
zQa6-WBw+1V9q=$HwlGzy0%VAC@okXP@ez78HrTc-ouQ?<$FOtxM~?V+y}Un6cj=*G
z{i4N5bf7mUA-gQZ4>R<}=ck0_|1xJ!!LeipKHT;)=t`geBMhU)F{Qd;gO6eYZjVX<
z09>SRKE1Vh`sh=#qKdV(uMFk=Ga9u7O-ti$-y7mZ>(9L*Bi3H~egXO*tL2x}j3Q(Y
zO3?N5h1Kg-BX2DE7AL~|_L5WtDtEP2!FK%HZNJF#+2D?>s&LMBzcX~@-8rHRYgKK+
zamPD!`NunrW~?r&2h_A&aB?*=KAd*&#+i)t148`u;2q{EGdJAaa|I^dgT+R6Fu}17
z%(rivc3V5G=>nFwM%rwIHZ?h@nH3*86;%6UGd6IYSE!WtOYJ^_@^gX%e^yPrbIsr}
zS;t_U=-weCsxwc9*Gc`_oEL2Q<HRx4;NsdF=v7hRWM))!GA}5A<urInA|0u)>bwfM
zEPLTe?~Fv=vW;S&E&8!DV=tnfT>aN{e^<Vy?yW873N~<?tkKk{Zl;5&vwZ#jtNr6V
zDKRJ>86I}eAzmHE3_jiuYJ&T~B+WTn>(A%EYs3b4C7+^Mcu@9ad?gkivFvI$SLa3k
zXbD||;#(gtW<=i|(Pd|uOF`X#mKz%pViJM5X>znMpzue*9!=q(e2V@VK8uCOA~Z3C
z;$Y80_Ne;j>=c{3iRENqiULK$UZ5d{fH#V0CHt`X&U>m5?xR+3{G-`ALU{dsV2(4q
zz2H67^$T}hb;rC;qOe@D=CgSruF}p2F;ge|bGUx)55ih(N1kv*x765E%|sl0tu#Eb
zIqb<`Bxg7XNs9E<s)@5-gE%l}_J(s>aD0t86N%#QF0G@PJA<0ii?xl2kn$Y}R+2W#
z$t@_6uO7ed8%L#G&=xB_dqi)ln%4k4met@nn`Rg)&$eH&HQwoRfLJTVayS=m6m%Z%
zR$zi1KKudml?TMvSjNxZxE*+m$9as^+%1?$w4n{o_>HU0Iv>lnsLjY2%o4-4ELQxZ
z21{8QSKA<jhw}6%>d?Ph)R1Ofg#RCV*Lx~9(3>i$`^P#p&GG&`@A=Ux6{8wju|9la
zwDZBjo1L8<96DNct0yJMZ|XNms$At#;tucY3~>B3bLr{BsIuBXknAetG#AqsQ;uU9
z)1aj}8yP?7Am@tUvE!)K2_kZ7@bzmHMS5V7Jog;81^E{z#FM>r$%+Z=0B})m*HBrR
z&E0zwdW(rpPTbw@@T&nti!SSXRCc3#(LJC`lOcw+<55IVcV$prlTlf<g_S-j);+%p
z0|>Y$hUb=JV77S99X76}CoXG(-FKtl*Je{Yk|sfSKW2x>mN#*`Zr)n6UC>@qg3z*)
zB?T|OgD5pacE63GV9t#!Vj4F-$8zW(Vs>;Ewa;0I&57Cwm`II=Xj28L@Op%u)KDx%
z<w*GU4W3fMUM~a)Q<Ss(KE8G&=2xQ#>QDEnbvu)K^kUVJS#5oH7BAvSz|L8Idf^<&
zd=1T4$4jR8OCc%w)Lvc^8U{%&Ec_fcVqK^K)@lQO%;+k31*G0iyf}b7t4vFlS?2Y9
zwlG5^&Sc~R6aF?Ako@?NqsHnB{iR>+@r*hGav9YRa8RTN;rA;9@j!B@hi_V!qUsU3
zmmUn^^R$Rd(xZ=`b9NKsrzv6R__w7ol(5gra6=>t$ai%dM7hAT=)bi-;9PDpgB)(L
zuke!AG;6T5uMV&_NRd$G6@l7YEj6>K69AY~+plUh8zxtZ(E5`P2N6*!GRxY~k7jD=
z>z^}UlBQsRvy!dn9^<IK^r`mrj1qQ<uVUb{%5Q$OrUA^_a}9F4yHd2^#a4k{2|Y~f
zOw^gcAq$)hfw$Vp9!7DnAv|XQiO@Og+s(N1Enud0ToV*{WtiakK-u^w_Id{0Y~B!4
zfF8$3I=(6Thu%Maeyi+Bt;|Moe9|9a_HiYD6sj}%(l;SF*0OszZ~i~}&(>M=OM)2o
zv9y0t3D2iYr>L>(8(es>^<NoxCF2=>4<;!eidwFy#!iLL)!|%=6TB@zjRGg=%U>l5
zLyJ;h(HMXCoz#1OfWFZbM`AMx3?%688EO{?aV$LJH9#Jx<JqB~dAHPkbKm$`r0`_d
zl2%;j(y_mKKDgqU$jxIXxoXah%m-Q7pIih<(B){xTH9R!2l2131tAFmhoMbL^0^m6
zZJ%gj;R8a_pT>pgES;dWi_5HI5Yaos`&{4D;3n1jTtZs|U~H~2c4l)<(q+?qr{Pi}
zZQk6mkGwiF9P847qs!w)>+sh2va%k5?eDT|_(EV6T_9rI(aQE0DfdQlzUMg%_o2JH
zq+zG35wZO<M{+JF1|vcEt$Qo6dK;GJttWP|`Kr9wVJC&vIfGH$M%TJxrT17@-{@el
z<kGl0hwsGWy-iSn>!a0k&RhM=JEcU46bIjJG>;@hi04C?YVK#`fV<uR0Y`i6aH%fK
zs~p$Igq85Ei#Ht#&^*y-6V+T<Ke`w8<a6j)P0BFWHl>;Ns=F&ozjiZqUiq-=bYjQN
zW-*imVQ`aOh?AfG$?dn+6An=~EZ~drV=MT1ozF-FCg?v>kz*JakLb;<1sF!fK-Wdx
zTlIf~8Y~^oUF=bGMi#!7l<RN8Iu+FhCnIj>Jod0uSN}h*-YP84wdvLk2@--^aCZpq
z9w0z)f(Lg9+PFh-hY;KY1c%1mX(YHq<4zOYUG_V_z5f53Yn}FW_1RlbRgD^TSCoXR
zHBpX?@<1rT@gY3%Q~GB!lcuZ(OiV_`1VOiA6*`UI1v?VmcHU+Z-T6Ol-bKocs10?k
zLaRPzPu-EVkp^VMWizm!mKY*EWe9eSK|j-(m~C|q`b3>zAQ{=6*1slK)4V>4g;PTJ
z@nL<Xj_oXd_WO6yR#WiKPLoWyq93I$Id7wV%;AQaJzb(%63NiLF2zuchHq<rPQn{N
z`2YzVQ@VvL_umz3leM}W2`w&}eB?sXo-Ic%(G+!mf#sP@N^OSUbdqJUPiZhX-|_F@
z^&djRxG+rEdlryUS{k%Kk{?cPE*{d?RwSRpJ3(n&oF}s=tgY_&sX|G2+ijhB?7kZ!
zWF}9Dh<ki(js>_YmW7=GFP(;4{U!BP_S+$3ICJ6y6Pf-30nP5DUzvq1G1B}<G|_bN
z?}q707{{wm$i=><&GPzM<Z6tcM9qjE_Gq$(n}pS^OsA^AlRx2LX<btb9k6^j64`)S
zv+wgAprYA{^!ddHrHpSmp4?LkzeVn<4FkTC&mf_ha^qKA8HH6<7oVu(I@rN0Z<!9;
zsdLz<vzEBXE!~h~ZC1p-q9Bb2{1N9RTSe<XiW0)~hLL1{GK9fwmRh>!O1OY-Lqbu~
z<_~Y->yFD_{%X}aEUlZZ$#$bQ54xtt^abv%9ZBMr>p*je=ETIzaDqQ9^_3E5W8ksW
z*bii^q!G!5TZP1ag0DQbLDI**#<mM@nN&vkUeth2sH31qCqsu%YCM8C*U~EFhw}>+
z4Tzd-6(uycS;RKORRXxi=jlIAE@d)qD0#-si<`aD8K9Lz{!6ujh*MP!Wc!5wR-}>8
zmBe>4NTB~OAwHu+>3+(v_A<Uw|HkE+RW{;(Mz~&{vaq2xWi{LXd58X!ad+YX-G68;
zeIojQ6L5f>QrFPXJ4Z)HxqQ7UBe+7PES*7Q6_1^Oe<Y!Cz?6Xea4LQIH%dW~TEkJ+
zPwwTwk5<^<_2vEH?FP<_^S2l{XD#t@XmlAgTQt0D*Wu~(_LwDF*YmBbu1DmI9n)vu
z_}&nZF+IlVj;qxYny=q2icSj;9+(+@EwPuL|8~hGb<XaUkg_4*^xT$UTPPC4ex}!C
z8KMs5s;(Xt0=u-!Rr#fTE8)LA;QrhPQmC4{8y4WNzt_stc@T;N2rkZ()~EC;UyNU4
z7GjSi@%@}HFaLY}6$iw(hfA{LA1zw@+<V6|Y{pU5(W(SMdvvhtb0;vImxqCCBNKDo
zg(cLdA9l&yH;K6&@K8UbwyHmGaruorIm~Cu`m93y*It`u+(XY-7}5x0k&ghef+6~a
zTM#>==$!z&^<R6a=k57LdGO(L$QqT$IAR^~4a;K8gA(R<^`3`({yW|DvjGv`h|WR0
zSlGc*A}$FJU=H#U^j|T%eMPW1o8ZroxkjvDUoUrjAlh{YBu{qZdAPa~o0z00vC!sT
zQohY^RF4f$11mM?F=UaNuI0F!IPiuqgIODkp2pTonmu9Dec?Gm5QNNrhX8yJX7RUk
z@<3zu?nhQu&{N~FQLH}&RSt5&y9*&fo~pjg*p4c`_{5fxFg6@5yt$%)<oCO--@>jN
zuGo!l;NU;4!XRX%Lcq>kg8f1B!XGfBk#C>fxjDtRoM&E?2m27y?m_1ePJ*qu*IbL<
z0i!viY&^i$?PC;0=t01<Qs6_19%>Pd1&0qFg$Nf4F<1|_+5dZHMq=JlT8NFCgAWI+
z)+2){hj>2C{BZT%;i={J49aNR=0&~M;mSR8CS`-6V7r4}v2T5TK=Tyk>>aoEv97`}
z#W3NX7iW!(e=G~T(`Jcit==wwJ+#=??_I*}tDBPtGBIY6H!%ONfrpI_cPhtW7?8Sz
zHg&H66g^~R!T;Z(CtP#<bhH6gI7Gbf*bZ0yVGi|ILLYnN!?1kG816GCeM*%~A@c`E
zp@~KL!`#-D9NX#agTL67NNR0dn3}8Ju#iSPP!4#i82DNM%L4_(-%b3DVU5N)p8)O_
z$CtT)9aA$24#tA?{<%G#Z(Jh>g7O`j@NEjwlIvqRm`<7uDEiRJ-&dF(S5H`8(wo@O
z^E+VUc^xd!mO++NF1|u{<Y>L5u*?sXHQ=<=k^H3;(1|di^}dcAuP~mOr8(8%qb9uV
z^ZSy@;Hl(0Qt91ms}&m}qL!;rYo3O2<O`hfieJ|Tdf~TOd)Rnp?TU~qwS%TqaMf8>
zn@m^hT3qoLwHkleBE?s4x=>S_QFtHyfe!;QU9L_0omAs>%dHBb?wvV5vhT6IQ|>95
zF5v}Z9&tu}Wl-iL&nE{hX+(({-p<b2{F!THxJ$S}D?Ihl@cC8ZF<+_Ugr=O+w&C%A
z42(=4uKo)DtMUP!7!Nj}BDyiCQZ4$|GxR5?D+C39eP<`6@h66RuNL5hq5=$2qN1X~
zKi;-!{j>O9)KX{!25Hm@mP#ytnP=O~`I0s9`)OcL=s?+OZ}04)@@Ka3K^@0uS%t2N
zkOL#V0fSuq<)Z^_K2HkOY?0lAIlCd@I>K{cb+%r(+KDnYB6B13gEopOajUAnGJs++
zI=cZ%@?*pCfgl_FDxJ+#7z5TT>w~}e1UK#SMb*Ybn8zOUJKLSA$Myrgpp}kn3|BP8
z5WhTH8|ror(XY<%^@pU@?$7g0UMiq?q81P8^jkMJ(BsGqAGCJeMsJ~!kXPwg?_)|@
zeCX)fJ0-_MBVXV)#H37U30t=5a9y<rE+Jwkt*-GxaoVr;!NGu+NQ6_vXO#PK14Nkp
zNI||TQG-ua6K@VW1PxGXnmzxIZREKX93qe@Myg?apko*vD;0)E-1M+u*xn+ptd_MV
zFmBM<zjx#vf4a{zhjEDJjUyQ%#B{#Eb!)d0gc-p+e|c?i+hCVr;v6n3*Plx2?Z}~X
z8z^z?8D3$yY5`k6aTeDGoqsSWu>7-{llG8SFRo~2VNSXIB{1-9RMtbj7QXmMk(tcc
z;9&PyMkBlD31xwAjcUz^XLQWKC(Y?4&kgNfk<2z`eh4Ws9|;+6xx-y7B*aelC81h?
z*B)NjHno`Q#odELnGRi-??>|%M{dlf(ls6R8!=L`82>RtRS(g4PRwyFFY$TlYwf98
zE(KQ?Ky<>1QqCGkY6|m6OO9QKdiW7WRg^TRjUd}ad!xA6^?qU%@OY*}%Wpi1M}Bn?
z1=H}Nvmiz5D*aU^2P?GoEx)ue`IcjUT53b>7@-#J-aW{bdcYRG;kUmRCC}k#q~vjE
z^tzL7x3Jitg2?2j!8R7+*xwm+9NlKwWw@Ouxh-Vx%<o5#`~0*JfjBOD8HCQvx4n-z
zmQ`&I?x@zQb<jJ#h4M&~af9A=2wY}kM-<8;JW}U8Z1a5C(@2H0?E#q<iHEz?cnITI
z4`lc?X9+T9*SG2~-P3*nJ0(2j0(_kaMSsHFs~*O4i@!|clKZ!i@mxknJ?fqB&N{QY
zwfL(6d}<*SJ1;-O-`7)j?9W%}ZXfRtXnNOBLX%_R+{sMGK+1UzJ%ir-K1gFBI<DQh
z{Ab4e*KFhiS$}?q&Fqx7w7jpW;f%+0=8C&5i}!YLA-m$Mv)=a27KQYJgaC1S4Ni>u
z*dEip$TZAFbhcCjCei*W;vv)J&@E`UyW)bTxKvXu8t|#);x}xarvEdfqRJ=>sUYUA
zy9D&u7+Bm1Q}~NEiN5qxM1S_v-SgmPyN@m7v4POYQq+JYl{iGF;V0YhtA<e_NCBj%
z=FvCfnTWpeC44>KpK&>~exEz20nEY4r7B=}ww#J%*jPQ>1kpWypg5X!K8!EtA)cQZ
zg4^i!_FS`YP0tMlbL5cYbX`3;A${0yJ>__O$%-ta*UVkh5;B_4%P3!)sLiNySRknT
zD?)1LrI^yE`}L&LJ%;(ZTi#X=tH<-;f7S{0nMnWjC;<)I`Ws+6wv6;$N%CKgP9tFP
z|8_HqEin+Vr%GsR6O4@fUUQ=}`1H%=AJcILbd*rDf5bI_bPB_+z;wz+{F4eh#%^kZ
zH&_Lnki<<|dkL@$99eL(;h6COL|hsanzZlbbi+k!(lnDI3+41JFIV3WUwhs`x4SzV
z8uLSW*+Idoa-xj(WL&&rZ*<<SEC>XMQl=8E`$XyQ9)C+neOn7O8YLBPog0KU5^<C1
z(~XA`xlgENk1HRqba$%BP;7LQ){$P(3i=4XiKxD*mW?2Y1+t?5%YdkSw;>Wd)+?dF
zj`Prr81cXMk$A<8|7T~=(sr#DWkf$i>(@fcn0gvBF^*3NfK*ydU2Jfz7PNd_`&CBS
z$n}SC<MZAE0z^NZS<ngDOY5+WFqG|6UcLBe#=4Lp8<b#g80?>sg+_Op;ry7WU<5NU
z??%u<Q|M(?>QPU7>o9UHQp1_u;daG-Kuyj+3+@}XsSOgyXKBS83Gpz&A`@~?L|F;u
zESaL_BQ~GkMx?dh`qKv*AywYAqHj)Z-g|OoeOnv*c9X{K(r*7dqi6+Cerxj@YUa5d
z71$U_aQ^kNth?z?S!KzkV<H|lIcbsraSk~se0z+JA3;-=9T};xA3IC9y|DdZYd<tc
z&vU_Y9TH+XIateiMugK_vA@QJAu~7GlVHg)|5EpwwSOlJUDV3Kf!W#5^w_Vw?kJiX
zt|rs$pxqDTz|G0SsA(|O=kz+o?*XYKLIQcSkXf_K;t<LdW!ZDHq1_>i8P@>flZ0hA
zY%6Gt^gu@l?J=A~6T&yz9>zW$PzgdpvX}b%JU1cDPy8NN*_4SR!SA$R6jTZ<8N%bH
z)eP77?io2{_;(BYd#$e08)SHRuA`Z>(>N(tw)QfXk=azKXiWW2xI9saTKql`sve`d
zxsCf|t}^mNWHII%m0xm<4rau6ACrNZNpV%eTc|LavI)HLPsgLsJ&|uX3Ll7>x5uZO
zmY`1qm<An55zE4@zR2^jIZl+BpyKaWH~jm8N_H<;H<FKchAjK7d9dEA2|i==l*o{w
z@#8f53`^jy!mZWb>%ryV=iW0U()%~`4WsM6HD_63L`aRXIpQ*g`*ekYjf4)Mn|_YW
z#pfV}ZOfBoG~k_}bqr648_(SMZNkrye=X@J?puVe@};DI%lNgJuPL#58SB$c=;eMS
ztT#C)A#=T^=rs$e@0b3zwOjO)=D0HPo)kzwMs-)s9a`Dps$#(Z<%;w10K5|qr#2tC
z{wu<iG6eWN-5bNkvgGn_E-ybB8Bxs4e4<S*WuSl8dikI2flcZ}>im2&_?p^U(BaEr
z7J|CQ0)D?37GBjG@h@;#s9m2BbgNr2QQ?<51+R<uqB#)V4xo|R-}Q}51NDu*N3sOQ
z1m~R^!j|!G@Qywow$6Rnbv#+H%vlq-X$!mYUNhh!F5+X^g-c!1X1hUS(1{ede!LHA
zE+sV-33FE;>Urn8YZuxPp^lou@g&~|3TWlK_};LB(5-VU?1|gI>UU_%eaNM-3EXnI
zNpasey|-bIV0*~7@TPZ3tNRg!)Mxsm|9vm$9H_@NZxT{I6E}PXIhZQhWp6#4tCsAP
z+f5AnQb!UJ+K}8PJ!L~hwlMtsCORLC?3Zxkl_jecX+jhF#|XS|X+_{09uk-{6W46m
z*!0|;HrX!T|B3TT7<jwmQ9xx;xHtq?DM`K7?Fush@YhlOsnGKHw(@F3vg)<1GfLj~
zfDqG{F0bXdwPUj|Wvf*{mF#LC^%DW5_*DJaL8q*cV6qGS&dH`^hxF`<S^7vQlKO`1
z{PyW!F4((Ob7FRkrB!r)$vs+ZJuHjw<kaTpz8ED3oM>YphxDNM@hv7kG(A!6PYN)~
zT&8vpCEjPk@`@PJWIn@+ivyU<hwq%}R&gqd9vs$kPhwX*2+owCLUwjt%j1{l9DHIC
zR+>&valtv-j?bKhvcb-H^2uFu1}k5h&5##vfqmqibQa;{2}{0Q;fksICrsinID-ce
z(>6!vbyNCTCc1C5=#u+zulwNWczYujK{bpWESveaMO&-ChV|&nC;A+TmkXsVHB1PH
z@mA-GP@ah_9j^M0=f<EK>X`ro4j=e%k<7PQTO%H$9dK^<W<4}DI%DKq_#h%$5wgBz
zcG~P7;=->Mj4G}ayE}bAuMTze+?}I6_fBM$`D>j&&*Wp0Cu;P<q$Lbsp9=H7-~(`2
z&bC0-qn>9CCk_%E<9gyIB7}E;QQRBTuayJGx-6@y-MLKjz+(_z$V^Q{|3RmZN6eeD
zhpoRN;@^k$&JMFI#N~x5B_5q008tPWY8v(>B9-Xuf4l%8i+lWLq>(9L8u!mQI~SMF
zp2ybxnG?`S;JYcFoY4pvADt!!ndP50r$$Vq-=w3l9QjypNp*OA_bt<FQX>|EHkA|e
zmz;_=YGqrp0If#`#%^fQ$iYGo3beI{&`>PInNlyqZ`f$13+sDw|1(AfakN$(d}DiI
z?8`5|u<DQC<(h-7JTap+xjVh#^Kv#v`!VsGfh5y1knMp7?j+ytX5W-Wy|I}AF>wE2
zhy-=EVy*6&&9+nLcHG=23g^14R<wSi7_1$hg5&N;#P#par0awR*zR1D&AI<CLC=Xh
zAT^Z$3WcseKRs9ru)~^Y<&(Fqq`7aS>3<!*<EsgR;<0W$3?pNusT{6L@O=(3!8=&U
zpU=FSU>=csc)AZl9*mXk1^qp$fE~IPFEoV75$teqcI(JQc<X*%CoF$4#9uA6zEA?C
zHoO)@_-&6JncF+>mQfQNwGX8v3*v8R;sM|u*by6{^1Yz)<^7K4Eyt%S6{h;NsjT9u
zCV<zmX)|g|^lLs7jp4Tg0}iA;IZ-yLUx`iGI!E+1+p2H(x#adc`T_zYhdB2l<%^X5
zQKFvTQVF)YbZjM@`jqh}lgc!=a%)U)*mzn>b&=F3*uXsrB2ZT`*Y_97Ij|L%YH(nI
z{$g%1*o4q>$H$@FCrEV3w~2iSkcAFi+&+!g6qui}iFpTrGu%ON_(Gk{mV7f@EJco@
zTbuqyQ}J8S4^A~gyfb)i{pNT>D<Y9T7%5raX6WV$!kY6<wm%O(jU*sXYqP$;?=yGA
z!48<kf5UMX{=3b{6Q|tyFj_U_optmB?>;l4_eogi@81VgV@DH}_KG2Az<W4CIF@h8
zSN|y1-&-2K(F-Js0Y@a?>Zki7_0jRP33?yD-?UbILatK7a*;cIuq@HI1^P$xVH>H&
zYfyDbZgn{o4u7oBdXzqPz!uGeG^o~L1W*1ii%`n+;@UzL8aX-2&UcpQXJWjU_$dD=
z?@r75>!s^Zh^Gs?vhJD+hUCEKm5=_N&NgzH2|I264jm)Sc56d^oE&;MVV#pz!)B9C
z6Par(QIRH1Zf`Vy6+XR+fTVE;xA~g%C}*Cf1w<jeKTUAe!H~JguBAaYE1kbsqoG@W
ze7-xqN$WP8t0DiMJcjZKxQ6n0@vyvcLfuKei1Ccilip<Ql4oHeyf&Ke2zx{)Yy8s4
zXtnGbX|@lKr0O`r3zj#Ub}H`ou(d6Jfnl)a6^nbSgQpR8d7zf|QhvZ_I-oOd-3Vx5
zS+df&GP7Tw+&XZau5cnYj(eT2Ej18_Scq6qxRaGQXM0&5hs|S{zU)8Wk`edyeb<+<
zm5Tp)D9_$=Rc#1(s`@pEa)(bx8MNzMB6oG0_HMxQ(fj^ua`^E2=b8)G>HG>(5LSD_
zv906L&xdQYi8zgQf8u^6G<>xiF7Wx+bo(QG4IyA=H?Ku5*ueZ>SJBs~u6&s|l?ojm
zcH_(z$iM*7N~`<0PEejL7tcRtqpNeoz3JejA=wVpVjq=-t&NY;u`pd1?(3Gggox_H
zRFy+z&0Vz{G0#*dla<#bx*|f7bNZD63aA4lgulbs*Ur{V96<rZuW80W=N$X%!>x_b
zyLrrD#QS^|WXVgz;}A}bQu)_^=24s{E_8WNiuhPV#qT0P2J3K=T$+OP_)6ubfuhE!
z>9+usTW(&nW<=Unpnv#x&9&CGT}Wf^QN@>>4J$rtK_?iNz<XJpc4XwhuwE%}-#W4t
zhYJ;O{kfT7%sG%5D0e7%y}wHyJh>lZEIkH+4rJkTL<WeYp*urz7_gBo(~qxMXUgb_
z+g7$Y{%C2p?{T-{yo>R{m_HKbys52=Z7j}aMT9OIX}{4;{KMw=nJtch><Uk^4OrtW
zf>w8zFni5ZDK*H`)eyD5ymec_Ty`$3Zy4a0Wz^bjs#(7j7!mXf&2cXMiWeo+C<%95
zQ+Q|8Q)8gpetNaU=Qnxwaurq)k@u#Alh}2h;fBH<qQB@R)TL)=KzAj-%oh#y#e6tf
z-BN_rIYexr9jXrvMsC$`unIeq6-khFPffg192s`L5&CvF-&MpmalWS57SSSFP_3N?
zg%i5@1E2u<N91+{9WMH}+$97??Z!kVCQF@aE+GBAy8{n}Dxx3{N!<~vX%bIoSaLn!
zYWgW4$0XxuxycKZ2Z7#Ub2Uz-gLG2(px1`k?OidmJ&Q&%QTS}X)TFCPKuQv`b35;S
z#@k?$_H&B_Xc;?t@BgN`)BRY4Ece)><gFH>t4$V{X`}uYxCij&Ra%><1A41oiqI(4
z!gTq>0spA}qYI(a!)H&QNS{zamY0E*>7PQqPpKL<0ru1l9~-}K;@hNdEfPi$jCD0J
zkw|;xd3=kak&IfJu@=Rw$bv<Ea{GSkBb9-^$NnRUj3rI4<?nt1Pppq8#;Uyc)=C-D
zpCVn%c6zdp3z2Sv*Th$6XxCtuzG-jRZT*UjSI?Bv^9k*cv)$4)hIn*w)}^;?4XI@}
zN^(qTv@({m2B&+r5ByF=hWTx{n{7+j1X4LSGrM;nAzd8%;HfAB#lMr=E3D6GK!_ED
zzSo!kuc0r%@Z<L+z)z$$OX}t2)#UrY%cNd>isDuHLm=jR4f8TX1T#sc9G!rf#^fFk
z{SRz11%?Ly=TCa8tx1i>CSWB%+D$NfNF67y8nweDGZ0M5EBGU2vb7SgC1W6(xHF(P
z`^X7UMdpFjH>-4vN>2)$8W8$L6KZNYGdA3Ri!(4kvK7ed=0YTf6myTNwsAwtNwTzz
zxOrBOQoK@)-_pYJiHuZ)wb%$~qbGXVqN6sp%~fS=&w^xmbXP?m&)(QRO(Qtu2nWm7
zx9o9gF{qUZ`=|KunvNfGv~k;X_APzMRWe-_4UF2j2W(Y->=pJId3Qw*u=zvsaq3PZ
z!sedA^;IX=J`@a$=fx^bSM~i#S@SiOWP2vVANpfOGN@XudW7nHOnv%YsW`gbJ|h9r
z{4;0GSNmvQwySV-kzCAQP_uo#z6+$42qiAD2|B}Z(G*DN@6YtM#a85|&||n$tI;RQ
z+xr>#NA&#4^7c&Aa$`VQ(e~a&rGNh}vT_nvOn^i0^R`fv?S(j~i9+FubU*UGu(Y5!
zVc5{eAPX4Fj?i0VWD*?jwTFSuG4;e3+U)<XknTj4xtnW@In3UDY$6ryUs}{nHI^K=
zX8;d;I_#l*-%uh{dHkHthCRU;wV?{3r#)b9U0uUy!=QM)XAy0y0fUP>zA}tw<HQxQ
zZ5rvk<k{}!)6Gi6n?%xXJ5#09MGNHjg3MWLCyy{Na9O+M)GZI=bvQEVx-*?L2O~C8
zrx(E?=l)Sfjy8=1s9SPqzY@gkXju~u5_TAl=YPvlyO35RAVek0z2A?!<tMqx3@)Re
zJGUWr&j_SjNnHX!NJ<}kY<Gmo6);n#V^fQ7wAr=WJ=~@W8~+q#R2vrGUQ(hD?+&eG
zLRja*WWO#}IxV*3pZ!aJp{V`Go@bGPH5B&0qxL`X@Y=F7=EKP)q~1G2N&nbI{<HE=
zAp+kE2Js9LcQuz3JpS#xWMp|#a22V&JyAfVcULS^A17;H;iK@6H<3M$d47ajT0z_t
zXX*uIrTtNo(V{$dqzl!hZ7NOTYz&@`PGYb~Y;f3w%}#1Wj7PuO+A^s&ltrEQ3=}%w
zp51Aacr$J?4K*+_|6EqT2S4zQIajUfheUR_Wt_=Yakj8B@QX^Ljp4%GmJM*A13;D9
zXBs!YsqT?FM!g9e;7HZ1-Ffnli|AejioBgJH!}aOsjMp7Ma|roraiXPbkWz1$|T=I
z>~JpA+uw&o5}#jV8RtR;C-ExD5znsuZ3Pggi5gt|AN!xWro{;Oa@z3b<k8(jqP*|0
z>4%0$yR6a4!zD(r6cTi_cpV~Y!XPBSFYiD2iE?jdQQUfwj)n28c3??!*QgO1s@sx`
z3BQ9ZMx|gqN^8sS-qWtn<sl1d)dur=gTP9oOq;bZCH8x}iw!%J?cuP}0yMS5%PUqv
z#74vR4Z9;i^61aj|MJ1#ukC>-22RLrcxBhiO)@TBBOxoN($SHi^@YJrRxXf?e2%=q
z<cp38mBd<mSQ)+ZC=Qf_T=Z1~9rk2s=p9Ck57NBdYvqHfP=I1Ph+X6JI!9DyScBM8
z*cE1;Bd=QB@j8dG?{MqY1$JM^9p>2!CUF#;(DnKZMCSaAASIzqnLZF@5&iDWIjf?~
z$UIJoAyHYzO|;}r1n2y1)*Q;tU9h8Tn5V4oSD;I~9nRdiUs~*Hmt@}C9WDEj<K)Bo
zONcI!bJtFg)3QnUb)Nbu)_$qQ_=A>k=g?z^Z-hiwxl|49?D`ZfZoT{}qud-t#?yFa
zk}`++v#Ie(sXDGNB0HbupcD3kVC_;U3;yISUT=yE+2j}UD)$g(GTE#RF9j(afXT$^
zbF8-<BEo#{*n*d8x<npmni1_zbVv7i8@tCA?}6inliR37aal+%+U^aWh$C|j#px4k
zohL1758*7~9oeZk@j3pKHS<_zlb@xfmF({_BSfOZ_xh%iOr<$Up_Dt=GiHvx-u4~D
zDVfu~etRxO%()6@XQA}OXWQJLD}3%G<Ov1jG_U?lyz*Mg*h&vir<wM_ob%=!=01LQ
zT(1(F;D|X8XUzINule|Tz(RZ8eaP`f=Fb&1wEAOx&&NSd4R+g&E0;>coh$$6OZm)m
z-+tI{#pzA{eM(p4aSLKDwEqSwAiz+_{}aE->8x_e|BYWTyx-$(fOHHIB?ZODBh0UW
zZIv?#o$80?>me}6RB%S3`Dcg}h9xbAzBz^sfbCEb^+Xr*@7<c1`o`aC$CH!&uAVq=
zJoAb{uSWgnQNEb}w$e97Jjm?9Duz>hs&M7j0e+VjRVcHaF+vK0uF6zX;3xGTKF^GV
zMe3;TvZ{1RupVHL?D7M}kZR57!B7&bN3LL>*krv%pTXv<B=2+1TIa^ZIU)Z+^Sc6O
zoo@40i$w1FA6a4@Ze%{R$8PEUgl-R<$pA8rx|skBw#^K95ZU0-2l-hoTM`IHX9p}V
zdYZuumpfz7{P{PBTS0Y%q6jRAPmIEEhmugRj44l3p#Z0uxVaCbkf(PuJ$fVc=T+j*
zyFbT{*)l%;*uYkd-3>;OlGLA(v`5#;cROF~!lTj~UxTw{H(972GV!Atq2G(@tV1J4
zGptBn22!;ZguJkISq&4;j9<5$jS@-3MP%IOX2;#c2oW?nE9<OznIAO~fVKE=r>BFQ
zo&$|y1S)7n{HxE_HrysKGG2!`GQ3@(xhsns?FFU745d&9g1_C<nl+y^lhWY8^Kx@_
zhXOK5sfsu)U5TzC)9Y4I5nQ=ixgh`Wh%a~$**@Qq_M`2uqnlhIs2Ue)qvDRsk<~Vk
zc2nl&Dn)atw`V-C1VPbK6jsYR3?Vn7ai#1X2%E73Mr#87wC{=e&MVE)H&`UtqhS;E
zMF^#{Q(8iqnbm^a<7RUG<Gmo00QFijwaKOo6*`{g>yc`^#-9XAmA{q)IGe8{#QfKv
zmgqg6w&|!0S7dIdg>EsQwdr<M`8Z4VVy~!U{oUez5{yY<rqNy<=Yr8*IHUGf*v?(w
zqL&j!!Z2oEbm3bQk>{8eTs~g)w;QYHHk^hKpgs2b&RLl7q}<gU6g)*%%8vV%l88xv
z1GjQ<=wX{%CuDS%dp*y9zwtLJgO1*l^JKLAnx5c!caejYL#D;j77>0*KsN7OLVNGR
zqcC2ls!R3pzhxL4)elNwCYnM?!1(rGMcb6}`q5HjRb5^0!_CPoqUo3;wM?vnzYH#=
z47Q3}h_f*mH08fil=Ii^)x}!Ny9a!wU7M(gZbX(FiLb-^vi*t*3LkA}@T+p1if|{@
z3iQ6e%q8(xi-TlKU=tYnGr&krm(LIYx0*<Bv4(PuPuu}^EMJu6uA%62y_Orvln8*U
z7n~#JXgjJ$qRTAo@>kdoJ^hlDm3TgEo^4&=+p8XXeN7?sIVGmOX>I^=;+pmul`k@4
zqZC-{h&H@Ucz_BkF(g^*w6kWy>_1hjnw!I(8Y_C95fVHjW^&7J82CFmEcxS*!5*}Z
zp<C61gsdw!A%(74%O@gAfZBK)v)>c|*CF+mmWd^c^h-?(-i!ZEO_ab6k8`N=+2v0G
z$kXg18ICTu>`$3bGg|*_GRg!V4^NC)&0|w%%!D;0Crf9q^4XiNU$6XqUY=;n6X3o4
z8@s81T9o(_ffZzAqr<vz7ab87wghKr`>tu|LO|36^H5=@;NBxm9j!V+hi%g*RWd!S
z7El!v!1QcbZ+K&4d|RBq7<$d=8bO`ww41ZzBo{41e$uLEdW&*$EH|w1-?1keGMm$q
zjz8i&r_wS?Ev<ohcipV1amr~^3_k*Y!riIrTfllhf~D*<Vm=1SgC1uso^m!$N$Njd
zZYSt<2y8qS{Jl68*=Lr~8y-d}@cuL0Q1n1IVgfg&{~X+pE-~2cBiPJgZlS-u={@UE
zXFJ;C&tLU3zBQ$lN_`X6@`N8xv88(aXz-%VcNL{otLeO%y>jj~&`*4<H2TtqacI5S
z>mLVEe5rWL-@+ujL7Zc{68Oik7+$muO}>WwMZ_62Bo7KLTQQ+n1jQ3N$IT0)zla>m
z*}Z>jXz?X|{AA4pba2QE3<gNqBr9n=2ND_V0(O<AzZBThZx1Eu7Q7^GGmbHfG!}7&
zr549Q+n1$y=T}~!f!1iG30ak>PY_0ggMF==#=9I|5A5yAz=r7`vsKlWN@0XI7_XFb
zeERz}+&v8uP^`ApRH&`KMYJ<bb3tbQn<@UI;Q!77OiNQ^u{!@Q3n9q24-b&lRWf8`
zWTxu+`g$w^f^eOBYjclITy}39AZCc?ClSrBP^j>oD6t}b=1loxr*vpO%fdor5cltd
zoAJVVAT4K{xrd<ij@NGb_k|giYb~Cq`aEF+fYfL@JUm1han9di4;L%<k7;tu_4)l?
zhWek0yQgPG1_bXf+o<!dfkJf`fB(!W@9oq=*`*Db>S32j6O0t)wM#|l#?w$yACf10
z2T6^G>vwl|5%9kI1IWerklCawa&qD~<Ts8@PDV&4mJ>BV;WrhU6Zd?M+*%mZl{Fov
zQvVf+e^fi*F(h(`zr(lmL+FL%-M-u3Y_XP3)y~eg@aUerASa~CVN>G#SPt*KOs+V~
zgZZto^=_ZjmixXh?o)f2$P(Pv26P%Up&3+I5L%?&jit|z4)X?j-p@Y`hla9Q4_2{O
zY_zFUn)FH;erTgwvSA0(&TFZ{k}hsYMPFSF(;F@dYQ_-6M(p*5IoZ;ss&z;U@HrA2
zu&M(Ta|3#diHQ+h;(PvL$544vEV@v&_jzYi<#I^wp@6JZL}S9gb9Sm)C>O}{_3M<G
zsvK)rU9XaS#!UQhhgK(YTWFZ5L|Xsj?yl8)s>F{YB?BsrNZPGWc;tR|LXs|r_Q`5r
z%>Fc`>86(cBTUO5H~AU>EoJyd!^cCwfci8)j>TW|4K?ufbt#(Rr%n<!6diCLX*Baf
z_r!0eV?zNN=U>w^aivv?vyO_*keLQ!!>PN+V_FIO`;G2;>SdN=UE9PV{?zZ4TfgwC
zz-y8_BHXmTh1==dyV5J3JI%mlidiDT0$TNCJIF4m&>0yBA`L=Fc!-v<iqfsBlr$-l
zyIBA1t6brSAFI;ZTLz5z5^Dd~5i#JEX9cEE^7x|G3jclk1ET5K9ae=E(z#`6*>Sp$
zR@;f`=my2uM{mm+Of=~Ld{^0501`0zq4pfaX?|=-1=$cVe@lk=Zmp}*dJV4KqCD6j
zm+J4(yu*=C+C)W%q!>G#PPWhal=O$DTOt71IFPxaFfu(`>9lq0Pz?oE6|gP&U7J7o
zd&O1Q2RXiM${xq^nT6mFX_my210_z^&l3JtUo?esyLC=~oJiv8pWn}^X}5fwxG?3V
zfTc2#*j!;}2OQG*SP#>(C5q>a!ZFCC<E6{?SOWfdAlUS&*udyw#{AKi+mQ@>y3V@!
zje+-t(LGcRUq~`Ql|2T&ArrknhTh@M^FI=oj*!f>prvWh_&kC;MQyB|3i#Wb7)17|
z(y??@*_z?IiN=%l8dv<|4nIiq<7|rnU4u1K8CmAc?agvakXYOG)<L5~TZQXiRp{Yi
zcx33k^^&tzo!#o=;=XzG`o?DXd0$i+kGd@5n){c@z0Of$O9`^GCStFf@A-=LW*2dH
zW0@PVCeyUt)v#XjAY<KC?FKUKmpH1m?mo$}-QGXl63CigW*y8$#VdOz2TYO3WIABN
z!uNvR%fEs@*smv8T2-55MW>v5errpN&Cw^ny(HB6fZylhD1q>V-6mJZF9z`3V-hd`
zjWO@6>OeEQ^=;#mjHi7?sd{hdDmp`686Zdq5V?@Y5;qR^*B4aceE0-#I1W)Q?vTv|
zALh&6+$h=Qrtm5MyjcSbChWsnQ~hA{UAJ#2`UYTK10IpqwVZFHYxR9)SXUC5{u*DK
zFfF$FccATJl$2Of<zrO)2g+kc>mZxdbi5%I`l3YsJoDeqGbhJ!0uWCw%O+iV{^{LV
zwaGf~p{xyUUAI&%hXArce#le!ODv91(PmGsc@59V;_NpLOlgaHm0uv)ZHtG%pIrFP
zzvRI=T&=M6-@SyfeKrT_ZCuM~D?pZ*5rTY6Kgki>Jlp+iC*0(Q>;C4Np=p_eG6(X9
zg`t#<bs!6caZoK&t1sPdo!tq?<)XGTV+y9-pSjTn$kAqC@8}WFF_42Hyz!kjags=4
zQbSh5QTw*TuS>^dUqx29)H^MTjy~jv*{?Y|R_c7d{x7@G(Humj|ArJ0XdC~OQyZTd
z|E+;dDgQDZh~;<L!^G%amIK&gNbc_L?C$&kCeQk9aZ`_LO+bWSu7XM%ZF`&`r`AJo
z@26XDa8E~w+U;ePYNju&Jjiw0?XTDGtukp;dVa{@z_?sA8I}O#%CIYQ<3Ax*O$!`%
z_z-%~pS9s){=`l@P<@5sXtwiV?AJk#bJN}-q|6e^nJbpP+8)~99+r2J>i<K?K{Vu1
zm2<2rBS4!98xj+r$cZO5WQ^{h`h}P|X}?FlAi0}xaG5=Rb=_P@Wvnz=;$_qAgF8^R
zl^M(5$hMHn79NZG65Sz0q5ojcbcu^Qy@2fS+TF@m2rm}DjGJxG85fLl3wR8j=t;qp
z@2;i@2Zvyi)KnWx*glHirb!ZXc(;^d2V>ZHpOFg5&ifWG0%{snwOSWb=bK4g(k5hU
zxIW+Md0%6bzO2yTjNio)1m9cG$&F={Y&~4*lJp;m24~INpOZXWPm|#2aO1UT?i&g#
z#~XH<d@s`b<nFU=1Wik4;16+1;RAr~m@_;q^ut>EkKe0`tQfHi?HnBgrGWNZk515K
zl`4EEHDh&^l9lO6?gDq9^ZZbt|3m;jpDM)I#yUE?-sWTHN4O4v(#BabJb!-X3`d3o
zakS(&mPzS<ARvu;UTOo$QH2VLI=yvoNyxrI^LjN9lP#=tRU>T_*l<o8SzzU-ncYr-
zGgS{(#X3m+p`^^f_h)ll#E+=9pobTDDwp54c%IXFzKja_N#~9(j~)F=kP(u7HYu9l
zo^x;$c{vkju&-PJxN(AlR(M!&M}Y`x7mj?o`mJ^j{FyRH7aW9%Q;=e09y61#C41`E
zYF_g+HJyYzdGdm68<ZP3Pdh=81mLuwj7NQHWIted<xR-U%ox5wB4Za|A7Jv@f#=aM
zY^7K81{{9Rz;Djuo*o%WIzXko>Qj2bVD10+&pVO@PVwrTKwGhd@q=&V@l_QC6rsZz
zT6}|7-BOmV!BeFQALqeqK31^F+0@P+K~*Bd#~X(x4-bLZlT@Zw*!_zyfrZ8q-dv7c
zHtR+AiyxL_hxa!=Hx|cTN7Ui3xv(ZXhA%C7n}SM^``D*@ID_Hso&mCa-h7=mMI^5N
zdLqq)4|Aw$3;4sjJl$P?JI*~nzX4i)#|Isnwnh}C?Bw@%ST3q(QF05`&V}9oo#X(P
zDl-8f^$k0zNhAN)mmv_)HD76HY;GR1w`XO$)<Ftj&_<9Ug&fbX0l#OHjs@c0fd0me
zFOJ?~a}BE8efk&*Jokm?%e{aw9U8ndyJ?uL2JvNDGg0VHUxA70z`XmBFv(Yei~*bV
z34qf!CZ6rn1M&JR%t@JcNd#Pw_0L5@RbXNmb-v#Pp1-3sSSX5U22Jy&WMhPLJUuzi
zE(ifhEPENXYyZVnQtwi0#%h5O$gf?HkD8(y4Do_ycS@|7>bt3r{9g}0emG`cLm99a
zi9P4pXSyVRVL|lXRZQL)yyq&@AO}i4Sb>gB-C?i+m!^2*EAKz7ofyEX$+QRHGfjaB
zxG_6t)#B`=iTa)E3b|ge!4kRd*L(!u&XWUuk4r<P+fn4H@5AjdXAD5olSf+YaE8lu
zpx)T5#&I{4&D>aQ=<3xR{kfQp*Kjg`b<(u7RBy?4dy>{=6QYLiQu#$O<m4hQXTtx0
zRS(^}zta4vslG<$cv(kwHG9qPxvMV7uj%d_PY$0ox)6`KE7yWY95g^Onu=+&hDBB@
zsUcr}yr7t$pXMT>=YRcLiC7rxS&Bi+-DF0kJ277S?ti=hgi3WH!P3$_=24k3Jv>AW
zIg0pfrinLxWPOX_$kSdu%O4GSJ8mW?3AX>n;M|0StsG}983#jiRUp?iu}jS@c$xyP
zl9c+SWPI$0K-ScyKAEnkAGx~HxCOY|6$bowANMli2Gl_<rjEy*z9=w7q0>`4c(KlQ
ze{z>{>k<MPMZX*qQx01g8A%KE+gNGtZ{G<?b*L-d>=&9~cUk=(47;UD{hwNldU*dn
zGWtYg_vN@ha$Arl)s<-BJGV9jmD@7gy%610{)$<`WK2Vuj;A)RS2!-FHCz{fy)3pY
z2cD$e_;j%>5$1o3su8d}?X>z`?ov3fiXZ%WxkrazQHl#`7r3?Vc(BaM3N{Htz=6bE
ze+JH?-`=azX*B<l@2r#hNeX0AM8omHCKoe}pHG?BX+b#H#=0xX3ebMNRb0~iBq6y{
zH0Cg`hd_LTuodac6!$%T5b|XV6J6Rt(H|JJ{x@5;2=4Q$@n+Y44DRrcLYi@{IEOAt
z8lpp9TQu84gfUstYC|!TwM8sp!`;tX&2HY$Imw##s|udr*_{+U@~=;i*|U<4M~M``
z(I&+frH^H+fNL>7mgOMu>~vLz%>EjlOMy&W1C=w9rSCrs(Eo|s&tm@!jB+TY{%e^E
z?_#3_ove!BGHarShK6?1h$4(W!zb!*JRjP=G4<870EkB76dW8_^S<t;^J4kU1kt4B
z(M3T^K8Y22EE@WBbTkyVJ;;Z;><9|QeX)q>3f~KL=rlA?8xasYe<_qwgf3;8u^NhX
z49pD%IiB>$;?D8)ewQL^a<nCXY^)T&+jrd%bfAt*()2B6CJQQb1Pt4noUwsMinG4!
z`&bSh3fOW5yaz%gvxiHj)3<)NB!put`_Pj_3;;0`o->6jq>=C$zbLY+_79tORgT`q
z?}@y?pbZCeT2t)-JJd^5c{!3ov%V6i?E}nR)JF3#;J!(L9g2drIh)gG`E(5D^UDrD
zlpx`DGoQ3#+n`$N=qNlK*%%zG7cMw8jC+%$7X&{$si2?G!$p+20<?s~liiDyy;@IV
z1E1U8rP063z|!JOm?kR=*$X0mgWV30b^1+x8|%U4j-vi_G)C8xINBbY(H~%2`Jp{v
zVZav{mm*YRw-Ta;eFm)|)1RpH(~oeI-*2Mpe70G3k6!Y9B~-b(?g)@JqHNdz*KU8*
zmC(f>M?XbsXRdO0w@)QK2t2zF0)Qy!Hr#75wADExSIPRIHCGg7!v-`@{2kxO#v{JG
zNUogs*PQkG#SK7LkgA?)ROFiEU2Q%%A4cYRjfiWwwoTm@kq76E7us<Sh|%R8sEn%S
zzBCUgIqn`BnegAR9{oz8uaxi`x*HP%Zd!4`>8D&z3?|4>`|n%}oL$YIHa6{dd};t^
zz$4`TqJ7wu9uv;)2oL)96jjr$oT-B%6$LcRS(m|um8BD&0pB@utXM;xH83FyLQRAG
z^nD2tOK!lc2lU(DF5mCTku1%eU9Yvi@>@cH)tzg`taDqVynD|~d%quU?dOoa(l^%U
zi*<Gm=RkiXQ)$Eza+Z1SsD%lsKN%SMJw%#6yOR%&aWm*%xEl`RgVvAW0k){KLZy6O
zXJcJe;l!oOtZhi2)p=$At07-bKEYih`mqz;5zE7`Ph!hxn4exOG-EdSt#RJrF=v^*
z+<-|!=g*&S-!s33JUlj|BsXN4J#ld?vrg*Oa%k9WC(^tex{0w6>Z3RqE~_nkTF}!d
zSrLU?C&ZC&vY(J|Jf)Bt;~zkNe|X5l{ePUt6wd!}d3IBghW;Nej~PFlxmFY0wUe<_
zsJk?rKmef8ft-fl9U5{fO4Imn<XH)TKJ2?XN<UNeNU|s**}QDe;bpG2ylYfKF8Em>
z10j2w)Q<s1is_#G9~|SD)h3DFyniC%SU&c&=RaZLhkP3fL4N%-$I_LH2dsQrj6qH>
zk6neET?s1?V<f|{{Kc6Y*Ow_odrc>BAI?TUesD?_g2Y@1kIAPokt;wO95a~>MB7lJ
z7SX;ew!M3FFPD-v<wDijM5acWefYQ>+Bc1xk$hK4U3G@r+!OH^F@`9lYG?Sa^0G{)
z+xkbX1&m8ER&C+(W|`3b!Y!UlPhP>+G>1uagwS>y+zxfWh(&$4O&(=0&_}U?O{s#a
zTFwq>Gx_Ww7Z`yTvO)(;+ROPUKyLdz_m^pNSS$K+F^Jkd7-@3ZuEO5|0mijUax<Q5
zs_mzPNdZkwgru*c7h7qonv+BzN1Qf066d=eB}Fu9XW-<&An9`9gdCXq_nJ;+OB#V~
z53TIJA~Sotg^6zjGCBG`Ds(jr&z}yAiEttc&-=CGWo|UR(dBnGS_%!@2z;}&-N3@W
z*io%h#GPD;k3(#~(pPD2dVaWul!S(6)|P5&5XY0zOfXSXN7-<F0zvQ?>aSlW95ba?
zlGFRE6n#f-tEh+z7IGE)gxS#NyY-}aGC@S0&hx2bdC?jMTRi}wt~DdgXN$t--HWY5
zc0^k9K0QAIXKXG+LcYo%z)=A7m__k+uFO5r4EIA^3)xRWZHkTPfSWRZ6^kWfHK(M4
zjjZK?!XO*3*RY_)C17;5z4JJR4&_sJ?KxxNi2n#zg?a(QJ=$P%Pgis%D`1)!%-pbl
z_j*2Aq^L(UWDLXyygDVlgkWiuOND;I<7+AMLv5PGml43y;QPFc9JN2ngAOy@c0M%Z
zdxrsj7y?&MOpG(;2=I5Pfv!!)K&1EdC#Wgv898DdThK8={#FO9xIy@DBGUFB<FWTw
z-Y&p=s`-BitX)}*xl2cmxigMltR_i9>lIjSYwbh`?6NGK?f22?-d}zdAQ+bfpH3@4
zXYHV!3@X}&Zq4U<TxTX|w6K<FJ>4`G&Bq(<#~pQbI~}#gHi>76KCTZtT=WKACspg2
zONHMYNv?{sjjI_k|0TWyi%Ngq{!_EQjl>Q5-zn99QlS~=KO-reRsg2mBb#{lyQqfn
zs_x#~_5*(KT?Ibb2U!fd$Xe@}cN#=>9~o;S8LYKK$B~@ly)SM>$=2GUb?GrS?NZm_
z=|#(l8>JsM)Km&<c$_$LL-)R2O!8a+Cj+US<$<#H9qmxnl;{4?Mxkx?#m+d!k?@__
zQ`LEXD@^fy#KwJYE5H}=)Z>lSkZc`VKKs3OYx)GI-e8l*%cjARX)U1rUP}X&o+W|l
zMC@6Y#BDqy_+m9MD(kX%sT(=gpCwSq@A|u9JY=be63P!I$DgP+&4h8JGA7C@5;m%r
zI8JehEmY)K#^-Q2kT>_p&hN(+GmuQr%k=PxFRR%>U*VQzx-8TvT))4RpGJ-e4<0VG
zbK=lzO-_aC)J}X<;|x&FcFIjs-?q>$#^i0{&#gva=3)`^qoFbCD{)TsIxRj-sgE8u
zv3{7dGBZZBV`-p&Ga`O87o`Qd5rOOc^-yD@b$N)0-zxYgN&Yb74eJNumK9yPArpzZ
zw5+AG=K>DX4u4or!<}EzDF>Z2WhIVjXm>-KwLnc!W6*>*M024Yy}r~`XMmTTmCm34
zd2@@tT*MoTnC+eUu{PuQpW;gF4FmF2F5~n)qu1g7z^YEK2$Ff{kw($`<;A-1j%Rr_
z)lz+ZXFjm|ti*_Ei?f<iw0PZ~2n3aE$c7@ZiFN|wGHqNERZ{Z%Crn^MCiKaKf0o1l
zp*uLlkJFeks*<%4dh0>b<cfb0m$g~ow(te4{iWR1?NiXZ9gXBc;m!mq<BSRl+u;QI
z!TPLZ-p)&l!R>DUeqTt)AlDDfmmo`>`FVU=-Nv!)uP=+2<>^RN{50wAo7=Vl>AVUw
z=_sESe&_sM$Gq4Pabs^}2hd)gaf6JTUKefnZmjW98(uMI-wguth+tbsQ4gnj6Y{&~
z{qUwo_hZr_rHUb{**_h7rTryNcY^q99k8ejkqdqmgaPQ?Jsx}JbNmnBiE&v6z@FAH
z`_aW{40Lz-lVm`a+&%64STgb+;)jpf!h$tTY~*X-VUO8HdBSd~>J5@*+0({41sxI{
zXJ?L0-rvY$k5XA%8@@Io>uEEP;S(FmT|AQA;K?*7iGM%QS%5Mx)k&cYFtT%&b*~71
zZ3`wp^C0m-_ggC5?4$0Eh(_Mcd|8I0Bf(uG=K#vv0qo|}YFH{u@#F-@z_M7}@xAFv
z9WN8H2FDv5h5H-7P$%s}zc(8W!Q*t<@+^}@%o^Oc^(UCPS^th)iu!+4AsrJJk^d*o
zU;kqf;k7^{fa-0A7RXLhlj3A)RZ_~+<^+?Afu3&tNDPj^n#;*kv}Q)uv6^rCe`y|o
z{}c_Axs_NC#)7lLf|xCexXmKL4++iS*}9}Fd~$pg2U^P}GSw8p>JDnG@QC~13V$>$
zP(y9;H0GOKyN||Cs@8X$9GNbtz}@nJxk-&;YT$0q`^$JC$W!wX|48Z5yR*by*lt9S
zwopr4kG8k6Ze%F>)9m#)eE0#ay;FJo`EAlx*@z_5JW{=u-T<)@;>O}uV`iQy8ug)5
zMOL0`x;~p%8;J7w>5iV<pxepZn9G<3pO`}&`)MTP$LZS$f@2=bORZMT)R_jU(!1W5
z$B4f2M;N8pg6h6%Em5+NtK{UP*Sj;5&8_OL6WsS?LIqWasi{NdX2@1MLw!=Ns_{ve
zd5lLoCh8Y-6MzU|=a)hjX55QT9^y8Yy0+=Req1&m{*ynu+V6aKuK$T|L-;YG8e(i|
zQHp(I>VhhLycp!mXV1K?vou2x;&Q&Bug-1X=|fx2mgSSWQXpb4>K`(S|Hjx(orr9t
zV6^JD-145oTGc7^+XodBm5OLAIouHp%^LMgwkWaUFCxJ1?6c1u?xt^S4CG@R>AO7f
zZ#A2Dj|4bT&l!zd+fQMSmxGSfD)TeiJEFI^I8mJ#B;D^LsI$3~n*&7g+p3)!-3>c5
zF>!0w#*gN(j>>c-<mKhf+nnRFyUM8tJc{Aa4%b&Jg8DrZSAm6QcWtZb&H$EO*AsYB
zJG&MKo<l=6r7TFkZy<_27BMk*n-@>3{Q)D<eDiS51XLXJes6EPU+ls*UOs9#K-H#6
zW}ews&%#{Q{exS#WCQ_Nqq4+m($p!Ee?_<zMF>DdM)K(Vy3xk?nx^jR&r3`9*t_s%
z7}RT8)6<4O%NB-Rz<kP9|H<DoOiS5*a4dzlBy#cE*Kue2eVgZd`0K^HaXSWMvt0HU
z`Zm})6up~pEEXVSeOTX-Nw^<E#3__rVu#Z&f<vKTNf=t>pj-eFEA-;z?bp9wu5H?1
z_o*z-5a~2ud~c3vM{msst=8Bhxjd{f;WDYJ7QD=C4{DPva>LC(xig%a*gHpmH8FX;
z+U`e#vtAvQiWwV|uQ@IHCrej1_XS#v0cA};Fi=39*Jema$LUyfTC`;KX;ju)a4zLL
z3|QT2upkJllF6y|N;{>^3q)`NCVr+?-xwpG)#}bBrT(8yJbZu?w9FPPpZk9~hr_#0
zDE@Kcm&n_t?s4&zXPy&KTtsu-PkZYvcD}K6EAIw|)v9eIOf#in(jthzthY2L7*!Ys
z2lxPt*stFT){#P=cYDN=GAC2HG8=z#>0zLafkxW4XXcWi?M9L1DQ=SV{~ud#8B|A`
zc8w;u26u<x?(Po3LV`QNZQ~As;7)M&;O;IPcXxMp*RzwEcP7t#=NDBCRdn~=m#=lL
z%$2w0%(xV8s=`quETD8RXa^3c%}tLSC%xlai(51bpnGjrRj9ledxdExYY4}QFAwSr
zt<QGwmoa}b83i1H<|cz%V>ppxUSpnDY%Fm0cLR{f34I(A1co;|X)NZRCbwMq^mNDs
zG+bgT$2;FAZx(p+MP<GxNU!V<_^Te@7kkovvYaB7`>x~t7$!`}Ys;`5UGfs5^8rwp
zf4JR?znR!E{IYM0F_B`o_sNgFP)^pNNI%?TaZ!5LX>Z4Zq5ldVb+D+2HxYa}O#Wou
z@dk=R8)iKr6`NC}log-%?p7Av82{Se<h!Db{7R=?>S7k6hK7nS9&M=`i4Zan5r8?E
zTg06RpE5R+lGt*FoVYJP)!FTbroq|wJ-`WfW!g!%-Cxk@z8`-(xAKv*0=Rri;L3Xn
zs-Su0SJAow(&0(m?N)q5-yggvV}`?f%&fjbdOMqj7a7S|-~;*A&_KbEFF%)pmELNy
z22d@X(}8n^j7QRN{Z+gEL+=kLqnPMCD{RpfLcJ3w;R5ry4&)X5Bv~W*sI%wCF(PKA
zvr5m(;t*^Gau1u-o6jh~1B`DO%4~ZCZ~RaucZ|R!&rl4AZ-cZyf}0L&aeRBzBrjZ+
zux`ZnPo|!$I;SE#vfWtyy<)|U38051(_7r6%VN-0PvO9tB8G7o$|FRUySp<Q@BrUK
z3t~ao&OJVCx#oyK9UoG2?pzLqUfK>DaCx`pAKcC4j<M|Z!b<|I4O8js4Mki?%t}OG
zy8-NYyt|3*KR8ED1Kd@2<m*1ir$Nz)0Y3m)!+?b4eJ~BgL05s-99_81x15;&PZjtZ
zDz%%8H2B|5g#Q|Vr~C%!uJ5W`Q_%xo4+O1#c4dHoFEc<rKdt7VC=0Cf-2u@g!&LcK
zFe=SLn(LwDY`a$D0!1LEIAS<Nelq&@h~GD8mU$fOuDUHgB(t|Q@=?HE3+KhnLV=_=
zm%Kw@T}0WUq>sLHx#0~F{{=S-zNN>TW*c~%6$oE6rXS6UA3Bn{ad}9^qYc`d_dsgG
z6`!Mys9F2c=lvGIANtyRU}fRbcEfWgG1iVrQeKw6cmxEBRLXB7<nz784g<@*V^e-g
z)|qK$wA*_OW1P}}z__<Bb2>T=r`z4Sd-Q6^h`S>~q+$_({<tmOz!p#1pbZ_F9zZE`
zp|))20xhD)hqcB~VNIug{P~ih*xvZ}qZLe4g&eC14Qfo1^)H;(SNPWL$lbKNP~=Lb
zB`AwuYsD;%Xj-;SWUd}wuP7P(_yX=`@VJOBgPMn#h1Fw1cK6j*aIK0pE(YDKn~?yW
zSiFf$!=oiFy*KBqB=6_C(hcu!QW~!JVaH;)I~U?P#RG9?_rJ8#gzSw)Cf=V_HCP^h
zLNL)cjrypdQgtJC4vyc57G|M8=F$e~aM{`0fPY;4Ma($`mrA)t>gSe%lvKw|#h%)h
zcRFvobdTewg1RhS8}!+kjrOB<*OX5P&UVX01s2An&Kk<BYiBt*wui5x_-zCqhr%xm
zMqNf|)H5wg4-2#wquQ}PvJ2Wr;`fyReJzO5ZW^7>AX@ocV_M#JnK`+oQ(h5In9AKv
z_P0T^syuy9<GH7|liC!R)%}5oT~;QcA%O6gXyr&hKQZ)w5zV-LH*!k#Rk8pf0??Ct
zk=ueq%q9&dp<3rcn+S7Fb6;v4J|(d(hgUZ@L`4t`<CKkYA?;((#UHDNfw1sPsz|n1
zC0HWX8C}(ZzFdu_19r;~rFJ9n1)uZzL(myV_L8^<AFMNcl$-aij&SRUB5up<qbL;w
z7I?T$*`d8Yn>5V&4r$tr0w`k_9aQWB(8@FB6B%1#k6S?+OaxA<yE=Wn7M}+&<1A_e
zj85vPzAOjvW}81gC&u%s6DnnZ(05GtgD9)cez_hcN)LxjEVU9h0TE2qCFSdupk;c7
zAimrZX(PPeS*0duxgTg<l_Q79b;2Eu=0O4Ep3-QxQ`Hd%`P0CdpV%`$39_wimzU#g
zJ|^yrQJ2*JZm!!U1rf%E*@<Ei|CvxqVg2sV3=CtGeYD2}X)}R3s!@$FeBt-L+}m*Z
zOw{i|imc8W0Doq)CDdO88YX4ZIhLS$lNby~Xnl6l`{E*N{kz%4-CDtYq~arUKH<Mf
zM?`J<X8fNLwOEe|mz1-!AJPup2N_*wh({FS+t13<Z@WV!InS?!iFE|psFyxt5rqv^
zJhXd}gDvrkxLC@DKEv>sr6|{1<8uCew?Y&k{kw)w0>ZI`e_#H?7XXoLPT-(GOE*~d
z&<e?5U(i#vgTQSw9nUq-e$-N)Z64;-dxYAT==*7x)hbu77&Ti6LP4yHw7ON<&Nyd8
zc=UvPn+M8PVaxc;Py~~kwnk%bCxi+M1oHRXxXK#2YyTl?{Q6Y4G^Na%S;tVn?k)ka
zH=moprq@W2le2I7gZasf7n*;0*W!K!QJZoz$|Q260=cDAp*ZUFr*W<S^Z%!Am-=0|
zOC_WK5i0mE&hA;_cipbsG#fzO*!9wjlT`-FB*w*v={Dhw&su4CJzXsgr<oa&b9`3*
z#R(YAjMXOiVs}EmMKWV5?JZmmIf8kC3^VSQc@q_P8VIfFv@Z8uk)^7nn6}7fH8tn%
z_1}2<u+Hfu4k{HbN=eqNPSC(BiW}C$g(%SbwT(iS;n=`4F*fgo#(osj=pVX#ARR&t
z@P8&f#3{cOfo2NzzLU_GU`YIWuF!IwchI@e=lQK=bo663P|Y%W^~vxgxVb?DbWUNm
zjJlg*>WyfMrNM@Z^TTW@ExZ#<f4(g+9~&aUjm_>t7JNhoZ4o8jX8drI(ZT#gmN#Y-
z#hxOnfDiS<6MUxUzo&INV0|w->*qcue4QL}p<4x#H^V7(MiT#+v2TG^K{NC*z6bm7
z8~**F6Tnb4cfG8HH+Eq_&G6>9nY6d8CJn~~9l}q(#8*Ah(`~XEpOUC@UAT_{3zbn=
z>?z*9PRL*uFrOkJKGoO6{dx)q&G~{NA76rIp-Usz+`aj@0T2pEL|apeU8{snlnTay
zksAlNKfYcGK?I6jVULz!78q)x1_uk3X(`@Qugu1JGkVJH`Pa*B<}_HFdcyGSzxF3S
z0X=sSTU-f%ceG=9@%vLVOxjOG?EhtY{&`33KS8K6FLLdWza;m3*2LB~fPwpK?Si+b
zEAZgpKh|O)9iMc6i(<@_>V40m#{kVghO`P8NkK0{=rhDDetYQ9XVr#e5QFP8Zyz|X
zwcd9d6qM8-Gf}Tp>iNHX<E0<;>#Al$M?o446apa{G8#lR0^J5tYp&i}*joW-y4m+S
zs(+w=*ZmVuZ+~it4IY!soEW^Tg!$ivNty8`WR`dlmzx})DgXEj$=9LRzNgNdaRs%e
zbB48YILz_h{7i}QNdsMh%>Qj($^SNv)GHD7zm%$oM<6~B2_R5XQi@4R0!K$bfJpH_
z>gt@5x0Hax3PKAQ*5-FSOf!r@T8(xj%*@eVsL&eF4n*IQJ&Ea9;NFhgNI?YlNUJcA
zL^A7}#1lNLN1Ye{DFWfJ`#38V>=Jj0U#6YDfm^9sL81#XlI@1E29mlye~W-K3KE~w
zWh+B_BX_dD-*Z<QyM-U&;q5z%`u5EM4%2Ti`KuVFcL}Dhd8|^-&MFPfJ9l^^s<&z7
z+Dk+mU5x;N1yMIV{c!iZlFEtC@&7O#?s(8q+8s|%b^NDX@Hv(sCLwvR{OeQaa-%&)
z0EhWB6`ghyTQg!pFB$3YBI45_oTLI-)!7sb#~WKaTf4d<d{wg6#+TLh(S`Z-E5}Zp
z0+Glvs<|d^jE7v(%Fpds{uzUt$U6oV4m!j_;&!Yi(aXhe!=lquCLR$*bXqXgwp(`B
zdk=rb)<EX1;;!XvMO50v!3kyV$Jn?WO8IW-D_;G@+h{)sz30St-92#sMz_>UMXUnN
zep@8Svi@ymNg1FAV|l|@3i#`?6pGL8NXpi9p|W}T&FyW#j~}8To8#i@p}&*Fno~zT
zm8z)ob1jXp(lj)gbL(G?f}l&0ngqRuUeE7Ou!f0Mx%@ekU<3uVtfvCr<Y^TM+0;0Q
z7Mpl}TI3?x=gIBYR{TIi@_APZ4LZ0zFBfl4?0T$n^jjVRf|`(^?@pb*>PMo^P$vHe
zgeb_9J=!=sU^jElaTZ)$Iar}rZfn+s4eN-P^G(Um6Ze8hs~K4*i0=C98~m;J681Y5
zg_+2p{O8O-f|r^w4ULV3pjIo|#MBfJgoswk67gUAiRbuu1QVpaVL5`x+>(dEEn|A^
zC%weE!TOkTgC4m{lF#E*!s7+)lV$S|u&*`+)2wCw)-=K#D%Q4GOWSpKvb7LzHGzQH
z>i#Wym|E_S?GC~&t5-rr&)Hl+{UVd>GGC~Q<sAK`TIbjZxyE|-u+QVjES%zGWPjKP
zj?0OYG?5{h^bAr08Tx_Za0k9tZmhrr+(w~Ta4`9$|BnazM)}+3<g29(|2tr{V|%Yc
zF$2P?zP`SibAKt*!kH`621o0ii01#}K(`d3f03)Zl5b?vB!3ZSg;xPB3rmqZTdN3S
zOgyv~=V<>^KW=3&7_!Q5QT>@>_CO4GDLLP=Z8It-i@pB{N30}7fdv`0t0PKah<Jth
z^2pAhkWR96l>B=PmuH0QJbFBFd=Sc&INMHzH+<a4L8T;e2$#&?E0seGB5AlIyWBB+
zP^Qq1L8SP<nB_jq5dMGTs8IJW*8hhE_?^5g{MJLNx_h47|Km%4f1N(~66-HdcLge?
zxH_$Fc~09P25(?wq{4@spPqrVPX5)boZUexvkT4@Y%HJ?sABVp>pLYSW=o9(smuBm
zBH*&a=?^7bz0H+>1^+7f!YDh&O#VkON>P&dTMCvAzF?0++~Fi7hx^^Gi=w@XKu_#L
zBsVnekpVV0k4Y%F{j)FeJ27H3G%mMo$lEi|K~~GD8m<VpFJs*cpUwZWAJFp($%ABs
zYH<H^@1RgN0;ARLMNn5)cYm=JxEM{Z-GmNOlN~h98Fd9J?ZK&bW-_5h2Wj<-gT&TK
zL87oMBLe69Cnk3O%5Eu-C>Z{RkB?IKx8WIuuX|MJ#pXST`JmlP#Nj%i@D`|tEqUZK
zE2`2&<kj7~@qAt>6%@ILruy%uuPf1VoP^~JA95$j#B)73l!{+zT3i2AQVGfcX&l|>
z0aEs>WALeboxWmKW{>tuE%RUjZAANLLiu-p-9PAsLj9&yhr4Fe{bf*5G+;vsjB`{j
zQy`t+M*BTI6O)9ZL<ig5KQ6F}y|K%v3@7_{LK=mtb}Cs07%bkn=<;v4+sk;K^Zz8L
z=)SvFSj;Qsh#Q>niBKvV7FF&W;e=mR(7}U}px-I7R$=J;eCEbXG571p#qy4V=`kS9
zetc#me24ZVidL0jdl2#?h3{V>gl(I3ccc|~SNxw2P;lM@x0fvCyMLTe(r<k#mWyr0
zzYk6d={+<w^zmw|rWFi0I5-3}bdJl#=D~sQ;w#1PY_HmF+CNEHxKl~w5nb*vDpdzw
z0VDW7zK@F(&f8Tt{v`h2#?)?hH_W$^URNRHU`nM>F!>G(o83Z8f-CocR8Dnw#$n<m
zsrQLhZ~0rdU;B=`_Kf3^9-pb=?n02aGR2?n#I0(}9fcgi#<77s4cF)QKjD5(XC~PZ
zoi9IPzrNbsd$H%%qs%s_KFaY@I{wAO{q5X8D}cVgLK6x4zfLfU9HzUwyQ8-k>h5e!
zHN^&`ZVNb^rZzP*GYY;?{Au?aCgrvpfIZWjT%+?duqO9~?1adSh&AEX{-LtD0UdRe
zld13Dd5Fzkd20(`WgytTU$X%j_q928a4UL>fO{o+O1o|edKF4XeEib;K__9X;~hQF
z|IRnn<KxGV*+DbhVxBLZ%%2>JLu4@Erw+%QiZg0Pa;-)maSt<KgYOMp-WeErU%nic
zDA2PNX?~=%*<f=5J<oq0!-wv-`l-5Q7nSy3mvaODX7?LTg49QyofYJynr=CN1NBot
zN{I;YMlar03OQRa{QSB3fw@@&Ex5^j8@fzMze6%R7)VO+^xEvzeX5K=sm&eB&0Jn#
z<iImk0Qy_A=X!7I!#(y#59+?nvv#G63~#Ti{Qb4Hm%7#s03pFJ8Z%8!pnA3W%WEzb
z`0u;=jtllafwaPUMbxWzl2p)Fei4=Oue<%7*ye#$vl(RNu*m;9IG<^>e`zht^>zM$
zp2!jnoQD2|+3x>ay|vAsM54a#R~1itR*C%PScVJMD_s33q2&1vJWR7Qe0c^8ITAcN
zHC&9$EK|7>^bGDiu*J=4dn8}<>Y8YDsy?xt%X=J8r7m3G)F7UE=)=Wj>p~0N{O2{^
z2B|Ut+!+QDI2a&HhkxM=w?qRhZL{b8NbDmimZ~!xbnEMxYlY_J&eI*2?GeN1N^|6r
zLI&A(ts-HGsn2tPxz-2A;SoLE8C(vGPat$L(a4;4gN|XFXW*RwJGQ!#u*(4Q1FwPZ
zr=I*MMXeMbStsurmkTf+qrvC+vC*W$sXw*aF26!y^AA{_8pa_(A_|JC7kRrN0YVtv
z$IPpPotUU|G$K-+&C@tT7r+yKy8mid3Kj{z3uyMLI9ft8Of%<GD+eE2{u$5p^6)a!
ztoyB9-P^WO7sbVnl4i~n?gxu;Rb(IN>PGp&t40kIw)3gS`rW`WtZ^b=*GE>X{8)HU
z$ob1ZNd3<Deu|4>{Wnt*BR)dIW;I5LXD~9F{}KrX2}evt1=HKxtL~k--~j>h4yO^0
zN}ZXC97K-0xH<+RV7}#A9X_?8wnxtc=;Zk}Up`oA2WI&z3vN%1WMp$VC@sQNJT&3~
zFnO`$`XcYou&XU;WkBF;O(yd=2C&FgFQ1e-BlWO93=G=|ho0dN;Wmv;oBn~&L^6fF
zh>Aw@#C8xv5D|K?Qr-UR&}t6nF!HNyDcN}{Wn!=h7E7ovCn?13@&d_yx?GW#UZbn6
zfo9Df=P!K<E)j%KSD*8$(qc56?08!1-S}5iTRYD3{!qDH?L$tBk3A-)uxy!1y_Gn+
z{X@+~+aaE+#C9EBMh-s=AKP#ISlYAwwg<M{>P9L)RO~0rU2}X3o$P@9-Q(eM>iN0r
zW?*-3RnGg~@*YoctKD#Z`h@E1sPC$R^1at1s{o;)dopmog$K(%zWfUwM#-UwYQ199
zoYi(SR~V7=SEn&Q?GmJzDjw>lz<tvHF&p*YW>Y;3u_OGKOVIt4Df|Pf!A0ryoy77A
zXt&Npp>n=%>U)0$#^0@zN~@y5)F0b##td8s;}!?3-@~?0UJl5d7?R$%TOCcY#bFf9
zik!i|bu88~J3k<&(;I1dp(8bT^S|@HF&*9O%dKMXWmtpDxx{o3>@PZabw09!ZR{#M
zM*CA9FH}{n&84$r_|tbA$Q_K9nuTUvb&aN1dI)&Egy877SDdO1LW~Z!&au!&Cbbew
z%vYh>?Mx1yKX1RcKYfPgL-2W7W95dk*s?pvX9dH|lcla4ZS;UZPBMA6Jri8-dXpH~
z&F{>Ur7|~3_%$ZKP&n~#POw0(`AmrZvy1G&`NvTknpwc)#FXo)R-BB0ZjON7*e@Ke
zMsK;~c%=s!4gfwMUYrHn#)<M$`}`pz{llK$Es1aflf9VoXhP#vcw!qt-k?e>7v<C=
z7K~}zTLxc4fSKq~amks?S3HA3y91NtniC>uq}K0|W6Ffhf4I0x{Bn0F?4*BN4KY}I
z49IT0C_N8T{;DzFNPIy1=({E-QP0-k^IYeZetp8h7x2Ia?SXtkV$CX>6!~2frlZ;W
z)6%)#Q-@Q%9165z!hFMz><4Yy_gYlb65PsX@{F^X69s<Y*E7}zCm%{APEcX!mhcpw
zf8lL#JVj@CIPIBJATqL2gPrl2!pstWI_=MkOWBPSY#D2~d0cmU_baFyuoWMgw!vb}
z|Lyv5(E1{gcDY*{)Iw{S6HMj7AYp|RR$xA&DY&RlDbOf;*3Wx9Y;@3}q3D+qU|ekl
z?jsVOH`~hUHyx^$!2~`!6&M+Z*gZb>m*)A8F*7$j9pf-&HF2)PB?ig6q0=kA3w`1R
zX890c<rb?5|9X%edxjK1@PAo*11y8~z&k;JNPzYfZ3i1EvKj%(hBv3v7oMA~(E81H
zR}VCQh^uC3$-j24?2N4q|CPA$^<;_CW_!K-G#4zTi4=aAf=^TZm>DtFTtBu^>G07S
z@}5RZtI`~Neu1~E#z1Z#4=!VAD?8`>dcx`y-?_&dsm%*zP^2nHD9d1<*ACV)<wl5s
z{>H#0nMgRK_&_Qo(v3OQUxaG2IO*(yzj(dTpakym(8-EvRD1r%)Zp0236D-&*akCS
zf=0H2DsZ@^^|SR~4in@(h3G-);Y8P8epKiK>H66jlDoS*DE4QnR~w(--b!3-_Shr)
zl<<B5^#j4-MZ>H#<BDirB){efw1;-IE5CNcp#>Vj7ksAYy?(Igp;?;#5g#AlOY@nC
zw9EF}5Y4??nWUuyx=ndXJ8{6MPDB6!zSsxbgEC{$qxZh7%io|IS(h6y(y5eJ`HEE^
z2JIdgdEB6}M7kc_oJ-97v>9rvPj^#%l888&I_iU92q)Z*iI@sZQhX+;znH^{dAx?9
zSa58f&m2|#a*Cs4%1IzQ{8|P>--s9`vwt3$QE?bVG)utiQ!vlv)s|WIAs&hFJ0WQC
zgxJ#buJj!Edn6IPw>gB3Rc3vIhG_HDji4{aBmnCUCXcu0q<rOncVXxXU+Vb&MZfIQ
z$th03JDOd*7>m3VU08_Jq2=-UmD;x#X;W#t!KBp=kP4XVY9hpk9uX5idPG%9d34EG
z4FuZU!KKc(L)kKS?Ku&&`eHL#lOG?aiMhKEUZMQVJenCX>0gmeWEKIe(yC_BNXRdd
z#7dLvpyH%Bz!vKY7Qfo|<|*sYJS79aAM7tX+a=#6KqCq0@0Rd(nWTU%-lI>$XlO`t
zU%l3q#_{}5S4$c+IK`(&{>v|KGl5<0jr9d05ivQRF5`1tmhftU-0lPaQP*;2)PD2o
z<eA5m!UGV^HEaI5e|WRKPyeaQ8=wbdSpYNNyA-@86buht{5+{^AVLcEd}|o`gw7Pm
z?Nou9z7_A9`3ds2BXF0V9ACLBm>bv#+Pt!*3KGR)r4i_ner<#Ee1b%C0?CtJ@{U&D
z9NC(j>%CgEehw(8w?k<FCOJ6D5G3u_p2Iw2UERNu5l%U~7%p_?R24qzE~zB0Rzt(>
zALbUkfih1nWEaSo`qPf9erLW9Xy&eAgg!Dx;ig`!08i*DmO^DQg9ckN5GJ-s_ci8;
zoc(TVC<gSuMxeX|w>dc_tD~Yyp5F6;vl9Kg--L3K5-0NZiowFf<Np&iDDn;J>_*vk
z+M*Carq~)dVYrUKU@kilof!3<cJEfd?@(IrgZ;6dNGXakzy3EZYI2hOWOP*IZop5L
zLM?=PZPEZ){qOi*3yJA*$QL}H{C$@Ubk%0e<BfSzIQY8?r)*rFNLs@ydede-ll4D~
z{&hIWK5H_dgt;9<(D5InCLxHQe|FBbX@4CFk8S5F^B_-CWuuS^7)uL%1rTk1w0E?a
zGGWqMHWrp-Y=7h)??{Ma<n<nhev`+c4xKR@r@F=wM0CB}jp@sNke+#_Qs{cI0PQ9d
zRCVhrXyLwyL%m0uof842m*-Jh8?ke{+<15%Z43=yNzK;*w)wXMHm?pnjAVdadF0ex
zdWkQVbZ3TuBh18;NxX-v;M*<q>(N`MGx!81yAcWX5~Z|7eK$tQZ4kf-5@r8qEHaGC
zIr42x{*`R0T-XM;^+&oenfzR8?JFX*hRZ#457%eEI}M`^k;HJkJ0g>)hbCDmDKV7d
z>Y&HR7)s8YGxGwi!YHvD+822U<wzC`rdG@^WoBC_4kv7w3bsNe-7}@)30eF6!vqhP
z2YDfkXx-X8JK^E*fkOth_PF4E!f!AcOH+wNBjuZqUM_p6zmTv}H^M*&#>Wq_9CgHP
z)M~m)x~ql&_nWgHDIR25nwfy)+ZYf93<2K(wzBg7hCuW`A2?J7gZ>JYAdi~+9TY-8
zzYAvgZ+GWU@zegN`zEprXQFoRoEMYx7#<qj+hk^FYktQeU_CWBtB9hS=i7hRpk|rM
zjpD`2aj7CMi)%2*!<eXAtrT>5C{2YK29TJ4X)c&3;O?bg3G)>XCAqwTPhDsRfA0l-
z{_&o}#_(%Bi8;HKe*i&~XuHG0mv7(|7m$$YedKGNE`TjK9e$K$O`6#PpuPEc87-ap
zw8ZK1qIB<#1~V2f=GgXJn06IxIM0pHc^o`fjTwfu?scFOq1BhtIrQ5PqZM^lK`lH@
za2NC7k{r9V`4Oc1m#!@^jgc_HKXjyX?5Hf1k|JYva%{fH=RIe#sLBmgE#o>neQ|%>
z8zX{gbj84mIPPgJ#xn9qD%Q$Nc2q2v!%|z|n5OfM6$3FB;Y0hb_sZSEU+^NlKY^x&
zom>^Rkc`z&WUtUPOY_-g2|5*(HsGyb)b`ZUGu^Mfp`t)RLq1R*H!S3rLBvpR2J|kE
z<3z%>(t)?)=br%T9Ef^9QrCZr(Eu7fkL@>bg=x3rr00)aa&K%FUF7&xrLe%4#NOdO
zo#0Ve-gfsIQo8a1v&Gj>I0!i~hS4|uTRb55#D{F}4p?gnln_%SBZVqOATLN+P?uiD
zwuCw`mOr=;8nV@t_^Q+v0keepW9gOnzPeUkq#?CMdM1F`KIl`4=(O%KAE_k*d$Yls
zW{N`>F?Yv}$#G+_%bV{AhnEk4OV+roU=fIi8(lAfYOeb*&u0*#yw28arRM%^Be*$P
zwd4*-_X_BYphh)G`P%ya%n@J4)0l2Q{`KHNg3ykg(ZqNaOVdoXc;nAUtv6EZ^5L&j
zvvp?I04{dxf?Kc0=&0|-Bw6$J2P_OwYY7)dB27&gBNKV)^kXFL5%NFpw#u%sCcFn^
z(!1993D#$G@O(1_`a<5SoiX*klG|W=(~F@uW5A0cAB$=8&T45YDJoV{M}IIwR)lSS
z2iw`<cw>WAXW>f<M=>Ab4@Wcjmc)IRzVyAZd-&=&2e3Qrah1|w!hh!3zT#eSv<$iJ
z@c{PIK(*aj0pWD~LmX}V@JpI4IZDEcXG68TI?j)Deo^(A%^U7Jexk}~TLnGSXfd0w
z%E~@(n;*m(UR+<>Dj+Y;UTR;PGOE@Lw^-o2&lbegm2KlR(Q(-(pBYv&1da8TgV;Zu
zXh|*=Yvh2oNLu~~aH0&pCCtwwHl6sk+#VOikp|r{O^%OIJT25a<$%A273>}P7Vb{1
zYKpgL<({3knlXOFNmzSW;Hj+j%|kjNa%;JgD_4g%RMKqO-rYWpL2>RBrnSv5Yqg4X
zFrGTaq?xy?rP2s|uZDY!8-X9{lbBtX^zjO6m<3j3_NN8A@sn}r*1%8xPR;35x^2E<
z?!;p!2h<G?hu%_ccK8R*<y$>yqr&CJD-wzpi$5b`x|MU}$&&4~JYJ;izt-gM@&;x-
z33)cYE^zDG0iGV)x$M8;S$|C5e3-S#4c=|ip2)j1-LE-<CEp2geYDBos=%S;3p0h;
z%4h7NB7@FdYYgD8>}&woVYO|0NQA1m?K7LQ@HpUs?S#a^hi7nZoY`<{K<NT07SAIa
zE>~>R`5q8xRaSFP4oy|MYelu&_vQ`nn83o`k#)fQ=ZC-VJ2W6hh?$=7e=hy|4=!PY
zK;yoNY=)A+n%7?mo1}YS6jk>s(ZJRe>GL#x*vQkB>=@RO0W5F1l*rFwLLJ_QZ%Ylo
zw78MVlhDo&lRCJErawGlMo`gH$hf$0Vt8F~F>5`gzf=-4p-b<{>;1?lU__14q5o)f
zw7(FK%??D6!MOB-RY(n!KYPN9et+8XPJZI`XxfG;jG$#ADi`-FaWb(%1C<U!z0alk
z0kK^NT3zXi5|YXgM`oP{S~?NRjX1qy-viM7iXW^WI}kfr@7j{KK&5c7@9MyiW43we
zoK#dCQ617+th}8<D(nZ$g|*Hr*3oiFbaH>A6E%$(cD~K&_ft{HoKHzCcCOuLL7k{R
z*w&X|Ci@Pj5kkdDELRd@UQWf?SrbLtm5XDqZ?YY*yKt1%exs3u;gq~i<7=F_9N5XM
zVQo+3<g!K2TLvM;cauy5Hq_Sqo1ecw;No>zb}KevwI`hQOlxBA5qzgC*O}l5RvXmc
zRHNZS3wdWP;NeI8>LZ07w85c~{Mjddq*qSL{Z;F9B1a<&j7~EAdr!Z<qRQt%);7k*
z#{0ct?gojTOE50+y@e?GlMVJ6WDUe4T68%VG-j=9%tZ(kU>wa?IOF2ks}C++bvHGd
z#M6_pST)fX6sxV!_4{ha{NB`S)oEe2yX!t`Ol}ZGszu#>R~u85^VTqZZa)2rN}_%_
zUgd;84Z@eTQd^VYju7=565V*mu!ERJZ9z93MyvWOOpCDcf>OVQ&JTn)GULM-2X)t}
zS#842XGCgFz#PU+Z{<i6&a|EVPuj%R$X%Rnbq5OzlW-h<h*9reX~Xn1bVU`aR@|9F
z5o~nB$}yRL%Q4_ve6uz+t1mYwoPyw!-%DRY+<VgBhp<m#RPvDZt1rut{QjcWGPo|l
z*o)0=aLh&-%b*=#6A&~BrcGgM(7*t$e~22eO-orx{f#6YUd&+9Q$<m!Cul$gn_~i1
zW{14}BdK-*UvYZq8(!s+mdpFpZ04FX4)lO452bcncz@usd^g~TJ712Ji)`%V=tUg#
z_V6@o4ki3OYY*9Xmzolizt20CH8?zaKNx&8#OE^SATY@roOyW65HKd;b0J}5`7(`s
z(A+2F89cQ8rXPgNL^lstAH%qf1CG)gp0*oS=Gja2Y%n$eo`SL|5Ld_Q6+m0FblV{_
zX2x=*Qfy5|cOE9rTKeH3;>r1Ikou7?L5hb_(-G@5O!di}mos)-X!*uvT*9>%Bc3DI
zu+8O)RV5zIdx`HTdJx$E{^9eM46^K(Io6IxlmA@&*YD7gqn!LTNdf*jr_^wRHvPW+
zjrFnLUn9H44U^(v3CtXsvFeF%-U9EE)J`meb9gzXoOrKsze9oy&8`;r+JwbB@-)|=
zvID2gDpFv7-W@qR-u2?(YcN%@6F7T|>w~W|6PkP<GBAm$PC2zVj*I%aTA$m%*Yn}F
zzqO0Q-p&rd$c>Uz&Mh2xzUYyGyzhh=jM$woC;@_#8t!8wcM{n;k@=QC%FRUIyz*nD
z$~j+m1|9sGCh<kz1grlfL-k9e7B|_vULvW==z^LT&$*0Jrd#?H&&c3|XytX47zmu$
zNc0(%VIA1~X%UNBjBl>Q$bt%`kLaB!qDl7Ecm3x_az?Dkmukr3RJxCk$dMZhKnI-8
zb#x2PuvC4Tq*ax7+w*IcE!YvmC(WGYwpW9$FiX!rV?_I=y;DiTPWX)$sRo}ocLwm(
zm{qxwVwhQ}Vf!2$AJ&c_8HIZ!M6e_ea7+Ycccow1`LZK=fgepF(_)%tiu6N&u=8-t
zA*Xh1QhIbgQ68w!&shpxS8mm;_?#>m{OIwRF3?KKc!~`~(%&bNO51p>GxHZFSnRaF
zQqj?p5K|ZXbn|?o`Ej|Cwld!M17UZbEk>%zvbd5*cM{!p#>)8$DkX^uTHV8g><84d
z)xKiqvwNq3+8uk7tiXu6b?2X3Hw#taSitauvcN0q`hxq=?gFKDhE*ezt+kZGbM|zw
zDHPE{o#f%Q;M^l7SS=C~^kS7lR1Hq6Fgv3`(xWL#ns~Hd1sG7@Ugww^aQCkcjH(4a
zeb?#;8lvprZK?&v+oVV~8)xNGGJR2{MK$4A0};9_M$=pfKlW%IU&K_X(Niy@Mf!w^
z?k7?nuZ<`Y8IuJB=hb6e1O*HFn1M0lCoQR-tt$6%>!)z*woNkQ+#Vs87ee`Rz_oeh
z?p0I0J`2zjm<)?iE4k}7LwzM6NMfzbec9A;<4ATA>_KvO1yD|x<<Zr-H+sf|Y6b6E
z49oz_u>6p0^))xR)2hl;au`a2Lp5dXI6MRee_A^NQSI^9h{TtyWP7Z$3cln~2<k>o
zey;pP9AGh1Nwj$M5=@V`{BT_)7AuZF#iHBZq3t7@3J^xX<;R1v_sq;KA$5KYw!e5Y
z_#q5o*)?6htusl1Vj8{YMB3$k7mvmWTRu{w?jLXE4&z0}?KP=#>^6!M(NMqA^va0g
z9esPqVHt5;;w&Y{6G;8#lWD-`yr2SJtN~SdJ^xt6u%P82a3WXhw`XQNF*-^@$^%0#
zX-3{5M46oiw(Sf-BSZHFMr+Cp=!ZAb$v+=?t}7l@+}>?prc(99LFC5sGm2zW476c|
zS8$RTQ5W`LJiTlmiNi43<#@)9PZrt;NV(<A|0$FU=-*&J3ixKdX3{2cc;j8@5*c=M
zlz`9kYgIIV+?9qBTU38OgK*dTGo|1G%aAeoV?-iDT{>LbYX<*VuSfNjM~t(Aqa+&p
z>06GP{{MKI5K@pBCv8Y|uYvJDenW)<(XGAVR#yO@l4a$!Z^RE>`mHtfW%Xj$<JFHz
z*W|g?$KKa$rk1kghlo<Ey!PW~i7#zwduw>{hOL40R@q+4h+HKh5@_^UY7|&VLD=%^
z2$B%mJ8^LrJX_pyNjydyq^A3IISe1V2{z0gwZ$MB$X<y?TLWW3O25p@rez(lqUc10
z3Wuf0_Etx{#;6oYyx9>NXl<_vsf-O?8*HRt%Yc_IoJHf%x->nM3Mg~jZZi?PXE8M_
zYCr&)qx7($vAW(i0xMO_mnbU#Fv`e+^I|*4O@z*umTo&9)G>e=oTk$>kWG?^&1NRz
zmkDZo>KKvj%`vPc96;0@Y;Re@FIfJlPAKMALfp2(W*yE_#w?kcm6fU(pv=hwWk%2)
zz!|UhY0n=pvgFVa>tAMCIv3hDozuK?gtgoh|9wPQ=U#QVb##6;wCOyKL0QAMud>|`
zS+6rs8UvHGE9KjPAI*%#qzkKVYggYf-E_%ZrrL>Q`y2T>m%V>XqWdL~-QIo~@x!n7
z^3pjRtK-t05{XmQHLv97eX}o@r*Ek-;Q8lNe#gsly-=FV|Az%gy0gst!9;66U958w
z_yB#jESEMcHBq38G~uMe23cI8Xu{QQ^o30vLc9Cm<ROJ~f2|BL2)Q`ijWw3!VB5~#
z_UwSQL84c0=pyF%<&~m0k<*gt%&t2>ZD^z?ek)(E$&Q|pDP_4H3!frFbc&}I6i(Yc
zI+Kn#O`A=A+Kf9%H%-^>US;OZ3K>7IRxYihU<}<UXdg;MeqYc9Dk(McmA-Lpkz;tx
z6qF3U@mZO{Y?kZf+#Q%%bUG)KELs#OOUiBJ_vNSJck9aeYJNiqDkh{BOKg~#2>4|e
zw~F*}wXJUS`c4c6OW!bf7x8fDg>BXIlurslq*6O@7dMtM$zt<a0uE(ZyUi1NHKl%%
zt*mR)>KjaTG*D4ucp_4Z%mDQ7mjy86lSH99)xG)5RW=|ZagvhyU+-ZNr8JgM8!~K4
z;Kd)y_#KXKB*I1d8n|F}$Zw#}cJ9{q`x?>eEK~_T)HS_%3QTkvQ{SU)4m@1LDisq{
z@&?$GO`EO3@0O(~mO@E;)7O!-ldT<{&J+7V+h?rYOtb0jV1)J_mfkxN<>l)oO?_N$
zj+27?eGDVTCTvDUuXN6A784OzvHb21NZk`h+rkqI=9ZNZKl)Jj%zmb8o~tFvz9xq%
zLmO184A+*<>4Ar(oQ5qrrkvljdyI63`zZ&hH^XqEDok;u969vz_1&0`tH~NaF9l&+
z6YL>K!Z{YEeDmLU+@&hlik|g;2;-}#6|R+$d)-4uBGYPi30uvXX~IL;V+?3gfj&)A
zN9Ifl%%!~uQm=-ohZwc`_QBGlURX3gCam@-+;0kyK9y3pkn3up)+CwZWhmE(S+S>Z
zDiV}}!ss_Oek%2F=&&Eok~cKTW#8`6lQumx-_(hQr4YP)gs}IZm)Hg=lR!E;ntcN>
zcr%7*wjBj@i9m#G<@4s))oQRs4A<ja!}G8ovfbNSES{-(U(Pb(<O{!<6EnC^8cBDP
zvzgy|?tNO0V6%+an5;!QRR(wOT8f{}NeNH+E4_Lzk(r3BXV7545^ah7&yuPq^m`5M
za+S1prh0wRy6&03A;znPz*7?0`ISq{m`AAIQX^MmpsW>_X*P;6J)F=3+-6}|TY38s
zWkkn76c+)Yht&XU@ns?*G9zfi1JZy?KdlTrpmW-UyJ)tAyl%YlwOX*HPlAZ@nbB7^
zSQNtImNsWxtvavUE!x_-C>&2hnk!X0V(%t&52i8$i4>R2=zRMi@t_4;W>5r5>t*D#
z0hba#sUMk*=eM~%pLqbp!e|~=fhqk*d~;k4xTCdJTg1>k1@T9E`40Wf<Iw1-k#C1+
z%&D~?ntnvcMlw7Ya(HCW1sI*lxdC67bD1Qu`E)5I=W5Aa63mTl?;Gjd#8LN6L=bjL
zy)~uMTR|*#+0gearYmiqBKH0k&o#H`zB7~ZJ0h#dhAwpp`%%K&RLybtG(1%_LuHn?
z<d~%3ms<0UqrOOVAzoF=X;Vg*N~#+>okV{%<7SvcP`d84L$q^s3xpnqRre>z=gKO-
z=y)Y2Wu-ZN?7T(NmhS*yAqjMApN{U&l_By&4Lq-SoWlfo$=cp9++`=l>TC1pKeQ|q
zD3^wOdONEY=whY=v$#9Xajc!jB^{iFTkKX>w0z0sBPMOm`|?dPWGPmf36C3pC%8gw
z^wIV95plO%4e|F&earyS5)B}B>|{_qJPm@d83Nu23imjHFc`71-?}lnZg8B=nRWw8
zvhQll7a=xgF~`m4qsFS}`Mu{(*Yd$S@eYvo^LnPj^&aP#l5WRV8+_V0<UN{E>fkfg
z+uMfk35{C1uk$)vup`eAQ>}cV*o%kK)`pMixcnsyR#snXP9$nANEWS07CBWDkomhm
zGpADR?2R*|RoJ~+P8VtTXRHxv3yMFyGODWwai&I;X9y@oXy2ZSp9sF8FIAA=l{dYN
zky7vQL+Wwp!t=is(0}jT9uriKc-dI7D7j1uXO-keF@8gt(zC2Y--#QjrlMXO<K0*j
zJ+Hv-Vl>$6;5VIWBziM~*%Ygrc11nSqquYo%2}{niYxfs8u96ksCt#mz@PuNr}Wnx
zF2{s#+?;U%JBIV;VMyClG#lqhOC*H3dgO+ffr!14ArsJ=aKh1y`@lM0KrU?9FgS#n
zn)pq<zZ)5-DCnEH3jFL!LaVM5=xmkiAx)Ca4qp$$W=ZOR5m&0Pd>G$|z23=x+PKm`
zOa!Dz4kdul;uCAxW7GT?pt+7Br+t>YceHaA2;;=L+vDO_KG22yrTnYz>9et?OK5*w
z#5_@bfDUDuGkmLIp^JhCKQgmcR3di7XvzxB7ZQ;MqUX!0@)p69QFb~eQ%3s=R1)Lq
z4b@b|OWsZ0igVJ*$;K-hQblN}oPmUn|2O|(_?`bW+4qtDvj(H;g~-+3I1WHj<x=s&
z@~JP^=~`RH#@a%fWC%7Lo*xA_>gfSn!$Gjs)Z_|i5!(uJ9GAGp;f{WL=uAVW323%|
zp?T4+j^8w3en>;LO=$P;Hyi05NceS0M=mA(cr%102I*#KUx8$`9+{O>*iiNvgS6s{
z)@L-vPK(=*$+xjwRovxEzL-|QWGMwv9Axo95X3ssV#}1c84gEcYIR)$s2>sOsV6H3
zLwA*(FlU>V^c1F{+!n?MaUP9tS-$oxpjxM5%P-?ZG#hzaV*6ThF%zQHahFkaaV#Xr
zY!XPZ67ssg|3&8zt%fFsuxG<y7m#vB;7bzI7^~gWcz<eK2Dse4a}s)2rfF}@5|>h(
zGsGguth9}!G4ot7BM001X0*$T{lEz9lvUktNsPe%pg;yrr6f5>RBg#^&yVA2h5wL;
z+%%4l=6<)iB<FgL91xmtp3S}qz_hOZ-c&_h#o!iu0XG#7K`Ib!uc68J11&VV#_{7S
zqptl11_69VV(IncKH*A*!7JUD(%j@d7lyo5pSDP%$#j0nmBX2enF0CrehQ5Q0cOp0
zLTXgt$t^&*-AQbiIz!ob7Fw-DBXFNAb&C?~2q55D)bdm%;2AK|6t32O_pY_CM~LVa
zt#9z0%l?Jpagl4BXBua87A_ev7)Hd7$e(~KjJuQ8jGxqU$<pS>?MTW_;~AmRH@(H5
z#OW8l%w(1Q0Myt)R4neYmqbHKJ$&T;F{}QTg+ykTcS~bn0(V=l!h;f{7?tLI5%*gy
zR3kdJug60+mVINb66Ln3SD7;YEc57#;UQn<58iU}OOIb{*1m!IrC+oJ`qIdOl*vFq
z8P;H|@}vGLCz=9KACq#Pw5N+5-y30aZwaJQNZySjPaLp?KT4W4D<-><FVlM^31f%G
zU5b<RFp)UmBR$)fP_A2H^n<d_`a0Xw>4+Dq`a~r=+q{;q&z|;y^qGKS+#zemeD=r5
zrq#IRf;4G_6<%wl`X_{XDpMp4=hO0vm1CqNH9Wn5`&O+YHB?%Zx;P5?UvToBkwn$V
zxDCkAue)txEEC-rTlw7m(+cu@h?4I$yG90CaS?IN2B&#zXwhj~O(yDsDn=H0F>x0O
zE}y5pZk|!Vdaxr|)PjOpwUx?JQN4BfP=JjgsT;-%WgVH(6oU6`=T~#}Y<oZ6g1Un@
z_oA!=R{^?S>&=_(UN4BL+GwWsMR9}cH?=xca^j99|Jw@{Px##>U1>CoX8lM16kQo?
zvJ4;`)12+PGDuu5DK&_HyH~z<e!fb7C>HN_!=h~@=D|KN{F;C6?LvtmU|_(Rk{Q0w
zZwDq=J~=1d9z%BwLSU%-zj%+lgtX5w_VFn2;@U)#<mheG!_gEK)vSb>L{@fBLL?{<
zX3Yh^D?jTGeu&ruKGmxo)VvFX>|!Uw<NTCpLT^iC8-NzFQN!!tuPudb&aSB3Aiwsl
zr6cP^i4F4TN)m~wkNw>uLGx~NS<e#*NZygbiz>f_&Q^EQe@cxjKsJU39azXt#yBJP
zMDg^(OXpnG5#2CSv#Q%`&pgtJCL=TGdP!^(AH=#0M+xtyJM=+Y9}P8x2N4&vKNbZI
zh;os6)N5tBTIAW58hY6#?|F&2t7P&pl?d7cBaj$jYQE`dPE)V*X)hBLsna(?NSPs;
zUo9S%y`Hg^F&TSyG_@8*NZYDXxa6~Ja<2Jy*abNu4LGx|mU{g$$In(`ys7CuSG+d4
zxa5<+Pv*(e-i8m4h@X#YS7;JRG!xEfBNY$tX32<SES^p!lrwpnAxCI!;0H@(fc(R-
zvuEqjpl?r6Z3|$;^Y9xTKBt6%r(2(;zvM?}d8e6Kk;GbZj6`aU_HJ8-!+z#u-<!@W
zESN-Yf)x#!5aI0$v@?^(b)R$ABq$8>Xf>a(t7A~1*B87Iw563>ijmA|Nz6Q2udZH!
zjyJW8BD!J42Pv~NxWnE4(dh88^+~_&r@bn<A!2=<pT>#Cvu9;r#FYUTDPfqst&k?7
zBhDzf<6C+<Hxdyc|1Y3Ts}ntLN@CE5F4X7xg<H-FibE1|CQ~8)VHOhkT0x2?qr0KK
zyZcUhI!BI<6TAc}8_yVVY|ikY81K|OV9giUN<0+c4Zc`gcbBJ#lrH6qeBx=)<VVMR
z-&cG4b_3N<OkKWKrY&xsl7{IoxCLKG%@SlcW<vX5>@otnuL2Jo;pdpH@!6=?E|4W_
zYTxg+0IV+Yk;0M;Xl>VK!v$Kf<%{D8=~l=Ti_4O`C}TLw)3@f+r4Bc5Hy6M^)oa0(
z&bmE>_<%mJQ~N2pAlbC4Zck9n(5bIN#3Lb(faMPrt|8T(a<h@<hi$We|B%SChPN{q
zU57r#XuGuBZb(1-m@;|WSZOU%+Vs(r+j;HcWdg>5J$o+LWxP{Rex2}c>u5Ld>S~w7
z{bQF9DRL}>;jekUOZ+j9xnMqPi`kW~*B{H=0v2}QCATCIEB~(*r~J0!lLy0z|FL3G
zu+1_+Yt3O-I*2}YtL_ET$z3(@t)&)2Fkv_nyXE(hu9xcucuImm57IX$1hw!dR{#yP
z-YAC+gw$VQVzbSk*k&eufJIUTg)vw%nb~dPy@uzP3utVnMep-xCQOS;H8al*Rqx3h
zE=V=*6g&-wZEqwW;IhBSOb83ie*^?P{RldLR?W($aZS3j4|Nd~L*R6e&QKdvDP!#F
zFlXD7k?aP10J&!ENl*i6`C*i^>n1ey`;b=;#Wk*cYS1O)Ccg9uxou$9RCEjS_Fyyt
z^Enz95T!D8<&x=;Pc=1k>=dGV%|#@TE6WqS`_svn)IiPkO8$`Wo}S?4HkBd`jMnVn
z#6m#37Cr1mbZmA&+pAOyFc004ITgbF#rC0Xnz3K@Bo-P;Zy&Ftp1|BHIM}m0ZR2D4
z*{$8kl$jhCQMc_Wqlq&}L0Y(8+E8p!Gsj;;v?-lot?9=^S-Q*RB?jB!oM=rBy#_lu
z<Kxv``sG>`B7=T(?2n7g_b`e41R!~$j#A}}kxD@1dN&})%>hVbPuf*{KQ&kBBFGwm
z;JY?#R*?BcCp&``)ZmSJLa<&bfh9dIF-qM41#RGTkE|_6lI#$czKMS&{~MKJ%O?JT
zI^jBaYrGf?F(81~4Zrtw$h7$jor#4%C~?LNvbSARRfx<G%eHxEq}(P?fZ);PNf>*5
zd1Jv)bl6LFwSpDWHC2k}psBLqq)N?j!rCsoFeW&>)!B*;5!6_HFy_z$k%hK^bV)vH
z^BO_A$0^GRT9xe(fYyGBTLo>!6U<#oy`K>@v`imBb|gxHTYS{D=+27EW-UC|SGp~G
zuE7-&vUyl+*0D&5DaWpXVfxZ-r4besH|ns0cyEDgOs6KJoC!C#G^=PMR&JX&*L*_|
z{t#iIm}L-`(MDKreJydVQI;W`^NDGN^xe*&YLwbrH_oWnFDki1kug&?%WUdOLLELi
zAgYPzvefou3gdN{$lWU`io4F!!7L(*q+R*gqT#spR5)nR2&C8S<_4nptbyIZOqb2=
z2hMGkudW%#ERvi!{TybN?e<7>|M0YX-|jKspEY6T(Z4xbd|yg1UM;(Nx<URDX0r7^
z=`AiqB6gYSFk0x(f0oRC$WgZbtov}KO}+^xkZyUJQL5`bFEhuHk!Oy*@b}5M7xLy%
z@excFfb}V?m!MpKu55Z_n7aO|I&!nml5*R+BK$3BV2nP<kuXj8s$!WwO~!4LP!~^9
z&9Yk`a*QPT-PS6#m)2Qr;8w&jXbMGIjVxkS%M>%UZz#67t;uc8?0K$n-!-o*7N``D
z_T@<B#<!!Bow<@0D(MP%hu~q2Z9G^0xWDVy`PSwozJ-hh48F6Mi@=tb(`H$jEG<sh
zAVr2kz2mq8O@d8l7<AimGq4nhfMdhhuqF{xTtQKlYgC&ReJ9O81A~T$R;UgNF7&TO
zUm5FA3-A<HJNSNrnD9$dCXYA3P_k4sL)MoIMb3{A3?<KV5IQZcp@m+ufNd7%U)=KP
zQY#C7)Op??1(GSg63>&KfB=(P#cXm#LP3<M4$<g^u2`UlmXj~Yle1ZLdj5#99-R#u
z5NZ^2ibIm;JtybR!G_Y-pN+R8?T>%ZpM_U!Wsxo)#K;%{rL&!R)w^QmQ}D}9I`6-E
zK8g{XjIJ~UgJUu1pFP04#62*Kd37()7)UO^`gI8%uI?WCIpH2c8eD*STVc-$n+?PS
zrPs?n_~Cm!fJ-%uf*1phgX~s#9SXpxU-Iav>BgzfC&@=EU`Uudm?*q_$z*|&I0WO#
zxfNo?B8YH#QoN_O$pQ)cW+!isX2NCIg^-_NSEt+=A50}j0}5V`${CSo8=lc>5e+Ho
z?&3MZh69WS{di&bp4Y?L&IpU8%@akGNY>)tr@lSjfl?&Xq6H|K9r)Yj*V`NW<;H<L
zFq=6?V$6m>(OMZq0Y7@NTnc2MQgkx8ttGA6lB?Wcex$lf+yy!)cJ<`3KG|+e>$JE(
z{anp1AhN9snMbJ?RG^i|mRep)jX(b0r1At-8kxXIXjnLeuGQS1cNU-CeD6uhoXV#H
zdAQA~N}5V}T^6uhu50ABTJwfNqe?DvVnf>(At<WHMEP-Uq9n*7u*pjkyoBTLnLDUm
zq$mXoT5Lm-hK-5&_iy~;IzUW=@pi$4fT*|(1)*h(vU(MME(D7abt#w~w(H_uk<j;*
zf(wrF5!W&IE$%OHq4dADmN!}*eg@(qQ?hu+QkHQG_P<RIkX59xh3WEH<JF7UbSMhE
zLOXxn;GTQuT&64jMChsC&WHciJSsuWHS$bO0?K1o0n#i3y)O=BpWPB6{H{M@<eWX_
z76=~?(;|gCJ@~cw0zcc=6NRv&^_dW~rZLDql`Wrd)<k~PWLW{UrrPnKUKzkk{T}IT
zxoaxW3~np#Kyut`@?olCJ8eS)Xx2l)HjOM>ADa@I3vZ&&-Na~>5`>thB9WcNBds0(
z_UQV=0^Vla+kosHwQ3`&^&%02m7R(ok*jlq!39)O>C7t7N*hiSVz`T$%^&7@i(I^0
zb%g9oq)i>>@pRlpkaybBF^##D+?M!KpdK^&vjRrJ@s?F`-TA`h<&@yp|3}tWhef$=
zeOpL(cZ1T+&`38*Djm|@4Fb|F0#YK~DJjj+-HpS5fOHH)$M>+$*?W7w_aCkco@;pK
zzE`jHTWeji_WTTBANRLFyTtYXZgFHK5b$0UL;pvRF$UWWnS+~KS5-yk<?#eh@lSub
z78gRlr8?=xOnhst;!lSchAm#f#cfg@eG#NId*z%JbCj+kJ!BDK)sRn=qDzRbmT?@q
z%m!$(n_wY9f`jxNIuLS*J3?w@oPdB_Tb7LaWI=&gURSd&CS@R~@g-B_=DMvK*VWqu
zjR0W4-b)&CTgx<!<4eKJ&(UorZZu@J-(D6a8n#saG&k8<qX&5=C{`1hn}#6!MU+x%
zo5Xd4<vLQUXn>6c-5-N5>kRcO+f~{EVaa-FvTD!s3MHGUomk(S^-nX;`|tz@ab!&6
zt@p<Bd|m@!AahLfZm}H=j5q-)mpF?0(E#UDEOuq(DszCHBQ$hI57&J*mVx3v#-_#C
z%MUJ)wW++Hsv%jyq6Y;GId)tARVU%%N!BMBwDi{jupI<9;bt1FBS1o-agQ_9t!~;Z
zS)8C`L;iWShk{sI6zSh|@&zGEd(R!DkOAam21+mE7R(}Vr8RbmpKTp_o37z`{z3$Q
zjDp$Ue(MT5@<0lfp5cg6|0LUZYO~yUUHRk3PgZJ)xT77jABEzY1iN3U9xbySyWuWe
za6_<q)t$EY9(zv=W+D+wZ=3fBxRixE#t-X5SF(_|;<~(kQl*-`X*Mt|gQQ;RU-ymd
zGO-z~eT*{T+j0B?`6Zi7-kkD3cIgB53!VpyHLiq!`tgq^fnCb0Ku29wb#*5YbwKy`
z(S5G761i#|<F3wS823>KO~&zAcbwS=tzK3*sG~wJoJ9k=ybUe7%Tco@oN{V8G>X~I
z?5+<p;u_)9Z}&%Arx{1zQ)Dkqv}g+X`^%GPss?hCu@{wSi9jv`uCM5Qm3uzwbdpYZ
z(4`b4<Q?Kg*80EK9ExIO>PLhI@f2(1)%VA_WDbR?Zl-I`<F`FunhQIMb%-=kFV{gI
zWWqnbnRu!bf(yRPZfQ)Eeo3``L(Ivk#|(-nl*jLuR4a~r=3?T8!ja`**!CG8jI6$#
z{3NzJ0;E{Q6ncHWnaIy!R^jNJvWHJgn->-~Wdlb%4USE^jI$MTKUAeVddY~WLJG#l
zX-KSRVpX_KsE|wNv*90SDB?~{=U(KUTTRcyaCpL*QG3Bs>2KIJ@-gAdYv$=*6kG&}
zazpx1+T)8~)S*MnKeba+-+uSO3|hZ?eCcD5pH{-e#FWw(yv1(Gn@yy$xN@LGA;E8a
zjZmEXBDsYq(WhG=U}*Wa-<t{LoPia&D>pa;Hzh3rdoeVWAfzh5EBbu$`$X;a%eE}H
zQo&pmyH7!1k_@Fk2seaPI(Co#rdmW`R9FJ4Vb-^fDMI{+#$Ya+MGPkf7lR*9NlZrH
zcSA?UjB)V65zjPf82O|hzi%6er=4CS+X^>b7{dk23xWqCp(wyK=*u(uGh5x5fSll0
zcVD&Nl<SdiFc+=%s2?=xtD!Oe;1+<^l5u9_l^VR-@x?T?8ijCqh_w{?e+y*I77f>{
zlPd@jdmbBB$Cz?;qdPcdvA~fEV-FZwQD2#mipkW5bj8n@{cnl1<$)uQM$x09|Kvys
zNh(^}r$Do6R|GDjrBc3PN-r?OV4}I<g_%+|ay=3v%{GcQp$d2@(vMpo#379=xB!U#
z*!p*V1~SF6t#K_Y!4Y)b?3v`ercYaM-*DV5Q2wOu_IiBg5b=%Qbw?!rN0esvfwA7r
zc7uRL<I-GZ5K;+QI4%%ZZg96n7sf&Dth)=3PO)4vXt#y@hNG=e_0j?LtUO#D4dI83
zvUBwl8b7gUG=4tC83y|OGBM@FS_a)^9;8iv!9XZL<{*~7See&#QM}MpJ8h>fH5tQ1
zi)|p6vVs^`aKr}GR*Q`?wd{k^1f9NM!LwZNScVZs9e18sfF=$7eFxwTm0YW=44_&}
z1}in`;u?&@c{Iwvr&)AcLFI8-P5!Vo_NB3gzO5!^>^^(p0)7=s$Rkd5p@GbyPiI4W
zh?tJAm47{&78s}0)UkGHer35=+}{wazZK0Nz?n7fhsa@J#i5_(R;t~n#n0$!oqZ_@
zgHa2S5)5h6u60l^9(1UXS51V1+BG-(T)1x^qqP#KTQoi<U!olzCE;Z%jqEKox}!|M
z``UH~`s$ce-#Rlw42J(0zcz^@3r?U?X0qa59!(QW9)N3J^};vE@t>YM9`syAX9q+1
zhn_JFDtL{m?tl@$)X@=He5x2)xk|G}pO9%KK6R~~sLSVbP3L<hzQ;G<Z=Nom-|}9}
zBC#Sjt|Ks75G=5e;{^`ISr_)(7rbIy)f<${L}JvfwVdZLZ2L&-77By+b)o`y*c|d@
zyn!F$9|#JdHB5Cg=sp<pzI(#RKza;YUO%<cn3`g?l5JZ)_GWKKlx<%+)#*E%f$l6v
zUO&RBc##d}xK8!4zn2%r*h7=7+EN;96_hU|t|CtpBo8RM1%M21_Z<Qyn?Zp-h^?=x
zPLL$c(%e4_SyATYn^Xfe!Z`y6sxlG&!0^tK^uv@fN-)F=r2<9GQ;<DEl5h>AU4j?8
zA0zXD8>l8O;d!uJzcSz?ZNY9@OsW_0tVLETOBR=x-)quWxyC$Lk*exb(o+SodW`Vf
zmPMGeF1~7Z)3$8b4rSO8oUeZ=UXz9Jq)@c_tdcP!Xawy5f*Zb@UarbR!e>QYZCJth
zHE_f=fjgu8(45DX4iIfb%r&$v)ePvU<ESI5NY{kV5nbD@>Gdy11}hoVV}Fqvev)gC
zvWKYc;>&GdR$Qvq5QoWHi-W<TZhfJt`n0~i$Z2t)2=o_(^8Z!<5FaW46tI4o&VNCd
z(H#KV{&-C~noIx(lfn_QN03ayYyB|70-Yp$QscIUN<kb9Bo>*&ne{{fjM1R@;tvH9
zwC2%)7kEhy>n`l{#0TFQX<6aV?Z!KnDb!iWU;j`dMj-zM@mE-rZX8MbScuuEA@Rx1
z`8-i0`>P9jF_Fr5O(9Ul^2g`O)3D3{Cv0*np%(`nu7XbN1p=I8DGZu5%E#HiDBFeG
z5?+ZSp&orhcQOG5a%Z5?SKJQsS7KZ8Qy$mU^*#O@3(&lNLE?8$5*N6R{;Jz5sZ>iA
zKx3I5B;&8hpQ|e<pvkIzlM$zLUB6Oo@+MZ1`70e#q@mWcR-`{}uc}78DRx%~)K)vG
z<?9qs7(>^W&--srdK@LFP4D*ahvCf4asdYS{FLpT!T`eZoUCzsEeH?OUQEMb?m0Pu
z&ldHh=8Nk=+L5H3Q>VRium0b3o(Am!l!+p~j<l?qOMQuqT@#L6gP$4D+T+&qKkJU4
zAIU7Lvmq4SzV`!>?Bpt0op^PQxC{on5_vnDfYyizp_ZudD>zJ_a_u~!K8^(z{52M%
zm3_0HlL)ojF^6y0?;O<KcZgf}P9{$)+t(}PxXnK$!YgnH+vA3=oxk^d8vz`6JMrTk
zw??HNve{7lcYWs-!aT~IEAs&XZYK$Tp;ssx6TbQ?$~~4A9>QLiI3-N@RZD$OGwbSL
z8;Wc4K38Pwem9P1-dy{%)ArO1XY^0dEBu^BGvX&gW4C<f6rCCm-94WjNcFrn;Jt<i
z|D;wy8btZFbkFih?Wa1P7J<(1rP6v6Y^w+ZX#|lvR4w-XQMZeE6{#zIz(fUv(w<rf
zPPkn7h}o%E7=Db~e<q6X8`$vkMR*E1Um*o5#v0C@QF%+ZtStXA)ZJ&^TQ>V^ns-|k
z18^OlxBO(e5?szZWGl_7gka>QvdV(Hm31u2fMIWW&Sj<2N1vGuUm3R?LfRz4&-F5d
zYcnA~Tc0b7-$T{MHNNxWKTl>;*V?=bI!*lSNWwP-l(dyq&<bbYs+(j0pvcY<XFP*8
z&uAc;JJPF0WYuvSNMVOwoH~)*kfSGL^!r_ghC{E=X1A6Iw}z34!P}l~9<4VA*Y+zd
zLFo||PP@Hk(O)<)m3edt!=pSiV8iTd{v;j-($97}j9m5IqyIPA$MK{m^dSwHUi|*)
zpKztsIjrB)#kg#GjZ_WC75dF<!^D@PD(wtuYfB}K%zBjN-i{O#1s53g;n91r!nn8V
z+u@Bd%$D6T={B8S8w@;C_Hcb#s(rh+El9evjxRH;<ME0~BbM}JBhJE!M*T2=wQQ4u
z2F6a~wkvE9QQWc2k3C-=*HRQ8ws1*M<vgu0ydvaF426P)*v#}-&w?#i@@3oX2=XMP
zzF|rLHV}+%Zp0>Us4$tw^r8@RFLbKb=|*$OY?s4vnC>%^z7AA`?si7UfzcK>uUvB)
z(<zOcbS9-<G6Slo6%hq(+>F4QaJkbk7?%OTi8<@3B&8MB%s?bDGNIyfTZFBoti|!M
z%wg``ueU&If^dpQvdG%)Pu$+nF?ro2e#jlhS2@rlFH%RO%yJ4UnpJTxT57gP_%ihg
z?^oYg7PQKpR%>rnE#I0~EqAR%Yb4E|@cHD=oERw5@*uKr#L_aKhlT4jjz|78wMoj3
zjw!+jAT=TA^tHmW@})v<Nf9hDVs6}U)l+6i76eQ-e~dlx_?`jM@mg2*y+4p@`j+qi
zdqU^DxtH26(c_nSW0(_L0e2{R;C)7F(Q5;F<E7@YjGng^9(PyX=?GtVzELXR&QTzQ
z6qM@|My~D)o$Q5yM&O)={rk*t^Kf8SU0o<nIS%z2-eC_Yy&HM4KO>fT^^Vr0cS5D2
zCH^2Fs&)LGZj^)A=FuCX%ZsTjMAy9!j)*pZSYH-8vX#V;k@6#F8MGEaJ;Gz8Z;#ib
zc3%5iFaM_iM`)GK#AH@YUqh;(p3TbKp7zLV^e55#uuRS&D{X7#I-@Pzn1Ye^`z8_J
z5biaal!>$Or3~2QzGXJ;_Q&fF;n2NPCr0{&Sts<H^vJan<~g>^4>8E+!wNfzZzEP?
zK9g#Xbvpetdfy5_qHkx(1MRQHPHbw#=lCnnCAyhLiS7G|aEF1FWE=Qm4OuY|gmZ-4
z%X-2_JE?Z+M3|pFCc0Bc?2ChYFFd$p#wLNvx*jCWsy>w5$<Kx*l>xr6&cou)Bkug<
z+yTQ~^v!R9Dx)oQCsmg^q5<S>rjjn39s|p|ovyfz-#;xcgq$#Z$pOx`N$h}Eiob~R
z%`cUAFxvGDiIuEVBW`J%jmp%P_Saf?^fw^+KiK1LAn!q@i}}BZ884j1fO}C_H@D-1
z6`>~h@bEAeUiYmeKQlyDd=cM1KbxX<Cn2Q_vh^B#aP75>GnV<GFXN@gjg-RqD{xhZ
z2e3xA1g;EAOuxpLb+_-~5>hva??wJWklRud`OyX4$<rkZoG(w`Gs*u9j9##{Kxq}Q
z^kd`ghP~q9rArI-uyD$&cOkX<DBTmoE^Yl2!{c2P>a65%exPB0d$|d}ROg%ZZl>DU
zp`R|%rQG*9%~r8ovwBVY#IL8@af)+J(;!VF0Wr|?cC-~y&7+QKre{)qAc7G#!hE4S
z(mCLl;TduiKQ~qfD>UXSp)3bSKtBTB>#aq=#Rnrs#Xa1-5^@nc3UdLk95Q9>{Q0TC
z_ezWPY(%BY`<B;qA*E9%9Me|wU)m4HfvgPQ2H6D&7%nZi2-zn-hFv!wU$Ab<giDQY
z5u8eF_0F^%9x<DHtQa<6L~u9Z*)gG$e)E|Ztj1Olf>meHXJQ4g$0bI{bDW-LuQ7TT
zxK4qc71gIJv@|&^5TY3Q`3ZFayfP2RN5VX8=od)FyQ0@^8QnL9c|fL41x;2VyvLV`
zYd!MjPVSm|={|Jd+MbmdgXDu*fy?Uzl-P*e1#TsL+z$(+zrju@PEELilgN_sw)g8>
zWS~k263vrR!|%_p>K$aKf+D-^^H*_+S*b%|iCEx)US@*;qr%vE%LXjo5+nrElfH97
z7`Zt+a`$MW-i}BtIra)N`R~SyV(yG4{97Sxurz2vQeAmDwGr>9g+(XVdwiNR$~^dH
z%oVBucVV`#02gNDmU^0#Ho4tLq514L<R=1SNngNyMtkeB_W-ot_@S0k*L`pp=5q%9
zK{eypvpF)sPcI!UJ#e>cdrkTl7alJ`#8pJGIcD{>Z)eEnJPJgx$72FVeoVswtK<K+
z#;Eq=jN&nOAAXTeD+Xo}m2ms!CXIgpi$u%TaS)HnIR1x|4Q9EE*SKN)l+E%QXtI&T
zU=%b8(@z!wdAIGkK~?s|DT`hmM=!sal0Mr&!Qb|#jx~>p$W0Z0RN&AhFquT)l85U~
zB}zeqgof^v_manEx+1PV^E&%*HDdZw`7K}Nc`_Dw1P54FXA^g$(&8UxA8fRajiUKU
zLF4tW<6+-xl)wXSr>hs|%WUsur96O6EKahmhq9#<0EashGjHu1?AZi0+Th;K7~*c3
zrZr&>zDyWH?kN-!Sn46RX_yEu*M7D$ZPN|lRznl<z#+Mj9l*%k<B02c!uKNhvsXjf
z`6grjK$lwD@}p8eDH_L#h_1WdKYZMWHX}U!CrW7Au3NFvqklZ<aG<TSnPZD4;*>BB
z&~32qnXg80(YmsGMVK8+rw6Uw?UY_cDU{2hyN!cQ!XMIBW2-5NiNPbxPpUe{VhrDa
z8C1bah^iVDI?5HU3TXicbCux`7{8X!JZ$?-#1I*_HFk=?HYkESW62hPTz@`z8fKh*
zub@%-PH}A`{lYJ9>JU6MP1H>sgSY!Zlc>a@yYKwnYFWLT9a8I8?fbQ3h+N}GX>Rys
z;})AvbTvc43(XQdN`9x%rQc&skJ0%v0=<B6%IikVf4EnHUuX}_P;wWOObD64Ah77Q
zye4$WWODTKLh>z-%O^QGOVCr8{_VFsmekqpol9;G@|Vt`y0wB&Vy5Y*SQGruZ83uH
zy99Vb1<^@6?e@(r>#Qo{*}Wo?j(=gdC>26awmrQeU$&*Bwt&q3iSSmi6Q?VM+_qA;
z;}H0DzK&ajo7(hwp^#IE;-*1{@?QZjJ-J%s_{8;;q{z=|+MjyA$O3sR(!UKfaw?_Y
zz-k8&h7z?q?ws5F0|QP88bd;1yGD!Sof^WL<)5qrH}`kv^!7kbIA^7E$J=k}NceLk
z%C&Zs&DNCVeI7l<bTB~UFVod7)3FTH(JW!wJG`4J<TyD4r`pK#jLx948uNH!8QxE&
zkb%lq!dLAwnDq^AqbQ{a)zz{OcRKr8Jx7(k#1Mz5b$2R!%ZC4y>UPYyFzR(&L*pmK
z|D5z|RPt;~wMxGZ!^Zs}2OdGI7r)?Jq*WHb)k6NmGoqK;wO)OCEH#hrloK}sN~2Hq
zZJ{a~t)Z&*np5b5i<7lRAs=6}U~aHEOn3`60Yf9YzM{T!e$wjj^LP8Ih|TlC6xKcn
zq00qQCZgJN*PRcJvbq}D<$B-T)ZU-$`vtpS?dhqvw5BR})uJJN_<_&QPv08Pon2xt
z%mwkwesR;fl#`^A^eJ|uA@+VKh0GQiE_pF1?b{wKJJ!LY&R(R2T~IzIM0c|6f8t0s
zi?9e?SifHOn$$o&Ydg{31RU%JRv^4pm@@zyLt98ROXsxaC#Gc2p<6r%^4cKks{A1x
zG?D=ku}>0Y3H;B6s8m+Vk^y$i%rj_5!nFA`n-t}<`6M7Z%~mCADTLlTlaLRnNRJp4
z%@ZF?721~RTF2<?YE3O&qw{w7UK^!V@Kic)55Mx9SBl5N8QcwvhA05L35~1ExXHFH
z?jhN4H%I*zzPY#dI-bQz!;;BQqN&#xBrgO=si8NVI*07HmtTwzhN6D=AI9#*gl6UT
z13;$cUm$P-%zE$F)Ws2X*hvIr@v%rVQae`sIfb$I|KEa?Bk>^u8ab#`4f^xRMI15+
z1nTJPleNbKa(_HYCH3`Z0c2(qy9R)<hvTiG^ciH}7o_^aZ-|rUNBahDmlgfhY8Tx<
zL}Usrx*f($dgYz_W*$G$oqe`;cvROCubBUdiQppi8>IgjhrRHl9e`FX1t5GkUr%}c
zlT*7<$&7FJFb}I2r?3N1pPBFffCriO%S@LI<xk)yrz5AV(lFTWa$15|TArHiAnz*1
zrGv_wDw36je#irB!X$vT;+7b>5;xVOb0G!<ktfqIIT1y*L^c-=*^>>R{UP#-+HJqk
z*-9zhYP-u#nU2Itu7HUyrX?PyexolaFGYWS<6~y}>?Be6hfyytB3Gepc{JE-Jv3p!
zedrdbaIOn#^<`DEF?hqmPto4Frwn-86j>J6XVB%<&QmojdivAYC~0DRe5Mn;WW8TO
zi3!fqMIy#?A;SLbS%sAb#&?rcdxG>jQ{;6#P+(!GcZ0)Fl>f%aSk#<DmF+HH*-BFs
zp>uJGMu)SaOZ4GV`cGz&c78xNHp`*OF1lqIqrsd^i+oZ~uRK#S(ffFJy#n7=y4zt{
zI|N_x3dm9zUlTeM$RRC-ddx}2U;0j$zEeNA@;TeV-;9V!FCcewl5(kgj<Y)@M1yjW
zEy20``eGy#0e@$nA>j5(y062`??tPelvjEN`M3@XyzKOXEkCb|#!^5T&cwnUiWzdR
zw^qW4L!4-ncVH<?C_Lbaykf45y|IQgB4f?9?sG#EEZB$~VkbfHF}3CP{6La`Y_6hj
zF~EHaEBpi7@&tzW4?!Flp-Xr;lweBmsi0e)NuzZVb4vg1?HUBN@%>hRrrY2k(di!X
zV96$HG`7(5D7Yc*`qz%}AlwQ*D-CR$WYiwPd~v-JDQ`L&!I6#b&$U&%z#7idXbYs1
zj41BDJLp_`fr_jWH(Y#4#qV__$bxD0esjjfHVP?kjomm>?UTXOurb0F+E*%1`;?4!
z#B(($JwqsMCdt@&2azOdr%Us%8*>Z4jw^8?1Ebjjb%%M+kJ;wrP#2*&e}xliPZqka
zdLBWH)}ORRSmoyX&z;u;ZcknY32z;*GpN|aM&(Xa$)iTV3F}5l)3uzf)WTwUJ6UK#
z)z!2bMNRpxsCl|0glN;PJB^)q(yV_o>9rRRq*O02;qh184#5~l9lw_gXp4B~eh9EW
zp>O;WlcVCVoW9#N;lMqeW;qaj*rP6QoF0x~%+MTA^RvC@iV7IIH|Td+;NB`z1^f9d
zvwnr{O033$n*hjzkT;`$EcXC0X&1kvIopC>SRCmN`rZAvzc>nhIsClu4?7Q(Ik_In
zoC6sieAa(9C$^<N><68wG!!Ef>h){Z`C(~k`H=mRBI;A<d}up{&s2m-2h?G^pwO~U
z2{?WDmDQE9x_5QB2D=%MWHEdnoZDQlsARv(wlWd$>)CJDy=XzW{9>k_Z@+I;^UDE{
z$LS-C!|)y;eSP5*jx{<xf*IUP{LS`V!o@3JG7U8TAuOQnz_BY`mPjcqu!(ht@AW#_
zuydfSl}EL4H7Uj|xngO;S*8Tvq-(|G?xxKsVRspFUJjcvbxvEV`+?<BIU^;eQ@31`
zH-fQZtK-!wMO#JPy#iY1`;U_?{lU5cz#258;11&0@ag^wP#@prE^Klj*D-@HkPy**
zXNHo&-G1D5CiVwTf(Mw+Z+drYtlB{rFDC_Yb>D_9DZ2_S{XH6Y9Q_N!S2bDJ4Gvw(
z(JLTa_4s=xjWVPS6sFf|TASx{#kb8+E5VpJ$x=JnFY`?~ehmgM^y^t-?uX{s<&wZA
z$+?USqc!5&HD+|ZUR3tn!dp=p&<t%Xi$;<j^ZOqMEkU#oasA@}|5Hohh3a4f8y<7I
zTBWT!4My%{UZF9TUdaA9{KYy#SxRN^nQ0&WW}VNj7LT1J;Y#vrTB~Zs#|{ubapy(9
zJux2&-B_}op@!5O5uloLl*Lc_JOD%d9ZhZg`fJ%t=(4bv1qeDa6^0%NV1|ljDn_iH
zwQa8bOk{Li!3v!R%UbbH-pL@6q8&~so9l1<@_n_MZ0D%c92i%atA*OA%r8@UW_RHk
zjyk+Ct>GCWZ8VnZ;=7WhMZiJ9_XW2rqGz8RaOyK$enw+U<;i3~2buNeWZ4aRJuUd(
z`(KD3ZbOQnE}IkO>cci)usUYYW)Y7Ww8>0&N+N~FCK$;ba8d{oDJpGANS9<iK%7>Y
zOe_T5d6fEMCFqk-9zaJl75a>-rfCh7l!#`WV4~K=`d14zQqRDCy?QG0A=-^}sAH*9
z#+?2vhN}I+f?%(!NvAiQ8Q=9I+(Mhz_ChTfRo)ONj|_h!#iKh{Vu@%c<AS4{QH*?(
z<ma<Rxjl{nmK#fOZW4}BnAF`%kcl-a(|pz1_xX07g*#EBtUX<_6mLLl=-C*sB{-V3
z9F}`+l5i<=I$W&w2ETj4wgXA>^GkKDgL#T&#lXUteqxok5l9^jsq@+H`u5~losh(^
zbFBI<1ljZ6aY(|MT}#(^c|+M#A*ZILdMjeSX97pJ_2b(}rs|kjk25(aPL{uVuXE98
z`K&G<-ygfN>S>uC6<Ki^wnXUchB$I^F5B+XKS@ZtDQ@;XKw2h<?g=|zwlF|!easWO
zsq?F;l}D_8j#$;{cfcE}H36s4ysp&=AVxNSImoPdCltujJdrrUG1x(dufa6=Bf~S6
zxiIebshBLSXsm-V=gKmvY2n+!VQ7NL#F5)7FtdqhbhHrgI0TyXd)f&b>kiCvF~TmD
zQF?xw{hxUSVu=s^&Pr)ztbeMSNP)oUXtY|ZsppuKqMwD-J~1WbDVgOPJ>c(Wm1ct7
zSC}BTYFQEV{xNSQ`jc9oMe0y|*eBIKeo_)N!lSL2Yi2X7`zz{9#Cs?lqkF0-e<cAA
zyzRahX+t=AqQC9>8!bs8>6PA=TBSZTK=q>|_x|$H93Jk0SsWRs%-QvMIMT7C6i3Q|
z@>`(Y^JAzo?{CHOLd?l{cOmJuAJs{?3?sT7t)D+x^9#S$LSwA_dIX@*zgG={v7NF@
z>8_}@l_J(jUQTjib?T*vKCND!0a_Ja%IRx;%yypH!8rzCO6aQJtw1i8mBZ6Y4ep&9
zs6K6?UIUE|igXfsc66JGdF}Krn!r$oCKj|2TO&*BryG&@W>VCM4SgMwfHN2Z0&7+p
zut>cbQ2lA7_0SyHu*IGD$x%r=rG1*!R)mSy>V~!NY3Vg34%Hef&9Dw}Ln7g8KSfaa
z%uKcCPnJ{rzNUefOfR>$IUL~<Yx_&J3flVQ8Tff(`CB@Qb{Z8$Pu2uI=-~tqH0-XX
zbfw^daN`%c1<}i!5DN=1I~Pe5Ex+Q|QnwxPT;r!qGKxMYXmrny*KBzZP*3~R8Tq!?
zT3_NEkcOBv8f)zRbv<Qql+Y?~a;^ROmlK~0(nv^VHYO1i`oy2PJQeZ?8bM)0fP5Gw
z$3A!;x0A%ISbNL_BQNl}HmUp)uKKpqY*7QG#JRe=<$D>_?}upAYEswjt5sPRskO87
zLc2a%kD&hfRTbx#kCB6eTr4IS<iI{YyzUEDn}VWQKaWFE71EJa@S)#6iH4hN<c=ed
z_xr2*EX0UIM{nY1F<mbCQ5X^*kSx-FuVFy>vt3fDh|=BOpcA#I&1`apSJyN-iE`2X
zfZ0(wF)K_c69iuPy<ShHT%tlMLPsC|HKkV9BT%7dkgCF^!9HB1L0cmq=u++-ks$W2
zAF9vFgkwyQ5}3lgRic?$+;(^+f_q2alxSF`;qcym{pE7~d4Y|{p#^JWC=sSqv(jJp
zvxUfh{_WblOt)hV{NuUb&6gOrp}i_@^L!~ZYh@m*K{d5JQ>vgQl@&x|w+;#KOr(*$
z;_>K4`lvOp^D}y1@))OskH3fbyHpBB-9kE*XrBGr3Aj8#K)hR!TEatt$?1Q(OCHut
zt=afhGoED}CO>7Dt(X02_)3t4KJ@!JHVBLl^Sl0KMSx}zo^gHfbtjg=x+cpHQe*g5
z|4$hOGK9dqWuTd^3&hOA@aGd0&Ol(5M2^i|MgPD+0_jGL*^rU3@tbhj*Par%9wEDh
z`j{28ODK~yNK<`PVof}PAz@G54CNHWFSrcJAk3Fth7*G{65pM;H`TT4B9k;=<f@CA
z8Z2p69YSCPyOl7X;(__qE$a1nhP_0jfz4*Nf}rZjUm-da%l$)iYmS6Jx)o(k8AP5>
zR!lsRJ*BuJ984fIb6!LTIz7caymt)&#N|qsPjPZzORFCoXZLowiZ)u&Fs}&9cu_;q
zC;E#px+<mE$D!eY#zfy3lyl)ZhtMB<@>6t8q8NwEC}O)_TtU(#>FJb%+OMH~uQp*P
z+v{)lH^P506T+fHsyOPMrLv%=jVZ~_y&gu_=APjINVC%*X@@tmr_FsO^#l09fRi+t
z+s8I{cAo0&i+O)5FJx{P{&?scG9*3!=>Gd^CP(wAUE+Ntrq_hOzy^UYzqIjIL?VLs
zYfFtnjLkSw^jzxTH~F6w+KwSDS)&5rm9=ii>*ULI86A>k{*MEoJmjZT#o~lxnK5F`
z?#Wn-1G-{aE^S8F)t);y$)6ZrdM8p~S<N|>P|KX%J>3{e#x2)=B|X`uO@s{&nSfwX
ze}&u#0J!mqD?d0NZUw1C^`^rcZwW7s2#%4_@kdklB$0@!S*9m69(v6#DJbWB&Fm63
zG3I2d)wkF_zh7)3uq0x$j1qC8JuYzhRz}D6d#I;cQG2!4T!d}8F?Guqhd@OplxU@y
zke%-{1tmV%>tfu!$Y`wRRTptZ-X4UslR3UV)1{+~-x^P_w+G*{^7BW9pF@Y86LCvR
znL=7&4K}5pIQAsOWY&6Ro>{)JX>~g3haf~YsI>MHi7em+%aW>N>-vWvO#qz_`{@z2
zCHFs`$~C(@HD=Z~Bb)qahfTe(7jpM|`BwF(zVaKfUqhtFi1lx__^l?=Pet<d!S7Tq
za_=o)MQSZ*(aw2O+ajbeB!xpa=YQ>^$?jkmHjLv(9jUS~>S<|&bBuI!KbGtp?SL_2
zuG))Wd#G1tzslWwTT4eU!ee%h;Iyv0DJm;d^@5#`D@%Kn5_~0&xUBRe-h$<GkFCWi
z2}mou6au}M2Fn6^JrL=vWVYX7FzN>Wo7HN*ahr|2Z&j4Him}BmffASWe?|ZSxp=BL
zz;;%<{aK=ZM0+&yEGRhG*xsJ)ybTb%HDFfy_|(+(rn}-KI#8V61&Y(ROvB<i;Y614
zJnnREJE@Y_j+etuc`FLiksbLlNR72i?oy?r-oKY^9;G!{$2glMq3m7A0yQb~$jQ8l
zt1To~A=jSuQ=f2D{g|n+vp3kJ+l63Czj2#>xl<wzsS5RX2EES@A+6F~3+kY=cJjs~
z*Ce1RwvESa>04hGID5_miwP&}tMFz>ICM5D0y?~W`%AE29$3l^wA1XSj0T6#$&pF;
zcilB(k|~g{mLO%!grj!j_@M4V2$bVcXztRDOcc~hBl~10kLdMp2$%ROy4z(jPpqt)
z;go1w@iDK<bu`U|v&8(n>!P+CeaFP)K#asu_I+%>#hir~UH88pHe%Su1vZPL%}K8s
zFh&F8_S@*hyr1Ge6s>8Q7`bt;ZDzXW;wZP_XLoHy3;F8uS0H9BQ+oDAO3{EoKZ`Nw
z#Rjv627{{=?UM=&9i4HTe94tz9UDvW;*6VGuaj@_^=Nt`o}9}Q_^+nP4}PvQzHK#S
zJo_r3cujkIx;;)20Ka)E*}Oa8#x#U_wRBGD$i#{5)EB$HQ@-STBXv4c-}F2b{yP>B
z-fT8kFZ{zJB&KQF5Twx?<*m(aG)CK~JX4XZjA+)dxHS!l9HsmBN>2qn)>a_l+*y}3
zMpB1ja_Mh{9WaJ_;}V<}XkB23@b_iQZBG$bhepx=#sb`njG-bS?z;8JFbZj?YbFto
z0#!k$gGDzc$Ir3(Q7Vc`zafNl^!5g^9(faAf`N`e<t4^yb<T{>KENC$B<JUeFzD0R
z7@}aHM|Gl9|35vdEC*n=qYKyTENQhtA03E)Ho?_-h5G7NTh5K|YYnv8ewXVg3rxp@
zZc#k}%D7V>ey<chn$^k4ZKO>1mGR;waZAgD&w~g7_Nwh!PcUr~h}d2TQ*@Ib4CMOE
zH+j*Iew~+;w#~~r)4y7ayQ9sXb0PYy`JOx!wsr}<)ul5Kj95opcf2-#LMevrXJuI@
zdEN$|$nP`Sp~%!0!d5eBY^qlTijT9G@4WRlH~poZ*V;ip#0Vk?RR>Qi=X_oWZlg<W
zSpQO_>IlP3MLOdGcdjJ38V<kv9mn~6(>ZGifHoTUZQ_Fht{$;d?Jc?>?X`3Pn8~!|
zt+QRArhRe*`5ZCs6jx=rAcN_Ls48BHA#y@K2VI#sA_#{Hx=|t_9xD6Z^mrMnWcfTk
zKXBR7l=_B&CNsx+^sbMI>wlWmIG#vHJXE~%7fh=EeDXRRDVmfYW2xRQk@QWbpj&)=
z{G$^_=D+5C!RIOv`PWbwhe>jgNSa4m(Oz!)hL4YUB|ipG^cc9McLI^XUibMcquqNk
z-TSD!wjjv|TQuMzMD=K{ynMTEP>1zfgTGs~;}aI8*o=@7>QkR}`KBxx^6`(Oj<VX%
zb9HyMY+LcT5a$mwA#47S){*MQmvR=mb*8d6PItmS-Go>m{{vp{nDu2dO^rd5xnz5d
zBU-UyhN*shem2a5{0_TpIW)qWmtRqKBbPu6o{@bXvpAJ)fg!9Z725x;Ko+Q3g!}=X
zd#FIZh<$+7fkv4~asLEfw+8}rq6AhY%hEA24(-Q)z?s5cH~_lt4t)Zzob4f4Al-h|
zY=9p>XWl)da9R_1tB2#22T=)}Vvz(Ul3iudTF%JFm1EY6-6g`PFONxwk%YjMHUBNk
zMb2hlqvnd_6i1vF7uSX*^J!UvOy=45VtDmGBLibI2>BC;>@|^2b+!pMsH^7{u>upX
zLNDU)KEyV{7war-P1J&9vd&*yEI~Tbio$9qkEr3*M*c1W{79_s_;o!ME?k{CePNAl
z@=;LNCAUNpq;yD*VEeKZLDr@24`KY~GvS=vG~TAiTGy7<-!uOIE2)3Dl6e(YQlkI4
zlCR>{)=Z0aHeu(7O+2KNJ-`MFU}HwRXgtkq@JD97A@3uq3P+&H>d6zp4YDeDtOmdo
zY{I$qERx;7u*Tnqe&BWrX}<g7F8Wu`k=aY}5bom_d<2AyDtTf|nw6oV&F<mKpK6n+
zuis5BR{c7blUD@4x6lScOlMacRy=T$Fa^Dz2wwiNn1C-DGF!!|FSB7G#&c1x6NoPr
zewTr~&T7ez<9^@s5E;T>a|Cv}nh6|S=FQD7%H7}5Zr*Ihg+ukqfpXUn?HV;(=+L3R
zd#h(<To+IKuND%~1D>A&{n=ZZH?tVO7xJyCh2OoC``@B;bjKKVN4KL`7_H5pOcaQt
z_#sb${-t{U5HRT+3l}%IL~}u`D)M`Rd&rW}G`b+X^hh(W-DA{0`4syFvT$CHqYaSg
zr8mA=YooGlx!=E1^*((*1y+G=0aLaUZ}NnuDk%-f1SA5A?S=7o7^dFp@NnMaLGR20
zoUbp;bx6=MR^q>SM^&WemI4c9@eldIpz70v5MFdB0=HsIi(^EEuCj$it)mBf^*w}E
z<fy-qk)u{Fr`>)rkETn@?LBH!727XEJZ%!E(C?Fd<>?Ww<ScQ~vZeJa^MV)BEepZ3
z=VF%Rpc_1cJO1B-GymyXsQJHKGW)<KZ&lez$^M5+!an5Z(?5Co)XE}XAt@*~mv(t&
zWgGG?fsX0`h5wtGh-WO0M+mbLguL+k%ZoX8DAD{_JGgoEt{e{(DF7~CqIV57y$L`J
zn20rN$#mE2ef3~EP4{L}xj_rcpP!e{i7jR8l_sw)<56UpIIH}~m>+9t`Tp!gt_gs9
zW$H<rH1aZ!m%G{6`z47}*oxuBz*gZiHXwc%i~tM_In^*{aOmE#fw<9sLGgDyQ;jdt
z82NHY1qqvYT3b0<X|9_tPUM=?gQS5$ag9j{ec4L}aU+0M3$ee~6{26;yG%K4IjMOQ
z2EJj~WBK1kC!=SPybt)k*{Pxb$$WMkQpW=Y%da9nG%sHwzj^b9h|hNZaJw-aP6MFE
z&6m*^;iXmV0O)~mIVwL_Wf8cHerzyjFang>>}>1MF<!y{^tntnpi(FLhX0Gl9zqYr
z2(9@3G=Kv6(0DZTzWlqf@p=HvA7BsyU=p*9wweW~-T?F@#IEhF@bX>jCNw7Oeo1jl
zZloW{PEs;ze(Jmu_58oXvOKEjTTpjqH1FKVYm!mvAHU<&6V|CCjMzXm8)qrPkzGAb
zw2jX4j-@>R8=*{ZPhIu+Pbk&vE+ingi=6Jrr}t;b=8qo;Xi^$xW>gXqlJ=4)8c0PW
z^KwTre3kbH`4*lO+39>%bg&Wma(gC<PvU@_^4cV8Im8doPH>x3)Hds8)}_m-51$O!
zP3SD<Q9~)*@rsxCPA3I!^SUmnf=nRpbbs=M#9jZB`0=J>CX`GcpqX1KAmw;c#}s0u
zh7;+KJ&Q3cF<-oPWuFYu7;A5D={v6*3-k`3OA*l3O2NpsTSe?6)wR<Xh81l4XpH)C
z`YEol%nH6ht|=7+URnLRq%BP@yL<LN>xKJs7KQc1E!Y0Fl;tgH<9iuf6>!>kvug~-
zUsGXvR<;9Xi~AS<f|0G>@2`02ecrgcKj!~+X_g)Gt`7luga=;@OAiYledM&3jJ099
zR*)^J*^03|mQj7kw4h)kdE?Cxs|f4nr{?1L9Lv?9rW_e}C58Rp+943r1LNl(jT7t+
zQoTmRcw8Lj;4PKA+=;N%v5-ekkf=U9da{W_MLeo^9z6T-%U^#%YApVcTjKc2-^2V@
zhnWuji18@q(tc@wQ}fsWZN!z&;lbJa#<;e6wnS@P>9Lh-)qApEElEc=abn7vjmtkH
zF<Os9ar&Y;-J42Z6_F@q4c1JulxA+#tjTaK_s=;fg_fqYDKYtbT~#a9(_eO-y+}1{
z3CIx8l@-0}Zv4IS<_(Hr!>F=YK4F>etMZ(rc+UaQMU1v}v2rO&FGrnUY_+n@R|_5b
zWpuw!v5Y%))Widh+@dRGh>~gZF3`sN1HXIyjk~ko-xX!!bk}EY1_F0}W}9?Ap_HZm
z=Y&+N5>M#BP2(du#s4|-!{IBK9VImzI}}JDqI52uk7oT?BHSq7S$uZt1r2F0d%<BM
z8X-d)`{s5bC_kLf%Bp3L8aYeGe`v}Cs_N=f>2|7q)!n6o8Y}x-2y;;AS&G`)!U%66
zmF*<eY_7*9ng`PZZerntPD2gdc`s^A%bc{bmK3e_BayryCaF*XVZBO@)fn>ZTmFjc
zOgP_quqxYBJ4c8iW3<0~1h6AZWhE4)Z0@z&4qtqLcuaa)md{zi?;c~PUQV;2<($tK
za$Y`Lv#T^1UotQrKf@A#^q(J$$nj`fbe?Qcf76y~Rp80H(yrU<CI*x;U&f@5MIR;U
z$<~*smvY}TtvI$Wx8jc86j*H9H*oi$#b|&(IDf^hg8kl(SomI7sY+>DxpWBE{=T7D
z^$K{yvdZdoB>QKMGrL?vQC{oc+1*kNyQap1(}y#`X-!A9PLXGETC?Gu?>?!cxH!j3
z10R`0_#r}2mK%Z0?|JdLNm6~Yg6q@z3G55UWtEn%U+;XK1Hae^ks-5KST*K<+3|rk
zZ0zD(8AA{W!GB3;<Q_Pz30-`qHMvAC;cUwlg7AXnwNeN|G-uz1iSGt)n}L3ttPPGo
zSFhsMmA}sP_rL{7+lYDx#kXJlqw2JtdK(>yYwk{r_}`YnqPF-)`Gt@A$pVNlHz_ce
zd`>41uQ7>^%SXE-N<?knM)c}%8?x~7D!08#cZ06E<)(7~MAeehH{|9|t*8F*0To5B
z=|-#$sZG7OxOhRQoq^$9+s47@ckN`)nVkr`AE$l1rjQXu$(a`HV<uG)Ftg-u8abjq
z{sdiK<C{eOmuTNn-j|~^ta>SGg6<wKmYqkrQ?Cv@LAFqvzNIbc+TeLUT47NBrQR-P
z<Y&XrlH^^j;?JtY>rHhP#=U|q1}B538T%I;qF*R0a4ARa;$|!Eg3GEaT&?N^e6;ch
zli|M_`be+ZjJ#^R(dxCVOAKaYv;3wz!q4j)vo0(5Es$*c!j6q^^}28R9s7Y5l#xHb
z5oURTboa`C6Mkdor>MRke%WwQW4-e3*0hDUTA->RZf%a2A({4H(yVd<(&aT=UXc5*
zbd6&5>){_HbPRfL4Xi4S7`>C*O?gz0Na4vrks&~7aOvBirH;?m*ASUWF8IQ$2gTH<
zV@hAc2bA<F<G5{*7hC+BC&C#TX?7OBEf1TT-Ksc(a;jPjp=<T`hF#)9P0}&qa{uE?
zYwvm=ahdIv=nPq10-gXn%EQYckg@-=hVaENa0Bn{%3O_fzt+B>_09Cy3GdhE?LSnP
zO#FFAIo+k)V)hJYdT7{29@zB10`zSaI0*i7EG_oG6zkWdve9e|39!kbD(T|1#uD%N
z_~aO%jU9$2)cFT5Ar4OM%NPA^P1{?Fzg!G?y}aHr`Q`;Qq$c;e2oqrrBiiK45yl`A
z#*)p{GYT!3>PKC8CZVJ8K%0|Y(p-hNYf7YTzZ>2Rjgb^P_blz<SJr5GuQkERC$*YT
zFoAz?nS}|yavR@tGc-qa50<@CT*)H+RL-S0d+Ek-8Q|Sw6McJOx<sQf(_ZTm#Hjtp
zVLFh3!w5XiRHXTbQl6n5=$g<Vrfoyj*)b>L{ETFNs9WC<6Er9)ccXxa5P_1ARy1yE
z=Ym_t@jwrv8s*J*6fK@PX3&dCl4q)3NaCaIS(GpFGyTgL?E*S*ZI0xDPY;!*BUHt>
zN%V^Sd_q_a4+LY_1CwLZ-Su)A-X=4CrXg9$FFu@R4J>sAc>YK#u&63hLEs)@tDn{U
zK58d~Zy@pgK@e`5=+>8N!=>a6_(;nz{tiqygZv+@2E4}$5F5Qag2{hK_2@D8gV>y>
z<}+nvRT8Hwqyv-zBdn<w7BK^aAw!sB(;H9oId}sk^m|2(DOILcu)PA1x1r$31KJNC
zqLB5PUFAu2YHRv_$dqbrkHrbEnY6z6x%s><QcGfuex1f4rlj1xoEh&BK|tsUl4qXy
zNou0?++Lfl-}%{@)@h(3lje@9*AG`;_rh4R@4cm$iBD%*C3}S9j1m84rw5fO|KUBz
z<45`_!av`<gXc+VX;&S3Nm+mh{--wE<X;<Ry4~G1(F9}1M#Aq|xCeY4dwo5<7AVSd
znDk-~DP<W{rthHOn0=4uvu`ePLUOo_d$|lX-m`d0%k*Ne_bb7+E*N2MM=SC&WiWJ9
z_3xQN%CM$IH1q|W#GL-vY8iE3Uj?h(K_pWv+jYOCGq11hfE0egl*U4wE!Qc4vbzls
zTv;QA{GMX>*c2FZKUAE6Jza}lA4+bLetDd5a#qi>HuChJ@BirW7&hQTxE%0`{KIWZ
z%tuqIV#PqpJx$Q9WM|gxb|bV#D@6CSfx&xR4S)3Ha4fluaw&N0uNSMtB8p|GydeaX
z+tpoTwz;ytpzr)VqvI**NkEOHjBrDb)1CR?4!&qKVsEg>(30CGcvhBjE+3aM=B>8J
zEDOTvf!*T0hocPms@3~7H{n-4w%h6TB>3VGFLzqP)NTEKr+eQa>cv|GBFgAEmMl_+
z7|QH`v3dqZ3xC(0pQ^Nw40HIP*<B0o%Qr)9Ak)2xhV(4+=*o@gssTXr{G$&XOCAk4
zs%&hn{j1g{CMLXZj{D$4?Cme8sKlL|oKE0w`x{c<wL*N(i?Qr^03&9k5NHgz4)C^&
zcP($L^~;h3?hG$aA{T4kn3w#LGUH^5;-n}Z*2Z&MEu1OE=ePZcy=!0o`V}T4WBQ56
z(7PA~W#vSNadZ7f&lStA<?~>Cg1vN8j6~E;iN%2Ll5FfvEzyQsTc_tcM8zUP4Q8lb
z4y{7kaUz#W6)?EBby`~8QbANo1#zE)0ZGYbjNUawI|Ki^iR~Zr`oRVj@Ddi!<7WTS
z;cToNepjyO#9UI^l?HX^>(DUD*)(1oy?R5yd0^cpH!er%d*A?$)DQVgdW8$bhPD57
z7Z_~vUDChPGAZ1zRCQWP!TtR-O;;6((QEv4ZGA$vVwsaxP4SI<IYh}cPdEo|<}dXe
zGNWQSb<|<-Ie169gCPG!+TIRoGGl4q>a$$UbI4Zi28`rOSqKjzAU)Dlyu6w*JIlq)
zDfcPmiiTxz+I2fk+FsY*`(q>mn;)EPjg%>SDsA@IcuxP|(7<%S%RMM)oG1Q=Yrhpe
zyE@x11$Gz-2(K(Tw3llec;y$dY36JV=;2rBzcy3ek<)C%X??7R_W0QeD~Ibgg&(ei
zzC^K$pBU@TK+1LM@#x`v_F)=Nb^EmQx_45=TWUz#UO3=rH3o?*=w-OO?%O+9@8;1Z
zRjBpgrx}>o+0JpRl=6M$uHY{k^V_){Zmsx!$I1(Fei!zc#B#JUP=>XutPTm2ZtqQ>
zyYa%BZYr0@3qh#w*Y(Q~s85|+UuEyHsjS9a|DC2Hi<UF>A6jLY4sefB51!jScx8X7
zNntso(F}o|mF5O81+cEk=JpD1J&_Zl81dj*lwoQ@^>&_5zI`Sv*P(`gQ2Y*H&;8^+
z_)dPF7c5%;W>G8~u`i}Bz+!KMKHuyLBlAH{xtxQ)U{>%_`V$E#6l(h#cMIMdwBiBG
zPjPDh#X~!eQD3VrJo<cycZg)D(!|O+yx-VFMAI2LiXMX{RZQI>c2FY%B%&0;Z<D`(
zs=Xz6pu%XVm0p;$AGkkXkE97_&-D0aiuuQ>mR|$9lE0<c75Q(fIpPOWh<ICe1|eXi
z>$drN?XNUz`}_M>Op=>bpa3h6n0H%a;jo(xK1j|Ht3tJJl!;*+rlBu+fUd~<ZHn@n
zz!FD+)=oW9HwDPm^%8#=`;|r;btfClUrG`X^S~P<Jr$MN-u|$?X1<j+{f#a#%%!Mr
zkK2CHDJIRa`?mw^sF`Pd#d~2+r2aivr$+p{-<^VmiI8YC=BBLZI=IMq%&G@xH;$7_
z><XPMKM>5S_}1c`hlsjwD(_()!<eu&*=)U;mmhcBSOo>lTw3;D-BJTM>V4n-P&H{C
zNzB^QRmPIQM&mJKQVB^(d~)(~wR{DE@|eq3$Jr9ctn--H`eKk9icx<+ABjgG?+In*
zFj}bLv+*BEI<#bWbDBo$4~Fq;fxReo6-5>l49xCquGKz0IEOW3xSL^h@4w2ggP||g
z;p?ZEF>gIQG(LCDl`~hL(`YxneqCxwPiMfDc0Ed32fv`jzZiHN)m?MJB6i7C4?0Q-
zqnMrmBi4lUV~XBMu<G8@&+Up%_!VMM*8Oeyq_vUjYH|Oj9B{idgjB>8{qh3T|0tf0
z5u^-|lSF4{r-Hn^ljr=0$4`0OcBaz>^3~5qbnxvJ7y%cU0yx}<%V*D77c$@AvvmVL
z2rq{rZRT=L9X`|V!G-TK%t3`Mg$nzNf=$wT6u|plcn89v*jMfMU%6Gtet1?hdF*JM
z5;{S-?IYyhyzhIPi)iMnuzg*pys>Xr{OB(=27Dxn&|mxQzB7huiz#G`$zyIsz`%9u
zD<0g_DXzNJHdXgC4yAkx)0J*F@xMgVlLHu(@W(qI{J)Rss2x~mGgqRV&L`mRusxaY
zy;EGMv~*jh3Dr7{8V^qLBPYNbEapI}6k>bv5tS_#g_V;HMU>4Sl$dkIU4)XdL^}9)
zq(zujqB(Q3G()^v|I~%#i){nS>Utv%(R%Q)an5Bz?fc?2*;e})xBrA)Q~tB=TK`%<
zNWPz048<GnzDT39?M%rYmij`~0ADh#GZR4ZvOEbpJ5-M=q?kNxhg?X$y8m!}x?kU|
zw?6h8IYOs)E&B1of6ZUn_<-BLr^Vt30nA@PBRbEMzmF?!G+p$Dgx@|xG8|B$`pxe9
zYWs_I%E(DgOv|~LTC*Cunjl4JI1l13(y<*P#Qp6vBFbS*#_q5{MfcrZ$}+>xZWV`x
z(jvtuDexGx|3+p{O5i)pC1(v=`%}C&7n_$cP%=zmuJ%?=ieV9o`un<I|7+@>mhUqo
za5eqrIn@0}-L-Pn)Te%t*1jWTGvH6Fa+C=td3L>dbIDpb=GJm>0i$4R1YMC0H*)m$
z4B*PvI*h68ui)mkHn(fnUEAzro{Nrh-sKhKzbu>a_~E&ohF!_Wj;;xlX&Xh!(j)xm
ztdAb6BgLaxL@tTb{_BaVsjK7T<L5sRyu0!K?z+z7(BK`;`tWJL0e(esQ0YbFpe914
zo-{p2A3}&@XMU9v{lVvare%*aED~-WHT0)s>^<)bK)Rpo1e(3;Ocr%gq~L9EE$RJe
zt(RF`+lHbA`H1xypWZFlmm`J6t-rVEMO+69!`4$qbZd7Pw#xqq>{tgJZLVxs-+nEl
z2*!<epDjAbbr7}t0<P3+`CP!`wQ<3WSicUwF!KMEv7wCzTOZQ2fS>%nm1#jK+s?gk
zi&3UDwW9h*HXl9dVHM}nb3%3T|3@VU`l4nCd8oj5W~R2MW;eoXL|GK`b&{2Mr)XU?
z8~Z5A%Fc!6_|;*Ss1IHbAyJiQLmS|-R?wPQgkdNcRCmo6T33E8fno0A1PMRfp$E?V
z%AquU<vSbK;>C_Ylr*lM%pclxepdgrl%u?RA=jr}-ycgm8>j>UrWLQQ0~UO41#`D)
zArZ>?A#+x#Go7;P`>#KS)R+M<DVxWTCCA@KS^fE>y1E(#6LW5+RGZx071$+A!u3A<
zB^Kb2UgsF9&y;pK(-kxdJg~Jr*1QjswCZbowz)S>)YPxigM#+1d32j~%&e!4{$G37
z8P!zw?G-^329d5vvC)Enf=DkGL{y|mCxG<cq<3%xW~d4R(vc!ai?jskih$CKBtYmO
zp-K+~NO&hI<*hYu{onig*6@J^S@+)DbI#uTSN6%xWuk#($KkNNWN7QNN#m8O^SLR{
z=1H3FdeQr2u@q<R-iz>lr4pP>xO&!#t(ZoTarFaar3oI26mijD5p|jJ3)!(QmC|QL
zGrq|sWKo_wE=H_ZdqqDE)W`Pk_7_<9ULYM*dNNRI$$a&Re|-e_3YPo2aGPtiQ%U&p
zHxTGhbHcFLb+n*@cAV50F1&?wEI@wBH2Ms6g9w{`v05P7t?@)+^4zwiO2_f`Q*Yi@
z`tGFN6u;q;+f^76S*n7MXHO^$zkUL|KX-l|q04@DvnlmV_PcL2)t_Po1m<~$er!7S
zL|A`&e(7$l&Yf4N!LKe~T`r=&YJYeHjf1v;YC-O#WHC-IAMW1!6Vn9x7K{X0Z|xC~
z)3RyjiQ@TAxbgl7NkjC~fF#f<wRl-S_b^pjPoQSgs0eWZK}l3@y_9N!PP0^F<0sC)
z$N;o*(3suVe1z?1mNZ|8RTv!zq||Zhkh)*}w86dXW35cr(94&_ejAG%Sy|s#^e&^9
z4ZTfT<MP)_UbDZAXV(<a|F*Sdh}A+_^g{5smthdcs%u(g&vfK?FDyqbk3<jWXJ8i~
zvVVfGdaatKh*^wr|I7@VQaYu!af(5){}J<^MM^|h1wfHCyT_r_`wMp7fW&WSIr!73
zPtK!tN4E(zH8ppiK@)G-uDC>u5AqNbVon1|u6HYR^=Z;M;yk~v3k@||=KyMST|&??
z9ZSt_H}7OVR1u<uQ$N)**8eE_@!u?PU9u%dXmjxL$4!SHJEG!Fv8QT;+smnH3=OT|
zPUZn4x`XR=eAjqF=FeC=!@`t!@bBNnS5*mlTEDZfE-<l9{Eoz&ElK?40CCW<*Nos3
z#n*>zpQjd^+Ns?4&aA5D+xnS&dyxk)tP#w?Z~xcdlFWB!Sw8Uj^Xc)1P>MhP_|_c5
z%W(Si=o;1xEYf<YA_KGmHNcM+<QFT8T0pX#^k$*FDt(4sPEuz4QPy-9HGnbHxu&25
zVV9XDLt_N_h=(pg$*qmCz2uQn%801ma@{4TO(V(IPLeiu!?)rhScQ!AhO|pI^p55p
zIl{)JaOb8PeP^Yc{#e%pe8X!!&5U2tcC8W}9V2Wvd5q0mHkd}XPMGv=ghXawYU(v+
z^lz|Yrpk`l8w426D}PW2IQ&~V4x*2O|GhWvDW5L<q0>}r*8*Zhe~FG^>?Ke%ZPUzF
zdW2^9A7}i4<i5)Oke2+4`dCdGs9lRk%oznVY~uJEr4hqEh;*zD#KuGi)z$aOX<Iz<
zr+hs-$649wy*ZNyE5!uH0k7>AnN(5|^m%H;(1hB%ca|$u1NoFpuyd?B;#_yaZ@2KY
zt=yEskA44c>XV|s)l5iQK=%7nPyC5fxww2`-7|qLZ08XZzn;+4@Kn|tt6^?Dx#tg>
zs2Q^TH9A5P2=q*$vNucE{=$t~5`A4C<98Go1O6T>3-}doG{jL7VJ0_bi0^*2T<6aH
z0e3piS$_en4+i}$B8MCAR)5^McAK19wQWoQ<MD^YPZyl?bkB#-b{|BWg=-ti0Cmj1
zX2wd78;j-t4J>n#4z-a;K3lDM)FknaWN(tj!gjR()JNO1Fl+V5FV8VEO<rdqrqy0Z
zp(%jAgd<`j4jg#;VA-i$2fqSMZFbj_StPLvkS8kf0*?z>V`g5*>8p+?0ZM?{uEViP
ztm#E_Vu#X=+-8u?G2N7e%&!mh&wi*H7h?+}_p*~%{g^dg536*wY%*5rGuHd1w>V&@
zXB^S!^{}4Ya7>u*u1}p0^OCpvL~T!Z6Zr4}2Dx!tGM|L#CJ0U4jwE=MjY;eQ3Zjvh
z$T;`!{t3hRvk@;3{`Z%aCL=l->#9P5YKA;f-YZ4tK(wXRr99@blJLpHhg`dTsrPCc
z^d!7an4bZa7Z)|vCd90p<;2ySypv*mJxPP;zlmIUZ@+XpS*Pl==%-pu0WwfZ6j>I|
zR<bEo<Xh1on>}LX*P0!0+cDU%_MQj8@Zt%`?1}5$6p)y=m4@@(Qpf+T$?rmEd1Zf=
zWs8U<YW_3ypaR$Fk%C)I>=$m$Axm3hzIE2mR_X?_!9+q@yYc8mJ)~}jyM42Bqhx5S
zW4I^C`L|78`F#B1T+{c**=q&0wRmsE8p8eW%+F10`QTD4fvaJkL?i6E&_X|$d?V`F
zCR!kMo80haM_lXK0_u}x9Sv{Mh&-sf`^yGZH?gjA%O+u~Pc#`4DvB-EAJ;tNr3?Fx
zTfB`F5b;hju5>vvVZ7_!viXG2hq>L=6#UL9!g^<0o@xTl*D{yr@$m<3$PD@Z_>Pds
zaRNf}QswEzy~+g$1mhQ=6T*S#C!;<#h)d}aA%&_xQBn;V01)1RcAE|tc_nHVR%C!P
zQ|z_q?|#UW5c~ZqUyE=5Q)l-G@`(W4Xu|U7lEiNJm_l^eo&;(UlJ*=w^}G2ry;bVy
zvNoB)5Vi~Fr?yovE0X?w=Dd4nas7t)W`?kSTPjP=4_1h$QQzB+L(ww{kriY)M!Js~
zM*_T0m=sc=YEn|oq>@z(FR=%g5V?}F5s8m<crOJoz-Fr9x5)IwypVk`!n}6w8j5R0
zGU%uEsY$1$$#qZD7uR6n3Tkui;JfHNmp6UrN)ArBm#oI&bFcUw-w%ch_wWtJKk#Do
zqAMo`@e6?f)=Jeil(}1WQyx*p@Ys*5*Ub0rYiilL64HA?tNGiC_6J;_;{1y;$RGlr
z$kR8n5hTZEn{65cGt&Lua=wwtKt422a5tYM7Dp5JDjnGaAGoB~+Idg-4{lBWFPP!r
zq}Niq+goeb#AlpFe5`2blCx-`NcZS6K^C-qSkva9mRKYJw=X-s3spGBrvwuGRp-~j
z$ljbc!L)Qk66H+q-SGNSIq{SQ4?<;n-W3zpw~k?j`la0_ovq-y6Rgq^KV7`^6!!Wi
zmhRfPq-)2aBX+(C*J?D5g84isG3NpC15LgSe3sk6lF0oWSdyJuATNo=qX58i&o3bY
zf&D=}7yJGyp3%C1L~T1Y!^L1X?rA%`jfXN0#cGZj2Yz_IX=C8y#s)W)f^}VJ63yJ~
z!=#+q`ux8gB?U<j$q)UVmFO)B>eWCZwYr*I2fkLN*P>V*C=;OpxQAs<g2+Yv_#bz>
z&L0HvKgfz!$cqBVpuZW#dPh#NNdakeQ5Y@J4Zg~5vz(tfanq*Tqs%uo*TJJYHA$~p
z=b9~=xzZz!zW_;2ouk%Y{4FID`!=UqVWDkayXY*<P;X<Cz?{hR_M_y5qQbWRl9?iY
z825}en(2y?M&!qnceAZ~m%7|`Ei+3?uX}iyZb?QH2C+5muHHnswq1Gq!gzgTFtz2Z
z*J6&Z#qC~d>zs=9Sq!sZtj~Ih{^qvtM0bF&4}EQ6kvW--=RvR0zfKV-CO0~*`H=m3
zI_g1myIoFTd*qdZb!sWgcIVu-+%S>|iato~XaM=do6nf~RZ62WZ<tFwqJ14-nqz*f
zgK_{jz1)6^^;MNoV3B!QjgLBJ!4YW+>$&{k$8C5moSx^Gv@r=k#u-6f)FGKM3DeXu
z9{Q;%IM`{rS6%lsg*ahgbo+Xyl6uGOw5Aj1N9hA_t5X&ZyV6+exSE9<KqD9^ZA@qV
z=#We;P)y!l=&j&Gx(om84=GhWZ4RGEKZqv&3&3}KFv!XL_x|{ll>-WN5xKxuBomhe
zjUuXJiM*L1Fk#D2E8nv-1RD4ZqKaAB8_8EQo=?;R%<#0BW%SAqJ<I$q=;&GXE0px4
zLrZCtCr=L0)0-@6ud<@m;Ti>N7?ahg1x2`5W2w1x`HFYB!;(MzxqP3!`LFJgbA66}
zIX~-9>_E?R{k;wU`^P$OK~&*<^Low|MN`X;{3}vR4YigZcRuu!^uI8K&8~MDhofwl
z$0lc5DD0|-Y8H!ZX>w-`DjaUU$T+A2ryY!m{mIR#ZsEOnt&1irG^j1?%8<7nEO+$8
z_L~cvw<je|5!-uTpUAW=Fvg}#nMuU(Ereq<j!NcRMgg3?{k0<)!`L{xDhf&*!}J6>
zRd><5$KgW|>8+ekjO`sCE<zv0jJiu<Oi)QtvL4Bfja58d9(23SBQD;J4=*9rtw!_>
zTjDSbY*J6R_fTcAATbUV6~8Xo9q<UWaGXe)gUMz(Z_X8G$Foy16)0EAr<RYEDF{Vc
zw@zj>THETzqC{4oCHGhq9)u(4{D2U}Toeg0J7|kd8jNPA`l<$>Vx?zgwUx0T?BHgH
ztG#NJ#&aOh*R?Po24&r9dV*Q*^Ek^GHK;?8<~<WAz1Lu0%pLXe9nuYJCVq=HsSWP+
zW(MPppj3e#sREtv7-cNcorT`5qn)d%t_@NZFt{-=_)D9}L*L>#`a9D!y6)erGJ!sw
zdM-Y@qa`O@w|tb}INpm89E4L2TOIqfSldz?1*?DIHaV#sQ#<$L+)h1vXjecz2vEPL
zX%%icT2W(B)`v^rQ(_4m<jeSY5VRjo=0iWIN_nnz;#M*CeG5ok``(QaKUjs!+!W<i
z5VA@Yy1Ad0i7K>}tJC@;KUlH}I7_^!%62fd2)0V-@}0ID_RXf%*7xI8n!{AhICPq?
zP%WmfNN*9g2V`c2FClS3f$#d;lW~8&NI|s3zNVLR4HFjEhd<;=9_)t>562GD32YZX
zhkrkbv({^*CuccBE+;~L!){YUNpi~pKC0WWtG^)QJRXFD`(W&?&_j%{-_k`$VBv?y
zwIsIvMg;)4I6(A0P8PB5qikIkslL~Oa5$?3(lGHD2?;$D6PJNOQ*9lb^%X$1w>R4u
zW>dGX@dw^922#S7)F@i@n^7~Pm!DD4-CRRiF~p34imS=(wuB1j&G(MY6^lwSelUG(
zPstadP|1=nE{!(E(&bq1K6={p2QoTwgnS9>SQ5_-{$}Ob8<^RR*_&-p3qAdv5B%Vp
z&^pKm-`YC^2GA;Sb!?%?bFK|lB3$Yd9Fc%{vP=b|CNy1W(=!Rz^r^&IHg4)nEuQDQ
zbBMrw1<NOfJMQ4ImwiV9(i!e4Dyplgy}DsLFuC-ZT4v`xGG9A-PEJTO&<oP1wCbtM
zP-w9Tp4bhJrJWY@i;Hd0u4S3~)xPD<eO%e=(SAigzmH!pOnPVwBa^pYlpzR7<&j!w
z;;p7-=_sH)H}MhJMCB&lqN_N}z)(rHFj74yEjT)Q2olSrZhi>u9)0jM^bxjIj%YSd
zUpX7o5mE?#lRBj;DY>|-<!KhQG9nvm@B&oKYsFG*dJhz1{Q5LNAFHj;ZM9oj*#0y1
zddMu^fAxTGm&3qn{_wJ&IlD*vMvF)=?a7a;J=dMLN`-G;o|`B@41?!EXRz3{0PYtG
zjDqS|$b<MOPMfc^I}L%3FJE4%^X6X0o%UL(N&tHb3>IA^($m*QsLqA?SnuE&q?SCf
z_gQe_G!n!_;bi=a(2}ubzeRn^s)kU(=Dw;(hd_EL>b{rMmO9RLhAQWAkNk+=(3g~R
zZB6mCyQ1~GX3MqlY+}0*e7`1>@2-^Y`5jk0W1#%?vW`%>T?OPw(4q*I@|UNU%q00i
zaBn2w-S~-(8p@$72?rJEg<n8lpH^US@Z4-~UjL~TEEc)7KG&Nr7ixnR)C`<EGoI}S
zMzK3R4?u<kvedexr&AFOp#M$kNYU*PVN0Wtg5I%uh+adazzA+-=u)jWS1{Vgsl21<
zvxRO;EHFUCc@(np)V}iG@tnlEN==knO^9VIW(2FF{$k4Y%siHS_sCJq2^J~OJ2%S~
z-8L8LYj-?V5JRqsKDyhOiBhu?Vf}Dc6}?N_bg(2RR5@gSJpOsa?L}5@Vc0{pQ+MC9
z!h@&Y!m7PCnmal=YNSA6Dflyyx3!kVAT}Fz@m|&+5YrcT3{x(L=qX@&ZB`h)c0zul
zU@Rn7vY6uY?`Bk7-r98GsCJ1g(ODXN^u#_@`hG%f-E<sCNeaTIG1zdV;ymr%ZK@vH
zXC~7dpr}Z%oaL;Nl;A`RDjcoWTK_F6RGvKCXO1k`LjGYsew{9xKhHIE0gGvZ%jM+C
zEJF;w%s<};@X=+Ea~?aVkL;(RgMB~*tAtg5vFRy^v&_u7tP-x%N`70>MZ__YbzE20
z(MhPP$}GM9wWTHWS4aEpU*u9iXTwBbE!mG|bp8CcYZOInqwFUdC$UlJ2X-)zr(W@q
za#z`R-L^`R>aY}=Ik>&%hovR&1_dRw<zVRqv+&qoT?_I$uTXx{6xiz|TYy$QVmu@0
z?k~ObL<g?1_M=B>QE9uAuqDCHf|3~qeV1Rd=`%0ah&X@b!Fc{*tDA_xQWLu$_mwLT
z$xpF3xI63;@au%_MP4u-+t^$OkWN=F$(u?-CazIGWF&V@7KNPvATUQY)_KmJ+p+bG
z@kpgETx|k;f&KY~&nCy7iGYvc-s4)`^zfMm+}VB5-8w%sM0K;m1sC!%rTp>lUQ&hh
z#*p|dvPN|)QqM-la}2iLBr7^vC(2#uahhgWxu~eJIB>L|-6tRP_y#Zr7U(lxRS7a;
z%0m)rYoFgMrXHdt+%G_xATydc#G*B6H<z2(8E9$K!H50CZLr{3%VqPb{)~&#wl<6O
z<HS~ymz*eCGanwJC5d>eKWbQTy3zfK<f|p`Y3yRMPH_`6OdPe*lb%8a<+&VZp^MVU
zN#ix|U`G9+SFRi*;d1#yG*mjON@2ydQuI9%cK>d^Zj}~z9nz&W2Vp}0+*^U<II+n6
z_`b}dZJ}-zAKf#ed{q`C0ZsInMONwJowlvMje6hSTs^eAhxNmd^a`yB^?pGFH;1(9
z@aMd!4+_vj`$G7DE^`V+-{U&4uvTa{XY*idn-h@|g9U50h?$jWy3ESlGiU_0W#RFS
z*p|6hHu4~F??azpzOGuW3^Q)K7z0#$lNG`dX+y$g)`6^%_6bAzSbbA&!JmxBe&sp$
z1x|@(HK!c=-Ih0v5jXcu+Oa6auFsiAx=?d)u0*`}Q?^!<+G^^SwV+38HvYQp{IQj&
zxbCo0U-Ud&8d3dADMY4+QKQ{#9-EZB?N9&4>Nfbssg__CDI*Xep3|SIGok)VrZFCZ
zYH+J|xBN}Y6YGpL_aW{miV^H|e$A<vFj-Dl6O&mO{q3Ehi_iTS4pdE7DceyIMt4s(
z^|2K%FU0e}8HAR#T(~$LXU<mV8=TZDJaMw(X!-_nUaHy@*^S>(XTQ;EI-nm#rFgK7
zqI^E60XS=45SJ6{%7`aaZwz0i@)Pq)im>XvZ;kY}KB1?i*dyhFnBE9Qg<%roXkll>
z4=B`$!F&*ZBD{@+gV*@^I;%XKY=<g(9oe7h@NUh?Svf4~!2<;V$2mer<T#wK$t9Es
zxsB1FGNg*UDaq61eQT?0thv~o)jO4)>fFNW_Ii}RVrVw(j(qy1)#krbquL9#GhVwc
z2pSFbmq66vHol)Sn-)rb@7+5=<y|Mgjl2$LKEv*0o9f2;{yk6@@>_QdeNE`NWxb~d
zKSam!=_V|wU6POJetpgUEKC6UGT%HXvnek@I>aE34JOI|w8!vZ;z=p^fyMH1L6r7@
zOp%fK5^&}~7AeKnibz6Gu>;{R?6Pk?rySNu2=i|`=}he4sTH*@*DTI=-u%j|;XYWN
zRu6>L2a?|7r_?%M@uX0`kni|y#w`&!a1yTr)^z4R6nkS3AM31(%|sMk%C$Dx&26_8
z(j<oO?|`o(BLf-mG4XEnLF!X*y*F+;r49KEx{Zn}E3wQjvCPh|s;<Z^OaV22;<dqf
zGmA{-<FADLZYOts`X!_$pYu8?fQ#07knoL9sNcBS=|OdMIp)`n1zS$L3yQao9F3$c
zCxkkCL^_*BI*UCjmfYBeuG9#ftN*2o1~S~I54<+uMFHgpRpEYmYVTe_(9yjXbxa>V
zvmHS|S>>3#`?W*Hf#xVEoxU04za)>|q+Rzza8BA12?H9T`RDf*O23ODQuV(zjmQG)
z5CFP3rgGQ25OpzKk%6)+u$a|ndNp3o2lH~|s5w!z{NYk~pgW&X1Q__?)!l|%;&=vb
z9qLdX6pzDyr8xkUV`z^&to1-19E|24|9W^H(3g{TB#Qs}&_i#=O>D^pu1;9|<24T*
z_}{KdM{Kq}OO&PipAPuPt5I-)bs5_cW&Te?IegVRQZO0uqnw?j|9H*AhaQ@ufKPz^
zXo|{20Q&U5KksS)m<-PHlh=>_A5N0}cexJ5#=pyTINbitu0#3xZ+0E7D*u0r;Vye!
X-(8V)^+?Yn;7393(VfCuPoDi3bPXiV

literal 0
HcmV?d00001

diff --git a/3rdparty/composable_kernel/doc/image/ck_layer.png b/3rdparty/composable_kernel/doc/image/ck_layer.png
new file mode 100644
index 0000000000000000000000000000000000000000..117a1b3a0ef890a0a2db9bf23f14f03278d1d965
GIT binary patch
literal 549343
zcmeFYdHn2Tc{fZEE!0+Rwbd3CZC{{Pg)EcJ4U(BjGFfIOlXW64S!XgyCYdCY$pn#}
z+U03|THNaf_*8*f>QcppoyS&f#jT*FDhRF1;mD>1Jj(js0sFoWe$LbO)BfAf=bV}R
zvfR1v>%Ok<^}Uw)!|9m&g`fMipF8TPqkdt;4`xRl_2icvb<`7HaqJVond4sbHyiNv
zGv$n9j{4vipLyu0qh9(u)zGTaCXABcs1wotqemwq&@wBl6Vbtm2r^6puPU-NI0Wa@
z%#Xq-@E^T~KnU_e7<wUuY7lZFO2aVtfO@AOByK-?zq1I|KN$!+1p)*7rWvjMY*&KA
zu?9Y12%JV4a0D*ky+;oly!SkChR*XOFauYNsuK|wLy1#R9Gw15k=JDDM5GUnlPC_r
z7Z-SO);;1E&hC^b1Ggj4Dd;0_>@SNf1vjxUm>F~m2E)iHFx&$-|Nl$t?uJi6!8z3l
zog(@_Ue~QUsGP;4E5_V<xfOOVL>yb@q{W=tYW$-st%IVBfS4W;+P}WMd$`)?!J}vW
zppHB+lkRK|BvU?m8oU+=ZwSQiN6w;fQgAu?iG;mfplTM6Md*0w*>Y9B(D=#w)(%+I
zBQNvfS9w{QXKNs8AlGgz!p>a}D|_q+NR}PU!N)`$?`0lEgzbq4gQTdpPq#=M+B1eT
zp?(=NFcG3qG91kN>v1$D?7mHzN{UY+&X@3&x1TM|@i-=X+)|q6GTrQ4VX;7#E4|g2
zsgtMc(X?U8#MI)+#)$WNw6OV_p~0);skmnqXsHXH*fTIe<JO#tL!zHYYh=5hF)L~q
z*K8%siPjjX)J~mF7nvKEq3ZQl`ZmMnV6WM3CH6Qmrn<cFazST}c|RPDm{84yLf8(>
zS`;T%q7$OsOI8zWrPsELVhg|Zqn7i?RNjQs46+6UkU%hQ$?0tO=&<$JQY8n4tY~2-
z=_(L)u3&M~#Y=m~#k*uF=X#WkmlmyNgSKBLdNy3L(|K)<#=RLBV@QIJ1+&=mVh#H_
z?e}R8O44DBD1tRNvQ!5f!yss|6^N<J)rP2U%vG|cWPc*z_J(45ZJhQeshSy3KASpv
zW6QIi#16>BAQ?u*H7O1z3&sO)W0tv|<@-AAAwg<r{A6fUoyJxzG-*p^NheD&!ChCk
z*t~@|teMPKt{^t+zC2;q1<@;Y(yM%rC<7vjXQP&5<QP1!@wQ<kJnOJdZVbVD^q%2B
zxLG2cxLq2`ZbCo?jSU@yYlh<r;j5Vv^aNHYtEeE=g#*UOvq;gR7_zougDxq5?_0u3
zo^LhAmop#SK>4axmMOD}nb~ZmD~y9U^F>BxmIiINsz!KWjd|^^yVn*LODSCzTrSM%
zfrN=8rti(0)izbKp=Iv54dqx1G~OHI;e;nf`iS>qcv<?sEbSHB+TgyNK$AJ5SZiT~
z6<|nNW*61UGv`#w1w<80ak-gCR)4WF$x@1Zs_(%0vZ?vbn2O$Jw=oNIoCHzs%bl`h
zz8~?-4tE!&8*CiLW*~ZyAbg=U`!XD28IshrBzj|xcNPPkw&u*Ft!fRS2Q*L75Ez3Y
zHgss}S*uT$oj`~^>~b_6&@|7_>p4#&Y3T406;>QfP0~Gh!ow2Dnh&$Q4@KfuRx4&C
z?}wgl?z{1NESAM`<&L^zhD~7b?MC2}Ma%6J=M#K5!&l%>aK684e2W-+s8TA-RziJ2
zZxEp;s;$Q^?Zli;k#y|_zM>44J(Nbcq}Z;eeIABIh@6;tLCSQx?%s+Rv@3B+wA!Tp
zoYE!$bD60+n`>)wmDR{>2q!Bo7#p)*qk~IM?95TQhnib*z1lIo4PPfi2iqWGKzA0T
zmgH{bS36gAcB~zt1+nQc8^9vMFiy6gngI9Mt%EOd80KMMA+3sqskvILfkRTnkl~k|
z5iG484vgQe24(|%7Q-iDZ!w>f%4|zn1JWMYeBo;{&Kvn;-9+%(oSL~<_01*lP7J(X
zQK_(uri%tsR`a@MN5EV9kl(IWv5{=*?PjG43ES2&J}g=-3DqLZG*!^>6|8Y0-R)>n
zi=^38+yIhHLXt<$gqU}-N90*+!mA20dj?^SsMug6Q<7(N*vgJ4kl}U%p0<3*8PlXT
z0mpm9_l>Zw$l9?|i&!LbJx6DPMJ95&bc6<<gbK$(>?TWavd@b|?5-wLn;!UqfDkr@
z0|(A1P)k&Fh7TolH)W`LO>%ml9UBA2bC`Y)Hqmqn-W8fmk8UMN=D;)kF|{K7F*sWJ
zG`o^0t2chci7ddK1Y|bg3S7|g#IBN|85Pu^-?RlPigSsbHUwRc`VqN7XMCe;jmaso
zKv!#S$?;3s=OPDeD!Wg-m|^uXBxLh;f~|)GV4uvoE-Pg{i-=;+Z&m|rwciC4Mi0F0
zga8s<<-W}rGjlP=>cu`vk+3EyYhurNPoh>yoR3!}sjskr_s}sGvTi9(v~ioMqmkhV
zrQc~tQCUs6o~#hq#KL8QZDyfvb74L+s{lbPScY)8AqhQP1_UdOu_~P*;K0-8Ub`Ky
zDmHCH;HWx}g!-iAD$<N2fm>WN8=fYmof*#pF#CN#5%Oq?Pjy_Xnt~B36C%6-$1T69
zAhD8(Bp({Ux>*>I9IFKfzBBG@w~;%)j&?8jS3>ZAVFz2QXSLpR>q(ZhSSO=A>Vt6*
zz#ksN4iaY%$jMKbfzi5dRrZEc#GxnY0N8`xScKe~NYk1)&5oA{ibPU5PsA(_q9F?S
z0vpDuu+2r0@Me3c$omU@v5IHXdd<X>NiYdArvikEwRs?V(OjwuzJT!FG@e>E43N$4
zEIi><P2IW5%;;^?wd`QAp!5SG4<SY3RoI8@e!NmVc8+n2J#O!sTpYH;*<cC$q(-7V
zpRDw7QmSn)1{4eoxZDeXK{A>?Y4)PNYVmcv-bKFD%K!-t!^q;B65vk1+3#~K8=>}o
zDU$enFN0?zYGw+_G!yGUtAr(tc1X&u1coEVqfSf0&L)rsozcW11cys0?38{4crDx4
zdJ`XP4>O&xMNS9XDQj1Kx+iFBxiw(Z7;Sx}vF9lnxBV@%j_iJkf!h-)Z*o&v_Tng@
z*K1r6r&$!6vJFGah;6OyzURV0kF-Ytt|Kc!_h5UpQWAv+s?ln~AX@W78!22Y7mOL8
zFyk3;F-XWJsse`Jn}O(Y8Ybp=Ozh*m+h%DyVJM&G=AsgtD|6T<q%Ay%v)G<-3cHjU
zVB+j(hN=U4mKWQ!94IWt5Mm%|4qn^y(utvL2`~0RLcrEYF11xC2dO++%@?>wkKkHc
z0FHqq5fRIp2PQGI<xVmzLwX)Rv)cY_xAItVzC(u+N%s2_ceWg?>j64|HrUcbrFgK*
z2TTo=mR*k)t8_P4FnmYt+xZ0JM~N>$i$S*mXpkr!ua``Rfcj{vwRpIOgfX=jJc6?2
zY3gNO9_CD%aVejLTQ5g_QQD|BI)y28qW1IMC|c+OW_61m3PAj<wbFbI?)GLjlQVVQ
zIH3>+c9;+zV(NvNl_@2Zo|@#eIdZMhXhikba*+|S6h>S;w`&zG=TU3SM75kSM2a&b
zJ`1841T0J-uXa*07Zx}h_m?i-!+3wMI&lPqCAo$&6vaISnCq^M=kWqELe<^N{>BG;
z$tJ_K)*nKY>{O};Tnpojp+SEKi3>Do6mO;5i)sXO7QQZNE$_sltocQaq*|}@V;Qe(
zF-zs7Nz=H)tx<0lEwUyN2Q5t$a)hKA?p(AqbX$bC#fm^>Bt)CB+mKpnst)26)3qon
zj*pEt`G_*LD4jD&v5FF97R(0bmh(e@A%qOfOvK%?CCjKoZ*e*(_K(C`_HvET@TR3E
z!+pFckfk%S6$*u<k#0Ae9a65gb3DOZv06_2^{jArTUN^5LKG1}98omU5mJ`%0j;$7
z)Y41NHt_%#^EBJ<EMBLuveYJ%VQfvzLS~Yxz@)MoR1>K*Hg)Ir3??3x2BNw`7_-21
zfiDOKo5~5x0>QVn(k#4HWkB(;)oiDuu7Y4z*=(YvoSx0B)xbza-t>r}O6TK}m=Wm+
z+fv;)++@GZz!8&xzfvr4$R)x<Gor=kI$bjG9IO^CQp%2dEgmg6aUqTg-4!ShI)W#l
zdN2BV>c~cDqo!)uo2?#>)`S3++yKsjM__QXVOAj2W#?HM!xL@Q21q5FX^eP@fe;%M
z;x&s<S;CeCe@x}${x*pz!JjzYl<-mRW6rQ&<`t@BzKs?&ni>Y$43=CwXlt-j9UW1;
z#kTBNhHdkGZWoN4D438r$GCLTsbILucxZ*UR@>@ov`8npx#=bsVga=pZNL%M3xK;}
zyrLAz60LzWXN_!NWV*J^NY6p&#|%Si0eCjpO{SUz0JWK|QVy9!D<cqtJHi6so(7n2
z*(w664iQa|8pWR>2Ao#3q2^_7M+*nfb$RFe@k%#PRB`<f^Vi)J`8nRN3o3w9DsYif
z_j`siu}se2d;6go5|l7dHyWES_8PZm0dxa+B9sB)49uZ4P0~eX3zX16a+uoGLV-K#
z)6<b!k5=Qr>6C_dNgu@HOb`~vQlG9ym<J2GFQJ*50zbkCGGB-sYXUhkjUIPVp4GPq
zWnoCzjIh}a*bpX~?kl+|ixqFVjSb_(N+8>TK&G~gS$JPgR5RwfbOxJl07XbEf9dHf
z)Hli6TIIba-^_(f8go=-08?i?bel#@h4jLmzr<*zW8+X2u?((A#g^J4D{r~);wNcG
zl(?B`l;Uhuw=|N#0B2==H`=t-a!n;0Jwo!OhT9kvj=(5n3N;)xcPNtXO%^^W`r@$4
z_tM%lX|>od{7Ksa3S16@t>X8`t!Nl?o#OT+&<2+85MblS2IRq7=VMF2Q5U{MOO#+Q
zXxHs8+(78%__W7w#&A8_^>DObDMN3oiTz>*S5s8YNEyu@Vg9?;u%{khlwOCWrsXZd
zRn<)V44O0%zRC&0rskr$X#J2MH_dqK^;Y!|LK<EiW0O`2ri7;EBWXx&;kd*QVvkr`
z*4IRYlA5LGq4m%x^l2vP9)JthL@kLir1gxbLEH{u+Qi!O<bGLr39;RG9%HQV%T}1#
zl*SkI0OBDEmYhjP3XsqZK~PKo#ISe_1Y3*=2I>cpUo#Ve8jh(Mu?!%4T=o&XAjD#z
zhPsK;(^j1c@thwm7)^b|VDgo-nomcvi_j&SJ9@Qc3{<H_y4$-1LOojJ)0MHDi%S&F
zp{?zDhSJ|)JzU-HH|e6w+-VlX2h)h%xoqc?2MVg-A`a(Ou2FJ3f>UQkdM!4y>Qu+%
zY$h$6)oRu+$hFD>O2txqr`zl!8mS~6aKB6hQR!%)qJ6gR0}^cMg}<)107uxFM@tDB
z+T~_9-W0JUu1XmzM}k47A*vZZgBuC8m(yt@$5f{m`;};fPK^!5VZisSo(dR|;wN&m
zZT9_qWG=0FI8&GW3?<vdO4rNvcI`@qG!I*R4}wCbH8K#>7&1!-o0&0zhZAcGPJ^K*
zbPYc9dRPs*E!#apr2k`FS{;lSZ|6adL`@~yGH`JY-~f3Hmrj#cCC=I}JqM)p9!<0v
zu<KLBsH}Uiwcdsdq2bwfy8`IMw8J4b?!{{Wooe@|E%e7iM$bpq-nIyEhVb?21`r?)
zbtO9B?1WQ>eIJGW3MX(p(~!JJhVwNEZslWjHRpvsG+8xEsR!0Jn$IF7Cb~mp1tl!X
ztW!7vO2QD4C(J?<Mtf}Cb1}xBg=^hjWdcjg4S^jZs>kBsvB{`OOQb)CMFR6?I|^ng
zgoDt{8bor$PRe&hu$*FCC(&g(iPtOwyOx621wWF2W_7Gd)<Z&u`&6)IsrhOO&nUK!
z&17ogbS`+3MH=CT#eA^0Not<MdmlsVt!tB6&F<7DCo^}7?#Z;x1+Iq&K>t~$1(IbQ
ztqxlhDuuO=Rd^^(J3lq)HW}F)lZN3rXgz}Z;nIhcilkO@m%w5MF51Q4t<CIh*iR8p
z>xbpg?40|mZrKSBRUyHQdfTP7HHlI}d+ScM*CatKroCv^E7(>9H}@7(MN`Ic9SOBM
z11G9&00hR&Gc%P$X>3t*RavZ}1)|Oo*o$X$Zf)E3URo4WgoA_0-U_GaqSBhSu3K^l
zUc$COlUR!&dw#gWJQIv&rA)BytS!N$%H-)hcK4HHk~=d3B)3_-xAL29l?6p%4!cRv
zlRTe5s5P9)Lmcg>u!4edSPBbL@Fmo1@L6L-2?Y!?&SsE3B{dkF8_b+RvaGW8d<g!U
z`)CCc#8N;AEF^?RX!e&(4GfuKs`0Se_esDd<n4}!kZ2-7t`LxZ5tmEc$GI3<CyD0n
zr$HN2c;nSjH$E%!j55cT{dy|~%qZBfoi{9iqR)22MObf*P^x=GFzz&IWl?0!`%|a|
zdjxKCMR6$rv!#7!KT^n{D*I#0>~t%LbaW77;mF_6euh*O&oPu7&Dv_LnEM)Ab2DQ~
zG4p)faYeh|TVNr)qs&hA*d3*JQzJI*Fx;Xo3XZ34xdFq@Q8yf0+hJiEss&ERILY}l
zaRcD19(RI|R5B$vUfDa#1=|hM>~&4vhPyFOMO=mfRA9L&y*7NSSZrd~nHVZ8kJfBN
zFKRXGDgEsVH>93`nG0<qU{isE%yv&r+@TgXQN)hWk(0PgWP+rOE#b@MUIuOv`1}qf
zx`mHa$_&ZM^iZ`;SwqhH8bZ{<iYO@vyu5>sUDG$`T~4p(P@Ss>Sv(b*@<vJNExlle
zqWK7TS4c{;OGnwvSzYsX9kQh2HXK?#n=Vo-R=`kD0-$c4YJ7vb-SBIQ*)eS6%shR#
zn8FPJ)#+-NW(>QJ28+(X!sT#X=KE^Dq-C0+%H=Xz6RAo^7Ke<?36+gk^U)Y2T_j?<
zpeA`g(S~!z?QBghC;2oU7X4^hid8n)1Kp?1LgSaJ2q?A=ZlS_Z0`utLJhZmbJhI2|
zpyVLfavHIviP4y8rogSS%%*@S4Hpo}mKQ|f=H_moRC7~ac3F&_PHt1WR^+h+;`vUl
zn_PnX4!_1QjlzWhiHCAAg}Z=$ms|6E%Q?;zWW_Xjn!7v$?U*VWbB4x>LgJ!oAGcj0
z3&jFZqKE<sYyitF)b|TR;G$sDlJmr&X_`^6Wn$T6xiBrOv<HkcWF~c@icBZ~&O%Tj
zMoiJLB#3${?`vY`QG3db$ti>l(oxWZkQL*Fnk*EP#eNoNJ`@eYB!~gFa!yak;&7y-
z;hNqGW<L(t;f!gOz6}rLK>-X@FH?O{&@o>{1YeVbN6IhMIvo=^wt++xa!1){gp~rc
zRRxlUmVONR^E_=9)lg?xB&SzJAdUtG?lYyH7KDz&+ik67lTBXv4zGBbk>;51(2HWS
z-mK(NkL%mZW@<U3q}bz>Zp4IPS3!T_2e8Q$d)HV}14ro@fr9lY-zUaBl(NSC&A<-p
zDcXfm5vV4Bua?bf9gk-OTkcjsiSra+vrE@NFx;E6312q40CaqWgF;kdbM!<T;=|nA
zHO!o+I_A#OLSBXDbhKM;JQazVh*o0OW-^gRn1#krNFP9&SpVdu3!fdBOVS*xz*-4=
z*9XW9b^+`vs?XYOFWsq0t%HigCi54u)7TBR2@`<c9J(Q=;xtbZhr*gYs3G*EU15XJ
zxK9QNQgI!a0yf$BzyoQzrmbCY0<5$<zW`rTZVidOWl{%LvZ9m$ND<7Vv72a}H*m_n
zHI!S-Vv9%=2BXS1_GUDi&I{oFK!&a7qtJ3X15wsWSnoD#5{h>NUzqmY1YTjIW(XUO
zz7i;Jw*WLZlIS>Q$VoGsx%w!gVKj8t5u}dmojO_T6Ajeh>Qt$1n^6Xn8ikV<9u8R7
zcgJQ+f~+!X?4*Gs0W2r>>WT%$T;R*L{oL-YfapC5Hfed&&b{6)0^vE{Thd%R-KZ?9
zGyJruVNVU)xilNjMtxz>&Pq?CN18EDo!zvf2;0JdagHGv9Gfrt^F<+mQ4@C$(R7sI
zmPwEB!N#o@?H0VZKZ`W1(-?3}@vF(OLwajovaC(=NMu%PbftR$1Ut8AXrN|cxW>3P
zhx%Z$Yqlbv0aht57-*`;&B)Sevn=w-5+yp6XH%IB0V`iG0ls!SC5O$12Ofgqg4Qw$
zzUif_9#M->!B4;onBEq4r(w0q5K%+d91ot-(@JTW^)liWE8`GEU(On<5LcV5XWKXz
z)FCn~_Q7)RMOL*7TwzliNsRWm1t?X;tqkch47ejB<%$*#5R6##;gpt1HFl^ay(?i5
zCk8B|u6bAp$$0~2u^*7+5O(VoK4YQIHPu?L&xs+F_U8?$x67~?mu1i=;U!0Ia~Phj
zHWIEP0>*%ZxLXc~W03rFF~!q~o~MN*=Z;`1g6ep>ZF=Ogj61bk5(8UiNvd9~Yka~)
z0sv|@1pXYj({@IY><sf5YdA7WrdV^W(@qes0tK{*wCpkjdAotP$*PwV#Z2FNFdEH$
zhW1qhTnia59=J9qi&>+|$ztQXwwo=JUOMQnry7w>ZFpo3;x$SeOEcaEF|w9WdpK<+
zqH|Xhv<T@0RbpvONUYPZTyC3iqA`#<x>2&o$Ph)i;EKdxiPY7rj8c%Ly0lAgCMFV2
zj2cCsQ=3^7&5A(ar%jK_2Ud!YMoX9M(diO061C~$Ibsid5SjTtugJhZhETjG^GIF<
zImn!tAQf)Oqby<P{3e6Ob9*fU4>@XePN>LMr!3eVgMtgUDcq6Plc{Vpk<G|l0^gxc
z;Rq^=aT|Ckp@2kB%=cBH6p$eZC?*>miTOC`4>~uNQG<r}9T3m)p3_!H2zhRub;d!`
zBH^}9=Q+tvuK1|%b9_Fk5|CQQusJzIlz`d9NT^tN$8tAmB*Ig~s%^3E2%A^wxXr>5
z3tI>t2vs$Nq+z&2RXP}%zEN#LyrSnOvyS1FJG5eb#K4<^qq(#n^{g@~on1NBhV@YL
zO3dV(*>tc&U91O1alA*1;W##0d$r4JY{!eX4HZF0zfe$`AqI<u?g2xqHfH7}@O%iF
zxo2n-7C2%<aAaQ{+HN7+=8j2;sj=!KvCvms+;B580{R4IPt3L{=OUuhf|A%`m1va{
zPxQ9e$J>ce54O~VFEy{^Xlq2zhg93;Will)A4&dl)3_+WWCT7!Xm!8>`a7Mn9@Flo
zEM0cOYAJtfk5^5OBx*kn*BCnj3C*Y~#<;NLw?u7spu;%*6;0S{591YGLcB8EDYKPq
z=d}y!4=5IioDz<iDj{p5HN(YVK8<%Ygh7nfj|~OJfDZ0v+k_>i{MwJkML(xX0%k@o
zhhxM>jMyL^vM^&)TYt(=DBPYa?V!=<eZFnN1u^MJUu#Y?gxi@s;~C|U#3&Mry0Cl8
z8Hy>8;+*f#*6OkkTYy9{Oc|m5xj`5kwcPYUevhS6fGRAA^lQ05gm|Mmd=@1D{f3qY
z<x0Ui!0+HyC-jNBt|qfeS(?C&_>)u<X>%B4mf`a=T9*ckG~W&&e3{R|xWoM>wj2q-
z!^U!D6<w=+Ze6?MbhN{Bl*}vBfo(5mg|wP3c`pV%N#sVHn|v%nASW2K<iL)-zL5=w
z`&CGjzR3}cz2!1}+|;zQNvHj$#57vWC7nz+Jkrhv@P-=-eAW}ng+9wLX4cM%T-<;T
zhGJ#SdMdPF`oTbQ0CUr_6!e~~7uFzySi)EbUNUpZ2ohI**+r0^XyIc-6s?$VAvtXa
z+#WO%Br@CzCD!MGB`;Pm$~wM=(QZXd=4(FWCt-ic?9y_yF66qBHWPD>3}$1vvA2@Y
zbMTR~Cx)I?n-K(oqInJq7DggbIgdd|o|?Slr%TjiHzb_TK|=!}_+1KhRE6uZW7Fs(
zwSw8mcc*1L$fMHLcuZy(lip~HH642k9LsW#fJaWxB{uHD)uP-JQ4_L)jStL)6iZ>I
z`SD=i#ZcB(i2Bt6rshmZfoPH;xE=V^nS!Fp?tR`Ae6kiu<R&TEtUJ~gdUldj4Q2DP
zt@qp@)t!{vO_a{IOKtBmL?8E12V&daeA|&LnB%lX#yi>oUd`8*!X?Zu+LJIaAJ!oz
zCMEWyi0Y@kULO(_1yeRXH77ckKd~&a=O-5EA>pfRQ8n{^)o5;u%m+O~0Kr5=;^<tX
ziJ95WOv}+`?t?<3pvx&*Sz~Tgn);5wP!Qq(^Jj~sV;~(jtE6Qnv!E(~3GyPHF4oAr
z%R+#74DgH`o|`5}P;Wt*%@tP82%syQ&Uk*AfykNJ_m}&Drdw9bs;r=tx&xb&#1~6?
ztprp!Sb5gO0*rl14?Bb)c}Z5|N}O4?tnPiFTs;zu=S?d+EqZCGD9S?7Fqdm0OpzDR
zW|9-NLo`TLa5K<oGv8H((@-YJ0<D2nG5rKJSH@I~b9@H)0K2A%2m+I5zy>f?ml&8X
z_+p|37;H4mMla<g@dpws?BW^Hu+A8R_7ln@U3?hJSdRE}!$^v9;wy6%ua^Zen8H)m
zlE>q9e<vZ*c<lw=YG=z)?hH4~Oc_J{G&Wmxi}6dh9YHI=A<2ni*oiUs*34K&a?s@?
z&}OGfF&7TUAaw!yP*&4v*oqTx5C~LNMFY~^18&J|_L4Q2=YXYDnhaXcKmkPD5Xmrw
z+|5WC<t;u~4rjB>N^;N(lm#VdToD^-<dxQ3j~6jCG>}l6%z8c!u}D2F0pVt17yw<6
zHrxJQz?xKXBY`v+s5g#`xuljrNc0e#h4HiwrEP50)6Jg5Yh%tun-#$)jRnJ+LB%3J
z%Yzjx;U+)u)T$|oiZYDafRzD?;rzy_Mlw8sH~7w3p=dbSwL_v1da)-$akV5Hf4}!d
zjIZY!C%|so;~co9;X+*QI4@HIDZ_<M1~aCa_0a(vZ>?P2vu@N@tiJ}qN>jBUcAsxc
zaTRHf<RxQQ?}ejSP)9x4swJDpGlN4wxhm_$iAs2c+Bco8J2%$0-4QVtjvc;dt5TK5
zE*Gm|=#CcC;f^yV9k4)*B1{+46=divJT3a2`iw+%R+?K}pK^hw>||{`0EM|Nj@Mo`
z=1tIKXcrU#`wEvA1D&wyzACXhr!G{`gfIy~|3js7ap!@Ul~n`4EU3`{M|ojyp@>!B
zJ<=ib(AW*4$i-`ca^dmZ9CmwROtwp|sR7v8PB`kO-<x}zK2geO%B@?=751R9EZ1<J
z5~t*LD)x|U7s#Y*M1i1T;c6XsX#!BNVd)6R)X;VK0kefgcbzV44ioJGIn^f}w6vgM
zXtGoRN29q2;?zj4qrzQz&4`_)${gfflC?5SwH>!HxoiN3Ie<qN<ZTLnhK-Z)dhQ7B
zFl?(4rwR&~DW9~voxEJsFa!!N(=9BJ;{~+u&21^K3$2SUG7<_0n||QIOF+e#sj`yU
ziMDP0jacnr-}2Xho(_0%M?1S^vYxt)H^e*$GW^Ex{P~Ql;8bBq4&Vo)j-Ue2)P>@%
z*8K@avjE}gp1agF9n>?GMT{nDWJEh%$2o3e(IAsc*EO;xz_V02C`XJf_NjDSZnn!N
z?f4|UhQ=*XL>{vpg+@e^D)9I$CtJcq+DLW{dX0F)Fy+v2_PteG?Tvn|)uUk=kd*}|
zP@;oHvffq>BXdDWRj<mN4OY=(QsP*oH*RJH-zBLSZSHURxGuF^$fHqUO%yYq@;nN%
zNDXe6^>zvyL9&*%l1;Shc{MdbU*#4LRc^dAc@RcceNd*3dmu5)@g+H2aJx=#qnMCH
z!L_U|V%^#HuJ!GSomDDEMm{<Wr&I%XwAaPQP&x}4igWjg1Si8f0)YxlO|f1Xs4eLQ
zN!??$e7iv%SMStVX!nb~)CYZ)0Z&V_aSOt#0(1*iB>}3dsOeEl&Q>};**Mu`f)SvB
zf~E6d%o~2zy>j9vAdSW(MlkN@81I5kl!PLJq8aQ7O!P;%Z3OCmW{Jd#%%>jck!rhr
zlBc1xqZC_1$dy>sohzOUyCA0kb^Fa~GhXhMiQ{HTBFp1}G3#6y$loXpJQ_rPkp{78
z^(SgD$-E^UhbgeFog=kC-E$CHurpCua>&f-Y4Fs@ds9^O8qLgIR$S12-Z@Ob2D_A_
zYKyw9l6Ky7<^}#W+(cPZH6R%j8L1rRAa&H6??b3%l{um#8<2-sM<DaiEgsxo&&JbD
zT8sl5MnPtq*#%>tvn^U0CZp+uC_$`-fZIy1D~J{SGDT_^1wD_;t<UH~M#m?HDIr?o
zA!*Wt`DWj2?KS6Bd%2XMHQDS8-dt=}ld4twjNQ2#SLe3lF=*c#*3f7kxsU+hj)fN5
zXi|1wb14Wriro0Uf{bU=KGTzAPu)7Ze39)}h?*~a!P(R-ui?W@AFUuSh($w+^|9$~
zLRT3iE)1-LNfLZTAsY<y#%w0W+Co>Y{=UmF;^4p7Qhh-&_Dc^$1Z>i6&uk)#0qZc7
z3akm)u9GNAb3WXHZo3fm(k2cAgP#uvv_~%xVo>m1f|{d2)uG%%_KY+jc3FWkJ~J&S
zi3~w)KZOA%F=aIcT#MoNAX~x)16m)D;w~-bghD3eQtuCh%_u9RwGq{FrNVJC7?lZG
zsOpL~dceK%%n&pfnG4WwxD$AwW6VH!qzQkxPvZ?I#{&rw`&~kxsU?RNxgm^kg(x5#
zPzJ<FrUSnR0VG}TK;MVkLpCB(vp%|t@eCWUruj@j?Xd=p2EM-8Zf2QhF#B?W_LXU{
z?+Z(+28}n}%;SOs?M?Ee9%Pb|sG!lI&Dd^3d&>-wCD$?5eTB}lSvr#XU7|wo3=8~g
zWKBZK-PT6N^BFgo*kJ?Gn3m1jFbJe4;E`o9q!zpW7R2LY3N6Nvsn0+YyDMP&UZW?o
zENWV92C$V4_Jhfk(i{Zz6F8LF%UZvTdRPo<Npm#we8}6O1t{<fsyz*ZHtmf^0pdaE
zXsOKmg(mPp37Q`C2|Du@2E7}P`?5Y7bh#^B4pyL6HTG+>uK;!k>={Jp7Lh|CWe{=<
zXKGGO7fvyV(IJ!W7D154QlhXC*((4W;$YtNp*0}%p0ko+1)2FtzF2iz543l?uqQ?g
zZLNgwP4%&jPIge@z=WxdyGonkpa#oD^w3cKO%KM|JOsV94D18>vQyiZ5!9^VXpCws
zF{(=0ZCXIqS_{L{q-Z+R+;-e~r<Vn`OQcaw<?43U2FwJPdA-`fJ5<7*q%`)Ra-x|-
zrlC_7<b^FzspoBRjxNkcwvxpZmR;#GyJU8Qt);<w2X@V-M<#tx9!m3}?o6_2gK!Sd
z!-S)Qt1ZR&QnYdw8$np?#I=s~dY-7!sMCZh++j@4F3c3)j|PERD^$J5eOF7gqA>be
zzfeG51U|RJMZ9Q>Icq9r2|D4Y6Km7!l*<SIY0wkCVE28TnFhV_W=PLrt4Eufkh$_m
zD2KbXXT$Xps1gRI%SNUF*Hkd6C*i#9!it*WKm!9{z6pj${!+$fLt@68&~`rDph&f!
zlopOxso!(@D6|*xcn#OvF5obmyqpx$)?0R1OS5Z1A1nZTFqnvfp2)%mIsR^aBRt?>
zbW?M}IHX-I2(ytn1h*;9%w~`g%T&7PVBE9<#EIS35IM&&+Z?aBKB(hqY=`Gfx7Z!{
z4~JyCaG0!QL6fqvGgBAj2SQW{NE6Zlm7ML(#cbALvhInJ9`WPcSi4b|keU)|y^B+W
z8b%rykMQZ%6^Chr&l7KLQ#*7u!J+}>2us;i!1L+~neW<g(A4mLy%Ru(<*?fHu{F&I
zX5rxL0B*ZTbhBaviSNa%UVs!!+Hv?swkf!=#>*9y@7AD02m!sGJ<F{UN0Uh*?`OV7
zpt~&GB=K+xaI({aV7m`daHs$rWXRC!7fdxnmPMY>GHyvAIc91(;;e>qoQrsBwpgs1
z)|R1}h1!E2wi+#GdvTerc+F7Tz*f;2Xa_4tBX?dQi+Da_m)x>d7<~zm)MkTM3W{dw
z#DGCy(Ix-3!FE0tom7)$gWmfw2WqE7epsyp5R!=y&dX+4E=d|4RBO<r$FiehwqoHj
z6|q%{?eInt#Uc3d$&lMg`&wiapN(cwS=k^nBAN4q4(v@dgpql^@tellj%~`7iDgtn
zZNH2EK%?ilbM2kNWIWC(0}Qq1N+B8;2V5y=tOItvm~)$eTZ$YG8P0IrOfeHB^gW&+
z9vp?0Ktd&GHC@OW(-yMc4qe`!hY7m`4UyEearlZMTn$1IzwuVNC|fNsAC}k|kL&;=
zC&&;XnmU-t4H00avC;t~!9yfUt)Juv2yZtE2)x}(#YUEi@t#j7;Ne|WD&f+Q`Yo<T
zj1GdiZ5~_tnA-wZ3xK4GH;OQ4T~}0NVwO;$mQxM6#>*h|OO4)cux!qt^o|>u#cZ?~
zR{DBH$}TIBGoLC8R@<5w?jRPutO$SCsw%3@hd2Qo2-|sCe7PqOu7?pinoPN3xy{F8
zf*DW=kLW|J4@x(1<~T*^3J^3a6_^owOdm4>%R#f@EMbipY>YN9wZMQnxUFWqJ^(Bx
zaHA=LwPi1-)eY$HFUNa*V>q)BmKg;ODlZLIZNUAp+bg+!M7srO7^GCyLfa~wj3l3)
zZ8v>c1iy7*)p_KnsMx}T>3+EZc}m8y)7?(nn`}yQjxX=WRMx1-Tt!Ga5_?=r_Vycs
zcl)T3A&%&D89;p}DVoYP=_V7lfF*|@+5lO`Ocw^W*6i|ZkBpas)~0J@tk-Fx=b&=w
z@3urpdWBsZvnA{eB-)(-Rnf?par?AllS;<s(%M>6B(>*@y#`Y)3{q619bh2yaK1+!
zu*O1ZLV2fYtOn!JxC_J?BxmLzTcG%^xsuvE--nCcehlKiVdrtAwd5>R5v01pd`Pcm
zZqWip-bGUw?2+Rghmb7zc}G>MQ)xHqjDCU7(^_1ub2(fOtHsVO5YT0BPa(@0<YIv_
zX3rh~m%@UME2CW-6rj~ilJB>=SWsgJ+4L}*PX%;_<a1B!*lGrsPP#XEU-sat(%_OP
zO)^X?whW=J)s|hi!g8Ij1ScnG-?uTQ4<(JW<62OqqVe&TSJZ6D3}MilXpW$Y*WJys
zVndMXsFvVoD{B`81+%4KvlWI}k}}#$c3$CoO3dL-F<LcO#+^gp^cK`8i@|27jDcS?
zgBI%V2y~$%_yjUR&!oOfrf#vulI3KBnfcmPQ_$7X#av*R*+StBsuV#}i@ZqYl#gVy
zi~^NlHZwfRpUa>P!@%KXoljlU0bB%bht@ujje2P=#dO4peuSzdq#6rd0YloB(K*7&
zv>tadWT$nPT4%!rv|a8%K8_`&0gi7&O>x#_f`MR`GD|Q<fNkP>Ij4qL1)5M3qYXW5
zAvA27v*5Q@PEX-QRjv5p-0s0$$&mHIFV6<zLaWRT(PhB=yv-)#c7TA)oH1PVz1%U7
zsfT8WZbW=iZp|J$7(yx8Z&#uig4kfj^p_E1!np*{w94CNl&+i@o%i&}iJ*<G<gP3^
z+|CG05Ixa_I3Lopkif}cy(KuhG~^Y$BRA=`zf2U+<tNU1Mi+!*9a{u-K*n;*29D^#
zL{w&6k%2;VZ3w<qc{#WNTyH^3i;3**9cu*P+R?W-$kbasO@d0jkgF83H&bE2@X<UQ
zk`sQ`Xp9Gv-k{j7`h>YyI!lna<LM#z&Ca^6D-#Krdd(Ve*|jOLrj<i1kBMcq+D#Qv
za>ywS?R9f&*^S2SG-jO})lfwVikZd0`67Ow%mMaC%Q+Zj?yMGM3d&rRmDU5CSJ&|)
z)o{h`X@!VV3h1B$yv6VBR}I<vJugP($$Fb2Tc(#V`L;H|!hpmb1+s9#)j=SRWGeW*
z(<lxGm>sB%gKYb}newHa2mO(>E~j#_-H@~s@?l0mEIRkiW+%DBlv?_d6>dWWbg8F(
z0u%y#CCJ+dBri1%@&HQk6-b~=v~`2zB!_mJ00cm=_Drv|oZjM~ZFPmn!2g<qGVz?T
zs#g}svV)t3(te`#G<oLCX&h|fVy^dNp6QL{Ns?p3NcD=noZv|`*0U_#@0w@}o)Usu
zONYZivA0Y%j?DgGnKr9(>VwtKr_`!H*-lpy)?zztD==fa=ymVv@kjVM4><pmpIf8D
z-}&0ly#1)7o^#Y_Kx@q_Z@=ld8@Jw-r#*DrAM&5N3;uulzxvwGJ?+H)JDwf<(Hq|J
z`Ja0SdhHwTx%;WV{!HfEzxa+5-f_ag#sB#Ex1N9KPY*rgN$39j4}SB3<2Kj954`H;
zuU>l7{og!npZTGe-gM?&XI_27ZQuDSbtiVlZD+kAzQexZ-b+7t++9~c{LuY>$Nu(F
zKXddk$3Epbr~hXU4xDnV^pxh#-?>curT_dOkIvx@#~u5U_}Kr_Y`TFe`Se>Zjqf={
z`p+)uUnTO}*L)QJPs`<B2RZ$QV~#oRa`vuI{g>ATg!8!n@*w{$oPR^t|64f!Eu4Qt
z0{_OP|JI!U7S6vRfq&!De{0TvNI0jT`ostR>B(pFlb8M2Z|}VH;C0Xc&=qfb=%K?`
zA1*d;{o^B$m;?CFeV+W*|BU|mZ~id)=Ki33{zLa%biMof7yiR%KKf$)@?$QP-}Sa<
zpZ&lO?z#P=U%K#5U-eq*%6tFe<%b?Td~<fgp>zM>2;=>$KmW-k=%k~cdXmJS@Zhc7
z_3tjf_>Py|`~JxV!>7NZ_XXnS*LcMpue|JkTvML<)!+ZfQ=ao>?7(Rc{otJFdtZ6~
zq3^yc8()3@O?O>)@rSPZ`o})<N$97BeS~gYdE)o~>YTG)@PSYM^B+ZDz;2Lk`_R>|
zwcS5FvAy`WA2)mF)eqhKp+oPCKR6<8s?__wkiGWJU}dNM!rz{Ix%}@kxmSPrnD3r{
z_CuF^?=LRC@MAZ;^2O;V|LOO?_U#A1b=FgxXWxb0{?M^c`Sa&N!ROv`&)w%g{F39J
zv-pi8(Vx?Q^?+{7Pdw|km)|Vh{J?jKYj0nl|0|DQUw+5Od;5R+cJ=qer+=(D=?|Xz
z@3X=C)K_2pfsYF>zUzw8KI`7E-hcNKgtyqYTzbr#UUBSGo^kfkFm&gSzID#z$uGU<
z%RjoD+6|6C=dOACu`jsd`=2^=%j=#zzWmP1ZhQL~*_TiGCHmmE{^Z4{+|+#Qk6%~c
zcITh|`<wZO=YRMg?D&HIlPLd;C!GKI^RV{lW1b+~@=t$o!WW<LoNwlT{R@xV-n@R>
z5ym<2#pC|+lIw&++-dhIfArS#(+@oHiihsI?a(!kd$P1R|E&79lglUkyL<B6*N{*9
z$LOMSuRHGKcfaAJ@w>_Oo#R{fH)<Cg|J>8xbKW(FE;|3bAAUXlP<#20UwkBGIp&?e
z^dn#xZ-3H@Z~yRxf3p4ZX^;Ql)eqfw$^Ca;e$nE*{S%Hm;xmsW{HOoIfgc|{|Fj2z
z8p%M{9{ObdOU<*N{?%Jg0_t$`hYy~2=7s1N$@hM9eZm6|{>BME=I!tu!F9j>`WL+E
zWk2}C$@|WE?LF^}KJ~lDU2x7Pe&a~*dhmJQf9U?ZE_(QZFFh=sdex8I`_6gzN9WuF
zHta%c^U2#k4y^8?*Pn9PpWlAx4=;WS`QvZD`Fe2o&2RX-gD*RL=_R-R+3PRdA3T@%
z!i#SE%R~JW?wj^M`;B)TVH<wxy<bib9RB8+r(XGk&m13p{gUX=%@03#^YxE=@=Lxv
zeWQE9L2&Tl|K+ut#esvb{L5bse*N_KeBkJJWmoMld*ENLz3sfa-+cZ9{@06ZUwHi!
zj!)ijlKf6>clL{qgiz2kj^5pT*6zL=&VKNQUS~3Y^VG>#t~~Tr^Hnc>MfnBnkaxoU
z)=g*K@%_nxzkc#*FL_q^d3Z~?ANky|Z_R+sMSqO_1CM^`CFM=t?ME`PpIYOsCm;C4
zH{0jm1g3QesDk|TSG{KQM-N~9%^%)$*K3Jq?QY!MTi<r*>L<SBaCPr*KJ#ae{=m;A
zmmRwB(C422p}*X}^MXHp<%^z`U-L83{dasKySMn0laKJeSHJs-XKX(G%g-}@{UI=-
zdH2|SdHS*HdYQQ3SHrhE?|%IyuRSyP)ECPyQi~Ve`kvQa_v|xoe=h&c7hQ@TNiL5a
z{;XerigOTX&u@JCx4ziz{ugG4&%AYh%cYZFyTJXW=Y08-*PL?s?Qc41a^4wtUvt94
zKl<$aAHR0Qi1Yw^%oksF|LrgM?fb#<fBD%T|9kfM3oicPWq<L;>z{PvRS*BmsSm&D
zhu_Gqe&^u_&%Nt4CqL_~d;UIp+dTYL@&w{P5Ca~2;MkX)eb<>cUwhK`KJ>96bRRI@
z+r{32gIE0}{(C0@3q1Xbw?Fx1Kf2{pHyM`mdry1ZFL~R=@js<-kFNQhzu0{D9gFKe
zarJ|DBqtvHvlm{aK6ER0+UmRaeCwgNAHV+Ob3XtO>h}&)pM7TOzu9K<8;`ispBniM
z*L~&v>h)iI@m*(>A3OMuiwEz$;jR~-0_Jux^7w9U*WC5m!<S$3x!=0@X?K47(1q{1
zP(1a@o(v56-tl+tI%25!=g<G}wWnO}|L+_B><*wF((&09o15<b`RksPUGW1Jf9`j#
zpPY8?ncom!_%ZFW+dp#cb--QX{0r`U>rbiU5s-aq-tu=RpD(>U`Kw`j%YmOQ4xfAW
zLqER0|Kyjx1t6Mr_>3zrdENi#4u9pqFPmq;H*Ift#ia+o{+heL@KL(JkBE|&sINYJ
z({2J_EWGo1Pk&|ePxk<SaM6>`{?T9GafSb?^W#t4REvvqx5k|>ea#O)dcOQu_uUOf
zyWxhzfByX=uKh97J}^G+f$yAp_UN|1{OH)Zb$`*h^<Pj2&c5h#U;W(v@#nq#@5>kb
z%+W7BIse@AE_&Zthi@W2amJfZl7&y){Ms|W^w#--$D-$>!g<CCN55-*S@cI=efQlT
z|LZ5c_yrGs{|`=k<_W~t|M?X+Za(|0=wn|su6^mX&n!Rx#@F8cuAATY>8HQyiceom
zzWd)7EOgc50cLgn_R1gq^Xo4q+G}6<iT44JxiWjoG3VWX;a@!MEk_+_kG|`QyT|VY
z5?4PzR<8U4a_tKa9lrYEvuyv3!|$E@nVXM0lAV6}E!V^!y7FyTegF4PeP2gUF8%Cj
zhi@#67mlBI)*bPy#2a39_E{fPj`<An&E{JdKJcs~Ip3d}{4;t7-mhMC;fKzC@U}}o
zab@REfk57$@&Cs(Pd@)O#n;b%_`!1bu9xhKuLGy|y>mbHf0^E!t^ow)UHFapU*A~1
z=J4e=%b$JL?hNtZU58&6f8fq50KQ+OPG0)4n=|FdM_k@xVaWTfr@Z5`zx{{rqz_*5
zxsP6R_uKK~pZ*)Czvs<Ii>F-bxA*)#$GyVs5E)Rn^5j$RxbcYU^P|&W{jT)mBY?d>
z_{}$dZE^lfuLKUG^L&rH_+t<L@EaFB@sv+K=Z5dT|DyARkACq*w_Wq~Pi~H=)Q{PV
z)6t`|EAKVz<^TG<>t28ARpj@+|ACKx{kL|{6ww<F-2)%}rE`C!e(ll6d}eUVKfFAX
z0cC9O{Mh=6AAIH#KnTtmQD^+_Pi?`Y8V=Cv)%QJc@U-)vxB*=3f)4=K@tik4^uteF
zeZkklcfIgRAiDS7`S3q_=g5EX%;IlPzwDFIe_);$AKmH5_4Lm^>n&%#(|qOrMETw1
zUwu`q9%v36{FA3%4(8!K_rqU$+5PVWo3j1N1qN^@)xmRje=_{VGhTPZNuF?Q@{zwg
z>Cj(X{J~#4@gqO}?z=7mu#`OKf>QvSyX2<&8-IM+;m*zv0q^y>Pks5w@bDN(9(CX|
zk9XgC@Y3(T^ZB1X+sFWES$`cE-tng$dd3OQ{r*cnbIU2*Z=b$6=2yV(btrPqi+&8W
z?`=Q4vHa2(&-lvxh$=LFPX6g%HsALJ{2R$vUiBMax&MZ%AKE_WrurY9Kd2u1@xgO0
zJN%==V2R&6{hj|I30{6o{k7mvJBF-W^o`ZUAGi)+z>Ocgwz~Z#M}O&A@7=m!n}u60
zJ$RmY{sZ4WxjgmoMW;RZyf2*g@B^Oyv>TrLSZ#h(Jx@RB`*&V`*M;v`oJU;=^zXd`
z5EY$w3DC(Wzxykndg7_?KCk`W`<26Iedn#_g?|jJ_?DO4c*J1ivAX`AbB|Up8q2SG
z_`VysbCG}9zVV~C-7DX6=?yXeV?YL<KfUMIU)dhI;QH4+`DNDO=7_xL^!GgV=-WPc
z`TDYd`u*eCbMN~+`XveYz4+u)zV^|Buld2JpME_M`MEE@%KE{JjH^EWi;uzNpYnY_
zS#<Bz?*r`*o_f#0^sjEcwL|y^{_eB4@COdUr`N|FeawrVpMUDOlivqs)O=-l)x96c
zE&x2`qK`g(S07QWpLXmsH-Gbshi-lEInTcL1|YUm&#kT-J?=X-K$rJ@)p&XOm6x7+
z>9^l>=;nv+xcLd6`_vcy13B#%o%rF;BiFv~XP@1?`YVUBOTY8BOYZyL#k=o)B&9xd
zP5JU?ocp^+9a#U|?ZE5~U2^^Sc>Q;~knvLb*7e)I`mrOa-7}tf;EAVR0j0pXy?5Qb
zS;j{uFs55hf8Se9x$M?IyX20Kf8!(1_;Y*r(Z}xFkzyn3iLd$A@i+78h0?EXzgB+f
zrNkYd{MMg%uXy+GfAHk*-}sb2J@>e$+&Vna;ms?+mcI9o>i_-tr+@oN@44;MXHHK#
zlIT8j{xu!8U-Fq>i~r|ke>HmZNf!Xhb=Ctn<$wI*7k%d~>UCG$f9H!==Un;2uRs66
zZ=ZbHum0u}3GoQC`YABD^)1%`zJLDsWj_XHaoRuq?o+Qj@U#Cfdv6|1b^pDM7bPW?
z2xTmVI-wFWlh9yHj^P-J44F<c6H-YL>XtdPP{$#2GMAyu(;>4+=J_0SPKV!qyNCLG
zfA_PV?|Rnr&$HI|pL=y%IPYP<_P+LYU3=sH^17jLtdyq3{)Y2@nr-iozlBgrgJR`+
zjVTKbIFXd?TNbPAWdCnhNvA&y7jmURq~fQb-&h1j81#=a+G9>LIdy4Jm+d-bc|0eQ
zCZo(l{`*(>T$MXIL~1HSJAK1+W7_xCyzd4j?!6tR8MV;Qsv-bwS}YBdO}v!LRK}4D
z^qvnG#V0J=jAH-Gts<fBx0Q9xcVm9z6}^NW7Rh*lr-;1v^1y6#td_5qX&2pH`x=m*
z5-R8UOt2sI@8{w7w3ld3!(_K!wb$p=)X&BFBfAG0VgY7#hvYKxYF!2uge5HQU%`s`
zR=@EZx8Vpm+=aaOFBkeBAH6&EUgfd^x+E4Lryi&oUa$CH{99J{yNvu=Mn9_-y8rn3
z->>+e4@+F8f%+!lx*gsk%P!D1IR7o8iOKNMxAF8ct7Z(k{~DJ6aevoKK|%^GI^-h1
zjh+~OuhPj>H}b>3oMCo$xYDn0Slzj!Eaz4J?HK+?V)7G)3v^=|dIjIfyO7EFZ#UE$
z1o!i?;K)LMkkGa0|G27u?)kTW_$~7DF31epbW(vJ(t{uM)-~SsF9+uOb$H+bIKyF&
z>oH{<$NsGl{KxBiFUY&luY5O>U*X>WO1@GMfxFq;9SD(9sAX~FfBP$IW8odR<Fr?m
z?kSkU^Ig?-IS%mnzrHR%Kc;=|+*vk$<9@eR9A5m#RsGxJn1EO7#I|T01K&y0()+L9
zUj}z0V}zn<Df@Ch{K>z5%kP)Lj)mVeidPxfyOqZK7(56)!mGExf5TsIl}<%r*yG)s
z?eRO(dP=nae|Pd<?=2B|wZ_$M2vGW8iE#fDsQ(kF{}%#vrn~IX-(G<KaxniVX#bzi
zuB`@sn;dAU9NGFku(KV22xH?tzDI9zq0mx|Jzx+4gm>$?GV7&;ji2KkyHJ_WsG`u6
zr|)FucP_l$tnpj@_$NpCIi7$Vse$UUXX0(;fkZ{s*1UWPpoYpe;D&}?0d-)}c^HA<
zQV6@MCfd~bY4T9-YI2b9<vt+oGt3ca<6O<8de&K3T0(eXgX+6s+G`nW-wtiVxsl<V
z3`*^lX?mKJ?SZ|wap=>-wtm}LFXpw#d2JHG?N>zN7}w8C;B<S|<<~c<)(DJ4G~@ov
z{`WshbcxbFU;WuE=hu>j=;Y^sL7B|VYd)0IpjSaAuvD}FEvou8r);8i1K+@flvM?3
zq~0Ca9^coN{PGvbP?%>Z1++dHB0YBz>FwSF+20QARom^PinSUMOR5V(J(1Vy<OdfP
z_wXtt%5MuKf3JH_9GKBLIPK+9Hn%Ikovmx&k8#s7Hp+O+Y`UFA=4BBnqM4WW@35j4
z4VA&J%otIW_~1&!=ppAnl?Sx2sF!XrpAoK+n%n)B$4_II^N9tcYv0+41%N7V9&R_V
ziYo{!5}#|PB5XHhzCvDhp^|nIGekxY4VsF$CQ3*@OAJ@!o^^GfFI#O#kTen6npbN>
zFaPXrTQi}`zI|Egl!yQ1_^W1-m3Z$rGu>Wq3-x({|K1-W#J_dGBGMS)V4owJQ^Dwq
zX2ry;N?@>~)jt$N>tTfp?UNFK;`biE{<Mp+ZfQ53=pSq$*<gYjpgLJ(V{9udDaSDf
zaqSK}T3n<J?>qelXZ}JrzgIlxivT2*-g@8eMN2&=53jH|ZrN5VwIFlh5K4i^<J<FF
z%S{rRKpaxHTr6@85-B5Im0GN3?l@@m^7$U1i;>96cl{vDJutV2=-Pidk9Mcw!EbFE
z0qF@oQeN@R(BfXh>Q1@3OYT5o`Y9GcX5<6z+(Y&Q1ON1gy9*W6P97XG;9KZ_B6Pnc
zhh`c$(U+Dl9cwj=K1l9NO}f`}BY-OcI11m?<2q@n6TUev8gSiW80ta%50}viH2K4O
z?Lp6OJF5+&GTU&fS042R{;BS*^atbb?bQ21e_;!t(`Z^q4}ZqNm$+EK+Qgfs17cR<
z6$Av_pZ9AhY=6>8Tcfv)<~7D|{4kmQA@fF~%yOd*pDHh!y?7)LmFwr&w#bS3IYiSk
zYcJfi=cu*CGgWE2{nc-{<S9KZ?QN<&ad6(lcPkGUm)U>p0d#cgG5q$A1;Pt5k2UXk
z0eJiKdf4YTsXc^@&7E&=ky%!&!@7+IOK2cWn&EZ7?NUBMZ0slf>hBr!{{YEE5+J=b
z{HCVE(=FD`c(xFmS0b5g^`^#;T!}3Ol=8pgUYb*;d)8rR^Ka1qFR=JK(rrBquVPQt
zO4TjQ<b|mReLB@!_f5DCOFj;jo_pt9xpjyBWbvCa>MYjx3TO+aF)|)Yk1Ob9iZ-h3
zTzM5y4`d9vE}pXQ&CdrqAjHs5Ww@9+`|%(^=u^lgeiVb7wZ;!o|AFUE-Tb1iFTK#w
z_{_iWF5INe;M~c)#y=3+o%l+``?yd}^?pRkZCddKS~&0XaU}Yjfe*cVO4|Y-=DiqX
zss=&jX$VA1opcU`gU=66g(OC{^i00Uf{geYO-ryeJeU5=$PjCIE|GKQ87J{8EiNPR
ze?1r6p7?2v{B7e$|9<|w3JSvp97tMX+YT<S$MfxQGAEt^a1FRl8P3N$A0buE9v|64
z&r<MJ)0f;a`H3%74bFq9&aUiovW@4+6M=Tflo+>Gz1^JB%Z;KPKrw{#t3|bTl4?g4
zm0cj7*LxlQd0bk~rIk*Uz5r2pD4^9sCbTsg_GmS%p!|dU*;{CsT;(QH!*avpZ0${{
zQ+gesi~v0y^THC!NerbN2)^U!p5QItB*$OsIMqzhryv5cIx6ekh`XgL*8J!*w)b$p
zIDRHqE~{|Zzi=J$UOoG9x9HPLir?eC=JJ7O4*gtUSoV1RDSCT}*IktNjR7H#bpmEK
zVY*sYtWUG2UOyboCDR>ey|WoA&EKWyzZpV;HXmm>*_Yw`red&VVq89JUnd8>=9_el
z;YjFg6Js;FUA0w9#4U)XW$a~dj^3XM;|8yyv`w=l(ntWo1mW`rx<~69zkUAone8JS
zq$dLLQF|o6J^K&@48Ri7=U+m2Q*vv7T<7h>roeQ}9I|@psPc8HEvJ1M7_&Z!!!?hX
zI8Rx^FEv+@ojR}VH*V}UAQpFb>GQ3$Xv-;+ub#yE`_E-zV_gv8vZ?;(D^DjKkg!R9
z+q9xncn1AS7xJh8<k5d3`JXQFf15{-EXY)^{akKImtE}_MRddB+|%rK71Wiy`eol~
z=(W~qzceXw)9okvP_)iNek9_MhPY{`%&IL4?Eqi+D|z#6ABiQp_Sx&PUkKrK&ipCU
za569d!YWdXKwlT7>{$oN^?ST6k~|R21PlblhVLUD?7%>l^dgBY^uH@xe_Gq*Ir-RO
z#oy&a5+M;QP~SNtSt5z9aT>70ONRXOqPO@N!o%uC^7U`+ra}P3tMs<M9qF@{0#|t!
z1kHHppIQt{o`X`ALj<3K=m1D5Db8I+PE81bnhw|}eeX9Z(+hs%%^XFZ{tkPVPCZK`
z^g78P5yb+X=H3Q+883~$ZxL?3qB2ZVF4{kCx%`r9r6l_103`i(Vu$8XVQjGN90&#=
zf;rU7jxW~nB&-biZk!!(?kJdz(KxFQ7d(K&`i<AllB~`;bX`$(&J)RG-UAt0n`Bf}
z*JkkRQ}(eRhcc*)MA~nS-#=y0&~IEW<EV1x@l#!X<AQ2!*PpFnc5lJh;tM(rFJBri
zUX4|h*x|u`v6V8bv=g#jYc>+DV!*j&<MoSwaScCphJii4q{{7aQ|fz^xythkUw$BL
zDM+unF{f^v@B3_{zFu5O!;lJOmW)tQ{~h~^Eap9rAo;9LfrRSlMLifCOKN8tm|>WH
z2=bcOHQDZ2M`zYwL9tB*Ay$}>ob_9?&9xtj+CDp;m+0Z*y(ca)+3s@YTaZB#HBjBD
zfC=rd^_UI!Bn5E4=>qX4nbFzlS%BX*@3sppe?Sa)!_$#5vIoYZ=ivdQdH*v8D|gTU
zRMqUYP6=*D2+Yf4z&2jl!FgK&qvx7c{%N*oC?3yEg*$xuem$#XE(R|w3uyH9X&Ocb
zdKn3|u0q;v@f_)$32pBTM7Y|T%HL|7gJPmea2nuNwfc<#zYc6v9;+fC+}~K4U$ts`
zJSFiDmYg&N37QsAl;ztG2XOwu`2Ee`?xjPlIUk(n<x2-Rzd<^gFWP$y61_sjCL034
ztD?|s?4QcZ`>Y-4&A2_F{5ns)P7fK;2cf3@+vl(j0kA!UT&>yF!IUXC@|aJ>81%{S
zm-I}%T?}vdACNMn6W~eMj5fXhtI^gPT<OFz0pR&E7;XA8a%&Xw5XA{}7{_5-4J7Oo
z{##*}N`xAf3N9rX*_AKmDd(Un1m=I{kA1qc15tkqL40BHenP-(OA3`)`h*;y;7bNZ
zH-7^2;g&7DJ)Wbe`~UP6`7IuD4~ZU6@R}<|#Gy2x%z0(kh969VhH<IQL%Y~}@``D_
zWtFMVkFk!+DEtfRID||83s>!c9;`2=FMp{X<$Jq1(!}#PA}?<mszCSG%xgSzMVw(B
ztO0R_U(AB?kuw7=(+lB`;AV;66kIo!eb+TTfn3d_ctZ-zUG`aDPSC_<=3Wlu{muq0
z#91n_!}s=}3Qdc0oLrxMk+BqfA<tGUmtc5o7337PKD#`1zVwC_p%dmQ4}9rf&pzx!
zF!vm^anA7l78?g?Uym3(vCm)>D(Jj=i=eBgZo0k{iO3xTS`@o?M>x(U0$xnsdKjtM
zIzR10B)?k^3qNuL=aXKZ8@iMY)SU8}G7{+W&Y8kX@zGOW?i<wAO%WZr+H`1lPJ1<v
za2=QrLprlfnHHo*Tk&C`L9r%>CYjYm?jMM|rf~*Qe>cGgq{|<Q6ujL>1e6#$6p!V)
za?aO)t0U=kviO;2!yrPbA*$y}7BUVXzee3z(c(SlG>J0(ad~BMqtQxEbY!64x+-vH
zI=b-8rxd1ucL~W>X)|Ses39tk?SbmPC<2n*W2J6rfb4Sp0CMEb9A{V!a>7CpGZ@fy
z2;la#CJ${;d^sw6a^Er|Sv3zyh~-ef?QMVSX+e(tTOU?6K2mG@(tD+6a+GJoT|i``
zs(6%g`^DHj+CA(pWos&TKOc)s^=CEAczaTLh<CZ%7`@_BvpNV(E=n3}A+LNlGz1I{
z>hE<h{maNp$ZR=E9y~!A7>QZ*c7va7Br~V9q=x&{Bzb+J)Rq7~wOkQxI;&II2Vy~F
z_&B_?rg~QDIWQ(!8-g)r6SWoZp2gJi&xw+GMtoIOeJ>FAQh&7gu2*_ba5@GjVS7#2
zKCm5{6F6Z*Yf;ZmIomhC-`l&*>Neoho8I(`>eA&MdV}2~K_c-#lw}Lsv8SAy*L;3b
z=Y4&v@w!E%iLer4`<0HdNYi){?t-raQ8yc;j$824YFAF?TtM*4wTSeuR8QFj&Urrw
zI_YWE?7?I%l^-ZdKgZU`<wabeOkUt=Z(ZE<!;~+_&EKRZ%ucmfr`nZ~2M@jva1WUW
zX#}zLSj<1JUvF?%1$cd?$BD&T4AC5=+aKA6AC6CQmq%QIDTDsdHWF}3)E65pHiZef
z&vQ-?T#^!UO{R%%zg3v`hoRGCbm@jnTImak+;$g~12FH63N8EuYo%peVw;~4mxHj5
z*Xi44c}?P1c5pds8QSI49NOjD#UP_<WwDJ8(EL!0MNlFEY2g^JWfatu#XHs>>=}+o
z(z=PKnzz;<IecVQ=6*}C^AY6$3ZT^J?e!;$=x5cV&%AOlz%__L`8wIj`QXf(B1F_f
zJW-50_Gg0{mC2Xav&49{Z)b{r^kJ!m6pe=bgu~sWZc}*<>w;bzoJ;D*_l-_c`7Q4B
zGIhMK*zKI!FA4l<F6z_NW?Phd@;>8bW8%!EsqOKjajt;q(?&z(<v5(fikbj;_jp-=
zNZm|#=4=UpVSG=V#9&IO{JQ(fU<o8EryXgXD5V!3k;bl<00gO$JJaQKv8znJA{~C*
zVTRRtE?^jn@GLYt@5g?My+XE0?Oof9!hJNI#|y`dhqm2YlGt|_cAI#;0<?_1P*#my
zh3g(7Dx}wkK@O>?zdsR<c_|;?BG#p!lC)xGg}fJbFX-BCIQfm|V12VL{)5gFxoF^N
z81{mlE>GbmiFL>}W$&x!mz*#{Iw+ia<W8A7Iixquy*=N0Gqa$+zkYSVa6A}GVtV`(
zOX@auGrm<QsHwJxb7%^|=k0a6bNV6+oZjbhl3j^AS?$}>*~7$ZtLqEyw=4#e09=0N
zmkd62TNLvUhuB>$4-Wtzkn{aG5C@!8l||~Bfc%e3uUF1qTchvrq4W)DAV$&-=%PDz
zUlG4yiLZ3p_Ji+QY>^qcb$iN?0QoUrWzUoG(r((8ulrDas*mN*7Z5ON)m~i&R<Gck
z;F6>QFYPwl?`KURV2B5l<-cfcsFjdK>sj7L4oPdTyPH;1en4k=oGMd8IUN50G0+H6
zKLyKmEDHC7Q?ng-rKmCn%p^kt67eazNJDrt^D;@7q!Z!s@yl<9V80wzkU4A}yTKc)
z;15vbJ2Z=aivp)iXj+(Rr9r~XB|PHTiDe~NKa4e_o1|X~H!gc;GF;r%*hU!3C^)f*
z;0mz3y>wv}cgk*a<cOPVJIt0eN`FJofsj|bmx{Mm_M6}g*$8R<gwDTg^)e+oTnbvW
z0lT+#Id$BaDU%a!hf)ifSGnRkPG1*@k4)ZY8~y(2^>u_+XzCom<Dtw%T{~>)mpypM
zcv0s3I07-})JVZmwOgXx)6BQuS1AM#i<m(er0b;L4zi*=INgW+5F#hL$oA_{!?t*H
zmnZJ?HB&qA`2$Jks}=H7LZnV`$mnCWx>`?@q*~1_DeP}qa5v7ruX(eL>4gb&+#H8K
zb^aN<olY6$tFC+YNsYgB%9X+mj!5vvrvot0(kAa+3GG=+X{wty9_wnHzuMt{n-{^4
z`9}d#HUdJvs&nU;Nx?5`rwp-Gsps2cC}Q6ppSeKq1QzO^)sH#-HrvtD>en$4fm_GV
zsiuVa)l%zL2PUQ;`(#=2Z^NJc33VgNd(<I~Q)&Fk(p_-(T>;dDQ9=pTEf&CgdMpF5
z!}d{JK>fZOt!q9~jw99D>PMTMpcJR`{2boV^M)|{B^i5!9Aic^nQWg))6(X4$Q!rV
zR1v*1r%yW+`5hy{r(81sz+k7pRnt-TxreMz3tAD2$Lzy#r@Ql>R@B)HwYlAHJYDy+
zSC(94Mj-0XdRg0A`KsWTNK0M<_di(BkPGIp)-vZyIbwT8;tnwL)Oi&|TS*NDv~@27
zk0nEidy;q;p3~GK`S`MJZjVpB&u*E3dp`A+@z!;JwSCBe-wg-8D#Oa{(mjPxz_P2>
zHG-`0R1NpB@AVsJCo-Yr1|jJSUHrylD2sE#Gj#5Pv}@86GBaUR_U(@=?dKVU&Q9H7
zLd*VK_?Q>{0@Gu5>>K+Q9$a7qqYRt>i<Gc)>$h&EElyl}s>tJk5q;naPCS2>t^J5;
zw)(iaKY{)>oBe}M*AI!|RZ1R@az41Joq5}S86|dY1Hz}?yoAX;#E*6H#*y{cgCpKd
zwAyk@&qF&)YtE_eJ{d@0jhResk@m=x+eaHD4rsGoG>vV0Jl|c;`3`rLhHXK7!yL3b
z%9p~k&l(qy0Mf5+tPdgX=JDo<A)@jIrBYv8ZqVO?u7<FStPykeIKziX1OL>FM(xnk
z7e!<m%5`G7F1BX-ji<+Zp1=$wd0yAg<li60bhd=7(s`+^SX#0&ldv-FzL`3a&SvKD
z!p!6^^|v*F1`@D+pxnzTp>khiaIi3D+x718Prjru9sU|U?BRa^ZGC0PB=Zr+#%_Fb
z?<Z<)5qW|yh!EXJ!pohH$uXQk8isb|<GW<)JbcI<2@x?-XNA#3E>SWo-%K>y+DYS2
z_qmp7Qo>6@l!t#HhNq`ygeh11rcQJ1VafQgWP(4gfn%sBWkWDNRKpGlHA;v$hJ_e<
z`;OXE1T{n+34BtLB8Q`tn0;7=rr>%frLY6s&A#&&F0MhXyj5I*Iw9cGV5U|H$j10h
z)+5)zXPz(k;$NSkzEBnjUDY9)Cji<K+7VCEVb!|<)F}0kdzWTU>!{skP5JY2(1M~W
zqeYmt1mlPp#GDC^X$<-@g=rcpxBGg7>H5_-VdrS-n3{ir_wi*C6`av}tS?b{=kN4H
zCkIK*@||O+y`~{On}78H>)}rml{k}Ex#)3+u_t_|Rwp%mr5I#Y1OV_3C3p=)c~B8L
zEn~}#wp}ei<lN~xeD@=x#M5)8nP>N*H0lE1S`?mnz%k{|gUe=~bq6mgSKH)4I+)||
zxDl|*TM)MpSS1CYq*HiR0gun6UNA=3crocP^9|FVBW#Er6Qx77A4uo$8;8zK%&8GG
zdHDXD(^uT_LUd=UAJSQTs0IS1amRyh8`%b>?x3oJr?>E)FKT+^X0EavRmnS-(s-HH
zan`*~=-u%B9<^yDOnw~rIZoER*&g24UA|JlQZ+m!)Y}qSPJCtbJ{QGvt#II+cJvAd
z<p872XOtSeSs&ANa(Bh|M{%P+bI7iZk4@9m9C;Q?^1$01?r9PO+-Zd0xc(0^bT}Ks
zJ|jgEbo*;=9tiI%>~n~fgHj#xeqLR?=P1QE2ycDh`|XhNJ78fH9E&K@$^RO4Hu26h
zaZ6qr!?Xet*`E$T&9Y>2UwQT`Z%k<VmA2mOzf?weA?;D5%skRH9dMjXcQ8y##r@>{
z$Amsmh+@1)DS%nR2ci6A?nRaBcE?!fMz=cM#U-4vxZDe-qUn_b;$(bz+cQ(RrSc%8
zmJ5r0w$wJhaQoE#Rwnwaq8bS@^ai8{10PD&_JMs%qUFOw57d>q)qZ`CAK;^;4wr7M
zU^gO?dESzx?TCu#Y&s6_(8Yl@#6j<{7&|!)1;2FAbK2s@=QT@H%LXsk7NE3<0HrOz
z4xGWTSBs5$xd{~d0EQ{+p5h3+^wxJNEa9njwui~-MVF}WBsXR68Yk_6gs$fGr&DKw
zi&izF#SOn1IotKPp~t?2g4I=4xoph7ZG+vM-vQ=;n0y^>`dd+76xhCpq?@x_rf9ji
zOVi-{=OXYGz|%yIgI^U$1Kf?CZrwmjJqAuD^x8u+q`3*Ep>n$W03n`72uVqi!v-DT
zHRn~>ec=<ss7kuo!z{Cv75YfTlvF(ax?tqa*j8m)3A3@*1nP{{D@$JCU3SU_nzE!$
zyNDm9LQ<q;9dRjp^N4Tt;fg6?^Y5p#a2}!I<0m1PRQ|yia_$Hhz)HSAEv&*77msup
zX)1)Oj<ll(3cn?{y0vX(TIvVCt?JKW^ggd=y&8#~91A&M5wx;LogXF8UQITal)96`
zsQOkC%Rfu+Rl94ig_RDr8<h#?f__z{uReAs%(G-`!m!nyran~6C7+4JiR(jvm~ypl
zlLA>5L!PW&E6rY_>2?fp52&vHzPOGnI)B^5-3{q~fd;yb?zSjQ)^!P*`ks*zQa{LN
zQA8hACp{Ti@pCLgX~cxDpKHGCRfORQ1{OUm2xBH>m^^yPm=UOWUim?unerHz`2^e)
z&|{K72P;=7=1jR|R-Xd0`tI8+GCHTDL<TX+Bz>&z)B0&K>1WqZ(zX!fltU~8O5TC@
zD_8d|F5#n!uk?wA17U$PKo1_LOVo9U<GzkwF?h>^yvrBN@D^S7y9;+)&M5x^%aPk`
z)Y&nXADT1<uHb#**1mI%UR)|nraZYshZ9hGddZWI<vj>JA-i~MJ<>GdV(?m=TPLRQ
z&*w*!>T0(IFM1aXUgR43vRpo2ySON{h^Ut7FYjFLs*5VBseY|(6~1*Onf#ej%fW0`
zyE+@CRoqwzBTJUdB6_Q+&Z)4+<w?-tc?-MF{_O>jW{9c_>}1-NIru?ty=Vmy*VMlo
z=GRE9(OXCMES)cEYhl?+q%Vyeup#kOOD4~sC1Z}9;tDH74-RP!Dfa7ZvkK}}<oikj
zeIY-I-Vm+12jd)iy2?2}PQ#wc;=bpB{%1)td6~>_ew8qR6`o?ii9L7cBlX(&s#qAW
zkiEwLWCUKntQ4>3nIAWH?K!4AL_iGfWhnviNfX&mtTL=$Dwm+Yh1{twkt6Sh;h0tJ
z7Uxo&*^DB7^^S${3r3mgbfeO_>5M-QmEO<=TuXsZ{rjd5d;U69@J8saf&j~l8&GHF
zZAAf3P<EBM9^v{{b`qV2CFz(EOC!6OU~-$ThLWgMRdSx(`VLi-^jp#U(9z*RqlSbB
zN6lEj2*kS|!ElL5(Y64ueim?ONX9oHA_PmiOZP0&dsBfIY(5`YQ1~vJefNIR@$TM0
znifa3-I<RyhNtjPTUt+h;7;o$VB=i6KBG4>T|c2#5wE-}h);H6<MsHi;RHn+>4Faf
zi<x&fC_^n$k8nGxz8Tcnv<y+_-0}{1<BF^UPCwWHrkhaCb}ddc*k!mFeY;d}4dcO3
zV7Vpq08NM-d4dIX7#UF%o0p+uy6{elU*{9ULIjtL7Jq511g~IxvA(3-`30o=7hk+U
zy7!0c_V^&Wp@-c_Wn`HT^sJgY+!!@{gx~J*ss)L)<$QFd=fONdOsQ=VkYF(=`?@?2
z@Y|-tywfSS?avM8N9val5jugk<uv;l^zC=xx>mh5hVX3B5yp-wOT*A79Q9g%<7*`?
zTWQy55+w|-f4{Apt}bi8@vXOr4q<=Oz1NL)vUuwS$37h%0=2&<oD#hLN`I#$+5SyY
z)IHI1i|_Upl%^0M5S?2UcMxXNAAc6&V~NuXp>I@i7b_3#?u+*!PoCVyF@(kyoohiR
z-~8__(^Q62>b^xlw`>r2AZ?i@ua3v}uDHQGpjW-Ti^@apsJ(%_w>uT@eth~Q)I10~
z=x9g8)>a2It!MU|S$w5G{@p)qbV$|r1}U&!d3*4aXb{;=t*HAwl9qFa@<xBo(#v#X
zwu_U^My0*C#ZB8IS~VHWlg}i-+iO>}`5Kk&zfCGv#^veF>B&yo_t!yvi&PIeL}F>t
z;Y*tLTt!n?eHdiAr$#fy9PRSE-Far97^je_DI2_;Q|i}cZg$(8x-%g|T^oFIiD^U<
z8EUjG-^*vjJ>TT6u~g_TzC_-STlEI<BhSr%WYd^q+Nq~TUk2)K*9_ZJab{|yV_E5G
z9;NP+900#O?#w;x2zTkWledsZkeN7n+T8c=`m>eh0WdEz?4%iCfMePK$GGwAHDKmb
zV4$&(YQvoK)(MC(*m_n9XQri;>1kA^yu;i1UH7zmmYlFuxVe_HbH6SYk!`ycO^YvE
zk>p5p4i^%I>Ei0tI+guuH^|u=_4URFhj9R8Kke=W>6~_pS%Tq%{brT5;9H2BD_w48
zDl)CLnz&a=h{6eR7>G%I7|t1)Lm0=ZTQ>e|J6_LjyVw?3-Im0Yv_O9%lJRNn4;z1<
z6k5p-<H0YpUq@E^@|*>TGr<!2*6pF)a^b|yjb^uU9ryWy_|c3YeQggwX||8~Xk<@U
zP69W|?P3&NrA+!za{4U_tsJuNmPuUgz31-fch^4haJx`lp=s?R>VHo3Jn+pDJ<WdN
zdZs%M3SAxejev<haCItL*?1j>A~a7k_(YE<b|uEXYk9ZIuCHfn3+gL8-;Y8*S4N7;
zJ#hueYC?5UZx7SS>pC=Lo@j4X`V$0Omw@ti>UE`4Uu-1vFR5?hT_w-@LV(>;-`bEk
zp%3Ay+qagvvYq9hll}HgDHnT(fWga^c~UJMvs5U=<j^DKi$F<MaOy^^nXqqQGP+We
z3&=>5F>4q8Z6d|yCd#m&VCGvYiENW%lw&gqLs$`*%&4Pc>#yh7f}cdd*l&k+-PztE
zeUJOea_K7kDR!#EKY-b5Ak!`1|DN;a&cmh#HKMdLwhcWRMK7m^%{k(scJ9=z_SBfB
zWD{s9lIS(-iZ)7xdMzK$OmCm%xD`vFu=-12B(3=BA{e95@p))1cVLU^r|dxhJ*53K
z>dV>l^B!M+s!-1KF}csxt!k7CB!}KE|M4BATT?KX(|BRLZJv2lZUu-u?QYxwhZd9(
zzPu=j+K{hIeGxaCJ~LjK=Tm<*d%N^Mj*flo=OAePaCcxvPeBB+m0q<MWPl3L%cqsi
zTqdh`Yv!Z{O$%V5YhBs~T*RQEH=v|TdqZoFok`v(b2D)lUuVhNshV!0E#N$(gB@0h
z_G9(%7dCXsZEG$|a;iJoz{DyN^STtRl{>+0?$W2Pn>RRB?t$A<?OlD{4^f4yXCsaA
zyxC6W=(<U08VtWS>u)`5NfBYL&vb}XMjcT7B-)HATrPPXkUx@`gYbh%LmTU>J(=GS
zM$F;4IeJllthDR$R^<oS<k`BM)6$-zm^OtsXKwTuejPX`H+2T9yn1E~c|XvGUE;9M
z*A6tuMHQX7F=uG;laTEbZgx+f<T{ZQkjDP(nr%U_r)3WL{7-SQB_l%2V>6gZXr%i>
zI20xJ*l#?eBzc92d2jWWbWhYPy)NHOL+&s!BT!2CVw6S4Pb}a4lcYkNKQeY;OfPo*
z-uSDmN*&W8QgfbPGLzOVOg?2L>U~+s!j3=cP#r5t`$FZp`>n_%)L|tetxASWvv#pq
z?s;d@XIua2`pi8>xNLF$o~QE>z(pG=>0-|Q!1gRSwABQnDGsT@g@@d;#H#Evjkv0Q
zs(ffxS~bLN<Tn<Pfnd^)mgXM-dCE+Q^kpuNa(aQ2PFaysTApd>ozhh5)C-XWP&{w`
zW3t>D0s#S2SonvDU*KZHNA=C5wvqV#M3^BDMqiC~Gevq*PWKRTTte<Zh1PG*mBbBj
zW>6~j3~U8y@+6Sab>Zm~aN~meCb}z^kr^_MYGm51@Q8IDMtkD5TDl&vi_wa#9u`?*
z(sn7LT6*yn8RlXuui5w;G22DyE~>>zi3yd=<cdu$qaTM5l0FKc%d`6HAMOpJvqOeQ
zdTa(fs%jK(N#F25$bHngx-Bp2?ID+H%K!MR*$xt2?@Y=c%^-!G<kc@zc9*!{X`bV9
zWPuZ<`%tfueIs`0Cx6W@`o#og%<i<uR;TGM9<T>y-hnkLq)Hf)Rh34;N!uoO6Ol{<
zR4FEnwgg=a&#D>2Sg!W0#H0^p__}<sM5GzhklDDXY~ckpUT1f8;Hu6T+VYt`xFyIQ
z{^O@k9MWeCvo^Z(`Pi)_f7UnMEOKjJZHW&p^@4=GZ%8BQOQ8=QT)-u8+6|VzeD-%?
zzo$^QeZ;rJ?rjxa3cugeg`OA3b=sj5>|H#=Vw^bOH=c{*84njI4^s8L?ox4e#tUJ~
zC*11$Y`k&US1m$W2X6G}+pj=%;BcHg<D!SlE=2q4;cK;sT%`o;hGSDHc6D*SA?y60
z&2X`G@&vlK^O;4P%?BlBiR`x@cKOdJw_KDai}jL*S8azSKj3l>-Idu(z>M8&B%ybP
zJn`W$t|}u`I?0#2X~-;A&&c+E`c|$j(pGl}6<sHt?1@|T5j1>W9`4gQp5QAN+qn+4
zic2^CjJZ5fCvC>#$t8&h7-0I)t%bxBH^OmbnevbSQkPHpfQg|_)A&6z%{H<YSR?kL
z)SrW9ucex3O54)5bqNNC(hzi6V^_6TW<ep)m+eNsA(t@e&;``iw-EG}4{j@Oow1}b
z4qr9FITe@~1kv?Z*QYt!;f1gkgmc1lcdG+&;Z^-G_b!6DQP^cSu1qU#mNs9|Wr!*g
ze9-P1ICYK=eWC8=;I1N4d}w_ITT{DnHudbCU{4CG%fl||tT^A$t*kko6tF#*y7wj<
ze2$HyPa>aJcG1T2>f#q1a8)Fv7V>@onIhi*srj=ugj11J{o(`g)&Bh<={CQ?kgZpg
z!{|G`io_9D;!sI)p;M=}>e!xhZ?ZDcJLnnXvbmT=SkzB0!HBG6&2Tol*RUhF(+|lf
zt@%h9jH`6&X!^AIg+IMg@A<We`0T-=!5qQcgE!wf4p{2BJBWXhcFtBFs*wQk^yF??
z_;0Y;aWdNUV3b`=uHaj}c`q;RvO2eMkZI-crAzgUY$*QS;0FrmK(=+*iA}Fnj7ocL
zi=QTq5lZy6tqV;O9Vbfhbn!vu3`M<Z$#)C*lS34iVNZhzJB`pTnQV|l4SWIaHyeZ&
zj{44bS2|T9Q@)>j9n`KFmic>cLk`-U-PP|C@>%_clN|O1@+m`u(iSvX1&3_fu*&#y
zN^+ue^b@)%*|W;?L1+Ac1=#4Ey(VuVK(;06ba@@wv<3c^SHT4eqWINr#MGgs3QjR*
z2&Bq2W=K3v={CvXTl68*l!YWh-xwZce!CW+k)&B<l&<9c2+_t<N*yf9xQ#A`OjiH5
z(_Cy1wwbkhkInQcp_J-;Rdx_z0Bog-`eIV6i9yQ^mD*)%KMvwxvn9S98LZ2H?)rnp
zMJ^jzJ^@6C2E)m>3LW-&GkybmWadF7m_#;bAPWYXjts$|bqWpDQK8H#EmH}<bGHp%
z3v>2jN@Jd0wjBbAOH}fJDcs?inacfu-$uPvE4x&Ap5i%s^BHZ6vi-Yw0}Er8GF0E`
zb6F;Fl+7I!`KHjMeYu7h`Wg^J-=85s?`)VD2|uFv*^;IOzs)0HinlzdPNLL2RMz!E
zK_|JKs~nWpxyPT~Mu%ZC{H*fzJFup}_H|$o|0KRUUGLgFsFk-S)9&n<vmE{bLsnkm
zJ7=NOJ)L}W)>=nN;H07Advugs2)yVboOeMAH$`%RbJZ6TZ#trVUQ3iDA2knOiJBQk
z*ag8iXTsRnen5Rfs!a6G&$b0M(8P&vikq&l#TGy2-|A2aW`4}v>Gi!{hFQ7ADH`g+
zvP52}PrdIOmk&2GX<Nw0V?Ag%mHr_HuK$EL9y;{i#R^y-KVW@$uKjTYl4+u%U$O=<
zSiDzo7UD~2yUp~0V;}H~rVF^DTrMV3^HAfE^Z~=*JzrpKPRuAZS)*!`hJ;JUuH6ma
z!Q#T*#Sl2QJ#w^^bU({OaZ;??NR**Y<+?M_vY(}rQANmLzdg(r7Oxsy3-P2?H9U`V
zyq#STm%MQEp|d7rLgMq#yQvm|$p-IZ?dOA8Ge84EW|Y1EmOf&@%QnSI0jpUH-SunO
zCXI*B+^R|!d}s8gy$w5eN)wHm;X_^K@zw!~(Bfcvl6K#0ySbl4SHO5@SQ33)5(JYD
z7;cuj|L!@VW0`-$dVKiZ85b-M(vVmuAl~IyU|sAD7yS+4q5MqtfP^3!yHloUl*766
zlD2!E^^w+OxnM(Y?ch!-I2xWKBe-^AL-qJUDl374xOz1t$6gS@@0|bK%=zw+16@4g
z+RH~4g={Jg-d&WCLHf(uvR#;4WfKQ)sU|q^jb#^LwFiQ0-rW<_BdQC7tbMEYpq1TJ
z1f;yU%urFZlW>Xhns_`F41Xmvhxt})#TNRp^KocOPwZq~;JgQi-CM;#!dp}kU#QHP
z8hKdlWBq1sBr$p1qZAB9ISF@UwjmuD)mrLTrL261P=JwysWk9cu>qL%d3}5gaTIw#
zFr<4XbuZ}&4(4U(xIiGaNs`CqHzcVrg{{tXsvO+wDot}HX&pGh1;cFG(g=%2HD4Gk
z4(>fx{~m<nqHB!ET)4xhcWMLU{;28g)k{(jQPS#&5S9-{jnH@LhI#R9_<9dyLK$e_
z)MK%^J67-X9pBYZwEW}5uAZ)#S0$PD(NsimRUFqB^8z|X9nl7&kEGFmuM2ou`kAYD
zk#p~X5%9`sZtiV>ysP>G*R|XqVvruR&r=8h%OBv*Uf%T74K6NflF;7=;E%!C9-F<&
zc<fbYW8hK8IZL|GYDc_UyVf_Y6>9s`mZZxa$!^AKi8&n_(u8j!hFoP_DA!258wCjq
z1uq10$O7c(D}4EDX;Y;>J&S+2;iQ#~UUA<+cZcrkOt%3SJwvbA0#(u^vO1;bO;b4L
zE?>yAn?z+@>@1WJVsMcO^?>)N5T&fZ*u}^>eqwxW+H@m^`|d)OGdO(G5J20xtHjaH
zpssZNtEY;Rg>9Nhs&;}U#U(KQd9HA9L|!yC*5Gwn^tAI*6n&YhZ?mqRvF_@boTX*N
zg#oli@p}sEZ0lv;THbQEn8~~tpuEYo66L4g?CNbBG9t=E9;CE6r+k*wk{`2vu?V_M
zNxNoGvp`MpFE1gn2$F<vcS1YRS>r%wd7JnJY67dm>8A)P+&MAJMQ>zh2u#g~n+<;g
zmOT<N6cND#)M>|KAypRm!JK62AoBjUw!p;ID?;J0E+sb~=$(Sjv>>Uq`iJWuPFZSI
zfIUF<#n;p@jmu1C&3#K!UMHkaEPu#@bp-~6Q_FT>yO?_6LF-$;=dm?Q)=;lTziHdk
zucw_nQK(gPZy7MW$LUyfvF2j*Rwrp%z7_zpnU)RqqD>PM;xJNbN@r$1{Ov%ZweADM
z`K)BcL$gt*%OJ(LV{0#e2J64G=h9~ZWW~<2g0{65_l@q}`cg_@W3Slt<Jge^K_~pO
zgsA(`W=YV<PmgYZq9Jdi92)g<7%+_H>4U0m;=j2M@kFeFlE@ZmWPwy=6j)03hB3&5
zdMrCB{^CA_7+Ac0UxM*&b3ICKgK*!HSd65i_A`lQU7UR41jo%1^1i^*CH@W}!!glv
zo=vwG;_wkYn-cP5Jx@@N)YT|4m}7GA9GMaE7pQH3W|fLigwv{JaKN1AUqAQC>@Nzh
zT-Nvh*aF8OALcxszNs_Z+6WYC3E8=?{hV!Pph=o<s<DN2oky~<3eO!GY(b&Ey)1c)
zp~2yI;hQMcknx~Ihm~%pgey{~ApIBg-&~3x2S6HA{MM)bMuDgOt<ufhvYdME(LX4>
zat+BX38bkk$Sur}TjX8~)*!ismSz+hDO$ng<wel@o6bA5G4WItdB(*Te5rT#EZEQ8
zP_m-G?btH<nQ+HzW0qRsT~T9<k6v>pGCvS7Tlr3Ze78M<t@5Iaz(zNzeZj+T@Qg*a
zZp-Ma*=kwVrlUiZ(@B%YeN0h8PK(JQt=1CSkANr4ZoH*3%Pd+BLIIsr6Y%fqapn_=
z$yK|dj%ZzrqW^)=K|RY?i^Z7T9qO*<Y3zs}^+|9ad16;dK&GXMux`QvccnA7G#PCX
z;o<v@(eBpx1qNjts`C&_Yvb_q+ty<exJb0Su;HuXuK7AG=})nbQyLwG%2+rzSNL!t
z0@n8f?sL{UPIrFA0T+{C%$GC5U0K6=qT79^@kQ;37hYWpQ$~oPFM=>A+jp(!46C|z
z)GFTXB)=A3cg~R1JtHxgt3jaD9IXI)*vRwSbJg7Wsne&OiAgHHHYiNb6GhBUnQ`@a
zP74RObEVcb%__7lvorG6uN)qQaP38b%yi-0MyHOZI>^x2!&fqo!XCN0bTBa%AA66c
zX?d_Vc<Z}&*Ce40c?+T|Dd>-Fnv$6vYH@XKqsJC96LMeaWj^=?o1UWJ{W|K#IXv3p
z=L34gT+tyN6S&7tHg*xeab%MfVw|2tyuo_LF1iEZGSdxD`%|EiMSaaFUCQ^WnJBgW
z=3*DQlUh@J=-{(wY_&#U@xI!aDGN~8QW>@DkqJ1<kFQH#99w|}bZN+pK3I=E_}0BU
zeP2{Ox0fl5d$$3FwxKXdW$HL~2wuF&A%&~gYR}BSP`Ml9v5|ks5U*RVu`PZGlr3j?
z&X%|p6UfoQ3om+$)JhJxuM1e<?9;=yp7#j%fQOojK%ttGQ{fnVRYY4n;-a<Fx;BQ0
z9BEJ(@MDTvuMWj+&ZZ~MLR$Qo3e)~r6OG^eN~1K&)T54HU3W-v6h|8>&Xmf%OmcFQ
zTIj#1|E90%qlTJI&qc1<a@aXy*A;>^+Y$Uvgx}BUZ;OwOKAW$d{-q*%i>XP4kqAkz
z#ydE&(y9MSrV26}k*;4><cR?bLkFxOdX66)oa*{B#rKkC@||3WrDw!CG;s-cV*Be}
zM@xk=i!E`Oc;A3&q4FM1NZp8Pgdko80mSZ}H1!hxR}hQ?M*Hg>;GYwGen=(IeHnUg
z<Z<D)VKfZ-5Md>hP;6W29aP68L2dXgTIbg4kEJZaWdK&_0Z7leggjJHX_b%%g&UR=
z7znxnE~X4GlpaW5Ml&=|SKC<XHZS(N(p($%xNHau_Da^pS>4MWFL0gB^K@?P<WU@l
z)pbWPx01t!(6c7sLcW?Q2;Ap22k=XD3NCwf+tThd0WumEP)J)a$6s?m_CATbV_`R9
zN=y?{d<ww{XP5h;9&e0m$Zyx++MGmhLXvOaP8o@z5YAt79*G#5oXgjVZpx;YND<xX
zXV7COr}oDDmWyLCEZ{(kK!6)eblAP>L$=3_?Mz7-vH&>NRNkY9h3k^`2IYjO3g+dD
zi#nl2h$#Qjgbg9r$~!BgWYRl>S2aYxc57&|w%QCE$Xi(Tw!7~F{b;7YWH9>I(IJLG
zd6g$b<nHG^HHUQW#^LV|!wX_*)-vW_Bb+d8JK1hQ>zH$99()xsO@o|qq$azz(-pA8
z;^x+D`YL7iIq;NTgwxXXxFdY$y#ndP{%MrVIKG$Dy6w0JAR=V;^6vF0y>5@VOc+_r
zZU>jK><78?gfIb34JfRAJ2nvGu{tdHx-mcv?5%dkN-L4HWZx!7Tct*u*N#w`#&SO_
z-<~Z60DB98^YP&zAeG<<tv2|=wyXp3=7mHZ4asB3)-v_I)yPaGyuA&i@fLy!mgDgo
z_JuwCu*~q&Dc5LMTmNz2E>GmB9bgLVHU8-^4+;d|Iw!3#%o|P3td@B;P>jhVY!y+$
zu*M|6^*n>v@|4qVBxs5Ekjar_K(s16xVG)DD28xWr&o(UbIcpj)m(0xm|kK|$4)&<
zUN2qye#kdyi`^jqf=p(6@bC>ID(u0yv^XnG<J)A1jDL289@@Czz0tDqLtZDp4Ov*y
zIa#YS`~y)tVXtFvhF7L;XxtuqaUEYm-P6be@k5~gJ1<o2usXb+EJXeZ<7|y~BOUc}
z(<)q;$(FYtoQmD959W2CQ1GIET5-9KNZ842kN<d4-XaC7NBOl-66~cg=AW%Gzyy>Z
zSjYNR!(U}=9>dmb$j)l-nJx$3?tmEAUUlI6QVR1-=^}T(GEC+rQ0-VPk*!$mb#f~?
zOm&6(4IRe5lyV44nMzyHVlDi~1(#us*-K<}gsCQ*+j%+E_a`4FT5N@@HxO+eqN``l
ztGVsKqr#~@O{L2<<pOceS9`(eJXPqgXf5sctXq;Zj<UIyvpur>TK8)n*NK2ir_uzZ
z$8m=BDo=Uw68kc>H!RVAIEdvN;+{hLjexgF*d5q7$YW7RFJ4z7x9$WMWi>=RH@sc7
z9hulYJF?*Fmgq)~<Ma<5cf}WU)mvCw7XCb^jjAxf7sjFK=noY5v%f1s_9*ESlI{eF
zvs5sK)5e&c>I7LK9Ga?9CQ7XjvOz26#-oGvXC*0{DQqt?`k)7kAohxDDgRmop>;m7
zhw$Rc%lEI11Iaf>Zj#E=U8ELXcBiw9)yMYdPLwQoPU9oJD-wX))sdS$m(hbQC1n|7
zxA)hN@_pePdLIecc}ALoC>L_e(JiTx-=p}vCzN;&IlfMAoq-g+%&!shLHVuNRGBfR
zO&0q3EAM_RUy#Rkwjd)+t4B|No`g;<HkirU)%KOV`Mx3PrqY%ra(+sAeYwsQ!9EWO
z5zdI&3-T5v6x)=13pTX%Sh27TmVA6#_R0FMLFwy1A%}%-NVx{U0eB0^Vy|9EGPELz
zm;8JfWmE;c8$bO-2BQXlz$U5~s_c1x55J-#S!kN;c+Ob8co4BAxR8B8fM2*4$CKaQ
z7FFlXrp;+Qk*?GoxRM8bTnK%fUgSJN%}B)|s~z_S>Yu%rD#+;xn_E-Y<`3tn$BGMK
z;z)2#DZ_UQ(2JR_2OolalXLd_+mCxV2qzuDD1R2j8OxX}EnFR7*~^J~mpD=U*w@sS
z5Q<j96&$)CP2(2M;PbZ^pat0go2PpOvElV!!Na!CTR}pF7m;Cx7;O2E%K_K=2K)w_
zpKM!-@ADkTS}t6?SGW{Qu<4tuR=5vb6rs$#as)WCABxiT9M1ci5l~Th^~q*GM%F_%
zW6S}gUc`lx{JYjwUT0t-g+07H_{}G-3)V`d=dKsRu!2|TD;s}#Tki)arTy&7yu1Jf
z<zcJ$%Kuz=Y`jdPiA+Ft>5;NQOXOQYX4^Vo+Ti7PFb8`^PJbDSrPg9_3a2N{x3xnN
zpXnc9)xYjHL-8UmB7>fjzCS}Zv^@{>Lm3Imtbf#Bk2V{;9Or0f;cG!E&Rj4;7Zp&a
zNXzI1Z_tZB^N%5;^s(=6qo@ZTtPJiPwj8ulZo%N6+4F@*uQrpxEgiFc-27@lvby}3
zyoD4&vuAF-SJ1X+=sr4rI(a1-4DRXe<$soTq{i2#L$+MY)N@n5r;q@%FRPc7kMqOV
z5AB<%t8{8i^8(bEW*#DG*%_y@gNj&h@k^lPi;D9uwisielH9S=R$NAFg?RTd84AUO
zS3IwDGFdxUuP6EG*!dkmH&i`t9f1+67`%cEGGp_eRO%H%4)HbfZ!Uf2Vv;D-;tYPB
zwtHc&Eq8_%l)7qZ$%hv>MVp{=UrYD2?QZ<$IGC&}4<kIGcKiBO!BVN+jYV?G2@InQ
zgv9hMaMNf23BI1uP<?_~;N|V}tZH}@gJiYkDd75~>44J<o5u4F0Qg4x&KufG$FX1B
zx9W_n3sZ*_c0u$L!<li$_$PkBZG&bsjGHSjiP=aA8GzwKtTg&J7!EK~196tCDmssH
zptTLm#yX+>J}LzD2keoZSTrs)<9Tldpucw5d?r5i>~HpqXK_BLzP<U<BaY)<W93ag
zqj(tau;gM+R=HR1DGiprj;b9z{m!3CVqtRTnHQgTENSYs_4z5!BzeDa&s_XQ8G1;Y
zc$^A$Sx&Lnibf<Wd5Xb`MzGI!l_E4*ly0_+xezi(79ieypiLt@BP9%db9@KCJn%%Q
z-H+{YQ|U7oeC;J@TMnls<VFj0p8hQaLXi-7eZ}qkJp~@bSbO8u@iPztQ}jf0&@RM(
zOjw7UvT8g#q!wVIb-*RFbD9`vF3se=jtF~Qka)EVzJ&D5n*xoI>9o4kAj0Fgu0pq+
zWjoQVYYo{@jBymBxe04~KR&kYjM*wXA0L2z4M@ZwPvbM~o^#i$f0!Mcp}#P`Jp<*&
znOf)snxln%y634Y76U{ILoN?Z%Z-<-yQ?d(R%@l!rxet7Gd-wVDnv3yuO`lK9Hy)U
z?DN}HQqXQ8mxRpT8V>Rn=FS#Rls7<Y9%8R*Tn<LCo?UzruwwBs#sK)H+yj9E8ZV=~
zf|_=7-y%ui2!$Q>TapY(KcV$Vg*Adw1MrUhE0~l~->R?Y!&=+Y$&GcbLTb<E^t==b
zanFMC*i-m@=GRk{I<94_JBGgR6&&#cwhXtF8-SG$pR$JvT7OuG{Q6o<*j_eLY>gz)
z2t+%=Cp6f_6odE~<sh^>U?4<%>W>0)rr~kZ9&;pPIIqJF32BD236O?TT%TOJnlcz;
zJS(md#WoY{YwG$(cWI?_nZiJ8a48sLdnAT8DZ?`F9}T9>Xj`P|X~<=p)9ds_M(Q`$
zr3&SKI#OH*B}mP5dE4o#+xe5wY`Jy%AP_llwF-7|#gHzknktU5dS7YdDeu42AZGLl
zO%bIVD-+$;*GS2<zB=woyu}u5ks-07BtN#0(*D|~evuu~mKIQ5TYo)?T3AT6{O*m0
z1ng0&Km>10@Np3&o*0mL65Pfa7k#?puC`!ixyohbxCoL|*+_pJy$oBp6%=kY*idSC
zg?4|S)ZT`N>nDVk??))1gjE%OJX6~9`Fy<+rmSaaS?c~HI;p!O%n6?uZ$%m#yFdof
z4(LT3KEmZ|BP?+gyu)etVQAH_Xp11@O1XYV6Mi&WweP=j@;Qv4hM+Y~SJ%Kt$>!g}
z;oTLe5GXZmzF|mPV91E?dXtC-Zl9J=zdIS%eR~NF4xIW<Rn(11&#}OJU9c?q)QNA#
z*n%)11oO#AwYlYTPwTlwW+@ZT9`ox4W8%bE#EA7h>k&6j$`qQBTaX$u^wj?>*0RD7
zJVuH%EwY392YxMMfX&p&^G1#rGc)hPcEdENAEhaev>!#?mAIcrz}(dg8Xix*6lN2V
zlrfZP_pYGpGXgR5!AoJ`%>uor+_U5-N_|Xkg3eUm3|eHACR-)%walquZ6K@3fmE9Y
zs`HyuoP)pz97(v<k@jj6X#)b`bG=pP)|xC8%n%Hg>`f)&_!6Q-j0_%;@r5P4>EU`N
z#be{i6Bugiw?=3KFmqzt;xTF29`}J05Wy74h&0Na)t&?36{h})i`haCiv%KTpo24~
zA`VK_(%``*hwt9#vxgU;9!9I5;F`~}EJ|LzQI^9)J|0jbP0*C&-qatwuX6ob({TMR
zl%mj3PiHoOGp_;8B%oiKA<HQqDab>eL|>7e%1F5*eTw7O7T73b(F@KgRj#YpuDNrG
z%*gB^2h*W=U@?&<<4|!K6Gi2m4B9<w5O~BBp`WKGf<x+Hg{vjkGs1%tQ;hV2hTfw0
zV-eE0E1LQ(Ow&i}p^BwFPimo#ZR=-$Hj^mA6;vPiC}Lb;m8&y2K7*yyd$xF@x<kz=
z#}<CV)MP45{rYqU(w^oOqVpFjBen}^v*1^>v90>JB!Z#8!?6X<JYrp40#ecnW^0R=
zA*`l5W0Pc7j-Lrqd=zPP=Mu6THRer&;n<s^_ZLMUs2Wn6Yzk)cr4%w)mtZLR4>Ot<
z^@Lf}j*>Av_tkl~*RL<T?1Mu^@2m_+CdzoY7^U8-{HtTkS06^&7^YS*&eo~K<#FMN
zT`23!KVBAHwc*q&<av#juYoDwJh?}fMMn}0!z@|7BQsZp%IppXp||0?LkSQp$#-Pe
z*G4x%*Pmc#YWvUnek(koE~Ua8ECHEbF=(`VNEjOlTRSz-mVF9FNxOxV+V&Kb6I2s_
zEPULn2?E0>S)<|NTPyE()YcC!4t|~v+r1if*DGVt7O%#Gi-kYWw9ztHrr;~yO&Rso
zR8Bs}JXSqXef8Ys&#xY7iRNUqv4`l4vN9(3J8=!5uLwVHc-Q?{Da-#^t7A~c0`r@^
zmrBCN#B74ET{$PPK22KOAP>q#Qj6ys%(DqL)R?j%3e~4{Frj7jr`$ZXfBr!`4#rX=
zRyie;`bbaT+?{{w?y}_FHFJzdeywV6JF>N+*o0X#v&fQV{47N-UqjL)$8`FI1q`IT
zbg2)K&k_c5@-J>g%N_Wv50R}~fEdYz7=4aQN0w1%h-g)V9J+m!eE|I6x|o?6;u;JM
z{DGK;dSmYka2s&DPN)t`lG{T4DJokV-<fvp&{cVQX#0y4Ze%NcD%X_?*aPxvO*3T&
zP4WQ!(c1gVU-+iDw@y7dJnK4GYFbd-Ay~;%)$AT6=xRAf%ve>poKFZUvbUPm(T0^M
ztWzc4TLgMTcAQO)d!mYq8!lnQi{o@8K(pUl<-(DMT}iwW5m0t0+P?;RK%>60lpZ$>
zU!KGL&NvyF{nUx?Sagw7df$$RF->JEBIb_=tPvM}7x&%3iqqzy4m??n<hU^zq_3W+
zfJC2>Q|YuZ9{byl)QNP3Jg$h*@|>VsGtPbL@A9-a_->ZLy!|W0{2iw8BrXV<mFw#A
zHRNjYnGUwnL@1FoiHaXh_?T$od`^&UrYJ%agssm4n$Sch^#{ZO;r-|hAuzC>j3Ul6
zD}~LvzUk37FY1}VraZGquUHhHGEZ4SoU6{{$@Fx0#hU`E54VgKb%HTAZ*I)MRKM4L
zxLEs38JkG7J+aRvO8E;IhoTuLf!vDL#b-ZhLH6E~da)GdP?hPpoCp>XDK3gx=R5K8
z3KiD}4&LLDEfcP?DJO=c`Qqbz62x}#*D<_89PIf}VJ(GvP~&2rId@;S+NrlnjS0F$
zu*ta@F6Yw4<8=zD+Ad()`Kv(swL@x)OnX5GXp0GCq9A=SZ<xkcaPm7s7tb-<?ge{<
zQuWmw#_^4OQaKlS7-n+5n9j3X!o@3ye5hDlsC;@*v^VB^6YHatBv;xNNA<FMck-7*
z+)7hcDx1Ef5HR;f`!_dmf!I~^ZrZXbCOiWCF)W7%b5!#J^!Ptqv!6T5Q0TUNYLkva
zavIAJ4|Ayd1n>EYxQB9mF`u^gewD0s0e-WSJIfuqojOKQ?mwOceB^Il&17wI?Xf2H
zOo-?AW`+i{zRoLMNbHI2DDlWqzPnme8XC|r#Wg?|FYO(5g8SX$fSG3K7A}EXo7c#w
zb@IVBkKHnL4^$QxE=#DUaGe{O_nIlsK_)+8MG7e#b726P=S>ASQ%P*!`^~>ZapwqB
zOhbCV@t10s2mV&z!Jgq?nXDWHe``A@>+1aDN^GgjN<p@_J)7G>-$`U0`wFPSEPhp<
zs?t@%mPt9*B^{9$#mUAp4Gabx{!>yr&xeqn9IC-SW2TMBiC8V0_ZC3}b4FWOZCe_A
z$S(h*6D}?dqn=Fvp(RdQVgQB<Tr_9-7k*~Bm7x#=7xD*7><GauTie(05?v|$+yjsJ
zIH+iOtLB%z3~$mk;(Fiu(zFO&>2BK@EH~fO^*Ofbc~72II^yj^{@_V8xx-orA4ht6
z0&>uGS2E<F8X}CGDk9sQ`wYE83X+3F%HI|8pI!8lB==74Y3VvH;A<y^_xS(Vd+&Iv
z|M!19qf!z|WF%>!qU`OIj0$Oxoydq#_O8>A>=Lq-WRsm04OC>0Y(*V=?~(8II1S>w
z-mmxP_wVohPq$~zInT%AaXqfbbzP5Z+%Hg<1CB|IPFR)bvhlj!J--=?h&T#jDV%z$
z6mDy<Pl3tkqR!5OI}_VaEVW6rf&6W;Yg*o=%rLow!Ws~0&IsK}$EBq~XZ9EbgPEN7
zk&@=m4X_JxQf3)1={zmB*m`NdMj5cI+9Kx1UHd)F96MhsUZpL0DSqq;81s@j;Lr*7
zI$-Ru+w-wVWC(S-=e;L7PqIRHu;`qbhAaOmgctFwDP7%L&NuVJ38}^}LN1m4U)&US
zKAQ!vpK9|+UK=MP!qy$Ydg@cQZ`utdvtkHg7ogT!)90(E9tUsrK1jK(&RFnG`G}5-
zZAVPoZHnviSy-ImZste#hq}Nz5vmzCog{Wk2#;3Sx#zuZhtMe&aq@nM&?!2+#)yUp
zncu+y=uRkt%DHY4;agy*I@`g7NVeSz39nqZ=Y5aJU6$sWfk_L{D$Ym#oLc88kBP^o
zUMdo9e1R$ljA;x_1lT@v3^NsYqL>f6^VFx=Ep_?@?gkrL_{r;v%v@6@k2uYaXR-xP
zu$K;bhyEyTjzM|Cmk%)sP8NCZ#%0={F^c5Z6}!}yXJ>Wsj)=KD?V?~-+vbXqf)b5(
zuwr-`1?)fczc=`5d~r?<nJ?TK+x8bRD3le=jNwk_?5nR~y+%J)(70<jBv54RK0t@>
zH0skvNwv?6pDK^AT#w`{*XJ@b#YI(m)>k*FR>ZuSDwoCP%!nIj^4K^EAl&A_MdkF_
z=0?bX5!O4opcP#PR(WWJRT>v}-=n~S@1e_r6X(GJ8O0I(8uEU4LD@Z3lDh#WMx|N=
zBx=g+bV{FGH)6l5fQ$##p|jJ5r!l3Z=^xbI32`C{$!k>#{%b=`-{>aJ&X!30NTEIc
z3L^gdLvz3>UE8Mf3Q(w{F68$rQ3Sgc&Dwruv2Hh^nRa(i6LmZ?vJX=K!;Q>PhnK$K
z&3lOr$>})dP+n%uSQr%Wq11qeaue(H<_3=PP{6{<QLg#?I=$^WI9(g81DtAt+iHy%
zp0@)Se8vllJnG!Gsvgr#hs;_!?y#;6(<Mupdku(y%2T-qnEL?~F!C&^pZY-Y2qYLi
zhs>R%Ob2^W!y_ZEnP0I_0D<BqyNLuz1p*}-*!AdhzI{KnlRKT<yHbXr9VmQ{AOwr^
zk6-GVx3)QSf2e}%#!RV|jUsdNJ!MLPciEn0#hF`ktz=?y>lA24#F%vNuQ%@9(LP-)
zVC0g>Jl;3{<$>uBeJc?ghWuLA@BMJ|3{Zw^gF*qVQA1Jt<mbElniw1s0P>*+5-Vn2
zE+*EjpU>%ya7U(I1j2!hOowoLus*{Qq|0GN>_M*7R_)sqMg@H3kB>z52%m}clG)DX
z>^=YS!mH|Q3vA9cpcgI9<@NLTRRKc}99NvXvsED`pERq{tn<q?s9tMurc0)KUp~5}
z@`1Aed1>aDx>LNkl==wc3=~SWKT4=+w`JsG=)<NArfI*vbt3xs4&8wPXiH-7oI6v^
zdwa8Nec>3CpKA(;GW7+FECx8W_GvYwntjsR$3RyiZ0@+PZ89?}<7#{z!?F2#CZb;D
zg)O$GCtB+_oat1Nn0Z6@>TpjEB{)+vsO;kdDzs37Cn%(Z+%fyIhUxOFR^vuuk^P{I
zy;>c^i`qS=2ZaaHYxeo=?Q3^?FqB|xCsUhLRfXsAxP|t$3mVZ(ph7Q%_eqbH3hTw2
zDK9QLPa{HrjNcOC&j6Y<6|ng&nv(*E&0R3mBzSBcBeSbscRTG%5>(U9<~dD(wlI6U
zUIYQpTLRX79j%cwV6vstxUaFwQ<_Fl7jG?cH+X0CV@_sdJuhKw>^oF0wa6;RL`m}Q
z?LOUh{RMPoK;h{O#VLR=S@6#?5;)F8AuX@-dIqkiSDKt_`vJY?Dt(8$w2nai?ajh5
zR~$wZrSr~+_noQBt~qGPXiXjDonr>Y-dFsTl<B9k@2|I}J*0M?yb{fdg6^I!-F@r!
z$Ke^Kp5+JBI)sX=y1%$RI8sdPaA!f6?)Lp$%6B$$cWjYzc6l;3%_#`|L5wKJ(l>sz
zW>hm#F3_h;Epl}4uo3N=0GDD;eHn5afBlmDAhPP%@cV)xNJYpPnt7SjCaBw1Y@;VS
zY!0DGUZ<E#&ptu?Cs{dM|5Ecokkeeud9Fj9+i`F==S}h+8P;~BD<LNygygXtj{(Op
zVJF*X!3C<V<)PByt%a`lc;1@y-}W6DZUs!x%K8h^k($Sf`BPCB90XC)(qi1VPXX~c
zwgvnO-u}lsgluZg-CIgSu{=oix?s9hz?#=&@>sLVPWN`&zJ*t!Bi~0Dd$*nzJu{am
z<ZNL|yUD%Xbg@(7$8NX?>zjD)I@MwK>~;4Xpl_U2zo!nYb}{?3hgiL!&L_4vZzqb|
zO}C4jaus!srz}O4sPdu0RxK~n#$!X<dqoyJvs90mLL9THym2D$tH3NkrHoyHeKFqt
z)w8K+bJY=n)PzIQ7g@C(5kO?fZqH&yI_Ya-^#amm0cu9EG+e0(bS`F-sojPLoTI%b
z{6npqVrd<j=c>EOq>2Y$j~pulxT5^QQoTq|Yim9(NK(_nFmi3%IoD{c`kwRjbo7nc
z=y9-0`uwb(-v7kcG$7FD)m){Ow#KBs(lWW8CvG@@>-SdYa4~=w$OuSia2Rn)-`Q@A
z0IS{>Y#Sl=9IKEh_||`Qn{B2-pi^r>O%Nr|a7#SsgZuiSsozQ}H~cbO!Gb*rN%tB#
z4;kR%DFxtKoM$t}QO}%4Qn&54UJo)hUMm(#(cbCap-XpaC*&tJ1-I`X92R`KO+@Gh
zlrto+wb?IDIvlxkA!gD?HZn4JSjFF{=$K*uFw?FMeaf(R#LgbN_xJkCQYW~-u*ly{
zd#BT+i~0w%1F}W>81o-EU2RSg%!lBQ0ljGsfI|g|i;h8}S^ng;QfQpz8=L>YH%`z-
zr~eDR*^%U|-NPz{DPy=Bi&Hm@!vl)k=>+&paaiw6x4N%gU}|^n3L;Y}zJ2x+vz+^r
zlBtIG<;^1<3XwyJ)X%5oO5E488j4Itt@qpxx2McFJxA|i5?fWT&q9~Ks?T<cOrh>s
zHAA_F-G*`>v_9&LS~s%A{e=;4gYNq8WOdSHhW*qx4t$l`{>fTD$$)mM9pb)5Tz)ZA
ztsnx5r<6qH{Gz-@Fc<PPv>~sp>{!Fd+ug16pL(qaML{ZOgsObKlzmX9KJ;jD<^&{S
z*E4$<k1NWBFj{cl8r%JP?x554&FDVS1(<er>Egr{Ghy-eOU{!BO)OcXbhn`MqRB+F
z`MfeTQelj?{c<r+?%EzvhJx^7GY5Qx_q6k<uYZA%%+AiK3P_>lkq;NoZ{=)w<5Ja*
z!k#S|p7qgRpG{{?<>ACKCgKko(}cRDyyZDoOMTYl8-1VWZm#~TnS=b+J}EpV^Xr1X
zWz8zQ)%mn>%ccC!>=}(_FH)Fk3j9h+PrMcLQn9&jRRkWR<fKy}bi2NGkAX*nJBfTm
z-FQB_u2f|w<U)_%KkfK*@>YYgOj!YBfu?7ES09>&E?ogS`x{UTtSje26P$_+*tBAr
zy4$vzsdN=u#0!siwg)H5py)tGgne1seHq83uL9osB5%C&_toFkO^j$OA69MZ{bn!S
zSYY!epRV0tWQv`5dcuhRY;y|ikePhsP=es%Av3#Evmm!6OgS>3<oH89`&_PqP#FNn
zOEh@C+L?H6vXko`|B`bB5niZ$qDlmg1UU3GMu{1jiHQAZ*o-=a`dH72NG%`$9Cen3
z=Jp@THc!Ob!gjhRdwPK_q&i3c1f_w$%6#8&a6BX`-}W<Wu|4FZdy3@O6k3cM+}mL-
z3*K9U*5)jGwo6~75}Vw2B<kM7p%xFN*GfWt^{}a-((Z?CbVMP?!wesl6=$g|w=D@{
z{8DDiW(a|e!SwrH$p>_xj5K)S+uDJ`cW;Fum0!mMFc>HmrJ7q^l+`DOAVc|N%7Man
z_sDuRAwhkAY4f1^#HId%lDCv7-A5QeGK?w5rH1Xk-oBcs^|6v|n2A!O=&dSv`7e~-
z>kOI@YExT;4iyW25f1(%R<pyh0yom^kJuobwWiIis$TZkY2}9e;Wy+NY9&+W@^^oi
z%sYM=uH1=1qxha(iY%aNt8~B%YxuR7W9feEK@7F<6ap=Z&8?|*rSd%djpON`Pw-fm
zYM5ly2^%Ffn)!^Jt>glgD^FsQXv%Q3RwdUxG+O=c)xL~vU*3Mk9?{}$=`Bi5oaG)J
z8p+k|4`X`Lk$^2zxY21+7cJ#ZTUXDgeiy5r!81BU?FEUZzQ$a}2%c^K-Pq2lfzOBR
z_K3E|v_eSf6>YKB3+<!N)e*j4(%Nap9V1Vx$da@mS{>L_YJRqCVcNuQPj}ci1}yI2
z$w(Zpnzz@?bqJW%*Pq<SoD%F%E2PC8JmEU@X*#52KE(EN>!F;0K6<^yww}5v)%jvN
zE*`_lv&{!BSbo&MBlqrXbW4%(FHE!gc<@c=y_S3QAX6tsEi_Mm6xPd}={&e}(bGf1
z_Z*++-8I-Zpa|)bE(=5U7}lRyYP0B6?`toW2#MM5$Qp1I7>QgH2RPO%;Ob8vU>2P;
z+(tUMPqaQ&ajSxBXdETnipGKO3tCSZQ!>1JJCxEqgg_mdNcIkRq*YNcdfw=oPoNa`
z2x*VYC$lLoUlYot6ecp#AtSnn%51_c!}usq^JAGK>pPD|==rNY_vN;9l0QRIv72P*
zyM-vNS$%?Ph9;X%gS)kt)*a>GRxaiy*KFD+HOeK>NV2L)l4G}sC%ij;y~G<fXGOR5
zDPLb_NpqbiE3x%?UiJ3+C*q2co1gXuXq+Wi%9^{TL0aFr>m`|7N%`IZrciOw>QC(n
zT4kXJMLawN-;uB%3GfGfYHe5zVZBFRU#Hu*TS0hU&!jvN>TOIidM+dSy@x#2F+(Sa
z%!^%^SVH;)5Jmh|szBEnx_$OO4|Hkpn%AQ5lWm><(cUZYUM1O8`&yoBYO5{;VPo?v
zZNTWdGF5#YS7Kj<6n<0ryU>JePm_bhpGQQ#gBMXr-K@;sa+cA|f3h=cSml&^fP}Su
zZ7?PEI<JW*DKa8l;$E6IHUMBCcd2-~bpgS@)eOqW@@qY0G7cN~Auf^m#nd^T&eqh-
zg-^}K+rGr6;$8cMuI4CX`?T~x)VBfV6ds3!)(~)t`Eq{R@^o;*UOKSA#(u`BexN-^
zB|JpphI0;pWE+dPc#EZyhIKmjJxY`?39Wbhi~v)T`c9}uUlqF%0ww3aPF{o@7{PjR
z&r{&8hbk7|_{rTvM)28)_gLJ6*F_}MBhn+b?hYdEpM$Tao;akz?=abao$@MzAc|Sr
zsdLG`KLs^I8*;>nTp7q{Na}0qeeDVP0<ZMMnuK=I4t0KUN5Mdp;}^&ox_vf2Yr2i&
zb8Ejepe$H_N^hfnGK!EM&Pm89#7z|*pKm^WORDR4E`WO5sQ>tHFO5ltz(`kId8mr4
z_lp5Y>%K@nnYGo?>z){E9<zv4a!zmDj)9;`rK(~LjqyT@O?zoS>XwX5Z#JH~6h~%W
zOd^Hq6V)kCds0oZrZ>g;kK}LvGCQqc!e0UIj(jy0TR=xjZtg<4S~>yk);MEFToGf3
z#P}0R&u01Ev?){ITCxO3qi1gWJ!omuM8Y+)<J^!Tt*zI+#;-4-@W#mL)94yi(TJ_N
zbtRH2W)98GYS-%*rocggoU6&-73M#QLm0{^_Z_bY!`iRv*rLI&BOQ>;-6q&mYdd@x
zZiSa|0m@SocsY4k6n5i`1E7U}lC)WN$s@x;i?)=>X7dE<qNIv^!LL(`DxLS+6&M$!
zo64P1=hx8})JeM@c~h!dKmAppeOgo0+RAkQ`Ifl$l<%|!yIpV)vuNubzrHA*>d=jF
z!TiCmm6_z}3pTAlpQp~%6kol0+Bx@&^Q?V(8LghUPr(8BXi^q*pXy99ruphtd0oTz
zn`?VDCHRB<u~XLdbe#fUpqM<1Kbo~abiy@7n#CHTx$>b&xQ(I)b<Q(+_I*{KZViv<
z)Y?9<iRQWlXk6=!F7>=DdAIvi^MXnRTj!KQazTagjJm*x-c-t4n`q#@Y@N{55nL`9
zrXs^kyL(G;_6d)yK2HAN6kKXOgsDQVRIQ^6`!Nxfz{qzw!^=^;C<#MP4C2T=i>E}A
zY~*GM+<JWn0Rr*@vzM!)YQHlB<&LXhA7w1QQ*UP!+f-%v<drIhqx3f8!ebZe#~wwE
zWNQ`?s5CN|v{c`DuhAmS=zQ%xEy9e=Nzom&>JTDc2TWt-Rup_+Chy4&xHy}+vwSzm
z+i*iiF||R&LfLBY_>r%do|2WdLpV62;uD0ELs-iM&EGM(QT4)AI$RtA56@Nh431yA
zmFX7#rHh+1OvZ(N*ErHty;SYGBZ16wJ7_6`iU!Hf+?y7ad2jdC^gaFi;?U~&N+bWd
zAKwKEV#O4#z&lb~%r|m0a@xyl$Q9ix<%1m{g{2xGP&Ihm)~9{N#yl9hmm<A$x^TS_
z*KrlVdt_XvRsk8M>7ejW=d`1000fJgv+y1aZt_kd55Xx2+Qm)?i0kgA>?`7|Ia6xu
zFb`U#8cI9AAG||R(KCMZ7I-fv%m+R4Zti;1f5bmUdG~G8ZLVf_wwjc<^FVW3z#`wv
zT<_6zJa;n^z-$3X07Gd+DPNZ7QXV*FH+K3ExAT6_)-Y(F`c>}9m-{9~<hOQjLKCuL
znoMXYuy(%fVHHGXy4bN^LKan^YN<xM^0`gR^VRQ1NL$Y!vN<ua*Co+8-urs-(C3sk
z+Jc&j5FwuY&K*Ot$#$QNg~Og9m)szQuGkUlG_Uraw}SP%ij&sgwXiLD`Bft2c@DsY
z)#uAcQeDW4;e#HSInLDB3J5<WW)|B)nn``V(0*oWY0rs9RK`2za}g4gnsBW38afFn
z`>1ty3Ke^0(D8Wyeg#iJwDnhr0;RPo>(YHEODzeqiI#E9q&ukMvstj?<_Y<C3`%(_
zPJ=Dk@&y-XGs4b19}0DL0zXC$*^Uy^Pa_B4e;!J<y>vA1h4`!BEWo;H5~ZIDZE-Ig
zd;cs&L1<>{=PRGaApN##yw5_cI@9W-LP3V+vKxRc9D$<vK<2G1imV_#wrs0WT)HHK
zfY-rFV{xCOIw4WE^iwJ0NeyC2fcYydghM3r0onzLFSIPH|EXNd+BE}3C-}UG9`NN|
zKhz~OSrk<Dn#?{4D)TY|jHP4vv^H+%@(qEgYsFkU;Pm1I8d!d*7HwDJdiHy+L-k8H
zNf+oZB}!_Uh2S&+RVXd{PgD*6AVp;hQJk7QyRmBI<4wv1|6jtw&`%CRMOcRNz(E29
zp02;u*G)_n<i)xxA6fW?AaB1{J2dse!=|O0S49LMr&D`^YXcP<iYcEUH&$7i5wK=p
zApoNLA!woTIjDd4djJY&=oD%~tY%fGbc$c?EqWQ23R3h#*ArKWZ(a85P3SK6dE2#L
zz!lmmXvuqrmStMG?1m*X+1SARF!Gp=#^X!D4D{AKa+JxzeoG4_EEHQ!px)Xg%nn!8
zz#TCrPP%|vlw&u;8fB^=Wc3)sxwJ4$L^SNT*tgp4KM4*_tiDHU*&mrMb8A0AIe|y`
z0Q(1P4ZGaD%|iT-$&DU1sj*mFwV{q1poXSDCbkc|2s{HgB<@ckLKc;+*Co*5-3}cA
z#@Nye$8PQmZT?dZ2}S}`5I1)HT2CAs5Z+BmKp*`A{S6xjxrrB?)@_#{B9=t;4IBl*
zOG&Qq4a_<$fcCSZ%-YWK5FP^sgs6xrUU<?wXz)S<P`$?t>tKHcZ%-^W#>Qj=IZBmI
zNB$y3>IMjjeDC^KL;26bu_rgo584!*g7gK89JpZ>dpdl1bN323=lxO-tH8+TVWaHQ
zzY<p{K~Bz6*D@BAm^%}HpZQk};KH^S3XK;pCN8b&J;2vnXdQ^W2=6!s<<(pC_+@|7
z0}W4e79XC6TA7PE(B!maViC3d&gr;CUJW%-8t<?7@S?9>MzBp*TWrhu4MLV+n{Hwj
zVc~NI@p?aB3Dc=(AG@$sY$>(%FdOjt(G$4{$g3A(LBX`kcojNR$YH^2F1*94_RN(i
zKjo9CJFuG0{k>=B%^*!mau($*(Q;{t#yTKJbv2-pgguAs>9V9;0W_^~RUiwEN=J5u
zul(4u7pFr2%Jf9nucYe;ZI~9h^5ND$JsnsMqdOnb(Oe2NEx5S^Lw?6HK7wh1F~~mv
z3i=9cL}1Pl2&Y*>C@)$dzehf>+-3_Jod)G@rdxxA3oWr<=EXe-o1z!{?E0@4?uOAn
zp{{$s2lzAKcEnCpBdtrI?nsItG<ToPkkko{GWTcHP6ynW39y8oKB%kc%Wv5Vcvk+3
z(roJ$JF=(88p-~=nT$pg_PBdn?eEk7^rqEGi(7wW6KY*dT=9J^jN*biy&ZSVS~BF&
z2?;7ZQo`i>y&}B*{O&j9=(6s%BP?1j@EHDz!hwXP+GQ+mqnxeeQvQLD1rRfYPIl_o
zG4mo#IA{ItJL}StqX5E{)5?y*3yBvpQC$A9*(D_b!;X@_OZtT11K(mz0SKW7a$gi`
zqEFLVb1qsQMqL>4Q1Dflo;1wq@X}(bwr2wUUL5H7@sqX~2^e@%QYm<e)NY5sE7CvP
zBSu8LmJa4%OJo(kL=7V1y;QcSkPgI+Kk}X8QPROf7yY?v1}b?kHbE(O&3eel7|&_Q
zPa@sThvz6mC_T43e+q(Cft?yli(DENj>`*$z)v4km;Mw7Y4aLUMvTy7G8@iPx#kJW
zC6%!>BkZGxg=5{kRJM~3eo3)q%U$)CaVQJVUMwC&JLdu$yAR*d=Qo@E{-qrp2IfRa
zI%hVB5&tD9mL~T5TYvtd2h-zJm-4)~273(M>EkqnmjuqzEsYz-i;SYRO)5(lP$Wo%
z<?{f|OI#uO&c+zN38#(Z&uGgYy3U~E#6etP7xHBG-}_%b0+E<~bsHZuG3iDC%G8K4
z?;+GA@gt1q?XT96&0W)mS@N7MpWL5C0U39p41MbMMEEAl@YT(PzUCtX>|PZIY<%6N
zcPkPEguVrP^7A7_e)xK-Qmrj0_D3L$pL~o;Bk&Urq?Og6nd3(}2a+X0_2)G!9+n+n
zuv&yxe2SkDh8bB=@vOrkCG>Hp2@Lu7yfKYPi7B7yg!Zb6uv>RI0?L>4_8;Dh@jxmt
zc5T=Jo&kz~Z#ubL#DU~l;`C*7`yYe;k3s*(pnspnRR?9c1T6~_6VZf>USF0Z@G;R`
zG9)PXNnLhbL#3e<{GYX}69IH}!4kZ$l#!<-b8)}4yhP6r8K&#Ug~3UCubxZbFI!;Q
z0bM3L9-Zhob#eCWEQIg5ugFs0`GHE``kpA*;XB7$QCcBG;SlNf^3Y&Z5O=~r6P(o0
z`Q7^Ka)tlr-Pk8&Es#Nc1>EtzgZTh_@EmpPmh7o{VQT-8J^C3+%dZ~5JG#JyusL3R
zkbo}wEopxXK{mg_qw9A(QV{NWc0yMeGDlHSLr7oBMp0ajs4%ppr^SNrPXGd(S9}%3
z248koqq>l&4LE{wJOW3I@;7Rrs7o}q748Q7ng9YOR&X(l5gKuj_q~cpb)&UQ@_qI1
zs<nuA`rUr<2b_?Xd*ZerMA&LjU!K0#8<r+!n|nN8dn_6{4O&7V`|_1HLd%H1G{5fb
zLb_MOcP$}cBX2Z?u4u?ss~8HvtwXW+bhtXo_kM*s$xvs=#Q&b%H!Ho8N`EB!p9eON
zLIAuCMr~0}ETGM(1{aID*k&wFBR~cRK5N!%sK(0MsbAMuK@(~?s}a!Af1n%ctds+p
zGy5~}eFG4VCZ2AN$Q|T%AW0?59emc|HRh#<0&s(U_>z~UHU62`vImwaxbDtfMF~m0
zlVi|(Oco*KCY7)ZhxCnhGZ7K1l720<2B&zeLoY%G<V9^NJ_#;sC-8^$h=|3uLOA8e
z8Fh?307UN3umcZa6daZagDH-al2A~oW<Opfv)^h8W-4p(y^Q_i;qgCsoc~NspHo<A
zMtA8~oaWYc65!w0+67l;7T-cO<3v9MmT>|X28@TVjeO1@LeS2^YHdJ3b3N@p2HN-W
z(TUKPcoi#&;(eYFB)?Hz*?B`^C~quIR%aRF+&{?=Y}f|Kyk<uC7~iH~5gSwJdFpO&
z9q!DxXSjt+bpR@P6|u095{Fb?Xjh9xpAHX0-~!mmqE7^0`2qso&k@zPyaQZ`a0HNX
zlGq8glz`jo*s~NaR7?hCoVsp@UoQ0%T_w-{5HP|Cf7gP0mD$OW_O0IFwiOrvJ{%Ab
z?l?diwtmx~HC*&$IshN!1do1MO!G9v<a|)yp&u<O@dQuDW0q@xYaWf_#a5b}a7q!D
z#QExe;0X0YwetwSiD40mKy=GGnX?^##r@BW2d#v*DhTe{Vg7!XRj_KF?e!pGE2FHe
z?m2+SciL}F0zkTZU%&-A8Uc9*S>^`v7(7Q6EsCwz@Q6Ljco@HO4$c>-HXmh^JbVMO
zlwxr1F0iBSDyheSqNmUeythdd1oIWh>`6@U#4g?GSHA1dHVIR^YKB5}Vcj?aOUQ8q
zx1LqFZcmr^!YgP7+5V<!KSo=d7!}b%pU~UAZDzlm3?zN-jD)0@qC7%v6#f(@mOI#i
zeCw9LNy=NlFYcTGIq1(1_+mNhP>@y)qLWcho5KEimw#_}WfJCcFRS{6p_jkpOOuH?
z8;)F*@@}g<;Hus$TwG{hj6$^`;C7Zb`axl4E13fU=9=F4<Th`0W5Nfp2Q{<;ADKf9
zzzLnb(D_?1e2c<6WKk|%(p!W+0D9(}-R?C6x#lmjVh<8GXjsV4el8JLo~lLNyA}Yj
zq^w$wvmBrSrRXhht$|Y-V*Zi62MXF_np>a`UXb&6<j#p+<z=qE-zo{-T_}i!lfx;c
zoz-^a#*g`NRJ5jYGPaSk0@}?v=0F#n8W0ghZvRe0OvL>{lv*r4jHP@AZ5HANRi&C*
zh1S7^B?la>!q}dg=xCJHWmS>^uZ;$r`~*u~eilME6_*1q5JUrj&cvJUhj?KIzB}?>
ze-)jVp=DLAQ=meS-=krmN~C?J+!9K0nkhtNCaSj)df6XFL=P+%iQQ=rokXF}9Ee5r
zT3-cs7Lr0Uy_W86zb}@ci=nw`<fOivv`l|0xEKSTHROHgQw_4fXO^rwltCMjFq5Zq
z)oH*9S$tPy+C2>E@_dP~fFAfDzHby#km4stKLWUx?|-3CZWeCGer}v;2exB$2}OwR
z%n%FWwd$=smJJyJ_{%Rw8H+j)?Et6<iqq}a-cr<{5`iWXE4&-{FSbG2#gcUxz~Jx=
zCK1nEwY8g9uxq(qVpZ+flYCz@BnzM)Tso0}H$FNFiC@phY1b{Ky8b>W%Q)vJNBq*h
zP`n`q$^PdkPJ3IRl=`S_jnLi0A@Szj06LIQW0(<tvZW7FC}qRvT5$Gu>d>Wn)vH((
ziLka$zM>{isEXGm(lcTZ+}o-|p5^zDKLkNozqmqzoPsDTwVJVQ*rnX8>*T-g`ArOV
zEy{}>heloEq2Rxgm!W37kF?%hCIIr?Rr<UJe2LpWL(*$C9LK8h@z39cj#VImSFayR
z$U?+Ot>pw%hNd@FklI^yUK>ik;s@sQAU>jeo@4F>YMEyVT3R4q8&$WhEt5dy@2G&H
zk<kW~5C6rIt7#iiNK9=h4rRn=NdtoidYL3ep*pl07)kwe)kaaeOiN*fo3Q;5p%iz5
z=H4cDvmEgW55=MFaWtyWGy$K{p$Z#fLIc%NO*d9J`=e4=OE8X~FV2K7UJ>f1=h1%#
zX+$RvXGPU^Z%{o}aeWfuOaY$WJzfZ8^Mav~UdCOTew{9w!0mhGUXF56?ZO3{=CGsd
zKo4p$KC)<PMI+YF)Cs^qN_vo7ieV?k*UhwDelX#B?|VIj7SbUgAAREo?uG&+I9Vc?
zt@={J)IYiwK#QWrA&QqjDcy2$akilcI-H({%;JmO-uvFXq|>s!^nEN;kb_FZU?)J2
zn>@&aiW&vugcNBtaOhS;dS<ZJMBIf8t9;*ob-oZ0zooSWbk!O(&N%o0pD+FxT2z2b
z%SClfx7Kh<K()zC`q23STto?O;L56`!$MA?*Z^dSIbBhYB)Hkrs1h9<nv_q;v7iSy
z2LSG?2V(U@Pqzj97X<#ug*qx*0%q8A)iixSo&M1zso%K(FCZtKE%GQ8i;q$Vv^10w
zhRntqXc!hgu?0uy)L|lG3sL`Hh;hFh$5J}pJWolgBQs#zH;m*7bNcfQ|H`j7F*Bs!
zA^xpnfP=z8fBkDP1@<T`L_WL<QGp=j%p^xao5zUI(h!0-CF})KB;}jPX(bkVPM_<!
zT?l*}<OQuMEur7b_!BGn|AHuP`(GgB3y2_#euAuS+fcJAT`aEJul6QD6UuZz*@S%i
zY<*TWCFHVLb6Ri`Iv(3-T|+}8prLvc1XdH@Y`o2APa00gsqy%$ls{p)`j;-iY+H-&
zn5dQV*>#~Zz3jspMS&(@4$Ez^g~X?7B>qhypl%7Jyw4!S5QO*G|MdzaR*i^JWB8i!
zo|+nzLcwcXpD>Q_6E3^33?uwNu%(nT=v3MOg{cQXg$AhXzxWemv%d}|_6NBffKZsE
ze%OaET1+#jP&gN@Ft8|p&v8ApFsCaV*g*9@x)6on5dhIRz>z#dZf9JW184Y-x&wsx
zz@Brsso_;-rYr<H4@9BKpwgfI88P0^`Grv25s1X>IKn!HjlFRaU6bJG2IQEjkpCAm
zCfbNVX;8%|DifA9bLvyCVK7d3fZ&vfCRui7LF`9n$M~{P0-;U)*MTs5!6$h@)%x~-
z;c8g5S<7X)4<&e29czzrE&t<cMfdN2IS>6WvLEUR7!soS|K;T~5RUyHiT)#3{EtL`
zi3viq`X7n@N2349qJQRr{~w5J4vt}fFN*^AP*P&m4d~Eq)aSVWzwrM2rL44lhyh2!
z+$YO|i+E;>E&IjRDn{blRJQH84RS*cQ);(g`Z*v4i@6U9`ILr+WN870H0<J3PJ`Yc
z9{BdhkEsF!@%H=ws?t~ZE4HT>_684ecdO^%`+aRXqg@b+d$wR5$n%c*B_9%bkB)-I
z-K50p@&;n9Kd)FT-to`dkDoFL5!T=waxK<D`gxLaz9*{BUV!JpfCTFemysU6*@e&7
z`Oh!<Yt-d2<Qzy>7z|RxN2z_zZn5R3Alrn3)2`E|@~iKV|LHH7izq?nP23q+cb760
zYh*=L9f4<>lRyK-Z`F;d5MmEVd|JE&dH$bX_h-hxV!l`-`1ytSHfgZ+vCi;3bxJvx
zlp8|c1{m{=h}h#&+OKCWJm=jTHGbB9rXY=ZH!f>-NGnUA-`UH<pt7^P_i@IJ0bzX;
z0b>dE8LQ?3=>?Ol8;QlC8Tt1I#^YP1lT_z~s`8sBQ@uot7x#rV7%}Ua*ud|al9io3
ze~1~+c^0=0Sx+Qq-^|JDule}o!dT3Of$`ZP<$7nO1&61a`*BOG+pWmLTxP!7uCs*N
zm6q(#18_rBJPvjptFnAJJo`w%^L*WLZw{0^@kECa6=#_58xFUbFG!yXIQ>C2*^;nl
z|1F*1Sm|8!Cjl+9=^a(zoc-%;ATI@J7`cY1=6*N;>mk;?BO?0_qU}B)VK5Ek#wb27
z92?`BH&dI79CL~3sfQQ5&nh$y&|qEYZL%=0`*50Ry3VqKYf<EaQ#Jj{_v5!|(>d)0
zo1_PvLOt&<S-;L-b^DD1KIN36r$8okAk2s7oX`XG!yuF(!Mr>_r|{<5(e1XGrDHeu
zW19(GEPWh^az${jjtj4JBhefOHe<P4Esr=s7M^8%f*?l?Q-rK=aiUE(Uc%zvPS~IZ
z<D1U1ctiWz0?n}XLgX#(_%-~!RajB-&#gigbyWf@793#+uaSy>`~&We5hAc@gYz2r
zgTe*LKEWGSu7%;}T5b_ui#l3MCa>_yfE5ku4suhf2IZPrq9*8T!%iWhc5DN2;EOG!
zf;Iih?cDwXqeW8jS~~zobb4JDN?1fkVfkWMY>2IG2qodLy@V8v*W@vC@O{H`Frxh7
zXvDx6O&DeC?G>YhkyKdQF%Mt91_fAuy&pfK`F?F1T#MiFWp$zWIbd%AS?Ka^<{)5l
zewcPMsf93_VJ##j@*i0)Yp_J*UC?xa290v(A4b`YMtO8hX2k@sTaajpdF*lwM%e@#
zMDG&<j}V=O5!JT{VV5WGb;+{~VGVLcaVEvxOFcK2QSJ7Y7rA2*&stZWx0r`OFTXgG
z1X#c9iH^pDNaSL*fQP4!ona&uhTNk;s&YN}lq~|&MpALqC;&&Sa8Q%b{V(DGY`skp
zPG==CF_LJD{~L+_=y(JfT1UX;RQh$xrvGPm0@wBrh6H|GrPi>SAgG(anj^fqja(pZ
zrnfL4t}zEr3vuV}KCT2*i&wMj+=q#9BbFi)mC@`U)hKGbLMQ?k#|~R!RNor2;?QGX
z16?SK8#Q68d{6^}yz|E-;92o1JnJ-^9wJ9Ye8a@OC?&RuitS_vL<dZ0>SF&)f7t_#
zTH#>%ukwPw7+eSv-OBB-mI&2L3~90>G2hW3sAruUZQF}6v?OogZj=SMZkOVtLJO78
z-|VP(-iWQ>Aov_oZJ*z44ZRmT)V6GB#Ko9xK(Kg4JzbPJ9RvEP5r53`L8=YjeacT}
z3HgOByFyCQ$QUf39Sn?W6mEE4y5cqXlbOC3=#}|01dlwv{0oMV%$#k)O$P%muclxT
z0Cvu1zw31#v~u>(6yG=w#b<>UweLAAltN>R38?9ac{uW%H#|L;63X#|Q*D0*8BN2+
z5$n-vpTRRrMk^A3-{3TInwfYetbx(ivOVNP(oKd~id?k$Tei&I20X~`a-o8DA$#kK
zT#LbzPbHTx1CeV3@Xslg^OC4zAmGaz)Zk;I_DGPCOPNr0Bk3k77y#lu;wBnneTrrn
zxnPy$QkL=0+2YUFb65)h(N)`Z8?}z~NXvyuv2ekp?PBjZBhr0_=c1(WHaZO$q+@NZ
z6KrhyE<lG$(+q27dvR_Bo8tGK?GEfa`I&v8OL?9@<>*OBjMKQq%7C2SO|aXhB&B%%
z{FQrs7%A{wUvrVEx9%W9prUI0W>>hiXkgWJfkfTGFv^^C{H2vgzHkpZ@<=!Sn~?`c
zKIZ3<CuSgc<Ub6S?^}XB!Y_^mqkRo%ay!j>KMdRj2B{FGQssJhra>u{(|o+Qh>?tb
zX`2Iiy&$9A7;L>RHGRweq$3i#v=6`D1>b1d0GiQeZ`H09>xk%rVfJ#p_yYT?eG5(r
zJHF8EpH9j8pQj|6;3@F{)=Jy3R4DiB&~HxwmetO0MR2lj;kYwfuWdm&uHTS-A+y8t
zX#`AR?V%7;P-77!Sl0Dnx-YHm1|zzcC+`0dAS`r9G2-K!e^!Toh<OZxzG$QAOGqTw
z0ve5;TZ0H;4~e-gTctgzy5-=HJOURG1G^GSQKCcQnx=s4xK;X>4CH(;I&Ob`gnbPp
zr)+E40Y5n`2V8Wi7oX+*5i%n+!qgZ}5Wx{5SK&pv19J}zQSxG7wB^!>=1Nlmy|=KE
zw0}^P*M3z}69<y9BOA1*VIc}IXY;)R+OTWm0AH8QmB6sq2l7gVf_>AF;97@;A}A1C
zE`f%dR!$cdABW_YQC+j)ibZ1QLC)i%Y3-5fJ&I)hu|_0Z6`f&KqYa<QETC^*NaxpX
z21LPwiY|o~KA*p3mu!`^ATSEqxI9RDr^vd)_>EXbvxq3zlT)nw(HUPt$iZkMb4)VQ
zMwgg<gJ|$tcx8+iMwJ*$m89|zzOyS>6m7i}9Z0RJ12Ky$hQ~YrlDlh&IUz%N4&FKt
zoQ-z@??Z;&wiSlm$1S#-`Ex}nZAV}3SUFWRiV2vHzohvee#1Z5a{v>aQ}6`dfdn5A
z?OYEwBVkJg&v)5;75tT8Zxa0YIAQ8?v~1^)TGSULG=v>mNZWyDH&!mB5f)Oab*U3o
z7Xe(J0S|r|{~&(#Hj1tXJD6N>gB1bsi;)7((s*!;ab<kb#~8NvXQ~uSCCZ5p^EE=s
z#t)33q|1#I!pmjyBC#G`kt|vi?^tmV@C)mfgG~?3&R&%dP<aa8tYu_`RMZSo^*21n
z(I`^rydU!$Pe})=MIB1SB9ZeD(+C}4Fu9_^s72#4*7>;sU5K;n;$45N{74COUy{)`
z&tNE;fm`~8j8UXgv2E}?Re#Qyv<Yt8rtYO5;H7No7<yy|bB;f%wIje=Q8ch6?4?GE
zD<*(hew^uG&Hx>W&nMObEyhp3!U}!_O<3+AEBLrML0g+)<LA~tl|M@>?SfQ&A&HDi
z=q^aM3F=?$s6GyY1cibFu#F3fr6218F`4LO8sI0Dx>$h81Budk|EUmxg^&u<!b(ak
zmmx3yl5FEhui~CDfU3kK0i#)HGKT`AVJ?A*M1GEA$Dv3#r}XhgazFw=Wvnl>@Tdn4
zhN6TOG=Xh=3xh1HX@I}{`VMvkXrxaH|BplI?uqt{hv0b@p7B{+??Ekx5Pe{WZ?oOU
z1mbt(!ydj(zkmMcJx~dO)`4}8?prZBh8^a9N>%wbvQhq++@$|>NH&Yg2!Rb;cp{dN
zx;n`Uvhg8J_40ZR#DCbc=vopNvJN~R02BbwboWE>WTXdT@vsQp#T3B9+}?Oin0eRf
zh<^OF#H_prdJn9aFYqQANE8k~euWhuMaaU7rQ-ZwAt~YjTRcdSfgj|jKwOqQ4|nU2
z%(j~qkk~q#O;BtRv>#Z)VT73kcP*svF2eJ|!#Wh_{Jl3O<**71WvW(m-QKx2jtcI)
z7CZO*JYS(I@QaE$plt(}%cZ^kxm2IiFmuOxsbs{QT=4At!6*Q}^xGohH-K@44!8~?
z`R`4vL?*gETr+G=@IEgA^B3}ukk+MIMMKnK50}Mn4@wm4jP7ke3ut1(fM+L+Oy3E?
zQ7=Dp#PTkT>_i)bzmqtH$p(Q+SUF*I7l$!csY>ksc?_ZLa1L=Fo)Ho#qz_v|<%w79
z_)W|!VBRs!o%!hCCBw7Bp*S)4(s^`=53&;I4M95j8C4EC{zPQw*@v%!dq9pYYn4z0
zJ$)UPalw&z``@LBaA`xl)Dj@O9(J+v981<T_wQT)_<NYSkq1!34~fpJJrtspuse1!
zx5Qz0v=0+2Z4|v11HKiW<Q7E9725#eY|tr1ag01GKKcvpE)=%J#yJVyov{J1#80He
zlBl?c^_4mpciM%-N=T)Ev0X;U*gh`{!hRb^<`Xy+e{J1ebkVumV;KVT+5xB2=2z|3
zV^jijl4^Mg2MKKqfOy#sjEZ)g_zroQEA<8`2MZkgjWmr^EBG9H7RVp))MFhYe;H6F
zFg675t&)dxB2T$;9h>m$@Flzs3|a?U>aLaR_&<5n;<npv)<LKKDWOx3-GXeDcSU@>
zD|RZv66PLtAeIQ7dnn#b+kp3}{DVOshxCa?;fFXv64IOlNEJdWowbUJR){9#h<Q%|
z^QukotHJX)dcK(2wATk+S4bY7gP$pG!NL%cj-Y)%?BORQ>m@<^t}Zb|0tI!`TyKaj
z0e`tey^!2GzezLp_xXoeT!{FHr0@lsHpV|KDwcQ{@*sg$e^ogD;;&88pm7hE4<a)-
z4;{a#EwgC<b<o`^*V+E!H8?i6=oF{WadN@<fzR%80FHm3JHZ?2c5tT|Y{lDc%PFRa
zU@5-V$$6ufc*3Ft?2}Oa6qXvyJd2fn!MdLwyE_(ytcbnf<g&Q$c7<&PV`KDyC<o3N
zp}MT4lIef)a={R1xN}A6u6^xA-lXMoMbp{^oMrcZ?KOv$AHkglJkf8mkq}vnj-=ri
zO9U(-gz5!Vk6K2n!RK{xSyfwD6ir~wk3jhjRNSmu@~{6n*WJgF@fef6glq6?yAC_T
z55F{EC}CEIW#Rd#Q7#=!@+0&T&}eI7P$hzuk}K}fM5c^4JYRH4FSYRA_eg)mGSI@#
z|KmYKvp2ayfmUJ#a!MZ0nca6`bGT?Sjz)lOf!q8=18E#$&^`qkyog+*gvkZWWhXmq
zx+<Hg8!<2?6(3?KVvt&NF)BGH;h71T3-?kmNBQ?<3;5zNzI>Fff3)8MzP@TxFv#(k
zwdE%q)N<K~bwn@A6-S_ESS@&-QakSeI~wyG#^&oYtrY@Wwfw{`OQgF=zyN2K0yq7d
zeIe?r1S6PbFOf$@3)8w%b8;Ym;|%*67{d(~av`C^1;um3;dyF<avEtW936u*<|rZd
zBlwF%UwOfKFUjNIxRQfl5(`wo!V>2*mfDB?TJQhOO#mA-VhCOW?JYk8BY*GyKi+^8
zjo^o80~$IId!2w<{AYNRkFYx!+!kF=8^2}W+^QO|LR+968dn7#c3M|2x^la;F>nq@
zZdo(Jm%1RpA+-N)B?Uaw{PJ!s{1z|By>1OWyx$l`38n8MRI6Ku=*8E2_0fuRikShv
zExo9_8@V@CK$47M5#RiSElgZ7;jWJy+xmnY+m-u)z3@BvHyJ?|z|K+<wH`t&y3B8O
z62-fo>M)!ff!VuX;1NR}k7!u+#JB#@mL3mt0(L5P6iH!6K0ekJToI9j#71KN14s>*
zhwb_0m###YRV{ZTZww=RZlW!O02EP;w5bMM7wH_$`&Jr1NO&<LgL}Pe&>u<j`=JY>
zOF}Y5l_)QK69&Zon)jyYWNw5NUWj<_^7tYQH(xJZ0VKc=UHlZ-YvO7R^UCdoAALnj
zvE{4LU9ps#7#_3}|6L#ew58oH(7q4vn{HpRxo9Va(Z>JOf9StUJJSVhB!pXAjf}@}
z6?jWYL=dtmrGOz0sM)?PzXxYuyf*?N25i3r&lY-MM(EJy5GcaKI>NP7DBP-QRYo5G
zeGyl*c4wo|brGt#*7`L57@korY`Lz|cr2=~_G8b$f|0`l-JHWnt?f)rx(KOwJ>+t3
z<S-Ilwg|lV`E4WqH+BT<PW@dF^~Spm2#GG_ru|tPyh42DfhX(gKKMz*l6G3U5G{*4
zVILwXcHo(2SelR!f#@QJ9(h$<I2cI&+70w(@Xx0OB4+xxX!l7F^C5{o)^_E8fp7ew
z%i>OlDdHPNA&b}NYxJU;^lJQ%x%upYtm#v1XQ^Y^*sJ`W2Fso>@^c$}!kD9S_v{;K
zx<~vR56`Ba(K%ODU+jO#Z?ThSvSD#b#Nd0`x3Ul(i*jRTO7Dfm1^riJGvWn277DZZ
zINuG#mNw;glrC>3W4XD8_^*F%X0O2-ByA$4!m#9$0zDe+dDPK=*ocTLj(*_gB;E{v
z6~XAxp+7<lK~&_TY>uc%md<JT!y2NM|2*Zu*x+K4n5=fM9B2)3Ir_DQ4i|dUIW*SC
zliZZ(Kl#L@VKsdfjcFRdy#5#0$n7K|O+5)T9{cvZG|l3`<nh%5!O!AgRbIUx&j|dz
zBDBB`Iruft{>fwTLvAT}>+^&?V`#e72jDE_{|FH)T?-SKSz9<xjbVw0$>ge==7{Ew
z>6oot5q_0V_rS8)5@|D7S1$|VP@)bNoor|&aj=re9P?UFn8ck^Fz(Cx+eWD{84LeI
zf;CcWunDv*!~Lc+J)`~Q?)lUN2uS7na7}AyWV%|d+lrln|JrVaUw`t9ggIejHUi4g
zuhmn*MZe;U8fDxEu<t5VGdtwyfP!j3gSl_frD+a4UW7)%u=E23H_vT6vy%V?y@4-o
zFww7tEp`Rid|XTdjTKt`7gzz);KY-hO?1H9U&zbG(QH7Ig|%MXY-wk0R!sdDgZ*B`
z9nG7A<(Dq*>Xlr*ne)`JF7Mruh&usldYH-2@2LW!XA8XcZNB*VfG_aC<A@DphMD>R
zVbiiov78lcErJ_Yu^^hx;f<B8LLTwQ<X2Axa~NKA`?}tF^!^0&{v7VwyD-b+Xus0y
z-HmwAez_cCt2qiY-9iS?7uI%_-x`A%Wp=IJeKHm*X`&pFd)Jj>HmsfsnKZoWaFPew
z%R5T&ta8C62Tc$h^tN90_Ge67;Gz7*T0LQyDFz}~6>VIUIie0(ow5WO@eM05;)?`Q
zt5pYi4RI_`E`-6L{1zGvF};fNEn)m%<S^Le>=~zb!1OYXMC@?BsIKh-YiO-<S(2sQ
zq-n{6)r9)_7T7b@4ZQ@|=j?7cNqprYh(XBQfkA%HstPg4z(nM%D;jp1_^dcOkjG%p
z$PfBz_Kb#v?35uO9bjJ-fRNGW!Yv4V+Jn}{Y*UU1`3!ATPL~iD;ssU27{+3O^6%gn
z%3OWT2AoeN3!_<Vv1L6Lfa{tUWLUi~umhHeMM7so7*~HP3%rjTe`%qNSOmO&VKsXz
z?BK9s%eKHQRQo<$W!(ph*CChAM!U8JzTsOq*Z#!PPNB_*8YoWSl^JlV&mAhCS^X*Y
z8nD>r5@E%)Xg!coR1y2|yC{0GJ`k|He?rfC7aTRF(X~%SVM2v6@Qn|;)!&5dB&7)r
zg<+|KW37-Oh$C>)-oUO4G4)}#q4&oD$)ZmNjUiqNJOm%4H3T%?MMC&Jo$Dx?=}usn
zx!WfKeoKN?<6+N}0U5V_3?W&)X0m7?<FzDmC3I|#0-tX_>3<!Ou^*99v+vVYNkqoW
z<i~{2Y;FP>7hXq|rb&2tAQoQ5;TR6m?!l0adga*FpWZnKvleW)8;AHk0@k&DZ|O_)
zR%>*=k|y*scY*9=N?rSe32iC{X!zqx^|wZeQ_@IBtlAaOe>oz*-mzjd3@_p|-O!jV
z@Vj9oi!MI&xD8*roCWLc*Set}t`D!jqP0fB8~*xC5x#NbGn}%DhYwbbgT^L136c@7
zYgetA3gDnBs#NyJAq_YWWa|&u(}@Vvh_)-YA-pkK6&4;uce4f^8g?4^1{c3(*uunR
z-qnjj@1TjuO$!h5Kmu<CfsZ2+@-}$JA+?kYo00$VIurU9#uO%@K}q)P419rMM2HXZ
zsScvxt(tF*7f(wQ;4|RU(?G$w*bgcz$AK5}A~|Hy8iP*eFbBw?c%~A(Z0!QsQ8sZC
z!vJ51rVc-cUldZA5`wTO0lnh{7xi#kM6N~ZEha?n6^pv;0((-XBBPIZMsORv>Ac$l
zeiya?V|<^*uy~^L?OG9w7?TDL+~m#>&Wazw+^nnKg<kyuNe$JbUNWnhhFcQIqo1%l
z4IPpY(9YHzMb7y7+y{oSaANV1Mr`p8SBq%s1}j`oq7Z=7)}Y@;fC1MKalt8`d|G^D
z<&xmX`tbJhGAli0`0S=1GT<oigpq)O=AOLYPG*Z3f!gaWIt^s-<<0t4-J?t%2O<gB
zFCd(yfC05?1=p>f$9p-Lb6|E-8Pc;afC0-g@Tac>z?cjqXy{3Zyn%jVZ?10#a#0<H
zFTXy-pc!^MP46<ntqKA9HOA7+5zwTKNHmGm2dbhuLbxj|n-;%O^Hi{wd#t)F9l(ES
zGJ=~CyRovtmtRUnmRelP^zc}{??9Jxz<(lsIufhr!6E{5GAI$|IR&4w$lNUCX&^?s
zeG>}f!xj`TZz2^1noW|GTtYiVigqg3(9&*meLxjD8>?O@3)Z1Q5;A31euz~8ifC*_
zulXa0(0>4*D(v<{Jq=POkp0#--*KWpx*$DZy{l)GDS<1JfDVLHCkq>;u@k<$YQ-st
zVEfR1|B<eUqyLCPK}5wh_chTsHX##h{urnbTix>2uR?Q40Cknt86;@k|JV_#ZbzZs
zQnNl$eSf&)=>4mWsZyC{?aoXQ$L?-K-7RD@jOw0Q4>X<4x3^+fPEfI4a|9?myeFc<
zA6g%O0IW(6Txk%W@#Uem{9={p5OX&T$C0A5DLPWIY8i3qMs+7Q(sB0ph%((-zO4UG
zDt|L2C=f|DYr-N)28kqB`qU@R3lkT1SEBxe+ZSOfbV&#D02L%k#+&(UmZ>JY%8Q7;
zP9YWsVrXC-2PJq1lRiT&9;LKThE~2kR_-`SDnbcy+A;SK+U{-Hdvg0Fb&4V5N@dBc
z4=z)QMmt1pUu@cbxUfO!xZYu>SL6MSH7}+6GdJ%Le*ZZ|H%j&at99<E1>(I<(jT)8
z1s|cBH_Skbf`ON6VG645vO<O~{urgRgR;TV`8*3;Ik)|IcY;0k9aIr>U1`n9$b%w!
zyMyC%i9a0pNpDdN+qbVfL$1r}g9LSt7Ew7~?<j*fGRhZX1kUz}1@7Rn(hIz>2k9=I
zKsLPY0vj=n%OmqP7`OheJ63ZqaQd)3aNMgiE=DXA!e(uGsxB0a`vCJ|IEtS_oRWc8
zmRBx_>!SF@;>$kQ$e0ALVREec_SHO>=Q-33jZA#0cT3AYId<2POVG72P3eI>h0Q>d
z5+DR|oVI{AO&2BsJ0t)wI?gF4s(tIu*)Di3qFX{#g6Vq7R?w!8Z#H!7!Q$FENg-3z
zTJrd1VXdKNS2=HP>T7#JWo)ApL1*LvGL2^TUPk8^18#EBdph)Ntm<%P4f;Ovz-{h5
zlDe^~Rym^4#er!AP7WDM1W4NQyuJ>o%H+OP-dW;Jw@|RqlHNQ7Rp%`E4pZfENw=IC
zT&^d0G76jo2!Z1y3Hs_qZri>=iTZ<+ejGDe&Kx8aLPf^Isw9_{;?7^^gU-B*6%$ZR
z>1w^*s=84K$Ch&~C#U5zJzPiS3XQJ;rk)ucF&4KecNZ)7j^rtm<IiOR$zCD_@8%wg
z4(@woJkM|8M$(1_N(aIuQ#R2&NQ8$pt8amD<-HsUnyofbSTpRL!KXk%o0obaLhai|
zZin-W{cGPB1{z<$4YuZ%LT@<YP|Xsr9iEC4?9m}xy6v`$+7b<{6Sp5X-2E=hkWt9|
z8RFQf!Mq+a(LT{q|3e3$!;^;wbae8<9q*V~Or}a(<I`G~qf@jIN`AdsP3M_lmhWYW
zl3$-7t|-vl47@g)Nw)ATuPy6I`1ZA|9Y=xD*7I&g_bJwcAh_SS%k-fAaz*=Pc!X^)
zBsYs1v<LC%!h5vPeyKaaew7FD5YQ83my<y_@fZ&Ta9!kthBFTV7Ws|2&NhzohlOYA
z<1d^}H#SSrDe3~qfy0v$Ai(<D3od4wwsHVcF3mu3kYNJA38}kK?&NHm2t1TS3^nc#
z_)VJL16Z}+wJj5U4FOOQFWXZc<rI1~hqyv`1BKY>aB<Dv`oum%1=6q^y0g&4*mHOe
z+RlD$Nq4Xg7k7GQKhax~Tewr)QDrleptk6NjkE{5-v@;vVJo+Fm?31(@O-l$EWBpW
z)v&Cj&xz5q@yuT0L0I)Kmjg>L>}=$#Teaxu-a(dxU_F*2^*lWIbBfkTgXaKLdPtLA
zbm{Lyy;5+4hB=@6;#L{6ybVrH)hhSJ7{FF@lC0b01J(T>uXiX)dV(nN1LWK??&+4~
zixTv_hCZD~c#1@Pzo%ICoW0VN9y>o?bGjrN^*7x9da7lYAw_(7us{V_>0GwzUDt`G
z-n#gJ=yMAICXwniXWg0O1jy_jHxe?7JsC>@ls4*o?Ac5l>+n3|1EuaV&!|KEBIvx)
z@9W?bJkcREV&MHMBHF62p58j+aOev%+kCUPXC<syI^dpQAi+mZ4qQZ509-wsOuY>m
z0cSvGG2K;qTI>OaTS&^O(sziUkY-g8o6pug8lq-a?y_8C{L+g)n&>=-OthuyxmP{0
zsf$y(?py8CTZko*B1yPQPB+PWg~AKel~SAIn0(F1{?lOx52<I4^YX8cy3kSiL~<0+
ztKPr7-Z@IcedC8?<BLnTsUKG41I&dJx6V5<_VY15W_Q=`VEqAw_1^%7!=G8q-ZiZB
zDkG1+M|aEUx6*)XT1Ojt8I1}_uK3IYaK0b(wdSn|l=fuMvvk(SYo_7T5Nj~=x3rxc
z3ewqwgS`(`Pcit8?tAlm`Yz~fge|W7{5-a$A(Nf$=eqWQ)DvPC-z?1nA<v2rGuL^M
z$+Z_z<KiJWEo~t@%dI<0+yeLST!7z1y8H^0@p0SbNwypjgEuF43SKKDOAv?FwAX9c
z)iV<e*PO|nYFaBaaL{_F?NF$QjX!kHJ7n4*Y}M;(-EiPSQb_CJdC|+n&6y^?NMdg&
zy^y4#s^vVR*O;O!qk+S@Z94!Ey?LyD97>}m`(igz_c`*uG_$6-2wfq$L5m6jnC_PH
z14<dQE{v4`Vu1rhoexkKzqL}8o)(D^1z1c$q6plg@_y8a8UIvfJ0YvB;*uJL5fUG_
z=px=)3>SwGp|7ul4;o*5EIX2p<YO)`4dRArPEZauXPb$7Mz5ZFN5JC=-+4(3w;<!x
zO6J23lUI9cV|@|d+ZolCmv7y#!22fQgG;Ezx*zjn6`mtHxjDTR_cre^B26ZzVe1I?
zQ!SLaU6-hyH95d!x(+W21~Yv3$$zx?49h<**?y)4@J3X7fRDneg9Xeuq0HXz_Ik>b
zKsCJ&5BJ(Ri<@`JzI~nC$Kki@tGD>9tfg~PhNv~*U?m~oxeWnF{4!i->{B<2%FJky
zr3qy00^1Q7<A=$@22q14C!eIot+omQqBshq01hW@<U=y@HM(Ax`#(Uc<0`sd%g2Ax
zklYU(V|nV>QN&M>QW{1QDKB_fGpwewT$PX^%Y81Q=`OH`A!3hSz-p@D^@+1!<iGlO
ze6Up=$+YJ4(EIwrZnAiyzR#bo+|v*!0`oQUn{)fZ+!W&_gn!c|q|!P!HN*u9#`oy$
zb@zWPW@pV#^pyf|+cV3a>cP`S=EpfC-^q9}J;0KzXCPYmJOEgC+~0%P(!)QhIlZPZ
zq#~{$dd+<hm|L3_(=j7JLtN3q+ykN@WV2N{y)t43P5NGRYrlo;*k6~LM&wFpRt}kQ
z0mrdLjQ?#NBr7n;cK>KB1h>b3UNl%ZfoVzlt#(${@<F^0*v~zeg^h9w@1JVF!L>ug
z`q5f)+IOCbm%1khn&+$)h$~*53EG@)KVg&&h*8IHj|l}^uL~D23#*O4P%(YsQ~;;n
zbNS$6D=M36ipDDwv74#6wDMFC>dMPeb76I`pg;IWMdv6-yTrD67h1#>34#0davE%A
zyK;D;1%Lb6o?tiv<Bb;IhtaOK2Psmcll~08!D2;^{8jHEMg{Kj`ZC@-i1S~vJwlc<
z=^h{fi^}@9;MXA9l1tFk(cZ0?XudX3%njCj?MZ)gRs|WIu^^xI7>|dMzTUQU)0*d=
zq5$Ue^ognjKXHXKTafql@$NG<gP)+Ac$DD_gp1F!xh2iOyF2F<^s}>DXVj1sD);#C
zQ8(H{L6#3-wd!}RGAt&bMJwChlaF46oQpG-XAp=Wf{h?8eF2<!^FCUtB3y|bo*_ji
zTq|Ut0{e8&N-qYl0+#tqFGIQ~?yCCr@2JwQ;)&X3HzWjzK^-u$nZ0=p&O~vCazfWX
zA0fOB$f)hmd+S2cDLiCkXi@Aze>TUmr*L1cNi+YIx;VcCwT#|rAqAmtG}&KV7%z8M
zJ}G&d`hgaxsCu-ovtVve1aE8u9F?6lfMVP`-U=c>dwvdpE0v_ZnGYe;FfCl`;ux+X
zfPv3wkLKk#$#=@wSXjFXpw#S*ReW;mTga`LiQzAAQ%zfi?cgTm>Cg-4S~L_ixU10H
z{Qjdu>dHoSaeEnLlg~!Eg<N~H;Y6;;jzW%>{&Cw@v*C`{+}ed}O76{rEK0Q=P=g+r
zHeFy=J_1YhBe0`q=fvHOMvAW-?XtfI8FxWX=m4H_K}k?TINSLakswV<5aawwb8qgF
zGv*o@%mrL@Rt3bd-C)%5{0L2QPx=55-iEb4Y01&72RCz*tD>${#reSwHM3eb7OYfL
zFW(1t-CvTtGv9PP=T&c9yhqY^m_r!NL6ypL_w_RpR8o%h*2e03*mS+8?woZ3+(J)b
z%lvVRC!k=ZOKb=BB~QFm-5(l|RM5r<lH~B9Z@Ba92f=VJrP#B3V^!0hYXDA=IlzLf
zOI;^a@wzqFhOwd)4xU63t&{%w1eL0n%(ZIC6v^fK?!JR_Pe7y;yB~I=<8sauUAQ_8
zh$!r4c_S_($2>sEj!%olF&jm+Zx|Oj0VmibStvBJ=@nq$jRM@}_J$NwdkUpwt=5<2
zLHs5h&`}uM1+e1Qc`YJ$7R0&7JVUh=!HeCZt3X$qTsQ{OsSCTF*$1J)57yc*D1?cs
zLI<yn(Rw@Dx%|S8iu#7O9hl8)a5z8Yw4jx)|B0b7*CzRkkHuo&lpo;wD13ax-Z>j)
zMZ5?$0z)|EtVaD}oq?hP#cWGuh4_MFrSSNT)51M__P-4p>@?_59RLWY8}Gq7_ZZqb
zs~R|#S2`jBVE@n5GK{1LX6MH`za0zPUx6ii2VR2j50R%K75Txf69v+wkSqkD&-CbE
zCu;|!2c(hgu|@}l`#5~Vp6(%MMb8AT|EC%8w^*p6jj_H2@*rriyh^L6J+2|5%4e6P
z9LXI9x1RQbf_TM1bzjMQn?C^09Rz5?+tg2MFv06KFKn{CA`cOtB#*HLXvwN{1^QWq
zk?_jK?Rd)<^4LK=-#&v{xh31eAnZmmQ9Y27nNmhX!l6y<HW~H)xFU|jN*P8<fnpAs
zhL!J(#?qAYZvcqC4r#pkn>>Y0eD7<(p2!C1nXWRP3QKd@vvbe5hvxzEZaWdL)rSq7
zLhM>3C4jvYL7j9EtTA6nuOVNyVaKoc_cOe9cA~}%QrY)!;B}k>JC%{^*04nyd3X;W
zFH*Y)_mhQm&wmiyyRqTQ4yU<EKZI}dWxURyHs;LPxby6#3*%!Z3$qi&&a<NC+*~{4
z8X`b{7w12SIh7kcQKcQUt{S9>E~!JPso#MM{^%mnlxY%tUh&YuDaorWQ_#PTKWt&v
zd|R%3_yx^eYd>+P+3j2hTw8g&MNOJ`=|!xszHvB+dkNpuT1UZfnKUlc1^3o@XO!?T
z({qQi!P)x@**mcY6e4|7m-D~t2yk7(S=y}|k;}}zFg{lxPJWMar|F%L(5v=NeK?EO
z94n^{;(cQ>;U9Pm$}Q)`j)qGDG}HsIM`g3kzlA&sHP|UZTw&kZ!!rhOgh#Z3zF5}a
z^vAw`@-*m%JIv5@HqO#<QnD}+?C&g{esuLXt_~cN-r(Jf42c)WAys-(QAQd08$aq9
zRD&xa3N%hl?l8nUk|q8hV{ZXf<=SnJD*}QFf~Y8M02ZKhgOW-qDjg!-y=f2wP!SZA
z5ZEXootxg2ARrQ(?v(Ck<EDP=<(zZx_uc#b@A!}5I0G0*g!g@(6?4utm+60iL8$@?
zpgoH(^Da&suJEaFnr`_%&Hve>UH^@Kk=^J@P}GCiTfS3I)b2~s!Y{yc$D~*wZE=K7
zu%q&`pMrxeY!QoW{LH7&40|gRj&1Nr&nHl&9m)!AiDb@a5<w~Q+GTQ4$e)hOk9te$
zn&+4+etf{sbD_3*ckK}|nr~gu{<pLn81gY3jcA<wV(=48XsgIbga*@zaCDcuY%sbd
zPB|2V7B0h(&T6Iw=cmH1cz$EdUcxSk$78?sE4`?r(`avwQWi+Fiz#Q36nUEVY0v~8
zLQJ6?eD2mAwCCzw?AIGUnZ7b*K{n-oCKIZhp~mAUGO>ByoFg-H{C0<?NC>vK3MDZv
zN{_6!j&FCX@ZQxaOSnCi8PgK$K6_ON>h0i%({GvFH|NcX`v<KE^X)k{WgBX8qZQf1
za?0W!JF^Feq3QF~e3aFQ(1RXL5hBZTk45rD2HR%N|KNe?_c*Q#t)@kq>Q*9LyK8sx
zi-YfI4P6)1(k+wb;)gw=%gHVkaAplhg^-I{vgl#EEPMLQ<p&xcQoyPnz3P_v32KyO
zC64-E$qQ*D74vqzzIO29q25GzVSQPoBblBsc0ejfr0AHbV1dSKkM9f7e?mK_$seWZ
z+%^f0!#b|3HJ3i6csdA9y7r7!nWd}uTJ?WPz#?Mqr>TH_YC_9+fTfHB2844YqHwj?
zYD~AB1miSIVoSq902c0TZ!Uyh>j<XGi6VlH8&*zi!GlmMujBxJ7FNxwokpt9dwKD)
z^qiA2r1;RGGLP*c1F@-{+egtZ^F0PSE_CNM4cnC~`WTDAy)&rlJv1C_N}tc^=EXP-
zca7wbI`2}VwO`@D{8+UZHntl>qlm7DWOfP4c?X6z{<H~jA`5}yPw-gfCr;hI=`h`R
zS&H`08^sCV&AATSU=A~Ttv_&sMO_JYqg9!K>r9<)=Kdz^mbL;#dcm+QLZBo{xVOqx
z!i+_I>I**EFQ&|oB7baSYh`RAAyk=LZ}Uqmqh5FaJIkJ>_(vq@YKLH<;*;r{ciEQH
zeS5!3`P1$Hu+gh(p4mL{JJQNIYx=l-ba~IkR>EupY8}(I;Mlt~^%UxCnF<l}Lw)sO
ziX(2lxyBCQFSfPn7MDMZKgG9B$Kb-Pzj>SG140rUh%NsAsXUT=0EhlR-ZQWu?v`Q6
zzm15Z9)tjDwQpe_VY@!vDl5s4H=&iyJ62|q{HVGGK(aL*tK17u1=s;!N<IOI;d<Yj
z!S&&drC|vJdH!Clh38y+mR+YQFo^GCT|+x9gPDtDnoUdJjP2@peOs=vg23q@e5c(Q
z<l$@3H3)eJ`UQ8`LjmJf+u9u5w;+fTfqGSGzl`)=zQ<OSZMnH-FlxRtC`w2-if?%n
z%mEwxnZ3>20;4FiO%4lFBryp)UBL<QK1t#F)bc0gHQHxmwo{g(?`r{IyFp4Zi)LqY
z+un!ioy=kUlrO$8umd$lFWH@i{z4VCOda!=k20QP`x18{ehTPD_0Tn0v{GWZ;0U30
zJC!Y4=D6|s{h7KVkv9#Vo69Tcmp`E$>GBvu3bWreZ|SGDC!=gn_(Np)5%RFQm=N6l
zB%(dr(TYQB-(#uFbSB;47l2PcWoUIj+RaqEOPKZKJcyU&wi}Ub&&If$1^j^m?iTt?
zA9<HWd$^GqJ1T8?9JWASgi8CpVO`L9T)6aCM=yJSzs&`4pDM^J7nd2rJQ`;(I17m^
zU(<!pnHL^-QQw*WhP%CRP7iW$-NdW!JOQS;y%Bwn-QXI1<0@#GDn(^?AHsxssNrLU
z9KW@$49*mI;BES|^Th?$J?=ESu(J=-^yQgfBE&I!4fR-qvjM96byr60NugoeXzzlV
z{DCxeuB-RIJky;4XmA*6`)%c4&yWz3jMVw=ZdWLwi3tdV+WzN_Gp|hI-~&aS$CF4C
z6M=-Dx=**mr|i;?JpT7$<*x!pTI#>LR5`=^=*KDFuP&TDi=|)^)gV1_Zp7+6*}Tk1
z@4ieQG%033r9>hoht`M9N93xnm_+$W<ZI6&(9Vtq9@5cQvqFt5ZRB`P?ZP(U1h=(G
z`wPM`{-G@v-1!!V9!q)W#02t1Z!cG$q#sn~<25TVe{Z(XmoNKs3+m>8ewib$2|6t!
z?$`L!4(&ku&!B%?jY(+3QlJO&{!bEG9hN_9Q^8SPP|w;r#dIFZ)~swZKdO;-NVl1M
zqpwA{#AS)8P+Y`i@iavaV&fVlm2cBkcf;t6N2#8Kl;4S4i<X!tS){vQENH1?Xfpa>
z!6}97cJ^_S{?0huA&OwAX2o#s=J)m9DMPMF0Z?XOo#s`qlMR$MeBw*@!fu>*Nhkg2
ztW2118%fig8ogy$8-N<&pVMI$c@k&$S(jtB)TsW{7>|ZjMNX8ipg<%3#n2HtvknhV
z)Ap9!AzC_=_~=WTpnVTrTZ%YjnQ*RIYn-4z0w|JY;39dZws7m8XIkU^sb4nlp6Hkw
zg3fmwypC_8ZB3%kW3VIduB&#hF4o;ncq#tN5u8Cu-GJzR<L1UG*RCb5c-eQ{dfrr6
zH*yjGn)bz#%~eHy`iB+j)26zMIO$!cPK>F_nDKdtz@mGYXb-j6I>H~TO-0mN9K!<a
z;S$?0zLpau9=<N}shn3vVyOwef*Ph-pO7H@(M-O7p&Lhq-B!yI@p>KP%{e)gXS2|j
zK72L395Z7ISWJK5lBWqy+_`%x1mLQBn_<6QiFM%Yp_FP5BPTThPO6;ypDRpoQpw)I
z0d2WK^6CI{x6*N*gqE`xcxxes>ni`(L1m%Gig0OUS8mDJE!d>!vN*SA>gQW_pQT{r
zGjF?0_tx}5q9&Dwix=!tN&sB+16q6WH@WC`#YX@U&QL`C@VlQk_l?ws#`+%~%7&I=
z%X~^DD`ql;oEl4guS2ZU{1AJjR?|FR!Wwuqz4kj4L(J_w2+oVQTZ7)_MEPmimj5vk
zxQxat@h_(vFj3$vf8Ihpq@fU;%{Mvg6g22F*Oejrr88Y2uf#&29?VMz9Wqf+fj+Xl
za?K^VfhW4*$O4hIO_e*d&rS~Y32lfOI%kx{MNW18{`SymF;ttR$|=?|t4ysnh%qW@
zE%8jOo(f&+Vq2nu4!ctsnpem2Y(cXanly~BGU;|%Pm>bIYc!8q^Mev*-t?ryfBJ1J
z#0=!^x-)g#AmP0n?C;ck;X}~O@-<B3c;GVR&*c77P`q_YrDy39b|j%B!i%{rJ1O&(
zUu#Ce%|Sy>W}k0ZEQP4$A^jssgx@cFlw!3BgHu53Iq7Ich`w-T>mR&M7uOXoZzeD=
zJ*>&QB?^FFF*NjJdh);5NqAnP-}ar+)OY&olj*oRXg*f^PF7H>qIzN|8Wp|SHmZqB
z@tN$PH6OQf7pJftM)g|<l@DJ&HaF*-5v!>@xwJ)OD)(;`E-3&u&0lJat|3S0zeEg@
zBaxQSjr|4NjA~kzi29pyBoo6|{wHC~QcBd@HK=(+mrk*L<Y8Qf&0vY_p*ubX5*GcV
z)&B2B{QPtvmBvuU^Xh~AA~Wo9xJgpXZ=l*lF5TZB@>rU<swgFeF=+|`_vwa*-|Jt3
zrRn`6D&TgV*=lav^qKFQzaY8itZ#h>0h*&a=ZKCkiviS^tgy!?Ibq{w1%H)a&UZ$N
zax*~zm%_*F;(D8R`IoKE61NHSMyDDqLR;JXIDq{@T@zBBFkZ|c`tTa2xOLsidWwbE
z9&9z5ADtVRYRAbggfaBb^xpNiOVf%bt!<nv^Zesr4#f$2@oUL20y(_wJw?*lREzl>
z8~P-<vw@|w<7wyBzS){%4~m&PJb(Xq9jE&}rZli}mi4@lWs-rZ*IutPZTD{e!!)$I
zzFDY;<v>x+it3A}HE%SBv<xbOpPvx*pJ8%bo5%EZD~<FnWn;Qdv*B3K_1=ET!AxQn
zPF-EN;T|Q4@C{><*ZgTq=k=b?x-+Fs<(3m$LZhylYjjr~i5HuR;d<QS!w$B}`kA5I
z5fDeu8(4dNP+wHST3J5X2MgfMv96~lh#8Pjp;!}yyk}WxlbQpJjZPt-^g-N6@*=;&
z5(OWauFdkez<cM1%6U}N{{MdO9fO)uH~Xj`9idd_U+79<d=THYAoPIt*O}{T0Vw_B
zV<z1)e950_NCwlZPT%ef;aME{{z3+OG07dGSrm}y*D|Zq+FLUG)RD&0>2u&<v{OLA
zXIt(IPcgSP5)T0=8xCRQ7a6<xbaQ%|b#sn&f5E4fF<3*-9%F>PtY_Ide6=O~mYHeA
zKtI^e%HWnv+B5wLRYw@P@ZY|6!m`H)_dZ@3H4#+VW}|T}NYy{y+Wf&?%zZ|~?2Jk7
z@o{!{ej$n~k2_9w_v4l${X{oNh;7n4^p-l6V}xsB1~a9K9&^wscWZ$Syo69h>;BTy
z4b5*{Pg<J4(CRvUNLz6-jfw5*g#;p67sjcp1rEtvd=?!y8QW)#spzhCuMBu+Y6(Sk
zQYUVIZ^Ed4&Cm<^DuKar8PeX1@jG$;6uviwE3<WOBHFPzr&JGWpI#Dn=cjtY2j{|z
zN1I`UcRAhXYzywL^Y%<q>}ud-{w1#-(Oqx0>#A{78?eh!8as&XO)q!pD|>mT!gjWS
z3UKT2DJQ~n3NGjI6RWUp<s4UFF9K@VmwE2>4i=hl6`w!gtZl>#^Y3HA<sSXbhY0QO
ztPS2E-@f=h4>Csr`FZ5_2YiH&<vD0;5SHdE0%7(ge$BQ*fNhGjMh2ws8b<igxqDw4
z4s^r+D>}e;sRB2M!I*rEBiv<1{}@NLdr~-N?=6D_VKEw4luInN>z+>*YzkBpcABO-
zYDT;M)9V?jHvn(vXjxX<@;Oa)(x`H<pEe5$`$s~%D>0p3X^fP7?~OaXdMwoF#up)#
zw^IKBzvNGVA<{Dzq&-Txh)`yJqra1lpoivd8yOE_RalK~y&GO5&B}IcjqTXDHpz4q
z6<hAHSk`v8zF)(OHAmI*T`BA|o%sspep<UgdQ#En{+j=J&{C}%<F~NRdXk{~eZl0+
zdOKx3q4*<4X2QG0eb7^FdQXkZy(M<WG0SIvFY~#N3FAZHcAvKT=zqOK(!hFb+slYH
zx^cxF(M2@2a`MsvWU{bpm|qcE=@^%dDKC|*IL=z4m)^|CuTl|PFKnB3YlHQairB_S
zvyq-A)bhyfnV~Z5O5D4FC|#GQX;X27e75TSfKDZ77us+GbRvo^u<TZ5Q;c}hYdzFh
zvsljOao1P9t|>}@p6gmG?b<Y6bhq4P$ZhI87wOk(%wX!8V+CCEZH`~A$lWi3kNrKo
zk`Joee>Br32yP#!qsE3v?jjH^R`)BUy3IJ+`JbxWmkKOVB_jH66Q6Redbw4O-v>Pk
zvSC-0F)HW<(8Ukgmv^GN%e}S}AU1J>Bt!I>`x5ZDfXh1)Tk-|@J#m@`Hd=^|9j`od
zOw{uJB0xVM07MkZ?6n-xtS+L9?(lK2rPF|Ue*(jsHs~6AukkX|6L0(;O^ix-{H&St
zJuImxtkZaf3z-qM?tD;$Ai4f_Wo%${ucha)z}K2x(r?|Ca2IlXxTt@O$6~Q8UYTnU
zD6`zp`f<t~$Isq)51U4Asc;Y!wV}wKTWl73VV2}Zu|9iTGxc$r$;%EWZ#vZm#`|x7
z?LKkw(#WXLoL0eiE?KGT*uP3^sGa{aH(o;wQGNb-=$rAxYR8DhTk6MSo|s}QF@7-{
z7-@Yqs{US0ERALI<0UCB?SosK_PR5o-cvjZP`B<K=(Uuvce%9k<THyW>9`tPcFtjJ
z>QP+E%coq?w+#JY`2nMS>~)LY)W_G=xjf`8su%*F;Q-n4A~hC76c^(NuTn1g(I3IY
z5ENDV>>F+#L?LOCc^!k0Phhhp>c3P1cXg}l<(vO>SKGLdyV~^po{f5?<&%a5x1FBY
zN{omMaGY)#dSzO$-;u2~)2+BrT300_0klUPW$qhqwsUm-HPTkvs9n2)I-pTZh=%nA
znVeUkeG_~A_`Up*Z_rItv~Fl%@iGPqY|B=t-7RXT#{v|Qzz3+WGr+g|?hPFV4lQQf
zlWpKoMb4gJeYY|5JoZUg=+lfl1+VWUo8*6l2#X;6hyW|JooTo(buCN(f+Hq7x|7dj
zgPW+3>xCW3d&{Zhfs=Y7p`TwdZnz_GEd*Dg%${^~#dWoG`E!aVksG=9=yD+tk%NYr
zXM0Viq<yDANYUx)8z7F|r4;dRZO559P+~jrr@6SU{5;=&<BE+-lLV~5k~I^<Pi`b0
z&k65<I4dJzSzy&0G9~xAszZ+mrxuGLa8v9ycLNaIio4HG9EZ?*7T0CVkbC9;FDG<u
z%nYfY4q~Pw>;D#4CY>AjJ~O@gx907C)mpjA$UU<plf&w=VFtlXog4NX82QQf<^Zhy
zfX4N#5a1Q9ni@wQ77JuCDhk*PT;b5ldykXY7sa_kO8y7{qRh3&spA$ORL{GX|E(UA
zm791@a=aw<p^GxSnNBakBr@HT3-2>x%=k_?)t)zXO>Nys5M>zQTednQZ-dk@Cpgv8
zBm6B^nVxn%zV+Z{_cciYr|J7{z&3?svPT|nyrZV~BB^+;r`!uW0QG0C!^AJq33HA6
zFYqQ&njCsB(r)`V3fK<aC>(HjYt@r|0r#j;&&J{JP^TCp%!ez-4|F()Zv3MKa10J)
z5K>hZUHM@Yc5^fb3h{Nl7TkLxl4C`y%WZ?5v&7z{=xQhC-d7Tx)z%|Iiuj1M<FzKH
zX>K!)68hzyad>O19!-Bwgz2%tmIQ}$1fspgLmd^Ly9zz)&(7|Qb=+~kP<v8h=ZWVY
zhi2|&T;ffOjw_dLEUmBM7Rv_<ZDaFnhxK4*fT>&kjEnfLo>qmomk?}fY9l;9!-XfQ
zGeUIZ^ExgjFh$dg%<Fh`qazWmmp7uyd>EQTWp@=ge%5nX<crkExIuFbC_(?0fB&yn
z-S#)<nd;-NjP}EXap~)U9O!dLWPZb(y1v(^j&muZJvX-Ern-Y(_l<G}_U9PZ8SU@w
z#_i8JwJ-bnQ99a&va5svwA4F8CHv$NcJlKp9YCU<5*tQ$CHQcqGk)L`4D~=l-6Nov
z?3G>}I}-w1vhhK96_i|>T|7$28#dPH`YWDf5M%++I~4H9nor*j-=+vGB0ocN##Mac
z<3nles9Ks@6Pci;OQq3~{CZ+^e&{?v6@iucp-LhWmf@Ndl5_iL_t}(8otWtAIdOnp
z+$)?H`uG+}XkP0njugnI@#W4AMI;-VK~Fb`9_rq0dj1qSJ;&Qw!bru3L0uraLCGZA
z<60YOZ6js*nk79m7E~$q$R47+Q2<ciS>7m!-~I8LMkXtNAgYtVo-1Ecxlcg_C0Cl9
z227D-g9nhq5&XxwVp>oQZwXb1jDGFYv#72ySi8kNhK(wZsA)JWt0ECwFT`eI$kwmf
zYaOFzsgtD4IJVug<c$~MQgGclNAU<&s5IDfE!<sKv=i}<4xLfpU&nOKmNbe{?>46%
zo{pv+{&@AF?CKOQwnIngoG%MR+oNB9MIhJT6-l!$*2yQ3qACQ=3Nn~w5{;x;ZB;e|
zMAAUg>{JbsX8#^PfWHS{|L^}K$RXE1Eqm`bmQ+2om-qBGu(FqVr0)eYin-^l&vss4
zjDB6>_$^N|R@vn2a;weI%&`kv&Y;qH`pS&zQhGnEp}GdY{oCs}1wkQGs<d1A3Na|H
zj-V2t4xU*6=C7FLh_&Zqi5m<lbFwR-7m2LI1Z5bkz$y$$v^T*V`><xW6!f#QrVI2o
zMYmHQrG#`TKZAFgLKd#&jL&yRNpmdE%T1Z5nQF$36n9s`WX2wB38C+{w_@qtP#Rr5
zglWm_%#VJ9YrishQctW34B`5`#`zDpg9+=;b}nfiwe>=I+iTF^B>-%rjKX15&WY`M
zvP~)SpD@%aOQ_q3D;8x5z)9?`rFeHQ_7`R~SV*rY3V;gYmfmUt*W*=(!4mPSM_K-j
zQ<R4!NXo(X5Fr@b5E=Hnt!S_f!|5DBUd|@M87*nGPY9J-N*Vvl6aTMgpF;DqV6pjH
z)H~lPKz4pYH+as)fPz8r{yMY3pz}i)=Mo7TUsJqEj@;$h+UT9>$xDqfBHfQ^7mcKT
z+-|D2(vh&}<U0?dlt4z2t_R%`7-3$r_zMx(#S#V1qLJqxT@Igc9O+FhPy{4prOn^z
zDEh_htPTE#9gXW|b#Epbc%pVMiI;rK^L2!8tKtsyu9WR!CN9iHf84lzO}(h!pxkUh
z95A<3Z?JwsO0c)SuQ;Q9HgaNIr+%JBdPH-swPg*?=J=>Ta~LQ9+^;>rO3cM>_Z8aO
zI$b~c`;|EQ*w~i?wce+O<^&ioe2a8~E5CW1AzvL#?}$I;K22l;U&Wp=^g*hP5=dy(
zSQJRim-_CX|JQT+_fwOcCC<&jVt)M-yH2n6n#OV<%&V$l!s?SWWk-bggUwsNrfs6g
z>3yfh)$M`CY%6thgsxj*^@R|_S163Q4XTLi2B&w(s6}pY%na_Nf`rZIjBXTQ%!o=G
z+<KhvAYCD71Qnqx4qe}gri4hF^2lL3lB6c?YzHrmyt-@4jh0D%s_20?lVYDM9qrYl
z)w-Ji{7is8FU-)m%IoweD92lp(`pKOdhd=D+mF?@#mO8WBa!Um5X65oXoXF5zkuDH
zOO3uMhH*INLVp>>J$f@RGPBb01u2E}lzyq}j${4F@T`hGEre~=@zHyMEAfLLwQPFr
z3+Qc?!#;XZE$af$y_9ew0Pgim7<F_^4)_0!6a+uv1O_##J}_fLSj!)ejg{dPB4e&o
zcKYYx5#tH>8UNEG&L-;Z8Pt^F8&5l&ZBzij)ZiA$da@PlbBW2b8Po!FrA(9+fGTZ0
z&lQ>0#jM-5T2fGCO#s^T5u!ogR?RlB_XKw>5a-!jB%^>0X~ViKjgvaKf$%BOcyH&A
zD)2{ga)dJ%QBJv+U5>?dEmiF6h27A+h(I<1L$1{iUqPFlBf8eeqnf71p;Ctx1L3sx
zE!%C=JNMhICTEVFY#aTvuYTz<BW`9n8czw*(m?P(nrB+uT<4Ma#kriWYqwD&fvjLd
zvOHmPfnESQh(V(t0o1BRb_R8cuVx}YJ<RL2yxN*jebWT=t+$<BgEo-JG65?W1*$p;
zF-t(TICwHn&L{gcx6@9-#>x)oLWgelu|vd9Aq+d-MGj>zqDI7aMfD=x>}_%*oe|Cb
zn-z{z%DwrP;S3_E8&?EsH|~zW+Qs;f*B<}cVy71X$x_j5BS@fgkx9rRbu=e#RF66J
zKS}k;)DY-2Z5oOx=_8c&R#|16pLqc?jl7`hK93+_RQ`UNj<$W!1k`}!muuObcc(Sm
zpjxydSsxlELk`d^jw@_MRysN~Bv&|o9rWQ-jG9$?Axjs43+IYmNC6!}Ke}U#@^-iT
zNabNzr1b6vGz)rgp;||R+m%G3Edp|MiXD_su9WIW9Kj*r=KuiowMz^WL#k(%*8?pa
zzG!Y)5!PpFy<!^wX)O=xt()#RN_u`pWx>&xMWeeXsx$}X+z0g?oA~M#n>KRPx^tPE
zQ8fv1)s7`g{@JAHU5(*KLvr3E=yv<y-oc|b-smgK(KIMTmig|&hKsU@(=-c$;SRV=
z)Q7Sw0P)2bE3Al1kB);#?r*M`X1a<bY^nO0p6A(JZqu<-$ef04#cqLpOr}=;ua$M)
zby3r124Dw1j~Vt*AO5c)A&Aya@IK{CGm=k3DkZ~6CLF;Hdv9l~_;128AKO~}u4#Q?
zS2qvM2-+*!7T1B$neP9*A;AP2cs{iyKe$x<`^|N3Zvx1@y<bPNz@4%~W(T*pNCD{b
zd6!J}zbY|^e?Goe3yaa0QhWr8k^VNDuI=5z$9eNFr<sj58}M4s{hku=OG{VGX7k6T
zV_q=3PhW^9-NlX>!RZ0zhi*y7^j=udFudl9^2Uf;)myuVu^iJ>IfH)&cSJws&w3TO
zEa|mY#)-#A6Mq5O=CzZe%g?y*`WI=Te=0kRMJ-$dtwbY_XI%3en=U!LZn<aiw4)g6
zAg^B$MPZJ$Xp$rhc7-<Xnc&<cgp;+JG_#n!{GzPANz3{Fwa!jyGuQ)F%um=mc^3ba
zOd@)rOPqf3zu&0M0oyUZctk*=klVI3?o}YO#L)D7v`70h7m#3#l^@jwS#0f83la*a
z*A3R!*CL}|Q2E_z|D#w?a_+g^Es^fM-R%ir#6tmcsC#gz54WToWJ8R)BPk8H@ce(c
zh0>hMrJ!vaYs#f_*aOn)g@eHW71^|pq7}3n15WKqwc1WR_OP7qb=t3Jq-AwPfXNDO
zm<`+Gi@SsLQ~&dw?UUr@V;zGMNQqz*)m0#1?O$obZq2c!v^2Gh+1a9-^HygBooIMB
zgo6Q)xC@;9JDawgsMx_=a#0W63m-r{bpCrc<bWEAI9*-m-JPEF693sAM2U_R)SJ91
zC#d_+8e*h`Eya*wBiQE{_dN<huwcZ9Nj;-{=eCb`pLr6iNaxEs4+>Fj=cMl8tF6nL
zAn^~$^!Nvc7&A*&61kw~)GA{?CdaYt2@3WA)P0GSA2edxny>ICI5bD$kbBxp5xhnO
zE(c+s^FV_vDJ8=dLU3sFGMb&)@e65V-tlxj=3)VOnHG(H+ZWP7m)<4t3ICzL62Ty_
z+m%hYLGi2&90@_OHC*Ti!1K-rLoTpH(prHmJ}DMA&hYGJp!vo{GB4#|MiI3)r>{Q6
z>Ul3wy?m?tj?fl`=-XnI`aXzml3Z~2j9cPIjjcHnmm-bLwK{vGfcn?jNU|tN0Nn_i
zy$02O^)xD?)cjJjcYUAjO=u05dnJMv?y8{ntn%#<A?L1WWcpklnLu^QdWKnp!OTb+
zkN-G5;U`((;?k^H1I3p-<Q)coG~ax>ZzD(}q*w3ApEh%*gO_Qa7UTv@RTzT${tDxm
zbDkEhqjueFHajEVnd$cr!>~P493*Nt^32T44ESLCG58w@dd7lmN_EY2@(UQe_%cy#
z;ZX8u?Bg`GiOA=u4w$^^KH~HG1z?Y7MPTG)rZY_)U9x*@K|_p|hUEJo>GXgePyM{@
z+1%Aj9htG=kQL+8EbDNU;g7#4iR8>4KMz#rxKWP|qX+0x&)<CVOccNbx1jcL|LG~s
z*QZ!@-i3RZdR_Pd3$;y&{{VjZRMh4AqhuB&vBC_O&rMX*I`U?xk_YAfx}&1Ql59ua
z-Sp<(-m$6JMYZCV2BL;u&|S9f%}um^{ny1j0v9s_*o6J_)u>j-1$doNYfC*jQY8>j
z2&ySPO*mSm;*{537<Gt%o0(!faxjR8Ol-u=HK!w5>k5)P%&uJmO|w1>UD<0eKE_-z
zg>$8*X{iuw9l?|vi_w7GHZuE7Bcg1#1GJ+z@?i-+uEDWH{n^@o8U{>IW$8!u5_H);
zfU!DbDD%j}zaV^(f8!Mts#z=t(7AgsiY4>nT(JO!y-Fmv4Di7-+lfH@Nc@7GgV$k>
zD}Y{G%CPTlmZUFHd31Os#zl7FP`YnVPlhQDUAttB=uwA`1`Dlrs_XOV@l#>Vw@`nT
zG4EbZ9a9MCArI)*+Q1u?@2o2H=9!1C;9x-T;>52{JfBHGj%0SWvQ_ABh3J@(FdC02
zTVR5WKhVQQDO&J>u*>3nF`Jkj5EyNM3l^~y1*Lsg#VDioh@9@8nfd($!`JnG2(+tW
zSvij*=3xahPl?q&?2_ZoXn_NP%4Fa+3ZUj~E)()H>7kxAjrUgmboF5^8SNah2J&k(
zgw1bU<2R$CnDZmv(bM-e{M(Di*3B<!CQ0B4rCo~~Fh}%ed!RMEa@tWvzvP42@q@b@
zUF>hHn6@O~UD|?hgASm4rQ_;q!4Z2=%OLY#^B5feUbmIw+$yM;-jXilA1^KcVgdqz
z+HAUPJmC~Y8tx7JB)7E+Z5Gzg!YFD8fECfa7mgsY`vaobOqqz&KstatqOX*qbIg7H
zdFF;Yi4^oi#U$%oHvt%<bp!qT2mL#2NdGE`j6z3TTM&WK%>@NXgsA?vt`_<9DizE&
z2<4)(JYhV|2;RjiBAgV70eMZ_?fA1CTG9n+j`LRdg?@I>w1k7?q2{((^5g2^zWTl_
zr6bG)g?$2kKPL+1jOH;7dIXR}Ry(<`sqII#?pLIx>sl%5S?EGvj-b9`$D};#el;jx
zPZj!yQk8a7<z$802}X2lHI)>x^=GuChGic`XJm#O>34j=u`f`je=X?Yco5NN`K)=A
zHMGM<?JL)=SQJDtH}`VwSDhPf{S`hw^-!LfH`yG~N6{1>vz?g=+v9uyBZTvQH+PF4
zhc8JV?uH4ix2SL!?omMBiV>|ye}hCIr@1aAppL>!!+Vy!z73x=pb(vllb%7kyRJX4
zW|-ETc%m4_H{3D~cOV@&1Su+k)S;wHvVk|fF32v%R8P*}c{Hc!!fs`VC#~b`Kmry5
z<f2<i!n((EZ{~JfB>d=hJsEHJgdRm5o)}tZrGxgQWGdWRGZJFA(^b3sf7w8c-A7*M
zb7<%;6`6fRr0Fm3v*h_sfpDjH>n+Fb$zL}7UQyY5w=0lgsr}paO-1wcft4_sXAHeX
zoR9r13JWx=^e`vm#JftzkLtmFNcez#DCj}AL`Pc~^!djj!|1tW%BiHE`Oc~LHhxVh
z{NYh(NU3@I+M8B5=l&4oeYqf}Pg4HgA7UvPKKd*UU9#l23ajzazdl#d^O^j_L7e$J
z-TSs22R+xi!f97uq|m083|d%CKRj-Vc9YGsOnUsR$=aUom?cC*1oX4_XQNl;Nqbd^
z=|4lT-iPqrpd>JOnJ%HB%v`{37weXFn>x+n6p4JE$M1AaOdzGZS`-`&ftc~2RV_=Z
zm_MN9y8(km)@X88`eggmP(}YK>F&3F&A?hrq$re%hv!+2K{QxZloH{eJ$@4~RwH4t
zti!@dYW4G`XIwW4&tG4X3;GG;T*jc-x9@(ZD}XBW<-)pYmqpypGN7<SHqWA@BKAC6
z<78xAg*JI^wXWf=MW=TMq5R5V)x1vk(+!5M`{m`ymraq7&<F_9pd6sipia>lmOOj(
z@>*#ZE&$Xt_0S1@Hg)nk{~R^cxm5VNAlr?wtJrlA6=>DX-t~RZ7NPU_gL&5(YdBw?
zR^psP2ug@8qbuctT6s>pYPN=fFeT|_aDtvF1sTe@X5Ne#MZElHz7J_vlaIdyaWe#E
zk;NgPQl{>KDkji}vv~Z})dz9X9o?6&&q=s0f0>75v$t7TaejYV|FU2&Od}%k`{qz8
zQw<!G)9UZwwH-Qa3&*6aIC1U@91{if8q%AQNk1Nm&P{4qP?~K3!}72;R+SG#f7U!c
zzuAi$LVHR48t(RH<GP`7ZBBf@tzA&<Xw$isFCFDScETU&>lG;y-sSq_^eJu3{ruvw
z^gd2TIZ&i&QEYpBZ!~oLg>#f=M$~uk1&H8Ft$38B^Zk@r-J)2V*U$FyVTIRf;~6Qz
zZ^dbz4>;&We3|OO?#uP=X8ANsW(@!GpX0CBUdme8?U3@w57(ya;$y^De&7F9xzcXD
z7gDk9tw;Ko=@W}QV#JX`ACtqEjOc+PVK=Z&tG6FIqw1x5+ka15d?$Eh&^{o-Ep`SW
zbiemwY>)N$&kwcvzgaIoRj&uR>AS@l6N9auXZrIRdQv}7d?YK5S@~Md_8Bn4O1*dk
z#@o^y^0gB)6H&sAWu002!&`&LDc+G5v(mVBm856VdYQbV<GaTjJv%Af8NO~gr9-jz
z!v>|R6H4AfVxDfOO@iJq?Gsb&-iKB9&X4@|@WNhyB*Qs%e^*t)!r<7}QQimY#&A&C
zWi*>qs!a(4MKEg-bHdRxRyy_}3U?vxOP;nVNA0Y8-my;O)m^Gdxc!TlxO}Iiaqpc5
z3^Q#^-o}JNmBMpBR)`StrKf4e)II{Onc~hSVAfJ;cU`ZvD2a7lYYK?lcOZao@L2b#
za>R+TtLr*{vK^)YZQPU66F;s6%%dpPxLTJ9AtbvtS+T7IFE+dh{tm6m@!H#DJ%F7|
zo<fkl|1n%Z243he1-l9Yh`wg^{YxR`G5LVR6l~{8trKr94Re9qH5{tcn%$$xJ-r_R
zVjuFt>%u6N;yByA5g`W!+3yGnV+#ESdZx{R1?I7;n(rb<7OWiX-12`pxTMjKha7MI
zFns35^xO$=AqQ!-to=K(Dp7OV8ljb#EQQm>-?eztuNUh2<cmt+IW(7>kA6s#C>Tkz
z)Vmmrx8bVUcMX|9aQSPvX5A|)ie5s-Yf3UBLaj!Wdpr7<c6NT#FKwN@!?qlq+u%=g
zP#S|b%VP7@AC!MmPQ5D4a~!-BEP4^fJPS{AXmxBT^0ENia&M+D-%>|ilu@xRR?t4K
zl@F&R+3I|OVMP83^g0gx)`jOS!#s`)pZ<_cAFFF>o^|1=9E<qUc)nU$FdRB|G57U=
zH2EW+5EeSxy95)iY#y#xHpRPjs=tQ+=Y5aGs!$41_EkY(00LVuD7zceBNY#=2TL9=
zR_uF63fSsJXZR8F=o9!2vim~FTmE>u)lh_ASxy%LMyjU>#Vh^Etld(!u5B3Chjp<X
zz<j?;Yls{&{=Qp}U(ELTilEwQ&8G)oN^4;nV+@UFTbYMTd+iCri8edexJmlgdQ5Q@
z{of2Qs)_e{l@*|eTd*XY3PFT|ehr2a*LnVUlX;8?UBiFqKInn^qw2m>){AHP?5xNk
zLaKMPFMikn3>tPp%I~u0iqwJX3Sfy{y<W)XtxJ44QQnhc(1*39Yx2}DjU<*q56&%o
zeGc#2(^QWq=Hes}o!y-e*hNon8XqD2`7&&~zNTF<HZ>Gon*QUvbq(kLH5UK%nCP5-
z`m}x+TOZ&<Pj`+E0n`18=>5g#XId&mTj7E3c5$|-^~0dP`Ac+s0X1OcZyq4VQ6mU^
z=z?`(ybBVKdPhB^_w~I(_9(?QXx*w?krTV+c~#?!<QLX`e2YQkT9<>{@8WP7krXxz
zwX3o$S4rJux!y@e0;^%k3gzOEvKD{IJs>N>?d<uzigmR%m_pWf$JUR@yNn1WDeoq7
zQix0J{wC5a^2~2vRl-9Dewz8{K~?_}If55-KJtbq_~s)@ncb&e%9@Z>1xmt^M24Vt
za-A0H`9Mnj0(qXbkPYB@fh-a2w-{q0VnmOZ!s>L{LzhvOf6d{2e>gp8e(NnQR~U%$
z_-$M$YAiM4LD~FG%8i0rB%=Xr;ZvMJG%mYUGw*USPlWK|q_|o7;9d5}3aUA9r$9>m
zdc$Bb6S^*kQ++E97keu)T`K@|oHJ0ze97I@(U1Fb+r;|Ile1=RBcJ`kYlE4&v{dH5
zbq|i$hrXly@Vdb0h?95Z=XZv*u3d)feQsN${_iZ=H-*xDz72s;sdZ!U#z<0K+5-S*
ztfXr6i1d}M=Jk4jk9RqD!N59hWJ0**Ni@GjvUXvnwa{ye9IRaa905!s3^lSeF>H%E
zsF*mO4ex+EBSOcJGl^;D0ETiw$qe&?88?T_(|#7PWuj&3#GTkxZ1v!4J$n|Bff>*{
z!jwcUo|xiflr=%@_KR7<^Ai}_glWe=KtSBAYBj+bCvjP?{tmgKlf{K<9ggb46@t*^
zbF-2}1u_XZCKA*eTfL4^nrmAew<1eWg+A|Ez7Q=OrT5+t)n@dMTtlK6_iYy5gTdyz
zfMuWhJnM4<+C)RPBA%n+M8LqkJ1Chsv`yTM;HDYFv;n7_^3ozu65xLb;n4RHKmp2$
z1rk$n|80cz-i>H?$%dneUWe1?iq>W0(02W{p5<U`?yVHskl9Ve5g}uRvu|k5a5%)R
zmt$V>mVIkpnR_g?cBD=PpRs0J!sgy}KGCRc1nq!|Y%5?js~a1As2@_nsbPW5VitMH
zx*W|DF6Hm@!R@X3&uT+7UQ`T;(t-#09ZLcXQc})TCFI=L<>`ikE&;TuVWUPIDn7UO
zHv4VEp`W9>T`U39YjsjYFrO5LVA@@oy4UHxbiJ3fGR+-$*g^aCB(wOy^lD?6_u+;6
zsfX7Q0`v@z%c4#(tZhHSp?YMMuzg4&?8a*ml{ii(kS!jGd0#(L-<9esU$)=z%OVXK
z6olpB4zTu~4$C0s{)peIhXpaCG0#@V+ta~8Z3u>gnrqx}cI<QKFkf!px{z}{Yq&f#
zK#9=Z*7D3CW0P>Nw8djP43SBNh@Y7XdCZ4}D|8I&>|vR>knE911l7Y70coMl3QIC&
zHtGBe0qKlNVEADnLc-<C)qi`x@WmeGhs?)T_+JBcbIxCs_?jj5J;(G#$|&>9{r!GF
z#a`3yG*N%6rwz(Asxp<0{jTjbxU`=D{m|^n;=1Gt1#}|x736ohSG)rtG0x*30O=GH
zHRznNoiNMfNYJ_2j|-m4AHpEwXyLtD8#<-~c`$&{<P9fYm{Q@!Zw5tPd4k5s6HMv7
znk|+!CHB{GJOoAq_*HWO{cOnZm=vRnDA-1ibjnzE<Cu2?v-{S!-*v^T^?uv7cHeLb
z`9}*dRN5=Z);d+#-WOUtEO_=Xc>UvaXb}|Zf4CKW1t*XwqkWi#XvoyXDo(i(W|HV_
zPjm{%VvGx)`F)P{SQJut38K(nQYXbupT7((tLC+1$Xw!I9j}2FDF|h>dW5bb%!{Nv
z$-Pg4GJfkt3VS*Wcdovd@X<=$S={!iA4xL#74peaECz%!>;>{n9dQP_#n=f7)o)I(
zMJpmcUj<WGrP=xyj5ePxzEW|@6JN^IVrOJ#nlKwSayh0wqB-Tb-gB#!mQHMI{Ip6;
zCb=lP$74IugMTyv+~XL&b&{!b^c_}v@e6X7ZI%8qK7JOeZaLm(@fSo2?8wsfD~6`E
zlD`%YXhR!6QA4ZGTeW!GhOVUheTtvlQvnJ>+3K(Bswt|8MX{{(7bzW-a}3$F)7oii
ziXua!?6RWlX%^<xM>KI?FLa~DJhV{mC~b`1RA18r5vC|-;vhy`@S&J~x|l@(xKaVb
zrv3~r3{R?_We*BLv(FA#WHu&ff^XSa-G5JK={N<|>~WFC5@(U~mr%^RR&DnUf!epO
zwki(p)Ee^Mwnn7)T={127m)eBsy^icMJ)4QMDRzL9D~3Zx?O3(T!HR>Eu1>%a?Xod
z8hYs)3KR$DC_4zaIobGjjQ7zs!}cX5I@j)ePMxCch4=MEg5-7mXQ?pN@xkI29ZLzQ
zLPLxM5emACg_m1GNX+}kcaZ}5gwyvvG1|6hlyr;&XA)Iz?)?2ZNr7#p0xipwQrWY;
zy6-T7)3f>m5abUq?3`GB^Q7tJ<@v+!q>!W=01+=}vZ{4L=<}Rx=K-X#cO1#XJBc15
zLi_XFLD<g>dlR(1Wu~!Ky?Zfc_ua<AtXCU8#LL=OT-T7i7UL0S7T&jNd{5!^ZyzUy
zxS@?|y%~u~A<Y8!vboo%K+=D<{)olnVo9`THYS>o^wfUC^xlZuFT=9&)P__+y;!mB
zO?m@;F8sT(&>h>lo22M`UJ5ZTjS^KtV8lG?-kp(NzVk;9pq7^!NVe9Cruz8pLLf|_
zR+3rH3>Op1%;hEbGYer<l)Re7J2rgf)bY#5=ud(L6q%ewH};DFMPi;=>z%wb-T7~$
zQWOQXHuZi1FVGJcS9X&-agxR`fh2?7(RJ>89JH?;%&Z3roP4GQ<;-q>(oZzbp9<4<
zL7Xk1_v4FhP>hW?$#gUt*W|5s;rfCk0eE1yV`P@J>RwFKUKZV*+%6gQ#O2Y=e-Jgd
zMY{=Iy_LH@m@m|loflb{XYdeT$nH&2VQv_H;iwvm+GG&-ECMpSwtle=U!QP*i-LUk
zT#RSD12ZeVYjbf?l~ZtW$<@6=Oe}5=;UQezel*93tbL3db~J5{zKY<p)~{#QVML8#
z@)lnL;%<RNhfo6gORR6O*k%OcY7wOOCXV)YE+-TzpSBmVTr-sg-8kF+<B^Pi`;eZ+
zBE9N)79bUmoc(YWdDs`I<UK~*09i+<riX7X%Dg$}^TcJ1_Sl7+i7!u+4LEn}uEJP+
z6#DLPfe8F4ryqUTxk@I=Zu^rAi7}55YRfqB_gq4Qqf>X#uY+D=hgvXceun$ou$XL5
zug&(vPwGEAX|_g-{Ywoyi9Rj^cOT<gSHfgfE4mULR+-=Q#mJTuCPmzq>r-+U*SM_l
zO=U-XO+hU)DAaxg_j<B$qd%thE%8&ItN5NAedEpHp-*zP%Evu*X-Lo&``zN2f9k&_
zSPIVIWhU}MYGn4Koh?=*?v2tM7b3`vzt|VMd@0sL<{M6XJe8CPmsy0^Ljid71If56
zzh9>FQYb2+MSHgr_n^QgsK)q_AEltXjMxUmFWHnwf4t2=ZuTSz2>V>4`Vi4==r#XQ
zP22G*=ZY{2JN=v%hCZRF=qobujuYQbKYcw6U!&&zG9Wage`dwIEA@K@<l=f5Nz`?j
z?fiBG>3(u1l||NtW8y=1cL0+V;v*|ZIHkVU?8vC8F!3nBX;@)c#hKyBugIj>!@}$+
zo6Vf#nh-7<$U0`djA>7I2h|a(4ljgg#z~V47@ynj>-_C5`SwT<qlk%<N7td!$L$np
zjx%0g8AbBj%b4-6#BuA9skKlyl#g8nJDG{qkryZ5ps1BQSovF@MLI~l6k1O2c3Sye
z!wMyzKJa*xa43b!Q}znu;ZOn?bdKvpYk1WU8|T@MTSP062QeJCbdeGLYVmL4wB1?y
zww~gRU}UuE$c6v?2`m>DVGh<D3TFFZZ10X;t3>$B(sv6p4;X;6QRdRZmYb!*G#y&J
zLN_KbhB_N;MjBZWOH!k7;8*6gY)#(6C&aApH!CrFiQl2&R+>uk(db^jMo(ly^g5Gq
z_w5UKO0SrLDjlc81Bj5Lhe*4(Wg}O8VSLuXK^ajkuuxq*45JOZuRbX0+k~^jQ5f3)
zImo3uj<x|@PkuOvBk}u0=eP^WajkpPe|UM_a9?tQq&kJq`A$*k>KqOWnem+xXSDgS
z*+h5m`*SyubaV9$fgxSiZe2a4H>bsKDQi@>#4w-d=-(^8F6s3Fl$l(GOYi4|MaQ@Q
zWMw>jEL>QscP&O+`B<M5x_@<pvKXI39+xjT2rk`!+a~&EBf#MN_0#19UIxU$=TcD)
z3>tKkSR{#1q@rfpWDwO?x51rm_VGh1L{J&>bzS9OH#@_<>2%8fsAgwecL?dDr#?q3
zaa1|gr^J@dN@qZc$X~?;ZbH0!U>oG0u3gB9(*pZ|b}~I(;FW77z(C2RGp2j(B?sk+
zi_6j5kLXUcQEq?)8og4{C*%obb1kPkqj@z6MWIFZBQa&MjYl{xgJmAcsX|QX`%im#
zzWi2N`suQy?)QE*j?wERX9x@}nRWFqpT%_ip4xHN15rj80$YM+WgPNb2x8BqLM8qf
z|B95ZVs}&f(oTB58GZ^BjLTb2of;yxlLN8f6Ofq|`{_YBa~zdx*Q*oTUgM@<%3np%
zE;hN|JBnY^lCTBr{t`^u*yxZGR$oFpEa+{1J5>yZkmQUNb<?4<K0kbNori1F<N2OW
zYy;itu3?4FrDKE`n6Oi}b5EJ*#i`Iz^!aRbs@Iyn-iwbN&WW}5?pidjVt9Gr46VcT
zA1X6*3Q-WDXhYj;EG)2lQUX*z$1R2eqYe`hm~D{iw6(=<{+q3cZmFyNv?;i7tt&|T
z!uNTGK!slMiVc!5$$wmg9jVo*$b2h{{IU0{uUlN+sln|^FutFfEN4~O-QXL(w&-MO
zIO93DL5>w;7IJuG(x`t!ioI-ltJIqq%@`ZV*v|GI&SN*4$%@#+u+V3BTkbrs!MFAD
zwBnCq3mrwVNFRM(z!9zJJ<}xF&Bo|E6;1AJ?G^82)_Q;MeqO)?Lz$=ui2U5^ZwemN
zYj22Cbv0m^d@(i-WB*x^riI1NZ$x}=%kYxmx6wlsK|e^ZJ9^ib156voxX7y!p(xMU
z&JcMLsYk9ImXK0QNsU}%q2fQh7@wU%wsDZYr4ixDBlVyX<VOC5mD-;^_{a2o^GMOa
zXkUX2^-en!OjN2}k81+grO8En>KHP&%mlOE?Oupwm^qk5B$JT;*C6XMd{<0X^i}3S
zPTW=Ysh0;O%mu2VEbE7toYIF4(+igN^Huy*-0tdWm$c#E*kboAiYs<G9NLyLUcANP
z=R>#sRYFd7Svxp+6XAj)E>=8bJBp6d(N!QtSy~df4=v~nly0+}@57YwmalTRJA1lx
z3eRjUCrfq<#z6BB5o+q+Z_^F-PGv^t5Gy`O78yt(%GA6`2$Bb%5)%>gNoA9Lq;i9&
z@ssNU4brJnks(Gr?{5$rAJ9e!K7G+l`M}7yNV1Oz7wfed2?H&?A6=?mlD7f1dm_oy
z-J8tpNfGoSxgPk%qER65Zh&Lf`WUb8J7A%2YLUvqH@@?i?bPoQmJ(4c8>ohmDv#`-
zqzhbhDg1c#ah1batas1#-8xbB2DPk{R^R$kHZx5qIHbD!jIBmjI~|4%RfJM~D>0Y8
z{apH34-ua$t0MQCdH06Eur}+UzbLD<O{pxwsYvEcw8j-14*YyS*GX|9)|PGlL3kkR
zV`U;go4WTjk7<<NUmNi`Sx@s&KP)nf1Laf(Q!+n(Ac`<3FwQ2c6@y=%jLQ@i7j@!a
zWcf(iq1#IOLg#NNkQ4>yI*ms1Bj@AU5^`JEjU|z%*#RF*m5H7{m<^C+v4;Ic)&2^=
z$d7#ivf?tGpNUjuO67EGe4=`mo;@}kbgW<%L7!BeU<`H~<_1#<@Es+!AR#2Z8te`Q
zw<dLE(gPI~S+Qi?{lu`nF(1lbp}6uJT;(guBRrA(>A@Yv>v!6pC7}nDp24uoNW+;o
zG)5L9f6{sG)puMdOPn`PsulW|8>VzOZ#c)W;@5md+4bPSn6o`vuP!WmD)+V<|NgS5
zy11?yw$?}<AmR&jqV#Joct0>%bOW5_Z)Q<~-x!jcMk?Sr>?@tWwM`7q!L}*#*1<jb
z|2EG@o&&gl{(@{cmWSvn0E;opJDtLxL4+8`e7vFAmgfhiDfcq8^4qlZ<3ed%DHq!A
zBfh~SSa2s8zT3|IuduM{4Lj?Y)YWh%mwPJ`Xe0N@_`0C_^zPvga_o)rca2dSMdWs;
zT$SrX5_d00(z!0S&5{=5H=`6q&`U!Yl|^RpyW$#!^~);jZO=1?7e6)!2ZmYi51jdl
zJc)D~`hl{TJS(S9-~fbkLem#|atI8q1zb9U+`pa!BmaMZkNanlU;Gf!;bT-CC#b@|
zVq~9fU$G4P{Q7`gAx7*hpFHtHj8N_%GYbE>nM;$jopJ?9b>YWr#zVS>;4J~-zjpx+
zs;Cm;%yNQ-8*l<+%KISFX#lU+aHmhk9>72qNY5ir-v}9vZ`8O~B5GPI%zgae0{tLD
z4aQXTY-Tvoi;8DwCulmO_=G6n5i|#-Nj!vVG;{b<tkuVYlP4s~D>I@xLXV0JT$4E#
zd#-TWpK|}=n0^EojcFOvPZH{|N=J+(SNLIF+wfu9t5(7|pT)jQcrpIlT|POTU2y9-
zs1JM_$^Pzk81DTg(PLpb{QNUmvstN;+c6t=I`C*7{9uPb$a*xOp#?eAWQa7|?!gu0
z2?l(T)FQ&cRK*7%oqyOPWfw9gLB_4>P(p$@smuUppmQFG(PDG<6{qefM~Hs^EE{-^
z*t(Y5r#4H5Pd&Qp$aL^8IUT^WfeSvt(cZJMiYm-b_uDL*oh@}PSA%5C=Nc+^hV$sY
zhderhklW~KZGWfW3MrZ^)vkcIV)0mtgSXpN-QTCLdOp1%l+!|9H<iLzc#={02GHHX
ztHD|OR~x#m-v3e9=R8WeC%_oqvf>za&HfE<aY>kt?%JuVE`$`>X<nB>pjG)S2L4p*
zX=2UaHo93B$B(3h-Zg#-M1$CJ!-q@U=9^L6CWL`8QJJ^VR~&08KRPzV*xV~Hsitj9
z?WSUn&cu@|j1;?rGb`aMRwhdZv#?#ZA9hq$G3*SpkBLYfIDt41+N7$>WhU!sHeYw~
zJH0=4S2ekXT$ROVugGJsZvWr-+`0Z;!?)l6@;Ur)3XISb!rneTXm$|u-Y?&5Kc4tl
zLo^_#S-Pdjh-8CGbwGK_9k>r1yWQdAulg5{A!OUd3-7}!*I?uz0fM`|Bx0?4H<V3=
zDClA}YqohiTM97NIuAgaRp1FV%IGzIb-w>;yV#aTAs;@K*XsE6;DwzgTQ7=_bh7Ey
z@K(?MGSeH4Ql*T0gvrkJGr`Xfa?$VQC{1#@>xZT5L@jqf9bucQFGo7);cs^H7OEfG
z{Q(Kj5YKFzhpNQFn}Zy-@98r$S(859;QX^X;m9|na+H$pLbK^qDJw!<cIKFfXulF`
znoRrpC#aA&<a(<9<yVd*n1wejBb(~Op6s{^LP2S+UF6_moe%oTz`X1G{e|Si^S)zA
z4LhtpE*@qh+b91E$q%kLn+_~|-TeN{vlE|hN}VlT^IgjKXNAzA@aS#S2Bq+_CxoZA
zlDl5Td2Pj+i&4=v+lV13c4?p0N4r4l8&3Z+<hZ%*5^k1*$jM6r@Kcj>yan=P02e#M
zk{p@NQg?}^%InADgOK~|nFM5(U3;U`@V~HJgOB5LRXbioJb0<i!U4-xcM0{N-C<(q
z0Y4D8^;WrR4q4}~r?<eHn*-D+Z__fsh6YA|zTfXkm!bY&(N!3UuD@{g2WNxmwESr@
zan!te<ZhwM#D|=AqV!7lY0eCr_sxa;VP)GsA@8)y!|-Z8*KD)4em-U&yF!d!vha>^
zvTD`Xdo+K!`}UNE*uu{YF3a=$#wgXp@=36}%Y{FjU!W^o;xxzaKXk1llX9`UQc(8P
zKhO}NhuP6*v3$7|I(EvM54&B|GeUz-2EW;kMD;NBWMo9|_{Z5xJe(33+P?2j=~cU?
zrO7Bi(knD0WULZ?7(c)nuIpKT`o{grKfdIf9XGer0I7a&x;NY4dg5SJ1eGqZ#BweR
z$p_`nAzvdBG7GlrKUWYslkT9{k*wr`qwal|n-bo}zpYiM;zW*pjLXnf0HyaG=0Q!H
z2=YV&f%x+G$4*~WmG%bEwrF;9anO>^kFuH<BAmgQWJIPVkN84#2qw+mfExM}@6<uH
zC7<|A*KS?ns83jh=*IKbyF#hjARtb>7kyF6eSp<DRc)1HD|s&Ad)LD7LYP6QX+Ns6
zKuaB4xkm7@`eXow3!if;L2Wx|&}00AeY;Wb>;^fP0CujcDS!6c7~qScgN4qa?yFDy
zwG-t>w=#dzSupQcc}5D(Ik&f$ikduY<?FF}IJPK-MxDOl!sDL|b3`<RO8a(3=VD3_
z9q;yq|GK2xe@=H`cWAv_A2WK~&1k^skHa=)g}*7XUjzU-WTnT&<1&#4d5?@30c8$7
z@@o`^;xS#@f$hL=k?KAC?$4ea__mUzAcZ9731g&mpr(;m82JNQt<b_@Z;_zc{BAuE
z(|nc=eD6?w5pgQ)ErEzF6vb~H0u7j7Vxy=tOpHwA-bV^?ssQ*-k{a#8SN`>^Qkgl9
z@wFvYbEuC0xbrRE(yTyg|3mIVES5;n!@}<bjJ0A*DJv@CbsmkTCt!uO@9xbFOB@tX
zS*7^{e5kle$KekT6Bm@af1c^&3IX*4mPw@bu)Y`eVF@K&_|5D{3g4-_FFfSSw8Gs>
zxHix>!|7j7-#9lcwiSL!&Zuo(z_@X}xI9x^1NPQkSAXAna*%_-eHdenM|(N!@8`HM
zJ>iehX4Y@K`Rh!pwsc-VQ3=asY4%p_z{MaIVVXFVfD1nZ9~gd=e*1lzRT}t&pDgz{
zzsoyaD7y6P+%RTy*JM%15jDIWU4|3d-xWn~do^!ox|CIHEPDBEE%06aOj#TAV4$M6
z%CYH}@k?3;`H624@82B0A;!G*=fnn+$UBY6Iuat5Lqsei`oyqbp<?#k;PXRd&)|=w
z1rZgWJ)0!#S{{(AJr7GQkgpyhyLp%lCwTPlzrrHAtzMR+O?Bunl^yM2g_l$C=Ayu@
zempk`oNHB~b{Wi9|FL0DG~}8I|L@k<ADfbgh$Sn154|)U7NLY8OAo#pSTO}(8YT~9
z?s+Hfxi5Bi6gBmItiw_9qq<>k{BmQiV@JTyy6lJNsXe(rwzg(??zy8c2dac~K5I)-
zNU~VeKS^_CbxuR8z&t#4rT)eE%c&yA3Tl^BK5f)wzFU?A!<%>}!*BN{=jN&$i$iM6
z`hE^QVo_11C2cfqUzFlCt5Of~9G_gfyhy|#DyY-$cAw6QcC^@hq0roFpYtQDBU|go
z{L&xa6-SGgue+HOlFJ|EPN&_w@(LF#?jZ{;w90vUs`p{s#EI&BcsJEk8NCIuesYcK
zd`vHpU+fUEr~e^X9-z#TLxVgg_|qbFDDE4RMZ^(UY-YziD*e}kefCU<m_;<In)<=P
z^Ux-fblnWQcR>aD8`%Qz)HgFv=C(^7s=|<J*qkGb&+uH^xE(-$@zUl0N7`FJRlR;|
zql!t0AOaE!f`}*~5=sb&Al=;x(g@NGDxrXM3DTXSgtBPqbdk~^oi0E`8t(i7`+xS?
z`<(IJ@7^)iSVI?!SnoS~KJ%H#wQQV*cT~tVbe9T58jj}NvOBoeJ#hi))6y<`F(kRe
zkBE}>%E$wUJ!e{@?zAs{Jg%6#^FBW|7Dcp#@~iBMC>JrR`QTwQ@kcA2`{VDy9CMB=
zrBq6NbhI!h3HmbG8plsAD|uU{)Tijmc-~ZYp`>4QGcH&6$ba%Ui3gdu-=n0@QArab
z%-TTrW=i+`Qrv6Pt>ve0(ORdv-rg7G=jH0B#TaO3r2bm_+poYdU^&B0nJ3;8Q!99C
z7gg=<{%JiuG;H~cX~STHOk8Ky(z@JGKW<ikv2%jvYXdsYCR7^*zv?{--N=DU>VJ$$
z#~dBdkLNp+_)eWJw)--@OUdsVH-NvR+4H^iVjTa%VAQMk-4QNg$M`MWclWnl8@|R1
z^Te$BT_CT}@FA~M3CAIArQfbTj=}J8&r=NinP63NuA@G@pCx)I)sN^rTs*ivxQw?K
zUSt@DQBQjuBgD|gU>)7!?bjg0*7s;|kmGm2h2tNQXPCrBwCT_FV`KdzYY*(m;S#?e
zYdQDuDsSoWn}VQP8yDe?d`jQK+7em{EWQ($Vj&{oPDf48z_+!{>dNQLP-i~8(BrUp
zomU3$r!><-B`w1(t23i)q~mifd3vWTRp)eHau;e)n8@XoJXO&Mu)K}ITf^~7@)Y^d
znb${dm2Wt|!f1Mh&oUs5Su^(rzV-T8M>)^g3pUa0Int%tCMB2B<LG+VqSh%n^$2qZ
zSCTJCYq(q?;tCnLsU^rx%17}vam?5D>D8Ss43&J^TQbe@*|F@C_4w6nqqwYuv~z8`
zUP4_`;!Y1*I58<`17_ucfMt|gHCy~<CM{puD@SXe56%l{XLjQe9QDf87g=Qvvahe(
zX1LFggiTCEN|wG>WC?hQ*pFjapRd75e`drDIXHx)b3N3NgXhX?aE>f30a6}taIA1}
zZkjG;ml2(kM7|>Q?yNo&y!!pu>t!=0_Z)mI>c{R_1aUugXv@i<3$AVT?#zL4drXgI
z_3Flq;{ry9-R76(fcqjTPiig~DGxT<-&83Ty=9viDO^IZTCYkuQnr2DKdM4j&PeK9
zl6eshdzuV=b3D)M8?=}T%5oPrVJ7ul-$eDX%_r5g(Nv{g2Ck*rN}5+v(puK2?qX#3
zSJ>tz^GKEI)`$;G2t$sAVTW{&ddUZC(#2Uzj||bTE=fBdM#>(Zs5ZQXGoST?DEPAH
zjLcq-x8MMKlFiGPNyhs#9izWDb5aYP2Kkf*??^acR>scM?!UbIP}V8ZR?Y6p6_kL&
z`r0J*Pm10&8Nq>0iT%cFCGmchI_YtGX`QM3<{1Mlr(YgeYX_F&unTBp;Hve%PZxP*
zSs2dC^;1?K5hBH+L@$=#b{#)FGsK`6$O_?U5WHA6dqzghtI<N04HwZD4|2cVD{A3A
z`zt4sa{BNW_4x=SlZVf(z8t<bHMn7VH*>r3Y+mHl;x*ousb9PdKAelZ0uQgFdt4<=
zl$jrno!A$4wALr}x7?gRH&4vz&s?jcr1{W)dFKaz;B3pOTgLrqTD~jt=Wi%y9BaGO
z6X{#2EMW8DJN`rk4X^c?<?rW4h=01d&WyNj{cds0n(DJWQ6gWuZYnoW_#w8Ch$z16
zRgZ>I^G<lQDLv;`lQA{1y{VWUy9Vizc(v-)!eG}~kM*qL#}ffvEtT6Z>m@ZZN>9@|
zeNO+@wjlTT@e2t&H?jwPzl95*^gb9~VI@Q)PN0hCEa0Q<yRp@dJmjID{Lk}?N;s%R
zmhgJM2k`D6m0-*#t{s~8cMrg7lU*dNA%Pb)G)^NT+c1w93VIOZJ&rpeS4(&)5?`YZ
z)xwUtz0BN}d@#?i{h7+5F2m<1>tlAJJ+k<AgmL`fAh&)e+a(RF!nEUqGj4nj%NwM-
zhWfcUVwjoc6=o|3WYh2-Fb@HNg?KevOZ0->j_AYmS<=GSHWm_BbIE<0qlbc>Mw*vy
z{CcmHHNS*gT_CSAeS$we?b}28%<-=tlO@8dGwP=vqr7r2#__OcKHTN`l0XSF&QVF<
z@n<L-8OI#w!cKqAviJ)NFs4|fDHb_YD88%v{+3GqB`J#j_c)VHbH6@)cNSgyS}kxW
zdK!eV>qsN~MAPA!{QH@J(9@9e#DP?@fK={gkMuDiQd#T=0%spAtQiR6xo9~ljVtKz
zIf7sL4ERNU5Yj(+!NGVPQ~hV}x1KuiuOCkwryJXp9a5XW_cZCUJV%&^mgmf@<%!I1
zgTzAXzh-pK8`*T<E&SP_=lv&|m+QLKwQ{8rPotc!85V_t5xtAOau?$A4pmeEoRIIQ
zl+)cG)Z?yudyD#J-TJbaNEnQ(Shs#^45jj%XukIo*GfFCyv~mOZMKLH$y}@yN-MI~
zB&x-+?Cz&rOP{2tz0>u`C5T&4$rBw`=ScZwL6Tf`tQWr_z%g!Xoo>uAVtMw{e0ek-
zuf!+ThYr(iA7!;ZJeAwpULh*koYAqF+F~lPTfmRywGaNbV<)U6V&5Zso8V9+?U10&
zj_!frq2T=U`NulQ+x9Nuv14$m{InozXq<qbi1@yE51z@6NY!(V7fvwF5K~O!X=u8s
zXA5uKhGRC{Cnswo(2DcFQ0sDdeYddj7CU{NK5xJ1JayVPEwSi*PPlWh#vp!mfhMUk
z22-ZLL^#jy7GZqW*=6JSio0wxLD&f0Gz4cSQPI6&rD2{kv1bAm(5!Oa))7SdcbE7i
zPgM6<uACZk9NGA)m@UpFm*zfH{BXAcof2B9-Ce8s!H@II*4T4OSN5Co3X-3`Vo2ir
zh;Is5khq3}Ie5_=c0=6aB7<mufk7fnc;1+H+kb-P%6ivZ_X|dea^YST2<qz4^P6(T
z2`vLdbgrBB{O#|IaXEmu^5~p`B!O?(k_1ui|LL(P3R-(P`&G<=_PfGSmIgA>fCYqf
zn-C;XrsvN}If&xtDuTG&slXmk`$ZY}mR-}Q><?`mImHQ95le9FWDYp1DA8wUo}kS8
z?k)`_%Ul=k`dZ8uG8IM?)V7%aK1Lwwg0K+;mcPYEVwkz;_}r!azMCiebg<?($VmTj
z*;*uXEiy}Px9ppxqfapn#*}M0d{9H-No`HiHb*gK<H|>j-~Gua<Z<04mF!E58Tu<`
zs8lB2jVtgEC08cb=ZTYCWBsy^Jwi8^K2&Izd(l*Ar*eg;RTm7ARa$N{`@h{9xb5Ed
zzUJiNO~+G%SLTy@WQO%WpNmTX=wtjX2F{a+qsT|>+>8I0c8(S~7Y`cU!?&v<W^u-I
zrs<)RY|GZtn+UGG0jYVuJFl2`Eo-{<KH4-t&r<)8V${0ZAT8JXMB!;~Xz2YOJ0InO
z#J{xp^;YOIZ<*n3TVuI+tP@z8s+YpYr?^G3iGPHf7>kT;LOlMg&+&z~WL$Pw#q5j$
zHadRor{vG~-HC<9b)|9zzj>cEsrwG+i>SoCJNP)%KjIOc;$H~RLT>usj~90lGk;D=
zFaVLF$5%jorauW?Lym6Q5HVY*Exyn8VBtwFRu>fQeBuUi*~J!c`0VcA+6Wremhk4%
zq5LTm1TCnyZW;wk<u>a4jJ6t9y5BSAsNZdrC7V@j)ckt<n|#Anq8@XRwc?c%r5&%A
zDn*Pti^U^I$MfWCFg5ymO-*rs(-_6rz-|GQZDGs<3(Rs1m%?dZpG#sEB)6{l&b4|u
z;NYK2_Y?`VUBR>$;=2y)K5PzS_0Sl%mMz6bqc0E{4+orJ-^)%dK}FH=-p$?R^nzo8
z&^#iCs--{Dtx}o?e@_x|B`$&Ee+6q?ShJY7tB29TK{$7$R)_z7tb#;h>SU7#i3sK@
z7@zx^gv5wQ?f8QLxn~!jUm}Ggs`L71@K$g~DAKS%lm`Fc{9n*x-@rzvqjk~@${QT+
z(0$3#FIv{z2rpyj=MBS|LerL1ww0dmda<mQwxspJW4(=MeKu6A!{#>T?PQ5nemeOR
z^`G9<yhKgdO=4m!M*a1%zx!M3OoB_^POx3NZ%#nW@J@!^y92+IvYWYj74WaK>kD7z
zvZpA{*-LUtI2iD;2Gh>vcG4;K1Wg%Mh)jIbPPlv~pvzn9j8a?8QOdA?6;x(V!F2KB
z|Ew})!SIPKa>%VtBMxd8RA#~Ob=2W`?Sow=OaE#J2Kn>nuz#5nWeBs(NW^N7AGo&V
zYcC&h-0YeGd1$$yCDInAejk<BZK}q^TdA-T&+ce+zBzVRUk)!2QnVR#sH8|3CpDA4
zXuNb`&1#LO$VE=q;2co~dgti=aPn%lj&4FKg^Ben-|zF^%C~>dMXnm5gG0T)H5MP&
z${gM<y75fCH(-=<<8_VY?~V-7VN+Em-hyi`yWh2Zhf7|QeAv~tR{<t>103`-^6v>x
z?!SWd@e*J6x0@77b{ftN*#ExxY<NACA-2bsl$_|$7MeW(oBK@e^tJy@Zm7X1umzk(
zu6~0Qlxt7PVvz(n)gfW97Nw0e9~wZL{@k-lKXvsVA!%bP5(Y>{zd0&tU}ITCI)_&5
z6w*&MMM>G9_|1nB_a+RT$~6s(mUFtr28oKK%O6!IH(Wf*geHF(_%w=dvzcB-t>ze1
ziaVAptXk3ebrc4dX9~o{BuVVp%&XvBp)_}JOeLkfM4Wve&f`a`ZMoGhNnI>T!1VLd
z-CZ^PwB^=Ql!KMHMb8vN!CB6Kptt{6;k<wKGLLKAXUzRq7<7r4E)DVt|BvB0DPmDl
zJNxL7cke`k6vZ4_X*Q4-#FX~Ef)`C&K)uxWR3SI2+5_nttbAl(@P-*}&?&@tXw+p0
zgF-Lv?<^cY7pN*wiod&<X)}}`&{J)lZm?z~y8bzg;bD-*pLF`!f?6S8?TV-*lH7<G
z^N2S&Vex7U4%y4RA0=+h_hU{xypVH2k0YL*t~_CCbXPzxjY*MhNdN6n!Q}V&_1P-4
zp-r7_jLldxM+tbA;WFz${Ccv$^su1*YPP()PnIh)gv;$EGw-&b&+o0ik>1^=VHzy<
z`@HKQX)Lipd7wO1XOKAi{6#{IgB<Yi&;R=~6%sTz;J*EdWF7)H;e;%&AujsZ$yRFk
z3ZCzqC%EupF4KvBhBQ&J5NE%ay0~)q_93aqb72S3Ylr@o`)+~?(|IX$3tvik&V8x>
zqVDk7uYLY3S#^O%U`*PNBO}1xX4(7WZ<&^nwxF~_1%Z1wxgFI9uCyMap{u;BaeF-_
zSJ!qvvQ9Z{(sg}q{Iy!7;ojNI<ZLuHu;hj5Ek6A9E`dm0NL%U??cYD$fBpOadv4aa
zh$w<t(woEd_lG%F=<DAJ5o@D_TGv-Lt<?+dD5Ty7SlW}D`=1}{CuZ?9{Bnh>=PN}|
zhGhKt#YZfLO*u{hJZFB5uC}U{1WebOb(i6J8a`{hW7sS}^~F?B%A`_i(|RTW!Y(6R
zYS!qwnSx5K#AU9Q*=3)Cv|1s8!w<1d01^LoH&)dD;6@J|>)-OtJw&_Rdnif|LU?;5
zzGBXQvtS7qc|Tyt_`yF@CMyxJi;MVKZw?AB)IXT9cDXsz0|)OW{+{dfVg9uG!*ZVM
zdb3tbwl5FF2Sw(x9&VME@5jGsCU6X{cQxs-WJR}lKuDMLYB^_XquBPxf7v|irz&hA
zyK~)n7oi?#ed8=TIyhA_L#Ldk$YhyWY0D7XX^tvNGC1C1*D>&Vo5q6$`z!77mkbWm
z!aiTP)xj)E(bEUg{O`b0z#9~n@;q`x%m24|4iDlP?PY}}4-eR$Fhi?|R|@Z3B9fn#
z2<;TT?{<B9Qfq6%yHsBG?R$0AY5X*7cmA{%6^qV2y~_5I%lw^Q8`&Y2i;TSVyiSr&
z*hsLQKMry{d43j*l>&S2Q63l-lCQQ<riGE(`qdQ|MsLU&S+>}Sls_4qhNy4K-;QUX
zY{RjFc-ub~^^S9qJA>hxr=e8wwKKgrn%>dhKRlWeCrQO}w32@;@B&kH*L*ZJSWc%a
zj+WlGp-Sk#2}gA=o0z0z%9}T5$$^;ZV-KTReml%?W<blTqQZ9Z{7Dz@WBcVzJDaPS
zz;np3z5;ZsH}2i*3nOLR#Jgd3^1mN))iWeYKgWc15T!FAVVAyI(8U8&dl{UGw#uDr
zjEJNa8J_rObT$+Y#-Y@Yas2Qei#@8g`X@3JS9*mW{=Fh&&zO6uK+=bX-BnS?iOxd0
z;}ha*=On&&4iUen(*sX@H%}wB{?CG76r29@rj?EfdUpLiarJz+QW{*D^S330l$!^0
zY|Eq^Fx@#<&drTP+|fTPFF4tO4h_+6|Fz_nz}iJ(q7ISKb<M2Pt8B@RvGa-(xt)rC
z{FB0%Oi6G=-z)cMUa1YK32bmr{7{pY5$ADKPWywD)Bfc6<rx$iqOZ(%rsCE$G~jA1
z_7~7g#d3$=F=)L}G6gHB!ufWVA97ocdm4)ZFYf~&olBC8y3+ydi^8xAGXaPK<>Gi9
zZJMIp_crNmXWL&^?C&|JQvyo#<_;e-!AG{2KxP;N1WOZPNu91u9u~tT!m!YS@LAu-
z2k$gKo2mKw_$0yzjC#c|(BA|D>*-~M)f+k$ynwIt%Qr+z18;f~v~X~ozeTjIr5dF%
zn@v1+_|so}k0ia`U8idFXJTR^bJQ{$4JT&=%nIr|MY!+p-i0l<Z<3OerTp5~c}Ew3
z7>Q^v(p~$<b|p}hUA%P3$msH?8J06cmG1n=t|nPC$%9}Eq|F_13shNe?;|%`wGWQ7
z-uTeB-1Y!nWV=D*&2$ib;r%oEn*4wmQmv<iqz7_Vb@^odxmod_BYDDreYM0!?;kTi
zY9U}$)maS!&+}97ou?&x)#(>d?zD=3?rgl2r6G%GeZIPbx%&1qatI3an_`;i8Vr4W
z*e#~)FK)Ko6efwg@)&WnU%=76U*@&gc?^a3m>Y2-ptxeID(zn=>~4;At)YhQqmQ%f
zxy3pKL}g_b@OT<(rM9zYswQI9vdCs^Dw8F$-P?CmnCl$;q`|<<2h|b8!Z0HBKSm38
z?VqrM6E<zqZvp$*Nj%2O*7k-GXHuCVQ;|0<2V^zupkp!lo<pYpEk?~YlX^Gvk=IMk
zYY8e2)44F(pk15}J4$aX=0JawU9TPwCTb!Xhcn@i$UMhkKJ*Ugj2^-+=cEK`=kn*0
zriA9B6;ePrL^PlAXtKUQ=GK#aGWXpr+cC<YFpX@}MMy~a>Rrrr|AkRvio{9$F}0gi
zdt}$IT@!NfX#qy8tXt2{qDwwLxqv7xJW?R#dRudlSNvc52}8tBNSC}hcUaZDew@Ur
zX3H)uN3C}MYw?xMj+~vz6j|+wDs+10YP(o|3n*}lPjSpf^s{Z_*)WvBRW_yoHJW^R
zL4({0se?Y6Q71i>+&42d;guK^fecj3!=Qk3fln*`Rxa9olud7rJDR9K&er8DI!i9<
z&TWeii`}V?t|n^xmz!7e3v;Sia#h!M`N*%Zj)vJ19vZN!RV3+LZQZr4tF0Bn!NpC+
zCc2(NeTh}~q(5-Me14cj?=>@7TfDFeWF%M5zBPHuDC)lFl;x^7`pDCe74;leE_TCs
zrqX&q=~*Xv2kDU$e(2Va(n_bnI$O<hCk}rq{y2nzcMAru5H23y!Kwlo!r)Q-K9n_m
zkx>WDl>XVw2i>_EkuYR02`CA+@Zf`Ai8}E+LoV5qH0^2YJGm}5YF^Xo|I|-@2tufi
ze+s#w|Laga7Aes4%+YWjT=9#jlSpjwS9_txigbK->OF%=hY4Gx(p#lUC*duZ$9Kh|
zWOBodi<Vm{!b2a(#8*nTIR(m1j?EoMqISpkMFDrQ(GN?$qy{Pb-?)9d>XeLpC$9S5
z@OgHF0YN$&Z@$P<mV3)G>29e%zK&(}zF5v756?WI<`Q?O@Qf0ZhP=G$#|5@}MS`=Z
z5mDhoibJl&#hin%r)w|+aO(`e6o4zm7yuKB08j6igqj{#BI2(=t-QYZzkK;JpG$D7
zp#RB3?&!0Z2;&?RlSKk?*6mze%08OIRykuhR>d?)GAYM*TzxoWQL8fnc|gQR%wg8o
zE<dKh)RK~N4Mv7ZS6b-Zw&v)csKXo}HE;H*H<n2!!5#|&AyO(JpIR>HHdW}iM5M_V
z3c^KKiBP#-D`WIJWmH7!X;_qU8~?~9)`It|Z#uAZ`!rwx|LX0&C<R+*N2*r4b`Lz_
z6%kA!^*`fe6c5yC?Cw6PJv1V>{a(bpXfoY#nzNthF+Yuj7itnW>!bLx&U0kRsr{l#
z+_Rj(U!&p=mELob<BWiYhL;XrqW#SXlD1WJ<(2m3H>V00+R?s}yj)d`glY_iFc?ZQ
zrZ#u#g!_HTgbumUcIZBawUcd?{imR+R{LT5Jdx32YaLRAS6UEoQ1)tK^6-TMARwa>
zdT}O+{i)o4Ubm$p8-qDC4Nl(#B%S{88d0S4e&!t!h*x670-to1&Lbn^#yl*gOS)(j
zzTE1Qe?V)7gO4B4MpRuOd<M5ATCgf(xz%c-F_@NthK7cYs$z80>p0L2a?*+9KFGNc
z$0ZXHM+yumuyLq6`g23Wssue;)l4C?r(s~^*BKJU?egIwQ<Z4gdXtI9zU5`ejI69T
zhDAhK>nF$hX9D;St7X(bo*)q|9%T!Xyu7G`a79D&7~}Z^i34>i2)uY<kowiK*-r*2
z!v6Q3mc1clbeQXJHxBCt281{MtJuy!{tHT7b!J1t@ok!B-$U{(2iS8S)Ht0~_|7=K
zzY$UT<HO(z?YY=HDy&pIXVEpzrxhYe));PA6(w8fj$5mLbi$R2s+6~6Q9mttd4-k3
z>C2s!_~E4ycLDn){*(>VFSGM)gTEJ2`DbTY_R}c7EQx!wBauu$L^8q$z71z#Q0<o)
z)lnN3`~4Uiw=yJh`t@u1#2rT=y>f?@SDZ&RN)x8D=+1X0N(>^z;>5{Gc_axct5a#b
zT-qch&LVCF(GfYLAAIvV6aK0T;+HzkTyec~h-Sd?Kp#~S9wC(kWdbUx*Y21J=D-tY
z@Fn}707ldZbWJfq;CP*ugpUI7Y|)Z<-Fe!WaEW&97)cTk!3kCF@9_ZZrPv0n*L>J<
zZ^(!H`t?V%>bGnkIDFT{lKe<`Jh__9VhYBRU)_Tp-)c_`ck+)_G`#09HE5Df^)$ph
z<#kwoti^TTVOd~Sw+)>#IpeLE07G3bVD*5E<En2ZySNIL`Mcjgzl_Sm#B?mpLkGJP
zM*O&_$6vALfk?jV!{*3fnIl{5JxiLY^-X7vqd15BD#I8GC?9Jbi2*&dsy>9h3!vCq
z77Lh&uUuw(U?t-G>G96uLVKdeIIflh2w&S&)lofw;5jHc*_nfy?GDm0+0VLbhl|J+
z2GM?qVR|P(+MFp-U9bKX=J+6C4pG>rgR;wu7i51Aa|q9^|E(!!z(`N#Q5ycpe^v^L
z?>D%n|5iBiMfE*!n3nE2$RQ=p4jI_DD_`jAv2o!5QMnle;f>#ND~%IUaN4U8w{KdK
z=Ll-?n$awGe0|yzomN~x8rCdI5T`|Gd3T9}twFxy)utxj(@#b{>)S>SwJyS+O-}!<
z2T_9e8aSb->yZ%Q#V%4a*fkl?CiB5=ff$AOX$HG-?Y`M;qoBA0roSM){Q;+5{af!2
zm}UqD8ZaXDQoCHEoc4BEob%_?6B4L9&{ugWn2%$ridNx`ZNl;;cDqH@2O|;{F#jYD
z<Oj*HlhWD6UcMTEIU2zfu=na1zXk00m2?O`kw<5j`BT^89@&ZG)VvfgUS}qq1X>=E
zt5>f|kEVmwr^Q_VHp9ftJz#vr-kS=Rr)R~V>86gn5Vtu+7xNIP$GSa(FNX#N2?PfR
zv+Fls?1)SSKIm9jW+9Hy>Wr%s#1RudIS7T|(>7j!Xy^nMa<%`r(a<j;BpSkQRy}bT
z&*EV*Hsv2vyKTYT^n8fW;c0homXO|ku20nWc+8F|-P%u^<WKAB9FDU7=&{`DnDdrg
z3N)6q?iV9#FiEX>B1>gJk4KySH$zt*h3BCI7lSx+jmnO?Jo%<kbTqe~HRjktqeiex
zWGz+YSBV{ijKlAD)dq|gN?vI=Hcd4;8WV1|^}Uw=V$6ia>1nv)D-=i}YCH0QdTT5S
zcjY1_xwH$PzxKo#yX5nu?lQkonbW7vJnC&Y>Pf9)!>w;h5_W_O=<3pVd3kksViU>p
zb;Dmw?<nXnZChNk3Pe2O;AO*G9;to@Hm%D_uxX>P6>3IX<xwxls@azIVP#1-FhGSd
zqj#Luwap)2e`t@3ZbAX|K{yiDF7)Qvi&a7zB1qo{jAHLm9140595{*TUlDc$(~4vO
z50(wg&Ro0#Awnwn<qnu@*W~73HH1>~^}RTbiC>;@c2>QHBMRUWgi{TRXzoJNkri;%
z2H2AXv(K+W<qXbtSp>Yp?x*=aHY6;1hjf#Wk4FQqGGXoSxBRQ|nT`|$WP<Ct4}HQT
z?B7d6ubkZa-fhfW%3K)8<u@M6p(~uIvP^E_YBTJ2)ZG8jU97AmCKpM+<2J#0E_yjU
z>q7zW2voj<FwYoLlxT7A@Cc?1A|$m7tjMD8-Sf1XX}n(YExD4aD`oT6?BF>b{RKnF
zR%}FDHATcrTz_w1gX0K6Om<j+2LBTAkmr`I8MCAN<&8~t+bi5nxm+pu6>fj--Bi*n
z$@W)gfW+3Jf0CPp-|bhJ0kf8fb%ezo&DD2}!M~M_SENN7Wa865pV;qhYFpMld)P_%
zVF!(dMY~$nAN$xbypAD=<Gm)375*aRs*||G^cu5Hg(N8A#h-mQBjIEz9WAauj#%EA
zY$76ug?&=-eD_bij*w0_xbr(~xs{9RzEc>;fz-;%3bTePgvLc?{hh73QRL7F&%C$G
z_f9q4-n;s8Oy$NnF1Pz$S&IbWhBey~kqkCE<}Uy0qX0^=oz(~P*}zm-2<E&wLTYD8
zI<(=T!=LTJ)5i`gdw8Isu%0FDpb8O*)UxzXJuE@4=&=G^<umn5DXFnwqJOrw#vC6z
zs#yQ<1M|GCtF-nLTPtqOvgW98izbrodJ<JovVZ#}^zOn9To(rDL}b3#;SeWS?v|dd
z)&}~LYY%4$PyVFgC3T}?5e}pf62=j?7v{2Gu_wRxgj4HHW!4#5)`kqi@;QH+Q*<-&
z&~-*S<GLzpY{E}bZSPP4VKyz73zH))(lUgzZ7n%WLt?_3-5c(BO!dC=H(*M?84&U&
zPdVuIGGoAd9f8|(p@BCHKj$Q!_@+}KzFRLDWMuFc769GtineaYmzr;<zxwQk#Iz){
z=`|=7>r>cw8nTgQbDzI-*kyZs9`4jEkbA?4M0YqA_#4DNefp#bdvtpS*He~Z3kPmA
zzdPUBY)1`UC2*T`0DZ&_co%d00|#X-X}rR##|u0SM+RV?*hiQj^DfRs4IEf#bwOVN
zv*S|D>+|%}6a>_Jy4=};Pz27t^+RK!;5W?KICBHzTUR||SYgIB!PE2P{NNWBCe7mZ
zxmBPi=xmK*V}Q8>@0Skp4%1KTv=^%@0cBwcGjhbh5BSJVh9K@-B<1nLTi%I*E9B=t
z*2f19XcW0+d%6a1Bsi&JLdsC5ZoN`RI-JDa`(LjAGsJeh10nXGYtM2Hdlip8s|Ke=
zQkot*%T{Ri>(?yvbIxG$blfgWYzM@3mOp)+W;;7!Cv_vQ+6n4JMm66~us{8!DP`I(
z>@c0uW7EF(vzddO&6V&npQUmaU({$=;{6Ph{*aK)d-mJ`8{83q|4X|xC`qVh&dSys
z9Af^H+V%S(q(L?yQs+Y5N`pU7xBHGH_vV6d-M~Ap1kEO0_GeEmZU)h(8>v>Hp}l;~
zMk%tCnpdJL>aN@Eo-BFkkflT-L$4Wfd&gcUB?b+72Q{G%l~q6K-G#62MnaQV2&yhO
zk@d=<tD!+IoX~+^^AFft=2%MDDkTHU5z@f-_6YGZ@LJ0=vRHmm#t6+h0Oy5&1e3-i
zC`?&3U9K*ex6(~Ghy9L$ceK<#P*6}1S)|ColMC~oOxk7deockCalP+XMScb_(OaYg
z9YU#@M?Brm1+N^nIvrX4KNUsw9?Lo5m+D&xRSOKxL7}4ASahB=_Id9F_-P*hF~b@d
zK+1r^m-lgJb%*Il9{3wy13%$-$Ts}6Oepr`J3~-jJo%SNnn84R^LW)Cxb7XKA6|Kj
zixUrV$>vC>lS!4Z8wWH!H({@m<6mv6P$c)IH`Zf1tb`_?9Ju>I%45IBTl-tmP`QWj
zQ;d9x!8V`z%CLFex34i2_92+Q6y~Cw{ebor)eqThfrfLxpwp7E-e;UO&Vw11t}YnB
zA6Uomr)+-GxZzZHT`AQ}GyCSb#m*3<KiU*Rw&$=zI?rw9oAu7Xx6ukW8uhJ{p(fJ=
z6_V6ar&B7Hs3U{`U8Lg+ZtUG23!b)4#Yw73_%BS=_~q6^yoRCiIn!nbuB(xx#fRms
zpVi+iEg_muHG=XMqySO*;=D4OBa@_;J<)1ub5n%R$a_*ozDi|0t{6)4vmT<tuagU8
zRgmJMKYI|sQ2!u9j62!4y$YQDnA^7Via6)a*$O5;+7oO0)>T~)R;F2EW2}y{qAOpi
zr(mA)T3%eDA1Ph8fF9emjS(tdhmS-EH;V!ko9>@(hVSUM_SLs=lN0=)iQ@fi&x8vx
zy&$gM)LBtm3ru$=5wdv!hTdr;BTG~s1zwSQ^5ltkX(Dpz_r=6o0AZa3Hl^hX<ymxb
z-l3WcR9%8RxqjoxL8-!l0{%0i%sP1xTK_>{h;g<Ll(`|N{(iizmVqEn%F{;{<bY2M
zR9_~dr#-EGs-J-Gs2I6?`*=hgKen+1%Pwu`eE25X$9<@bw$ZuQd^k~O4~WOiL51Wx
zRB+G6it03lhHe<raggro?O2n#r#~R7Ou}7J>5;S~*P|bcBI@dVuw5VP{JUoCo_djO
zp2Um}iLF}7PJ#jPO}P~6Q>RZh=A+!Una{=YMj|lqx^_hUAF%cMT)7)GU#9iI3dO(g
zBrv+fFAh$Zl-u{NL;8jCXqLMGaeOH8-9n{K7)?zwtRjV*Jf25P5W6?K)WB}MhqGl@
zBYl7D9D0m-A>M93=#G9XHTf;p6)q;#a<8;j9R?JMr{U%uNvfXL<UJXT+g7h@XV7zk
zI~)TgOJ|=A+bjNdRLPM&p0#b~cr~Ka)g6I~@>(4SaB=4{*#iWq1P?UQIYK|LzbOPG
z_@vk3!M&&a2wDhYY-mM<g0YSZGg=v+rILpU`A`Z8S;RJQ<#co;i+JsqTTYxt_zR4Y
z)JXy6aDj0}y!I_j=!(O#-H>U$8=see?57jz&ly5K7o=CN@Hz;<Xm~OVU<U+H@ngb>
z@{p}zPW!v-UM}#?=^j*u*d9=$V|f74Kpc1@VM|NPS}p#~@qizx;^7zO<xB0Jy_AA_
zc#mffulH&n@Yn}6_z=<0&cwTu=srcq#f~^FFpVvKHtKkVi~)<o2!pswftTSY<~i7{
zkP3AM`RUeXmh!GL%V6v{lDO2k<C#lWKjHVqxB;<%7yc(M>uL5v*USwhKQL;izCaYy
z!+g(t=$<Y^&r0X7MF}_a9Q7h!kle^$LyrkR!<cO?OZbnOdRQ}`y+D>$c;2mkWz~0y
zutEo@_>j=2x^R-xmxM_i_(O=S9DwijqE6h6qO?nqz~OOr$9l4Hj9qHJLh9QbOt%o0
zyk5P(n7J+DliD|<yM90LJ)0r^+G0POelr+wBcRe`cipl{zpa#|nE3A9rJFzm6E~$W
zt$LMLVpv6YU2?t+1+campxr#<xini~h{hQ`W4|<F%eCP0c_S4M*raW?q;D%8`%k_8
zJc9Fk@<a8YlEsK%smHM%J06s?3BjnkrUdn!2CSf|%<+GDRc^$qvRmdH;eNve-g|SL
zoa1U}8Q6dDxueLDujL!Pd$|x`qhul%TUh~QZ7pq^t(1V?vLi{+DC+GpM9=mY%v3KI
z9yexqxMh>IZwYY0|I8N*Of@1(B*)hKBn?1B!HY?Z5e?8WkN)zD$9Kmd+=Ns8%uVb%
zQuj*f`wfz*67ECo<O5+V;)0wkR?~vKr(SzY*j0NkGwR@wgeO1xop#svqM{aCgL52x
zssyRPTvw*E+{yje;oj^ovbDV!p5Cmxua0^<SxBm5qIk(Ds4AKRwwGlqOUtQ#-X8a3
zxd9m1=%hhm2ZK#B;6r{@9KF>^v&-;|kCip}&tgAqV<M~xmLxO`pQdyF=7{Vsg3Ud2
zZ}f~iQ{*_5{Fi|j;hAqh015ZT^%e!IuUydPDLETkp##-oCC6oIo++B}!ex7GdYKD;
zDKO)%q(x7X01Vm>pyLy4tmV!E(KqAwPomL`>MdUlR;)Ib#}r{%#hA-GCf?CX_juS;
z%sd(D0^AINphF@s$E)m${}!*%Z&~w`uCRKA3m5FfV$|o+z5H=gd(2u9_WjTfrU5mF
zEM(nYk<!<6=p-Ui`ubEKlzoAeZsw7qVi<KrmZSN{1oS8=V^eeo*>-dIEp}MX3jA$x
zS!<!FPrr9yPa}??k*sV`T1f_B+-8OUn>W-Lteoizj`x0P=NRy&XRgMZx_#%-4lqQo
ztk=0c`hkXpK{AlBd&?I9qzu-bjH%tS2QbvauJriOhW-gnp1uCB-{i{^Jzh2;kvv#<
z39#_!2oj1Dh{=)z$+WzE=raCKI0Us|Xf<j)xNLgZYP@t+Hq}h*!8a!58<q3!Z=XFa
zpF|D)4AUXG!>C)4@-)-nG7k<q#=d~N$4FFEL#neNdN=J#jh26<=Eo2KJNB>-BSQIv
zt#>S-_X(up-x&u=<PXMBq@%KZ0wcbz)}K#PS8Dc=9&`U)JiqPA&Z}Dy5tJ5q`EJ6k
zBInA86vhwTUgXsUR$@os?5RJ_ar+g_b}d)7%%*jxQeb!+`D9g`NOhGnb$L@L<!SPf
z+vQjq1)Dq|of|;GD&a8wAsyCFCrZ8NG@cG~!om$38ykD2nf>UyLdaVbn+Yt(@)4HL
zCAEJO2b`GHx><xDe}?(>X{-H!%<his=2WD*T*_O+>k|!uq!{c5^A>e1t8_(@lMQdl
zlEo$`IW}CLM=C`OIZ#p7tmyWKoqAM0oKcseO|lJnS>`;R(HJ%7_k$onJSUsQp2WN>
z&A;8TXJs7I1dje90YQT4_Ku-@#f*ae6T8I^KgGsSz^)<qWd&^0h1CuxOA$Q`7UQ1~
zY>rPGT-6~QqlP9zrkOwiem4ymN!wvJVWNTBleaO;t>gLOnu->JH3B4yh<fqZEk2K?
z>*?Qx<g`2S;h6!L-|lk22vi=Vh)nOjdtzU!gFzH)0jd+?KMeL!S9uF=qgleDV5i_s
zN9~(b6*C)CCwD9+yIoWkm@iyEl#CZId?3(#f0iim;P`PN%J}n&=Bn_4EyMzeD^Cy=
zl|wq)^Y>7u7Y||5@I@UY2vxahjNyaeLo@EpK}Q3LwT6z{J<a>mog-~K9&b71`YDB?
zT$jkT+$4J}b@3+eYJGD!<0HK-o0QJH?4`w2lC%4I0|4fq^Fy^5H@uuMNl$&trXUod
zF|=qq9Y7&eloiqSJ8&PrE;cpb;dK*(6a%bYhokYTj<KV`A7hv*GO0#All+-AWvpxT
z#_aAo6@Ej9NHUGEHS}OUuruBd#0&>(-wIM0KIxx)<vMQ?-GY5_kshi(?DiofNPrHy
zn9A5VHF!h>MEk|gKS1F*u=SdJ93XD{P(C!Z4EO+P#^RP4@Yg~-0p!GSn`mWN&6yg=
zzxQr@V&YtH73(nxS;nH%r%yN6hljqk7%IMMW5O9if%_S%g{iQpk^zI3Mx{tc=qp9M
zTzfOBOR*0e?I%}MvsGJPQMA-B7L%cz)-h6Kcb*b(EAePQ@|po^C3y%ZV|9U1FbI+b
zK`v|G{r;XtK6&(7;c}VlwjILL)<RO;wN_}{bpse}8REsTqYyrt_Lp4ldpjkvqGIU*
zEbo`YmdhMgbjOw?H(WC3v>A(@xUnANzgtX!idcZItze`P#_U;rU2Om6l_ytK{LkK`
z<1IK=3e-yQ#mKA7G3R^pfi%d#4E_9VERQ{onS8JMyf8gFo<AL^KXHbqF0-(4#~y^Z
z>vLTq4xdviz%L{_*!kn2v9YnM?a$h?-klgPsMBgt;IC|Z<9}3`qW*yo5;pvRnDrp1
ztpG_J{KDCYhnRvgAINzuvPi7L=|Iv3PGJ2Lex|Gs%?jn<hJ&idF|0?}AIti`=&_A|
zK$CfNQk<K)6F-Iic_u)vJBb4sIa3%PRCAa<moIr`AlJWrMa|SumqB1>#@4N=nRGQ`
zVcO{P8^-YL{)LOT6kjAAIZjtMx9wf-qD2aWjL_n4-T!p{WXo{74~vJGh$5vE8~ZPh
zz+2bShf((Ne7jaxkZ4mMAWdt9tJBTWHGC!_YgsDnKTlnN<a&0<LS+^vt&B%I6y`&}
ziXysp?6?LJk!pl}wFvo723g5SU`;;&l?jvSAI4)*r4som@3umS1mIX1g`2SJh6uxW
zSd_CA#b#SuNOASjrDK!g-S^Z!JDXIvY#KpX*tau0>}l&|LwXo;1#%<1C)K)&z*Xgi
zU&qVO%q*=uuE!PVRSAseQhzA>7)Tq0P_!06;VPtdEpWoT5aLzlsm9=k8@WJ<)c)1u
zj2E|<7SPgD?}c+&8z`;6CjXl2viYkvgm31n;S|g!3yzqi+Cpa|ujaBF>=O*ir{^mL
zw8qMtlsU06HSv*2e+k<n5@+dO8$gt=01F_LVm-P&YE!y^n?zPp5|5Y};ML!3p+O<v
ztcdy8GM+h6XRVqh<M#cUdj!4Am78k$DP5|WhE$_xcC3!E_V=%#R(^av0sDl57TwfR
z{2|70?f2{x8llC_cz@R-;%y5J$39QO<cGF0WCKjobTzEHE->f-%D;>o_a(5=;Emaz
zHWGAf2U+WKx@3-3YgI@21)DdP;Hzhe4H?IIX<+JOw<x*d%DYWh?hgRjTO?a2a0f?&
zD>-%Q)SNgPT!CW4pQCX9&B9H4GQv0_A|mMVy@D$3^%BoMG(JScsSs|!uv6x@nH{{x
zFW*;=S{@EHy?8N<K!}{Zf~XGav53P*@e-axG+-wjG{8N#>^z17kVL}&0x@SJmD%h@
zJ*`9eDLDI4ap04&R+jRfloN-$ISw#_aP{OgB#nfde(_|k_51#;i)mIoSKoe&i_0Q{
z+PQG|1>{eXTm|JR8|T^U{<I9kB;~djNzQ;cKuR*cvT)U#5vrM7wa=-U2`!yfX4v-5
zDsWu=c7N-sJp1LhJbG`U`X<*38ssv^$EyX#S7!>1esV@%x~#WV;)L}F$6OHba*<)r
z{{}iL@q~s-?7hl+ua-8yh*wHS-7#pt^5V_L()71I5*ce1v$SXYSGtH#gd)BOrW%kj
z$?v^YI(Hrid)K|sR7mH37L}^Dmzu_AV4@s%F%F3g59+cbE!CVYCc2S?W1FyPR<Pj*
zPy&^BB1=rELss1LBIes$hZEJ7^#}u}?y`QB^)I`5-RIMCW9(~a;f8~Wc7*bJGE5Z{
z;06V7OkHA>wGtUbxMIVIThxXel`xEb43U#kLuHN%8AQ05nJf{_8$$(!aXj|sk?5Z8
z8#izIs{U3e5xGZVZ}+o1GxWwNBucn~Q?Nyd;mvFSsWo(NBdLq`es{^_2hwpEK4N~M
z*sFjEX<4&hpAs9}1U?=?m5uH@D9tbGtXE^AZc|{I<M_%ylO7Yl?~E+p<#j%BZBJS%
zX9C$m=gAhq__L))e7--@311?f@oGWJtv^Gu+>usjorWSPH_;EJ%g4q}Va^BCpt&*T
zEgS@)4|?n>i6?0O@G@zYuzPuX<AgYySVpDJKomvDPmzLA$;5BQR8&$<NVtjsAFd|B
zUh20iy}p$S3hwKf#9pa?X2r&$pz6VO%snr)0yfKt!>TGpmtZf#F}3;r8WvG(XmUXj
zQ*3XsDrRE`_DN~~PGt}W*^gsADnK5?vI6@);d0PTb4QAqI&1aj2Zmn@zJTE3oRE5F
zAg#SR!9SsMIju0*ked=%guEYMxYw$Z%y;LBeM<RBncfBdnrwX(b?2m<tgI7`=-PRQ
zqp|7q(sio#aDdwXK&KnSd@j@ol-Lx0ZK)!lZSpb&pXM?-Pn_vCVKVq$@85GVdzKsp
zJKHM?Fco$GB^B-jh27=fTjlRj$N3`L7Rn#$YqLnI>1@3wszC;bb<XcSq~CY-0lYPq
zoxv*+#<r8u-`(Crd@E6hOPt_>%J&;t-LZN@5ATh4zn|c)v*+Q)>!uzQV=+{b?Ra1E
z!0HoxJP(n_XBj4hG+0!1NGY(i;@8@IcAh-c5Km%DTByiuF9ywmJiU5@vRzP4FD$OP
zZC}@I?I`YGfbvi6f#wh@C8|WiWd1k+but4wk$rIOYFxIFN(|nN)yS%?-+Au4Ey87r
zv_WEGx&aGonBn()+ly><(%Foq`6r=ki5F>uT3Cj~)WC5pp=H<XWgQkeFyEWUa(1LC
zR(I;#z1}yaj_dP1EyB0;f~{cYUXauRB{A>z!@ByFM&OnE7kQGKW7j5&z%Nl6j@SAK
z&X_d#m8P@h0qAeekuoE0f?-VU!~5Iz*Qd8Na!%wasu-mj`d>VT8K9)3T<mt$lrEMI
zn7|&}43*&{)E)}DrNZ35)f6kSr#%+Q-FW}}=H57S4Rqga)Lt*;L6|HHwH8h7lhhI2
zb$+B?qy%qleGGiUN=Hd<$O>5Ti__d&^3ajI4IOmhtA8MZd3BcMx@C5_iPGpj?$f@m
zuQR!vHCf!aaBxR6S&Ckjs+PoEh(nd$X<5;v_(bZT5#F(egO*1J8MH-Xulo*d+D!g|
z1*`=8k)Rlho(f9^q78)A7Y~gt!EGV1*jHWN9aLtuMc_ZZ;dt2p4ICPF1%jM4SvCuN
zPs3GJ8i1U=uz>c>;(J>TQ<cAuUK|R<-=3EcxaId=K(y(-YOhI39#SeWkV=$z3;>~A
z8Fj<cAX*o>W8<KYL5Jlh1F3m309)sM=gr5s7#Mt@)WmF4K4l)Q+Y3$0lTT!_6jHUv
z)Us+)$rGg+Wk?!TFY3*4tkrR3z}lRr=fbY(Ma05S%(*c`$3oqMQg3tosK8tnj|8E>
zi$4oxYU!OqU;U=gO|`1w1O@w(${}yno==dcM7-f*wojRwd8<gj-0dYJDoXI}H$Xy#
z(Vp_mCg#(SeV-koWO4gi18(xoDMfp2n6p!}n<T~i*vHc2Y7aO|L3*1^KWnq@j9@>;
zOyw|}x1|Pc<PWB8%O*V+CN1IZoBd7l7LaUbS;3m;C~mt&(boEv@g-q5%SW(m?b<{P
zpR2lqV`G1TVQl@f+BWD<25fxrz1)47zS-8ccRMzSQfFbec{G#JH9bPU=egV|?HTIK
zrP~}VpMTKRxf%QzRk5~9mHz{c)X`Hr<EP!+e4D=QU{Sa^iUaVCxM`j^F6w@~Pq%zU
z+)M+UgqLdOdUfRzF3W@*Bi3oeq`{~K2V_awT71_Iur|@)*T28TBRX_tLXcP0e_AF3
zNyufQ`)5%}4VfCy&^n%T@STUb|KFd#BcY*bG0%r)tyncdJud0Cx_p|Jn<F>hExV~<
z*_OS<Q8U0rYvM*Ysd55jB%#WA=38vYlpHP90&!fbrwE;#JcdyW`>8nV)HOX0yjrUZ
zp|>4lSSIE3pNc)GVpK?b8?7;DjORDU7@hxk+QG7S0XofcfTb1I`xW(Uw4%J7Ze(SC
z2vyK7wVfKN-5KT=Ce-u*THL@<B{Xz=P%cS<w#;<Ev-WX;*G#%heCO`=`Wz5-5S?`3
z8_b+i$mIwQ5iZI%bS$?RttfmxCP+`dgLMwW_x`#RxmXJ3Vw@UZ0K{%TpEHFAY4(o7
zcPk*Y!|Ba!e!jjJN;q09gPLDY&A|pSV~pFa-Z3@wZdRlp*?NfP+oLj$miFny=|xM;
z%!2H+hlPsTa2&0B5V6)?Zk!k&N7^~Bp0@5+_SK>C)QiL&{$!$l<Df~IH3_tDd}Zko
zs}PT5mVzKR1t^0bwskV-rNjwCI$vA)dx$6cLe=BHlpW#5IPUNc2Pw{IBg31dB<C0o
zl>HnM#yyb>3gJVP-IC1|YSf1VW-sy(WzP=hu+RZzzak;@*L2&_5B*EuEocyBSN5a3
zh%w1gwv+xAk!nbCs{B*eDn!;=C1_5gc7Y%sD4Zr{Q_EKue0WSwZ)w;LO=3d((TGy`
zI6(1CToTW%{fOtse4DVe1F|EFy2_*+e1`}S@utGpxupExzP%_Unb{(Z61DYLnqSVU
z_2u!*-ze5=z)wfegd``~KADf}vRu4PbGq5~Wbe&k-=8Y;Bl>bXz)`O#_;l8q0VX|V
z!$Y$bGX-I3z`0Ks^K5+XpK#DNGh+nQwYYJ~GOCXZ+GfSnS~syb;d0k(qNjG3?G1ry
zNAsFG%9_@e-*~R@)@cpQl_)S>UY?$BzKcpbzy<~iDEP=IJ-dA7lM%}wt*1()+L(L%
zfRfRD(z_z>WchI3-!JyTm6{N4`$Cn$%e-74hSpyj%yM5h`Z<!T;t=6fG0&yr@{Q~K
z9^Xy%!i?p`u^EDWyxM(e*<U_$T;Wg?c^|>GHOT5+MUMaDDNxzZH%^2cTtXf)_Ol@I
z=G--8KKZBGzwmK9Pa);B%fPbw>(2ghg=k~|9sManeGYcmCEs`cGb`?yng)9n7otsj
zu3=tqoAuq+eTiPVb5mL!GVIwJU32IMUDT&44_QGxS->Y)J&tSn<q?Sos+6Ur5U`lN
zO5&NC_okP}YUuRD#dJ;T#6DHYqux^74NzSSFIf!3wa@qoIl4mDqj%TNgv|F1e0=gy
zJ^v~8RBKe1u~g=^gh|;zjh@{Z+XmaUIN>5xi+_Q*1GC0wM%dq>QoCO-)J8|w-O^N;
zIpG`I@wWzbzHu>>+LrhOV)@<OI4;H1i`w3aT);!GnS+s5xNC`@wXT0I09&iX#^;rX
zhgJ!1i|ZG$SXWxj6pRuVbM%<$JL)<wo?bLG!#4EsOqet8sn_jK7Uf)w6M1{-YT0k$
zxeDvaBo&7W*KN)si_wv(R5GY^5OHVEG=++sb9;L@jCkz$f9~WUY`l&|5PFkt<u5Ei
z9LHGok#1sx+F7#0C<oY93G9E?3ha=%*q^eeCJ*r*|9Xz21>K9&WBMGTO0S$trq)v=
zXByWu0f2N-sC^SYm1^*#(;#uG<^G<f)`ByJf?~);Vq!X*e&w!y*}UZ(gMIaecT8lx
zfNn<G$|$R;&S+n1D&JV_^6G+gYsZ}pRtUN~!70Z683B$d-Kzle%!!V=C*4=!%7ZL)
zV(<==bX^*`uN(Xvs?upAI`UQt9i+t1L*&<HFe5{!l=8asPL~i@vsF@#K{S#Al;as<
za+hf9baL69_V-*7&|iC7|BC#eR=-v?+YD?*W=Ga!n*hIjYTOg6U-_!(W^CG$Ra|1_
zW&+aM5d$8C%&^rB$$E;*@ak@kIl}pOXDQu^+%gGXo7(tg2oU;OG%m+zO_L9FT43v<
zR`I%;!1sY*^}J=qFEV`Zfi&_0_p39fPn%SnVLgt-<?#4V-R5A-3IPOC7`{1s=*@)S
zj&T0hJ8DAiNFICgf0^6F(7_#Hrn0Ydx0o<qB|||FFiUfu!e<79auU`^EraenEJG>n
z&eP#H87gMg2Q)cDQ5ZTGVvWWL=?>rxOoeO*yHXy8YV>Hwv%`X>m{6iUJ&aG*0eRQ`
z`I$yy!IufTKlvvzKN&Oz=;mhxm(VZO#oj6EY=iPeR6Hp}pxsQQzCOuNSxJV9Ee?xT
zP$?{R5^Q^aHyuD+>3G@lsW^bImj!0{@Y~p0?>0;8S4`RyUM_EQSxw5uECtL|e19h|
z`0EUn3$@E@gO&&>ODgdb2p}F5$&&oK3T>{lta?2Emo4Xd)`q`<M;L4|&ntB6aBj*a
zc}6p(zU>E4{R-}h4La-q(#pwK{cNU1k$|Qb^um#YcpMhgm`7XL4Ebq~Md^u(%`~lW
zRj!Y^$8Tt7s2AP*^e$KUib@OR#NmYrREZ-|$x&_va{>V2jN?*1iPWTYknS^S_Ix^m
zO?Wo;FPs50f-}&%`I7hOM;%DOfBn0v9I57)yYvXLs2AP<V5D4|Y!mAcT^OCR1qzC$
zUKsB)_>|AZx$fMNpKZ@rw~Abjp_3rFsq%p`oLMVc)n+>2jpMJ$4!RSL;zD}C3*gXg
z#;Q;09{&TR@YQ($S7#XSxkW}1?A%5YcR$vzyI=2R4)mUUTf<Kel{>2t+_|i;6i55L
zNm=~IFih=OidfN-;`%}~V2$&{dwFMP)0+Ig6Fcnt>|ppDMU&NCF$`w;cB8x@bX`w=
zML|OA?WLaitz3Ng8N!7Eodztu22*H2)n3pS5O^{E>3QU`*(54Yi6mnFnt45$^ywS6
zaRBCj1mTPI7-P{H`{%3^sb_r8(Dl3k)^<6>1;%MXe3K6QvQ05ot>`N7&RxJhuPHA#
zn9J*!#N&l5NyEK!cXS;X;}c=A{P6QL`n1)n5Qe`&JDQp0WL>b@gFU4w5QTOj169!5
zRpvN<!~`Z#^$qTC5XIeiP(o@&Jeo_|o3lv3vWEmL{K~^3>v&L5t^OC%FBYjM@hF`@
z9fq_2e15TsT&7E!6HS^#WWWU?qBTjWD56W3M1i1;?P5)6XzHYF@|_4pO-x2(lB_o#
z8M{TO{tUoeOylS|zM743q4d~VCK%bR&-^GZz88s!ROf`*ALVHJNRNKRNCt;&6+Z7}
zQ#OmI;518s<crHoVephH!x`47E=Dy)$!S0%s501RqEk=y5Yj$s|MAs><27hB(~7l6
ze2ak&qc^hD*ddNfduYWDmVAT=7QMUH%mqG8If%7K6LTTXpLe|o2d&<aJr?5h_B;N9
zfjj+kKFpiU*Qfmc3=#9Ovv=#4D06Xfz0hnu@`Oq)DEU#I9V?=W_znF#>UjVq+G^n6
zOSFy7T7hg^`v+Kb!Pj#D2g_DRTN3cogT&C;yku7WNWEf>PkJzU0*S$YLv++bD!^HP
zzTmKuu?H1DHtARB0}3RDe<_$WqF}LEFE#&5ss65GRW1nTiBI7^;Z@%Kz-Yo1|71~Y
zv;lUM)BJlU7%CcoW~e8o0-!6jn0L>nQ?KDJgojR9byFxT@fKKHszg3cE%EIUmz126
zt{IArg1!c-_~AYKok?j_Q>5B@<glk<>NYG2p|#%3fwe)6nWq^6DxNQ#@39VY>yv8K
zP-3~Rj3I|mQ(9>bnV=B+<t#eC-IlX-vSD$EP+=GoE7D}E_nk<_7{9Xb31H<FQC9k?
ziPbamqf`d>S(NA!p9IN`MMWY&$YmPjK8;oT9gQH8%m9T!ta5WK&lk+ScLyR2QXGeL
z>$_?{T|o|46{(jVQeeP@5Do-$74l+>Fm$WoeD<GoQjO-wSq|TP<?_FrB_(o}_#$jc
zuwpNaS-Z^W#zp(9JAkmqt6QFYDr4`!DbdL85iCc@boJ^>8d_Ro@UPw__pHyf0EcAz
zbldwm2-KxD6fWHVNkeAcY1af!Rc+C%RKbCy)V{5x5+-e&tA`E&W}cYVfL<v3F)%W|
z#lpghaf>glI#0nVIa|2kAS@)*p76lqhEz0r7K!(A4i1s0;TcK5osnt>{b-Vy=ObwO
zO66rZy$tx%2M%$Z&&Xj}u8h4SGWQ5BISF9JasU;vZTNaL2urYxP{PI0_}e0%yy^;K
zA(HVM4?3=82ujo<kNTGk($E{Ifv5jI<!{%5xMgL+$rFeT`S)WQ$xW^AMH3?VMQtpV
zzvhr}Adjt63y>F`a0Q5q^pt$p@D>~o;+cK@{d+X$5evt%WZs@|91tBozoy$H7^UrI
ze?;SY&QzWvWCs|#PXQ%qwx=N+Ivl{}ck7mueD4{_xU~@kG(^fo_9)vjWYFU4*RM8U
zj>AUbzV5+fhA-SnGB-bTD`ky$-PpD+;C;$D5X1euzNrJS1gaAZQjs(VdtB-HhC@e9
zJh73z9D{eAfkBMmH*6%iu|w%J-_0OwpnGCtM%C>lg7dOzYq)k4>;L~cOQeeVQ<Zq~
z5Tg38<J>|Z7;bi$cc>>dr(skwDzP02D*!r9?mO1hr0#XQ2qjj>k$oa}fOoQfj<}<^
zfUQLj@GcQj$*{ag5kN&_y|7$|SyTsQJ(C6>wX`*~`BZg*WA!RBVg{07KzX@`wKk4{
z@oRp!GQcBxENg+rW!PX?XLdQKf*q3gG|p!O?MNZzx=-vE2vd<_TTH#}&w+4*Z!uzs
zw<*8jX=s}PPU{hfXd?WDJW@eUnO-AHC+%I(8H0%>a`J|X!)AKP%6ZySOd6jR@y3^`
z3uqHaS+of%74>1+-S1)h(GEmopjz#V7kb9i+gskj;U2V9GXFDcO#Q<c;*$GPlnV!g
zOhq8INBnRvUgRMRIk&=V`8>bqQUMgps}@27pZ$sa%XtlfR0<&dO6jom@$`7*#+Qjx
zWBmb252QhK?T|SY>7L~{O=wy>n|Mhc+TefM1A%i$Gx%Xsf!>Jo(*1z~!@v*15Uw+v
zy*Xk7Xey~$#u)CU#zxV^f)67tDqe;p!N^n&Y||tVt{K*VK$EF4M}4mpNDo#*uhV^P
z^7#Sfm}Izs5}QWFE|kWSV`Maa&$H-wR<NnJ3x|M;-6x(aAd;+lb#(@945J_Hmg7R{
zeSkkD`Q`@uN6t+#-Bt6A4B1%v?5TJ%f;)K!Hir7K$ZJ2R=8#6CexDlK9IJWV4ZD*d
z_4AoCyvLQ$cfUIR=MjQUUQ35Kj+9YV05VY(%jh<$B;{E^TDKwi9avC*9w5udkhPw2
zFQy|+82=yI-UFWM{*51Q7n+1hQMZ*HvRA1H8Kvw|C@W-dQfW{WBI_L0ZDwR=OGXkW
zdylM>z2`BW>m%K_PWSU4&;R+qUccA<J2~U~{a$;#uj`|m-!^XlbRl~4fpzM1xdZ8Y
zPOl|k|GRbj;Si{7wWjlYCR<Mrh-jULGG?<}QN9on$;k?T^;6YW+js0B<<jSGNxxPR
za_eb5TMKK#$h#t5d(x4%EM6#uej+&wt#Bhap0ExeP2H&xGhV*F(a`oGll1S2WN;ch
zaPOs6E%I<#F4M&Ue)FG-iXOX1%J@r>co<{l+IK?AO<wilg();biyW<Up^O#zD)fM+
zaY4sP*sM`3ETH-?I+#3zqv2*IjAk69prE&LZ?sytH@-AIpo~@tjJMh3iqCg#J$tVz
zkRz0e;f?;=w{PJHavcRMC6`VZ9R6|P=9d7=;2!U_=<Or}!vi>baA-2Tq(BugNT0%l
z?y3nmueb(UK;DR8OM%)ui&7VP#`B=aa{xh&lvLdu&kUk?b9w=fRfdRjvGCB(brhSn
z(4yG<{`=&$-2+W`!Hm-!+|07ZrF}>}wq7WvLeWKPI`igVPV(2DsQ*>A2!)h{!*j$2
zV$|<w!z2TrSxcS2aY#x7REmrW?{1dFtCX?X+FwcVUWoe#k3i=0@6c01%!3K2DKm^l
zIIk=@MlbBgtWpU-Oa1V}q1^Jt;L)KcdrM-yaVa;Tb^X96>=vZG+OH~~`$f(>N=8hn
zK+iZux4_Ny%iO(Jf5*=W4Uxf$(0%M8`)Ea*20ETo_@#?#VgH?56_Ixa=)>cMJ{-1_
zF^r28=c0&$mJ42H)o-?2;thF=t#+`?@a$iCDpERC!PMKUxeKSEFy~&z#K=g(b?rxc
zPNtk|!hC*h{?@NQD?GpUZ8eEN(Ov@8@qDk9VUS)x<h37DY3TKs3OZT056rLXQvE^X
zFeLlGLUAI&hkphjYtKXfPWn!?VW8R^0O>+yAqR)nanX>l2(yK3T*WQ0mBuY;#&8T8
zlXm3zweOD5T{*2+yo;L2MZu&l=CwDaqTiqo$#?afwULIi;)dl_Py;Z&8v7VPLx01L
zi^~uS$KGyEk=$A9zA}ZKhSS=jZKmXXPB1bJjSPn#Rh8vDeGZjjGQOFK+EPz}T{|mo
zkMBz;;%#(<lf5OSua&W>8};Lqy$NlUIu?(#bq`=74cY)Ox*vO!zA;iR2)1Rs+6%pr
zgHofk6MWDR)Z*EnI%eoLno?LLArP5gCpxxt*5TIKU>wYX<A1;~vt*<FN++$DlM>7>
z+8(oKF^A0y`-@kO3R(T|@|IZ-GGM6;fJ5XDO{~!y+C8GJ?`S{oSSvE1*0@E&(Mh|t
zsO21txDZ7xRmjt79qavxOZvmpet}#@J@|#T)ceOrXlUdh3ypyC?=Is!II;F>f(D-n
zz`9iJY!xb1;wz#qtCu1xe*_?6JrABoHbO%~2}bH{j%l*ocyb@$-<+bn0xU0TjF`=1
z6O-AK%$rj@5pvjShuZUg^SCyl=85jaLBJYhSM9dn+h0{3s~>`(t2sXHUbc)0dVZPE
zZ!+Y}3O$7lt(!ChK&~+im<B*hO=%GJXx?j_opDsatR(;(rZf<+;p;=@_bwH4)ZTx`
zatQ^uG(Iq78OVWS){1a^khj*9sG5+I<t|xQgA><&cv1`zKi_>-e4@etbh)os+9G<c
z_qd)!(4pnSRdx40uCLSkI2yM1`TsU-Ke`rMvprb)Q_Rs{&V<3rnjV9uXLr@?Z@eW>
zk|4pgeYfu^eU_c$V$quM492BK{z{q@@)(a-bjY_K!yKou-l<*YZJG{HkT~s}jKo%@
zI2HEOTx##{BEU4cDl2v1TcIF?Tv6t1^9U~KuXFA>erp#_uO_*gH_`5X2fL_TeK2l!
zMHr}*b<zF<2Ts_yD)=)!(-39pak=0SI7BMyIB9T1)Zx_jZQG3cKNXn)@I@<IgzZUw
zK9eY%=lc+HxxS(#_tlkRm;3kbolbojdVEH57lwZBX)kUW#!#ndX=xR}5ZV<P_2M~3
zyDKM{6vBQS^mmWCUM>ppLR7<{kL%<pWJlKGuP1_t)Lq9wl%1u+pEIHqrU?b$IjXxB
zVj;0n=VQ&>{a}2UDt;IA7`o2E2MJ2t94(~K9@<7*u@2e$HKvJdJ^b(PS(LcIFj%NU
z7n&%(vu0|WV6OP}5k8Y28aNQ%Wb>xoM$pNU#7iYic5@Xh{q<GEW&Zt1!(F(=2Zj$>
zm^>%PRAaAP*#B{50q1Ta+<Jwaap&bh9oCl>WpwVBpl<Qk=>CL$$P^i<ZaxBFi*VcS
z@9ON0y$U5T$*SrfHOdVXpNjhB0SS2LS3{+%p)Q9hKbM#wjnc@teHeDXTzpZ$T%r?c
z`0A?a^$uwRgTIe|F>!=##g5ARSF`9H2Th9~A3%e#FeG^JX0HtAQp$a(P@Zg%9lN`s
z9L87-ZQWQHg>*VsQRP2qgu@BC$((~zLC)g|Fh@!g)J<{Z2<zu>8jLov*mSjfd&$UZ
zXNopu9^Jj|=K@za%oqhV7uV?bzZ=x#o<JDzhpZ>0iR%CB|D30T__5UZJ^kj<OD(Vv
zr6t!fRoMxqz77;~iv=ZtvZs*CrJ?4D<<84|s#mV~9T(qonCF2)nBWN^Pfx}bM#d%R
z)XVz4*79-E9}LI)XbCc5O~ngff{jz|j&vnpm=4IjF){c8$inKoo3}pBnTg6*`)p9n
zr<$ncr}?6y56+GwKQEBAaP-xJm1+Z=u6W7n7yz^EkKa5I;4Y*C#U!fpL%R?2#0Yg{
zmpSB&vonpU#H(HIvIL8Nkn`V>fneKYzrmtB^LF7lt*jKiONw53vbev*t>Ji_Xh5`P
z_lAo(+nsf~#jPfL#V!oaLA<co>#uk|2Rp^<1UBp{jJC;$V|u&IzX#(qL5kM8$b?^m
zRchJxWsLO$VU2!6>?2&TE!6%abgCaD_M1fFA>XC&u5n2`#+8`RV(1}KwMp+6V)Y0p
zQ@v8rAw=%+WAG2?={Aw9)dSE-%**NIj9B1qsF$+!DYRPa2AnK#7mE-lhTH9H?O={i
z{`!HnjembytG4qn^WVOW>c*Y}4l7GSmI8$^W+7a1__Wb}*NGD90C88R$_dwN4*9l2
zRLK$9kf%{Iu~!3<>B5ie!Te?#mv6BPV}-1x4wcme*F-YBZ;|22Q06Sx>dWS_E~`es
zcXFNTHy19L&H`}lEK?}A{*H-Y1y@Vz7U<+%23z$~R)#$`<-=h*x-p*Fgv<`|;#ola
zqLe0F9M5aLG;euUvLnsde^^)+e}SbOhTO=Gm8pVVKVA2^VqZXm%~D&dr(vw~alWmQ
zK3%e3L3_>v+AbR4f5LnEI`tj%p-Aip=Z<&DB&*6e^t-6F^`=%FyIYdj)ph->0jFwW
zSRx@{Yz7<+YWo{hJNbEwcp3(Nh7gVQ^ZQsXMVl4MA)6}4ak8`X+Yf@nu_~|7g;7a>
zb-ji3oO6!Lz2SqUKAMc3D6seUTf&OF%EGyZQJ3FJRl|9Py+wW^&QlfKSKNb6im8qZ
zt=TXrlKFuhRIDl0Kqm6K7!)iT8aEwY(@!oG&XQd#rrid-P7vD#|M_hhZZ8M&KABfZ
z9skon&WnIL&dyO>A6rf>Y>ijvjb_NA=mj@&Pz2{S@)fLO#}HH`(Dtu4Ddg@|;pd?M
zd67Mt$^K5B-N=!Wk`j_PX|Dt;Qf#2^Zw#e<1wnQ3a45S#bT2c#sd<gVKJU>a#N$f6
zH<pth$QfmKt>$;Z>FxX7yf0h{l=VF-^q0C};7Rpuw+<`I2N?zK*Voen*oBTCw|YeX
z)F=Fj<Hzqpo#8K=Ez}~sdOM3^#@X#Utuntsm*I+4w4S|qd9Jp@Eq?GSufd1Q=yfkn
z3Y~;881n$s#O~EDORT=~vG{>L=%>j8!>VOKD;K|sQlqvlBLramAx5Jb-IqScQc<w9
zTWOGdM1~XkvRUuIe1scA2&Pb0q#W*EA{@@lTom5DaFaC_8@$3G%FM(>%H@A!L2Iyk
zw10{BJn;H)fvI0XkPodG>%}UVu%Qh9f(x?9UyyoU@eiOe9CgC2ZJ*O1#{L>GUt4g#
zgm#p=$7lPF!p_5V&p!Z-7N`T17kT9~#eXG>aDE5G%LUj?&STh{-yq+s1p{=2obNB;
z#S}QggIrD%0w><c81}CUA*Y`}d6L=2%r-aRZ)*{mQoAoDFHU%uvWuk}#j6}uVP&r{
zf<Y(AvRjN#YPhNFO>f=EhdjftD5R!kNl4BN*+0N_HyMAp3>CbRog5lz5imrglySSc
z?(bVhPq}vCmaMjlQ@43s8`f*~zgY&AmZ4iOpFcnT`D3i^^W#)X!<QVc3}0h|dC?`Y
zVwb;##Z0%Ca_xV`bJOd1M6`UUie)x))(2&Wj(pdg3VRX@Ob)5F{R9AgSyN8&bU|-G
zGN>Dxzc~-SkENk<*Q(2RhXw!fuY`W*SBEPxVy(MOqO+FrPQkIKzPVvVX@4KswMMc+
z3n2b%YmI&nXpx&N7(*--ZR^^WGJkl7|M1Bob%ar}9d^7)yzg)_3a#h}jq}v{&#DM8
z4|YlF*XNnyuG}NHe}D6a!cbpD5pU{sbz?Y$v4|^=+Mbe@_tw+nJy6D}hG*D*;+aS|
zldfz{Rycz`4Ed<|R|Ih>tEs&@-g-Fe!_cj(SL@(>8s={*4Sg$xbPvy;zjK_z;;@F_
z3!%{7+s3S~V^zb-$UlZ1;M-04@>}sP!%tm-_k?a#8Ta2d>c886ZOY=NmUSV?%R?9D
zlk90|s){@CVR>_e(L4!f4%HxQx3RIgC2^dyPkO`hWG|B>HE(m)gu2Wo!bBOJufKvY
zi$H8@I+o!e_R_LKQE;nyj@kXPRtkp|a9>sw(y)GRX@_&h=@;qgw$!QlR9kRQdW7&{
z6i{C^vFbCpVoKiMZX6x%^<8Z=q?naNh0DUqG38@o&gc;cNzDNL#9N@tct0<JM3ifx
zJ!I+T78xa>;I6g?+-S}iX6PeR=<O@hR4;8<RJ$A<Z}cz>yY#-0rW&cnc$OP1_HxM?
zB|ge*x6^bPbM7>K@0zzA<6psSui#eTbC@?ud(0_Cm8Y7(EU_RHpEuhwVjOBnvj6y*
zyJt$aZNsohjU3;yZI^*09E?O^rnGJImThr5XS%WmHGgc8gZFD)so_5}_QIRDG*j<)
zsB-g-i<MNyyo=Q|VLWqp6MTFZ`YrTg@N%II-f7rrradbw_W2`PCnIyc>*RH&mC+4w
zq~8UeggK9$w2v9S&HZ@mOI4Zr{UJ?1(-hN<63FX%u>`;!>O&cu8;s-*XAZ7AA(785
z15J!$56j^EjBzp%lT{*P?AeCpMSoj>-d1?l<bd;tGYr)Xd(iJsk{6`oWPI5_+CR3&
zFw9bFH?63ky_@y-uGwuEh?^?4j7IYbqZORhm)w700fJhgcG*VwP%TB1PRK90twiR5
z@5P{$S06T<coo|;E6UZHc_$m^Mm2RU^@mzuLRs@+YO*gg#^8Nsx(EL6&%NeJaQdOS
z<oMq*{u+FH#7Km(E|cZSpZ1Cz<g0g7Z<NgrU>jd5BI7ECi_^fY=au`aOu;M8?Y502
zaHv#kU}Hi(?rWG3FLk21xyBp`bJ%mT1t$0EGCxJJ9hjipA@}7WUF*%Oz4m6N=T=4>
zJ2l&{?pvd_2RZ33(V+dcDnR|1Et@^?tVVw1z*ME}Teje?ePdvxVm>~u6{eh}LLwR@
z*O0$$SDn}0LhA$@9+A0IeL^C6N+cUOhgf)BeRFMi-oE<3zG-zO1MJGOR$icF2hHto
zai(?Wx*SO^lE0B3B@jAS#i^s@@7*fL2D$LJ+tgl+60*pfX~-&1)Lr<Z`UX6Nyr}>4
zPFP7}^}6MzMC<KTd4}&FY#Y?0%X!sF4;#h?N^+U%iNWR@ty43;?o0ZjGj+-nZPqxx
zT`F!<p`BlY$3FD<WVz=r^>GQbW=a}TF1mHk7E>q|14JhDT_2ngU-_uR?@0b)eNw+<
zL#tV)_Wipqv0HJK@&;nvkLkNpDmd9>3Ui*Hn_Si)5zRa}%U+f?%pmAqcFg`f51&00
zN92p9%5_y<<Z71Zr1dJ7U$@ZQf~&}9D_{1L`LR!IT4_|m|IC<QY(sgwA4yfYuGI4-
z7Y2!F)5NS>p-f54hAK6YS{`_a^7>c0pO#|#B}+?|E@c?X%cY;5>tq|`8_$imm~^Ih
zETeNCPHrpk?{MgO)`pw*Q=AOaxt%*v`eL5ZzcsVXs=_XBY-Au|T^F6VZ^N{*{cr?*
zx*7zmvYnAj2j+f`gV!n|ZxdC>c8O+&#Q(hy_#RMu3m6&!@9g2{>thRsQ@FPtRs3BU
zckMSiTWDjLB$d7^?cj&-Q#T~due{y2c}qIWcsVn+c-=OA(LF`t&s1wW2MlP*wK;52
zuhr$4klx@0jbc9PE7K?bQE??4NY+f7t3C6X2wJ$;q2!^_OLsSU)WCfOb=3K=kn9qT
z-nNR>-%eyLl99DYHZ@`T=cj(YFi{6<p)qQ1)-h7As{3r~KG9PGmUiXL!JMHFFskY|
zU$*3?X&#8N_Av1K@o^_zSm>La-qg7cXK9|-w(0dM@2tZcbR~k1=7aq#0o;7B`-hs^
z6iwZ9=6~$buB%dD$atz0;3>|ifGta{A+ZHXkI07mf0yGPrwi{3wKu|PzuNB66vv;t
zf+7vi#T;Qf-IsMFC+gM)+lGELB5?a9`^3J{hW_@D&}kDMW3$<ogk9_h=0_~tX$Zrq
zid*;lx3|u^u|sPK2KkQ}VjP)p83&xaep$(gQn;{b?#D!FAe<D?Q-=gy;-n{VlI7sY
z3lg}#O)H#ToLEpxvj1~F{CU>-Y64}*q*g+>vH+3BDw%GH5^2<s9>R-r<&H}Z!({bD
zjT~1WSNMZDm_0lU*=-!%E&)cRom-&e<J03K!s!O~1aFZuRV7l!*QiBNu4~s-E9_wq
zH+AFDF{-(Cd)<|fPysBlZPcb7O3X1I;!h3vb?<l2>;aS0cwJfh3H+(zz<!5WCuEAw
zumEA@UK{yd+64m8x&xkIIb7lHnOS($)Xuji3Eu+nCL>|lYaZvHAKT=iL2;@*t3g;p
zV$Rt7OKPFc@^q*CkX#qeeZ_hhj@f&$%{Iv8lD#=W<Xc<XZTB1JTgAj=eo45JGte%*
zQCZ-xGj|Io;re-W$O|cv0{yO>Y@w%sw+Awa*&2e_8j=%jN6c2LZNOGf&W5>TI&KSn
zO|aOtCkv0MIQR6J2>-*q$Bw}4evHM@OlBqeEzoZ9tyK`=;LPgvTY8p1mp4|z<-Slc
z--Z`grO%9S<Acqz;fHt(E`oY<ES&RaY0m6$GQFME7x<E`SO(qXPvPfb2QGi_KluzO
za&!NF2hOc=9LRh>f;+cun%2lmDYVWGr*u$g>>L2+tP65V5R2<4m3K_+AB_$FmyT;P
z5wjgkx9Mx|8P`V;{NW{^Gl$z;q16jD%Xk7|RhjTmymF_YxnhLWHaMNPpm_7nqx+FF
z8kK_njxQXGT7q#NRi_t#+XoZ{B@ApmmgCAx`}kUBlT+IV%W9a7`9j&l@Gsd8Q$Zkb
z$2vd?#Fz)z9<;eW5S}gO9ME6*wo?Mk&I+|xk%~qSjoED3mcj+kx=ee`*)vL1)p=(<
zv7x&9?(RTUeT$H-0F}e?Ok>ZtbNikF-s#k8f;VUONd>|0+mtbRoZI5Gu0txtZPoi(
za4dIfQb1_<{U<Arb_gD>E{uFpUb(lOtlLLq^f5-vSqk>*yai7*Q7uhZLhma|4(IYZ
ziZ1l}@7b#mB?9Y)XOfp^TZ5vH=a8@S18!atZl3jah##=w+g^BW*X0<1=sbz^^kE1H
zyh7$nkgG+J@J$}WVD$_Or0aEFv*7AG&dyN$^E*FZ)U#gCwobQGXs;w`vuicW`ex0O
zcY5K~<gnNYC7LaYVl!8F9~C?mrHIAkD8ULYedt0bmr!Zem2R)N<83gx4CdLN5d8}G
z{>CRgXzx`+LZ%m}dG412oQdEBsULb%$q72)kqbw*2YH}-l|2DopY($BS~S<r4+^=F
z1ND-?fksK{_O;LcBa86tOdvlfqT`Ud4mCf^7USrWn?c%vx|mB3<ntR~j!G}%{liTj
zPf5OFB_yPt9y$O}i%M1gTVG!Z0#%o_sDIQWhp*r%4s%i7=%8K5;Ym4WT>q6BTSmAw
z)e$>&_|pTp-?c~iKfYg;u9vDh<?@bh-A|aUe5XfR1riHZ7MqOTKJ#iX65g_nhhe;M
z>%QU>kB*W<)_E&24uS}gQeQec)A(emhjCwTv7qp~%G?nP=iN#qW)H?Sxe-QE9Z9Yo
z_=lqxFN5}7aFGTcAKwmwaq!zQMD-&F_d6thjA|Hq3GwkrF6iru0`r;U1Q3_zvC9cp
zjlMv*RRhI`e|`wP@Q5dSH5xXvH#SLZ3pg_iBnF^#FI#Uly}!ZjH2vd8O{AQ2#ti%B
zEtFKRU}_R-ZVcwx15s_|P=oV<S(aid&D)p1%f?N80KSxFw3@f-O@QfletLwDmA~1*
ze=s*9?q6dYpbjqZYhea6Sn5)c8QzBdeaMx`A{Uru`Nf<Slyg+`1d-2SLq3Ns+-Bdu
z3|uo_$A`e@80MWDH1#cn&2#!t{o{6-A8+FaOM>%cuy4K@G<$mpc+7h8y~9NJD+oJK
zD86dF<gL5(*Iv=_LfFl5ZaZ6S$6FAd1C84D+HGKOCz|wqO9*2Hbmiyqy{i&^5%>3g
zk++?tK6}-|{hGorsaJ9X>QB;$An{cv2NQHfpiUA<I!Fne=;(;eeE|+LJLv%tD>$wJ
zD>!UxwTW55P716jNsK`b?O)T3K6P9UxTrq(Py_z4@v1Y&L5TNhC;E$kaceaq@^u<<
z;OjI7qjG*{Io#~QQ|&_D+LtMauXFcIAn+u$ZcjSm$r4<&Prz;qd}{D4u#Glbi`o7l
zn4M}Ovy-j@W+%Pj^4!1S*-y8U@)wvL72(%=iJk|D#EvXhmWT;G2*Lm`U(Uir8a`F`
z7HB4BmnSvGJ|1RH9K4w%K=5Xo@OJcH%kz^N1{WZBi?@+H_B-9so9f6pSlQPh=kO~y
zM9x7Pm<P$>vkv4ON{O69y&f`6daCWjytt?g7dkx?F8@miem*GTi#(94zK$auY<R>n
zydDuq+lmN$ewU~-naNscr}BUCGH{ihCKpivd5iceCg@>I4BNel1S=*DJl#twd)hzc
zfDhQCYCTs6!tT)kuQNNoQXtxXyIKg1%L|oj!Idnrz@gwuV<g`BUt{&t*u12I;Hsv(
z3YblN3%7s`CZm`Z<mQ~<bwel=LE(NvAi8sA`+p&u6?z~m`*swrY~*;%1N1;jb~}Fm
zGu*@(4#)=@uL!hz!dE;&-pFXw&!U@Ie|?z|I5xm9OB8wsq8KeOXUKfxRnX5dq8RN#
z3NO;m-=E1u22>?5dw>(ZauUKSt)|bU=qA@+d$F#0?#sj~ggke8$Bj8+5`AF=70NOg
zk^B8BNZ=~qmebp`?U4#8Z}u>>*p41M04gV02fb!LKS0*U48nxpVzwC+v(4<yp(7%Z
z5-0fF3Pw}i-w28x&O-w<2-1vf8u-djwxFkuI0%=IR7#U(0+Uf|cAv<~i6Mt0X48D^
zU)#S%Dvv8*<+Nnw|Bw;TeSf6;wyoOmmHj8ihdI`+egy?|lGU`q5SA)r?jfSvj1tgo
z=G#qs1|l}dvH;y!_(m@LNjF8f@tHa^d#D#dq+-AJ0a<$B)Dh_%gEL5jn1~}4IV5m$
z(e3_30%w#20yjcAElNa9;D{J#O{4W(J49zr2Ty&^qRkntLSm2y4>=$|^2!hXUv_*|
zNQFq`bXY;;bc9DR#3JVm*`9e)1fGbXaOFYdR53YYqyN~?_Hg)=64Ixd&=zh+`2T>e
zRqJ?S9wAS(PVkN7^T^osYP5}t5BG;~@7k{h4tm%OsemieDAj`YK%hw<xIfOIcvm(~
zAfx5vj-A9NNEs1OIkWvGF#)}SfM%DccK<=ZO&*cJr;Z&Uk9DX3;X2c5h8~I&9>%J?
zC9@merDV3a6G8M)dFY|?Y};Ik=;WaX^y5_d0#Vo>O3MMfauz{4rHf>v-XarZ<4hI^
zHfThr7qW$6xgR}<We>T%nFh-_;=-dBC73%kmZ3kcvu#%zScKt`|84T=Bl77twg(fD
zPYna)>mZJbkVHp<nc;l)76lnf9Pm)}H7vH^nU1%@>*i74K@AAa+eqVi*eTbOy%-iC
zY>0_rrB-PGKEycEo0$2G<nUDIpJxy|dGMvq@YQVh?VI3QhVc8g4R;z50fP|%x0$wn
zK?Lk@X{I2m#SEr{J0D7)8YT9or@>z|i>%q_-<cI*13w>`nua2mI24l{$^Kgk7`mR3
zUb*oC-oGIE(pWuV8)hu{j$3>qVQNnRZRo7zQM;DyjQ%;dzWVA7$qN|0@9c~!pv~DM
zoM&KIHz@UtFf{?$+{$83_nUVXiS=wbs(X-pEAg5RvjXka&S^g)qP<Zn65Xh&<NqPL
zOmG`I$n5{mM0a9k3$8?yPlsi^%}zuy3Z<rQ`>uXVxyyp<AdZ?MIzWOG=L8Cui9{Q2
zK$Zg(&*U-zJnug&-pG0?$PPF!UTxS8mzyFLrM9Sdh%B^j9T5#iRh>mKrLz@VqJHYY
zZ<fG`8kpYExoHjpcB1D$nGWBKy)c4`N4J5)hB*>($Q7l7!B&-vu2WY6{v5xoPQ6Q%
z8pk%Sm3yrR@>R3UDT76z&cNK6?G@Cg`kHAVWd4nUbtwOyX>pt*uF!$;jD|}J*^Yp`
zcnkx343n5yq2_Xi4}t|%{l>bscz8dX>tUhwf7A87%Vx;kPy8-ZchBS@my{H4Z}Ep|
z@UTL4ZRN%e_{wJt{P?WXo4?>wS5cVCXD44-3GYV}^xzT4xc=)AJ-1(S<h*UL=9xDr
zl#(W@m8nRyKm7gK;UmXIfI7_sp0bF2r~)T@Qsh2-Djj4)iy)R5N&~jF)x-ney}TXc
zooSMLr7LU14s!|BWR_`TBHn}Ocs`3qzH=j4P2?bOZ5fPi-EbG{A2E=IudbKiSy9GL
za?=n#6@f}-rUhlT*5@G3)8v51D*(c(JX0kV&1|po9{`?5cW$LT$b9{awWWdzanM)K
z1s-^2?%SR}&0q=}$a#B5%u~3<K2&ac6eoskyE{DkOyg8YUNW#T0W4F8ZdaDWyHd98
zm{xOq`83>6?U)Ct7saW*EJ5aXL_h2?h=g^~`Ns|-2E$dFAt$Rcx44S304o$xoU2C;
zHbw#nyg1}o*`Wq*zX25Ezu8zq_G(fq_vK5XA3kG?Bq!WenQtSh*KX^~3{nHdq^q)U
z#a}TD`a0PEV=;%}vNoD5KvMVEEx2|S4er?%<ItuR8k$noh6TxKt#oOdffPxb!6t6q
zuSJUDPqL=^%UZ|W53Lebb)M0Txd~z%MkC*Nn`sR_wKptfp?>OHhIQr6Xw~~9IPpGD
zf0@gbJKIkE1We|IiI^_ZSd@kDTVY}9wAB5I;Y}(@^;_K6tdIXu@F{SjBwd`k9P8f_
z)Q)c;M3ok^*yHC;LY6oC%4*{CkJKvs6pmdZP0-D{Y}s=5X5&HIv5$m(k9*y3Cp6vk
zTv~LTqAq6iH-GZZ{KLB#u`2P=_ku>g;cs9ZJ8PyqJZfpBTF`a)L7%wph?qQQqv^By
zp66nAUS^q1ry54nr4p0EMQp#c6Z|iDTa6X3<}KJQO~m5$Cy&jQX^!RU1{ia&lRKbV
zDDK!L$3(174$JR^j!}iE^}jp_2ov&WvN9sbY&V09RY}oZ1&`uk0$szz*}ZWR$kvWl
zfVk(nJ8y-C-gF)PoVSI$wf~O}v&im!D~esPcVx9-zInCnps)Ch%SS^^Qc;VXlE$9#
zJUiq4q~^KJsa(q(J1}?5Qw?`FyU)#kENr(rG1BMi7cuVb9*RlIRO%9H9_l_&R(kgs
zfv;&cm|Ew_vci#T<N4+HrXKC_^2jl#J|^sPL`7$EVI<G}X)R}NT#T4`?vJlcx9^KD
zq+9n_`jRoX9r8CPoJ++uN-Fw`?3Nh3+hjacX|;J?Ph4hxA-@d!gp9dEjryaZ2VSp(
z)HH<wv@tOIH5HL9iwFRV6Lh6khe(8xkq7~j%-%vOpAQ~;=v}8dWE&n<U|><NjuH_}
z(DD9OFgJ4+id~Ml@Qf<YaQ{2T&f;`cA;)?wKBRUPclU3>%_+FD&JGpmGkmWM6epv0
zBd4&)7oncio=w#m!<A(*AEoVVy0gsW(cRaM&)X;Ng_Y0c_Bq-l;U^a8BChk{RH?4J
zs+P@6<w6plz-yu4zeJ~(*m}zGx}2ZwS}F0RJ6s!#Q${^iR}MSVTC04-XSVd3PNdT9
zdBP`=qv19(EpV$i;(&fny?b7L?}EvgvDNF@wi6`j%;FMxgNq-=k_|~0Ri?I}bW_K*
z!TgK0>uh_%@1FFG7zDmhP_aX$h~EvRM;bCdqj#r}ol*zU`r;&l($92=hqc~?Cwy`y
z0kdp_%yR1y2XUi0i!v7JQED+2^Y*LwKvri$Ua;jNj=A)c`-+@*i9q7aef3m&DlVUE
z&hfczWCh|q?ShumlrJth>P*Vq`|?&n^skKhMO6(-JDo`va*dUFe?`3!9n;>q+nnRM
zGMzKsAXUZq#--bC+OJ6}<fzPZ78>=x8sG1+ewWnJ>t<1~dNz!|KW@IKb~UFhM}Yc6
z{_2FG-e+}cKF7os=Mb%!tM(T=%E;Ut&$ZdD>X)pZy4DQYJgey|D{B-793X!B_4df1
zI5AzIz=BMxiD4aNS{{?gw3gK;!lz;cfJjb54tqLP!L;@x)7lwWXiZMFQX<4$DvTYg
z?}{)G9@QMHdb$(qz)$Pk@98i-fY+Z5bIjuF9L$t4_nTDEG&`cJ;hey9qNlfc=j!zB
z-QuQNKt0tVzY;9F#z9=}xU7~veyb&6t8baqkUNWA=R|07MTTbaCq;2;dakVOkNNtK
z=#DwNn7;b~$Ew{=Vt(Cc%(cQ{o@HSox>z^$Vq2(v>nKg~kez0+kEZ$2_MG&i<K~>>
zVYfRzjP@ofyR909iQ5%f(bQ|HIhJpo>gRea$JBo14(n_Qz0b4FBC{1(|E1en!V3-a
zCARjb1#_sJvd7F@%f>(MEM;xa;j;{1_#W@j8;iFdyEUmesl!y8L>PO2bF_VUrnx>O
z*I#B#T5DROV$XN#)zb#wtA1et`~&kKRB9f2;zf!a&w2%6gG-OdM)ik8w^)>X{mzmI
zClJ(knf9!x#HtDU-1?^fF#3!UN<G3ix*q5=8`9aU9sN+2`Bu5iXW_oz`>}*MF)E2H
zD=YEdwkwn}&*vWP9Ia*$T}sOAct+u7rM}eH=YE=CB<Fz#%m-oFw<K0Sk#(?};BH?w
zUwwJ!>0kwIqc{tm1V@-F(VN<{cv^%qJF;+U;N0G)pXNFkh3y9elE&6mxWQ`ZnDt=S
z{FruznVmfM>jB$>Z){9YX{qv7O)HmZc{{=nM{AuDr*)q#*+EcoR`k!&u`0E!o~zKk
zeJ826{ln^nvE^u=zt#7desSbNtLDF-NN%jsEHf0gui}z<8ZACmU|Bp)=IxXeXZE7d
zI^FhCp4_%ej<cM<t2QVxs#E?)bsT8BG0O;0fQZ&6&<=&1>*>Ae@F>Kx{Aqz(aW9`Q
zw@x?eXP%7~aK1{;e$0LJTU$UsBQ|5(Id{?C2*KM)W9RWza_0)I`>VV!t6uU69d-P>
zITf6W#Bvea#C<wlwNplU2DL+O;(rO4IW;RL*Sh%M8K$IHe3~Gztl|I78mPX~YH|3j
z!lGMz_SfF;nVqZ06=<xlElmupnk7eOYFnD${$?0XRe-&e)-m??WezuYtagSsudf(`
zicYMlTctm);OI__X!b|m^l!EOTUrzL6k+*|R<%xvk!;~w_!=xAQM!VYx18aaQ|XpY
zdSeI?OhD@m#)A-UJ$RJ?kNumpU4b-L=$hLj;${<MJ;4wjhx4`Bl5ETk&qOdle8+er
zO@<1@vUcQNp*3MCh}yxN^E}~ue6wZM&Bo%9eWE(kAIsBvy`_&*V0EWb7YsE>cXO!&
z2t;ULCeq*HnXZ<-^kqZ$u~yBJ&pms+uLY9OQ442Il*)|sxGjBB%V|8aC%?W%T;`(F
zmlFE7IAJU<2`k`!R*PpK@WbH4oR9=UYiHZ(Dmps@6lIMP$i!C!{YaWBc2xR5pZzwK
z^pv5vt&N#vGhJF4oLs8^s4bJ)ZBmBwgY(M#&=K~eSqwjQlLF_fOUdF+y%#R*{O)fU
z;#?L&_Jn#S<)#ZrSt-rmxqYP2TH8*XD933-2yQYKvG}{wK{Sn%2gzy?j?w`|O;d!|
zu5K8V{&*;YGUt)-L3Jw397VzSg51KXDoqP2=g~CVc!FoTUM1$8R3l|hF+?2sZ8D!Y
zuGKo+vwYoUX?>&aFxXaQ&O=zTOsXlNxnlb6yJK!_bXycK3uUGjrO#7KOWv(Iy?e!B
zWg60I=s2S*+TC>My?((CLg&mk3d`;v6`^9*<11VLTA)*BgJ5!Qc}{%(q`)30UkkIB
zNn*j`F5e%}=+~IJ8`=5Na0&DBw?-(gIMaO%IcBIc{q6(dg@k|_gneeWUuq?CoLK73
z>fbW)WO+cJxpb?0h{`nK@*cA7n0K$`Om52*1{jVJ7j2bBAY%Ei1}!*XGZ3Wu2JRVD
zYj=g$BU4kSks)+J)hYRZY+Rf8Y9A0S)B001&CXY?<PL7%U%HoK%EGSMs6RVb2*(fk
zuoSqTR-hAk{X@*p^BMCgW?fC4<wv6n;exi`+>&d&v?uYoOM|A|*T)~<!kHa?uF#<4
z-gs`$>_wA;@Amsj4pT43mD6gPO{UK``#vlV!D@4?<~3-fllDq@*ZwfJzA>56k)CDT
zCZp@_*fZ9RF9@lj^7|@IKOWZHa5DZPWTV~>QcxDO(T0Bxg9JF(^vOLUCk&q&gz{w8
z-jHE0_`Djl6q<+DbTSg<P@ZqZcp!TVGp8{T8ef`g?0AaTU6C%oE1<P1H_?~gX!2f{
zZ#!mnetK%nC{E6oJ{&6hQ0smuuR%?$ApcuHZ=YFP2uVFn!fND)-lsj?WO?GDomP(i
z_i4t$FWtwu5M1R4-TOUNL&6hXYrYeTpKsk!j7)8KW5ad(%kL}LSub>D@rlX%<PXJF
zC$D6i3Y?tDoLk{m$Qh4o2*heithD*-MJMo^jt?*UN>H4u%8rpI+?+Y#fseP7(k=m=
z(TWeS-AD;kCu;ytSEV6cN8IuODcFb7#h<m`!Yil%zSeJ9FMw0F=tJ+A+psFE%@}rU
zKE0-*&Nzn^$F3#pOm$XW3(3Y@#W$k0m3X>JH`g2VSOH4$+gQKus;rB3Pg3cQp#%!k
zX)Qisjw$E2926LfT*CMH2^=$3-$b5;1tmW0j)Tou#uc~aNy#!#ezT)b&CHfL5-dc;
zoIaM(mAx#;qUQ{EF+IIka&;)dzUKi}i^|>j;5`ei)&l(3XP1K%eL}}v)4ru^%q*ul
zw`e`iQ=%i)9v=pac7L{s1$M(c*KKKdQteqKJ4M)nGaeMlxe9?<kZcWc1a>A8g#rO-
z)_+RNuG3KUnU(tQz&_&`fl2wSA#^<v#!Pha$J>doytrcQ%5?M&kx_sHT8^YQO2KZ7
zDP6?|Ez1&0%g-ks6$5yH8qmdN(v<acF_5H-Q9zPjPzT|_Cc=#PP;p~Hin)tj`ZX(x
zXN_uRFMnjzcqJDZil?>}k#4^h`YxT%1AlIANg+(ws6frQgmXdoR;~!ILe_N`(UoEE
zlDEr-eU8j@hY2&CCswsyFDlYmT|b&X$61}F9KFO_DZf{I<#}HBqn#r!<}~;NTqhpu
z>lgYgPvUUFB3x41+(PEr_!mo!4a<^AuTD9sO|K;JP3eZNls-LlTJLu5Vp67tD~^3G
z%~CM8K8S77!)lpv(*;`Hz?)leH*cgHqf$-csSiCwbo8);4tbloBn@%UfI=$Nu?TGv
zJr9;T4t0H78d$<Wn}bm|SV;t?s_Ed&!|PfA<FUkt+jC~7K$Zl1<(!8B(36F40O-ej
zwXfSj!bqmB%}@i!xlDFl7HGPra@<1cT!6TCZAYPX5}%6ir@0nJyB3pLtq)mBF^-Ne
zDM@!{3XHpKRKuVNr1Pud<_0LocSu`}7hNE1?TJ~=5A94+T3*iV<dF)WURFyd)O6mj
z4DS@ITbg{ARC4zty;byHIL4A;wa2uCldX3p*QmOCz7kUK<{)kVqzxg2@qq4W%f@qx
z#ZTUHaSi&^*VX3Pbud*4t&D3$<cDE1I~*k*V%%?vE-Y2_iAKF~CR+<1Zldtv>4yyB
za-J;0q{s-FK}qI!`v*4t8fbPxcgibx=NhVe&2ZSD&#d7{wwhgJ5|&t=u0@^x534H#
z@3@$(!YSCO(*r<W^YJCloMCcmtvhFLr<5~Y7}2csM!^OK)dbD<h%#@kz0uhzFuPL4
zU!B(~TS7OFeJLa8(6O|A4~FC5Ax`P8EHQs)nw>FZ_1IV10(R$J%<?)MllI-J^x@@N
zyw<k-89Sw$vu@zHmZx}b^z}0xk>DTk6dX<?n;gQ4O*|It%IWli^#Z;5fr7g0aYtZl
zw8^k>B)F>JZ5@MjKBQvsSSO6i0<;SQ@F^4`#6&px0-ONjy|g{Vv;k4$hb${|b0s}U
zV2mDQS*+%m8i*;LpMci^p)W4k<`2fb(}O`YZkV@t^Uj!aToh()1&88OLx!tilI6qX
z7dCysXM68GCcdYqzFQg+w8ZrUZOSQ6f)0<l<KUwIX~ug+XvLuV&8wL)p$T6QE&{DP
zlev4>#3Ad!%gHlQG`-%xNJfFtq;XK-bXs|TmitcQ$fP-$!~#QNeqV>6`;x8Dqf(Fl
z?R4%(Rag3bKG;kqI7YW73JUQ5NL5(DtvU_03w~z~D$&(hksdXDR<OGG)V`;<TD2Ie
zD+wjkZ#A4*$F-JlPQR+hvk<~p=_p<WrTdLGrvcT_O!FU2gX7vjmo*gNCJ$NEiqL3t
zM5L34KOZCr+mE&#;=Oe#%yqUoA@PkRlx$1^6atyJ`Ka=~*Pw?g1yYQz;CCIW8g^j(
zSqfF3igIUkOq*6d4EJZ5u*O&V9v*YupR_WRFkc{GSged&E%=qm?jbdGvCG<u`z@ld
zO*6C_+n5oxf_j#R^qWAojw?IxiR<e~wcz>Xj-7NR>%mfB(!zQbsc&^@OvqZluFt=x
z-<i&g{FMR~A7)eg{76hc#tp^9Ld-`yu3vGmzE!e%z77XD{o;VFSX*5H-swk*$iMJ1
zluLRHqPBcCA=>G;@Z;xOsOYmoo4pxGtO3zy+Yt^^6t+gJ^%6Utd6H3=!d6#(80|NU
zi?HlkfbEJ;&)(lr0+ZevZAmbUGVxA#nROAyT%veWSDu5bI7W_K1pw9G?rReB5iv+o
zORgZ@vp1O6!cBC1K5ZTwI##ZDg!5{Nw(2vVnMQ4qf_SCtAMz+Lp=4!lr%dg_9IKPM
zN2OP^V%nHGCB}N`^$S)+4r;rn+<7}wa4;y$Qp@Faqu~@z-d$g6_YrP>>irf+Z@Deh
zH|iC39oSPvo4b1V;>yk@KU#JX#%#;$PTz|Z^|~wV%RDK}Z%??@x78-_FK8?cOP_=D
zY7+wsI>WN2^y(=6d!n3ajcpZ&20&Vz2JKwi>I3kLKT<V`XlVk{_g^4gAhQMq&@$a8
z8_>?%0>PRM??IWtdQmuek8>ML{d3>xDtyIfs^TB53P*7Y0!pmMxs#xC8<1AXHL`3a
z0K7YL$-?}og91~o0-V>FnAIQhVZPhDXL+vEA|nnU{_LE2(>1V>N17(CYL*_7(pfNN
z4yz!t7CGH*O%XWZJ$vVWbXs>CQdm6gY+E=pu~WKOjH1eS#>riL^%4karROtzrSATq
zOWBNx+20dY>M`sXm+u1;mGXfrk$XG`b*;|mPcc%2j4kx|bN6+>s5T6bWU?2VJ|8By
zIJqO;l@J<Qy70;8DdrvLa(kA6?DV{~qE~oE3wuHSE8{XCuz~Q51(yhOe(tQsf(RGQ
zCJS5Tp&vk2VH3M>Eg-9~ahfR+Sy>TTNkgLy{uI#=L{>mr>_=pUhQiN;>layHI_It|
z4q5(q{iyW*KJljBa%BY}Mg=(6R?Rda@@xuB2bKczD~QHH$L_6tha)r0CsIfdnsFh+
zod9fmb=vp}^3e(L!h{OwT8=<sH?acKFcur?Y#|cI7s&RGy8KY&p3Y_SmN@-aNGlu3
z?$ycC&Jvt^{gTc6tb&6AY3{@uUj0-kg~X(1DOtX*I8BY=<R5oge7+S5$@VblCEVwu
zo@QjZ)bB&+c~Y>%_tQ%|XSOv`yDV91Erg16*OX`SvqVqW3<FgU_qb8w_tDf{cbX<k
zTH>7*?T@EQQYI5yo>5eMW+u@2{^K#M%SxZ%cyG05OVjY2`+~o}erP5C<KDT?IlK~H
z45wG-Q@Kh@R97=Lgx@0vz)OocVTgO#UrvCR7P>m}57>Dv;aCgOBW9477MNO_(LoS|
zbujZ<Kgm<)QkWn&oRwAA)yweM1#rkEfPOs&r*D{#=OH`#2nPQSfw0?ho@{sUnm!<X
z^gxidox07?tZ`wIiZ}?c9dx!pfVW}hS{y?K@$;YPgX3~~c?~<3KHMpK#yyZBaS%>X
zJnGs-Yq@m7`2u;i)~t0;`@S+?ck$%rJlO*d=12Jz_By$i=0GmX*4-SmqB4Cf<Xu5i
zp|v$d2-flYONZMn-rT)+H%|nwE_G;k*z`lyfo7}Sa8S}5lxKSuMd%)Xn(bgz(MXmX
z9g@3OXQ*o@?>!giiNQepDrkjWsm_{;-aOemwXu9*qv-D`ay#p?)}1exCz|V3jK{S!
zB_|Su1wBi);8-F=Qif3=AKtWijfnp`u`_H4;KBH+kLd5n%Y9*hXn{$s6}02{TcG-3
zv!^K#rz+iq*OcZ>a42WbXL{eKn&w%NAIl|{2Nd-IzA0Z+RsJSp8^#~+_ibw3SnsJC
zgwbbMYtG2uV)T}@Ge5TsCUP1xzd;$x^Fg3nLAwY3U%SV$_>JL|*g{dQ{->n)*4|*N
zonVa9m_Bxi(7K2|T%M-B(0-$xYuZb$N|1BWZNYm_12seewAmwG*82CwhkCKAjhbVn
z5<1TG8LOo=ot|6;iyFHRiN0wL@q0hNZ(j?&%KX=8kBQ!4gOI6G-DNAAAKy;QrMu&T
zY2I>Tj|y}WjJ)Se_{X(winW{gCzasbPBSzt4V&y44Zd^}m0!4>t8(8ghOlS%KGlD_
zI;+{g$3LkOaskd)cg|@!6P7QGDgRSr@P@wVe*jIC_9EQVYZJvk49w;&W?*5=Lqnqx
zr4xYC2o#P`x0;CRqh+uM<SlGh*b~LM@6f5^OqX8m&vTkF$q2~+>8=!qgh~*W6ZM3c
z9#Npl16!mIyTp>lM=G4d3auSo$woNGwJsVB%xO|MDBO!!tLe444xFE5V0T#_qkWpn
zxnf)WNXMb2YraE4Q##t`Sg5G-XrZ;$_0-o5j#K*G0XaS0@7y)VX*udULkg!BKJ6$e
z)XuEfX^OAX-482Red|>QV6(cTvS(;0d0tnK@6_fW2L7rm3r@R5C&K!6xS2*qql#IR
zYncWAE>`yxqswh>tNNCMuilZ)o>o6KymM>MPQtVWkL}U`^sVTY=8IIeM|%sO>bn49
zEKoRc6Ln#4r|}cV^%@9y#rBc?PsG4+B`L7>Jl!p5isLZkrd8LK15ob7im=OfYD=Hd
zP%~p)tnMrOzqjU@$%<5T-hwN~6fBQe=wn@Hjo&YSenC&aU|4>&hLks;PtXIO@}4R$
zw$QrzYIR7VHBYm{Qr5X1In$*2uDd81k%HR##5i^}rR~0RLb6ks<JS|(wJM=fc!lr~
zZOtd}jROwQGtp1rQ?XDw$3rqds5$lwyJ{kvn$9?bwH<z)wHQifn#nZo@&nWPM0w{3
zHCsYcWOJ(D_~ftr-aE=(n@_kw*M20hVD)FwEU49AW6YS`X{9!mt1s&Ls@dWYhep`a
zvWMloPR@BASC?j7eDj(c_A(udNf@XqKL(w(RWak6GbB&~aTQbM294j4g~srKi2fJq
z6~uA`iRDkge}pZ(PMAtr8<9aRy~M82<m(^r9TInOV0evZ#KJhYp!tXI57l7cKr{aD
zJP$mK=$4(`dF-IC*sR*h(-Y?&_747lVg2pfzc>f{A*2}LUnNASPJHWn&M@uLH)CI%
zo++^BA=BCK441MSt6GZ<Rfk~6tnIp^^0Ux)wvL+z`dHFVSK7m_s<jLP8IDoM=~%6c
z{3I9WI&V}o*|u`L!N-Ydv}HhT;90}!6)h<8Pk8EJ&Fl(Ly%CCY+_dwxX;iozhvH=J
z>t$}0ZthjN3zqA}$CL9sb+U^4T9#Wazq}w^@O`OjVufe;x)P1ofqVr^g<7`^W2QCt
z+cOAKoQ0SN^E9iGGzsDRwf76_ll?@X2V8|D%+IQYxDsb{A3=Q1OzrxeA!`vb4?>0<
zXlwvZeMX~iJ3`-}k(8^i;q~@Ve#pE8XPfl(YyOs5P=@t9hsqz|p#&9KBHKdaU_RQm
zdFzheK{d1Vde)Ol9d7|AHKb`|?}3+;lq=|LDO}6Kf-sg^)9Uys`MoklMZeMvnSFkW
z<acz^%f84n3FATwj$@#v38g&UCzRq-!Am;jr#sFmh!r%G1@MfA9lg_0F7Y_8Al1wz
z${OGOF;G3Y?FP#63u-`LCY7w`)Ki)(lJcpx?(kJJ*Dj+=QC%935IZu@;n<~Xkw?S4
z)SEJ2QL%c`AyZ>;;2Zty6Z)OA7T<#IZK`Cy*wI%Om3OEt%-WLS#(>5WfTaL0ZLE$E
zHv2Y;FDl|`Kq)txaZKwsxkhB%yybs{y=(oN4-2t$j=GG0oIsPFtL=JB98mXY>%g;J
ztV;y4wU#uFJ_7Y&zHJ7Hn<G9!z&q7fgx#;aGGUqFY)xZ_3o{g~SC+P0R*^iCo->zY
zo>W-R4|J9rHRUrnZM%eoy5jZsJzHKJ`!;xYc!;;7((?(yqTpse%J&q7U9g9&3h&<9
z+E@{4lh<1a2!UEb_of$>3gN$cs+8>*AkV%Xqf`>2*L|VvHF}}bz1b6`z<vYoKb0xO
zG{3jD()Gae=h4!}{RI?y7o#pq6lw7>f!Tb<S7*CH^M6ukn+bySGPV352eO{jfY(N>
zWGETCE`%JBix^0cH^6grIwf<+M5)O&7MGFNk0zM43Al;&7S6Xg<-LVn!xxSTPJL3h
ztZ-HC={9u>CG$%pc<ON9n126bS=Y25i{-V~(K_}0;fL8-Lg=u%5BQqTcAB>jMK4|N
zd^76z;0D%VG>Ln(Q2>9hrd&c!JYf$^v#RNgPn22qS&TlGWA4=WHk0^Hs%&v6vHiNX
z>5$sw9gZ0r|NNm6%SXe9QITBheda4g@>0bua{wA(GtVY`*Q1Rc<bx66cJUHw7C<h%
zPzof*5iz34Jc9s!;`S5r%24(FrMJOg@v2<B4_*X+5(oA0#Ot9oq@uWBUS<W+iDESQ
zv67K9nr*7(Jd$1w1Bj%g0dpr|eq)iUMY}8O`L~*}Oc<3lc%+@tEIPYbm66=s*x>zA
zb+G07mC4)%vhGgn{!-o53HOOdvAE)dIYl!bZu_My<`U>pH$-%y7PgOY4_==0a$@0B
zihc7clz{duPNV+*QG{`EWg|K6J`Hu0BDUI|hn-@F04yOqH#4Omo}%r6H}L&c5V@a|
zKlf4Ea<S>jL?Jec@^<ffquKPmd1(%BPXJkISr0B>ZG54hs_u$n<IwU#lPTUzueGv(
z)^tvTK06^zFF#49(6Qs5Z_g*+m1FJ;?n$Ky?(_QjnP&Iir96V(aQuxgX8tG|(wjoL
z(?)rNNweDURuBB;-~YU(s5cRWkUkMXCZ3-Ky6@({fF=mn2GvOU6SBG{U?To#!9t~l
z-ZQ=iqh!DJ97}&`eqw<yfrPyP&ZoA7&Uo3mePRLbZl9l@;D!^fh#Za%8yLa4X*8*o
zb`3;5!3y(^A8Aa2{veUY)U$!c%*-nMN5H%ema{ttKf{i4L)SH%{S$|k{8@~OU2Q)s
zJ;rT+7Ua^PVVVFV-XeNal>!!)L9;tBuqoyU?6~%YHAQ_mL8vt9*Tr0}9+0TL>YXbY
z6@-8uwCfe6^|?}d%0MY_AvJ<6Xue=0Nl&uFP2Qkq+eqd06zXpnK?@7awMGQ>0k0uE
zs(?_l2_(BFw$(5H!UBk@et26R&<_idvOmV6IXC$`qma3#)~7L#=id2VbV$g`APh3t
zJ+mt%Ty6)-VFF_jfO)O*f83!0x<mCA;*l039hl3uAnv}wJ^qI3jD<#|9EXVnJX{?@
zVR{nKX<%iv&cg@LB-=_aiY{P(MD`C@e6)6RVVCS-yX1ZS#p!{CJnvvu?Hhr9N14a9
z$V>Sx{u*z)CA^RPxDqp1FxwiV#;*|i3}0BccmL4p3kMJ{%X38W-z@1Dc{jHGHg-=9
zltE~UT!Ud$?Wkk#3(E?Q%+rDkuk?MMaO;<zhZSidSL+%Ir=A|>wNehV$QjAl;XDdM
zc-)rV#cST!g)m7AH~Gt!PJ$o~YQs-<blzReDFD28qqP80#QPEfqp#Fd5%+uHF}~S>
zqGYr>!ucFq>Q6!r_e@)v7X%~))**yRD*C;jfJMn6VJ=G(Wo1yic?-K&_vAMQ7EZwu
zTv{;M&qB?!W6Oo8wQVkk-6<leh6TU}&2{tXw__>{HNRE?nD?amxqrY2LsO#wRW|Hw
z|947yja*kHq4B&9Bi$&QK<+aHgb=~wje)OMLxHadBmtlK90jjQT^*q_(G&C5$E($%
zAg_D?&Gm&@R|zE#M1}TmEl1lyU@tzcLzl<=gs=NiMezj#|K-f)TcNK+9j)S1tAttF
za&x6g|4lGmurLjQLen{c2=OtC|HNPkA)J16Vjl8vM10=}z~Dpp%5y<=$ljyz*XX*p
zP)8Hk0ijmvtxA(CXL-ImLh#@-=Dtcu2Fp`M$`-#LHg$MU^pxQnH#WGebwfc~);eGn
zx+^@Rxl>iQEM5*1evJr*gpP)30X{S`XhEWkXs{vn{WQwvVd22<g#903wc{>;JZT6=
zDgey`poyDx+Q$xL=S&`Yis>EfT7ZP}IYtBa$wGNM9Q*ieb!BYTr(kt?Y+|y4s|-#`
zB^|Z?%3W6cid+8;oRqM=QKbi9Z8kYfzfVp6T{O815(-+?4P(V?W)S94+JDWX5Yf0E
z!P(Y_`&*cx?#nNrh#c}6G)sN3^D(SsEQ68qAVPE-)^)C302Lk8tSk8{T(6oC$V8*1
zup&UIjq3|8diJ5xl4znGk@=oE3Cb;Iw>GLm#OOF=`87nM&4EH`2pVB#jzR9<;~|&}
zmVyH?OI<1Nq`<@l$Fsa+uRm&AcX^@~yK-6uYNsi~*q8FCWJff2`GT|p0hJeSkHDIU
zjW&7^%?>&Zq&fdydPE;tUsAF*HLDZ`8nF&@E}R5w$UFWJfU&<ze>tn{mMT2P*?jkr
z(<ifG_hX7{YBtmkv%?SkK8JPX#sai4II#)gzl-^qJcK~x3b*^S9^s(@BOQq^6|#52
zt$9Idg0rfe6>f6B8iMdY?0AlNgVrdh>NwsMU0s?oOJpYzW>yaO73zOA-%VZ6)bAQM
zaux(d<WE6C&37H$oZq25F!=?7@*m@GK)-d|qb(QQ!D&>~u3QtXbq`zYjHDR$epA6F
z55={STG*#UE&3XEZY6`SEkz4x(5@61CV$(-o1THDu6+K`tz4vAjg2tZw9yv)+Iz7!
zkoMa<;5;EK`!(uV3Hs!`HlepZtHwS>3o8U*AN2$*@F<3zS5wijvxlQW-7p+QA@RAB
zUv4J;MQ8;%Zl(Xjed>byT%Y#ca6fe+kRtxq>?=e@)Sxr#g4MM)pod0j|0e@_AQ)(5
zpdguo_{m5>Fielx(jk`rrnMeUNHBsFa+u-g#i0#7%Y;7gKi-5vPpKkx{fn4yF^xq*
zW7qqIss7UYb8R}oTf&IvF|Iy4GiT0L(%3Mh%{Rlkwu9jJ8#t3k^7N=sN(md8Xk%+@
zS3c-X4oGBP!RAqL+#WRu6I#kWqpjoE_tIe?GDy#xBjYiwa=!qZ+mw@HYqq17T<&*A
zDEV$&X25z$@jCDdIvT#yq<44_>3)Ci|N2V;xgcylIoFkS&43DoIU))a)igVg6Ah_&
z0iap+snfzvINVDRb{)|AGb{SQQM!W2h|LeG7HHVmy3c>2(}Oy6dBRv@lI~Hf&nGOq
zJa!@&gm=f=7HTwlfi!5tN)dX>DCEpBdH%m^@^3cg-#br3#-GCkgAD!%5>|7Ic2stF
ziArF9+mRVAPhnTTE{~`*YFW#>T_AVrfCT{FiJXqJpg<PyVYh4tH<U>ai8|b|fpcJf
z<i**(M)n1&h#TiTuiS6~Q#sHqLOe0{o8<hW;6|KiT_W&qWVWFa?pa9&Bv{r&K-d&k
zK3YP2ToxOR0e6@?%p5!oN%>A{?e>_<Z@a<Oo7gSGviW1M64#qylXe9i@<(jM{%9KE
z{!;D#mv{+7h|uM}P{|8yk3k*WmXU-|a77V*@L(_Nbo%3Aw@4XU4I@qGYQwPEw-jrQ
z6BDQso%;GF4h~P&*-gXGp7^2Li>h%9b}(7ODgd+v!mB0q2sMh}q1yC+jg7NBhGI+O
zpDgFUKN1KY&NV21e&U>!axko;)OcUCPh#m5^hFNBY3J&rEn;(RylV>}V4c%Pv3Hgp
zz)bD`ADLMgOf&n`jb<*Poh)0uos8Y5#MWX{1poF$T5$$tYPitrW3Vqvg)_nhmLKpr
z!}{nXTC^%{sh3&e2`sb7y6y@4=DlDy1Q#;r2)tr$H4VEFI)|-cnVc8w?U#lvjAWyl
zW>8G49eD-!zhUvj3({KfR0_C`^8Xd>-PZCKgd9HyPb2KM`xRRI8q&qWogW5CHMx3y
zY$v9wu5uej*{BSTDzWy!mtq-yxRf3?EPf$Pdl(KdFtWZ$=Q@3v!H~gYp!6d?y+MO*
zZ2<@9P;0Kc3LV;B!27420q-N`%>29a`=uL1dTitk>Y-UD3UD-!7D_?=Pncvc&=>wD
z(3coA7w|Zy^twrG$7eWbqJK*z?A4OH&q0_OqDDJkZs#xbGYXlYVkl^H8ES~ZDr2kr
zVv2&_A0QBd4V)#)V0iPUzllo{Y|BPaD~^ymC6ua9A(^Ji%l8AsBwOqVJ0M133kd2r
z_sb%yCfg$PzeL|(W6oemBUH9s`V~&hK%}RI_rSvj52@9aWqJ}@_J!%uc9F#V#la@n
z#U#5jbq8YU(n61|G+~XI7M&j;0nod@hyUb)qp|tP@vtbjKccPw1HOaF&dfBIx_ib9
zJb0sH7aTt~kri7};f9nO<-0a;ctWqmDl7@rM<^G&;ietViWpI$y^wJFVT1Y~BE(z}
z)a^IoqjYb<XaV<R&AT3$bBc2Q$b)giJ($wZ>%ii$;~R!QIeTI=5^Xn+T^kZ^gLx(W
z`O7<ty|h&xzm?Xh%YNjo1(i02fSa}~K(Zh+dz3i+&)kv+E1ID1?q@-Zk~Xea+6?j{
zyjK*qf**glkB%0MXGMXDX+pY6n`POt<c0pwpph9<^SJ4OXe=!#8>`&s?Az2V?~@Cy
zH72nbO)nDM793UsoDzYv@yh^YgZ5URS8~cwB~nav<bbcd^NGD8Blo)L%u4_XqH<mR
z{~_!xz^dGu_HjiNu}A@x5CoBy?od)dz@SqQP*S=>0TGdulu}7a=>`QPg)Jc6xoI}t
z@SlhCe&6?=^ZNe&*V)&(#EtOmwPwxSbI(1qG#s$*jqX$j3A2zijL+axSa0d>Q~69P
zHXT*2XC%1OsYknB^*HLxVRQRDmJqS9H_I=l!lFI2y(c%X-Vvraud(Xd=^UA<sbSc%
z_58>!vb$U&EqRY*_b`vLrOI=$(sGTGWy5l|sdcNmcaL=l{pg^6k7vQ7bD7sYC!#9I
zXSaM|%fD}{lbgkb+3K0&Qm6&b@`L@-0FSnFm8bxe`NANEzHk<0oO@tiV%eF1x_e;W
zU;SFN`p2fj&KApK)lbVF{X4G{rajy>(PM)t%Q$Gmqb#-ft+qX_4<k&i2Zt8>%i+_*
zX#IH8eV4Dls&iro*0(s(HAjaVIxIVu_N~>T?(Wyh!deow;HiZ*g3vsN%h>@QibWfH
z1Ma1&iANC&L4VrNr`p%A>~?h+S1yxBY^-(xoH&U^hY(`aDOpIML<lk1$-mK!;n!~K
ziOGN9Vo=dEFU+faZCSA*W$|GRVZTHTfc+{RwSx-`TkI{8VbWA~fn)6&q~c3#QT@;f
zT*N$D$+&NC*5Ob|w(0dUcS=e(K!HMVni78Hg3)Pc&KlgP?wknL6;;Y!{Ey)fPu28R
z2}S}>>MqBPoeTeiz|K?56o8Q?laVmExpYl%C4b3)=2YR2fu~Ri>(1)9f6TcPDfxk#
z{LZ|#!w`2i%<G(iJfn4=3xPslR7HE|Q;MJ)l;yl*&bexe*8Ab4|9Zl6u3M`kKPGb8
zZlMV_|3R5xbKLyeRc2bh9C%<+pZ@UDpPcO?AlZM-J5bVwnBUdK)&(h=o*xWt_-6Ct
z6biDy03bx_LHV~P>IeNY+sRy(!xbv``*c9AdSlj@L=^gsm*47{K%aA8p66CNW`O#y
zc^oj1Wa<wCNoJ8@NYH=%k8~6{;2+815vG*jE)rmG{;hqr1;QH%{Kwk4e3l5h-8CnR
zkPdY2eEHjj(3Wea{>?>sdl7aGnt)$&LC&WjJ{0|T3%gDMq-&>@1LJ?L!b@3%<+$5-
z1^z`N!V%g#e<B2rQ-IJ0uus5ryfp|C=79^m!aR%_*`NQr4r>*C38pzAb?J=EJXzD>
z5&0aL;8rXQi?T&3Z(bfpuStT@x$G|Z*EE#t7lCH&RC2ia_s5WthV<BWFb?`y|NaK7
z2_XW-{u;NEhqNff>7;qXcV!{@$i%lb7=fk_0x^W;-4_JeL4p>(Q(^#*Nvq56^g<-?
zEmR*JBC|H95(KBC4D?}oQ||a%rXrIZW;POdQo;ZBq%6pjhB~nR{ejXD-F$eyJ>lQ~
z`ZzD;LXhWQpRMEn0Lil;SW*g{`1hp75Mr#>+ogFOeM3iAfL}!lCW}tR8m{(mV@WA4
zs>R(MMzNyXGLwS<;S3QwT9jA+!LXhKkNGc7_uq_2Yzon6#}>Z-L#RkXH2bg1I1v^Y
z=`eTRRDi=G<^m5qFN6_-OB82Dsy%*w;AT3aggNMJu+_v(1?Yz!I#9;tTmL0K0m-b{
z*9{;suZ`@qImo(crSAd2ZKnEft9Kt+z26Qze_y6s$S9MkjY7e{|Ml_eF@qHve>!i1
zWR+9!U`)n~NUbyMAy{D>c#En%3sz58l<*dpV@@Ej4aZPszB}vb@&H^-I`NLP?F}eA
z@)F`4hWQcY8fdTzW$jf+cnD#a&p$sCe+@_Qi05FV*ZkM4Q?QO{CZ+UpwTr9G`>GY)
zg!wDlstchpX;sJouj0`O$F;?+anSD9PYx{mPxfuPA~g^4yoR=R4_}rEe^WQsP*DeH
zoOb9Q1f2xZ`<%6PKfv$NQXAEybUhPmlsE|Ni-z=%dxc8#1)(KP5Dg8;`T!4BJS?jy
zJY45?qFfseR=Ne3i|5jN-t9avtDHTvWBl)0_0k;k;vcWAMwdncTdRt6GW!jt_WO{E
zpV5JfLdEvhk^hMBzQiD=+*iZSLe9Ul%5vQ0chxEV2KQ^y%SXFfnH3Lv23^8kr$KSI
z?1p(Dv{8M;W&u40j^m|~{dQ<)_Z->9!EJO_R6VWmKGJuPZ(C|wmWhdW9LwHZ?p(TS
z22<0(pJazhG5ljJO#w_=Me}@gLeApfU5%$$I06^dYr~KZUwg<ui-~?s!1u@?Q;T3L
z@6D`)%CY*veCaPuOpgCXOE0<$gD8XWqX?weF$>%PxyS@;pNT~EEY))4AxTR>qy8vA
zuI#0~!>ZgZha$hFAr;26VS{6Mj}ihC4p(<od|$qA^<AiT+1DGjXt(L2Y!PPK5yp&<
zq%@!o(J7AWU(kslNub>p+)q0NZ6dt>+1e_%98pj25V}__uNaTl`0<js^+aHya;gqh
z-igfO<hD*<n-p(jxg0X+wtJ0JfB&MxkNXayT92~=NR!=0Xh*fpa^Cm9|F%VhI^eE9
z*qwIVOs+26ZTk_gY`Uym8hbik^WdDbr&v0$OsUPwFaEw(N+gIdsK(^c-~amfR7Zmg
zY15J;Mzkw=hz=mUXBrmq7DAxMfji-yPdM5)M%V*(IKMa9`eAA|CEWD>u9ati8PpTK
zbENFJ6$cwR9szKv05adSXL9Fn@{CG{ZYpHnk+}MKAu8UvypqwdME%5#9&_aq(JjvP
zpdS619aR$HMJIr-K{kqfLRGnAyIBHE(I@uv{dns`PVGtOHH4xXK&HDrzx>ueyQ)<0
z(bnMedA^Ot0~*|u#hdxGuKS%C*+F}hE-9lEt*7Hrr3)YU4u4fnWaw&?D?2FOyiWhf
z236_$@ViZ<cJQW~Z%`b*@8U)rGNoYZrbF@c8Y0xrlyj%RJKABB6jnhkEAz)eE#CQn
znI`xOI2q2)b3T8^#xFw=7K=mn+CPiDNPH`R)k_&j<U>4*7MSB?5)Vvu5|adnb2|P8
z?iGco+tg?9C<n*oiecvIk59goR|OZoio1@T*L?%y@{3FCVIb#9x27dBV>9$T{|FG(
zEZLQa79Iw9XPvvLGD)9^?F-v=B9}t*DFZY~9l6{Xc6Km=<K5H0o=q>B2#}dQkRLrh
zNIMCA{7Q3aL_u#@st&_>;x@n=G&Cf*5+Y?*5Bc1eXfOLC`&jEq_jZG+n$C<!2jAk6
zm@B0FtFKEI4;vy}dg@=I9gN+f8i5Lc&en|YHcb)}yd&;xMyNO{ZIx~L{T{UQ^kM!A
zy}AQ*(l$IN)lvQ(6S?VT3@tJ72lFEapk^?&>vd{6{r>jBJ+$_N+w!`w#-(vNe0z1<
z;~EJqg-?^A^N-XyivT*D8mqmQJ@}Z0-Iu-2EKgQ?xH7Uv+sZvyo?RuO6UAX0<jx)1
z>xC)<A~Le1aQOS6uQWqMV^mtc^arr`5hN}7yvS6x4P<Dh?t#4~WaBV21SaQg@c=a@
zs35pl089(hTjz3Kn8B@P5Rm+`kq)mC2K%y`_ZNKL<U_*y&nWhGxH;u>x9R{5kS~G}
zFoA}UOyOn{TlBXWY*XFC22-o88y<n6+KD(HbMs}#P|*O*=61KrN-}h>5vnNM#Mk9P
zI;}TYj!t9wZTF5zqc_!fZb!Qce|ggH@8Pex6m#`h?(=;b5&JBZ4UyzdL|g;e2a9i*
zKd#Vn`M@T-E9Pac>{f~>ic>W9a*)u5tI&2`eyK<3uT7O0SC(Dq&UVchx!FFSpTS*f
z`EcZo|Dkbdr|&hKcv2P(a=44cJ!g_a=BA`B^n9OrIs9~Yy?t+)(q*Z_?!61?e5}tN
ztJd~C312WsfhqUiP`!OnKXNX4xM~cp&AeZUj64d7DX47WiD}RgJo4CV!){C2J(!>}
zxV>rE7$L*uy|7nhi7RTzAowZI`&OR!<4|VF5W0}K%#cj=Gfx)e&Nfwb-X@zka#lEU
z8JZe&L6y#z+K8ezR_BbD9Gt3_(fmC@E+gnKZz5Nxc!Icx;(~|>Z#(=r<GSA(q+r>r
zl+EN}?Vh8}X=~nx-^5JGm?u1ayML2v>OJq{b10GT&5}TClneQag*r(PUO)Nw*UMn)
zZ%b{|?@cJfuw^{|{1O|BmK4LH6~p=@=~-j*Y$_&|?j)g%5cVyhM~*zIYf~+eFuhlm
zS+nRT`Z7o&b0(n9DFr02Tb1fZo{t#vTdPNlrnF;yq26yQCX5f>`DE~BV9CZYV%547
zN!OYfE>1pH9XIel(}nr2$+9TGTs{}2m8mD<+M7%#OVXUvf4M&#!%PuQe_<~06`PQ;
z-OZ5al2sHURuMF68QRvvcPxvg-fLXWq*V-YC|(ry-lPh&SCc8K$)EZrcr;wQO%<s(
zjWVTZi)?G8a<I&h`5w^|y%zb)e8$mNJ1D?$c)-Dc-EHNn{C?BQo<nt8YkT%R3W+2c
z3%g+=<F<)Homz5xJY#L)A5(H|dFOr5-wsQ^58V^PJ8AFKE$(k{?>`sqTNHNhP+yeO
zlgF@Te||k4|LXidT7Xd|EaG@K$%y&T6S%>i?6^vfJVrWXmvMtn;Fi$9+SlYO2gvwN
z!)18-u&CK44U0X=`D<?oZUi&DFMK+)`RYuZ%ozVyKlaa<%w*GCB_iDI@+Zp-+QBS2
z9<18pD8;G=0$b_YHr2alll>*sE0c^-_rzU3-Cc=XNHSKf$(O0ky4rQy;cLVR345I?
zcfnmZcF9N7>&t^!iyyjQ+QxEgZ;eyx{ap9Clx!$nlRxRcJKjac=@!ybIpUanW_bDc
zpe&(7*GBI%^ySjMmgnbh8&92yL@Uq#cAd9iJ`)u4$(Gz$`()r;b5E>~u#a&h-OrP>
zdSqA&t7cJ;Z{z>n9FMK@Ma+qE!SaAWqCc$dU#orn7Oc{x<^uL-$Uleg_c1PRx1vFo
z3_A?IIES)7y$x3x!H{$^sJJ#jb+}UH<AVneU~2Z+58S4&i5VU`nrkPrO2EQJvwy&l
zQyr0svL0MaP>4^&m5-?}q^59sBq>y>s=xf>$xIJtuI&&{Ik!&M14{AoV#(ctFReMH
zgyH6-Aqpr1oT>n=zC5+cNi&LBdm$%B%gNU8Arpo5sv9#rwz0L@-%ap;OLU4}qLj`M
zxuG2WB_Oix*ENR)@knX|ZmWnpy6f|=ZWc1qFf;o(wdtZ+KhFo}@*SY)GVI<t>JWsx
zxN|({bTSQN3PkPN-+lVp_QRaF-3HaoFHE}DJak?)`6(RJ*9mW50CVJ#XCym|tieBC
zGdJOgmxWki{qv3g{XT5rljl%(67EtSFRv}ZM@;5<Qk?gQJw7?B<d}k4_)4}sSn8GF
zzKc8JzQ=NjQfS7$eY!2~8ql+1dz*^{SFX%%3Vi1Ysj<=HG~ycD*`G3<;kG8+q!9js
z`_9(kaMdfj@dJON(|+Zi6&d?;9+K>xA&Y>q%Kk({uelFN+Qui-vL@smmpVREI~P5F
z`-ZrwwPm>Kn*WoL{WnwBK6A%)amn1zu5_Qjbmm?3>x+W=3*8Ti;)nXmO`;>&=MIX4
z`3;<zx)W2{8WdZ4G#6vY+RBaQj2j~bQ+KpNN0v$VU*#C}I@B3yANV%Km37qig<Uv@
zDjVkDi^9fI@YIe!6_0i1PlwflNYCMaZR>y6>)-!?&4R35jKCQM<g8o3+AWaqDA0f<
zzw{*fh>;)bGbFf4OFIfVlUlyfMYzY<J1UASH7<eEu-W_EWk%#u`Lsu=mvmfLB)xro
z%LnOmFM`wTeRuV~<8bcsEsd=`c~WBTC)TSK%4EDYxkCcy>n}?A-Q8&wjSMXCcOQCn
z(jaB7|B)rm=n7S(|Ld-1jYX^Nx+`4n#&5FS3}4JWCQ##YzqN{Pzd4bPi*mtB5L?@l
zni<v%7}?k5dGf2|ii+u##fsE6?Yg5fyp59go3FZJ*fAN;Gnr19*iD(X-NX>b+l`dP
ziea6k|Hg_ssd)0#Kcfh8Y=^+<sq=Cu{OP^pL;h0Y4up@-GB@z<AS?9&j&S$8x2KRp
z{!QsPnAm4rpO~1i+FH_nbHUsx87dP^??o5a=eo!Jo)UuRz1i^$ZfySa`wCrFjm<30
z12Vx;^QSGLO%;QcB{*clb|;Nk_U=2-MBKj`cz0(l3e7){sW54lkq}O3F$tnGE&hBo
z!%(&rj{DxX+V{hhoLZ|kRVAsm@4I|Wl#2X)8OK-rOx+U`mhQhRx9FGc#2_LC)7~Ro
z)8f82B!dnYD{df{8aPa}jZ?ipqnx4G-SsH5YSJuP2n<O!+UwpoeC*z-KK$o;_<!*E
zAN|3n&DZ?+{)hMevb*4CaMBdGzD(Xh;;<MISk;HyR`|%7B?~<EX|I`@>j?2-9gI3;
zYVyz4j<9W?)@@ZwI{#K>Y^Uq$6nmzzVAB1#=8D4&O^4fhKe^0$2@*{eub44W2olD3
zd8-*)7et7OEDe)ba;5uy{hB(Sn;6($G;T5I<r{T--dH}mTDC<+#Q?n3dt+3h4*KM@
zm6Tu(_H>|Dwz8v<wlE8Uu5}T!1bkjgZP;gVe7#6x@uZh`{{1-o>(8Dyu@;sxafxt3
z{=*&p>pR3g(~6DpKS&ApMb64eU$B6DKIEfc=L5?~!H~y>U&DUz$nla{=)0Rjit)wg
zCX86SPfKQC22-re*XIYxztB94>Zx{Jyg5w8eWiUbaHjv!8IS%{q7OYeCh=$){p<(X
zd?ND$xFJ_MW{1U5*qST7Ts<m-0hOe_pL9`m9d-lj%sKDOA7;?J-gxHw`wI;@2Uoun
zosxrquj%D@bydZd#>iL}|L~gpCCx}8ng93x`oaF1*3{kk@B53yEQPS=Q=gCz`HdR(
zm-!NMFwQDK{AUotfv|b!H*x%_m?S0GE!w6>Xx4F$A>X^SdX=gPmvY9krY{CF3u#5q
z$vCRK_ug1aHr!9n;Qh6iZ@JR&eL|(y{D-z8ttCrMetR3!ZC8fzyzKah7&d-Q5^f)3
zmgTeL;#l>xaDMB5o=-ep@#8~z9B_}1tQ>6EPv`%BdVk%7?J8J9bwXJ1ri7>e6loJ5
z1>0qXk9awLtH*gK74JYiy@zWz*7rd5DT{h;IO{1EAwj}uq6|#tMn$Nc?%NWRRr6cj
z;W_-8!|#dX3oFE}F5r^A2Y7molkMHxKj2Sh+^^-F^DqDK4zi4%*moiLqkpxDDk@^p
z;yZ&8`RcSe>4WpF18>9WgRXnfbVlej;ymbn`xq+<0A@YRu{Zx=;9=*;lZPRPyKW0d
z<#GpC<3b^J9Sx2y(T*}eFx?Ehjz$9DKVF{8XR!?&&`2PJtcOSZi)JzFq%mCER(0w}
zxL5o)P?$n9wxxW$P@iiAk{9X!czef#=dZx}*AxHy|Apwp#`+$;iiL+4!}^ZE>lai0
z?Z-gS@A%ADTyvH1tjQ1hW}?Lt=8fhFFiq4&Ps>~msgZ98;mSBY;+dWGEBiTxH~sze
z3l6{k?gtkB%?}td1=aW3J62tD!4s$dS+{zzy{u=ky$pBQJ&_=kUd||f50z{>v(v=O
z_hO%I${5al-J~2rq}@irJJ%+FHw;4aRd)O@u`ftMw7eL8>5Mq?$p3h$V86;Ohebbn
z=(sVb3=M2J7q9V1RJrclEz3j2@_RwO6VbLcMzJR{`s?Ax6A==Aby_ou6>=g(>eVvn
zFkkxrv&jE5Pz)Q`+u>B}5VCgHf59%}2%H*BK=yy|#xX>$ZFbtmP`4$B27wBE9ONZ@
zwlnyX##JR9sfsy5$s+eZa8;>^TsM8_c<HSDTFNI{p$xy0;-z|hOvx3s1E^~GZ5Jrg
zbLHrE386wTT2;Mr+sBY)XExGgX(*`n&At5G7p==>TU~<RRU9{;T52|F>hrtrl3%so
zX(W$;BlmLWVV$A2aFz`C$V``WDCd#WeDT=tq+z)NB8j9V{<&7jhC{^6BblqQF*9QZ
z@v<k-M$|Mk$o((Uj~_plAz24%nB8sH*7a})eM`Hnpi^N)iX<NH7JmhjVFvUzTp9=C
z5?&yeAJEz~i~qq~0S~{1^YQHkIvd$4{9E8^PD#e^9b1_3anLPFzut+{3aXiUxasf3
zMEyk+qx{4|nccj=!3&)Uj@MVE;7re+r5oFsN~?&vohBIH8qs$_c<H;ieAEVK%1_;7
zL4tf~Jh4*!)2(adxTr6WB=wqjYcY4T7p8Mpf~X82`C93lEdAf5TW+cy5pP|k%e<@h
zSNP>T`uPkU{_S5b%L{Q?6h2Ut!y;G2s=tA_tQOoGe{fkZ^%j<x)z2G_?Lbu>y8BN>
zuxO?bzW`N>xQe5Q-CWq}uNk0FZZT`QrI<V&bAmTuUfhkP;mV9yAGotU;g9*n{(U%a
z)qXaj$G({v4DaO>YUmiHXdgT37*pv<Bd52Ua39-X7JV_gboa&ZqR%Nlz`iHOPQgN{
z%dnC0lF^-;Dt=XE(y6e1mH&Y%&2L&&NS>JYYtaWvU!D?V2p{q}l4tF{`(2px9J7_M
ze&+|FPuRY8xf4#0<<K2S9>_#({Dzz%jlIxIX+vhd7t~R>wT3dP{T<>KA}Fo0CM3OD
z(B3w0ikwm1d)puKzhG_&&6nBnM`2%(L(XW-)iY99w4<KiI<oB^aC=<pQZ-)OB@!kz
zyCDc4Ee`ukH%p)nw~o|m(i0U<DiMWm+ZS;#vpMVns*{*H`ETuem>&aMzs#4Gw+|+N
z)VaFG2O#IoRpGi?x#0OIC6f5$#1*&`%Ma$AD1QlouY;&W14&pk^gwKI?!0Qkm#gZI
zMOn{*g77#$dFCIU;;%1xkf2X{e5eZm{ang&T`Pch{CtmnpDed-mFpEyOK^)a42{q&
z&DA=2G)x2>+!Au}`n+fA{Q)EN(UhscR0C)1-un-q;;LOQR!S+<R3*}Ldo6UQ6F3)G
zydrLTAC2ZCir=%d2vlgDyCJkcTUESVuiwins=xF4S?gZ&w{%0D`n0W)yVLJxv!?k8
zi@Uthro5IDUToD3)7m&(%USkD1X(!8Sx08Iw?)C)t_B?R168T2X){I@KxvkmmuOFl
z6Cq$I)%QC9o-~MvW|?_uSHg)7Oi62ZC(tWDEcf1B);cRSVIgu?+0i~y^=w2$R|tcU
zuk&flK78b=oA{YSaX#mf-1Q-jvyZXv$fNj7#j!pj=UBNvxDMLUi(rhEcS1~&2qlJr
z-$*W0s-J|ctKK~%@)_6F*MEji)|0afW{gtjQ1y-mZ~gtn;SwNj93ldOlTHsS2y&E&
z;=7)Xd3)2V<+5LO6FpLR^fsw;#rw|C1B?$kvtAI2pQJ9hy6(R&t8*F`e7K<<U4D_m
zx7Wq$!)C-CqBE8LyUNYi?}gG%^KUIRJo){5XZ!h%bGhAbxK1Lj??I$VVq8xT4eR_|
zb>14c@gu88j@|qA-(S+|IU?!*HZi*xSmJs3AOR09@pq)wW2BPXzhZ38%7uDsd!C9V
zQCuT1smV{Rz=~b+v{W49A7#)xf!Ifs*lh=ghynyAGTQ+xA*jB$@(lUt|M7CChXA|W
zSW1mVa)cW|dyb>0k?$Mjful_vAD-O?Waf9cT4soqR+0Xp<7U`Lhv`R-Mbl|Fsos7Q
zT%V6tE3=3ZpD|5wf9+ao^Fi0)<^{o)*kQZJK4T>~QAN+V>rq<cc`{O4%db>TRm3{J
z-p<I)9;Xt-m{&Z^iVaWO=c7K^(qk!I&6ZL!sCs67ar>$OPljcex`uld8iQ*~&ZnUq
zt3niS7}&VfNt#n1HS4dnlx08dVug3xX!q{VeW$EB*Bfy?RIreYUu)KeYw3UOR}!$A
z_Qjp-Xga{8KVOocB8P<}h5RF$?k>1_c_c!LXf7+Wy>dt=k&56VkiCsZMGq%TPop9v
z2=Ha)cP7tSp1N-0_=_QD&OkyIEQ5<}<6wLb_b&@1jwhPEMjRyhhb5)O$48>btGE~>
ziimwf4lbFqg*@`Dl#u?CkJseeI)IlIbl)Ds>AKq-C&+u1>zSu)I3qRO1nvmI%+Bsa
zw?^EPCr@Nx>cwQW-xQ;6l@L@=o;bwL%sdNYQf&)vNWz%KMd%U5C+}fM56Z8*M98b)
z5ZrT_?kL<gd0loQmPPL71*hVRCJ}5F^nxvSv%QU__uj5o7>qo47sYjr9u@DPw&o^K
z@!J0W+;zM3qpCs6vZear+Q?3l@zl$kLve=hZj7nx*lhg-jnVg_O5b0Hwe6QCxI3H<
zsD<*g_vwYUo(*AJR=1yRdQZpMGlD=V$H5DkLq!eq9u`WDn*tA6Ww2C`B|i<zhtG^G
zIqRcvoyOOSIxJq1HWD=VLQ}0^ZR8ayqJxz1<13q(ZeQVT%LZ$Ie_jf<DY)b;wu^&t
z)0zKW@*Wo=#XsRHT6{HR$y*Z2k2#rY$Rsv-t(K&bB_F1S15oFK7IxYE2mqqQapkkJ
zB$7MBEm`6@I+ek!+5*duHQ!|2g*c8I$y)7m)qeHX8(7(?uYm$|{*8M*rT)nAH!WIF
z!G21hKCjB+RTeHw&EYk%Hwet@Ue${1#fp(H{2j^n3>&m2Cn)xU^t5d+6fKCnkyEuC
z1z3=#T3SG(jt1T<U9f=pW3qSw=a_o;3-nLI0qKbpgSG5a3Z@OW&>Oy(pJl5Fn873!
z$qV!CZ3>6gPJ66}eCzZ8u^!EF2uz<P$r&1N-;0q4%8C6hs~DC7?$;L1^WK2^)DSz<
zO1XRI?L~6!13?!~s$ng;qkH?DH;CJ2s6ruPc!W%gz=C{|y53tZC|9+94plS(y+Lo^
zUS?wa5)>5F0QDs3oyJ<|FKpTWHk_7ri<nJ^94@m~q4Zp6Gk&%DV#<7==t*CmVQ;xZ
zj+6OLX0BSU{;h{eZ}RM?n#1F=EtxjU($!nmz66A$&a&!Oy}KX7^Q0w$h0hUl9cJSA
z!W1UP+KqbS{)xkby_unM-kXZ?nfqxLLuInrG&gR9kBpSs;8gFg`q$jJc~goYVk*>^
zI+EehnbsIyY51)Kl$1-DH#LtarLv|Np$P#djo~#VYZa|Dp|o=I2GO9(JV!rv6VFP-
zc|C3C0*AsT18B_BRkG4!wcVE&UUtK!dfz~lT0njpg60nx4xugLs!1lN07VXONKBYw
zr$TX@&IiqPWhBD&Y0b;;)kL&q9}$!P37gUeM<My@&%UWXBtCzOhXqzSee6Z)!EsWv
z3bpvI43*jCXAHx&#N^P?Vz>Lm<wZor(P!8@(QXq#pL7f}v&dp6!F^NJyBzeo5mpg#
zB{+BlAvC2<Z%D?}&um$YT!_=ju;lKEV+>}aHV@)|<?uQ{Re)Qt3<AhN@5X`$FvwLO
zs_<=8EPipdo<(m&V#w$0rr6q6uY2Npl}+@SMaOlIrr#HAH=UHYZ#QR#?_(T9U=5V;
zJwQO(G~*}0EuKPRV+pMi*+p`f!nn$4Jj-lDW?98VTp!L|e;}m)$8037v)ejExH`!>
zk~rC6M!&0-S|@jMG=?z{fs8K@Z0CmLTVhr@sk(l+Ann$xIZ#Y#2S#q-T)gVds$F&+
zYQpvtm2^Ta#uuP7{2M5!Z#yh$MOsClDS3<HY74={inuvL0|nBqluzj0g$0P73x*@$
z>@YvL-+0j*!#utVDq_!^>cf7+{Rm^{sc_XWPJg}}dlJiYD8j_wgD^$m)iqv=lAmCu
z=Hb1olT8!WnHQ8Y?v&C`9*+IdRR{ASq!RSZF(8{L#soX3anFkUm{jtIV2Do=KZQF)
z<-YxCRT8^p9`2`8jhBsQH;^(=HVu{(M?41LYxPDv@U%dyq-%<Zc<fxgM1FLBBO{ER
zUMT~^Oe9dHf@fK5=KXHAvrPE}ApVJBa~I55)E5OqV|A_j;O~pGZUt}g6F0`l^*q``
zX;-|m;*8Ym+g&UhzqIajKsWy3qqm9n&iggegEBDzVP79QD>eWB;VILV3%?V`3rjVg
zmia~pdAYscn^W*vE{H*{JQ^zYgzLpZm%Wk5`jl72`ieErhnoB?HETN9cEZ~PG8cy8
zrkIEE<CANxKqbzF+jP?%rg_xBzzOB#r=4W8q5B}S8HF~uE^SB1RO>*09KKW@5y|Uq
zJp1%u$YJEk=(m^C3T@7(!cg}lcUz$`Y3?!im^!W2j@d@1qG(q|+}ZA$Pm=y5_6f$@
znvNn?<Je^m%Yks2rAgq3^~d~tM>$6k`vtNjo{jojcAIbq?WgCRDXgNrpD_z6yZn9#
zqW;J0)mzw?kFum*P-MG)G6-??WYnYxFZ0q9vE-VXwZ(zr@YDk6(rZlLhRXonK@*wB
zNruL{0tSQ}<OH7MlWyr_f{Kw9xN4rQxP{H?&9e%dBldR(PjYEwTi802r5z*qf1H~Z
z-|7x{0J$#yr-v<c6r^oBCGVQnRG*B^wQ8iKokP*2C_suQ;+jEcqRKpLeYA&1seiCp
zxxK&VRASp@$9tmi8<yMfOWpgObuB>86ZhO$e+?VyZ&NkM&t73H7{*FCg{cy@TcSDa
zaZ79Q#VTP0);;^|iX9>562h@ulFo)WI4TIi&8?xktC?6zn_zW${->|V@MLmOWPUq9
z+=fslghTYimtjVBGyWW^<GvFV!i@U(0l1~9<<c-dPV<~>dOxvTeN?}@F|S;{(-tp$
z4H{TeOdj|>7Bc!tHub4FBJxeLVM~N0B!M9?2ddF45){vuLL4N6iz?v4ZI-E*KO^Ge
z*|f{Noz|wuzI)-|e{uZ&>NFp?S`HSM&4p*3DHkO&xwTi-jSlyBTB{>bSCSHKAiOMW
zzF(>(RDFqcI&6h-AX~dUZ2PysY5$VJY7dVo@}kHrj8>_&VRHG3H!jgnaUaP-lP;yN
zdUu4}_sKd8HPr@|=-68mM5&-Sr84lw6WR(0r;ItIS3wsgIklB%1Vdj!I1FW{wmJgJ
z;rSaQSYlRP2DT@Sfk#oc&aw9+9{E@Cb<Xe!OI(^<IOFLr6s1W<pA!=ah=>e|!xHl|
zVRo(WC5ngjEgs25@~U2EE-EF<US3;Ha9QH#GVPZCqSzug_Yt&ly`?s(+V38s;Cl22
z(^}Q;6d?D?_g5SzC8$!)P*D(NOT=Vq(Ci<B>3=X|w2ZKcM9KnAWXRbA|67wUCqaIU
z$6q5vueXo!&M(!fObir<Klta%gwQmTo|UKCzLsa=MSBa9?EAz%=+yoB$IN~t7t<4!
z4hI{9SufTc9i11zqY|!(^H?6k_~l0(wMYR3la5}0o}fU9i%N-;91K$^43{O5$~IFE
z7EtzfvfLSeFjy;H;)d2G-(KU2I4tU_sP1mIqFah!(ZDyq+H<$`^XU+t`;4>~&wDMn
zh%TLE)hZ1P5Zp2VXldc!R%CT%w_<oa#V$@Miz&-ht1C6buB^!MbnBdoRG=}<qClZ;
zsnz*|>Zn&)E{A)eQRSpQ#tuzZl(;hpRk7=6#RbNCgdAH@o6bODsZ$L)=El?Fx+0Bo
zRBXnH3_>nJbgIW}26eDnnP7{$(B=61m#Op6g3A?&whfo9dRGZyDq$?YBa9Zxda5=1
zTF|_o>*>#^_Ux|`!{!}He&I4KSTI;LtCcW&ik84mbZ3$X1j$YN27a^X1P>|jQp)Fh
zHaX>`IkKjna;q0A#0kh2&k@qeUT0?hX_3?0t0sB#5T<Howy}*DHFiSyt{5Wc{?yAQ
zQ)fufD%sDC+mzCityB3nNc6Dxtyjg(1?R1$MWwnzVO0qD;r(yky?bn!D(y&eWzV0j
zN=ZOpv%3Xg$FF2L(j>XaHVeF+*eT3dZ7_SNcuPx{x|yjn=tZHo>1!u18#lG)Q?s$b
zFOTTU$q@F>VOfVRHDc(a<lV>v3cnFTY#-x8e((wZk;wTafLp9fFo+LE940fYuaWR^
z>idH5xX32!Xum=2^diDZ5AX*$*;0yjCAARFCidXa<&%m&4^`*a><DER0Ho7zURZOL
zqJh1VYCdB?eUk7iaIpL@xbOTwLWzw0P#j$h1QJPRlJ;%TbSYhSH>4_TYvi!pnCnWn
zPO~*y>P!tP9o4#*&7oAMAL0xeSeG@Sodehqi=I~1Gc%Xtr!Ekx!kxjb^Qt*I+9A<4
z(?4g|r-MsPx#O-k{f`!)JM#`S=41+p`!j38K;H7wKyTxaBp6-r7A`^+98~0!K|hn*
zDTh8g*;wdBQHev#jqO0sqil~sOIs5(AKihrojCiu5n`4@Wj^u3E{wS`tg|)3R^t*j
z($5A?&0dB&kQqTF$aXI^S%%WyBhQOy>+SFv%$x2nGWQ4Up`kn(;d4u94Ml*Q<A1zl
z_K^iVBwzDJHt3jV_=}o<_5Lp~0ZMJ$ol&Q*bmi%}WXwY4HBc{Xsg>mu!Wl_B-9vs3
z^*x+T%2Wt^1wx93IRd=58E<7#l*00R2|wbSVMxuh(cX;3pRO}4I8L(@Q)XN~#>=^r
zvdAwCjw%zB<AHtkKfz29_$WzctUzp%Q+zhUv@h3T`e!u$o!_g#aXd=WmwVy>2SlNK
z`or{cf1%v$1r~~gDL~2}tn117t$Tl@s5*k4qKIae6@q9*_xKP06==m)><mDPj}?1M
z<tnS${X2&25bLq62s{9FEm(_cUIG9JG2y(@N3oM#rB%f7k!-pam-F-6hE2Z1UC)-m
zto<n&Dp?w`4H#ziytH`99E=&$)AMMM(kg@$#5J|0w4xG3b6iQPtR|$Uq_YM5%jGF?
z{^3af`ttv$B7f=0nEK)=CsitmYbn-V*YOemiBw2ew;Gm+;}7mcyU-dT{@4ao0XGPN
znYv-*!wzBeEJ9~yoI02I(jcdZbL|rw)j#5_H_ywuZeXQnzhD+J<?_u7W$1?>7*6D)
zL^!LtfNY*Jj)?3x_74=G9$2yhlHYcrmw@IzDI`)mt5f1gw4xS2Q$F6sB;@}U31;LX
zu4L@&>}T=tkRJEPh9=Su&7fVm{j?E36&*?1MOHG`VOENU(W98@7)$-ml>5_Eg753y
z7=(_1-hO+5`6swSFUYPEO?$FGjL4fAF#VMp1wsDS#SKYbs5n92{m6_<ttK)3*<+dc
zCDUFu*Z-q%gHRQ}z*UeJd@&eBjA@hv*23b+1TA9ZXZetvhFM5%qluIhrVTt}t%E)!
zCPh(?ovFxEI<em=5#uxph7!@vo`C}tBR>8VCgg;F<i}qw6hi=JoKKKn69ulxTkNAZ
zny{2(qLpLk0lOv7q)fuqFQL9So~sekF(}5L*rNI|(Glm;>v1R$Q5Rd^EN<);`jv|z
z^pjH)=`SF^ZUu9n;z*Kz8P-%2GFhz@Hh*W|V5%osaLa0Y{Nld-I#nj0Bgv77<KzSR
zLlP-nrg+lA^FsbueH_<~U;l3AH%BopOi=!q_&$LhTfM=y8DIX|uA$9f=Org)obvKy
zTIQ8sd6VbUHuV)I{p9_*7-|S8NsO}b{8pw9aC(nWcKgng<VH*TkCsX7Cl2E96i;2U
z6**{avCZjz+wAulCDlKak+Kz_d8A7^ADGbaN;p-hO0jVA-f*oBTL|gJbHcu)c}ZS%
zA83TWaBNuB5YfFB%^9rEJC(|A(~w}#{=l*OZnOb{c)xKU8=ArbVrX7@SKHwd`9#1p
z(~-Zu(e#S(dNAv-nXR+y>l1O9IY|JOBm1~p0XJa3uSVJl{3IH4`sR5uc<0{x!>yx}
zvDzj>NvG(9%HMW%&BM9us#}HZj47`qejd8p9AW19i}5;K!w<uIcTqazf8<m3T(O1;
z|DEkzpZT@AY;J`|MJ*yi*<&&M9ke-mwSjH`K{|b>oC;3T0~dqM@>0xAaoB!DdSxX_
z$ntW1s}^|}GZb?T`^f+YoWO-F<;{82d{2%vST;3Pb_}b72=jF1Q$=_~FXPo;_OvB-
zAlXJ`JV8l2iU+3_IljT3+yIb`B;iMpg?JVNxi$Z@e}mqcwbw|*xiu|z`>TZ&u?tTJ
z^VhO*8kz~w;=>>7iw(CeChn-#a#!li`k2kASDs~@YJDBnRLk9(OHRBuN>G$g^$zzM
zp8ot7Iq-`!J;XGt&erB4RwwV@JW6-+Q$2^4N_~4;X}6SFCXOb0+9~3c&rxA7U#7@D
zuT<8M@A+Z_yBy9>PLc%-YoCXN2G+uw<4zx#UKR(DJi*FQO<G2){0hsU%ejPutBL9_
zHVj2v=}zHrBvegKF*jJFDs(e$%@Mnj%+iFGGvZ|QEY}IweW*p>UWgR7=*r4yP112u
zVZsXeT38jNb-;V+Jr7l$GXTzow_Li%MZq3yLIVe5{=!tC7waiR-Gv?|%T8VW-0Th%
zF>}&p-{;0QD7e;V<W#+dqARtNwVQ{<z!@^Lnl&gyQV;nSYMIcm`53!c7sY)|HYl|v
zVpGkICHXnc-l4t^A^jVayuuO;QSlT^Dw$$Up>%dxwSQJ?_<4D@!-hPz90~Z5w=uwO
zRDZW_*k6tu<8%&ui8eOYhVV|^O(8r+LcjGr_nsf|UVjyuu#im$YXE2YSj3apt&xeY
zugcBE6<V-m;soqQnYiBE00-4O`M~iKvxmn5Wm<LP&V<>Z7xq5Xvhzi6+29d=y$Qf2
z^DO+H8-$ohkVPs=S7}JirerEVEu5m$pO3duW!LLdSP*gPVu>AoQ)@j@vfZF&H)DB0
zeZ^PqQ%6Eg+M6k_Yhe;&3(-I0@`tdDIPX1eUAcy*H$Gik7STx~gqN<0WvR@VrfI@`
znr36AXG2hy)t!I8xMkZFsFU+3sm4?`r!RDJOI_bL<USjIAsx)d?JVP668=_PO;sN`
zzgeDnKI@c-Tj9HoWStyumU+1z^hJ;O2AA{2Ov~4x=Km1Qhq&Aq^4Hnz9U?lN8HK7r
zQViWC7ief`xWa=L-2=VRf(}a`rGqGu5~p%cD{wF--I;z;0arf(i`*E_L|Ri*!}UY&
z>LlKndKr4K#Of5JXy2Q|nQ~ostYd{;y2(=$*D`QhARdVWpyG;S@AX0Z;!%ymc+cK~
z=R}Y-XP(8RWITg8=%OAXpFe+oto|&EJ=$y_T3uoaG+JVFawxhAQYgy8q=FBretgLN
z2*NEdnEx19XsGd-Yw|}UJgNmix%X~=iQa73?s~{m&j$FQzhti~jl5ShtKk$BGR@)(
zyn(<fgMSg_`-ab;?RBOj*}s%y3bzkhzwScAXn=Co9g>R>ac;i*e|QLC#6u_}9wNX~
z)EkWgv4|fF(<hTEa+%ON8?4mqq3^zO4*Em69=0qIH{48e>Ywy$NetAIT7KLbUsRAZ
z^T17{`pxj#A)VM-uI@M1W7_2=&A`zK<?_M{s`XGb(>(fm(%?{)@`&lMVmf=IjfaH6
zWt**>-K7UNb!CTxY8X&lsG{y_-u*CJyNC~I!Mri5!<<juLf-XI9pS_oS#}Se*2?mf
zsEh9|V>FQxo^P+ko^&%<F3oUsfNl?yha}2t=&u7y<BKjcJX;R??yH}tT>_33Vb_Uq
zX}g!&Lgf^FjHPR`69w9DBvOjHlfyZwDKa0>iim5>0km4yaZn1GGN20w{)QM3MdjoT
z!~9FDr9pncdDuY6)>C=xe>#f_WEHN-&%e}KG>_wji46|+gn?E}SS7YHBDOQZG}o6$
zs;K~_T#9yj<QUPP{`gC1i-wi)w}rm^6+k6roZ&YAU}^sP@>Q(ZZCupN-;vqF-+<_T
z<a}wh-(Z{eV<cR<N3S@hZkYlSfa7RK3c%U71j8&zNA=iYvH!B@(SCkueRBkh{}oQ7
zk1)`)Dy=+N?LF*vCunKD?jPj3d_1H<2jyGg5n^xvXCQzBVPVUK@)_YLzLWaa1Jd18
zl#%=*cVAx}+w>rJ#qQA6*746WZ}a<mFwW!=J3_W&X9Rxszj?2^oR-o@Rg-30=ay<h
zU0nXmGz%qYg~txAH>&C&Dutm9(hLDzwv|`Xt7+!=#+HMU=mZ`dd9kesj9HTFZlXI~
zP_ni0_w63W@#KXo++7jzcWACy^4k$uqK3xvZj|?z4ZBl?IPxS^cz&h5r$nb~U(N!G
zktWs#s1bQ{ysYaVQn!X0y!!Zi6lt>8I@Hcz0Mfp17zLxFlLc-6B_VwgNMDmLw4g4b
z9ikD*tUmtummP_18ZL@PU#`b!$5pM2NAUf33M*hCW_d=hBVOd6ua_qXF@>0l<(Yf2
z{2uCSxV1Gk!GaB(jbS1zd7?yqn528kXiDJ?UT62!fR2@DQb2{z@X!9<nuJoUl-o+!
zab9B!L6|{^iQ9R7<^hc5I0?>(tH%sX@3zVBTf|H%KWFUB8n>a|!^479Wk#sVapu^p
z5BaWbT|emo9Nhw5GdS(vNa=#f3!sh+wkwNJLW%P=ryHGLnln2XU@U$vpWOzJ#gIqZ
zyQUcaxK$=M3G382Z7N}y=awbS{_F?qeLm65N92_z#8$Vk8t7?9*Eh=f91qV(XjQ!s
zR8i+u@3;-|pY!cxqc>nbzQqsB6i<yMsKl!i9MbA1O`jClMemcfr(FOjxfI+~+knew
zbHI97;d<Ufr!)XC8AnjKjR;jViKMmB-R^=fTIU+2S%)Rcsd)K_D_Fc<&<goSf%2QN
zt3pO0Nu+JC#J|O1iK%6(E$j^2(|bWbzubuJ_I;BihZif=pdLe>hcOhXpccBu0>R!;
zHq_QWVyVd|&)C{o*Z+mbLI=x&u*1c@!_t&@ZY$9jZiNXa(#RXjPL+trhg!u@@jP?a
zd&LlCsi|VmfR;1cJ1PwnZPQU+VspY{*wG>hb7noS^Yg>WruVPa<L(dhX8)+CHti3y
zH%<XFSd8<6({MLdOn+&K`3rX$#*DF=Ni`b|)akc+W;^_e`Ky^g5kkMLClDZMmuyrW
z?WJoXPE6kU)zPI^CK8jaL=ZnUc#X&G8z9B3w&l4+I!L2s8!Tp{+<<~J<nu9B{k<*l
zN4NhGcU!|z=c_a48F%j1ECUjzC_Fd8s@EKBq+%Vn=TTwf6I2ro$<<?>O*VrXG4J6)
zUj_Z;8ThwNd<c+Qo|rz6V~`}5W+x^lu0RfH0P2b91f+fkgJ8rXS+&C~P-A0bek&rF
zrs!Kjd)OJ%k_<woQkM%(o!THR{dl_m+r4j~@iC6T#>EZ<_ka^94BftjU+G6p$J7IM
zx;!RsJ8a>~4_p6!64|w+e10P9?dcV*6s}91H#s)0b@r4JTdW!0^W~>7rS^zew>R}O
zD_P8uBg=IT>a_UA@t~|+wqb}@7X9e3xZiGRx_XAs<vp}0W%DMfJ-R@MTD90v5>hql
zDv_{_5cNP%;zVSEY+K-{iZk`K-7ePG=bZ9s2B&LnJ7f2PC=Gu1lf4^p{56#|1tP|n
zTQxJyUdHTCI;nkgi*D^;IX`B47nx^heM>~Q)P-ufZRD+8!=crcgyOZa=N19+ik52}
zL%P{zU**hy4!Fsd>K?ga8lUKlWawZoKW2Hy{G8@-L!+Ch!`8iZwXe#L%o{&uhwIb?
zpP>t3EZwaRXg)=*La}>ITxEX@<97e4nG@cZDBhSBoO6xw#PQszzDV4;8+B|*&o7dU
z^*w&op$<81OK+iRWQV$X*=e~{L+MDFdL%{z1-$sW<c3fJpA{BJ2O40|RNU}oC=!YT
z53E#X`&_al0woIF(Ln?^r1W4cz;x-S>p~3;4HGbJ@6l6ZW9_nJPKQr-eTcu8K5YtZ
z_JB#)>ahZL=U6lfZhI6h$c>dk)v#H3ER)mmY&FH|4C0sm<8>96Ox=-(;)(4Mt7D<j
zX~Ef@92?ShyGAu=O^GtSbmOHSJ$~A96KH5Ro1Y*m6WZf?PzuIOxe3NJiC4QHe6^mE
zEk8%Pqh2OV%J2{yx_x6+mrv}_+yVL|<YvGG8;6UXxG1yKSRY<+oEakT{BNO&9YG^@
z2paM66!kF<xlkNi_69F<={N0=Y;jL+j<Q&(kh*fRN>+{e@>RA(k2rmMZSUmH`Tj)o
zubtG^GJ#6_Te>I20h&68D>T!P1dY5nk#Kl*g5yWLY;MFgK7qt{1Ozr6)3tAD)3<g@
zo-o*E?T6O3^=9B51@IbWI+SKYO>xx<Q)A~^;7VE)Up8rdKjamCQJVK_ADfKkL|n)z
zckLS}ip_a3Mj}GmEi*gB_VMxU5!D&23p8`p)wCL5GX2@ZqDdM1qnFEX1Imdkyoq-|
zs;}sbPILSNyM<O4yf9xlO3rmt8XS7Ioz~B^Hzj?y!k-g>3h0G3Xtl-863Hw8S6KM{
zrqDn{Vq*bJY{OH@(FuUDMg|AYdq*RDP>z5pEZ0CZllviC)iDsjUiZ%7-f~v=7p+nb
zjY5;tJ3p6p&<WBoX>MVY<NjI6F?9jK7Ic2frPi9e0=(z*VHDhH<f;}<l#mzHJu<2W
z$VJ@;jce8dDFl^L6)5uGuzcRAa9Rsy)2+gHM^Ci>QRgN^hkv2Q2D$^Hgq=^kt=O*U
zvYMb2s`FMCHE7;fj@KV`g^JVl>;E#KSL1qb5Cg);LV(OyjM<*yOR;VjeOE?aAGDYX
zF;J*q_e%wV`d$HQmd8Di?8!Kis`v?33sK-=6P@&cC@$FIl3+V%*B#vz5oaqAw`-(4
z*%qd9c|O!Fwf*_#&D2J;ZJOV%&jQt8O6s-z#WM+9j9#qa=V&0-l~)$Zej6s1@^C6X
zDY2VdDY2_fzD8R1b7yEVFLpTP>W7N`iga66LX*;V(Q>T;kS*k`1#(8VyR_Ti=csrt
zt{NBY5Lf!DY~<PAw~?3~j4ewd`q8$FvsBAxE4u`$D#ud@Tvfi7=N`EJWyL(BIc@*0
zeb;V)h})G_{Hv-xzg{O}EW+=mZ|yDa{e_zd+<ta(sN~3{qCsL~%_t1oCZ8}Q)4K~!
zeR+wQ;iCOmUmkPv-rCYoIS5km76wZ@_a8#_H=Gr9g;PURO(F-w%wY$-K&y48P@NOB
zv(9v7kUmV3`80b87pXE@&AvH+D|n)I9%rf%fI?HHjNi1>l|cv;poA=%MRoDQN5Uod
ztQl7;>Kgt3Sw}=*C&*)wUCxiq(#(*W0k!X>pnF75XrX4i$2ze?5XuR*^bVpgb_nGz
zk)iEav^f&w9QJ#a0mm3#Gr){T*KoQYY^YY+O>C?{H7@XX*6E4y@n9%Pwc6aEp?Pt1
z=+-)Mz#k0x>2n)4Fd03h+C!FKA4PNxJc5GRl_@7g(!Y%qLT5?(QT=I}YQ*VnPiqg*
zIf?zfQwn32^8*rvob#_mQHgzFg%)%Q*sZf!8PuxvT=r{671r0d8**MS*t^J)Wo<oV
zm4Bv#tKK2e`(xbix4Fj8zFiF;dDWUV;KzO$eCNp&+zeWBwM|z#QF}Gw?i_J9)n?0G
zkCAOs3M2EnVVfSU?D;I$MydGB%NKUm*WVR21qU{C@C!eWvNM*xg1%?oxX-wgXYjD4
zxiOl$Q2Sz?h@<zX410rq`-Z$cnT?F$FF}zuQvI!wIQ(yB2A9zktn+f_-^N+(l!QGL
zl&@&kn^sm2_QxWTUgmluuJ>IKRD}@uph1(Y4>kMVV#@@Y+p1|wC&3EDVTe4pa`@Sd
zi&wdB2Z7&5^d1PQnX%*Sjcx$pENBq67xGa*O$*EiNI+#-*wL=uVz)djEEyOtV3!Vp
z=ds#wJ9D-jE%~y0{$yrlMHWNp>UCX$BPxvc8`0_A8+hb=o;^A)7wHvZl>({nu6&mC
z56f=W$QkiCqJV;)U9@87iGdIHb!zocn79VlxTxrJXa@?s67=7Kq5>k8t*NbjP~m6=
zZUzVHO5)v_n|SX!oGt!JV*iQ0&+1@(rx8+RR43D2Uw&oGVfgh`CFrpVmbTVMzQ5un
zdKU{|_%b9j*=L`-#MR{UD3Bkk7Nm|f+#p$`8VX;)EZq2dH8L5KNkHrijI}qSz@?6H
zrW_PQ+H1s`^ls7O*I@5c3pM-@ttU0zdb05PltqaBP0vHS#Y-CUdl%_*hOJ6A1KsZv
zgm>+F>0QZ87A(m3lvH6pa^)Fx**>efG^{$=?_u#ne@{2^>*kE8J5fAEQ(*_UQ&Uwn
z5{fLWieo2buJ$lK6#)6!`3*e5z{KV8XT{?`Bk3rn9nu3Tq8AM#GLGs$p<iXG*-XZ^
z6bgrgnUiik&C+X3jqQ2Qsj(Wm-#V*yekVVvxTQU@0XHgd<JOt1-YE``wljQhm8Ua>
zoiDMZWZCc?1a$1II8EfqTzkUyJE6=T$Ffarx!a5Usdz|kwW;#yzRuK#L%Q<^Qv~D+
zfa<^_)8U_k0K+Xs94~2~ty^7H`raBe4PlpBp&r!z_`Zf9);vM{@^GcU!-$)zU#+d%
zMwjZ_u*?lYh9cN<m7!Fka?CzR9@Lv!X>wq+5e`tR-QM@h?1Lx-NkGQK9ppg!HRj9=
zL3~#?$j8b;(XsqCCV>xVH$mkR)i5udni9bG=1<<{`m?jx{upMl9{Ga{JJy;<f{ji2
z=m$A#Cj^I!^e3lGc~=h8n%)dMV28_HzLQe~`?JNLBp*V)y3D#YK{>kM!=zSfg=3VX
zh6qd*Y;IW<xEGQ%UuN+%Q|F*IlsE?k+egzv3NlL5M{X<Uik3v)^g+K?N7psV9tntJ
z&)Pp;z51BJtD@9e%-yn|6BO!N)pj@oo0RqIKTNM{O`81G>67}7TYJFZQlk0MgY=c{
z(3rYkNk)>78MLw4rkzNjSBP5j?c_A4j2I|cm`Y2mCwML$I9)xjVp-pIAkCd`LsHF5
zQwl9<mmlAHaF##bPeS7DWOM(+Q%A$)IlsTnwU)<rd;V6;P&N5jC$|$!t$o$|=%|*B
zO2pY4=|;{z)qv@Id-21>rTVBYLc%xuvg8kMoT~ni;!V3{*kH7wlq*5~v?+|T>HWc0
zbP)ruOHgcG$4HU>Vv+C%Q`s~9(=!iTg>IrWeXKZ1=QX6~a_ka6q&-eQYi-!_c`!0%
zccp9ZwT34&>?H>rvyR_xL%e>3bZ{mPg#6V<5>b;cr9=+WaQr+ZmYdBvz3?J{5|J03
z_ZR>eLX0maGqojK<6Khh+^$XlW}I{LZW0OguMmvu=)N6SY(al=)IvrcCgQdRjjGY(
z)&%ic{Wy9K7LL?Fr7~erS;tpF5z95PqKEUn(@j%p3Wsi(v&(FL_4~r9v8>DTj#W;}
zCwG~L1C#eqB^2XG0>KY-!H&knd%xjk-&E7n;v&MriH4tt)79{)c$s}3H`Ok#qQ5Aa
zY5a;NUQR#1+tbQ4*4nYNSa#cmaV|zb25`iFj+#yvD@tQ{WU%TvUkGBIucE}#K+i>d
z<$+_>i!rQW*R}g*{RQ{F&{S%_H20YWpv4es9>J`hnf@>cli7{C4SJp*o+meT_a8GI
zY;QYWpb#MDXjqk9s9A#~puNP(U@b&00-AK!rl~=f{$ZFpN!%91e^%==oosr}z#Cmp
z@-MOd6wx-GGQWlrRi!nZ?Om7)UFrgvN-kmF-qtdb3|x0w1ubgRAH((ag0moM*ZT7K
zNrCCZ2IjJKLzj{E_l{ZU7hQL6sobV=F1Cc^QkpvYwMS5h<dUf4n~j03@wH`iDT|1M
z`N}y^Vkw4rYisAhEd>MYtL6M@+MkkCy^R%1{3iQy9@+f9rF!A{qYTmQ-u7q-EMfaO
zOILn3YS3}c%+1*Ar{A*`#h`U(2FC2Tvb5U`FVKi4RKB>9pcw6IH?Z|#i#XaIBbfD|
z=-J8qq`6zyciIkt{0<{82}M)jqND=*RtI1H%&DWUS!-N&?+71xZ_K8*8_KO#qA^gr
zDttNKr`^}GP@?r5U6?QPv(hAsZpq~zGtbWtx{tglz=IC8E2GECk=saX8A8A8%~U1*
z*I1y3!~#9sCw#Q{uaH<^`BxERjzC-bZU)%u0|)(CA>rq1Pu{6zEk7@!q^Q%+<2?;S
zL?)k6NY~`4eCmzHxB@Tn`_MUAdVD7-BD_64tTO5ow+MeR@7h>vp=*ey;@a$eolXta
z+M~C^mJaf>N{5|P^Z+tca<0E9KpjX2v74A*sJDE4RP!xmIOp@THkop=bz3|K`xC|#
zNre^FPNy%Cg1i7+_6>YSaM>*2vZ+Mf-#w{)%}|pz%cxnzZ1gkgCDfAzvQj{lkvzY;
z_1k(};eWIMBOaEN^bhkg!`F|>Uq$JBy&MmXK|!E38=H#MEgHDN%Bn!1M}PmrQ}-<R
zk-=?MAR)E6+UbRZciVm6(RH<nemc7=1()VdOAhV5P_-;8fNE3Z03&&hTp_yy<<#et
z@_Wm!1KGV$uYY6$qio32_nvEeKx!!KYdjllUY6vNcXqMJxM*kwpax7+W|Gh=u6Rq#
z0MAVFQS2pFdTw8LLE?VQO*wbC8tVqk-><KTm)Y!4F9W(<A=#L}2Zzel)3W-YTS&D`
zRgNN@QCap;_J4^UD+JIx<Fr^b^h~Gv>ZHr7!0heqhNvP(ZYf1=J>Dcm-=1G7RxFV<
znTdrt!zBThMr=P6nYv9`X=hIeT8s*&J=`_SP-x61!GG?s^+_~WUz|(CF(di^BkV1p
zs!YGPVMSChKt&XkQdAm5q#FfPQo0T(DJ|V4A_gHM4F?hFl8!?egw#P&;t<l^UEhAx
zncq0$|9jVBAu=P}vG1MNzP2PKT5zzV$nc=J{h7hm;Hg!dnl!rNl;UWXb%|f3fWX7)
z>$YYp)TGZVh8O!cn9qh~_EP&t1vFxa-au<PX&0}Y?ATtp39pX+%*a$>eE(U+Q>cc-
z0|)skPSt$mte|xX^-ETJ`#y8~j;d`?g0EckNX~V^lW4*F!?(&E$lt%%58%biwHrTT
zCo;$@-cWOty>-h|U`1>9{5TUc^D#(Q`)+e*Ye6P4>h|BwMt)E$%3ZMeOmM>&;qH*l
zbW%8V_2Pb=qu7>%cAKku<7-PqfdO;$vMge6XusaYJJq$2g!0h+#^xe$$^yq5{WU?Y
zp2LZ;7CabL>fQ5}4b8h33eTC!4Fz-lK){a3niogvUD6Jqy7h4GzJQQ+{`r1Omu~jv
z_TuQWzBK2PuJTDd3Rk2|xXor~gB^|f&R}9yJ=hPIo2!$nw{{6T?bKF9CG!}j&|P4#
ziLs0ASlk`8+Nbp+=6K{MLT~2~+D3NQF|??SZ!44)bq<8<CO~DBFSl_RHQfVLB_w_`
zizlPMBn6aktxdEZW>9$dKCKtPk#WEPs3L6Ho*L4U&kNXm`=Nw18(Z7!<?G*%rfU@|
ztjL~-Q~lhHj3h0~RHnDLb(x4^!}<3+4jnh^FGr~>Jk|-qj2M~3rkSAj0NC-Mr+}p+
zL(?dPB&Y(MXglR5XhQgalmt<smc2!B&Nqz~3(ZxHgGR)q*|dt@e{!JBXKlQu_HGI2
zafy+HNW0dCf|j(j-^H|^(GF#Y2b#rNXYAW1_-p%w$XWlC7_8hQAviNwY#T)HM>$^*
z!-M%r$YKyCrzyJ7fm0j!LZW{wtlD@Mt5g0&o0d+rL_kv;;ek*!4Rz@1>QugeyR*ai
zg^J$C^jOq!<x3;L*D-O%PUdKR#ZgLLr!!dLsN*RL65(q$pQR5u2KSr?$73r7x^vFC
zp<*b#imi999CGE9Ir)+jRj^PWjj(_$ge^|Ix-T<Dy}fk39G!m}s8?*Hx<lt?UnC0-
zO4e+EJK_^~a*@eO3*)3Y;W|4R7h=LmvbWA2niVLq_jU8#^8YH_)5}7uXJ}J~?)GDA
za#r?sqKc9L11CplErcA#kA_3Q6(3Nvj>~0&8rEneQp2)wfM{U|2ocP-DcC<YUx3^|
zT(BwnHNM|1sAMd8uh*=b?Cf`4=n)DG5Kg)w)5fO1BW|xQFLhYQ=xsbhGL$Q_x8~Bq
zw`y)PKi{a2SId{>DmiWHL?p*g&wt7+;l*4-OwCu9AR}ree+vyFN~Y5y;WBGwzJVlH
z-9PPj-8A*9&vQh22I>t4ERM}zkIhNoZW@=Shiv+p7$ZxG<%xQW=w|mj!Ec@Bm2;2O
zhSzDo?6w;{GU=|jXS-a;c;<7>aj6T+iw`je%Pvetok8^`5Sf&Qw^|S%!D`<YL^S;^
zeSEPLOzyj;<d3<oRu4ZuCVXyY;vz0CE-FOLVYeP4+o)M!=4DG#PzGX8OfB>oXv&ib
zYC^-(+1}=9T@>!j&{W;%AH&8<Msb<H2Z&JG!o|Z#e*oka6>r|3kT3JtbB8Pf$;DGq
ztWGleGs3&;2QK2(U&hO~^^gsz+dTSQ!$()LFt%^QWU(+E1#uhX77s!xTk*4n;+kp3
zloXXX*DTZ;dz?;x6$HRKltP1@{Gt6I9?GhzHT}6p#OF9o^bMK4AG>AiR-Mk);&ZB3
zZ%L3c;?(JQ9nciY%W*teBcG8`IrUvSEh<KQAauK1`<f1)xTJ&xpR3Wk`RLl+<^ijp
zxrg6Ai};AVC%^*yQUWM;f*3z;=4JpRLu%dgf>S9YUlW#Qp{?nv!^k1#{uBOT@r&O>
zOh)9B6d25Q^@N|uh8v4d4$Yc>O}%>@LsY*wcRMT-_qjS|>%H?5>&-d-w+(5k#n&fQ
zSbZ21v?YL@WhS|QV~wgjx?mr44m{Kdj^BErkNRMt6Lei%U%SNOe1`=<;h>W8P3nPG
z8lFigQviV1`B18X^V6(_(B$<UbH8kw@1o|lp7OP;1=G9)L6sw^+681c&9c<Xc@{_O
zEa=|X_4@O##?_yg@}SN4`^q&(NGL}`9TBCk#%bSDRxIv4nXIvCfSK)I#BKFB>{W3D
zkvI1Amb#sta`=3&t7PU*(@6RXW0+VjU&w-40kgf8RHnuXrLj<wx%DG<!}#IO(1euo
zVRx-!W9TN>t9SM2%+fg2b<HTR9*dA&aJXGC+%4@GbQ_ZO9qXJ@{@{z!NGU&j>-bLj
z)|<t?3l+muFJ07hUvG)9$kqo8p?B{?3iq!@b1!HaihxJS$cAs90Dyx_kRfLFm<0k8
zK*2ywH<ng{OanYD-n_-e)>ocE1W@MaxOmBQaPYQdH$au{I~6+9qnUC;Gal~U7m!6^
zydh8jp;R#G;r1en%+fN^;*+bX4swD+I&QqgCZx5`tg%K_OF_m6)QB5p>P|T#7x;3T
z+1~Q~a`m}Hva;e(bTsezsYcg2p*|kH7q8d`>i0r?d*chjdL2e2FzziF6`{-42GxFC
z?F(90wvxW>{V&wd*07K-J2if{jNlbrex{qHT_(3V*J}Qe2dZ3mHWs}`AbV<R&l>>w
z7$t(o8qjT(cGt(~nan$li2xu%q@H3GJ*kMe<^Tumwv#K+G6OH_`Df-+Y27P;{}(N}
z8kGsz5)=T%HQH^S-DE-ei<5<Bez`bO2hlVOm=VdK=PYa)V%hOz%S{ef!lPEzfZp4r
zGc#jb1-qG@oxKi7+tb;abM4uBk_l;R;3K`B#U)DdiAaq-$e)7F<vUGhShbQjodFpB
zx&#|U>x-Le34#$l{VI>`%EH=$YMTm*^H9DYbG37zAsTTce*gF&Nu&1Ih|^2`X4O~4
zv-di3`ll;_??e0-nbTlte44SSWQ#7ae5mD(I{2Z>E8}V-p?%KJHH(%pgwc$SB^F~#
zS|+^Hkr|wq=9bl85^sBS*SN~LUDu`{$@mdJ{Avr6S4Fl}s1gf)p$+?j8-|Hwn7jv1
z3pF{#y4yO4bzWOE3JPyXoK$aTHF}F-vbW?)eYVD5W>J`AOnv{o-P#4$>lP&!BYLz0
z_^kI_=7c1qqy4K)3ok{6`-khOu3X+Ahsx9S6y^P2K4wtQyNzyNwmr{B1x<XNuW3~&
z(FtjrKYCIRPF*>qaAPW7@2cx}=Js>18Y!z7_E^i!=Pu;!J<O}|X#Q8T1eDTge9j(;
zg06qFP(Wpc3OfLl=&BHe7}eO5w8vPW4V6M;W7POI)M&*-9nbVp)}5_vhKiApfi-G9
z;lKjGpQZ)e$m@xs7fjZSED7REh=%UHfBHr|w$bl<@x>rvDd*wFXxYKn0C=k(VGu~8
zJ)&>16C}r*57y<uarM2xTvc}St-E2I$wnd--Bt5qhEo<^*ZPChrcLkmpT^J<okp+d
zhv^}CV=s|paGRo`Z8~)Q6kwxa(u}yXkmLND#B1RATzhhi{1Iqp8tD-)zR(v%uTl>m
zBlAQeO5>KKy$U<X-~q<^s!FET#5FX`0!F5?JojX}m~KvIa55J#$Nta<r;(8$Nw0GB
zW6TD(MXx-cwg+I$Q^#^%SFB{R4H=Z*>dy&GlPWRsfGH3!o9bSD0&cf>tD)*q<By7b
zHc~Z)`BIM;k-Sp=(Og?c?Ysv;#hUpu+i9v<x`4sCQ>vj$8iAV3fDvoAf<Bg<f2Cn|
z8jUiD%|a@j|N0P_Cm}dASZrC-*_^HJhMC29#9*Fg-l+b*EF3m{YVejT&So8-kBDMM
zp@@@(XDu!8b>^qgFjs1qU+iOvc=A(i%!_Yr-;hqI&nUdmDf#HmZYPBBN==ksln$8f
zC(W-56HyCqBzdwQDs@wXS8~RN2g^cO8x!j()ot!|ayJ*;Hh-+?D(f-lP9Y&3IrGvh
zzSMLkb!Ugp?KKgmAnbF6;3z>KIai<Y%&o9KHsSH-=@GH^k3%zCS_Ic~zu0h&X&gGo
zX-GS&8u&8HUbwnOxH_*YE#xr$gJNnFA=X5!2u*hR$!?718&oTw%AAWv$z4+?MPO<3
zIaPxkhvxz`XF)NpG*$Z2Nx*!WUHpTH3L+xv%sck%b|nG6n7HGa4)3Y3&-}_qFNk^9
zPSJB`6H^V=CTObJ4_r5#T_?G?UE<<&gRH&8L8;?d@`A4HX*aG@P_g%=m{7a`7=xRa
zj!G(o5WD4PN)4*Vyw0*YKkizWQf5{CRi-{p&j*hn*W5fke_z@4rn{7EvkH>MmS$`j
zc3GSPk};W3cD`I|1==T0gyuAKSZe#T3PhQ(0X!0*GYGd8bntos-P#bqX6ou8ttbHn
zPWenN-s{FK_z*GPERx|SG~#4dSOKC27?mI_!ndtqZyB0dMnWh1`01C>u+O<RW3ouM
z3IIWUZZ#*eemy1m?9^qbj2w9C8pDcf^^L0yTff;B($oqX^epjtt_DWKGO_HLjoz3&
z;CtJKQCi{LmS}4s5<T^Pg<X4x9Rv(e^_*0#&?4rQ9Es_d`_KS%x6Ik@=`{1H?1Xgp
z&1K^}llD61t5N}%B__gW#!p#o)KGgUJzj(mR^4s^nqp@jl-oWJ<?*S>)~k745-epB
zhqQ&lxS2>U^XEAR^?WlIEighpWAf`HK1P@RgGnK<!voe1@!pxvY+mg&)q^J)(y518
z9(@m#^7xb#JeHg&FXI|Ach^JEg`qqOBc|Bh@0x!_XuKB>aq$89A@Je3|JU)Y4Ts2%
zm!j4Ib5F_+2i<4M`Ugiry8Et){~}p;8B0r&jQMUcL1^ol%<8b#vnJi#I<wL1&hie!
zSiQY2wDzDnTcM#Hqux9Qs=Mv2hxh%HW37tK`#EWYFR+MtN2_yaTKxjFrt{7%&PAtr
z^!uP=rEQA~Z9`|f5>rrRymcYnkA5+KWN2&T*fVVgs?0HFs-!_O=)neQGf*<U>1no#
zq|&EZ8(NmziqntVR^jEdUStMOC#B?xc|i1!SJFo&<CW50O?@~pUO6@h`r5np9l=r!
z4(TKH1p=}k`Je+TuFu@`AUtloioo2*W<d<Ak>37wUfsuW2+7I~wlM*ALNYJUEFU9F
zLYDx(j7d`B!|hC7g^Z*(O9<Rz(CyZC;I7!C%N`l`1l%8q+j~qT2(5m1jkc5ltbn8&
zLzWCApS)%mIGUuy<Y762{pwWqq2;`3`1hqzjYLlW*K*IkoHw#nI_o@3Yw2gFR1?rg
zo!?Q`@r@o{M%9m(A&tBYU+Z!w`+pZl{zfXzlkSfv#I~Tg87*0r81It#ERIQZx1}fr
zx-Yl&BMdyL(477#lHPHnTc%XlB<G@}NpUri>-j7tCf`eEsF`E`$8(X-e4i@*3OoNK
zBY&x?OP-lL^k3(4khk8Y!D(=c1$M|x^XW~IqF1cbl;iNvb*6ic#Wn_{2a&(LcJbh)
z-O)M!<|^q!Y*E)KU=8e8n|5_omN{qrDIj&Y_t}N=<p(jYutx6LUf;q)ZV3##a$B}y
zOPk6W)sqGm!{Oe%6^ZYX1sewXUDnujznd{@W2#uO&ftj{Ys(3-aI3oX7yXZ}?|OTD
zHx1zJ<A;85kP^`Uyj2a)XHGWAt+M{%R*@dw8ys^2oP_wVL;wZ$u|#Q>c|BR1P^mWR
zIG}-?{Pw<(3{?6|kuB6yr3<@jFFE&}kZ9*h1Ktg4AUDy)N?a6ZSX$!cf2pz%{Bg<S
z^#52+x`q8XbcLWfG+E~ecCY=Qk5Qv#6vZt_ioeAd%g5L}O^m8R>Mei%h($nV(=P_t
zP-R|$)iQ-serCmCk?YRCX!k~6s>qN(OCGaP_t(icrgQeU#WARv2Ok_d6#rO8Rh{!>
zhst6|W{ZtVdQDCLb~3L$Iem;_Qo!_!Ufvvy##W~O$&o~+TW7sjo+jqJ*671RTV@lr
zzZHFr<`WuJQN{KfrX0hj2fNv~I{u=7n|J`!rZ2e9wdbAcSh>dS+z=`9fNa=*dsfg@
ztWcT_g+A*~ZPLCU*X5Q-*Yp=nU`!v*<ra%<*-CQg6H)PtLGJx?=dOLe@YB7N_(45t
zQ#rK&?H^B{Jre~Q<{+S{_pBMp*ar+OFW$LK$vfHyo;N~@bVP8SNH~AJhyWUTesEZc
z=+c~c{xO%q@~%3^04~)>eyA$CHb^MHhVrHO?vCNo1wRdpoT9Pg4YGXPc&lpt8=0=N
zL081D$*}XBU*;JkWbG^Vok?Q7;GnPZx^c-5_wWdI62B}&Gk#Q2cE+DW0cgD3VIAA{
zRR<x^JjJ`;-a&SlaCG6fsPjiY>eqKF1MwfpIEKz9>d0z(`j|%7JDYr;>JU8q6^I`q
z$5sAUwY(Y_=rVtT#Bj%tVM*0k<CFc%NCH!)g47w8Uyj0JgNdjN)h-oqPuyR5rM;m-
z%w1T?6Y);+dWb>WV)@UKXGCH%bg`Fh_nUSeLuqo|%$a|<gm#WNFK8%pmGju*^@>@c
zV<KXTB*B#JC1{>@s@L!JXH=G{1;r1X+`3@0HorMuHqNIXtvU7OlD(t-c}lt;Y#Sng
zu7G0B8lE5U835e<;r6-}@HQPfdY1JjTEJ-@iWnEJU2B<JSU0(G1Y3SF%^Y<E+s$Kj
z8n7wjNc&4k@z@uCIe1w}ly?j7*)tLPqu5X<<{&1HKIUL*(10}%so`e*T=%}}N%y)E
z&~{GSsb7JI#ljuYmU9i-JUTPJB}m!z#ye`2yHHEA;XWsloPEDI&|d_=#IBEbW8@uG
z<Z`*KO^1JY7C=3-*A5B(Fv~Ec9TaM}-*CK-gR${{a0-9=h@Qkq9gzZ(PxiCIPqKxX
ztAWx7^-7*Z=WKePAG90W@TwkyVBzE^=R9v)HBQt;O=2>R<LB)^KY+`U2lm#lSvEHB
zp5MCfK2hhdIlfnQ+6-rNYolvwCG{SZFoGCITHV-mFwfQ`Wh8)Usbr0^IiIwfGs7F;
zp^Y{xb%oAF6GY91AM5Wr-OZVQL9d#(@5o5KH!~>C>&y)TcS?F>f6@`E;O>CJ9G2%C
z<XzRKc;YTdBGqZ$YAaw#fRzYTCm{$C?Q6=SBfcPy;g&@U20F(a{w5Aoima$hNAwPt
z`a@mQI8$Af87_aoqUJ%XHh9Z-$EJ`6cu)EV2DL}oP)zL6MOsKyo&@-M^_Sqxy|=^3
zr3M+`G7W&AB|GN{eG3YXT$L_2wX`S{5~WjkUp;+F0NuwrOO75yP(qRG?uBX-4H-xd
zQFS`|sQi!I+bd&qFuotsSKvnVj^BFe726vXCE3@mtVb^@O42~HQX?V12h*9khMq_L
za@w|hHRnGjZybo+yu2XWF7}kbWAj5se4}_>x$77ef(}(wog08q=l#g=<=ozMuE38=
zkaJ60C0-P4oVn*crfBlHm80$)0Q-L=<h+GOBQ51A51e&KiMRDEEYqbH_~=`2ue+K_
z!lz;*@cLt}>Q!rKK=itK#}gMrVyx@L{C}z9Qv@gW(;mfM<b5%X*=mkGF=0Qp8z0Hw
z#3+z<I}LEzB6HaCDsU}K13p>b+hH`g`n`;(_9a*TMoq!BjPCM)OVTW&%l}|{Atuh7
zXKpba+;);*uNn&iT3LZ87LFW!>8#q3P^;7!Eajkj;Cu7agI(k7{Aq$7qpDec7UYH9
zWefk{&r72<Aw~a#vsc4HuCT^&5lsV;sQr1te?Db1R*|4!ZnJNm!0w?3<wSQD=_6S8
zs+orz5Tn)}i2q@M{s$`VMJQNjbBCniJlvpsiy0*i;XI)6d>?8+ec0tnnN*5e$z`^#
zwX=MCd0i9V;n^A-cs4aPN*ct1V)Wet!s}0)dSlsFN7)dl%+!q@Kj8Y6D_^-(_~E4-
z1F>Aw14cGU)7q(A`<-X#qj`dyS`MP7(F1H3Fn}jsFSSyCC%drx;_RvKzqr9zsnKs1
zc#vu`Ac-X5vb<RhjsYv=KNaskzCI?0z}9W`glu+m42kyoUVIuPNwS(5r8Mg;mRl0&
z@hU%o77RI-Uz>+e_Y_8Qiy&$<NT5qKLovU?$7sU72iUpWn02&*qENRRn_kHdcA*18
zORVi^CuZ)L7i-fjnf6@3wezy*sA9M&>=Zje4l+{cu4L#5x~=2+7-hUeazTKf2d+f`
zoFrTX!dVzKrshY+jt2qric1JEK=E6*A7>CaPfP6g2>E~i*M0La7K$+#Yyby@>&pmB
za-b@}#lGfSi{{Zs5M$rnes+oIobW%gQ&vS@;9#n|PDUv&ROmV>>p4Zw<0FVg;?P;!
z?hjq02Lahng7&@;tH9Qm+^^mKV!)fPD(G5-n~`-&4+3jcjudip`#ZyRJhxkt%H*oe
znANg&9*${?U-$N47n5T#&_0Wt0IK3K8d<6vf22_SAxGYb-;$oBn8AJWVA}jP^w2}C
zEtT~7JyLXno=rbdR=n}d^IK8f3!myJM^M9~K-ZEns^D0VL`px{MX+IaiQd$SjnHtg
zE7<1fd7u(ueEhSQ&+!Mypc-n62<Hj(rnAuG<o2F{RN}%Y_~>L#_f8{9?H}E3Mo0md
z!#!FC@6sE8Krfp|n@K)f-~I)>Q~tUx+k0afMd2^b?!ntzdH&;rlwS3#Wp<pyb>Y<(
z)Iry{q0@j%*sQXGn5x*Jc`tUwbnOX!@EAFUQy;e$c2jQ`tti{+(8i!T<*NF@#XEt!
z<7(B!tj32|a=b|xqr3)#zw`WzKB(t`TBXmP>xGuwDU;K2Xme8M!p)C4U60D?VA_&1
zQ76y1Gm@+M0+iDH52ZwN8ThnCBYB>{;(t%Y{r%TN5Spk9ikeVb3Z4u!ILo3t`8Q9e
z+oq?!-0E>9-!Vuu>pEiX>A@P@rV!-B9oiXsjNiMnMzzc5aB}MH8bVO#`gM*)iC()_
z;1ww9GLsZ0Rm&UB`~v4MSHyKAxbd{-qd)A$%b%pVenT~Z@lYcE2ILbWsilz2M0gRq
z3oQdrGfY<ukiC3U`ra5o%@D~FVM2f+;%=r^@t)qC?0|Xts&IgZRl}0fJB5q4_0dvK
zUYXvy{3=DkO0oKuk{IJ9hGGU=AAQEl44o&)>rY=u5Iud(h-Zv!FDrj5c4pdUb$)tc
z+GuXaNjqz+WQ}i1-BfU{-<fE{b&U<5>9P9}JZv1dfdu{uk;+UweDsV4`Kk}_xs(Xl
zoI4>xLyJ!!a5z(IbrCOeSDqH7dP%3`c2ji4n{25vhl51o_u4dldn;rwrEWEZXl#G#
ztE5&?n0UK!J-2>_ab%P?<4`D}z@8$W&e%nbNn`Zo(_@`IU3`w$ZVMV_w4Lo)!?$zy
zF7A3fbHzTHZE62uV>TvKZ${?K#DRkZ0-B`y6nF%m-|oZ3=w+u0kl+!}c<%f6r=YVm
z^V8*rX7S#*tWhfjo~<#im&H;q&Sbo0Ip5Lok}3wpcP5s{PQ$xD{5!!~WIe7#c&;nS
z1x=Q&m=wG%bXCs&L4t>HvshP-p<N%wxs}UqQuyPH)(87)uHeX~ilM`6N-=fZCCB-r
zs%@X${=)ZgJ=`6Kx#IfC;4dtIPQ@Jmn-xuuw$Aefi7wT7TNmb|%gx9s9SaoQ8u{7R
z8yO-C9Ygx(o(NN$+*_e2aj7Au-qOVSq2DE#a&MH?Jj99kjy?QxjSaSTLVG_R){4rJ
z%i}2(hoRr*>dg7R;84R=?mov)CpTGuXY|Nv`g<F#e%EY+MOh^VcyCqGSjq|6sEkC}
z5n7+IkleQ)Usu($(q_6Nl<ox!++mH1pwl`0Y5EgWUPdH$xZ<i9RnOjyc)MX|7bn0N
z&Veb|cDm^hpn1af_7%G#nKh*nE}TI5qm0of^7s-pd71G|oQ+90lRRE$Im{^57gi8#
zl9$;HUY?&ub2oQOND1Dsb8i$tmmiM7JQTI`C_`Zz_qfJ6%&t6e{S<PpvjDwlfK4U&
zVq`b3|8|v1Frh;Bib*5e;p;#iRT$GO=o2J&IktV^;1s9v3bE+STt`dQiakbCS->OH
zcIkv8(eu!#lOi<F501}LR8{Fs{ptP*zG=a*lr6mCDL)FchYQ&SjcxHmxVj+0D4VZ&
zF&g}Rc9>1U7KvhOEtBvw(%h;Jl<}V^9cS%5Yy`i?h1A<b9u|su4YL`zj2C;^n0>3y
zmU3trWR=xTh!61YI5mz}=Eh8#Hnk-`%e}r5ytlCvv+6^~H{o^0|26MY0yS$_(F|Sw
z1Sd02OXUD|*2IdB{r0D7xz=emmGvJtJHV@z)9?bM_AzoR(TCA%Gd(~V!DkF3Nhf4A
zg0IN5<r>MgrzlEpER8w13bM1`Epu7x06v=NA(!RxxFA|dJ`6`Jzk@6wKWuCRw2L0b
z!C2TtXl%eVc&y<Z_($nA@*X(KIpQ19%~OcF-o{7MPGKhrXzqD{D6M9l<iA--&^!`h
z15ao0%+rxd8}zh#{yU?iD331Sl%!o**$N7FelU5>yw`@FURN&W1qb8&w6WWgykf=1
zT|vW*y)Wg)q7wa@n#3K9X$=+%?f7UMb*#t%>{kW3r8X}iJ&o)Sjj1VH86zQ_Qlqxv
zn!7ydhC^pdj$ho``tHnMBRnmf*weq=FvIsrLrJwz(?UsXba_iNPcN`r_Rij%(oR=Y
z;TYZgbl@dDxa&}SxNE8QVf7OtAj<##^ge8#m(kivvg*ne|LMJORwkuNUK)lCzrOIz
z?7j%{Z-<{@j;bZj^!2`1F(~UTd~xJ#?U)vKc;)anQC$0UK_n|N`E2D9-H>mpYlNv*
z$M)8+aT%X+sL;AuiVj=lSrQ$UiC34S*G|Cv+_0}+M*Qt-y&FrBiwhUa1nxi48LK;n
zzM2<yu08X+b^oKRj+C+a!j?t@&Uv#E9gpi4Pct#8MK%HVR>#+FsF!+G=8Drlxd5lm
zMM5)A0kd97nN*V9cQz*iMuq}~X9?JW4TK-pDJ@?N3DqVYZm)4-p>>hVQlYjkrNl79
zA$$rLsq8j3nX*d1k2A--e0eXo^Wt@wr$}+GM#KpINTyi3;3-l{aKbSl?9NhjHjPNw
zm|F5htf!~jL0#T?vo>MJ<e<pV3O>~SS4}aUwc=5nh7+N0o^$N(YPMx+OxW67LJMY{
ze8FWros&T|HrMs$5xP7+$J(&*`?cyjYJ4MyEcbVjPdQ915LXaQZls1Vr#_pRO4E-k
zwHSK_lZ5CegC3oJK*odUbq*!%#HB?|^QnoPQr^}Wc8|V;mWOdYyyXuU-tJ=TlQx@B
zL-t23xA;NxyN#xI81-Y^o8#NVGHt)QMz*BHSukri&#>U5Gwxg?0u6TCbG9j4Nd|rJ
zzaju)u|;^oaaAig59iZ(OU~;SQ%KCprcJ6sxnF&A-k=^ycr=Io`#LQfhGCb%kK%4u
zJ)UWkt(AuuMC<i78kCbYRpw&etV7SoN|&l>`&Po<dcKE)O+@7qCReD_bgt2gU$y&Q
zD*1KOLF?nL=C2a+Irs$3xJYGO3iRB<(z}u~u@x36<qY1LWjWzji00B>scOxZVl~72
zOa{87SNl~rQ?%QUux@SGPs!#_ar9;hCLVKNs}f4pOel?Cz{SOT)Z0et<T_!hWjE;G
zY&6JVRp`^Vqh?yI_y!ISB1V{|2`p|64B>_GO_RV}Q=e_w{KhL$HYWAdW!9ZfHzxN(
zN1Qx#PV1N4NYa8oTc92}yRgEJiiNK1jQs}=@zjKSE(1th87Q#AU1tacBn#62s}^X)
zJz25Q2k}o3B#ERSMhhO4^Pec`-&L4$ID4>Hu*3GH_PnE_GV@&?)HCT5IW_4ao4Q+6
zfg`4xD|}_na^?|E3c$Lc)!d_F*=e&>v7+t_;$rE(B_&gRaMien;i%lKW0Zz+sftA;
zKmQXoLi;Ey^r|uE)tRFgn?0>2oo?I>e7%8(lgn#s6)>VNGZg4mlc__EtVZuRZNGMI
z($1(7o?fv#Cg2WYB;?GeF6WjfL2bU<KD5Jux!yrJl4-rQLW)^U*>RD=+Z0;VeR0k?
zU7NQMm(i$;k3PW4%YD&P<n+;I9PjfNH617aq)ndUfrdoCxNbO5(Y{@vQTKY1&3BLI
z;VnVmz|K)h%yz0p{gZso^=Gnmv1<xLoN_TbxD1S2r$zo0zgChgow|0RTfKzyW>7(0
z2#0UI;NUTVLK7|38|Z0fk^<RcS-Q)nZJ*U6H^1yL&*6?ucVzmYr<?DLj`1n#^x-vp
zJxwL+l=z@KmvZWSE&s%<Sp>}jM?9VJM7PXDQ;M+Hn0sK}-7uFZ#tVZjhj*xjt?t~Y
zF3XY&E7@zdl4FzDIY-L*Jw;Oa-97iP3m;tPA4NPR!n(UL>0Sjd;@)}0i=c26vKjTG
z^)wU1rewjC9ezHZq}{`zZ902^fNckI0|b5Eq(jAJ77o+VhZm?g=Ai?BvcICDVVgk}
zjMOr;R|EO0Yk?DkSpG?H5?>uNqr$^_6<eJJw=8g<*VuXrr`);oc(zUf{(Ot43BKK>
z={SHb3}wwM?7^je->^v-;zprs*!mOiALYhZiX4knX&+jtXOc{<D7qanju9rBDqxRw
zj_Fjq*VcP|C5pOTuY)mmYtqBz@`|z2-7ehppxSrxm8}Q<qh^~`V*;~>{7NvPd~KzD
zGfN46h237<kq5B5ZM0RFn=5U8%2AW$^G?B#OuNT>M@Q>sSJN$e9B!f#Svs2z)NBjY
z_w^}qbiJi?X|r$9&02Eq>EAedmeXYjyIB~EYokte^FFkfmby6BFluFvhjkB&IuEb0
z4)H!l4(f0H(fhdR1r`Y_Zhe-F3l}=ZrWs$JRcc;ba%FpK^`-VToVTKmZva(iYgT#M
zNWYoYhMmLUgC&ju5<~Yy?!LWmEhUDHU)WX@+pkk5G&Yh`Im$}6ykBU&zIAkp(x7oS
z$=_LmG_!&Ww{*(U{_>aF%r=Vy*pr!au<GTnZ$q37mv2hVNr1^(9?g|2vc1dA?L75y
zT5g~VCD^c?$YfLaA!}<T?x3YW4@$-|WAK$EUrhMBlO#qan@Lj4f>{_*cqoZOKXlGp
zaQugO=A*;Ill#hCOaUL&bVDssnjAV8B*AbkJ6@HG70Lk8DwFS@r99*(zh%e&VCHL#
z>YVcG18O>;s7eDar6%?!zrFRLz;@u+d|FO0s1{vm!=sdey*<!XLP`q;UI-lf8Y7(3
z<T&Y@dYk2rfEe0)4qj@+0_XR@x=f+(B=p(`?(k3eX=~<)>xG%Gq$%}%d~Sr*Y1U*a
zFs^;HqCK{JhHS@vO%%7(6efhJ?sFPac8-@3ByAkpWN2qNdbq`{!#0|#wIo18rBtFT
zM}Kwm*up8t0*%1a>A$>|=s2E=6yi>P;@(sjrZ5FV;+fAL+fzk{+;e25?04^Ppk}}p
z4jtIPIP#F1v7iOVYrt-adMl&48pySiT27s@u+bE`QPP3AE7yPbi(VzafeaotRH{(~
z|HRXi`%Cz1Ze4IX_8;E`(LiR*A_HA6R)(cqx7wLr+mY+LQ@Jem`kWJV3P`96nA{83
z;7W~Wj$qqCErEf`H=8jOTrD*XE_h9dv)jRW%s)L;Cd7*p>Be2@&4{?h8VxRe6Gg|=
z1=3sXU%DTE7T$J^KvQq&NNF3sEZiY+v^Qv|tpo*QdVN>ONh3=(GJ#k~b8^Kmf@UXq
z7u8;5PioK<{e`!~_5!t_Q$eC@Y3<C;E<U;uf3_qq>fUOSluBBEtpjt}VK-`v4Bt8P
z$jnb0Pt1iU2EF39kH{vWURLHNoVrQ?M*4jWXagQ10cw~G_cdjso~^=`ssRX8r@V=U
z<e{o7NTsqwOnKwUF{^F5kyOQ`fT8kNm!+Hqhex$r@0M8hmp56#w5nzx{|phtY_YNy
zxR^3qJOvjj`RZ}nd5SA)iQ@ab!cT%D#gQ;9`u!GIE-NK7DpIdwr=R*(+AKxgnauC7
z6L|7SM$KvjyT!Tv!e66e>y4@#$ssOEJ?dm_jZqH_reKagwP}v6J=5F6C8l0;&A64*
zez2f_Wx3RDb7`HSMgwP2d^aSE@$KqdvAgxm7;f)!>&dS;yM?!TvxNtoXLBvxgrbTo
zq+-sumF>!uB+6?S?qA3=4c|yg?@8&iQE_qWcrZRsk?KtL6eMl5??G}>7nrsNR|7V-
zH@7J#(_9zCsIad#JRhl_ckc*K(4?I6NlMfB^7fpq5otRnGyIgo4vqNa!YHMSWh3$A
z+Jc5uzg6SzZcnzC1-(k;5T;-R`{~^3#-YHtX{N=oj^;D%?bC^AW%jai>JS7av><n(
zdlQ@|c7<D12ma_hMI8m_N&C&#Y^m}AD$}Ix6ui>WfOo?jkxNxa{38YG_2%EJeHA;`
zH+8=~%|2M_TUYNKQ)zFhC`;MAzERk~`{KE*Pm-L5b&`=~4~u*JqeRAX_h#Ry_)pQB
ziaqLK2P3HCV!$p#)~%V4$vB%+kToVlZq|TyDfI@wIsGxSdhcv(WiFGH#3A&ITK{Q(
zk|fzFVUE$Mn24jj=<{-O@n06xyVuL8D6ZBGaUbwFfK9wYyCt0XT^zNS+5Z>~WKV<R
zj^(C>W5ir@OY)|67IwZ3ThW`>u<MiWx*~;y=VzM?h48Rkd;SqKi1$qe&J?@^LIL5$
zTH15`_UYCmbt5YY5khVye755Q6pDbv1ddM~K#Z+y<d}Emh(qP3Lr!>Sn9K#pe@ub8
zj#0B?g(S{ZR)1+fcu|@(c-Wv7BRdfK3ef#qolhw*zEnY6DME0id_@HC(f&OzSWL#V
zokw1`OV#L-xbE)M-j1kTtQPc*_Ns$<bg7HVK6al*4%+Eu<JHYv7&q(SMLSB|5b7Tw
zP}nSye(H1^d!}SJXPe45Qt)z&;M_Nd)yK>MJA&8kBJ?~M!Y0maZC9i+s&cJ)c$rpT
zM)M}ER8Y5lG<V5ct`W{#s${R3QO;z1tI?@FOYo$lX!IIsUC1Z%<_-Ntb;&W5gQ;uY
zhN@k}YDi~||C36a<jMOqj|c@MZwNMWJZWJXnZI5>c*nmiD?s|{w;qG0rKXAi)1|%}
z-Sb<b5t);tf*JO9C&4RJ>(tS%G<C2hY11j6lC#^MQ;?#JP+a|<k?T@N*1ED?bD)M|
zPP48rvm)KoMB65U5vBfZW{sF+xG9M90uD_mu(ior*ZKdB__E-P`}3U!pUIkZzP(@4
zoI`kLR@hdlzQerZ<4FI5#?hGoPhIcG5DG7xW}33}+1FfY>N6*HtotQxv6L$<87{Ni
zPUijjnoDQ0THY$3rx-LQd2gms@kDP%SNsU0Vxknzv{mRn&jGAqMw}pZ+s>ZJK#OSS
zc1hQmhfs9ZL~J6lY2?*ZLdODd?Zl6aE|u&Z)X7l94q6toL}Qx1_2IP`(|xI(e^}T3
zwZI6|&TV&G=8fYRo6GXTp33CZ@qT`>s>DPG&dhj;e8#kUyQLFNjpg2wZ63q?0c)MD
zBbEbQ4ND?RPn=Vw&X3dw3whg4+P>;=FGRiQG;8UcD&*_)e{LW((WIHmU+xf2eC*a1
z(N@oe?j8p&JZyQz90wkveB<Ga@UX02W596B$9HwCA)<B436-tuGufFncWpFSPUtRF
zFWk^ekb$hnPlNwHL}9w39UI*90MvHv8ShFP{T+46{rJ;tV4rq5b$tnnNZ!<QjKNPA
zV>!Wwn3?F!dpFwI+<X$6s=RpUV)(Nb>I;p(eLNm&xi=6eVWP2XQa==kmQ{%@PUD=z
z^lQy-V0Q|~%NaWvqzthhL8eYY*AIB`E*Hjc<xhlbrrs*IZFOC%xsT%-Rm^IYyfM%v
zAd)<JQhK0k`qS)+G=qI|sRA{cOF6Uk<!HMSbL<+i-mztYsP{|Nf`{APM(-HS_s14X
z)9gfC2?-H4;_|Rq;^}bhUw_cbW}?0jX^R>n6!3i<ejEI^BS=Q!pt$afKl*Q<pTgOh
z`cQPuvFDp9_2_lpH_hE{C3kL#_BprIUo2$FJ~*-9?UE`xa=z@4r=W~J-Yd)=&&$l7
z#k%8~g7&$`xm-<G6-x98%QmJ-7;R*p*=llJSor)-pOVM5;p8?~7$MKu+AzA9jO_0B
zGjz&X56f~48YiRcZEI(m=L`x1DW`V{1#~v$o#Z|^Lb}0L_Z4Ipb+|)W$(#9gA62vY
zx#+x5wOrm5I&{jBi)7V(Ufa$Xr^K6cc+FvAv`oFs7(+#UtwT@JDYjo;;W(s11J6Jj
z1!81K3tX9}kRR9pAiF$Z*7NM9RK|?mfo`u2W?;o8Z2>&?+|_%Yv}!Td$0_-gkY)<3
zOS~~65GpVJp8_#tD27$!<Cgcnc3F(dnTzrzoogA2&3$l(?MX{VmQCXn=Q}66`oS-s
zbH_M!zh3c2*-!M>yc5t&K7dQaEay<~DP?t~QQ|U%XY7+K7*GPZlLG%s3Ofp*zn*d=
zQWku4T)>z14RxJ!;%Po5ZDZnqiRJAU6sJ<jt=*U9X=7|!wzfQSyscjD!fr)P$Im^K
zG$3sZrfp`RSc|Y|rq1k7Xs?*=Oj~P%lX!iFo7Pk0IuZ~@rq%TQIUr<(xT#}`bG(11
zeY9yafVIg>>u?}NXOZ#JbvO4-q8julM@*okz)Es|oDqrrZhz`=J}bt%YdEL0`sF@8
zs>A2WWYv?cVi>Q*WH;{alt~xaV9UJ&ESkT3F}nUv(rBd7ZQ@g$!=95!;IVTv1SN@O
zk8r}E-n@+TEW$P;Wxvq&?a0qWa<?5%$XsMvHf#D?Phz-!j6Z7Z`1N_SS+}B<n9a2v
zmR|d?f<=q)fZ0(L)m3NRTOXSY!2wknCijLs<2O0TGg2Hz<LopFx(A)sW^v8~UWfCX
zS4_rIxbBYuKZST#&c?1%DD<UG0F{M9pQ#;ct4s{-aC3k+r~TqQg}R3dq>;D4Pbl#4
z|8@ZnjX%EF&?U0ZOa91b#R1-`nS@ZzwsHybk@~@{ex3CkE}X6hzqMGFYw_(|uH9+V
zqMod+*2d1qD*e087<AO0c}cPKEqmRqO%vN2gqQW3D4Mk!WEim54Zoa`*%_4_6Zn-E
zl;&RZfvik_K7YHz?72Ww2`NM=&`+K;LzF_(iZp6m-m#dUM|BvR5yz66)+M9LBq<i^
ziM6To_n?dUx@_s*QYlwK=*-F?S~m53qAHYAY2o?D{XRL3zGtOTYw>7wp^AB$w&Vcs
zCFT34z*(07XI)aS{Ts)>iwZwf#aAX!h85j7g^WkecIA>I^;&wxJ-LJitlGJE<F7Qn
znxH>lYw)pStyf!aV`+5NY3rF{66Z>AmGev)Z{%NjLOBWWryp@zvL<<<dN$hK*2Wh|
zHXRoj{oVFxC5x82nO+!wpD$;0sv2oBx{7j>l#Wh8Y)$})<F4<QZ^zwjo{zF<k1Umj
z?=R!)Ps?26R_wqx^)kY=agXqJx)<K~<RX|}qFle_$|i&Ui2qc_Y3)Y+oZeNCDCHC_
zW5KV3tPvt*8@vWgVT8L;@~e*{j$U^ev}P()qXLzCjr0KW4m2U37BEz4qpSNio^}aI
zJ^SX^O?5=_SiM~w!_KX|p%QXc173TF2_hT-<;zw&H0z?cv#>YNT=Tu@>KrmLe98bJ
zU3Rj}I|d$l*PWA&0qMeNw4R<j5bX6-NKBvDF9%72Q8OA(5&DDQ*`dz}$tKPg_Pu*z
zkVQbhZpK$N%jc-oTE206DAa@6_O?V`YC-^KmTFUaxpxd?eUwp&ZX;HW0i2nt*e9e)
zsSkxJ9)&Yt_4%4)wRXIvJ15)1CQ7zcdYns`Et_mj_gu#<^F1t|*;Db1+UtBh?Z`!Z
z!c0)?q@%1?o}OUq8D2FuOSYGUXVxCbi>}-uRctZ~+x^%zDdoVt;MWivqrIsc9Ixn<
zoTYu7vDVhLJV%5|YA$L+#EZp)s^dMD`@xqcVG1M@JDGZw8js*+FjT1GQ#7VY|KxFd
zjYEZsLhU$HC`&?zLtpAS(z;Z~uOA_2gvrIEPt-bpA2yaetV?3MYddkS{=*$h3_I=>
zf8K+4&h3ojli$O5`DBbX_K0_7?+C>U@A0KyhEhdcsd&@M_VImN{C&^+DDBcx+h<mk
zMygxtf|-)6^(&S|rw#k*<me`Sb=*7mRLt8-FQ>KGW!s(Y)-JHfrE*)%x2->n@!%8C
zBo;^-kjyDJ_Qs7KAS}r%SL6;UR&iPywssOGq7X{8>)(0v#ipV;6Sucrme<bot<t8B
zFF`6)KCRkLIcrj;BvtKh4S)2@k=WR0?EA;waTtLk7~H)$loh#F5u%Z2;@!@&$r(_Y
ztD|*H0*QZVMFw~SAT%OA@+%fBnJMVG4rBCCR#l|h#3`8fdF?KZpN$V-VJM(gx@OVC
z`#q%;A1oeM+Pr`C0Vg-W%OgB1vCr!pa;96Hql9{z;CW}GbkLACzNYfVZbV_TQ$}^_
zN&V4e$KBpK^vlbr`ERxWFc?j@r{+!Z2S?d&1(RomjeBX)_V_(@ilVID9aQGUVjFtN
z4tz<^7W_+%-7bMKEs*n@!YT6$9D1Q>S&RRAE=}uB2sqW>H>?>{3-ArTK-_$3qt{6?
znKv4HTH(zxG;`Yuf&G?!mQhiebj=i(a+cyy&Sj-cHBr*WB`p7|P3b72@g!_?>uv=}
zzIp#ix80q&$lQsrdlM|U{_Vt90Rz?{X)luRHBC3Ib@rN-dr%0*yo2SoG#nS6%AIr2
zj0_e9wIq^`Sg;PJh9xAp`HxF;PuNTTAbMs^f4Y|7F=O(xE0CxQ0F(YL<+9wPeRvC(
zf1ZDU@VAP+az2OIX@QkQ{w3fqhg2~t%kmNCy)6``I6Zb<l`}Nj=UfN{dWO!+DJZ<M
z9+mD-`O-RTJzRT%#JzAxe4r~p`Mw<F8tK4h+NU0J5v}xqa^sd|HRZonwR-XmCfCj6
zgY3VnIX`l;PtU@?&qhrgoI9r4WtwFzG(7Q6-#?N^N>O-2$4r<<&KF`hP$M=x@Xx)*
zPs6=KEd_>nOa}1JHa{bwi>7mgmeqhvEFmGGwwzt}`#f?1Amz`u?C+a$Gt2;jA5&oQ
zY_fvF_QLVUmAQhGm8{5*@;gB@AAPbe-Alyr;P@hiwvHiEJpIQS2?F6+e;TSw|JQzg
z{dE8ZA@`1Kt(03yvcacrJ}17r;48GEcYO2oS0oxy%lSG{89MQACBR*He!5|<bZ$Po
zVx?jHwnmA4u#OEo8HDPRNTs&z6`klluk}+ucLppUDT+z58l{feOT)mkr~r(2cVS#q
zS^06`zP-gg3R6+qimZlS^E+4su3135y4LSe0qFigidGml;SX)zS+gJQJ(Pixn}vj4
z&ILL?B%pmMXJi#FGl$vi2yBwz1<1tv_tzNk$Qhw(bZo?s{hvFB(>Xy1Vs`VTG~!bH
zq3Z6Alwe*k-8fV+f0ItI`EAfOqo!dk%dV0}=jE=>ZH@kh)=cpm@JhOz`kQptlDqW0
zw<CQ&K!IBm5{-IkXo2IuQph~S0<pp#eP0`$fK-|_sT^Pt=LDVe0KT8vkg_`v>)Z@C
zxKeq5_yM#iH|elymlnm>+W<!ou3Q{))Cxe!#|bU<VOhdU4IBZ2&{TQZun!UdK`aJ~
z5xI>0=~DXKAD|rS>FJLcqaq%&<NJldqnSQ@=ElFAz>BOR4r5QlJdgij!?5x~1RJ5c
zS<}s>7#QHdB)MwV)ELp!fK|aCYX1uhP_~OclhJ+8_){Wgsot$*Jm=%Vm?7v~P9_%C
zNgS)$?QEeqCI;#Y5#uPkkraqeR*{6L(?A!+0!&4b$k51tHL)7K3QcM;Fs&vtlGBv&
z!i5hb_0OI?o5W!Xfh=)gamsKShC8O1w5P(fBdOQ|{pvspUIn(&d9e~Z>02crLhLKm
zV1&(hA+GO;Z9geexr;1bht~e7Cn>yA&AYnCe!o{w6og4;VRzmW|1Jo>Uc`6sDsy*C
zw85wFr(7<&qow`Xzy>tO8}>Qlr2v#Rx0_Du<|J-|D(ro3@nB_KytBz?{EV(roI7!2
z0BrITWMPcRIFBMq>jEjwy!3$jmoHyn^4!b#_}rwaOw4SH%UoH~ql0X2i$UaDUnQw0
z`dsE@95nk%ohbK@LMI-R!)$MNww^E0GQ<J7EFbP0Byx;br@N2}#VulocXW^^>^cIT
z%)ecaG;lo%7^`@G-<a2}BOu|gHq-t8Ea7lmPKP8`U8FP4bt;cGgq_QXh(Ytkyvy%5
z@6565H{MwjHtC!drXZHbwD_kgj!57%GTzn!Oo>oEr$``tgM9Ie%~_M>Vcu2Od-gv2
z@W`)#>iv|gj$}&-Qho&N0SAydMJX^ZRs%X<dJc5_e7_RBHlwy1n?Os}4pTeapc8UG
zbpw^4GY_Az5f&{tz&q_oiv(#t{|(aAAg7+I%uxL=+i+6>Qi4HkM&`dC!C%^rX5I|T
zA&<ncLi1KQ`N`Bb+>W~+54x4;=WqJE(v25~gJOKE1G;Lt^CcZ9hjkENh}4}HRnyht
zYN%YKfrM3hjSE;=euR-dh5C^4HNlnS%SQ7Iu4yXe=-)=_2*6~qw55&dhl9u={X?L!
z3pj-?YL@DYzgsr_JkY}&8qq8N(9J(g$InkR8E}|e=x29E?wG1!4^zgCNk?e$3u@x*
z6c))FHHx%u^hcVkoPjeD@jEJ@8TWqCQv2kP+?OHbkh1AS1|lkR>r+6^a?$Ra#t|oM
zCQV@uCzy=j9`1qOVjDyo9Xduf(0Tq9x21L&%#wh&iK^VY_r3EGxj=ZO%^~k3^??80
z#X7A>n2|>w83O+JheCPz70=e8>48<pjo+=!uN`^iArIB<iuG7s96Dz9E6#r^6|zR5
z+@>Xbd-F?k&UTAtZe{LiBU@xoVfkzT9r~8VE{A;0fNaI4B-m}TjX|>0fJC>iWF=O3
zIcneXkNPmnEm_$Pm^-)f{RNxG<g(evCbqKWR{&5>eEj%vxGU!Ws1jG|AXc5J+VOYU
z0rU9=V<SPId+4{_{e5$YIh&TMrPAU4Fs9-^Ik{S*Y<nXmErd>0cBB^FSj<cEP*(a*
z$F8^Ja$`pfl6*Gr+IS(zg$3JOsRCzUu1v&*7$3DWKV)s_K6^|Qw|VzFn24hVU5<r}
zD(T}Z_fSO-MWw=6+kqc;1$OC7p$Y9{Ws7x8CeW8AZEr8@^bkrj6~sOUG`|cU!RMD?
z`^Xrt1U{<VI|{&<_`lcu^Oo*OLR8q3%{Wc?*ZcT~uz~C&Xbum~v1gFFEJswg!;X_P
z#F7Yb3m<*M-C6Nb3nbfGL2=tc{~FK(NL|@yCe{?fq#Dqb*_6T#iK*=e8=M!sj4F!A
z{Py8l(ZhQndeKB%<FgNi9(vm#aOt-rJ;4bA6A-D8efRep__aD+ZFrXrVc~~wf;?$~
znzf<%%TbaZ;N_FEI6WN+!ubM;?swAz_TjC9>uk1nF8c;P-btvy(U|o9%f_Ezg#+-F
z3U&IIdHnVKKp^n2yU*FFU;$4p;Onm_Vf#SvEg@V6`P)YCUKQr`8oPR6D;)jbSAmcC
z8zLh`#J?Xd&1={sQNk6Qe`&5?4@cxZNQ5`(usa+9O$lsLQnl<G3mntSi^wnKu)Glt
z-}EKy_@0N4W)WjQwjcky3Y<%CDEe^V0R#mn8sv2IxAXgd<pqO4N_ZH#0-u8n-nuy7
zz>Y}4TSd=G4zICsf#SRGO%WgC6e!}&Ts}y<fO^c`G5t(*l0X6?k`Lon2%PU9q6J%k
z9jd(C`tzyMw80K<u++ICo=O?wsbGYrA*MMof9dBACI5eRNP_I}cAS&+Z#%3^gx^?T
zeE-{--&gWWA6Y>yk~iQ5HXhPqT8Q1O&sk~)H|GE>gixTSlD%Mz7QSnSg46z&4}t8(
z=?{AWfx`Sjp#D0AL3qj(7WYRHKYJP?kuL3(d)&x=m4Cb)L39X!luf|I@jV>y+}&Q1
ztJx#!3=#Hlw;OL3by^%I>^wy-#O5`2jLt#*<sU?g2}JAO!D8xviPi}a=fiKS1%JDa
zzm@U{M1%*=7$TxosSSIdQ6{$kp?H3{DP1G!K>d3s)^+h2GB6b2S!#lkqJgF`>MKz$
zHWMPJ!-8R`E3mPyJOA}PyyU^8@t%I(;Pv<K`*9fel))gYB#=RD^QTpjv%uBFSE(b9
z<Y_tx;!ha^zSp0AL#_?5Ujxag^npXinhX2g&yPDDSqGnYf2+@O^}BbMMxX$&7|m-v
zEDtH2kp*&II4ud2T#$pCP%0REkbcfZ;{S@)ypZ_vA^$9*PX23fz2-quh_cvnX<;As
zpbE_jpTDO@jxO@&(%vHjJ=)7$94L9MKX-p2T%gCpUO068As?*nqt$uX?g<|_<oNdw
z%*eg=A%PbSMd805{rY7QhhNI^kEeqOuJ;FZ<aIoK5BniF>p23i3Vs|?>g#|pgY6}D
zrZX_)xKi)*wmE1H?qgNQ;CaQ>Qos)su|*(r0esm~_+l48WaR0u9sd0qJlQ~GwoZNj
zGvJiw#BC6nU+xsm5d2c3QeGy+X^Df75{XVff4;dgi_;{zB27@`bi|S9=gmmKNsQ=U
zLk_D0j=&s6#pNqfiyVQq!j*p=_bucI0uKDXtiOci8xncqm_F|!ioO}SkPNz9<P6D?
z@V@B)_7^@Z59;c+MyxRR%cQTg80ZP!fuk)8BCJ&Iy7AGP3)r60VUkk)wnMrW1k@cs
zl-(-W1cplq*fTVRu6P+~hWf}LKlnaUc|W6aMZB^Aai-0Uzmfg6v=bzVje2%%@x-4l
zjNo$~+)_9z8!;m9??7rA;{MX62qd<~k33$;ZSx&}nM1FYueSS!fpVsndc-w@=S{7Y
zgD4VELi=A~qm?%dQ5P|UNOJlokfOTkzA3@u+n{dmKQhqxO<HN15Y*1tVb7X>1@%8|
zIn5o!hTM+(2Kf=16BoUVm`<zgHZqA_Ki~R8XY0R)1p}NR$^xiIR@>JXhEPqKz_Gdl
zK?yKHTa()n=~+`E27u|W$N3KfkPmkJmn#43%8;Kb0dGF~z;Da^`;9$4<mE>I?5wI9
zlp0(LD@P2Ji$sJFN9lI(YgY-too^0y3ZCP%3N4u4zI^W6$DA6dxSP}l$As{!5*02{
z^EqqDRcvO4dgNYyH?=nDnpM)*FRe~;s-@9op!u#Kr|zQ;v%Xn-83z|3JHFfA{n2lo
zUa0!m#*%ZY1Cp%$owl8FadsXnn}R&AjlQZ_D09;YHbyvWcsQ5OdfRemy+=D^-9BxU
zMg8(puNL~CMT&|j0-Vq4v}})R9Tg=N-Onc=c_KCed4@D_wl!7*8*pmQ!$FeQjhZOJ
zXs<gjKY%w8)zl<fsy3>EhZV&;<izhaMzg;d>W=?&Vk$G>@Oa7oB}Q?l!LZPCMjZW<
zPTG(Ek>SK_rA@4_M%oa{KZM(XH7jso3GG%g_UwZd<+&A7M<IvV(kqRqae^nli_RXy
zt*MH|r9BKom*>sIM)%4qd}ht9(;!DP=3qGdIi*#jc+f>Aa><8l?i*>{;>{hcs^@1~
zJSFEkXYA4#ZAPUd=Lhl1gbXa&gjKH#Bl<Lx%0+UYmm^}{5<axaQ*;B-LK+CpCQu&4
zw1zM!R-g3DM0g28pHA^2e)jes(3lj6P4#88`91mpugYO;Jws3qJhy)x3ZlyL2p@CI
z_r|y4Cc6uY8?YmQO0wvSBihkk<z!B*m}QzbquTAs;BU{w&cEbs%e0Nxb}LDX56;Fg
z0g!KvGh>c`Uq`t#xnFcF&+y8+z*v12=eBU`4VCx@=-O{WeNErAM;-lg$C^$$7Kn;M
z{^$*X*?zbK0ujUyG777-fq^(HkdaD;VLkgP=qzp{mMdq>>%Z3du@t=1!B+;_zdZn&
za-=$|vodP@N4Myyz1PEo=qv<jdu%2{L6x-dCIWn$6rQA%($lkh8B)0v7hkiooHl2&
z!L2Fwu{mw2Y|`0-!$9WNLw+3qwYu1rYN9l_LnZ)~Gqvq=RAXW0%!;$ItCj;ZH`*BE
zOX5fbkTDB5j1L=D+=wz@0Nsv2C*8M;sGNpRN=lBsPJ@Bm(CJlj<SZ*i3I{pI)G5$N
z4F657h!jCyiGo#qRQ&IEji{UNQt<E>lr7O`{v^F#56%*(4ipce3e);&&0jM~+Mm5h
z;@G=WvwvpWE-St#jj&r=J^iVWdXB~%c1tJE=1{%|C><c;Gi?Fn7^jtrylFzOz)-$u
zs&-s@O)((9xBrJHsGnJ)yAt~q!YteMZ~fhvy*D<}<pRY9csn4|3Qf~py?v%)*jQHe
z40qZg{?CwKzm#QMO9!QmUpDF5O>m>3oQRF)jR$1$1HdFVeX{Q{FXY9;s%lw@KcT2p
z(}wK4ewx&O1B{OhU|t_iUPBHR6yUE<C$<n{v2*VKW9z#Esowwp?U_-MY^5cl>^+Oh
zESr$TvG?XUlu^kjdmSWuXKzW!Iym+^B#yn0efYgjw|npB`}=(UD6VpI&igf<&&PW1
z1pen&|M~OQMA>Qr1eDFXP-)_%E7N5Y&}x|9+8)@Wf_uTgynvLsM;>KAmFo@-CT+0z
z4<&4b%XjOU8rL5eynP<38)<%u?<zB?_k7~B$2u786)K4CqL-Esta2E8XtZvZm;a7I
zI}1Ps3;Z%wGNua<?>TJ5>mu!PHBgwsV>kgOQd`0e7Hc!`9Q?iuzTLALxTX{b)&ofJ
z2@gl$WRY&}2emY~YC!7h|Jc&WxGkN{>w+tm{(Y^WY4`ZMY5(`De}5)hKpx1wdGa)s
z*BV*gNIBAGCzGpMY+9;#`3u^K!Bbe~I0f7YfN*p#a#>t-lEN<68JzaSrVR_TR;%iX
zT)7fz##W{i%aafTSeY}2B0Va}Y^eKAK^FJUJBO;A5#>23Ns8t-bZ?A;f{$d*0+ltF
z@<KPS(BWDHc3If7z<uqMGsob;N)M=n^}$2lE1;~C2dmf@M-uKd?|=bMt|;if6|kf9
zmhpqxrEW_D-$%w~SEO-Q>fW3GQ3N}JHd}16r5iMff4@rbS!))~*Hu;|N&lZ;{b!w&
z@<x6LK>yL&d>6NBMxc2&eStW}Z@fV_rZdL9*JpjwOEs}hBjZLdh36fYx%VR^x!yrv
zdd2)svFh~PPaocBPICVK0+7h4_x-P3&6Ym$`WeDbv{{cDR&__k%n1b02VbSMv$w-|
zQ{W2P`DqlaUz5PDrbTrP;9fif4z?rv8<z(g&I*LSdoN|iB|~u2s=a~v#%;_z?oeFD
zcM6B6cd*fbh4B9=!$5mcD;Z)Ke|Wwmq`5FNbY;!GD$n!u*^;m;|6u|A$TG~r{y@w&
z37WG<<y~Y&Pv0FgHOB<$HR%gk@yJE5YMV}w(zRzdGutianT4%zI`7SCCTCM;HrE%a
zYBhD!FJ(=7oRru0=%oWKe2}x)0`6IzTmX#AAIj~&Yk2%5vX?JsynuX)>0ta?l4K3G
za+YgT(&5ML3VY}LZiGnIv%h!O2`$D(Yw%4roA!rShpaB~&(|R)g+ps@d2PNkHKv7~
z-Zc&k^t_RXC5T_!X^+*;?%w-D{G{PO_Y5c}`pm<h+t6<Y+d6UvK$tIn?YRUx+cMr&
z90{@cr&?X`k6-Vlk=l&}NuX#+zT*1rcfj|S9~ekv0(n#HgR7~>GUWdIbm5l=Hwt5V
zHxt=?9gdHnhBHNfc0mi-O+^6Z<amdCskaKQt^M((tw#ZX;Ym+-(#Y<o;HcW7m$ra?
z`5*5L(J8Kthlan(Zr)kkx5!ss$-3?2bslX2bu-?>Xkpo`GI;lewdLh6wf89~Mc4q#
z*dR8b?uGQ`YVTn&XFE>Tg{vM+T#cKg-8DRf>I$>E;v0@>a@#}a%FWSyay_hJB6OaW
zkfHzkJ^&07FdI2w-dGNB1BpBR44mu!1+>-un6VS4bcjjuhvQ;!S9Raf@;L8F9dm65
z#{}YD-B9QDGqQu50BSs|zZ@njNOAx6rmfJb<m}3tx9Qg-%g^SBCmi5Trg+zSNeX<!
zf+Og+jX-g1JJMOg_dnkQ@L*{3nq9`9XB4|>HrfkS8v<z9&2mhy!oE@N_qD+?7IWRj
zqFoROaQbqQkB*=y6N4JR{Hz|6Pgv3jy{T7b&&ILY(<66=Zs$O<>$EEehdL9ak~kre
z@zDiSu+vOLY~fJ(VJPbgsgD<hBn?nGaNeItv$&4F070Z618Eo*F<dT$2JXWvh}_4A
z{JlP^3_lZ6VVPId-NO%?ccNGgmQXb18kGfc;ueNdA&n*MHWMmR;?spJV`X9?@W1Sy
zpVZ@uoK6k?@o<QEbXX?}+uHT!E{#tC-;d{{drwkVt}5~<PitMA{*G3&Q=@=+=P%#l
zZZVOK0TDwF>s|(K-{8ur6jKw6g{gZG-|3&)jlts6Sn4}WxyBU~zv*v%)sAg-OPNU&
zX*ZJv*4{2DUoB?4_UD(DUUBA?sBAh`6x}7>v$mCeq8;pKqn+S9XEQ7Mj=1)kGXV%<
zc_S(t(-X@s2t9pTcx)e{9wAE#`Z=xeV_a|70Up?m%~jC=U=>4xu>geugrwpMh=mae
zx8={^HeL*7ws$-Y;QgW`?5zXbe`5y!t-J_$e*vI;ijT^u2|k>EbAP5rA&lSq)MQ9Y
zP@Ou0#!g&<&ON+4;Z3$Sz<e|-%r^AFM`Km5t3)s4K-p5N@c@o1`2K2}{-dl3IG``_
znW=eNGq8-EX~=t0TJYt@HND$fFSZ%)=r=qmu0KwnpDD|3c4@E{$#uz>ct%iQB~t3u
z&Rem5Z*X9`o@)D&r?8wd-Jp=nT)cDPbh&EEQN*UtRB*-ZmH9>^Ezth7H-MKt^0|f*
zgLHtk0>U|E=U-|#6$Q_Au=KW}qFTX3I?s1xfiWH{z^oU}amu^_D==RXuIo+_cC~N3
zY^nOYe*ELqdNz0PR+EW;9U7PB*2$o{xR)EUfutHST05D+Goi7En@^}WEn&FWmsT)_
zQ`9uNGh@kT_z}~^if(zL@?$!@bf1muEix2U6X|#N1MbM2erlwd%C2)L&aba=@iLkY
zOvS0YitBgMFBMNOve-zM&(^d{<Sg`vh${5LZ}2BXsK*uzY*xvT{RrXuY!uR57DLy1
z8CPZa0Z4#cQbK(mOb^<f4PjsYaWi%B17}g{u2z|a>Xa8cB@YaIa7OnQ0^vO1ykMs@
z?C-e2pz_XCnWk=7dnfR#eu=sE-|{U03dEt`f3#pFy2=>|OpV}}1fo35@b(5!$t%Og
z&NH&=gt>tlrqp<L)DNIYztt20mh_<ID%S~z{@>!|?+e`je2L3U4ML%CXsf7AnlKoC
z+iSDKDJ)4?9f{`K(zo>pKtISI&R0cFJd<y;y7`!#GO5J2B(AD?#{tIDH!2sA-#Y&w
zxr=B#<{+VHZSZhiB00S)+6!D6cX<gAn7rxFAF%95^B{`qXj`E1T^2sTln{YQ?_vMi
zqEe~9FHn?ct&1x=0DbS;5w-fhq=oHBWsgy<^0AA%YNRI+aeA9Aw=daKD$N`gDY^|7
zD>Ey3n8WzgXU)`is*k1bI(A=^@^#Q0%Jt1K>A9c^S(PhRAKrGT%<&6*7hbX>bp4&a
z9}%+j2zfn!$l;en>_(|LFg^TX*!BR&0cf!K`-w7(JbERvMzyZ}AjLYyG`r&8jv?99
zCfMf+4r1Js2ClCEA1&^92<ijKXJMNwk);HTjGi}ElS?big-~EMK$-8gG=h(7hRp84
zvh_H+(wH0Z{}q#eo(HW2K{|zG)9A{T1p96--P8ISZAn#GC)8I?ej^C*AVpvc`=cN>
zUA-pH{GczJ+6TWpu}e+SfNHlGiV9#FtX_w=cT0oOw*kIh^}Ji{0Se1vzV#{qPUUk5
zKm({?X^P=EN3Q2_uX!(ABAc9@w{Jyr;i$rOytjqRC}^s1gWG?_2H|JNLr9A7w2REs
zy|wb?6M1~mv$F=y&(<w@vyIlewNHwrr?hWa!kvkk($sx~RYVrJ*!c#4#Iahkjj)G#
zIhWTinlECX_j@B8i#r#WKkFaOlx6ZN^c=2EH)cA&M64fRx+oYl(oein%y<E-4sO;?
z*Fm+Kj~}Bj!@KH8w`O4znjyb4abGRY>6Bcs*Y=Jqt(s`!kN4JPK6i@5?s2asf9V)z
zhYclZqOB|=dhv-XRrkzVx)Cg_Eec-XZvDeSF#bgh?v%4idTM|%v%>(*aQoT%_Ps>V
zM*#TUs%roK@#Dv60f<^x(hraedx4JbP6VRkYi7XPzqJ4XL~wzpNfm&xU;-H7Upm8C
zAggvcuuWhLcu4tDYo<muW<~I`mk@my`jGFP+Ktqq>Uoq+>!+R%7N%HH9O!m4D7?QN
z?COl5`BK_9F90s7<g_&k2#0?P<^T9hlt`Z5Q?2XpPxH{lhXC{^p-Iv&%~N1n&ov4S
z?))fL)}@DRwLMGE2!H+yDCj#}e+iAGRB7o(8RjhwThmXu7h*2<sz`Z_r<pAmCvdES
z2E?ijqKsJ@5X#f5j7~)9D0>X%XU(JhvcEd0Y#plTkQ`9Zc-V}{14L+kz)Adifvvfp
zK-N;BkYK@x!tF?rs<}%$W%CYdN!eA9Ze2-Yc-<80nYycGCb+FLGX$<2;yCq5zNTTI
z&o3At5#+Vg7*p9^8&0tvfmF{Qc9T+lUOR>UX?DWbTewClbT1*95z^_-;Na%35+h8i
z+!**+kyJi=%$~w7{hVOzM)PF39~Ec<?i$VP$4MOzVh-;exo!AG_^2KmAJPppZS)!n
z&6K*^qn0w-)h)jt6^3sRFUVwBSTtD_TOii=l4|F^WK(5N6_Z;+I<KM2k~nIXlI^7Z
zo0F*ovX`bO_e`j?n!LZePR)d^ZPDFy)1OsNcQ5lW!~Xa=6uw~!G;)35O+^Ak^AmR{
zy()S&SF>uK@TUbw0Q&iEH3PU(-z5Kqm?Nt|uBMVM8W>O>{1ODv0+~<k0|2#TP_N>|
z{T*cBX#T6pjRVkFM!xatCH#B>Mu5m%a6gs!$0t`&>(+LK76-JZktcAx0_O(j8^i!$
z$kf0!&O-4&u5jGXK#X5U43`RGiTa4J1jUY^^)`=Eb^nxhf#nsd4NS?hb{&_Io;Tw3
z^)hw<CJ92IfRxe{tsYmP)AV5@uwk;10CB_8jv!z668F!^&|>OtD-pYvUo=3y2xtJY
z#^_oFH}JfB%Fgy<{PGOvZ$zZ{s5Var&ci^9cE_mh5zggMGzAOXWvQktt)t}jF)Y_b
z`wU@DEjkk;XIK_^$<Fj9ESCGX?BgP<$Wh_RqMBt*db3V%Cu^K7TQwtayt2`B{-CLs
zx(F$g=ibF$;f=YELP>(Vg}H2Xj!9VQ;|*%-q3ZC^x4G!NFx1Ld{0vFy{3;shyHks*
z+4D;~an%h?&`?G1ye?>G(gakK$b_}Er?H)8Jy$%&!Bq?}Qa7zf4UvK#{UcW-d;p1{
z_~Sm<a<aj{YqJ7EQtBL_v^x3~rOv;*nQ8pK*`;;S=+cc%+X(RzO7O1#&QE~^^4oCG
z8UZ*UyvV!Mkf?wnP-0VMbj<qtWd?E_e4qj8(EtthpUVCg4iuw!u<&11*Y8RU937)~
zHvdL^C_~A0)Zm~CzKJMp70Np^vmVO{R~AJEvdSD=7jo{du5Vv-odn(kDM`oZ_Q;A%
zDt(q;02j<k_#;tH?oSP`S;$gvqtQO(9}(RUnQ~%>?g_P@MN~DHEfv_js;BQVUb(lM
zT<f}UE@5+NUFK|HVXba=z7x~SJ9kMr**>0pPmN_-_j+;l<e?DqfoErX=<aqPV{DeH
zkGmA{hj!}aB8#daSA7VuoeJ8X2&S&Y-(4?%g2-vXjMTn2`eKn$D}7emeoROuS+}yi
zz7ayR3;SFsj)I0LkIW+<{Q&mJ%5Nj78-zNwtx1QIZmYsg_PxE!suP%l$Wr5`xOm<n
z+|3flP&0Ub9JMo377vc-KN6?sGT`k6fhy9l(|W~lmRwLlEAG02U3RO<#R;GrIfQk%
zcNz4OfFi@G8Tvslg<i~Olu&GrlLfY*v)ix=#VJi?pgLc3Vs58HEY9<ZShaMnHC6)q
z5?Kw{aFJ2`rX*lF^vc@Cg7&)@z*`Ic|Lc6lHEm?0T$_`P=A2U4hlcZ=3EzI;LJE3J
zDixt7UK9K^<d8<G&v7cAcQNyYE=6^qKBTUU6)`tWivh6(ybbHMhqZ=9{~WFRsi}~M
z$G%&jmL}mc`nsZXrx-Zwc(i`Fbu6`B!@>4<3@YSOd5z{uS4yhhwx2tx0_S(vb=L-`
z33Y3vO*{fSSp1~_{5bE^yG<^(%SBVUHy`@6sb=qfbx`ZC|DNw#WJXfGqKN&2rn4#3
zgIi6N%(ta)`nH?4Qydg%yo{zun$|}`XsUgRo2F_t*N)D49Cxp`C6#Qq(4%tNLmduX
zo%j6`lTMb#?Ox4S+>UY4U&(mDJp|KQtnHThEy&rM9GXDC=KtMaa9jmv*PP^Lz1@T`
zNO_0iL|+qxAUW6=c#0K(lJVmqYPJ>#(~GHD#czQjB(nxrzo83_>MDnejFRl9glvXI
zCD%N``Dms+ipvb5WdloK*2!Fl=#vpZ4El}b_9A4xalmq6a7u`8qufT}VZZ<f?RoIm
z&YME|X>l9z{58M?$@YOqNu3x;{@?u%ZYMV1AaJLUtfTbvrf@j=X){(*Ng~t#=<rNt
z(xwe0wirh(@Zxe~2By||7d`)?VdVRae4%SX^DBS>q3?<vz+GbCesEH=aZ6S3k<K|j
z8<GXH;+!YzJkG9QJe`sO-}8OzZ0_D>@+(iA#BRJYAp_~0rE+#~;FDlKG2wTly3ze?
zhD_G5VUiddks_LjHijY2mt)*9KA*#CCnvPX$>UnaDH{#4;op!UE=af&Qf$;Z1X32_
zt?l3(n7^~|BlDY5#7f5&-FWSeu4+Qn!>0AA4fiDQpr?X@BT_{(ugpz*1a@L;CZX)2
zzN-#W0^T2Ei>6*K9sA4JWMee~qKTh+Jl(H2i#u|Cl9Yup3f9MJ5SqAT`=(GM{*c_G
z%3yy1sxMnTb)AeUyMcoK_Gd^cR0ovIPXHOY>SmqsQUqtNytVV_8OSnvyLv+p@ItgD
zrQnieW(X;~LTd^NWu<C7o(S1D#4YDN?QsD^J|?WiDJj5qHe3?KOfC^p4Oc;@g6{xW
zh#enzoO^ksb#ffWI4-a_23_;NIR;gH31ax)u|LG)nx4qVLKY3D$$8OybNqvvj(tWm
zV}ZIG-X~t%%0VyRgN>+f+0tihtyTBVMU}2R<u5j<s2v(#G$No#GZ2=kEZ)weU-rZ(
z6*hHvHYWz2UEiKbRqlh=<9T{}261Y0rx!<w-i|=`U$>oj6akuPen-cn;$BLROnlRQ
zVEIC=V670<OQKRQeyS)%imeI#4JGj<ZEoP&+oc_3#D0wvB_RR1WZeWLZR1mE%J0w<
z&6atnm&Y&~=RsM5H&5*cOu^XtIA1IJoO8>Dx4Oo_FiS<X^$+EgS86zYfvM6v?Exp6
zlrH{WI<+KncC2OAoD6K4=L)z}^REC2pnXfbXVzBD(gPY#YaorwGeKdHPi;N@>>94)
zC{UXJ?rPpdgHna_I--^=P>EykchC9yW|ZLW4kmH0M}5}omc@hSPvAL)YMd*}1RsC%
z%(NajUZC#dPV;}1wErHXt>ykk1c)r3)sEAMr7!Zt$<vR1Kz7>(BUo8o2x}&ZNu`1q
zrifW}?ML%SN#9evT=HKpW%6_%!5`7(;J3>*^r$3)8+BQ9GmYea+BPqGvh^t_`i_lG
z({k<E%=xfHLFH5){k5IO@~NsGsz&)nDjV(LE@)kjV}NqliQmy0O~ty{eUSQ*X;@}h
zcgg#U0(U}clqxjAe-G3z%lQtCNN`-1+w8MT@7uV^as)s;*nKzua8>zTweOXW@$nZz
zt-hUdiGaql#C?a_ZE^ogX2ldSI8ULY2UdyL?-a@|yZ>^x$#c*KRO}5kA&~6xM{iXU
z0<IgtOM}tY@*n|f1mtxXsm(3+XAA*PpLu8W54LQ<#~wJWGzh3+|9>Obhl`;3xQ5YZ
z2BLvL{!QDF5?mJ4P}Z`W^|Q)A`^GON6u-KQlJ)7Lw~K{%jiWCh;K<U}m(Z&Hl1m#W
z3T4741yMgB3ZZM}l{oh~(nITMl(TN$LwGW1E24T$suxGUft-!quT?Tuowe+*r`5;0
zC>KxjMwTj9lou0Qykc(pyt=>-aqqTa`Pd6b_MdqiI3L3!3zlq<9}3g~>%)Z@1b%j$
zHF_XL%<QA@di|Wt@tN9SB~)p*iPc94wz7Us%A?WRY<{JXJ8oc2<XJBBJ9>?`1U0N1
zI?Maqn6bywFgX~DL~fW5wbd34Y!P{<Nnii!OS4t?(;gp~NzBtD-NSr0qkdn0^bH>X
z<L+eDpUIrMJxjGc)Ycq6jbFC~hyPgr^Q(a5)Bu(9AJNMHA#na$o})U=yx=LD8X}>4
z3H?#!u4`HTC*uFajlJ}kJ{$R9wk)&FS~$7c(3QOYD!l9x?@05=4dQZl4iSASNnIYp
z#}1Omfkiwztu@^N3QCy4H{=E&8fYRDNn$-k+(a(vQ#=mZh>^k}?!0l;5sB48qO=}H
zKeqv%R8w7#+4;GfYE~W9`?%iBUL;ny&!JWxc5uiqC9~IL-;VR)hxwfTa5Wm#q>+Zr
zz6xiPN#F20(C2QP%JWjzsi9IgltP`7!1Hx)<=;b`7y@k9nb$P)bxjqv`gn#=-PG9w
zMJgGTF;%$4yJ_2k2C{*lWKo7S{}flWskp0Jbkp}>(-inEj7|IwGk12uR3t@!OFaR}
z{&PEFWV4z8xV~b6rwN*Cyf8yh5f|YE0@^iRdPFCefL<B(a?;=5ju^DApre^S+qa@R
zwQaDlI}uZmRH@#|Z4+4uch<da5^n2Pc<3>}2radX2nM${j7?^Kq#>8ohhk6-*a$K4
z7=+18d9HH`iAer!I=@}lguB}&H|a2Y@9eyxV>;-&>N|XAOX?Mgj!P)C7Pr0@mRWb;
zj`=(8_A}DkP^oDBoxa>_J<|te#;4Aj9j~k1iz3JqWib6rJ(Dvi{X!3;dzgRSV$!Cg
zac-R>xb(VZ6Pm>^bHj>lG*-u<YtyO6;dU;sn}Y@>KlHYT`<Zl&gn1bk+6LrI03+;z
z|8$w9)Y$@b)ya^!J>;&Q)J;<L8&)UV8%EPG(C7K>>eJ#A4UqLsC2$7Fz9ioJH!ZF=
zESg(a4Zxf_AS05qvIdBZ+DRem@pr%0gaC1A9+*jg1%=Gh&etk7mO|Hq2?;~-_sjr7
z(HmL(J&-I=o`Vw=p|L7ryb%8<`@nZHT_I$-1`Q95Ye$RSq1Ze%2>ckm2kEhIJhbRZ
z7SjYQSB3PW_u+ytcdPO8TouBP6ZUng3&#k0-Dl@B)K^YR-Pb2`0VB9vx6$Zw&Ju{?
zzNI_Pwm8i{pEo@$E26)?VfHa9s^Z0W3MiObVo+ug)_o|T3oIJuek7H0uSu5sGeWk1
z{h&K}(LJN-k5YzSxYw*m0KyvTC=U~;YZ^>WBO0fU5c`6#;;gFSz7pUNlbu$E@vU+B
zg1H2cxyVYJ!YU2Rbuh-lY?ot7ul944<5%64L6-wV-p@pEtb}P6;ciORum&cnNEcvR
zlET9RMscaWPRWSg9|FeA*XRyZYwC&Pr-pV%-)efVWA@7(zl*T3{|Z>3!f3?_cgEd}
z%M6in*(<=|M-`~GzGOZfOMs&BOvUkcpg>7JQ)NPFz+zjghdS&KSW+ZCh5`cVJHcI`
zKm{wE16-rv)*(P_9q~OQ^!D+c1TtKfykYdUP1~t=A3$yTdk2-RO9~PEqkeyG5}*N!
z?9sytADflDS8&*^*1+InXt_?R!?D!r<t*i87uEdbQ|pmKR#&yU8qVw}Oe+1v?($EV
zA6^%#OLyileOzR*CQDoE?ntJ3|6XMqJmV&2v%I+6Z(pwcE_VO|vQh9!X(fhow1r3$
zCU#N<Pd&4HxXeI>*%H~bg<E-cF@>xB)FGUoUQTvhufODbR`(4E-5Z+FLn8-@=E{pT
zFzq4L@04P-ks-_crqiWem5YV*6DNuMOo%h(J2pAcFn;`o2ry}Nfc;DZ=krFt0v5pb
zAe{dP2VsR=j$Q&9l9`MM>4Ap*>7?~uV0MroXiwk|T1VKN+VXM2-bnS)z9E<P-5Y~+
zCJN;nRN5{f=I&hzNF(Ck#*(>aY#sQ4gL=2TwHMli-Alr`Lcj*deg5^8Jv$H6G|Ex;
z7tm%JH2T)KuZ?Fau|o&voQdF4x-GtUl~bjmEQH;uBgS6y^yXa)C{nlJy53nhh$VT6
z*&mlZD_~w{QtUdZ^Vo2&hp@v=O4|X@e=x^?b`lLnC0cs%Osq$tl<f?egfFG}rw}6G
z@^5w6hf+69OXQUDeYS;rL3u6EZf2L1tJGc>!+D%3WAcaKcvuxkzCi8wDQf!+wt4LO
z^oycua%OJo3BsHw0NtEF3i?8~+7<Vh+C0x%n7UQ%4McpC*6)IlAdA`ScN*9bDTvGY
z3+^5<7qe6CX<c}YKaOEs-?=hgW%>E0c=^+Q`9o;9s#LRNdt+0$D8Z2C!kv$YvIHh^
zP2j$Q^o=kD_=0IITtvsv=F|-__?~JATK0A)P9H1li<3sqHSgfeC&21^N3W#WYy9x#
z_r8V$0DV9hJ5)eq)RQ?j03G3ce_L91aQc8|oa`l@+2hJ{k$~rf2zN?}q@u4G8@Ywz
zneVu?WKa(B!OHII+HqE0*mdXCJK!owL9zoe;=5mEzC0=-lcdvzKUp^JY>v$~DYIVM
z*J{)uPY^IP{)UvLd(3)t=&gFZknN9t7khL-=P8_0gQ{(UFUK`reF^EsOG-NgwZWjz
zx;3O(>iGff<iRd$%<e(qJR!xhf&O%0{qI-;9z6&1YBLqV6d2vR)WP&drkDo)oRVrW
z2IhSht^EXUp-<YXLwr$<lzokJ*+X`16osP|h%eWFo1<Z%c~~vJJa?X~(4YL#?gq-l
zp?*uT_ARrBeF1RZnSh)onzc`tAYC1@@l97~=>`oNyekynU!EnCR8JCC1-9v`gO}$N
zI6MR1`^ub5*-19BgM&%xPN~2&k;_eW_$pb7PO<TnO`N_&Q`2lw2QHWpjNcJW?oPeu
z>m~<+F&nRD<+}hMzmVIqR=dBt^x-vL8PDr*kbHJ(yt0w%_6^&>k+>itvjg;-x;=2I
ziIqwd@1^TfJ1R$N!Tq}hAEvpR;uLEM3fOV+24@{4%&zAiKu^g3Z~d{$PD%{|Ew{$&
z&EmS;jn|Lg7>Lw-XKclqBc$D+e(ymn1<ITy06mjuS<6eaww0DBl9S0}<!Agg%4?aP
zJBGWk^fjCA+b+JL-Cs-hseEh_7kH1xZB!8;D&&z1gZQgeouU^HD05BXgst`RKW;Bn
z4zGz}>T%2jxquAvmzasCQEyQInlL1MIrltWxy6oq9QLo{GzQs@qsZ_AWC4y2l!aqc
zO(i0*N2h;uX8~}o^<!Mv=?5u_{}$miQ{xn?3iP9d1YgOC0QG*sU6}Rlako(h$QPut
z&$zT+r1K@Gt+FVeIkax0^@S-#gn~-P=1S`R5=x+PFMqBGz+mLEm3f|p^O38$zsi{E
zpEuknJpEdu?hv3(=Urdi=_VJ%jeNB#?JOUmV#20@Cz7~q20}rtMKNE4uh2+jyPusR
zIB^lHYvKnUca=nQd{o<a+gIXf9_?E@FdqU87gx_jX{I<i%w^1<gd8QuZzy$nU~{_@
z_<#7)sQZA*X}Q`3q!j1@(@h!iwO|ffivYWc%B9BRb!0fJ1Pgb9<4z+qPI~p_`eaQl
z8t&6lPq~kTyyWOc(RP4PEzB-bT|z`{M(nE6wIYU|j{R>Yi=KWmWRk}|1~1G`fTyMa
zfGL)1;}zkg+e0T`dp8`axq+pVMe1bp8-IY23-}3CBr&O||ELV8Y`@%U<o<UVW*q1R
z`mUt<9p$dqu7LPnFzjsXP@tHUhRyY)Om3pmPNC7SUZrCrbL;8Sfv*}^iGxz3X=b<7
zB0vO2*>qrf!(fg(J$hgP-HJ*UR(C+eZ=49xwfx=JZrLml`bh&w)4ORm)zuIrdic_Z
z?N2Cp{~$8aG`XpkGCU)awTR5Lsps8Z4}}im)C89`i^WX*3L%Q>yLy^C#OkIVCh9Nu
zCi0#b2{lv%+t;2l&G#m78iiP^yTtDZd#%fd!6n=0_=({2i@j2Ra+`~tMH|&ZAj&Gh
zw+Xa-Pydp`^}J=r#WU|N%}JhszU#wb+!&C9z3*bU8cm}*BNm-_iU5Q)Okf?X+TRMI
z-x5bMEr34-bULvU?42hAi5@Mz#pQWR6%tO0!6~Zf=SZmv5D;eUq?)!Jf^7#L^(%JM
zJ{|Y=j=Csve79cl2TaP>berhaq+Y%La&)Un`HqM8-o=h*C*RofoK&AQ)y|)~ZdS9!
z*N<L5Shn(`142I&-YWbd7Y1I-jegCewGMLD1Bw2vlIxW7{caC-@*mnG5DsmOue}WY
zhQ4VwZLiQO8A1z4;G~mJ^re_}HZKzi4>z$%rx(IlK9KXCfmqIOcWh$2^IXl(Ygd1Y
z#6^;=h}~svLmBS8K7yVgXl5*Shtd@pscfDo5IbA!RPQG2Fzimu(j9Fz-@_h^IDudH
z(0V(`BbR#fDXvp2ZP(ZMl5AG}mC9CeaWg!2=PdQuVe4>~aKQ)Z;X|`BF-6MixA0!8
zp*APa$VdxhPRWjasdl##LH`8Rx)EFiQd<jv*Tt$p4DA~kP+GK`^mx5ECRd5SD>qf^
z=3?6&m;z>W-a@AET;DyJc}-&tX5n?FNO%N0NOKaJo(9VXQ!!_uQ(@Cq^1z-Mp)|=C
z|6wDfX>{uZ)E9gE<*jX=FoigGKiD^;^DxwTaml*kht@#*Cb3*tHT@L02EQ3ePxNeT
z)IsV<$zCY@Ig($G8gx%34uf>ZL~jrx2E6n_T>+W#3gq7i509nWcQ!IjXCR$-DM!UN
z>;DD%u5$#%QrXx%jOgWMwFPmk8TU3Pw~8i)Ha-Y<#&FglA%E=E$!BNN688NrQxdn-
zi(seWYalfU3~dC#A-s<uOaP8y@R9mltKFHeaq8O{Ef(c$fyf#=YoAW=pU}cWPZ;=P
z53ec&lQEuVT|){*pJS8{z9oPA<5ks9T8Z@6W3FD^(c-@*kdWP>@#a0_)lSJ>Y1B|%
z1aw5-XQN_I3_b0#aaf36@<KJ0wQU>~R_xV4V8b5kPrbLW=Nn(BXg?fg@!~bEy`TC%
ziOb)$>T@D8#fRo;7o3ntG<2}&k&$<ivQ20Fm4><cV^$U4`7f^On2FpLy9Zn*aMAP#
zLvhSLqqoh0XL|bW1Oh~!5T(cx14&p~P`>aHYpi>X#y-Bg=WuMIz!!o7Jnm+S0}X%R
z<|Bn~wU>bqQzHB2`{%YInxu6;ej+18%=Y&?+&L$CtR3D}N>CtMaI3l5rJ!3dB8r~O
z2;bd{+DM@&fd1iQ;uAcCc*)NCT=vo-V?N#YG5vNa`n$7%Zgoo)Vk<_b^V!yM>GoTh
zT<gI$GQneL(#Y8Zx4NA6&+`Vz1)`VmI@cinXQ|?Htp0ED#|hpM-%7e;Cf@00M)El7
zHoXyc--M}RnPFK<XqKv74U;cUO4T^AEdi}QfE>+t2yqO-tK7aD<|q{oxEjJ&!)NyU
zn-EX--*#1?g~fU!s5hSp@(P5lxH8a(c3rq8^7B_a>Xde?a(&;dsFVOYlXlX)_JI7X
z>GtflznU3?&u}JvG)Y2Vssnqz`aI<?tyqG^_g(7K@Twp8vQ~-nQ6EE&CXK@LF8#Rk
z(;%iq82rJ9V4m@xW5{2$XAwG0VMihb4;hhJPr}hB^W1_fXy_np!9Z!>^I4n!<y@U>
z5jXFKD_-{8JIC7FOi@B`<yP~0@0B<05A*}>ayf4QdbB)QBcPfpZ8WF{EF?_)FW)%c
zOajBi52^rmIPmL7RIXO;Q&Vg0rbJ7>w^JAS``LxiFNr@k5!6e6JY(4=>%Mq5sJz=+
zUB5;(039uKwE2DkObOpT*+Ebu&wmExRx0r5E^TVTyRF~OSF}UyeiG#`A9rM@`aO=2
zKC=Aj$`Nevb{@l`fMh9C2)i_+PdpJ!A?|KB^#<;{DgD^twQvNB%p*iSma%ZM{La_p
z5DM{NPJb0vI3*gPU~+C%m3U`?T98iy+c;K7Sy7*@ICJ4WYL7ij=@=&KLw$07a9$4&
zuDw%ljdZt0p?e%v(sSna_`ff`W?;*fbQ<9;ZgM^3hYFisqGFR8+SqfN{#y%B!P>n9
zDVA1!Krwl-<EFwgd;NhZ$g6w&Tn%AT{g6F8LG9bkUW$p!1@D$es#8Eo@*yvl?}1pE
zxsGwxRfWAhwRi%=?dO;0V0NQ=N;%@muc_VGAM!TuD4RQL#Hw09luGt?Hu}9GN(g4P
z|7?j2vD2B??B5G<dfomv=Q(eI5!bU}m6!9w0??mR97*6&Ixzbb`sc)-Oz=tI8L;r`
zkOJBATg`wY;isFV(P>U4$Z9U5UOgATmte(hb+|smPK?D2G(^XB6<NV%6i?*3^0YEF
z?62=0kJN+?IxH|T)mdbt9bE&u&CZguO_lw!zFzp1AfkL@S-<YwP^}PO!|}HpiC^x`
z+D(aL2tQ9dFRCveuQ3&h;r}{vT3H^<ah1FNjQqg+7X`F^_Lr!Pc=Ha@t#hCZ(@9(-
zp|IgKv9B3d9A_yJz!}NbSVq(^S#s6<i6|E#58D?<frc&UF9)ki(hB6|52R`9CD#p!
z2kjD#t+d{<6*?tl2<#;n9xn#)y>@sKH}Q7wfb^owQM$P6gz9orI$RQ@m;6k<qkcf}
zHVVEmts69sa>~!J*AXLut8#waDI+NGe@n2sZu{vH+MM`mMl<k8<$yt|_#%*#a0$yk
z4*b5|l@Y$aTc=bUJ#HgL3<=?)gPqMTpE`TON(oFg{jU_1Jo-U@osRhOz(DFP5;&&9
zxzlU0?+S2|v)(bT*Q*vsBnX56?B<g9_a7y}dd5D!!#C8L^BN;~^uGjJ4;Lhevc5^z
zdEq71-Y3|;s*yqLS!7%)*sx=(+ZiKe^PE?nSj@!h=N0cXC!@zP%j2j}_YE19o>Wh<
znOGa4OScy<hvV@R!}}K^8>e!g6iN{E%o#aVlflIOZv5C>Z8|$u)@(FU<U4mHKQ;wc
zIBE$QlQ@zOe}2M$jcxf^o~*0U45M9_tN`GBzvnM5iAa+tYX0(gJ*vPO+TgqIuA1A0
zrMe)RHRmF7(;Y;FX}0YO+AZSMekbuYX><-!R8LiM@Dr9!{`O>;epx#5nLpYOtXm5b
zI47I>B<>Uki&p2aMSEUjXQ3w}_Kx8VM=#t2i~}><7iSOI*74Lkk1l(vJ;uBRzZ+m{
z+;sfu-}n(<_6$E2`^ol&p#L+N{-cDnbu`h#7a&&ghg6p}j-03w(5^<~`S6Z=MH-w-
zg$$oht(o;bdQj(iJ07cOPGv}wn$~L;XIX<7=f<-<c{IVKH5Z!{-M7;zr9|eFsyl$s
zuNq+MlVprBlh#?2l2=#PH~kzCSGJMZptJ-}pBA}#bL6leAKotYERHF<R_qp$ypx$<
zgOmBPLBlnpRPj-)9s|MNyGKz7h=;>N$L(mkwEbLWSVqy@3CdY=`>VlVfvxHMQA=XZ
z_Bt#)$ewF_?rdH$Cda4eY3HG&+xk_i_y&06vzl)<@#NAWNSCUEr$v1Ngb2lt&KBVH
z-_`J+N}+1~Sx_Jwc(K}fo`UM~U4J}fy&ulp$6$PRE&$j9P%ay_E4)X``NfM_NkuiO
z_j3%%^UP<&?rALu>ji9bH`D4I2+cEP7~%()@1jEjAke~O=kZh|$QF6fCkU*A>AK11
zY~K7iY;6T$?2lN8FBHe<86)0i^H)V{QQW88e7}WYl8$`Gz#A$1PCryv6A{3E{1GxR
z<ENFbRZ3*8YdiUvM(WsI1DgISyixoX{WhVb%;>KDGBT6iTqTuPBWb=GjE?%JGB{XG
zEkiF>M6?eKdUZulg%Qu_&mTb#x9N|p+h}&%AagDE!nCp<dwG|6KsrANxui+c#r?eD
zw*EOT)$`4l$#^aeKlJqL!7pgM7M3yeAhklBTccc0r`CS&HbH!E3MaF9k1rv%dKThc
ze)c@$Uem!M;jTtm9hYk2V1Tn)(Nf2A3dxM_oT?)Y%scZ7yEbdFCOLM+0gBKHe%Pru
zU!G2S^vtdBR7q9G@Q`oEk^RV)Mw+{ut2XbdX@9Tf0{<+6bCLp^Dh#~O(tv^R-Ne^I
zt&-<*-e90Jxjo{XlugkME@sfbyHe_Mk-)!YUK{xawg8tlIE6KrySJ<L0y=P~@Q3~w
z8*}0d3ix1t7^!>Z5%as9{H6F=PYTH7V5MPBq4=pNG}c??=)=N>=yQX@4jQ}Qid?zF
zFUH5S-Fb`MLCU+^1#W^bf+PwSiC?zN(%X+Bh1a%*l!}iO`qNy7E(oSQx!xJQew5a@
zCb2~$C24WqZWK0If%@U*=jxAU_Yu()-{(CUlG&t@^?iA4pv5skC0Bduy?b$pl^@w5
zG_I?Gg{2==(WlWEYG+q7pB{fj!Q}&L3%w_oM~L0Q?c-$CTwqU?W826gcVtv^RE(E$
ztYR{6^DfK1P^ltxUbE$Aj)ty`NtnwO+-|la2EzxlvQtUfuTzXp;rCx}?grrXXD}Ii
z0HoJ1fo}gZm|INZ4aCn{z|iXTX-_&4pD<YEg3!h4(wA9y#St%sO(&-!>7PGTHj)49
zkUa$7rBQ6$(AuspgB8Rj*F_84jrVb@1q-!qkI$6}I}r#cHbq<uCwM#Zn5P)ss+nN;
z_%`t;a?jWlyY#mv`$qJF%L^Bio&`W+l1k>Qj~zpO7N-rg2_|0?_rS$`j2Z1q@g(E8
z^`nD#*Ax$((JE<QXU%Rdj(G%A(h3jmS|h>-vrN~BdVkvUb~yVuim}P$@J;EIKcC^G
z)G5}=62%o6&(+ki&|s2t*msRa)dLk14Q%$0BFe2Fj$}Wg>p;`f<WXqf+a5tda!@mT
zR=op4at}bkigzGUWf_T=H_Vzn8F0{PzO)8mOHUsAa_yjF>1q%tlS$wTMpDn<;Ck1?
z*J)M-|IO%mY0K$X$9eDFex)T4AVl8U6(%zM5cEkIAoE=lj3_L=rT~sq;fp-oV~n2E
zWoge?!GiuQ3cjZ1q}fUMg7|6??wEb{!X7-YHw55*{@T>*DK62KBC&_?S{5=VhVJid
zoL>^TeTOM7q0Ox}^xgOCC-+EFjcT1^g85Ktd_Ct*&;R(Wz&EW4_a5d_wxSQO+o{~h
z9_C!A(?OHGY=L^UEl3Sob-#$>z232+BlFBtsqiSsTixyK9wOIiQKEEbO6wZBA}O}8
zJ|1=R=AG=Ci9LxuDKP}7l<V^l(8pC^ENV*oQL1Khfybp@1qh$<rwmvsXY)Miskj2S
z-|_1OJDEt!<}%GyPWx=Tqp(`vhE|VB1L72Lbv_qKoI4Ac2<<mi-AH3blA2sDIX<wm
z-k&Gn9C51~wLcc#URq|@CUPbQr53FwMM<5DimC{W@H_f=VLPS@%l@l<-@w&1z+mDl
zr~l>UmFhUWtgjoDIv_2d6J)GL^P5vNw*Ih+=jU)lNbQp`w6KUd_kp*52>^cAcIQBS
zo4q~Xr5VN~lmpa&uI;QI8i|5d9h7%PeGl%2F^hzQFlq&x;ezwwg8(MHd160aVyYnT
zlOR9}BYlEluzx$0uS??_i29N)2L{r&Hsv1u6mBSPJi3hm)k)-oBG~F^>0ptOgp>XO
z5Y{MxPy=#+@cMs&Ak@IWmkILPHOnk|Zs`_13-3vm^!lh%peLss%aabmAp61oxqnkF
zne12b(@ct(TTkQ{HL}%MM`~BYI_{2xpE}o|!rFDAn{jp*1i4!R5L9UPsY^FKuaFXR
zU)TC&UYndcmS3a3vq22^8*(*GT&nsDxi_J5i6+W0-*7R#TG!k+nSjQ+@SO76BB{p?
z@B$(2s-_fkpJwvg2}hoRf1%uIPVs1O!r{BG-FP|IOj@}6ZLJ(x)g)nh@s*-`#Te6#
z)l}Fi*Uz=%jV6XO9z+iSv8~$PgGnyKJm;14qC8<4z`A{roqZ!$E@$OiM2X%8;Lx*y
zwt?u~92QJsO7zqT*^=69CVYX_BX`2~DbO&+6`C|PE-LgUpEY!^F)XwmwMYltvmGr7
z&?!9Hc*<?V2J{e|dqS&Gwa1CpcjLJWK<Shn3PoCz`XE3a4etwD8}Y&T{fkdO;II4|
zF@e$hTv;HoIlt2Oc+}-_%<b#LH(c9m0?Jw(<R?K&m_`9*Ys1i`-t&WDn23f^EB4rZ
zw<^Dr4cJ)*2)cVGK=BBIl_%{ce!=>qIBElgVv$2mz!0GaQle%hC^_|Pklpv7R2^4P
zu7*xjInb<z-Fod54Kn0bX-3bsnx^(1GIqNnA5W+L!YU6B*qnlc@%B-;ve8{I2e*hS
zIVK)57*Y{**c*E%N!Kj^k<$D4r7VwG3W#Fl`YK-bg7R2ue77uiz5(F}ACuTX^XxB2
zCB^QQojZ=koP55Mb0-6u3Wys0=XO`8``U5mY>x<v=W#sV$1gXs5M1f8Rd8a*blN={
z?n{MGHbgpD=M<8pUE@LC!`@OW-aXncB)dkNxYJ~SN+@fr3=&#)-lrV2Y9Q!~yivpx
zgZky-Rn*YgdHk^+{R2f=zl^-Iv1i0xx*NJzHmGc~3p+hI{Q175{c`WaZX1AQelIfz
zL+;C)C`F+g3MM5H_7jz=l60d@bb4#1G@D1%i-K9L-m$v_(K5Q+w5rWzs}E01Y)_R1
z7z#g6J3#ZKg76e&ST`@?OP6QQ-Bj;)*?_7k58G1baclKGsdM3F#EMF0)LwzsX=WIB
z5Lp9;{VEl!#OuW2LdEYC^K@s_tT(%j?_)#WJFm*|;2|#8QxvqLm^@A@K@+<}H;y@$
zLm~Bx4IX%}<%ZBzOJwkOxfxddWMH#XW?j!$vN}C;P>pYL5@66coOt&%_eBuFknIFU
zqGC-dGMU4jo0A3JUtAzOvRH=Z>r49(WwpEORXcSxCgXs;U5#vNriJIGOmi6plVXE^
zKf8}t30iLrda#+f9#}fjLp@uG-D99zT5(|T++$$8(W5WV*C->e%t->Tc==Q1Pai2w
z|9mDP04=-7BH^xb^3eVapMw2o$VFeTcJ}oxLZ(xZ(~hoQ#Cm>s03@~UY5co;U{1r8
z8YXa+KP;uVyP5nVJC?MK;=T$%TEgSYpWWl8eL`ICkJg10XNPOQ<4&<{5U<2+k}pfT
zzjeq#<ew_##eCLf(*E?`>a|JqIr>}W<L<Bh+@9G_`FEEw@FBU&hWpVn_WMkEb7*_a
zH>ui@L3+Wt?1sjbv%S(It&Bxgxp_DLnO~AJ;&Uzpa1r*CC?=|;D9h(xHhy4k>7~o6
zz2l43Ex^i)*?&$>7@({|E9;(RHlm8soovp0mV~1P=lXr_GdmCTvzvbsJC(84P8pvv
z=IIC$8w@4mDQvW4Q19uYe^LtyJC?^<&ideWIz9yjxEQFTKkLd8tDIvJd{|9n(DZ{K
z@LSig|GN*X)P1#_8V!LocS#+&lF9K|^F9*a5yv-;xvhP{5s(5{I=n__BTpw5v*q!M
z{F|P0!~ucK`~lP;=uq?2*gotVb^1g*i5|-iu`H5cwtk`Y*R~V=HR73#w>RNwUOR()
zR`QT$yaU8?6jyZcy$<mV6rRA!aN&y3{Sib;^J`WKcMU*w61lF&t~cx(;~ULkiKAZP
zuQ@bfEAf)*FTAJ(adKBJ8k!-j&j+j#<eqJ4fZlU8oF1(d*^nQmeFdaGH;1=aP_1;m
ztJBfsy7N29T1UyB>7cJ5AC4$eUZ#;b;TSH8p=AJB2UM#mS<M3F%+kKq)pwh%)XO8-
zHxTu^1Ih;XUhswJfe@YB__bgB?-j?hgVb?67$>$Ud`Z0Yh)VVt6d|+>;_=TO`g_EX
z(6I5mK>0lid`z=QRl;^mw8!c5PH^-64g;;UEDyIS4WK;g!eYCDcu+AyYELii9Pgax
z;e1?7+(W!r5OH!53W^&8rx)7quwU;Oe<biE(ISaIcU<f?_BzQi-ez%kP*dG1!BrIh
zJKp#7(vg~1dnIgju1oSu1DQy;ro;^8qAw66{Je@Mb|}@QZQt*&iAQkNzP>HU4{x<i
zfQfhb9wYO@IgO)GzQ#pwi}8wY0_S;ee9a_&$bZ*n<hs`LF$XW3XYmerlg8xDq6_#N
z_~ZL14TM24#3<V@VZA?HvP*9#gfgCXVpO-jmnWHc6lva2@9KUmSwo0;n{MJPrkt`c
zwtIYW+N^CSuGi%Qaz7k<ydNs=7!YmPvZENmCbbAcyJ@`lG-8wsAU`yQFUDPecVmWe
z=iVu{(bXs6jQ3~hg=rX@m>V?}lh^~B&S_{Rw@|KO%+m>8l5YavK4~ZA*qqTsyIq3-
zbS#~K`i_f@Y3-Fc7{mOM=&v@4&2zUfDUiAWRj0AnFHFy7Gk{=QB~h~dgn#vzxfO|I
zP1G!6A?LgOig25oJuGYE64t{hUhSzE9brd!7yLS&W}R4U{`i5#Xg>LKt6y>{>J9f)
zn77$Zmyud^k86k!0nC-_pLNt5vMP#Pj|*N66hCWtjN;XkgHAIRabCf!4}KJ{u$1-W
z8-))$Og&05_SGd<3#kR|Re_LD^4mYfY&K}G+V5_@m+_yaNkb!2X#!~)9ch?_RaR&|
zNt>zZcdyYT&)VL+xAFosx<l0Vp#bk1N;QNK>jf<80-Rwnm2>?HH*FY2NtcUYFU9^_
zdPm)6Dv%Pg#m_3^0;hOcC#O%IY+QWe6Eh$xdIPjBe5%tWNo=Vz*r$=HUQGRsmxTCq
zMRRT)VKCz<gJuKHGD9RWPS)tjOzz#i`Y^$l{g>L^C;e4K_oe@IRUR@^%c!k(6W4tu
zR1ncLwT<TMah=;tgEFifLZEo<JxHE>i!f=F{}yYTIkHf&FI8Onw2#ZrZ8t=*pXDqx
z)n5LrD~{9hw%cUC+XAXbE-?VYN-RbL%JX;}27ANW$ykTOnZkCeHfFu)%x`w|&F5DM
z>M7#aIg+_FiyBsMEA$&nY2gDl*5lGlJl!a1$b)*KaOvD;>gu0g`@K3MuNe1`O2dFk
zhnRzffdQ@8MoB;}@{&La4RWU5pgL17q}eZ;U$2)ucQ8VAY5xN8dLgIcd$Q{HdD~r)
zq%60$9Ps+Jf4urQ`}F<krwAQqXSnNf_UhoRI<$8fV#$Tx6OV)W9yNnn3Xwk=yN_aE
zs9M2G!{fqlsOc%ytg|G<D@Y==$!Qkw=a*&JCdSws;JqMBHU!lu(U`muyZBN5cE`*Y
z`}&SF!R2ADP$5yrnP#H(HGY3<0?9dNf53REiPk6Gh$r&}ahXN;UJM<!v%~tz5Yje+
zBey~;sM0ReTHaFYv^R6%f7~}j2yGQ7ELk5?>So8&d*4o(7B!)}#A-Feh<#Ad^04J`
ztwvLfiFDT#?+1Uv@`>-4j9m!G=<ksqXiam{F?+Vb16|HiJf`~_FSjzW_x=E<V5DPK
z&2@SWU(j652{6Sydv3B-`>SQStK14n^zeNkn%R=~+rdHihCZ*!-B%vil8w0+q*KZ|
z1&zy+Fnup<H{;&T<~_@s1_2ODk~PP-{*PLcUtm1|Kv+`sS8<eg6gP_C_e`bR3Jztu
z+f@6*?ySG?8=Oy!5b<G{uO=mP`y}J6nusvYFkYIHqj%FN4d}m$5G;8ctPkE@Ob2g<
zWQp!l>-ENBvyF}vi7K-_Q27il@u8h`ImZ3b$15YB(RG!XJi<2(-s2U0*IeMu^iWcB
zcJ=I`i%TJq9(tRndpiZsUD2z$Z;w^->C*apgaR?^xAPFc|HRIX4Wk^18rQ`sB#)Ox
z@!Kgj-`6<5GTmZq9+x_%LGH5qeJ&M1`-oEjNAS&xhN^H!lS-C^LeINFu4_)$zMaMK
znp}#r(&uPOGOM}Gmb%>cip>uBrbyjo?G-dx%Fu60q1aC{R@e?Tf4R+36^8i{-tuTt
zx~(g&xbID_%i<K@4X&<m?|d5VV!az*A@<FsG5zF^L*V(scG)ArVQxX?K2nn(#Rc<~
z7OSMN&;_0bkr(_>DFJ{wupHw7wrG^Rh8o=JxQ6*A4p}sv(QgzI<%%OGL_Eqpt*FFr
zf^=-6)^x<VphvquzXw~YoRd_VD{xE&4K_#O#pg3(i59PqTPUuv6OYb|9d^_2N7I*F
zbR>jNX%T4EzZ|)eh~1I)&XptXOqMR>mzY6Zp{(jN+z}2oa=ket$kO{`uXnG&C>ZE5
z;W$3Y^4#3W+9!Kn*(&D98ru<5>+Pk<NLQzPSnn`&w>fT^J<FAuiuj>w?@k8cq&WUm
z-=(`JBQ}1S_+=v@M9C|?n$A58k9|$M`<c!v9D0;f%X8&w1R&=U6YWCOpH|2T$iw+X
z_Pqd15F5vhd(DsV-8~i}0UMxz&$#D8VDVD-p~U+FfeVb?c&*7GleW2Q9^`Y8ruV*Q
z6ri}-1QP%Iz--bBj)^(}thn|O5a2|K?V)JA{uz5u(23xerGsl|Wmwf;T}}1_1JSG6
z*>Ip;l71reEDuy3td?1zY|yH8U1~mSJB#qse^kW2{=Axp<%cxMN!BUoo}LaD=v$o}
zY-v?*R)p5v753-j=b<K}$HV}Esi?bgjq`(g6`Dw0;#irZ#SDT^PDd#^!IG;`V*d%y
z7US!GvbV&$b&gi(i07LZsD(ZQv7`0tXE~&_Ut(0Q_+(c{V_<-r#ZBA>9b$p(d2Tj*
zMt!p=%(v><0#?u^))L`0$U*3@QA?fQ2U&Wwo0pFQYGR@olG(s5=}%CUgf~zaC`z#C
z4*oBp6@hLfQ!ksTc&9-RD)tpK@i@JRCIx#x2HlsgKp?IU;=B*^+TWh8j+M1MU}m5g
z)hvDSy_J{-NG#n}!_P$w=V>d;26^o)P(}9>!}&}adHhwX_s8g7oGT^cU|$)t<He;B
zl`c%mKFZd+pA1^`ggmX>tIr=*=>;Nhr5W<2jk>bU-Z)2V*mw~S8;~S^6zujYu;F}f
zPQDJYguChVO_g|BqMuz@r*VZ+QqH~p8-nfT9B;V1uya=ll0%2*8Sw(~Q?KGvw9ex@
znk&BAvCr2jBsk_7HUzU{a4wslE>oB~iykIGM(C#*CkKGwCnUjCe=kVe%XOM-3)ohB
z-lUl<%UqqUlc&XnH}sPQG~4Dg=gL;zSEq<Nz39ti8nuC)9zs6QpN%x`&#07|eDaqf
zlXVy`Kf^-x{6@NSsq)#?oPrQ;`(oK$tgB`*BeiY;f%GD8%vR^(1>O_a&K}+;9$hk~
zeuc&JAsLps&@9n>n)hp;2NAoLXb4A22kkk=evK#G2|`ng8k+fjvGGP`seZ3@BHRPX
z(WyPRshr5WaV0D72f24QLLgOogI<?p&*2+?pANXVtiOUO;BJ5NRI;miifMdi0H1sI
zS*uqlbm-1-{T@k_00krrW9z;&aJ?l+xexKUn&m2JHv`V@Y;0{`r*YhGn`fzgR8{XA
zdo9@S#}-e8E2(4fhHmx2eZR&b=%v0I%Lm3bquHAdd&;q!MxVSf%Rg<ui-~}o{rMV-
zO?Jd%Pw5YuFQk;0+wP!Ua7Qv^cQ589+o#H~6(YGVJ#HAFEkku2K8Yz0(>{BTFF4$o
z!7|iPCOK4~B}^8r2o<O9|H0_lb~$8sjrX~hHo=Lc>E6|4nU~QvcLOOsgQt~8?6U;W
zMT~uiMj?(w@TM295BMe@%@grfIc*JG7vCB|x&A^#?H;FRxMG&IKV0Ya(dkU$CdL1P
zh*ZO_ll^r2-exyw*G#K!X&_Vy640~Q){=<DE*g6M!S*XFdGxr=P*6EhSl3eb*iQL!
zww)ksUeTb^KD3$xkKBeby!W8V%h`gmR?_--9lHB#MZpRYAQO%C%+Q5c4{^+*a?-jS
zz%4Os)I=%u-JAr)M@dk`{{QGY%dn`|uI<mzB_Ju1iiot7bcuq3bT^7Pgmm{%f>J6V
z2t#)Zh%n?xNrR-MfWXk*`CjAR_w(Ezp7-lM4))$~{IBa;>s-I{Ts76CQbR~;%f?&B
zB33#9-86g;4{WD+4})&w+7qo7k@_;NN~2Jf?xM~L|K8lpp!MuQRo1Za6i|YDTMGNC
zsvqUz+Jk)*gt^%RTQR^utNQ0jb6m|_to2=l06JbV-kEwEThi#uUT_u-Q6i=J?+a@C
zjNyV(Eh4w&VY*c@FWlbWPAsn<)<Q$A%DR8CeRb~t@a5~+dO+y%1&t1eQM;?$LCLsJ
z+$Np(hNZ1oF#UV6SG{HX3)!}O)5j~tS{I2~KZ^n~^74#>iOz?wHvM$+wMHo7h{jAN
z2iv&xtUH=}_uG$b-b*1dI_5fFvyo<sbijwv4GD!mW1E>A^Z|Q<xE!_5?vc)u-!=Sm
z^}&X1C0-%^p2dW1b)XR?U33INaV<twceC%MrD7$K%v-i>3AVtc0GNY^s%s}wPFJ=7
z)fvkkZ0u>t{=6uC9VaRPe}A&x%Z0}+1+l#SLS%}h&V7S3U73qf%3B-a<~hgE+1p9#
zh5;T$EythB)DKtNj`7YV*?8Pk@C(K#%VkgV)m-VbB?|0d38F1tSims^-1(fYN#>)s
za8ggN3lPbyoh}=Q>fRW@b7VEQ;)1S2T1|jv$QDkZrvT}6#Hqs8g=52d5^tF+>Nq=B
zVc&>3<Tme5VukRB5>Zp=bUHH5#)1E5VwPzxC<&IVfX#3KAA!%oP?7W*+FrF7egdJC
zRTMqZC)~lmhZVix4kzBGMkzol3PYSBlo0PpO7o`>zJZ7CGC=0zYdGKSZk{&gwZP}z
zNL#^*z4ymmOsYNRKp*_t>7u9Lxmo@UOC>Z^Ha&x{oZkPe>|+X7gM>rA7<Jn6U@7>5
zB}x>=2=jPcAz}~1ygNTcWaUvPn#8xxwJ5<^S$MI#C<R1_S54hS0rLLvV9ZPnf?tO`
z$mHYvMQZ5?wLfK5`VR|$lkQN=@6&vVg!MDcok#tu{zx)?lEQ~X*p$iDLrFfKy^6=(
z4)?^xu~PdrSuQi)Pj!Zd$CihUKW9Y<-{3;5IMIbk%@zA@H5`RhW-`8W7FT@f;a%qO
z(RHkhWIQXT<_VORZ28Vf5akFPg}&kdY3bIX*bll07sF9JFLEwV1zDz(;)B}!haOS-
ze*Dl>ei~WSuiy>BS^h&Wl;xoxlh>?3v@~{ob$`{E>6WW*Z3#c38!~zwedvOV#^yc$
z($ZddQf0OacL#DbYr2gN(IR*Cl0Gh8vj!cSnpfgOm89S4SN&wswROY!*NTy1CSL^{
zW+ZLKD=dQqZqP2_!r~ejC0TiZ-q};WxcZto;47qGwR}07S(Zw&Xx5p~80#h5GTf!t
zyRRZ<WWQM0YgDBRMrRvY;|cK*;+1`GfxkwRaHo^@mFI?x_*pJO(WA*dIk^=J<fey*
z_;((F+xPoT)MK$jVwG)fIGRY3wS%_9yuV$LZ>j&%q-zb?vElPWo=ob_p}Mh+<&`vC
z3gRPv2=OrzjaX3rQoM7gy)rvZxW#gq=n?C7WNPD39_XNbVDBwTS0u4=8Tc(p{qEZo
z+Wu6RQylUy*gbReR4^fX(2#uUYW<jQ2p?24|23>BohG5(W+3G$7N4~_9}h7Q^89*s
zeqy&5Wty_wSXadED0OI*(7@0bTYPq+w<_{7u86YS$Y@cCCI={s8jVw8!4*)>E30*q
zQ`hrnS)1*8yj;cuP@Ax%dgCFDfU(i2<xc#qtbI~-(Ora8fSl-~n3JWjbY7|Ha$Gn2
zaa2`(pQ!8H-rsqU+pFNcFaKgwfDwUVCPIgG-jZE&OV38^zSS5|i~Lrw06p_fSm;?R
zm`E1OzQk0u0=3R1#j@O(k!1P(ZprqzY&gd@$$R&oeZj-e_cP%xb~{rB){o9SIo)FS
zVq}Ysm~uX)^%Eyg+O55-9`Nbi%BCMQVHRR+VmOH?te22_J){lPxYOP56Rs|Y^g*lV
z8(_#HHIF~xw*Iw3ik*cLwXGQOZD7qszRqjB#9xHPKo3E?#i1s(7MJt2psYbQy22FY
z;+)3JI7?T4v|({j;2Sv0tgUmJTIc!Momqwlbw#-5K{+o(qZL~bx?D<elCLy3r7ij+
zDy|)mUEh0oIi?;I#e5@WvQjn)goYw|R~HEirN89-a`EeoPTq!+8S3VcdQh>iNMnKc
zxR@m4H{j_&P?>{B-l&CA6aZ>*YY3|!%W`^(-9+6(Xo87k-0L@Cv|s3jtjOU?Q&R&k
zA=JS_PU~)1pP*K#jDc>R;cJPP0KzQ*)UJ{z-35*n_VNkX%ShN0HL0({Lau#-^JS3u
z0CPQx^xiM7?~kcaVe#sCj=9P7CGj3$*$Gx=h;2c9Ix|DK&Qsi{-%D<q*E727#_P02
zQmXSMBtU)G(N26N+G!Gj_&3Htx3|9h=^gC@S{}nYqcx1tke9^za*N;aZd~`PZZU-r
z?xZDC^0Dyt5bv%)KHY^xQ*M)>RU&1~8fUv>xEXawV@!GHa@Sh=5r&qIgpl^<Y}6sI
z%a~QuYpi7B9mOE0rAc}V2_{=Qv(>yk1=yKgeqb(gJo}|Y&=koBafO6Ik3Y#h_Av_Q
zLA`}{7UQrmr8gNBC-!8m?p|xXghR0<`|%)8I$-$II`)&g0T8C!GE1XMjwEf!Xs1qE
za(yrDzPGBg8rh%-!euF*KAGWaFS+c#DPeLO+QEn}V8p}!`tCu>*Yb(uwX&wx$G5|;
zad~Y_*V?ug_TzY(f>jz2*v6UgDK^2N^O9~G$DucT627rv2~CBDXa(*UR%*hc3K$d`
zW-W{Q4`HI!iHeu%ea>+S7B{cnD8HKVKihy$R>yvFY<X;Y7D|d)zq8p8u7%+6Z2$JL
zZ#)E+1U4#)%1veZf@`Z#IoNeyE2^7tURcUjCDko<*8p!Xyn&`aCPz8JUD8lH!CMsh
zD;c5YY#NK^DJm|NuheRC?UwM<dk1c+3vQha;bY!6Tkky4qdd_4_O@#^i0?}F3D?2I
z0@X$@3PbS1<fb|C?ee;PDgGD;#8tIeQgV`)pai|gk9ORL3;pXIoepE8Q$!!j|L99E
zgYj;2GjT32kHZz-sKe3CCwl{cI^#juEq--Fo7zm2G{MMTT=#r)??@y(@5iK2E|y4}
zm;aA#s_U<89l2BM;$|4dIuj}@8?LdVH-i<U0m;!4Wc<PHbnL6W&gu|S=gEMeIod<<
zS*8O`joFi!8it7`d}=;pRJpC=4wIB0Q^U!4RYdRlXNjs}z9n{}PxNkYXyfn?uD1Vp
zd;1U2rtV$;Td$RI%s>2h%8>N8UIS1|Qbd~&fFg3=g^I}7yKs)=yLfUfPJq@L%oF4=
z^Ms(gv0<RO4#X5~IGHZ7#c(wY=tI5!T6nV2ncus_$+9cyR!IfBUg@QBtjK!q4|eXo
z@u845gsky`;n^Ag&GPvP<+(YlFa6Vt%byU0DHiHwwED|9FHLKFLubsse4aUZuJvR(
zLU&dW)M31iXWJ=k%ZqQQIU?kzBvg1u2X_R~)(5Gq`@KsI9D^5hw7BPd0jiv4Eck%U
zeE8V;6%E*p7CRxh<aPk^f?`~oxS=VdpT`;(#Gr`4zs2X39%f)b^|2DmQ57p3Ixx-w
z^~ECq%^WhHsc`kN8=u}Yx<r1bX`O};hxy~ReF_-{9q27D=n$ajsb)FgI$UCMtar%?
zR*9$u<QVM-ilyFz*uZ-br~2VdVNxAPp|K-$ZC-}DCpReDHN)~NUyE1m`+Abtq^~Pg
z+uE~%ypE)z*OFL16))gO6~Y}vk>A;L7e8_y(Zku6aAiZ{5pW_(h$5~e_Aw4zw?LDP
z@8k*ICI%py(_F<NPbwwrC#`w;K0}g^fc$;&;!ApFW<C~I(F|579KWW(*&zPFy;fEZ
zf2p-?`&Y|)UtD(rpNFeW6&W+^IwE84IC4SAlv=&G&Lh)QwCZsACu`uEr|4Mr`lpR0
z^ZcaC-eeKtFVy4|njrZbSP7qMgGw@eoim+wk0e2nFZ~GEocB9eW5`k7Tm`P-cec5M
zV}a~Iy+&sHi>I@E9_My<0^V}%H6y$31Om?uGI85hX<k6i97I}Rty+}*m}(xGP_jSE
zUM+UT&fc)lEa?aHVah(@>X%`Y^{69o1y}`syU)Z0eb_DrR8sqX9;2L`p-NN`E-?W=
z%w+$$a|SUY)hQ#QiaFEbCq%Mnf2cte&*D9r)E3ZHjYQ6kB2pm5!7cXG0FM@89`nuY
z)jXB!t~A=6#uKkp9uIT=iA>DL3)E^Cb^IdYFvmas=;rc*>-FFv7e&x!Qy8-ddv2RM
zt%RL(XPnLiDi3}MD@^6z*qGC#75x=(JQo%ipzG(L90_~cWc#MOfI9ruES*%TMgfV`
zNQPKEr|y%<jT6v(fyN6D`Vy;W_$lD6#O*v632{9YrvHPS232k0#!j436`=%$Alkoo
z8hw^8Pj_6kO4s8Io!(ZBou||}egOFn&(hSAHi#eTj{N>#H0##BjO~uis@S$g8LyZ>
z>CrO#ga{j?*3&^(<ix;89DeZCt9xbKBP^_{$#e7KV&MAe>v`@ZN!Z&_a`|cekHjF?
z_A^7A<&Y+iiYVnjtd#6NhRP<v?mrIZX?(vmS<TnUy1SI@KqE5feR2I(r-aP_Kv%A{
zzLV7jydVpwUMB{XQo^88!>+sL`hk!1;3lb*CeHTnskckuyr6rMFp6sOF{8$UB#Do>
zbzLIFJwD*|n5vnTPR-Rd9exwn;gKtNX*=}Z_CXPfxW^TXxuI+2@b;j2j#(J2^<i<3
zS`!|1fsrqVQW<ntym#$0ae~z(bHGoL<fDOQou)UvgeD@<=S#tLJ}E|fKd*Mt=|=Mc
zS~J{q9LlF(<1SByqR-%}Mzw3KhJONV#=IZWimgN|E3z&)IctKMi=JHH5AYk90(hh{
z;o-U0S5C+;B(N?v{~YJ~zC*{SsVKNlk|!nVQ>yRO%lK$f-G;enb<v&ON`Lgp0YziN
zURDFCju-oKi?#pD<59we!BE*^_lKd+)X9vSHyh9F%LZ=OA4>z|AxKN`1;sY4Y2uiZ
z$2ZP|IXaOrQ!xjiI^mF~Po|PwuE=hJ{b>~TnO68u^CI&czmafm%Zfy6L=kywBTG8T
z^Ez>%x*j^Vsw1BW;aOOz4aP|ur(KQ($wuN`pG(PNk#T!iAJ+iAf_hE^As*EQQ_Y~l
z_`S?DNb20Tumzg6-OCl6trZla5IHnW1{@Q28Gb-(<M_-{-_h_WfjwuL#RipkZ<LW7
zCc@KnA3HNbBB3F#fHydfgVNGtW8yJ@zwWe-g6(ew3Q;%r!5I^;we7xE_+o<aW&%1J
zB<wY_Nr0=<@X!1H58w|-_+xf|rL2LR`RS|yHJ_tdn<(NVQH?9sDA?A-#fvvy<HWi#
zU6%kRZ_KW-lI~>}a<HghlcTt)7m|NEEVKhVezAktS$dqW1q{PrLw1&%Q{d`tidGrk
zhfYenmx~L>E3TW#h($;^)T|%ui`Rw`;ble|<M+uIp8!M+?@j)o+>yWy?My7}Ea>0(
z@R~cgs$;slNqlGZMf-Xwwz;s9%}RSsneN0IA!c#cC&s4bKqiHU)Y678Z#v!WH_}b1
zMOA1qyLK0y(r{dS+%qVwmF!v@<nyxEMi<780xkcp58Rh9ZuPSie=3{X7y)`ni)8|=
zucoq$pw0MnV5_+ZSoGW)hB$tC(3O&>&oo_$DFbf(mMq$VwsaUTrQ{T-_6Ni6{nVV>
z#X%NV=W9X@_pNYvjH`ZDMwF9CPPWMvPn(orO`GQ8fCT`t_C0YQ26O79p^h;PwD7%^
zrv8CaKm3(2W|4tNTp4e10w2|&IJw0szr!gn@3V-v`v=uCeT$|W)I5nX_qRs4KILM0
z&EJy;$KvqR7P8-_oY0#1(7<R}KDB0j_-yXeJXUb=X9t4Xdh?Oojp4@rWYcLjWD{g$
zDEo2?%eklXBqCvXcqPDUugu4)ovhn>CT2})?JJDnFZ}M(Bp`MVL&KJT|7?IM6T<QS
zQSHTUV%=vu3dBI}CjTGEoeKlGOUqBm=+4UEx&m5QkneaA;>J|hO5Iw2Yx-dSpvCQR
zit5S{WdH^_AE^Bu6DowOnr`tV{`B5D?)|&e+v_8I(;26@sTOwrriEsKOa-PN1@iPH
z5NoSZzGz<4+>s{coa-{1>0ADn)7e&;3nl)_Oy1dsPYX5n?{;yTA_xL1Yj@~Z4i-f&
zVUpftZM5C=cOo}yT&XR=CBgCclJLj095RJYpRF_?(Xk-&%2{S^zaEH(RxAGMGA|Q!
z855q25YN{Rl&?O1EdqOvY$@Em`<`<(R*su}R*@E~{1mhm6omNV)hJ?I&qL?6xJe!8
z6O2WVOP0MwIF~$VEHf58My77(G5W}d(?f+${XX=(gg+m0U&jda1w^C=NtAWD0_qF6
zLg4W)JJ~xmPNmN)?>vL1|0JOkP;c-)z)}$n%&v$i(V`c>&u&V=1cv*^{Ys8k=f{6y
zq)dDozM|0i0boHV;p|g(tDzEUuNf8i(-i+V9~EgmbyJ?_q6yO1Q>A<~QzgA{UXq$u
zg5NHV*Q6Fjew$D->-x$0;RLJY>|lRto`lR#EXi)$;qHxTYu{bzO`Owi+v%@TNJhTd
z@zTmZLv18Z3!Nk{Z=Z{e)2jyn%Ac)$jip|jV!7K`fEDb<7w~1`16t}P9Iei{zQX`V
zTN<kaGYT80X}##2!83Qa|9QQaLTOFOY4fr#1HQELm@h^^-iwR|qMs}$O!aG0-*BiM
zWBIl~4A-U=V@azYbv=CTgXlg>BJzyhugpfveW#104yW3FJ$0QIgSE>M<-^hc`k$If
z^(xQHv%sXOTyRWM6I0h3{4IW6y523{yLk<$l@fF$lI$gyQ{v!leAEP{+=^EWPGkmw
zR+u*~%kNE1Krt$S$P)^p=__;4CTzPy@6U=fXShGdQM7P)cb3ZU&*q2TLjG#MslG;4
z_LIoBeg0gV^M<0|nqGS0<6ZTM;ev-VFTUtLU_OmCO>Rh~Id{IlaQd($-G0Y+|6FRZ
z!1xj@O4nKb<FXdRdxN252S}&+Kie0<@nX3NK1Q@vBJO4|GcHI{dkSi#2fRe=Pih<#
zjKp1*NThaJDTQo?^JT|e&B7DRY87xz`vy9)umn`5O0*ZjCew!?nQQoy9_~Y6&I`nB
zJWMAwA!y~K;(R%ew_ruq$6|30Rr!|k%Sa9&R6c<Ppf{H@=#Pt>pAEg`3!cLd#P^q7
z09}=)udi>q0i@8PJu(9<3w;|9{KtN4rokGl4-J&u0&8s{St_x`X`H4HL9)*1$I_O<
z-v`eNV-U==T42ueO!VDtv9@AAzRGmFR%V8e)vw2IcNOwlzQ5(lj=}{15o{rLI(Eka
zK0Z<Hz=!bZj^|Re9y+ZfQ1SS^H+TWYH(2|jf?wk65upvMBx7iB&ocOw&%k<<sI4aM
z9J-y+ThAJ}^R<U^|M?M<2O<c#>h56|`5!tpSsc}p%K^!4V{KCgZ2G+0a6BIoXkWzp
zZh$V9q)0qUkx4Ou*Tf1;;^X)&6jEM&Jt7)G7XJBW@4VT~_$g2?z6jaJ(sa7YwK|Xy
z{0Gsoo!<9Y(z(tt!eJ=)+Vh$BnNZJEu(5kNHvXA~yrR(p;emm672l(l${tWShqlzH
z3V6{B7O2JlP@}IU?N#*~<SAP<y0fpm30ncilh(xtp3)C@E638O3bgVF*1Q1N7+9d?
zR~_G^mt@M6bJigAvHzq07;%CNaB9G!Np)Ggso$H_m6={7-2IEW8K0e5?;oc6(3F6W
znxRCn<>JL(^4R1rdAuY{aSL-hzyFdnBCuV|K?Sn_VtgXTs+5@G%_Pc2u8SF(7CSpz
z7Rx@u3{6#@!-|40Zfa)(=iNQOq~R`UvwG)vY13W72Jy3|grxIqeH<Gd@yA@-blrPe
zk{-INOlI#^r;^m*R~h0g*q4m83ovh=nY%K|Y}}SJR+^ug>i_5lMfl_so^{O)!K~Xh
zW%A9V=~tp3B7+KF&Kjs$0%PIOTNmIIgd;I#pHjMna3<W@*CHO9#(zk^<2v*q#M&2c
zo0tQ}$`zpeLm-^xv6je;Co6EgrK|$*YFvorAsl}X-A}f#@ju9~rkQ$ZWMRU#%GB3|
zT~qRS8y3#$DezOl5TD7|&4@+a&+4}*Wl$yPM^EXtH7|h|n|t09Fi6cCpk3KcCV9FN
zv~#$mf_M(qn01D>vVJrp_^!h2<*$NhFY#_Lex^enVkWBn2Blr@fHm8JVi}bgeF18*
zugyxdqWg8yU&ZG>K6NM5X?wr!{TdpqS7xdLX)Tm}i1ZivrB1s9q*|ea<)6V2oI*{o
zOttM2_+u@cI^_egaPP%0g+AwGmpAtO0Cupk7Vw$W-Z>;IX6E`V8cZ*o`vDA2yIp5d
zP=I`Xc9g_hw)n1muhjV<Zp`j~8#ux!nuE$mH^EJBsY)xK&(PT7_ZDYL+GA@*C4!cZ
zhBn=tw|K!#cdhh#k#x$1;-d8GQ8VGlsYcr;Nm8wQzaHUx`HDwx`1d>O6brxpHLzSk
zXS=ygf^}m+X}2<F?D1`JE_FP9P;Y%oyM7_*vK%Qj4AU5?G#8TW&gZ@BJlL{e@I+Bd
zhe^h68LRuhKluXR5AXYOt`{OTHe&t*NXxmO_=NdB{iw&i>dzOoP?@dAeB$B)KH3Lk
zk;-4N-sYfymh#wbtC-wsv%g7oh&Whbpjn5WvWR#}l}H<W$W)>Ql6Nq52qE~PSF@ji
zk|Z)w35z-2>Dj>FPSn5toZ$|3|8uG#MIB&DC69{0D6G#oK>LHC$UjksCjx1Mf!`YW
zLYL(=Q4P4|2Qmr;wdzK*Cfatzno|Q(>YiM@D$OXjbw9nS%#B^;{{4ohI~jA?jp(ob
z)I<x}?(q>CQJEqda||M<?h4W#EAMLVn9ojKYaCU+7mD$Q9O>PgFWPb}kPcW8U_3cv
zu)F$mhiAqsym7K3hBQT-wU(-o(m5Nw)#lTs)h})t%uwfK-SR#(c_nIVL<SKi&O3eb
z;(w}>Tucj7@B$%|L<JO4d@$r~u2F^$L(q?mPJz=pAAD+o_TR0E(iss-4HTxEn9{)U
zV<kmbZz6v%r(Ws0{uCLDdD$~wht+7yksR!6%Z>V*+Ia~sRn`}#E-(B@35dRsF^~e}
zjT3_Y4}$K9<EC^MRk`}uzwEN$MLS_!+_cvPWYqNH!-0rXQ|0PmP>mZStl{)sf5WWJ
z0>6QoI4a$!=~;KOtiH1;qu<M8b=>q0-bR5|OK3186<Ua5q<M;K+ZA0FjCYNqZ+eWw
zo=q_D27f>NVneok=ooyYyu2mwGZ*l&1AmZEv}Wf{R#t&MlgwG5{d5PuJ}rc7l>V|s
z-|_hm;3Ov?jk14=o3jZHe_*f&2*4t|uMbw?s?;ui78yhdinB<FNuMml?)OP-sRQtV
zVqdK>yLAgpY3TJuM3pf;QXegF+U{BYQ9vhL#bSBtZiJblD~7UHHy7{Pnv8sU>UDM!
zOw8q2t_p%MjsFa0`&YyG`e#KtSXw$VYSfK-yIj(SQG$o}B?bS;a24=Q@W1cnFX}Xi
zP;yOu?Doej?`qEflSY;GE_e12+OaBQZ@Q&Krgl-DvB~~q!T-}6i?Z@1WthF|Xwu|)
za$L-CW%lob=0yaC1kMVk4p+tPisx36KXuRYMY`aL*SL4{4SyR$EwA4{;vC2H@fDvr
zH}Cyz><KZGgZ4l)F)gq9??yjD8OYoZ|EZZ@z<$ur{L+)riwn$cs}(-<d`G$58Ck5w
zlo)K7e}z)wV<v1aaz+fC_s`hYQIkaz<&qQRwEwSjPjRQ3Nm?%D2XBU1^tq*^oN9U&
zZMfBxdN}UwYcCaQ{hLti{c3;^V?^q-87-I2pXAT?I>oVSc`K!?if<4|9FNhhK6p;Y
zGwr5d_bi{TVQOOI0J}@t`ILJ@YDZ*4+C6FX(m(Fg#wjaAr0A4_6H^Jr5Hrti^W+KA
z^8VcWU3}>GGuW6JyFWugGy!0R(0K4*^f-N(qE-)lU2B20POkFQec+njZwO&?6|dXs
za2KbWS*FKo3}Ng1gS+r)MsZhymp`56ow0*7H14oaPJ2Ork%SF`ZQkVRsxlnlQ#rlU
z;>9T<JpOjL)obeIlz`|ge#l-Byk1&lrHn{F{<&)Gfh7@lx1TVNLxprRR5Gin!Y6Hg
zlhfA5<2NV+ZdX|k-M#;M{Ou)R;Ic3)nteBM>60G5fjr&0?D_J%g`$SKu1)!S5N@c1
zD~K~7OmOMoKtlT+?M;Dm!w&$pLV!&e>%KYVssZiao$v_0r81hIuw=TCSZr3rGj0uh
zubb(>mZde<97^e`IOC(j`$4v8b|MlfE9KAY|MV?5qdoUc)|`71;}jW0tOjJ)uY*ta
zb{tUo%4|E5h8@bDP$y>;U-KCUhSL_p-##n$&ZC*UjNt-FX_ngF$K%QkuJJNQLvKi_
z)<(&ucl}S7s%Et24Q_(={vbsOR)XwpvOaktshm_w*|${!s_L1WPM|8`pz_R9)3b5c
z6IMF7YAO%LKK^HR)|0(VTKzc)^Ht)i+$9SIdh82~ie$my+1^`J<K}*LxIu`;T<sKd
zQ~2HPBC)?0U#{*Gf?(j4#Mx<9j{_TR!#i{^H}*S>Rh3GrK~hwgMT6Qr0sSaGncf2T
zZ`~gmEopf1VD&dJCO<23{{mIJZ?VU;D)g${sVPO)SX<~HDu6j^E%H&|6YS3)N`Lch
z_w4;)RB3a8<vf6-D=*|_A}iESg=}V93klaOCvt!1caL4p5^OQC^#uC+4Hw>A>_q%5
zUrvMd4FxgcTkp32wz`bQ&Q;0Q4@0+NT~6xpDmVx?1Pg$!J@7D0EhXniJD-ZS=#1ws
z663UeR06(KhAOQV#vyGMHZjTW?buwL)bV=b%$5s^Mq1gW;pY;?;gX$TXCA2OuQ5v_
zOam_5SP8E6eSu`r@EB6PzsC=U(R~f$ItB^fgrr5fAEAwa0U82XbN66<wwvOqf%zY9
z0h$@yew1(@F!aUF5;ic)MOAPYSbI?ta2ca2NlxZ>FJfda^&wqP+*u6q1m+BhSpvYe
zkK1L7myr>CUp7-c=LvVDi9e1tqG<)b@(Muo36J&Izt3T!k+MN}?46EOV>ry%6Ow0c
zjMxD&!7C&N3*QkSbH+e)ax!II>vpy!Ev`2vb#55sc%ftf3LBp!;=pwL+fgE{lKaK4
zK^u_XqFrroqUSfaRPQ+V(0iwi)*84*CS0Id#G=s-li#9dxufDJ0%0c<IFUnJhSjaY
z95G&7BSi`Kh&)cJXS^c@b5+*F)$Sfla)B|Sfi9!f$3V30X(vo0zh9v6Ge4-eXMSZz
z`cDkteo8sbHkwiUT|erhl+G0__a7EOQrT^^)ZVwlC977ul}i89<e`x?qW@#9x1+Ql
zC?ZD3-nmqO%0sp+%t$lr)xkaBtPZ?ey>Xf{ame%p)|aYQZ0{(+p^Z^Pms@wgU_it+
zaIv~XV)q})G*4tPij2<3RAfxWrOxZc{D9DTs!kAq6JULS9>rX^r{H22J4J{4bP(NP
z$JF0qy);L{qf|g$C*cW`QI9o@<9i4&7>V)JCg0FEJmX4by#@|<{Q4&>myzC!94!am
z-!K2;tEas&LR--D0q-NLNt4;^XY>#5%-vU=k77peJuTs)$uSmr$r85su!&Jdr9|;u
zwk|s5fh|hWUi7O7pWdVK1TMWRw&F_kt73OI$3vS3IpUH{+D%vo`6h1$b5-MIv1He>
z^sX@e>p8ZlQMBq>Y#@y1<gbK$52KNgM#QB<5c^AgDYlYSkYtcP$<Iq$fF*EFwdwxy
zSC?r(jaf6km8kuc=(~sr!JBVLh0IT1APEQF_9cz-k&*YAz^5SCb!>1L1i__Kcn38v
zVA(|u7g`<#?p@iRtswKkF?GF`tL6BEw*w0UN)XkB&*TZ$Ek53&Gk$Y=d*xj|7J<rW
z0>>BbABompu?N?AXASUTH<E7AKm+Gx?u^?pI7&lSKw`k#{*5nRNnU#)7f(Y;czS=1
zfk3$dkS1U?@zuK7ZjAJx%LO~ZY@Ha=S$yYcoLE+PLcJx9?hj5larls(@2{qTfOX7D
zaZ1Mxp6aRF2>Zz@TgBf2SAOY=oKKYBjF<DNg5DbFt2c8$oa;puBgb$V@2-u_XtX{C
zC^eR2d3^##d0Y;vk=w=E|A0?Ai4lX~DH%mtL6x4sYuCykt9bq!EQKAy=&qpX0yAqF
zjw1!6^pJ37<y8o!z{qsB%8xlc{GO*Ek_;Q&)MaNV4D&yuRpn`*F%V^Wglw|6<r_Y9
z?fQePGi&nS2U0%Xo2zzu2lBpXNA}oC8oM>gh}HI{K7H4|+f6<hBmVh|ZtZh<*HX5a
z?CsV*9Sw>2)YUG=o*5KYy1TuAdbutsfXlv$*5<@fQh9at`^i4|XbsG``Z|e1i~@g5
zKA_?4Yho5x%DWPirxVNurgWNINwG>cb6sAO#%F0NiI0O7)vkHa)->MGp3MQ0VRvp0
zzo;{lz~3^Z(+Kkh-1wUhe|+5?hzNw4H$=qx6Zz$*P+grE4d!f&mw^q~VZ+`T<Aa~c
zehChyYG-A)1>md9fVoN4w-@<XS5^;P=>3l6j6bDQm8RpS%ay}aQ!-Lx^=^o@J~%VG
z<=ymHr|T2HiF(5%X<D{T@T+Ad@}Iu{bzVSlf$wFu-L*oeZWX10<NcF}VV_l{q>3Gp
zfNNHt3bjG-%D1w|xE0gn+E@ddA5FrRziSY$q!A|w*=obiF5fC>Bh^dO$uzbEhja&9
z59&<Tz!1{I*|go9=F%Gze?a@uUU%-fosIgW(sw>z;9{cjLQ?3m+@%J(P9oxvN@V#1
ztD+{jg`<85zi1^`9WID<LO!AXRS6@j0$J>V7WY`#w5)Di0K3dbNRlc-^!l<rC>^Yq
z2QuP7<&u7Dx$)7KkUO%T*-8yJ6|#<X=apJ6-zP;Z0jELe{hLz^(1Bl+=C5RXfDT0U
z#p2uR@q0RjA+(G7Cn~1<N5oGHPMiEMyfFbykUn5NIwP>f@r!?c{-jZarN4~%*b4l&
zh6{DLCMql=Zh&|K9o#3R1mr6q0(m$`nRc%Abq-Kf-T=Ssi~4%YY3=v9ou`9PylHW3
z!x>k<8^jChvhFU`#s-VXatYBaPOr2tYjz%Q6imy+uAxumeDtqCOcXh-^=uHKF$D|z
zNhU;;x!QW@4Y}eI$475ybPBZz6NYb%Rr;N|J3TOXUp&(9v(=D*v06Yy(4Dbi0xvt#
zUn|pg#Rf{OCj&2OCIy>Ph$Nm5?ps{&#uoYVUT(LLfn?SOAddlCL8V|HMRx{WC^X5!
z|2{KNwQ>P1D<)2~CL85m@@y=;?-H}(+xzka9g6)eeu4ObE4y*yRI~L(Vzj)yS#`aM
zS-gDdP>t(aLU+;3#-7Wo)n%TvUQ=4!{MERzN1E9E%UQm7&23Cqp_4f6`^6((qftvR
zx{>o_N#d2(mo*aufiI>{vy7yNch)|sZNdu9x!KKpv*_syD;26_R*QwJAIS6YjcuR}
zH1#O^^2$JUb!zCOs>?D(1aW5?$dbomhw0FBiP&ZRFlFl51mT(R8znJ}sTQvDUcbO#
z|NT1VVR;h>8zCbNUk1+L>z4!I^__XiJo{7(0yKF|zuW(}GViPfl+M`6NT&dc$Cv{(
zurnv2JkWiOOY;+nHKXNp&6TL|iEHTtSpyJ@q_XZSLf!dUEWZzvH}kI>oc%;d;bupq
z>xF+HES32xBTN~q(Z7awIx~DDLa%i2pKey<)Il~YJ6ZleVM|fG_tb?~XjnTLJ>ST$
z(r)}Agw!FLR|Yq5A_cdLuP-1l^&%)9!Bg`tE*pXnoXgEmcKp7#I#T3T$*v?V8uAU}
z{?@&xz5@1lEsw*}asX+pUL-x8u$&U|zWXfjP<!?buH3|=uwS_(#rwChq+rh<HG~Bs
z$Ayt!@IXeHtU^khHYkp6-@1I&hv1l!gN{k<56EK*aekb-qSC&^XS*GHXW&LA4wtku
zL^ttt9CWi}kSH%dBp1KHXnzphlHq&}*11hW^(m#hYs&aHzz1UOHFS`X^;VUMj;b2<
zN_I7|3@EqnY_ta|>J`4If#bs-sFg>8EIM7F6&R*tW;g@ZYeoI5!%Br{vhnhpZfmv8
ztNX0OlDn4p`DA_jHIcWarA>*6hH+D4K);F@FY&{(<1wgsWK5Nj{AD!NYl*HLAE>>W
zHS#S#e_cjvy@2*F3-U&x7KwB#mUL%=UHg2-=@{K~i*%zLy&zKvH%%c(|1!HdO!SRW
z0r;cf%t`=zg?J0Wf<cu{JiMK!+Lm2!TYaq7-b*4MhPARXT1J7t+3I-y?c@3#`#br{
zuTE}Xm!X~PC|$A#e?6(BXl1<wpy|95cU>mGAvU`9{+9ebAQn#y&LMGqq<D1-)H8yP
z%lt1+<HX%oWf{R(uCLNwNTxf#RS;BcG-h4PdDpyTybnh7f_ZtJS7TNOp?RSFc;%p1
z?dvR~?8Zwm_?B#BZzP0}vQ;opY6YB@!-qUuy49PN+0jghAJ4#x(z`!DRV*t?o&Ucn
zJw+B(Aedav4?iZG&Gtc9rFAJ6z{<cWVNVO|1EYdYL-2ESp1DB6mX;$wLuT+Blm!vE
zrGowTghd8C0FZg;^Z=tS8G!dEdb|J~Sc!EVb6QyM^62#tg=EXV0Ds3>#gal|_CXB0
ze0bC~5@5aiW_tfP#B$=#=bdFKDl_h5fp6V<WrKO$-!z&&+g7bPcYEr+Js0}w=GKMS
z^gcanW)nZ`7ZZ2fC1JckGnAZf5cE#%*0JU@TC<1^9-QH(X26tEnPod6Eu2$`*mrG~
z8OfEuWyb}wvwFi+)h{wCA4b<*->}u#a}Et&eabz168uT0poOt5o_DC*oslMiG9+or
z7!@nAO=5>IU%<4}-N4uzPA&29<AB}PE3kmr8hM|Kyb=?x(4*@2j^Q|0B{?56-(`qO
z!6OoRB<j!{FJSnQ=p|?a;-6oSr0#2REUlEkaqNLPYx|n<dPu>#6`q!6)%#H^5sM$-
z&k*gm>PEqG5F=C>ka32HgzJAvo1-8*Y4KAmY3eK-I$<wG`<<slmag3tB9OH-^W6#M
z({p3IzHrj8R$50OO97Wn!MO%aoHY4nnP5K5+8>{Gkn7B*2c5XP)(whwVSXNmT4~oM
zS@wR=5#x%=cx(pz7UU5nhudNd2u)%e&}0z@RmfvjRgHuOX+}|1*ab?ChX1V~9z;;q
zrdZ93%<ZP&R4V<0*Ef#!nPQ3beRez9SAfMV!*ykF<q5|wWc_<$MiK65b2%f*;oeg`
zDvh!2xfa~kqBS^Y@W*@$@M~d>R<ToyzFIx!I?`LLNKQmm`j#3vdF?N2Kt0Oz%Wun_
z!#Y2<In^%I{9wE3SWsQ}$fx&UFh?0T-=2-nN(VCwAD#|aQ06#Ka_Bg#p?vvToYX)u
zWp~#rUEBk)q?qTQSFu~k#0$A8<1b0g%W!`5I8WTs9_thYMF>!HYKPC3k|ubyzN;E5
zn3Wp|rLyP0C!<``5PVODI1%1a*wDK2?0tq;O1HoQf0xs7Gcg|ldKujBKhL?d-eJeW
zZk09kthmV7b*%NbE*e$we!S_a3f`qvq`VvSlU`&8EcfCzyN|)pR=jha_#*2C_eu_K
zfPE==n3j^92z6PEiQ;%nL!^CI2)okQ^D4T2lQCw9$k40p>Afb<i!~<Wk|mLJAqVX{
zc$_5veVk4XUvY(8Tz(Mq;tlLLNCImttc$e3Kv|sAi3xMiEHMXdaU5M<iS^;{LCfPA
z>mfSUVJU~B0nw`VTUs}*i&p*We&}Typ!G3~39GOiDuG@nb5mIP=RtuM?zMou?=l5H
zzlcpl|Fl27X5+-}R;-E8G@&K2-8%lmS9_MVb;DDaAW;*IUtXUMv#1r;+wRUgQ$3d_
zi!pdaf5?1PcjvJf`+KQ!I4tk*XEi_Ycd@fy+!#Do|7B;AgEM<l)m}YhA#9%|R1kEO
zhe=3VyM84$f1+Z)77++*Ds&dpWpT$;c)wE4;v}6R){vW1h}5}tlKZ!k%waum&ov)h
z@@VI4yDx%qchk`+6lc9WoAWowe!eO0dJjaKb_%X@|H=h|ePF%nqMIpqwTKmwYtpDg
zRSt1e?OUj0CRmIt1TJjZZ>>0~mgM9RjwN!{IpC{Fq)?A5&B{kFDAT>XfF8(^8C#{t
zF4TB%@=0_+L*RSKZ<C9QX;1T<(Lf%kMWx5m<Swd=h4nEkFMJpyMrqervVNQGdAoqa
z*Y!n6R&*g0bOrl`e?Ast({;g_j3u1j%Jm`o0?-vPW<NAMK>swH`8YzBznH)E1rQzF
zvZ=fv*mBsm=c?v0?x9y{8nic@i}aY7cMIXL^E%&^X1}9p_RHSQ#LZC}UO_oJWqchH
zw!7QDbIrKkAauNq>D|lkaZQ(}wpi*0=pb@BiU+uYI_TG$(0+K+I;whTX63;cyqy`C
zvaq&o+1NYuVI_4{Ia;RDKF=UhDW@r?)d26e0kOAO>|R64grrQ8_jn`qzglUe)16d6
zM=WoP(<<vz7f#Kq03E|$;GfoXZ>n8;XK`3>q;eNL%mu&>mb$?byxM%5n!w(BbX7V+
zVXZP!NfI0vo*nAX=U;D?$xzmqjxw7l+02R|({c(U%%2_JjATB$#vAPg^bBl``Tg_U
zx>3Xn%{<>1lFFwhg2t*Ct&D-LSay)}ezt=LNVB4v_LNspN&#7wAgoOlaheQvmp%fS
zgmR9aORYp9THsx3DY!e<1|F$m*?G*o0^nK%f<VkI(ksKc9>$FAgk9o0WHTRX*=6+>
zgIy^8e~LS8OmX+8*eOk`77G-{@;`i(X|s=vbaQIx2PL|^x61^F?V1G)YuCQy?K_%Y
z)__OJ&EidZ{T!A&7j<M7-$a|%_jM7xDyM_r0oKEXX8A|M@zH)!U$I@X*T~xD#Mt`#
zYr@tvgEd9)d(Td~^_Ek-l9z)+lcoNQ@w9w-ARNM4i|8|{edy3X`9z+^j<CSLtl1v5
z2hu!SZ(d_VmT(OxQ@~->$1?wwX4LqS8)O+N$Y_Vz*Ks6QNMiRlzFd3=1~**5cg-(#
z!QsCRns}-K1{9i-U6q_61uBmHW)0JpMExcU$rdEgX}Fh%C3X(MI+V`u`crKQ5By|Q
zWRzCz@S7K&f=OI>-%BX^)2n=HMmL*<(o3?CcE8STO#uufGWS<fn#ha}%o`y{Ktu`4
zjl3n4>~>@_7d0l|p7GftWY@SOvso#=W3%`HO?wqx72)qG7c)R^KBuK7MQt@mn5ZoW
zlGpahF`)=?VlJ=xdCeMq4Qg_lcG~H|dG1U4=K+(j!SqO;L>BMm-0!wceGeRcC(Q1t
z@g@iAKjS`W_u_(O<>86=`;2uVtJOj-d0tK{2QYCTZuIk+yJN?;&IosN<&6||Ur(%A
zZ}Hc72<D@ql$EO7bt<1pT%Mmp#O5vwV}aV(B?j%UU=`S%XRsbCP07ppYLoK(Ds-)O
zI!c^B5VmiI*9dH(D&3!Wh~XEtJDg<+fU6?9Z#vdSzwHh1NCI?8{6`{B-S1^pwqrCd
zDST$q{R5gQ3Gj1_hbZ-EA-i&P`sN6SgV|c{{aV!v)r%J7bhU%oPVsX##AD2El3&CP
z6iZl+f+$}uYHodXD-JW#5cDbt>AlYK{7&qJI&}c3AXh7J^G>Q_e(oD6QAdD?8VdAN
zz^2qMP{L?N<8$5)xxiNpiq4AZOZo*k=boj8N``uL%;sG?5g;^4a2;`BsinRK%n35e
zi`~lLEcOK2G-*hYv6O8Zy1aXuvio+2d&Q&IQLG9G3obAJzDMLnn1Q*`UT>Bq7>UPW
z01<vt$*n1**3@hbz+rGS?7VFay9_Ai=z$Y?ZfH<V1&9(FQQes1$c;IU+`SKC;;Ek)
zV-!56gQZb@C?75brKf@!Ce3yFkFu6wD3_r0)sa~yTOH|7;Bm7-3dU@?(oSntKsb65
zRt5?Lhh27wlA8X&6G)=dVtl{Q2y+?f8(ei((k^47KPPgfq8C}O-BGN)RENXDKWd_X
zJf)5FI+)2C-re&u+$kbr5Pj0WItuSv>b}0uFFXEkFLoAfoe~5N*gtuRm9B+5Ri)}r
zXxF0;kJEmg@D}P-z%}uWR|ht`aU16s4%*`jY!5yvdH4vmqg<ZMb%q)QCTam<+QPoe
z3#CyC?-~&?RxjuhGh9Th<CA~sF64~qhhBr+<S?oI`iC154roM)@7Md4z@RPz%+V1b
zIEY$X72W3SbBj)3$9Yhw<o65>!g|C01-!Ez<dYCGnlyOpz->Wd75V!{DWq6$DArRk
zm2eU?bPD2Y|9$yxUzq==R3c9WDa=-SNQJ;lg2r=_G;C+A?YGL)bDLIdB6()53^Q-P
zOqpxsg0EM|aaLAufa*Q*Jb{ZHAF(X;24F13zYqK)4%)uPoU`N4T6qE$#O#N}tuiw@
za+=7OhHjnZ4i4BZry#YF_(ng^N-;AvNggpJ)+6lK5e+>`eD&K{axOlR`Lcc~iHc09
zXgKx8S#+l*rRg3LnDU@|P~UG9r&k~*R}1JL#t{|{KaQQCZxNrk4KiJ|^j5ANJMS<5
z#3Wz+Vmef>X~Lqa9?hog;ll36jAU9Lf-t@KgXCf}9b9L`pLuGlJ&@24v0l4n4k%5f
znun7|3??@iw&^Fz<qxdto_!V&vL3v{sgW7tyuN!=qphqVz4Nltep-A^$^Lhl6B3s<
z1rssw$5>+HKuvNdq~?ySo5FJOf%*|qT>-)TB?!*W1i=X2Jp4Oe`4?K&gjTd?ETb`6
zL7yDm6rN3>GeEI1&+rVuFhQBBj=~dA_*R^DzgNc^o+Pu3zcHd8R2<i|*W9j$40+a%
zeg=@54Y6Bcd#JMdcJXZ){KV9%yZxCzk3n4tV(>Im+7`g!d?qWQ-=vSPibkyVl%LZ6
zb3%8fZ@@lXoMZqBbN+S$P+@D|V5Q!2Vw!cP;hIBULt6q^HmED9SXie;e39~xfhNq)
zk?DVBc}yuufte?<_1uoZ@cl0|Fj^v*CoyAzq_lhf#^0{;-!Ut0hfRQRea}ny^$21`
zSPzNhT0Rw060lvWMx~#@k7>REcR^mH*Mi0ute^Rqd_7F2zzc)GR=iZGOZ19<*nGs2
z&coKk2xXvOY%fl(<b1Z3%**vcHt(D6NdWh02%o6|?thI$7Yz~+8D_!nhItyPa@1nC
zaV@!i*KeE^9E6{g9{eMYWPoK`ZvIoF*Njuc{>u8uLy9j?w~=krYWP>#hRnXUnki-V
z{WB)f-XlE9D^%rIn22B4#9u+u%nid=+XVb5iC!a$Ulsq#jy`U3J}^tFe=Z_rzx?Uq
zKgJ1VJamF3Fqu&&8HZyUR}i)EDlZkEDTk<oS9dv=ei<e}62F(+h5Hkb09#>{+mVu<
za*U;h5To^3h<PBm=ZVr$q|Y+~;M(9t&0;9*i&#7iawj0XBK={F5>w^<n0qiesoBd0
z5rz*eb|;WI-wVY9HCGmhw@YXQtkk0CG(27X7Kcu71@UL8m;&3Jr^SX|DR7uCDgF`^
z8&4esgi&ZmiC%&ItszCJjClLbuN9`So`K@mD(DmwoAfi<rVoHV?6UV|X%9ncf{3e#
zh6N4G?xNC_%!YRVx}?P8ohC)<5ip$uG5rk)1kZ6f-gPXdi(?yv)zLaZ^bD!VamQZC
zcGPU6lpthtrY1mWPcJPnPWnLDcXV-jKq*@|opU2C<62E*w@(;pJ~}7r0J%!?T|yp2
z|KtHTm$4T=3gpVQu@M3??g8|U7F%@$fp+xRz$P%7=7Fgk4z$F|0gz)an5)XJz6awI
zD|#D<(8;!Y0wzI$OmZ8N1H$I=>+Y`U3F}fXzwFQRd5gpFvOKRhEnvgp<0oE7sHG^i
z9RqOwqUPa83*AD+(xOPaKQ7}s6KfQdu>j2Bj(kx54H7j^C!0z(OvdiHZghw|tPPR%
zD3K^3F56uuo+-gr2h{iP!Uy%MM}}pphA)`gq;X+hV$xr}IYmD=8MkOrRu&DB0ShGv
zd$hI@VN%ewp_%9*msDY<YBYU)Wq6~-Rb?J5)k^^}>Ez@@c*%H%0-@DG?G~2Kb^554
zLBjQ>fjr0O#_zmM(xvj8fcHbCr{O#@C<rJxot0KX(%tzl9t<}O+cM6AMl!#vCdK7f
zgwxXO)R!%J-7A2{eM`L#puV(O30-G#<#If*E+t`B;0-|fhLrgl9)ZA=Yu=ZrP!h0W
z>#lLn^MFH?@%+cz#jq1Ei1=r436QLQQUlhD6yavCeQ@jz)Y$v}DT$Rdz(6=oAXXBS
z9uh8$=^CA8>7X$(t8=<P^HRMzmt&yEA8@Ma|KP(^UxCN|b*>&*=dsOT{}?F^e&mTY
z_-Ba0S8+dDh#n<i7$B3q?+tJ5a4G0L-%MD!=qKNd<Fmc1xT6LDAnaFVd3<`C+lcfz
zvf*r+#RzrI_+G*57a*3k;kehvaJyTszqS};%qCn%!It)I^3{J#)44tVtB8U_KfsdH
z*H61KU*Ltj4JchOW!S$A>cs>WZ6wMoh$K?eLVfF9xvnKv_2sL|d|vyl3zvSalQM7r
zq>Ov>N6L)^M&E%7PGJtCj-PYWHrzQN*DfYm`FycR&w8R=YS?E&i6;-OSoB+=e}oI|
z`8_xXKob6$si0OVFM!lZ{8*~r%g@q-%5uWRAgzXk=0uQhju)g{5{H?LUkQN+Cw&i_
z;FnH(_ZZmqS|g7Au!;hHLsSIa@j)Jf1^)&%<)oN9Z)=Ln<^iUmXJ^-Lnzp;-ouxVP
zRW$7q6w4&w;y}B>+s%6uiN;cm1eXIoWEPLNm(vpo4gikHegABb1TxSS&lL<ufvI&w
z!WoqM)-9%4W6l<-V{YvyN>u_=vyyxs{f@;#1_3p`tc~3HIQ&@o1lz(>f^$8hASi?-
zv7dON;?x^t{mk8ldEBU$#@;5mC`rV<K>#xp^Qy9}XnY2ji=V-Cvx=fizfAF2oQgff
zzMoKUZi7%zjRytJSC16dlYza#)s4Q2bJ#>i^*+Pu_r^CzGTdRt%T2Gx1%l|y5cLeK
zIxHESG<?9|HBU~<-9bppZT_1_GyBP}Tq54LiDY!^SfZi7ewe<i#7VQ50{>@E-_zcN
z-uNm}h1HB(zZQ32I`RH56W9k@yXFtbmz^I&8f<_Z^$?$K$t}1M%)=jp<*V1Xky6ha
zEYjMLaF={Q)6RqZlQkix5z8AL#21h@dFHYZ(bgo$G8_A}D9$qwC3V-Z-qR_ah`rkG
zCA`w(PkR(0JT3kUJ{aJt!qJZ1A-Z6DEe&(>rVQL3EJ9|`jw%4sohwL4p?u*Fhybi-
zT5TB#*fpDm9!c0`yY6B7Rzbfi=6~3l?jZFp#NBGFt<6O8yH}<jNgGoQ6dMSR-Ax%~
zpC3pP1orN|2li9T`U&*gkq<f1Cg&2VbF=0qAZ%n&8h)cKhYikLLX1_97@pn_u9Hlq
z)+igQsYfiU^cxPrep>l_(opdhJ=83#nym&4S`#!eeyer2(a%F5E5_+C|M1qi4+zU!
z-s_hE+XPOZy7CzU`zf1PT8`+Rqx3wT#>`PUzIoYqgAg>Cq|BUgY#Lp9;8<nCieapU
zdbKZz?(i<Terc=*Du4smdE>w4#%wCgq>w|f^A9F?LSO5z<r<rR%T<v3GBHesWma+Q
zXRd4Je^>xefM6Div@uuch-9`+U^&zC0VJIM+W(D&FBF4HfgL^mr!h|m7v$W%d80PM
zq3rmf*Rf$jiq6Y*`f~)jPl<KnhgLyyg@h^@WccZt#}Gb99LmswG3@F5ONFmiFJuT2
ziO^Do2y)4;OAiH`8OC!(3%zGpIc<ST<_(`9t6k;1n=Ew{@7=h+wmTg#fI`w+O>g}$
zt|mpH9NqnA4>Loq%;LW>N<U?6G<*UE`>JtvR_;e{aL85TFE0|T4rW+~x9X}=E#y!7
zbd=SPjt8~xokd0WZ63O-UoDR`4Gd}N)<g><Td(r*ZMEJ?8ZcH(`7Wz@+Z}rUP^QB$
z1Z58zQ*OpLj*rky+JKw(k-_aL0<QPBF+4gIf0Sf6RW=GZ2txkEF`maiUr-M#a$b+W
z#nl9dkBA^8S!iShAsO61<jb697u`fFn|qZP6K&$h{K}`g%AAR-jut^gG*xu}@dkb8
zSAQ>Kh^x5ACIY3=YrB5T@CW}6H0%+|6Ylu|@X+#Lu15rkKdfzrf?;KD&|Z4~{>o6>
z8T7-mipv<u-EcS-v`tJbuImE2PdN4p4ihKbtOb3DWE6v#CRlwc=@_gqRr~jX2$Vy%
z{FS@PWtnibW!c`<A5Pc~gJ$)IhnjZ6<8+1@lkcgqf-u$o(xhVMo;iuFdU;F>0&8%_
zD`srvn_7pyt<3?0oIy%8NjZ3=O04zN+GxPZuH~HSJ24HULG4KJ0TgWsHwOmn5Mq(-
zw8^^1HRM>lgEWtFC(YLBAc!w{g&*l%UoM7Sf?}!qI1O+D1u?2{X++0QJ>!}5Scv5d
z6O_rY5TMtdiZ1yBB&^wQ3J}RBAd;gklsuMQP4@x*F~b_c*1KDuH)ZYx^UGmJb_z8@
z99wqiXy4Va#a5LT{wS(ve05M;{9Wy$*5w+{&Q$2tGfI>S!t!X^BMmJPpf6v8f~;xb
z-YA@Cmk#&*UG2aOT5hG6J6NedNmC_2_iH&!ddRgdRU-tw%lE_-RY@J;zaQ*gV}*r|
z8A^TSJ$mXz6O{|2(;*45;5&F{+hjM32qwONs4(r<k>p`~f3VhVEgvT}L(f@Sxg5VI
zk)O4rYMhI7TPzhiI}i;Ea`Z2Yn|xqU^!9+eAd3vdp<&>5a_}gFa$Xrz&f7M6eZ^Gz
z&X`L7&0n4|+=vD^z?j3Af&BGO24`N8A?4+?e_`wx51IYiqhz2y5?2-Ga{%wepeimn
z!bLLGm;ujE%oz2jjN+8ME(;Q4K<f0vLtb$_4$Q_%%}7h$<B-PRudzbQ+3-)DE%wzZ
zE#%!g(qn@M+OJHtb?<lo)cU%Dax!e%y766D#A0HtV6a!L07R?1+HMw?ZdNW2{fh)t
zF&MC7D9B4kNPl8XoVZMxNa-_^qJ-=<tvqKE33xw7U50!TS(jmgY+m;OOJdR~kd;5}
zyec&B;t$YKCyS<r**XLzJ6-|jv&Do^w9X5ul1Y!qrPtGH!~s%IK#GI%LWw#HyAB7!
z$e2MLsm|M#%HKfYdWmjko9ijHu>fkpu-Ce9-Yp&WkILtDVjf0+f{G5x6fY^f4Sj}6
zkQIZBq0Sn}bFsLczzBvxv$VTb{UCM1xCp-X99Q!j0Mrl_=mf<CN$U5vZq~%kw{3z4
zJIqd2HM&77Pc;YR9!I#JV`1sC>!1YN@xqB&hQNNu1tN0Dvp{-ge~ZIf<H?1#-K$26
z+F&t(FP4T)HkpT8zt#8}_CNkV!oCBX>i+#7sVG@d3fbACl#!VoN!go>PT3rL7s{5h
zvP;Mw$KEL;>yVI5Rv~-O|9$lQzTc;w=lA<x*SWf6e$IQ`<8{C8*Zuj2RtFIt-d#e&
zW(mJj_$LMHO0gWXuQ|?ZVn1p9bh>_pQi@Fgjmx+`$kmduMsp1m)^8=c&a!G(I_4O!
zds8M0QQaLkGJCNUiaXRfUQ%P)dJg@@rqFl70pzlT>C|7YPo1PFa_TLr`mwV_*2v>?
z7VlJ(r9gcpli&fM+7Tsvhzif84uTDYIQqtwCnaOEBeh;$nJVdosdi<CN|zXt8^%98
zJB_7!@-hL_f<yXg9*Sg*`R=kW@7b}XJRqvUwU2@BDB5(x;eA{znqDy~?`qm>*|#2s
zC@^%U>C{i*e?^S9WXNNUV#(3J$~H*nE9l5iH%=;jF{;4WNwY>o8{NSK5*1g@+hj-+
zwA?}I)3R_eeTOd2CqY`-SDLLF({{tuh<baWbEP4b#zt;a-ruE-X2X^F{MOyxilLf8
zeL900!@`qVKVuf8t#)F}0xHHzLW@HX#4jgJ66yOM7Y8_7b|yd8JeCAwk(<9A7iJ*p
zB^)vMcpM+5&?J~|<}Z%DLcvT<Bg?V!Pq@x9Q&FB+2E&Q`*`eRf5OAG6&V@t)Q$flP
zbDhz52Y6n&#an95*-%6!G^czT@BhpWIQLic<T>D+2qzuS7CB>1;TEQ`QrVmBs@Tfh
zi+bBvwvBWG&pVBGAJc$BTSDfXahtHE@X;fSGqJL*M=18Gqp?vZ5~su@m`#P}_8#-H
z?#`rJo=Dt(cdq6tf$%dWg<Oo<C`d#>Zr<iHb+6!WBSs)Mhv~B8dvnw7fi8kx+o#Ux
z6h`XDp+=Cg+G=HfCx&zaR!lm+zv@#uQP%-QoRo#L(I@Nf<hdTvlMigL(VeVJ9YjMh
zP4?N6%hIc6g6G{Df@YB~uaBjBmR`ju^=)dD`6;JYa1E(-TQ@14-ZU4;Jda^I93&AQ
z+$t8`>Mb6gBs;`vHYiIMHvN=5(>uu|@XctTarE7-eu<m8V+6u(8BIwrxCRSDS_nO+
zPFq2S_2mS!QtJ=GY9G`HWLEW_7uF!=zznjn-&z{?#)Ht9jH?VwQW6{Y?(4Hl@eHzr
z*jXSM?*qc{ve4C%4k=fFD3;X;0z5-9!WyduosCz#;78elqNiTxR1<C8K2fsQ%Po6e
z|KVtfWt<JO6ew|U2%IrK96dZ>Hl*=>UT!xff0t^EGg*Wk`yKYRL_uUd;=7H6@&@6%
zF=6p%Hi$Kx&ysqkRy)S+Z7_e>`;*P<r-?RSCHJmb4yb+)h&jzf7d2i+_^c0deTZe_
zaCNZ-NDLJ~Fd6?&#l!B_oW9v-pLSMG6KGp*bLLfwN_45Kohvxe2c%FPd~@S(9|IDT
zTQIumR+%=^B`=0w^=0LG_7RO4U^&|eG<cQ(;SDP$*qK8uGr}gmAyKE$<krt>#Rcwf
z)N|NN{8;Af=fa8Bqe1*0N-XcH_?<(B7iU1M{2bOeWRVR_xSiPeSZszLfSrUNtuk-B
zAwqrUy-iJ%r@!wyt_{BVEoO|kR<_6d=)>+Ox!Anp`*Dz<2d-PcIaVhgY+Ub)Z`3|G
zZm0VID~1!?E?p8ZxJ>w0D;FvhbORp;-PvPgRwVY(l=K#Tjeo81!-m_&{A-xLd;;si
zN0w_a-6`9cAFK6s9?&1@gq6OIBX?t0D-Si#4E9C2ThULmtG=V19=JBWpX-oiU^O+R
zKzVzEHeqM~<B1mWR+v*7q;PT4<J6VME=@bf%3eXs&nTjDJ=;gb4SKR4d$in$Th<K{
z{4ikeroMKOxzWer2B3f>J{yQzABBuA+TKrSwc~Joyl=Ow?gm`LQ?BRgQl-kov5YH!
zmJAkgqr@pI5SH)IZ*J{#q{J#=3ck)ZJHk|!$%ML9Jf3@7xHD86w2;qNGYMj$qx3&*
zb2O}edwlHV$MR~x=aLcF!cJqP?DPDW-L&7Cw_9&xl-nz#jH;=Qp<%S0i?^(rgZb7m
zi4gAyxoCPh*;s$F4CS}coyXs@Q+qxiiKA81CMG7n=^9P=IcP03*;~%~1OYMLxagcS
zB1TY<7kT-W-PP5fGt1?DCeg*C#pu2IS5DksV%S4s_|Ls-NeulEUabO~3OI~o?6@q#
zD$abP_TSqj703JI)<&~AdurBe@+0pBkTNMB6b|@@)T2mH^)YhZxWQ81^{DW|_nLBb
zPWbsPsKx0hp&$|5DpujsOd^p=i~GA&B2kM?JG^~VPa@^(Q_RREH@<w*V@s|HbcdNS
zKT9lX%O&zY7-E&_G}~|P<&q`(MRQNfzMa|TY<aC4O28$8tXE*8eKs%nj3h4fhoBYL
z;soJq9X9>Nz5TRpE$zYAr{ob~JE{6E3zZzhL5efEgF_k?LWeo>lM+ete#C3)Zz!hZ
zGjATcPhQ<PbfFyW-H)k920ygp!QWzr_WuoZ6|#>}9EXvZ8pmwBpE|P1vVmOBmiET(
zBn2rg_Ma~vlH{c}6nwil&Ve+%bFZnmN+xX5C2Ab9L{?Ze$L`g(Q)nxkETr}+^o^SZ
zl(4v?q{&P3r5Q7gx9U=GHK+Ph%E~G5ACfMWEOFfYkwNda(JNV`U#6g8I<@xc*p=+Y
zXYZmG^%zh2pPy(G^E|*2qptDwG$#@giLUg!?Sx)58`r|xG4qt^;dq+XU`MO<CPrf^
z((IcCjp+NB_NIchuQYd(gfs|=f4DeyUw!wS{TR--#*mFKX(Eb4o~Ohv9SjSjmoC!0
zk^CWiZ#vX{|Kcj4z7(4LBy<B(qP;RUKO0H48?ihWRV%{XMsOR9H%Zn`%=Mh?%5Qcf
zgAQGK!RNSpaadVGyG{q?ZkXk<dFowpaSkR-hCb>}V^z}{9%d1)jN_SM{>S2V(u<G3
znt6f?$7U}G%MWqb!taDRa6(51E^TQ(ZRS-yDz0ueKB^vd=kB}=e619|wre?CZcDl(
zfUWMbAz;TPhog=gOe>5al7C7g%Kgm!hsr=(sH&#E<2;f3_UfhK3uFcP3gvq&k(`{Q
z%y^w59bcA{5Ztuqd4mV+hcYrdqE43|owWtxL`8S^D`L{TCtpu6Z{HO~>^uxW+jVe{
zo^GiB5Y6MzcOe~}Ep*adf8N8MGl`~OnF%TX<@(QCY^3NBubUa#mi=#MH;bO8F5UV3
z@m~Gm#DVd({ri>eL4-!D+l%U4VsF188%sw&H>mh7tMJnpZgI;N?0-AsNA8DkN_6f}
z#Hms6?n&_AFCDso{~Qay%%=`u*n6NKDD1K#^~w{P6(0^)6bnTKUS;+}gc(lLV~*+X
zFE%)KE{w3Aa&@(Mcbzx#+>xd5Ej5zf^}K!vIEb!rDl74`J07tL&*3&$!Ph($;Mk6)
zB~mhDhn+QyM1(QSJYgJj!hI`Y=|jKx+Ngt|If3u^lM#7espVtD+*VY)LH^Bh!X6}#
z2g8zXX5Koqo95;aJ4hcbSZRFfab0ZvwCM-L_K0-uHERaF6>&xa|7RbUbJ#-_ibvA9
z^}h{#@UorHbJK+D7phmg7;^2qz6$Qtm_+|w9u>2<3Q@SXNHxV2#!2$XG6ea4-rAfV
z?ksSvX$vgV-5UNP7cTI6rJro(P%L)RUD$Pps_CF2j7j#e4$R3AF?Kp5+&q4q?1g#j
z3p1X#U=?V0aO*#O=&E$Gh9Zvit4Wx@UJvsX!-0?sllVT$okHKKGt&6;Qqy8b!?m8j
z8WHc0Ih#huu}`53=oJZGVE*@p7{1h0&zp2|6LXLCYeFBSFU|WCdnv)`T*REtUHq|w
zTXiVR)kam7U&QR?e_!5*B!&)%!&X73+nCL}LYEKqtfGfqIDNVMPRkvjA$bK9aXtni
zJi=9O_vZ7%$**j-)Booqy1}1Xb;tZ-Z!gT#8f;Hn!CQ8J1{diRx&|40rsV7?^oSy!
zA7XK3W)}WdBeVE;|CXSUFXNa3=Ha3XKUM$t!Hyo1A{=fq<rVhe0L)vhQljPHAOI=c
zn?Q{w0jA6LCIm$HKl+oDIWNB<H3T)p0{3k@GEVI;#@H&(8x&?-Qh%OEM#GzM_QH-~
zmSF2VEaX9?^+FV!PbWB^aK(kv`wfL&FU=48!g1yD%DscbXy^{7I3$TdMh>S&##=4X
zgEM7_=x<AidCT|k@+(mHuD%yII_u83K;C}_^K1;E6vF2-*Oy20l#n_TuPf1vEKx){
zcJ7b23IElE&9j(a?aYxDDY8~Qvn;aaL>8SlEFQ_{5pu?_>&Io{7)BCTPm<uF#p<9&
z5t-RbO`$c4hT;dRa-K$|*48g788G&Yh7n9yJ&Nd(xv@n+J<4enmbh}S{02PYGX7m>
zKF?4JI{dAR7)w<gY$pc0@b{MtUEOSl%qQH5BgX<0*Uy}h2AD45REKtsdX6?8KK@N8
zZwg-&hWq|)+UV~;`s?FECAho=VK{yW<C&)BuHzb<KHQW&;%W9bw<hx9$ofO#If$AX
zPoXcJ=$JZMDd{8A>ezMf+ch&-<!ZOfEW&QlVj5tac|>7%8qp*SjMxT@b!_OG7`26?
zV}SX=U$442T+Lw!0Lu9=+d+|32bF`)YBwioGxbt{oA1LhfyUTpP5ct${S(dQw7+i}
z=7eFvpAbC(+nYecAQR2U!n9Ll>D!VT6wB^|vMVLwovMz32MQC>xD2cNI0o+8?Qy}&
zGtZ2FVy@lr#La(OUV6;(He}l={joqJjIg{TZ_&j^rofR{v)GIrxDUs}gOh|^o&cLN
z8FZW<etAcNbL?1kA^s_}{`(GT%=iBNRkw)Q#|^k*7V~=)S1Ja`6^tIvoWr~QqM7QF
zAA;dQi`$iZ&KsB9a{s!}M+c&NW(G!3Iqz*<p{J*R6O0)sM9L^PwiaI_06XS{lK;2&
z!G?JsJ#BUdf4q;x%kVxDFL$i}_F#X1NH9{onBD70j9Q}FVI<fMO?V#^eXrUtW?_1x
z`q`G_LSf9E56k*+TKj{Rj6Z!QOEkXqhT<}2gWmnRh)b`D3$~#TMibg0AzStP@QjF^
zUu@6}8Xcq{%!BUTM-iMl=RgE~O#wf)9`zw*s+W)~aX2vg<x9BQSn%gXpG#Ri#x-Jv
zw-K5eVEpxuxABw#-o{e}{A2Y0hy?x<Sn=?{?>TXY5Q}qsyJKH=Ym}c(z~)<gfe*r7
zpAH_L{i|`$8+drN3JMO)!^gnpq%B#j+Txu){Q-o_c#Inf;zd1f41_6yb?SoN+ak-s
zb2o3^>`G)VIrA2dP%v!PKkdwIjGYmz9D?}l$PoVZuQxNs&M>ne&-~K>|K~FX5XSU>
z?SK30fZRBRmT3ztFnY6Ok?qcN5%XQ0;9`0GW>^qRSFe%^InIki+sD$Nb?pxk1;X8_
zwpV9CU0BA(h6e`+hiW72CpFvvejm7;tw~w1v=m8T(I;m7NnUt!V;r<W2iGP<LVvd;
zWB3@Shu@;>^ZVKT`w^*&hJEI9<YN`46cUpO4#h<5M%i%NDA{R(pJIM)4NT(g@&4It
zVYdwe5c6pIT4209`RQ!+Oml`Oy>hGA<HwKhgalb8u)^~h|J(EB9X%hKXH)i%=i|aW
zA6H>~^#ApIn0w~v`2T%AcXcVpb4LR_T+vEJyB<JWZ_%#d5K(as-B`ByBz_9b^}hKQ
zydU=%SU8$HT%|HN1-@8W!c8xvuH635uKoJp1I8PFNcF`I|NFN8{_#f~j6X_X#N4ki
zl9ZGsi;vk9?3dxcEhkgxFlV{?$f~M$$V*B2D)*NwC5cva+*$tC5Dvp^Sv+9q+A~<r
zxmaUBW54&Km|>so*TXCmF#G(k>nKJH`|Q<xbyMz-b(F`fqkL?9(EqWHNk{u!5<)A?
zpP_&0(0mx{(}=v$@wxY6VOMNE6MZdQe8=Ul#hca;ft~Ptx7sOZN9<sewI%^Y7Sy}Q
zbEvB@pjTmgen&ATWexxK3p`7otk?BTe$NH>h2-K{=m1gh>B}>i0HLZfhfVAa8nI`7
z=B48+#grA_kY6&CKheM2e*`VbV(c%N-A)IPe$?oV*bv@3e&DjleP`eEcvk5`Py5HB
zwCep9{Yt*PpxLQFhMvJs5LoI;Rk@~SJ8gYAzE@;*@&;c{0IA=~je#dUi<>lJLDSW>
z1%fiYmRnj{s;ujDj7}?y_}`lNH={)kch@d(?QNegSrB^KGq_1N{{xwBWQbTKrg$z!
z{vzq?&xD3@wxeq;4h2`R7THvZ0Q@8a3RLDW7d3sqB}q&S)OkXnYljZnCLEcj&U5RJ
zLAP!lj5g#!bWnjo(MEOvGyOjFezO3fu!tx>{G_t;hLz*Lc~37|@SX?HO&7v{-=(@p
zjQ5PZCV2jLBlGvu<}G&QXe=miI-ir6a4nT=KgC4W-&;s6U`t4SLg_P8g<1PD+f`Eb
z(s}98WmL{|F*6o_&OPje*dCd?Et7;VH#zo>mn@1sy)t?jJh*x3#(Z86U76X1B5Tbv
zue#H(obK4TO!{MIr8q98((Y`O<xr}*qRuY)CpYK5sOCt{m+i$#qC`!@yJ63njJ5U5
zU)|^@O_a7+*p}uoQ)78NdFz-1dYfepU)^%>*{9R$vp0w;m><YSpEdMDL^DkJVvN!l
z3EUP|;R%Sy9uAgpLKgz%ROgYAJ8?z=<g=dh+?PlvGIe$qfMJV|?|*nY4gvYS#*-&c
zGB7gYvMRr|akWYz!Z>?^f3wsF5@4zKnYxFVe_urL+xYnv$lLgK!oPq1_r*Pn@#6E8
z(#|Dn3EQSpxYHbi{?;>GsMDX-Bfodo8r($br=;K&mf()L5gp!~+ffT=xsNj(z>D83
znZ{N3`sIv6*x6;L+i1S6A`=qak3s2f@fj)Z)?V!!9u$Jea7k=U^V+PE1?h#&tL%lS
z>W!Io9Xt=xrxGrOLgXn?OYtJe69jXgacHK|843&P+$`BjdR0&IYWGJ>Ba0IT$b{y=
zm1v#6?0DoMKfqDAJ$^xKf9<96u+}hW7CnWGnhw+JD_RB`f0EE2SO5#if(#(mE43P?
z$h%*JkFhBX|7M$Ok8H}PYn!)zw<*LbU{k16o=`#l@^_!|_m9WWVBjc)c4bHs{giDY
zl{(Y)CRo(GtcTtFM~o3SGeyj=hd->P?8Fx=G|AstA3hn@=~PcOf5V-vg@+S+>*IE7
zAQs2y>b2`67Bq*uWKNfaB}^vGQ)iJaleey$4pTZU7<Wdv#Li_LBkp(E8Ovq4$D&m9
z(V*L1`Mv90O-djZ)y7(R`-imRQFH!9T9pAOVty;CxCs*|H5Q(=o%ne3{m>ykw4$7(
zMOHWw6EDepKG^)hyXUc6(<5Autb0q-CWh=uWMbY8OZuQ8$t*-x9kRc+ZDmiE_*b^=
zZ4D=5ex@Gz@F5Ej>wN%i)k(0|Xz54`upTOVEr$lOL4kG_%r1$6v2GA9ddxmbfg@@U
zM&KX8#DNs%h_Vh2vHyM)#qA-OK-#~vCHVdGe?OvF_*IxN=WF_Vxq8&|i^Hfhv+8Vq
z&bAY=?*qmvR>$y*D6ic+I~{>deQ*7xk@~lM_jHeizz_<tzJWd-#bhR|pds@YUZ(AH
zr3?2v0z<YQ{dCN7V2|W@5~^rUr?k;_&%{)RJhkUBnzH*e)9IO@tD@g}*md_-Z}2T_
zpV<vdXvhfiqPM&&-y9n$4?Un9I;Z-t=E2)9U6jo$Kv`Vk%h}g<^d#HCWiOe3ztZkz
zGdFTfl}nARm@PuJ_rP^%lh~#IQe|tORkD0sjU$m;^t3}~&HZ&2iS3`eVfm|__@=dP
zp^55@541|{yHrz7KKUOmz}`Fk5`n;JH?fvP26}zg@(cWK83i;vVj&QyBwbD%gnYn%
z9ZunH`I0-d_(0d+F*x#9X&z*J7lwX8sEssEopcA?6=*DHeftfSF$D}7lHAi|EMqI{
zRf2SW2<y)$|J7R4W2|+A<wE4|)|&PPSZl!>Uj&X~qyH`P9)YyPV!qq(+PGFG7B^Mj
z`R!4(+3|%Rt2V(&l?1CAmIXU5K{p~hQEEId808-xlsC6PY<{y<bi$&(lo_ja%ZC}w
zj<fvJjp+K!bAe|Lj{?9U>U`>_|Hh?RO`algu3uM2VK1^nIo$QzUZiG;ZFGA9xdNHY
z=tfAmpvT25(br1#JP&$}I0lP!klEl=FYPMDW@TM1zMXqL`qH>d``Z2HIEElWQIfVz
zD!~vghV+4)FS=P0`0m3cbR)MS>&?LY8AwhfGH6Zpd(&JYOFR=p_Vg%=zX5T?26Eg5
zBvT%Cz47b0_v2Su9q8?iMJ?N0&P75(LLuzT16rk~*KNQv2IBox03g^2b?e`f8%j)a
zQyPq{{lm(*g55^xE6>XUN+r&fWi?VI2nhd!5!HpE@&aH6O}5ANnNxaf%wMQ*S`2;u
zoES;qwz(h;5WU`|m%48;bePpD0P~wZ2AmCtu{+Uu4fw`1kAc#H-yVCN_oS%BR~`Z?
z0iF#@UlNF>e+%!i@SP(dMhyGr=z1juA6<wY?*Gi*>T&7I6WTv#KKD?57dB#(%8gX#
zadNypzsYnp^-STc)Bc8tJ<atGnQR3YgGI2n>XZ7qEWg;V-Y7{9Tod-N(THG)Xv$PZ
zRj*q%%m>?ZAcN-*j@vzGHz6PBLPd9GjtVGZf9>_O44B*?RJQk>S!p<Z-yRK8w~kj=
zzjzH-WvTw;o9Z5krM{>poBp5}<A9K*UWvZ<$KKN9+iu=ZM~j?<G0T34&6TP7BZtcc
z4tHPE*>in1a^<r%@SdkvBd<MDwiqhw;^a<M45AWsrvPGUNhTX7Y0)0EY*0BRIE0}U
zFgbe8Imn6h4s;U$k=eo+c`?~w0CYnjT;A1bEI0TgT&{*Yp~cO*!GJwVmk*__FIgJg
zEl+`2o`=^o;+lSu(syQQ$tK&N_7e03%EJ14%-pqMR>w8+S#n<-5#i7-!h=3`Nq{oX
zMBhwlLI@ZV96@aVy<T<9N8#(mVqb)lbd~)C*IFEpY~R}c0RPR^fh2lb+Qjn1tY*3F
zp~CYXO=xvbKI_}u5ghU8*~OoUz9J2!$f)sjhryq@#tK;H0vwy2Rtr35ZtB^3>hciP
z>{p0@ew2!bW~MM&tAJ$U0e8q%**v~8ofZ0OSsp(^`t1szCPg;%SyIyV>)ZmceP>><
z=2bPWSWZZVlg@6)oyf`WcxY{72zoK`Xjg;Q<VwTDE(6%r=>u*C`lbEP20Y`>rYfcl
znb8<bPOw+^2tm=k1xY{6X*47T(%`tSxNUeHzUIs^Y6xsmC{3oF8}!q4`99g81$3Ki
zE5^+R%5p$ySH1{{o%ch`yt2jp4dbIN0Z%^j^#yC=AD(=S6_XvXe%5`48!RdAI}9~$
zV<7R|d-B4m{PZ~vanRqYBFzu?n`v&An;)&E8FgO`p%;cJ%)~HKdi@kL*Q_Uw%(=u5
z_4WDoKa{69F3U&Y!7yxSIujfwlAbvw6Q(r7_ci2qGYNnw)IE?z`7RdE@A)A}bZ0_m
z{pekgngu-gpZc$+O71@Gp&R`KuK!J+)o7T$F+y+>p@22w@*Q`8FlNlA%JgK$RlRBp
zTlVFh((dzJWlha2tJ$wa30#mSU_0%5Bk*fzx{ghWvOT=s_tb#s^<49#d%ZmaivaQ!
zxUXJ9iRU_X=aMESWGrzMg#7Fw>}mCQtD0Z4$wKBeS`*gp=J_Fz({1f(hvy}F9$GIp
zMzMlH$t>FozYewrW|x}kS0Mq9h^~U3RlcNj7~^aRYJ_qxCfm;l<=(JJb19OV13reN
zzI>W|WJ^!0i=`+)MUr_OkSzZb>tn#!xu$F+`VTv|0nVQneG-3*<}sjV^a27DFvnAa
zkK1lFN2j!n7848ubI0{~Fl<nPanm=L^uUJ%U^J59TY+>4o3NPc8Xh$29*1dUDfwF<
zH%GuEOerE7Mos%C=z3Hmp%hw|Ay4$ver0r=r$W}e>gBBBy2<u-z-n|jpM(6{GxCp)
zZ|>1}y>9i-C|lXIeAZd~b~qjVAu_W~EVq9aw&fYx+~ZWnP(0&3iPbRcQ8)jE6omLk
zgbKSs6|u~wMTn|3Gw<QN%hu`J+qUEnIQz%G#D3Odck<G$^{YrY!LFQd+C5-x-xnIX
zlwIsZ@@4aqpwGd}X$j|oZ9;5y=ZyqiJ7!oROwkM~6YS&tXCtGE^?jkv*mx2RrFV=e
zAAa8)^5umR10M!bIb{q%F%K!gLozKVj3MpL$AFZ=M1L$Z;}Q*2@uqLlVjnURlafAx
zMg*3)^Zn(Z%tCS9EJDlGs=tsa?M^~qU7dK?v8#nk!tPr(Fvz&M&9@%a6!uK^{I$%b
z>>B%#hxMqgiPi6P-;DnF3s6`9%(dIfWL&FxKTnNE`=o~So;Ow{tzu+S-=grl(`8Xb
z?BqVuF`^F3N_xnc8y+!Y%$GOzKg-u)49Zg_1Ao)F5!LdaL)m;a7Rtx)ztv{Qdu)!2
zeHU>&w{b~i0v57-mbrnaYVf?X(2uxU6dB<$=6&ywN+O;1^x{$G`$|`wTimmUZR;|M
zQy7F58W#RCLic)K#1JJ_cq>*C2~h{BJVv3K@bckZ-9^%2j=K#jEFLB|t6U3r<-NfE
zzZZf&U~rrd>HgR+es`G8{!kup!@CwuhkcCX`D_-P+(t-CBHpz`aRvcPjy*Dn-*VvX
z9*5JykY<U+g_Y^{x#eaNR$-w^!BglC(xBmeR$*CgFDM)~g{1t~=pW%!N0kfX8D#}$
zLP(iBnmWcBTJfS!c#UldW`8Yg$`F)}NDzc^%Um%{(`C^lBrzxphMI)R+nYn$pBMTS
zH15r1@kj6eNF=CB>rV-XOly@mfOT@_Ow2Vezik9Vg~wws?$YrBwrygS?mtgt7X6rO
zQ=YB?-~0iCgj}&STSZ%_S@sSrChUeS)mlr~jC~YL?+puG;yC?1d@jkawyvkvdgA-5
z(E__U{q!khI$wQ<6+<0V)g)U3FtBl54kkb?SfaXkF%V-#ph<mNQoYvGqb*T{5|+nl
z@cn9Uh1KBmkN|+tV$0TD|1D7y!z{j6hv+7hLjInF(Sq-NH3Oy6mvz9Hx>xCFiM&1o
z6+&Ye7#RZsc%Z29@F~3*w1idmoF^5z8C-t^T#CRm&AlcoCi0v4Zacy9heIDd$~MZs
zyttdfUSDCJZW>5l*KC!O&BhYf>$<;l9fR-WJ-s9IwR@&Qvh#)GUr<b!$TQQX<Hofz
z0TR8@OkBfdr}r<nU2%#Vs%6NG+-K2Vouhusq}{5?fQP2fQ=%CnXF4_2weFPtVzARI
zNkVFk(=E(!eZ%9<lqNUt22uphWiF!S!qFO@{k4YK>k>OYM51*sm=vA^E3`DB7L+O9
zK(;I3vhvpDjEfdG_bh}m^Cb=<B3JwnN@q*;!yxwDMmoSdoWSJydsJ5>NS6>e*(>wt
z{+W?KaD4y-g;uGaLpjcQeU==mMW-@UEzE$T?pq}bJ(M(=>N$bTN{KABnII)_jziNM
z;2#$<vR{Q5M~G2EN0yV9s6?!&q<`P0qv|!KP;>zxI8Q`$dStU))#ygw7g;#b9QZ)C
ze1EZv_s-*9QtzH8g}i5HNfL{91W->CYv}R0x2H4+V%B`roM72V?+zqK1Xk%6%IRZp
z286=2s8W(h)AFj<SbFhbzUk<_^u|II0syTI4(r?_JUADFLx=ex1f^1B!(;$s&hF@j
zhSIwj$+v_8@T$7yq(B@(?QFlhm(mH;+tQpLM<>x|C+<2N$bXc9yqLkl9(dxEdHr^3
zmJBd)Ms4{_$9`NLs2g?_JyO}*naewc7yha*N9P&6bl}**{_YlZ=M{h`BQK*<ZFf)0
z38?RjrpOG4z&SMasYkAr*^>Ppje9R1B~!YM=f+*+(x)HlrU-v)2r)mtre?x;E^qKL
zoyvKE;{zIgh{wrHR)FE~F|h)&<$ka9>WCY^M@0M$d6SP>8Upq!I2aoXIjF#F$Ee$5
z2$i|TXI^0L(ijHG-Pd&<sc>LtQ_D1Qh<M`TLjqocj8(a8%wJ1s8D_K;Sq)zVZe!t2
z18~Leq8=zl=gqX^!}`1wh4uZX!xzFh{LQm<`@dxu-iM@^I&Zs()50k3muMNlOliRH
z%Rpp7*K_ysop^s>aL3XWTlA9x#Z%hQFxkq+>^{)NT8M40f%p>i!Lp^@euWb-0bM&R
zs_%&9qq#y(oVhp#C2R?CaV#o)G~~bapa76>J0SxhzGjA80$h}4w{EoPE<vGwp~)xv
z26nX^g-ABj1@_2XzncKvMkYOKFtYvz*XTYd)K7EGZlCAUeTO$QK0dBl>nZY}E#@(#
zQVgJiz;f>mE@+!ShQB@mx*P@N!BwY6YrVuUr6K749-I91(s^s?b-5P?=4RL6@{<H>
z_p`&a++Q1zKPPVaD2{;qe4f<Ln^04-g~84y9r0S9)DmtuL|pfcHf^xEE~yuQ6~TMo
z<#oBP3D;^PvT!9NWqo0!ItV6)-K-I!-3>>w>;Z)<S@$cKAWy<(3?6z#c(@!XTdr$}
zP81OU9cPf_x}Tez2+0l^hbHm%>WoH5j_}Eij-L%d=ewZIfoaq=mes3d=OR-LjPp%~
zFn5|KOEu#Ojn^Ith<^8?6p=LB9M2K5qBL(mO`QRPosUh_gE_5;(NC*oZktw9$SEku
z6Zd|e)OsPmSuX;maR02GC*{1@Er@`4{%yjbkFHB&H%-gpt1;q7F1?el0zchY>8C~P
zMjLqE>-Yp&P`kN1Xfn2VPPz%L?8S;s#P=WD{ll$(hr+y;7`yda@R@q#x?RJSx5(2~
z397q`lLs!OPw7V6n|2Zo_qUamQkDpY$`A1lcQ50fyc6H;(>5d<l;Gc_HXvZ&pC{g7
z@y=lCYy?*<&5tL_H`La*|4L2e(RWPIyqK!ZQ(ldudWFU>v<p?mXz}@Nz{<Ct!jKER
zGA{SIxO4S-pkN8~G-n|CH<`)1JVH4y8=17n_NXabFuH^x6dd~(LV-9AB#f(Me<asi
zkn=nJ%K0HG<8;v&$#PsAp|<_@@!rn6)SIoRI$Z_L@X_=;fp@>A$P!8O^q8f!W?AB+
zrGw6?P0ru+IiZDLGg~gz94P|%Ysc#^GiS!Nd?j6Exs=n@Q4yV-UfhLclL{Y=!>iV>
zg^OUM;!UfQlW8j@i+m9zc`&5T7iPIX3?=Wd8!y{44!rfTg-Icbg|uQMt)u)Sr_npP
z=3Fy+D|GDj(=ggS-_27>G~wE#7xbM$ScA}sgTym?t;fzs5&*43ejv3bv+en0<np&U
znCBLO%yM||o#l&2X-FT5p48{!CC1jYeklTd51(fXvA46IE37ECM;(c4tj+c)lzZ3d
zXX#g0(2oVZUATSrl`C`xJx3XnW>Dww+JC=0brr9BrZ8oFvi{h%MT{UaZ5U-4E1Rge
zO%-|i1nK45-ygajB>tXv(Ov{UDSIsT<TDCA=Z@GboM{B#eSxcB9^JAnb|4W=qGCI>
zI9T!h(XC3iutXUEoa{q1oCbKz88KMIfi2*#9NMMBnJa5mSE}!3%6VqJrt}o?&1FcJ
zHdjQPxR6tiQlEe1qx|N<cG$f_8_~(OSQ6G3%*8tUS_gzCqXoQO9j|!zzV-=%MNOD!
z7{QoVL$Y7y6_5z>$<*Mf_6|FNDbvU*`$t1%1q;#FHwcUOWaF!@&BKIfNVlLH>pheh
zJr?}Vz>)H7`!~3s5i7riR|)$MM*;Lihrj5Fha>`S8ysN9Z4dTbrbdfRraL?DrVh-4
zRD1i^FUmy>iQ4rj|7Y~%11}O%12a`gqGF1CXTTD$l1_pRqvo_es8)53{Q5JXCh=Ti
z>Kl~mPJ*aMMZ3p4X+@WlTq}7ZT0<mc;j*bOaz*g9IE6tIUNxhQlfFWo_tqy+K2!gw
z9MJXUc4RhngXvDSwrhx0q=C-dr}DFcgI8tT3IeWyHj?g8m5WLSnT1)XFQPnNuja5b
z&p>Qgx#9+)i%jkAf)I@MX`I!_B~A>XNWyzx?Y61TRvQ=Vp7+-4@Svhr((drUgGp&c
zkWNvdqK`B&HOYPC)|9D;Vn+iAaz)6uxtdH4mRhsBp&ZH|=UkuazX*BGO7(f!hGSMk
zWsgB)d%*q9&!1UfR-e-!ryoT9QR*y<0uR}@V!v)7b3Rda)551BQeG<{k6XV+m50%v
zW;A#L{I9xIe^AW?Vou^M6BdoI?WK0tklyPfg)im$QL-CS(`(iT2NFi{1rMofbjk*X
z98|-4AKonv?%C_o>+YvD_F2jvQQ2s_aCdn^;+^h}WK|Lq!{l!1>^Yz+NTzeu4N#wq
zm##kmI@)-Dh>hTN%yyTB!()k^Pn@!#7PP~)qb%NGDyhm2A9S2b`xY~FiJiNCQs#ln
za%xSNwtg^131AC4nuna*s(G)&X|mzZ`8SHi^ExkJKD}ao{QG69Gr*MjYfWNz86UsH
z1?T)u4wZKK$rv>4M<`@+>d49Y(+|9gO%2X|mZhy3<x~vlBd-Sk{`ZtHTIA(hFK}!3
zHz-T2M?>i(pT@8&n&YFH<Rj`J25N5&O4i8Hz78NJ)D)}ga`h@BT)rq~Qnsdk)h0)#
z<}r+GO-IdULl<ZC)ne6*gQ;kPwpO|IQTOeYSBH8BUIDFNzN{+r^g{iZy|!fN0{|8I
zjz6`l_BRGwj*mhK^E#T&%^VchBR1#hxm$tnI3ru-yi6PY94Icz;rzuPS0F`@><|a%
zf4%$NxkOH5@~%~o$oAPYc%?((|DA)h_#?-vJozIP^IwUv*~`isSg&M1s;D5<3rn52
zK$hcJf2H<N6q1|#E!>YHh`L!2a+y=|XD^h1hNFBjLIuSvJXjg8r`h5T0gQ6m3nWv|
zm#N!vy{I9d=3M#~d3{C2Nv-V;A9rqvL=uwEoZQ^cXZn?OA${bsn{2*+$$D}Ngp4y(
zDfzA=#E^{E^mK7G%x@en)a+<Mi}tmRh2fwxlzcIL*FE??_46bk*HMgUZ&fXBmByXc
zfmz;%i$mn>>IC5QHALN*9t21xA|1|N&`M<2`HWKKj7Gkx5x*%)IqAljw)(*OOnB?=
z+K-95*(IC7%TB*#SmHSMFdT%944%TJp|`FsdwCAW7~R+ddXj$dFNkLEx%DUoYU$N8
zt<1$}1r=CyTLEf<I|e1GWW~efiY%$PbKwF{opy1ona^Cy+eN?RA&6f)cDta3+tvLZ
zjU}b~(=BZ5X-~Kk<0sUpjeti`?)Rh|m7-KB!}5avnOa9N+asaYr~7(Jzi4Ll%gFCL
zk%A0I#!vh<<Q!>B1m&F0OJna_6pOTTAXaS28CfxE?&2(tnlROO+2#dNC5w?NEu@tT
zKBKb3ZQt1#VXOrzKzbAGsnNjQzQV8pK){kU#!d?z4vFA@ybSpfRa{|%<{$Zykpys2
zjU~d2Z_#d$Izz=Jf~{&+zCLX@gZ*YJT6%~4v9eM(X_RSvPEbBb<78FQ@O_FcX^fku
zfQhg*dn*d;uhdOCU>@QG09m2R4H~{`B(%kBg*QkOdv?hM@AmRvmyHbWS&eIiVE0yr
zx+N*;)Xd7rE{FljK-}B*CM!RjEt_64F4iAKZyU{RpubvOfmcA!xcPIikK<LVCJ7qG
zEu0DGERNN7=36HTSMODb<I8c0nUwSGz|bzlgk0BHFcB;$EV3RZ!hAyxos!!mpQ*2h
z^I;+bkeXuEqi*=|P_%V)a3}PM|I)D&1b2H2Ey8HT)VZ|I7if8Ws#t@QZLVe*CEU76
zCPf%V$!`_WX}YMqufgR|N|q#Dr53ap<s<#vdY~v{dzVC;JftV+f}m)4U>u0D)_kc`
z718yzmsONHd@cD*B2iJLd#9>!ouh>!d}PXvW97Ei^sSEa$(uDzZ#u#z+m%^LgH;ry
zqC`HG&q353fn-iHfN5{{I^Jd@%&y9G7^ldDT)X0*Jq+_oR<+14=w$j}OiK&GCkHKu
zNK}(8A6xum6aS|=r3%%l-8-|J<aVa5CdZd{e!6`ZHG8^(3fQlaN-MglUR%2~<6uvD
zof;UB`KR?2urTFQf}|2+sPfseJ@Odjmw_S{v7xP2TvW3zYOjf6Aoci*6E!4dNuq9{
z)(^PfE19S0>wjm#du4q$eOiuK4C^Yo`q?2%VadnZxh2-D@e#7bV=oycnl2a93VQJC
z5ku(-ha^w`{!+xYx<pyZe#_P?$x3>Kr~F>r%*IhxLw=rTrv)xT`t}@)UD%D7FZ!q=
ze*TD|q$PNV-)hLaiWriC^!5qU3Tkl4$}OkqU2d6eE3*Q}3Qk4?v^+<|SxuSd^q%QS
zJ3eWeTpg{*B|M}5DzF}<kso(H^<U~9a${#8nVWv}A3EBv50)G-3Gdv6>4Tx0hESFY
zEMx7_ly*MLs^Qx;HfC64*3&|*P+@u}d|&Db@2&U^+*`%oY}kS-1$|NZ*`+VuiQIPu
zGFsqRIuo}-e4$u)BBE_;Q6c&B{LdcBdelb-<rM#?p|KS!Il<v<wZ~79*<2%o(!P&s
zGM^fD6n4E*Cbju>PS`@VAPQ?%A@h))8@wgHuItXlhSpgtWgpI%+Y8uD;?4HuS+9HG
zkm#g-F3)NG{8LS$gh{xS!XeT;gOGDn1;jFZhC5B?p^2yyQX}?HuAn=$1=_0Dn{KdU
zkYvtBE|3$p`0Yfg<rW+In~YYuAT0(<6t^9qzmpCj2$FZ|!w{U?HKu;9c2|<l!WY{1
z400dQ@#JpMcb|o>7}Jb;4&~CQEh}ve(cMWq><njXvyzT*5E5=_vs<{Wh;ATLX4TEi
z&`~c*<hhRn^}I_7vcblG;D8oKC_u`6)o~ZyMaOqquW`R=>BeF>L}^^QA+N;pG?E=3
z>7V>bss7BJyZej2w2aZujs5S}9i#M?v~lwH%ANM}PPR)FC$<H&RXV$SO@6Z4NBlTb
zdr6`%ygNIs8DeFLrw+0uw|k>IehS;;Sf0GyiHcXFA*soV=CoRO@*z5K39gx#F>HE2
zq3%)5DV!I7gKSZ2ei!&65}qU8w^lU-V4~<>;MXZDF&l!PCl&3?E;%}K1CX^a)!Kgq
zWa&YG4sdo7XGY`D8|l-PFAB2=-&94!ih0#2Umwt*gec_!u)F^OS<yb`$HXY<f5s?%
zm_lhDB4!uK^cdrVjo$fpl*Ia@>FDTKH1m_$qaC+(wXeQb6+Tzc5w(_aZX-UUB8!Pb
zNJaWf)RORnHj1_Reic+^0n@b{g|a40$<D*<-T)jx{csh9hegFiIg$)6(Zw=hmn;v8
z@!qb4Me*b^3C~4FnLdW(nT0+z<45G=)CJ5iE|+*-^8q)_qK<0+0%dG|5P5J+;>{r7
zoKf7;;?u#fDBcxr0*4~5axPp+F+uvv;@5%{dO3(`(Ko;Km~{Y#AEvZX2GY`?8rQ_(
zp^>er*>J_}f4Bf7P_6P;KBLuc48m&9<p<Ip%x!-rO<dO(r8PK+nvpbdS3#-y?ZPx$
zOhbfPs!Mf}-qljo3lOTH3Y}ECT`!wZ2Q)Eq&1I}u#wvM)KVj#-V6%PbZ7cB`j^&>U
zJ47jDD?^qTr^+36SZw8OVqKl#%i(ot*lC_Xd*cG)1Y6<q+IziwZ|+{osvULNTU2ou
zLa?688NGkDCJKuqVNulCx904$RK{})b9&An1NuGFT_S_Yo>QOImqqbkI!8k>awkhz
zgG~$fL#4-KRcoEY>JIC4?l-n!;rHs-)0Q~i`x4AOJw<cAByjF~zrw>FYpVK0(Z2oX
z90}Km{Z^-WHlvXRlhFoO3_lZ9%oxgUVlsEnM%|;sF@&Mk`ysLpn?F}z_w}4qGC`Ev
z>RDA4JA5e3>RW4c2YC7BetYb{pv*vk$H2^qc(5XFxM|l|0?<L#dHtq+sd4Y93twUF
zKNqiU@#`#h$7b|D-)FYTtuJk~5~@N`+`n`Ak(ks+&R<{`1qMc?WF#SAP$wqtFNY3e
zd1-{KBAaN?JA}y(t;`qlrd=TE6IVJR*H~pY6{vh`8*;ztE3T$GocH@LyuM>zrN7-p
zmaSbw9<J;eMK66SE5cRMxkRJ9?`j!TlBFL+$hX^FmbN1o)KqdnbPJmWXfb+b;Kd2p
zeq)i<8RnM_r3$Rea;gS!m`yDZ@R^LZ99cV!pfcA+0oTilwY4s*mtnA8pn=!E+IF~V
zzaA(ouUO1I(nfoOMfEiemrK=N2b6}K)u1D;q9woDAr)sQ`KC{so<$B5bv$`u$fczD
z9}vk0*0J~Gu1_m?uzFWhL4fZn*3Fm*gErnkTBQlS(t@%u_eZrO##AL}(3Ra|FitWv
zB4`byV(pOp>lbXMU8-%yaloyo`yZ~<>Oq(+VPliuewt@Zn0sFGcLdo`5mR7N^6m1S
znSaGQRtd!k289#sHl+(9Qi}yBOj#l;DVkLsE5)s-Q79w(FmdiLt^$`q<?B~iI`t^~
znqLlmcv`dDw{Qax{pWlvsXZ4rr1USWwjU1AI|TB3<wm|QT@;ZzyoGcK=&gUKT$}ft
zUFUL^1BL;(Ro=?u)f2_y=2*Jq{&SkT=!e)Pjf0w;PnCKf%)YO<P&3NM)a|(~6He!T
z^<PGg>c%1#A7$w03OxPApNOsel-Sg+A4t<@aec#Q8qo#GJ<1nZ3Gue*Q8UsE%QG$g
zwwCts6{Al*idIYeFgAf;zX3<%C$v*!$;ns)*3>o8cn^Ny0Db^<I2kL%gl#|vh;+@1
zTMIt8LzvIHl{Ic?u?jbB5M1@d-0dYAjHebWpT1*!N1V-RAZtEb|0bJ=PT>11VRgm4
z@3LH?SnlBjo#T9>OKX|+NbP(tJGD7}>jvBDEaT3Li@`B0s~J-|I#-j^{bm5SH2(29
zHPG;Tz`3*IS}UA0j4Uj28t>>I!W;-nm9*PeHj@CVQDz-h|9hyZN2PYDkG%QN>BSJV
zaP;V2DFc-m`AGe0(n57V3pdh2-hxN#4(Ur&-Cs?rDkomfO3_a!kUOc&4VLB=T0!o1
ztu(o&?R{^<iL09X*|M^^(Oa%--|vS?XErLVPvhbr?0zepzdzX=t{(ZG=>Sfqki)H7
z?cTnfJ=dKz)(Xd;nf{$=mnBL*GMu-~|0C6R#_&cKSDcC};A|~>+(?!w4<i989r8P{
z7kCi;GVVf7ucjS$imjha7mG=Y*TsqP%F5I<RwSJ4M7G^-zp~+r*yjEDV>Hk3YmuLT
z?Lp!6)>x)#XSMV&%NvL9FL{Dr6;rhG{VWO4&lK*<Q9ZT%P1}I}00^TG_NoYx(T7gh
zvFhwGLP94TES~NyB0^SX%){Rc5hT9Vy3!w6k8EF|F@6|*PVVAX8cwN5zxGD>bz|+@
z)L$+LqO=tTl!sjoW;tV)d^~=X8wJ!;d|5`m^<s`Bj}prpmDs&`)-td+v_s5m@e*V=
z!yh717?jn!@)zaTrLn%go(G0R2UA>;=sWWxlrXa@B0Bnp9i^q=SX`V1RY*{KE%b2>
z&;Kh5MI4F2H&p?a_=@RSL`Y)V6%$V%xs%6V#WAUpWE4fpJ-oIAq4SW|I^Lklm>z)H
z4e3qgZr)X4Kfh$0g)VG_eoMI2!&#smt2aw`N44KGefR@hw7Hr@aj{d*a;4+Xdgdhe
zH<Zz;MVN7lo_ou(t1(Gl2U;287X8AU!jKMVg<Gkl79PBLvk)TKf8}yLs@klxr2hGA
zDF_>0Jsi+CwdQ=UhTi@hB%XQS>3iq<AwACY*;k;1%><1u;pObW8-lnJTmAG>D5Jx=
zDd7f}%w4g$4%GzyhEhPhF+4i=OHk8Ib8n1cYnt01vbEnRGMR>kxnyXaQJ3tAoS%xe
z%aDMMXBdX-P<42)D_?BO_nTv-#YSV?l*@ES_B$rwprXpfuNSJEFO0XTR_w}OH|BIH
zmrjuX-p@vHi<GG-wl%Stof?vZ`g&Ura{RA_f;!8Z)!5;-n^Ynp>q10E-HD?S@*O}&
ze2*{{m3Tk2NGI;V+K=G?4cE|AH+B@ZN_<NYEn}A_YdI^e9aYOf8sN11c;YK^vPij0
zdq;yyO?5kc>!eGBmj=Hq`5Qe}VXyy?fl6{SIjchBx2eM}iQTK;-JczeDK_yD{6ZL#
znWBOkP*+Qwiq>|EN>X5&`z*bJ;lTql)v<ZgnlsM%G@U}Pamy363iKr**QanE&^FU-
z7(8JN(^<tTJSd_9sbnLFZqXUgey#DS?4HR32ibBSZ!%R8sx$GIXyQ&n`Oa$N-)M=X
z{_?;gbzy_x*aAo@ZK&56rUd%<FEGG@7FuZv=oEsxSa9FcLrh=79{4p71wB|_%WPBb
zmD~@Vfc|_uLc&G>gd&re6l2Mtr1hlAz~Rf=S*z9Q_5zr1g6so+yAkwkIdUiUl19N`
z*e6P2`(L_Vv@G^ah;5A>^LLJgijv+$>PvNpk$WjJH%+Wtryq?A-guF^pm^xj1W0SV
zAW}xK2{H+lOue{RkpL}D**WqF&WUc7SG`{jAwp}J!m+!Q`XfU)H`>qrkv)t~X%S0H
zi}aFRr51L+8-5^d5;-kDu4UxDwL~Yi61gbv&%s3=q}V>9O>{|E94ayeE~^hx-Ru)d
zWZgDp+_mfTbI^ni(#se&{&0l@i}G9l>E18s3XrN+Je$ehlvSREt_^;9n)&L3w(oeQ
zJ&Qos&k(L9MP)ovQ>e9_I5Ej<3KL&qR~6o4g*QTc83ZgWI_S^3qZaGP_Rta2clNhe
z{!wqwNeKmKN2ia!`U88txS8zc+Ft90Gf8S^P0{SFQefyAoV%d|XR=(&zfpWHKR$BD
zOpkG4He~QSkMmU%#nv<?ELo{4oNu&&&DqbM?F~Pc`$qlDBVFshmdLYtC!<mBaCV(E
zio@--<yxiXpYf@hMV`;v@4``Sr8iF_k{qGZntvncxIuc+p3<rZ@hdHoikR(nrO?^v
zVviOQ;@+*q9Kk-DNsgl<_y~@mhm>>Fd5iGyaJe`>=zP92ABAM(HuO5&cZ$E^a0+b6
z`+&2|wA}W$IYEs8QV80@2kn74$4{}8e!i@fIK)}-@#!rn8u2UqYy9}|D1Ib>)#8M|
z<-%12_|$ayueIa30fiqv*QXRoU3hX+df~(gme5A#E9Wm`4w;J-47hB&ua>QeiSxai
zABV=ueV`8}q&b*YaJ^~;73x4?MAjpG^j|bjYR-&3eGz>l{i25?Po6xn58-!sQFii7
zp+R_WcRkWR^Rl_+4nTjz_qPD^hpJsJy#Ng$#!$R3dhwvgIDsTFg<;by;R5G9m2WrP
zw|LC@a?g;=rAo`@J5CeaP%$04mZxnKLU>jX4TZZ=k225w8yyKk>65diqn@Z;tqIog
z{)fvMS?30CoUcbg!=Qlu_vbITWTWxlC9T8lVn2UFpqmPRe|Gg-+feAvzzrg0hYFBZ
zR6kGoHFgfFwaSw-fXK6)m-`q(+1FLLC-QqSL;M<esL;y-L}z#dmxN8lWmWgz#zgM1
z<E%XC(YeBA63%ro#jWJ#uzYH^i8e_%0Fl953~5^J_Z6_B<U7uYu~Vj=DyYWNmyt4@
zSQzE8Ss|)0L3HcVixBN)UI;wjMs10$$tDyQhaxjCSuAaB?x{y5B9Z?M=`2GPGgZQV
z^q)wI;vnTL(7hb!ynae&lUh1y8I|xYXO6hwr2=fNBfiWX2~3=N3DZR^>pKHXT)53Q
z1;tlz-h!xTi^5wdm%HxWwS9jG-@dVM3p~I-h^2DkAQd3IY66b*A;zj9|8CX1yD%AU
z7w^ZdU>RB{?RF8~s=58}ZCY;fT;^9`x8P-jUgXvnZfDgy@`M+ZJv2@15wZTmoZ3Y^
zL|vpOj9)Hs7_GvH2TWJ;{6KN*1+wXyg>#9fuc4@&0<y%@na$9Z*#xFBhwDl-!W5#Y
z=P1VZS1l8+e7O!ML;x=tnWT5ovWAIc0-k$iiO*BkY@?Sf*0#Q={jgL!P}g9L=Dke*
z!Seb#>QcgLa&9)Kxt9L)7?dTApkSZ5@`@wL#Z~8EZ@UYHP-<SiK=Duy|Mnskkee~L
z@;@)bIJm3ENrWQbGP9hLbC-r*-*d1xQx^&}ao^cwf7okYWR;<Y`)%=ry+kO-g=`7W
zO9wbA=W`p|lKSGjtdy`iVz}@O_%a{g{_$j9-9Y_7e|(zSm!sd?lrm*=Q-$;{a4x&D
z)urDmCP+9H&mhLR^F0o)1aBW8{MY8+m1{fqGH{K0Y!SJJCHPg#qNVV;Z@8Ts0qgd0
zC^F9I;~)w9f#@{k#!c~bh@%T_4Ui0#LW64S2l~=0t*4v7uInVw>F{55n@HZCz~q;4
z4p17NLiYgj5sc|cUPgYzgO4qaj07PViR%E32%pu4-@2;%t4vJs3X?ku3I(^n!*$Q7
zM`;T?kRxT)-T)_6M}Ha^LOJzFO-%}jI{Zi&WCYYabh&tdKr7<7F!bn7HM1|G_#wC2
z`@aY$F%xB8hc(LKPT(<G(3$_O@jHIHvJQ1vARh<Qh+XPVSK>cC>dMC$>BPXm5SfWi
zn<gCZ|GxZb<imet0vg(kfViw4y_B`O3rcy05IVJ_?w7=JpGFrV6vFjPtitowua(Ce
zLP@1LUJ#YI@Mi)CQgVKXgygC(?Su8m_e1?x3Pz)wBlwyF#s3>3I|knFzDT}Q+xuSb
zV$4Ugq4mMdPLD4H@d932+vN<{n!Y0suCuZ2OS#{P;CX!YPUh*u)h&bgYWp(^p8Wyf
zYi^JyRd-x(^$KQ7B9nZ)iK2HYEPc*$(;bDyIrmh=gRsA8gdx4aX4p~h4eF7I>~6@y
z<tJHD?`N*F1Il_0I{98l=I^idBMFjfI;{vfH{&@T(ULOXHcXT}aOXDK1n_q3<F;77
zW0NPQji=(ZUHnOWg65=CpOgeP=UsV(xqi`}41M6O(hw&W>z|UZp&oY6a%jz|jz+o%
z#xsDK5Y*3IfT}}}lpqa+;(2i6=glund1AyVKpN(kR?07a+N75UQ0ELf%8fDZ*IyCS
zd%<TKRB+Kx90=p>q-nfY0`kj&>4JD#R-C8%Km@Ah1T>;UoO9Wl1vsE$)Du4lmJP2L
zayyB;&)1{HsV#4Bh(7G3MV5YL5e`abQ`3|_QPKxZ-*1}}G5ms$Mj0cqS54*pbRd*%
zlN6y7ZavQ{og7i7XbG}JAP`o$J=$}%My`HkH%3*nbaLX%_%@i9?-0quAU7j^4e&X#
zxw_>`<IiZ`K>Gtm0bF@s8|`vkG5dh<>80_$@5@>eFI&EXL*qS^>u7{|q0edGdl?j|
zYdl`~oO-%3lsX6sxC17Or^dC|ocJGo#@*dq)abp3)X;G5eR*Cs$+&B~x_K3tLrp-b
ziEKSF>EgE4={Ut6o;dXc4<Tlpa+b2`*IoMSU+-GDsPo*fLiKj-qp_$eS$MQbZHRzw
z>qw;0A&)UY3zuPAPWmD^ZM6tHpc=XKi8<-3up}f*yUzEV9k~r;Q6O3)<zEYe0838b
z*!TBwZGcI;3tB$Nc{H7qI8H96N*j621HlC{6E77Od@;3Dcjy=p+SQgRr%@t_5pv>!
ze#)O~cF?MJ6XMSJOEbCH5O3TB^;8TMXqA+{K}PnW+x806okZbBOl|*VK^aDHmT5bO
zhF)W--^6HPxUM}!x?@;*c(qZAoHqx~J~55gmUplhEJLm8G@c_Nc`cV%NYr}yDoU<x
zB~%pPAh~e*;~;R8<r3k!^Mg6;G@7&-%oRkhXND>q6ao}{<y~YL)r*6<$&;Z>5CYL(
z`I@bb3~fpA5>SHf0$bCcgQLB-VL{nX=j{-bqejd!Vc*%&udBL&DZ&nf#0M&~j+gZj
zeu$K&^y$2FL8&+!(8Ean^z$S~@uTiW;kJ?E1Zj6>5w_5!^c~WA6~qFl%RTH&#=esz
zniI^95t7gF&bR^bN^}>DFPqv^P>8*jL0=a%xzG;3>h#{h?&21df@!aoPaSf_eEa;N
zGtO;n;J*$8JPGk2xY<z43#(zytirppiRB6y`rJJAg6>fxlaV}FVWQKR#<iPZ&MXs2
zPNA=OxX*_{lKa1Dme6j1xt{r&B)Q;?d(^VP{%EH|7XnG<=plSe)PfPRa5@sQNMxk7
z8wMJx12^lNv?%d3HynO@RSmh20M(0^E`|Qpc`z;avwg-4@`cZafC$q|!7Pp;Dj;Jd
z!3a)rON5_^(kb*TTM4N%1msujCgr5xe(A1a5*Bk=!G#9hYZ_@$tMri%Qg@2vI)P@`
z1Z_bPobE6oFow@Ec0uGvf9aZ5z9F`D(Zl--3vv(oVx1tIf2W%9Vq2rc7va!g_6ZGC
zoz7x&jY_p=oPGA`M~-~Hn2UqQgIjviGUKg?^@y;kCZ@bGBMjvY$FQe6S{fQ5y~x}z
zGdE4P`cVjW@`_X#mLMmHhFMiSN;BV;aYg1T3d2;rMj+8zaSznAGBfa4Cx83d^^H`3
zZaLM=QK1TB#ExFIOmJDORjj{<Kq<i9l|q*T!Y9Ef=I7%8pl)%3QxxS8HXru2bomUt
z?9KB3V|VKKF-FsRGQ5%&TLZJzrvhvEPx>K7TUnis8n$XMID`jX(9r*5?Y+af{@eHQ
zk`hWNq$GQjkz`ZI2xad*GLyZfi0r++l$o8qDaq!QJqlULmhAbxo_gQ!&--)#?)&rC
z?>HQev^~c4xUTa$uk(D#gAFui4DhIWpYz0$83jfCJ)=fYM{wmb$}JnX_vU~zQ(+#!
z$j7-Vtho;9>(EhtNdT!;Zc2@Yb4=Tfx=58zsvV8xIbzq<o$)}0#1HRArzW#%K@t_d
zX=|A1kh&XigAs|914HwzbloyCE18vnvPub}aqP=<49a1?E)caT6-sXtNTew$k3A;C
zx<syIWZcKMU|t*2j4K@8!OCoD!O;_8C;szjrDioMpxbp69%mw64Yjxg!J}>^=R2<M
z<vZ_ats>vB3@pOpzf($%HCeQi6fE3R2Yatns?UKzg;{3FSmFH;HlG%ZvPMTiWB(=8
zGznQtf8uMdrhBY6m0U-OD^O@=x|6TNj;b<cb{un;!H9w3@zEjT;iB+NLueX=V{%QI
z!|Ga1b0oVnRjO=_#mVHOnW77LhyPs!1<^?npp-zupuf8Q!}@!0(+Z6u_F=<8eF(0L
zi1T1|unJyrT4ETosn7oZ+0+Vb{|%N^(Pd@VnD%a}jB8Sk#Ma=o@j{8b&VD|E)bL@3
zf`hV7{IH6#veS7q5Z{`hxk7OD>d&<&{+;xV`oNj&@~{f#f$;{~Y(C9!mGRX;$y)OP
zCN3<vhd@s$B~%n|jUtoIu!$67;hZZr?g-;~s^jP4Q2w$twfF}#4m<BBXP4(X{3xR_
za^SX5@1${B3|6NZ2em{^<p8H><bGLlD9oWZTuSWb<ZfFGJ|v}!E}eDnI7<|gu&3^1
zlGm9pbg8_#@DCH2%a7Xh0EEIMyfMy_`91+6R$46DCNl4-I^|e@Tl~VZ25zp)Bj>?n
zaz{{h`q!0$A0Vx@#l|K%a^yAp^r0$A7fCHt=YN<$&gt&|CnZS{_A02of&wzvKu;72
z?|KR715p4Y>Tw7wfM7yx@aA<4^(<XM-5{7k8s`Ve@aO&r;O)SOip!r775vMvqlWj;
zPAxF%JNC`i6bZuknoz0OS$2_4-(m8uvYTVNm+6nbbW_Mj$)31}7e;%)MubyqSfjWw
zf1$Y>vyi(mNVmw~?80Ea1$Q*RkEshi>p<{;a%R~UMYOy)Ik)ojGNP=5Z|{I}m!X#3
z%m?&K8tJH3{RJ9bs;Ua0KIRnC`y!Kd8!Ka)1W6ffyivFq^@m%UEV}htyXy8i6MWwe
zj-9^nmAK{_eVxjprSyCCDixu-3bw5SZQvfv#X9(2gV$0?I+i1*OBtDMDLE^ZEWCRG
z(R&#z)O}j;Mhm3Iz~USWnZa+>ILzxfwM!PuHUFfxiE6GASQI}TEb~1!7nH?YV>Tbk
zqO`Nu^X3{~L6y!JYo(^6L_w5q+Ld6e%%SB`<Xd7;rV$sz1<GYcG*S^wlroP0*gj+{
zZ63m{F_c)Ebb<91L?<!tU4m<M6lt*%1%nh7NC|m*QavR7o6i_fL%aIa0c^be<jbje
zPUGUex^H~^@LuG8^+&Gl`CgRu`g#Qvp^|azZ?j?@zrtL_*ejKxrYk51eFJ*<=Y)}i
z5*mrlk9;0V0DZ5UG7J^>F?wjevrh7wCRFsiYjtDT_dI7L6@BF??FEq!$GUvmd)#XW
z#+c-SxGilBH}!s3Eyj8M+<vmczJ(`~B7pzVx8$uf>S1dPA87<#P$;5}M&Iq_Mkt5Y
zk@%i+JXg9rb&*AI59{cn7wNQYi^Sw^zR$0oH?3~(;<(NTwZ`E*rI+(>BA?12shrtZ
zls@~I*GVGXlbHh;?=Mxu3!{^?YkyxNK6Gr$F0mf9D|26^Qcb8xTT-9G33bjmu-pH|
zl~lKhuZcbmcE~EMno0UO#<P#0{m+>ZDjPF{EgqqN9cF3HUbQsSKn_{HGVjQZ#u6h$
z0ZoBs9Oz{u55ELo=lKG@cX{e1&EK-1&f0;r<5xHTsu(GW;-m{gSx9=4%gf&|EAZDq
zL8maGK4?>Z#xLs82@DESVrqwCOg(&qHBm78F{{YvW$R5L$_4WZ&Me7wI#O+!xFP-u
zl&8jew!5l6XrHvtp4E_@QOTxaA-Sei<JgrGNPIaza5@*z=A;4O+KpaROt%St_#o)2
ze!&lmYZS<|YRd+eW?+1OY&YVFKu}PYa~`M)ZqH{@U&H&wJP@4$=`}7WlHO5Wj=?>Z
zzFtL4y;%#C&ka&eBY&zN^m0kVKV0y4)#~XDUC1q!afc$(HuI=GW&|W1Y&zz~2aPZ3
z>+2&)Jp#{%XtMd>d-DSt=!hNz%z#MB>(i@?KRiQ(IyMBJ{(G6Xh9GXJbJRde0Yw;g
zI$dOzMwPS81Id)GBTqY(K&~#s=kk|8lQV7&90iuj=khOB&A`X1$^K7sw~`{l5O@aB
z0fJH=RIw|ioy)n=1u{I3`Hf0Qb$dERe`*&|NY6?7iJkssuTNC8?Ky16!oP(j@l>cl
zrav!^nL3l3#n&l9;?Q<g_tNoa65*AmqQz~8{KBlhTQ`JXDxs8MlDJN!FdFOq!A^>6
z=HopjbS$ZM`>bmE80N331#<cupBi-yT2^^DW{3Lj+AgCM4O#giCm`lfU4P7~Ya+fr
zVD9zJD@Z8!lKxlUhU5UvncZXMgHzGAreein*h$y+cXn|t`P^>Y_u>sme5)@PT()Gn
zcyR2z<7BVCr+YT<obSY^@8;qZy(Rf{riUvHXm-18^@pP_qb3ass~O7OlG&DK;jeC+
z8vTn4a98ws-?I+h5wBUi%q4-~I<Iezt^!y{z4JT^fs3+U+X)h~lOtmb_nAW5Gx{^>
z)(a&YIIJ{Y0Ka%XDkoQAbF}{(AUg-ZVo=H<>u{TR{+#F7gFa#vyk%a&6o!VfO7xD{
zLx`odKem#IA_ho&0`=}%jovnD4f{y@TT?DIx&x4PQL!ggGn%X}I4{ADNgqtR)!yrQ
zn+R-;$JX3;uK%AtGq<z=!v2KGj4=2Y<H%+Cy{w({0r2IsfA2~+!?_J)Fo{V9Q`Nmd
zwcMD`#sr##l}e<<-u=Z=(a*9+lMssfFDd1@Ggx?_v5{d=q*9_9{8eSZB;2-)&5H-p
zB0`1~LtAfeB_~xLm-4>=ke__Qko1xQ)7WlS;GG?9iVK%X=c9G&-FT)$gkzi=&=Dj0
zMA<B6q9gw5T^kC-DQtF^h*uZWX|u$lkZ%Ke0Z{xLGMH-qVAQH)pE^1=cF7zTYZd|=
zOu<8|IY9$}kZ812en}248~?vG7RBJ`$NM$<BZLZ24{TFC{|75r+KOJbxrscP-+AEI
z_WjpbZtcuCeuYY_x&0H|gR5qdn(0<!VR=Mvm43vOm^MlLNvhzFNFdO=<cJ;uXmikh
zU_x`yw*EpYQD+5b@<$o$qcQEvrBUXoESx`6E|D94sh(Kt7-=ADte~Y74t4e|(v;Ji
z3Q@$H`_lIjrRA~Sd?qTnQU!e*pS0F1x)DEqKxtWHNF*5IwV2+>FX0}qWqs`zm999d
zxM;3x(GoV3+geQgfntL1%?WNZ?dg=Fqs%5QrMxn1UMbme{Mz<}(uqdE>(?{33wat>
z86#-=xsNv=9H$O6>ZVC97SDxq2Qo(IZR%Xrdn)Qed?#sF#wF8$%6$<uUXY~|xqwK4
zT=@Qs0hH6cp$=u_Og>+fsZwMhF8<~+?-wFj?IP75QQYG114&jJ0u!`Y1;8ClBYS~m
z>TUMx9Tvh0Eqhj~DX55Spq|cUl{Qu)kP~+-i%MhV=VcZs)UF|>;?=y9eT8Kp7O2Ca
z3K`Oqz{8YyfqAUWKC^TX^BcBRExR|5_7)_5WR0Kgd$xhhe#ILpWL1ikvlYa%rQHxl
z{hTei7$+VMPJk`hH-JduuY@Dg*^dV=&ntLZ^YIkiP0_IP*J6oq2QQJ-gKCi3l<uC%
zJ5NneGN7EyiUsmP8O1x0;}s~qD!1e;)H+A88*!}bEa_mun;Wy(x(b8US(5vi=FM9<
z&l;S6-5Uu#zY`;i!i^MRTQ6}JzLx-rN~{~Jy5=74(<)1m;d!$K)5lEp9j4JzMlmj8
z2SfzAs0kRXVj>C~n5t1qI0iU|OhqI1_r_-uU9nxQ2({Tp3>V_r-Ds_q(K1}ykKWa7
zXR+3+Vp%osbg@c9`Ej_<Jf}Hkbg@d{c+6=+*%KV5yckmmAa<kQDLE07YRYF=3g(UY
z{cp<nDL}6k^~4}~363Lbo?*sR`~SrLj*n&A2F*XQ5*tf6IK|+syXajxxeC}w$|D+Y
z0a4630`PFRh&w}6r>6<leAd2Mj^hOt?sF^<w_h_1S3N(zzs>vV&=wQ-l5R0>soO5b
zPwnhCnOCc44;*eMm2}Z!lj=<0w`WMEZy(pG4`?bKGv%=VMBukr$`p=UC(P=9`x;U9
z4ZQLmQOsX`JU1D7OG=d8p{>Lpk$h`>7wcHl<wN$>jGU6BY{HRVsu0JIPqK>Y0^Ea_
zNsc8JeQT0pnwGKAeU8bW<#m-lxMfGqhnP)|G<=!CTD<j2#`vdMu=Lx9TN`X&OE78X
zW|s2(e%2d!1?9Y094F&H>Apt-jyEEuw<L!lB>5e-I-t~q@hHc5CeB(j*P!tHa|I$E
z04Ulw3W!m&H&$T|;);uUZg1`&O6(pm1Yam<563LAOeYaVWZ+l;OqZ7<saa4M!=9Y1
zTSr0=2RW&gouKA!)L~$fW}%K?#`=5z#NPEebSRKNnTSjVx*+Ard#jwSG1CjerQm#q
z!yy<^-P{o$W!`@O)rFn`UB1@>-!f!KIup2M!lN?aw9YyKsYe1j@u`w?A;eJ*5<yp?
zI>C7(6-js3b>kgp6Do<VA+~+lcASOzl4b!dVXo5aJKP{;5g6&;^6}$G6SIl_r=V}q
zUh=jgJLNOn(0f8PGmT#P=HnW_1D>;dJ`Ifbw-zHZv21T)@K0qR%HbmB8z<J++B4e#
zHtDSP|L}AbrcXx2o%uIjBp~M(u0}x0aX-^nboV@xUqnb+gW}(ZRU})coK%-9tW}DL
zHC$I<xWf(VbgjhDWSH@OeDYW#vKm{M#s#7filva5SgBF1K+K>AA1*ik&i|6-n7c3G
z*Iwo)|K&%Pw&PjZ>^Ol6b#)`@Un)_wG?{pTPKMjFnmy{{=ADIO7E>=TWT$r8ZMSjR
zYZa4xWjnBP3NB<O63gP;r5JD<>q*zJ2_?cJYu2wS&-Iy`mo}cWyd%E-v$Z>7sBKe*
zO1}CETMO;_DoWUTvg5N`QjW8wymL(LAX}JnMLHSr5#dW7vX4oJB%S+_yq=co<Fmyj
zv>&*5ZcYW>bedYPYhjyf$A)@Ml7bFFiRm`(#$dX?e6FDuozo5eBq~gu)j1<M!!-t<
zar*N$Rn?o}5G={&6oim+okJWMWZM)63t<*ZL0OF{i7!WOAusG;Z_~nZv=ke$Of{kX
z4wdgc1KAXzy2i$l-LiH2!##%!SJ_2MdP}Ms{F|WLBm%W`IIRqEhHG9m;Tm9jL<Cx6
zkYm7rq%MJ-Z(_F8tXcRFZ0~YT+Pn=4(c2g<Is?XcicE}OEo@&w!JL(p0!sJ`gec>B
zgv?2V2AFr1ojhC?j$eWM>OCfFDB+lF_W&b|E{f=YRGmrLuPlN3wPx;%a)^m#@{?Bx
z2@Q_Ql$3Ofjb2k_ELv%xQ<Wr$<LlsBIHG-Kq*0;-qi4>X*<{P!z{SN?Oh<fdzUGr#
zC*0<<kjj{Yx^*h?X?;)Gt2||>{EX#Yh3Jz<+4=w180#r)28e9&SReoKw8KG)erSul
z>PHe##D!v`GfB8?Emnd^xsvKLhBYt28GRaFp%V*nZ~X%Ua*RJ9Q{8SaIEN67<Tg@V
z-+or`0Hc}myY-}c)6DI*h$kxQA^t&46QPYZo0FHnxg0$7j^LfubiT#FtAIsU%<LZ!
zSH0IyRJ!*|wT9i4hxKLk*%I;i#m~8t567S3q}{n0`j~Tpr2W-12jS+guP=+mKaeLf
z6D%%S+Jq#TY=4m)PrJ^7_!+Lr!{FL^h3_wevcp&&_iwVEsV2Sls-R<QF$Sd#9s+o5
zuO-Aws6}fl6N1vH`DKjfc)Qvuo>oUW{@#MrR;%*x63AAEUiO+$-K#(D{)%&pAXAa9
z$fMq8yP$*T*KX~NYG%Q&t|8^xmEH#GFUWqh*mA|L_@=gfvr?IQ1_Jazsr2lclWcA;
zv7h~TIjQ~^J{6lxea7cd5)xav^(^_<T4odF*Zq_#k0w8y-OzAgmq24;AzSjC8A*4U
z&1~5S{gW}paw4Tufycr1RK%wo(o70*oFSd(b~|hO{UEqkVO>{sB^%Vrq-PbwPmj?G
zWJyQkP?70Y+MI^|NF$OqtbO-7mZj^}0ft-88__cv=DBsCgiab#5XKUX^36BCz*(r|
zNu;lwvnd_8a<HJtP`m;MCMu_uGwXa!W=?6b6M@LAwlfU<A{C?BMcpdf42AiT{sf+<
z1(MT{DwZtFu}!Mm-g!19Cf<05n?P3Fz=0;cX4)+3cNG|gXE0T@7PN#OZ<}i8bI}l$
zLz_NY*iKB0L3uDEXA1yiBJ5Em-zum!86X=K+F#<Dt=0G;5*XP!Lbdvp4{gh4bZh<F
z&AT=q7Tm5cm`~Hi0Hv6DZh?y>x+25C30gP~HwoO(sqUTyniY7(Axbo{rp(ZfXg8Gt
za3;!BM##&c*XZJFQ6@XcT=?3n!g3a&HymoqmZYxOU7l1202)I|ux-6;%4I4)0I09S
zP&9(1N8=JqH4hx|pEn&e`geIGD|pylI8P3I0}^llA-`U72?w1y&3LPaT+r1=`1nAv
zDp_T0W_AqplA}qOGY8A9G{I!HX_{W(ajIB*ptPxuyNiG%J;}k&bZFPDbymN2Tv-==
z^MB;WG~fS0bOUA|<Pj$m{7(sODnN6EC;ahHO8qOw29&pCp+>%b`1Y47OQ9L5#NN!E
zysQ_ti-c9|r5cYh&6VVZY-aIh8vWu*e{Bj$bWHIQ@W_$nnxP)lHV(7Wvp;KLOzczb
zj3$`iW_Vo+H+75Pmx|GvSxvW^^1iZD%tFJNTw>9*)YTCpjpXd_*8w`vN}hVM)w6%+
z?3i5z`P;>H4@YKui{-w@(T2R5!Z&x%eBvQ5V`g?@j=iZH-RAeEg3zp(WY_6ec*oCL
z`@({0{Y75at#u;8mQYh}znRU6wb+0tmQ>pS8th-7tS?Qp)H+nivHD$nL+jZG0Xnxw
zoGj+b1iUXI9=sNx`nEl1W!&(fV65bN?tIeiUG|zT^C^XkwUC(K=+3~!IZ3~t<02pl
zE`=@5A5^gd|50G=$~%FTm=C<!`sd&Insx-NyS}s^m_6e<WGMXSH~szj7y*+Sext__
z2>&L6TE*I}#6ghbLH1mKa-@gAW9|ECgmw&Vi^Lkd>vt|dF=u}}s$mEi+a-9C)=Pv(
zQ?%<U7R${<cTc7ro2=pJ7&;wY1$M}DD|nwf`$RE7MLrUSS5a9{g+z|a{ED_izv6iZ
zzTwT+XH@ElB?WRN?q^Q5nJ-oQ%c|+oym#D{iOzD0<!hurh$|;a@2o4cbI7nQs6KQ#
z=D5~9Vy>JqR2&+l4GtMN#?tv!KK^;Oq~a~KZVsxXG3}o(FGfGgxtV)lVSLe)W*mm?
z@8=MS@AtYu6(n;Il0(L?$)S-WE}y51@wp_uXJ!*!LUOfog`JqZe#2`i8GBfE;HbsR
zM1h7TN})<p%*QAw=;A9HQo@UD(KdY)uW!${)4t~P?Np?>|HBB?VNalNh05s1rdRzw
zj2b0*Y`xzaF%_EF6cDT!FKI?!+j5)m5L%q8R9W5%_nM^kN{;*};}vzxazy#28-uq*
zChHn)#-1=vtKWom1mktDb&v7gBZc04$9W3Wu)IIj(MGIH^w4=GkA-oA-V#ssHZ#I#
z0j+TK4R4MUA$kUUJ$uu$lAR}n78-mmj7otzpzLVao73dBe&Sf9xE^AG0rE_-VCC2b
zZOAfD@Zr?P53UvQ#wgf>jy4gS$-fcB)+AwPA3SF{5pBVSchT;J?A01@vk<{RMA}c>
z7YO+9U}pP#o<Z3a6p<(gXibZ>19+5FtcdoG6zRRJ7d4ljVt^Nh`4C%7#`#rJmHEVK
zuS+{l$GdKm8}@`RhnDzR9L{x>Cu`?2bov+Y&(v7;8V-(zsrV^xJPC3b%}lvXFu<ga
z{kioLhs*ja!F4qDMAZrj&G;m1u32x1`U{({#_c$}do73M4t`7m)V=n8=CqA${P%y5
z52hx3b06c*JREoZ8CQC}rhj1co;ZH6u$7K@&bOEeC0CVTi~=fw8lvJ&?+~|~%zoy_
z3{Sr4q;hP#ysH(^tg19>4-<Mf&a3GaLodN*U$>=pqUqBp`@ARllWf%HAH2XP@&hDG
z8PJTiUFc`b9|au>GmP7pvEShjS3Be)mPar|TVUMsc~8)7E9wF%=esN0v!H>6L&EiM
zS&1{!eW4a1ModM#F@bBq`;jhaj28Lo7&>h6u;6p&Z(v<PQDdqUf%Qtjn>Y|Iqc8@p
zmE1`7+y^NN+Vkw~cG3^Iu%s|fW2Vvy96PQD)w0jKslT5l8RF8TB~D9C2|c8h4_0Tn
zjz?COSFv(~r|m-vlgxN-$+_9b1ey6vubzLc+{hPAUCOp2nH?xp>8LFG5$?U5DwmdT
zTyse(FdSD~e0g)#FyDZu{ds2e?f9WM3Pbd8vFd1ux&i;Cx98Dv#dhx$IQjPU(*iV%
z3`!&?bOwnR3S8YhNe`Tf7@e)y3;X77Nux&=jdlWVl-DNt%uSnicQj<?N0}vVRqGs2
zI9V|`&7g^ttb4q?qUln)oV0J^h)~?@DX#t&wYczR{aDcbo;h9|(fS^4_I;L<HsLvp
zpp{Nqr~{Bj1{yP{#|8080+b*Y_NK+IVx;0Or467%^AimX_RFKni;scNWZWD?s!;Fl
zG9(DN<uy`Vq^9`Chcg9i#~A_QlMTRxz2QAgAFO?HMF`%4h;4kMolgh`4w?WXR}H!^
z;*UB(u_mk-lESQ|e!K59VT71hXL)He<6qKvpTg+DNA(qGg+I=9lpe}vzMtylHcrM_
zG<L?;0!1b|OHj1)S*pW$rTk2}gI8Zsy$Fhv&i|cuW<eNf;&#<SqIOF8!Jgmks{$<w
zSIaI}DpRxbI}EbD*qBr_&6sYQEnlzFoUC#itI04d<EPck{Qg8KERT|(dWt)htVUK(
zu)vCyF@neRQBPoO-%QEu*V$loi*G8|!zMz84wkXfN7l}kZ+=~-R&M*2<E?y)hfr+g
z4qa>Ca`R~5qGK{@&(f-@V(p^sk4Ev22_I%pHypTVs2pr0c1|4Of=Ur{fG<=owS*9U
z3Nh9HNIFnNV8*a%Gsz3>rh_V<t_5BH&)V=M>ctRU4F83Lent7Vg@N2MxpK2U3DAF`
z87aZiZJfQ01J)%zJG}pE7s5pD!uiIdmmNGI=lp`WV|Rcia63^EH$fKr{n`Pb8*QMv
zU<^aMwoAhtj7smqzIkkq^+FQ362p!Vz0UcQ=RiB{G<hJ@@adTFz%g=kk<nnM-##tj
zA~!!j_gyhe<hIY;qYQ)N`M8n#0;MWJuP+x49gb3od7Y;A1j|c?{)A(*wgiqxt-e8m
z2A}a<T=tk$OD3`<q+Bu_<0(F>X4F5EABb|L#W`Ds#o>qqFOFf0;y|j4Qah2xni)e&
z4l!Mo1Ss8XZ4FeRW7Zp4!*wS9w0hGOX11r?4D_nT-B*>4QEBvLPln(6_a8-%U@u{X
z9Mf90u!TG9hpMd?e|=!erWH|PQW%MDm>};h)xbu}8F)wr7vUqT)^z%`gj2{%DLs-u
zb@pux=pw{G%_fBsaG?Pm@KG;dn`!TD%v19_t@N*dbN}7NY21E=QMG_9J6z<{X)%oB
z=}_U)%}8E5lPa6Z7cg`zHtsxqQ5~j%Vzv`33c@)In_H<v1qcZRanO|dhLnG25GOzI
z#z!`bfGqF?5PU)CoMvR*R%NFCVhc5f+#K(I;12Nm`kJ!1)xKE(KGMeGkbFVS3ZL0P
zPBb{i+yiaLw7NAT9OzrleHrmhqw$@yd1Q10?sXe9*X%gaY%OrhPGJPCcRGk*WQ$W=
zYy2HDaTJ9mbTq2iH2g7CHT=h{^`!(d<;kjVT`tCA9SgRK1-UO5&Wn>J4;RGrvw12$
z(w6IqTy7dwoMUUx)y<E2X5r*(Oi+LLMX<NN{YC_b$KIR0m=!*Ti+<fDmJgJjU+s)#
zm@u}5bK5=Ao(xNVJiit*knwE3F>}+a$}zT<-ONSJS-s3|Ysz?6??{65{SH=#<I)xZ
z6}>RGNs;AUo7?H0o)cwnzl95{J{qx$W$XB!p*g$EyLdk*))(0)3LR|rYP;I*3;QQW
zJo~|SNTAo7F3Yz9y&><op>MHJ28#Qg!{_~)<RR2Tc0Ypb!Oj&*ZIk#4poy*jYPl&B
zLdHwB$khR>OJn&bCL*!$&j&sgoOj9>h8)NqNj|~|0^jk=E!Ob-qzfQ7flc7)ShX_b
zc1&L@up~^gm2+c)z`qrTm?dsKkw)p%71VuN5|r%~)TdA3WpK=BkeSdn{t7bibowv>
zQ7%5$8&km#xj%nBmK&G@IVm}fT-PcHHI4<`M0HT*3YnG%QrV%7>l^KL4Y=KpayOpD
zYi8(1y2+lMwDF%jS{|&MrX$@^Gj4Np7<{vGSE`b!@wehOaU^Q9L%Q4YNAd4hXC9<O
zutTNY(!_4ip{Bq7R!w|aqW;&Sql_Q+gnD=Gn9!c}Hl>Smt5!ytaZH=}7}YJKrCUP>
zs;(X-<9)sulujXMPe7JV%&TL(og(TtPD0IddJJf(g@I$VM&NhL82o6MF?kztOIJvK
z`Mr%Tp@`U=s;-Cf9uPA4J^<IQ6?P{BYB<U>5LY<)i2rvRpYS}0bIswuDLRk*-86{x
z2eyFA5}4<RNW%b0(o`dt&O?X!UfL&@DW;(5l1vLN9mYD*+6$=05)#+(4rs}DhjAl+
zmIqm7!3Hw;Uw$}-mz_%$13MkhTM`%TzrO!@AsI=d<R1xJDm@-mVWOMA==HS9hFA6+
z%P7@q_t&3lUNzF6(^X&EMhMDeKCk_mmMHmAg_E6XWh=QsYVX#l!3ariF#|px$G744
z(Jb!k-_FkS4QQ-85~4083tT4}scME?Gz;#NJ8uDwf$Lh6nTzX^g0~TQNdH=a=1dx0
zG|iB*`imURlGXS9q_@8olh%0LHtF)3{VX-%_f4FiK(p}5@2ZtKYEI8(YvX6aGk8T5
z$dZVKol4_>fX|luC)u=-{SMUSQ8zEiu4X0B|04yrR!}Z$a;Reo<HH>H+_!Q8-|?5{
z+=sNC+F*;Ac65AV0eHq0j6ojee|*>=reB%{ohO>_uVLQM&H#PW4ZaGi{_{+@uYBVL
zhEVPWa)G;gSTpDpUC;)j1iqU5STez@a7AalGh&yNq1-tYia6#dynv02iZj8Bz?2jb
z3w(5q?RAg{{QOj`NV=~n&%NRswDhM{#hvT<WaxIce6ccyFAkmKo~5f^51Lz$ETcO6
zV!<XLAqYdI_EttlTi9y3RH(Yer)srKAx+Jd;|Cw#@S%09S#Gj^qw;7NGa5zRSe;MY
zl3t$TNeKA(Zpn6(h_uSquU5b#damd;vs>ap&96F8>B>xQIA$agPFktDSofmlX*)Hw
z=0MoJA?+JVbMg=5rAsC!+E=LQWxtjrsOBAqk{nZb{IEV>m2!2zbIT#4Q`RMF;jY_c
z#1gNo0`bz0I9=YvG;d?MLR#lZ8l_M1;b*u0^HyX+2R<6*^x&yjA2ChAZt6u|5r7IZ
zKx1MZ>USBygFlUvGe2a*U{=WsJ3g2bF1DGx#|zveiX1y&EN4JuiU6(+`X#6mkWw=L
zbF1LTHb~WotJZnK;FJ=9-6DKQ<p=R9EaI<6z2&BtBp^FR;HG{$umMb@*MU7K?TztK
z*HyMnn{et4qk!Lewuhe?1pKgRazO)3QUmJENJn#Cb~qPj;9dU_0S^vEm~<nTNht)H
z4B5}AqT#9XDGdyQriGD{%#Ooy?noX>-^J9iIb?*}>c`O2cju5s-m9v=yf*r|`~QS=
zzxQmlkIq^!Px3Ce6ID6$!mv}P-dFTLGFR7Yq1s|gW&5G6YeZiK2cEJVte?r4Ef?T@
zu^?bS6;yh&;bF|O5bQINDSeSv*>CwV%}&nLa_t*xu@B@^A_v^OVhhWc3^z=bzjp$)
z9ymHdD`c0KEP*<5vKjyJa}-qPDveDKkw71I8y1efOUaf3zG8%G4ax<I$toBTNLWjP
z=l;r9yvZAEK)G4B*Mvd+il(>}czXumZ;&e^opY9=D8q*8!5l68AmtjSf~<3aTZ=5`
zT2Xt%_Cnt2%Xv&J+m?^y_*7Ua<2lnWF%&XA)D-tST09V!#Ya_o%Bzu}lGz3Ry8nsg
zrmAvfFLSqV#5@pAxph_Z-TWsBq717H1JM_*?>2oky3IWEw!#&8c@F0v7v|Z&Q@K~n
z*ZRZJ!64ase_4LlyhCT#V8BbTD>$*9>lP9<z#!J%l>Xj9(*Ly@1`q-4EO$!~xlBU;
zI}kp5=(@L%g`*JoD+EA^HV<?A0PE7}adR}<Io~0H6|uQcMvQ}g8)3)`DT)&GeR-8l
z5~d-rpF&O%ABFb}JwD8cX*Fg=KX;q7O}L;|Dn6cE@z|xfeqK$NLWQ*NA9s&+F=5;N
z45L@arfAC3-OWp8?9L%+BGu`?KiR5&{2sWDPY{{>mhoR)fU1suB2L5KLbK1_83kpm
z1xbEMVVz)m+NqJ&tK28xxX3jleV^fHX-FESK-$-*{dJnCY%MKYqb{;W28_U@IvXjK
zRbiG-4HJ~Q$CVRy9(gBcr53>gZf&T6C*6NPJ5)$2JwsY@_GGo_;qFYE-5Dl9zIucZ
zI*}|b5;}>{)zQXK>Es^Lq6Vxne1D}z-h-CI6d#tiNc##PJRK?dkV-TO!%zGYey~Id
zPn|Pzl$>3P?-2U)U}Q50Xi9zfN8W19Cwp<VT3pCnLCRJ;n)dteQf5`7{+#4qhXv`D
zdF@SGYQ}`s(KqgXa!KJ^)3I&EV}4|dHj$LQja*cQDOEO)3)+*UfDO4a65h})70bwc
zkJy}10E>&&$xdfpXWKg4e(_?&2)^03uF)^Q4e%<+4(^X1nbfg-+qGa&et+Ur;Y<aQ
z4z&h?-f+17+lQDhya4a?OsJ9Th0lXI&JeF@Ypp@dbEyB|&%*X2x98cU3kab1!elH$
z==u=g0uJA|N%=I(EzOx<SG|DW60)2Mxq{}AB}md!96^d4a?ERR3Eq-MndgsDpE@GT
z!bNHcB*iC58-M!849sR~sba~tB;^ZtT&5;TkW*}++m=oY5T~vCxa8ZvKb1bq>Gbt<
z%e*YzBx&l>;Qf}FG~jikKW$5TCymIqG`LRayKMa44e2TD!clZC*HOTq<5yv(+&)fw
z(>(t0u9H@nCn^4C*Dt*{XJ76ne!F6LV#T=j5O57aYNNA%WsQGw$A9~b{|wFvbx7Pr
zWX(jh2=sjBDiiP=Py8gJ5Fp~&<Sz6?IRPO@?im@sV`qg8{w!(Z??`zeMp7;lC7nl|
zJ4j|zq<!%}pPe1#?CiuBOAz{#AZO?0;a~Z-XrImZ9o=5XB|`OzW62Mus-|r;CH(!7
z3TN%GqSdRmoj&qrN0^isGAmo7Z$@qI#Q*L!UciYja!HZqZg-&~nk4-@hb;cREj-(`
zXN=U#vebCGl}U6dEQgO>y-IYdCR`<O(B`cx^^BpV47y^Pn!a|r)TDh`s?1MX9pFW<
zDj1Kip-f-@lVw{&?*4by@!wXi2ni{7zKxt%HZ=vnsZt9!MRWsbz6Zk>S4dDr?i7OP
zL>_q^*Lm6zkN|orF6z1hsw7CFzdm>Ah%mPzBlaKzWC0vQ;8TpSCZI(AUdW~O3K_l+
zVQUS6+Sa&?gK6+@VJav)TdQWrZJ0ljD>iMr!cH{DrX5rNn@8$cToc3ds7X+K=I{Yi
zE4TWY-}isqc>g4@mj7v5Z&bH)L5LEw0tLgZ)Wsu<QR(<f&rk)yocM9ipX6q`vmq3x
zi5u5Ra-F2-7?~gD7GKVI>^M+AQC1w598TdQ?~+-rR5|O~x97?4BB2wvQO7d;E6{wX
z%w1)%kJU*v_*K|QPD*%!Y7LtK8#!b&{Y)B<+KJnlt4d6xFU<EbhAs>V%Oj~ga^a&o
z*iS^le5B!2-luJ`{cmfF{5-xkEJRvIV*W|lnTK4k%?s~6kWh9W78vmkjV^gQAv8Zy
zk`+WQ$fw6G>RzB;l8N+EK@^z@^ohdAcx@%C?H6=8FIc(IIWoclO%{evv*<x2=L5gG
zxt$d1tuVO2oCvPNsin@0-IO4bgM$vc*WfWJ%INs1OjMEGYyY&_9e2?WXuq;|bipzc
zOwGdcYT`cS2}A!v20VgVt8aErFq|Pi#vi05$YHmn6-aFBcJ(9R*q;bqjn+~%=D2YN
zbCj5x1cyd%;EK7A`M}<%=vycs)9JW?A2llEey6mTB&rVgk|MYb&wbT549T+ElR_QY
znpLj7Cr7?%c3OyTT37$CB@4TYWDDI|MOh~chzAS!Ni^Ky<XUea7g>K2aZW6T@g!cS
zm87S$85tR7J*fSy<=oP19Z`(bkOwNaTf}AqMO7mav_IChRBM&sn3r>j!l#1~q~bY%
ziihmju8AB#7qM5T(gYs7Vz!|5R=X^zJW2X(pSwzRREdRY{vz@3x0g-R4|j~~&sS8;
zbx}@_4n~0A!a3QYzhWShd>YZJjF}ndA9z?}wpxi(Lp(~zZZe{l+b!As=+LpxJWtFp
ztF;}J&eRiPYsPM;vbTE{Z#<Yk>~f7RuyElUU|=llXQLkBJS$UeJ;Ka$?OwApG3wAy
zk;2r`r9k?bnaY-9)zN~bC`JQnJ1dPoV`S~A@pgD5<rh7IlSh#n7S%i{?*H(qgM)%p
zf6j4dw@$=tQXjw*yP;M>dvX%#A?h)*2pmJX-V5SMvNj<OIf_{!?z^+G5s;s$Ini1!
zLlJVlJ&Y-Jjg0UPF^pYEQo_b4J}Kr^f-6pez#ihF7;O(IqJ8IHIgL@><j%fwC487+
zQ^-#DnM<xHZxxeW^YgPddk5n4yl%-Cbc^o0=-a!!DYeg5moAPI8@X8VA&_AZmtT<P
zxKc>1U&!-Et$lp;Jrm1u)6wlmzNW2{NIk)it7WRO`tfpBM&Arkv0lq9>CU_#=Y1<~
z@Ud-gCcagYPv@cySj)3}zH?yqn@M4|{2;UUcbgqa`eJ`Hr}4CWxJ1hyQb9#Qsh9P$
z$?o4>Gyi!ov5};$QSJu+Ntb~VNbi4r>tK}%$L9~uT22Yjx3_X0XiHHsn1D;Iq&l)L
zjtjgtu(SmFQKoEX2qlFuvRMQS?=&DenKo%(uQpc3<d<V@d!AZj2Q#f|n@1?^2Xd{K
zsfjBCURcntt?VqzyL_K6oGtuK@~QaC(x`2~_v<-pK_M0SN)KHJZ!z;0o8Vn*iH7My
zr<+kKX~c}dmoYY*951!zyVzn5ILs8<t^aw8z2C2`D@dhcqpw=@xDU0*0tWc?;V_^0
z`c`9rF)W1SDF7Ve7yo-M2H6B5$ZyOwCEVINN!n_mhq4q+(SfuGL@XhN$b^;Pz2_DR
zpz{YN6`4tOf-8vVX*vsXsQ5R@?jj%oJOQR`X9*?QFtXVMG+)>Kl|Y#(uu&0{RIe;N
ze=T2QD=v{gnl1j(A#w?G{2vflRdz4us$VdZ`TDax`}r5^JGf&eM9u`L#}oc40~z|+
zjhQ_!(c0B4FJrBvE9UZ4ug(3d%w5-!mWiaCu!<xv8Dtumj>=Xkx^Dcaq&<RKhK?5i
zRI=U0yaU^VQnI7Pyu5>Xr})iJo-$JW3XNRpM(1~(NZP5z*MIa6>K9n58+Cj%RB-xY
znMKL#X-)eZ8QW{Q0H=lgJiM}4&MYU@;y+K%KYz>t<Vs$N?;|HBs!cF4F?ZG5v;_Gb
z-x1wa#3w?~nkskdzc)~xV;-eIl(J(C6R{<bboeKPDDrl9>2t^W$Ae(N!D;`#YOjU=
z=LXHWimFLv`g;l(cIcS#o4j7+3qD)W9U*1n^o~9Q=LcSnjQEu;CV`iW?}-Oyek>Mn
z+~_FuLdD=+u8)M)3U5f6=}X(*B6IAvC`yk<<E#xwlo!vT9|sUG_doh80hB%aEbHgj
zaC{Vp-(OeQ^wE%)&{nL#AEdG0Oszaca}&sU^m!~`-TPex9?vQ<I9*h--$A6Rx7xu1
z#>>15vw;%X0aC#kV1pZOJzjhSY(r@Pql^Lhsuuu-%UDSBZ!LN30ee0JXtnjODQ+vg
zNcY4DbTO8!ivFcK=5LDOb@0m%pj)c%``Y2IN6lPVTmP`rX@oN0?k+<yGA>OhNd}kp
zvT4DvBJ$PR-#x8l?Eh;^4iM-o_6#=7C2u7X6i4JVDu;1X`U*}$KY)$I{^(#C61+Xj
z+3|_9ruq!o+8GAD5C_{(w2F^@Ax%~`MU~be;gw4n=DMWJ`lO_IDSXT~=-Fb=@_QCp
ze#%`73ec@?h(cndQAZTzyBn=zflh}^vO<Zgcjd%%j?&eQtBw-%c&W7v3X<RC6L#?v
zHkOQyibXB`l}?fEju>Bs&X$?Z%GMi}>NR1#Uzbn|ruxXY$c{8c3a}w`Kfy-^{D2KX
z^Df)-aL=sq84$TsdxF1D6cLIz3wa&Yx*u*1MKCJTx$I2(88*HSuBHx)iJ^nyD;lKd
zJGjsS(*^8yC*c4G8X8Hpk+qP}kP3U&mOh=(fh~%zJu}<-SGh+Kf)pKTZF;K6Pf+0e
zUAg$i41@6!5>t?_3cT^WNw}{a&LYqBSLOfmZljQQ8^xgs-vk6_76V#})&1!x5yv`i
z`H<)Hm4Jl^sh2CecIlAx-X0OxmGJd<{bn%J7?wtSAD{e7X~dP*__0#2-xW8*Og<fl
zIGV;Vk~MNIBtusz`*Hf4I9}Q?M=X{-(eadqp780N^0IWfUx^PFd7Y<9?GtJnRlYl#
z>xtzq^uAZp@7p8j+Slz<X)l#^W6gaXwPs0QP&n?SMt<n-*}B+dedAuO*Rx*R_?qSC
zm-#$T-jf0(a4>6i!ezk6IDNhb<W(ZQ&tRR4E<iwTazPDXR^GhVH6GYx=VK~af<Rf9
zgmy$Oqd=@OSNxMvfG!3M`G&*jW-G|J7UqK!iJJ^)eN0t+S_1SN%HBV5CRDE^t`A1~
zpP)autyDcg#A4f$`JFH|jRyV>QG5`pwQ!Ly!!$6jPV$!hNK69^Qxc)PFbdG4%|lAU
zky;U4<3m`z%HjWye+)@l%Q0DzAN;8v$`1_ap2<!XNs3<h>A-I&9={+9O__k1F4E`M
zi@6kBU*^00qd;oyGo&xpk*lFXAT4Mm+Ow7rcRA%>c&aWdJ=4%#`joy~Ck&-q7=26S
zc&L1Jk`P6y>vb^fVS7I$_u<$5O>|$cexp;uj8!bY@!yGQSNY}_4l^Fd?4^T4`60XB
zG#&N%NBKTRU02CoUyIH@KNU`h${k1PzPeq=v<r4MFWH5h=T%(Xx`A^dak}Q&)1%Ex
z+2dX&lAFaR9V;~i*C_xeYPcc)?3Z(6Re`ErgVzwqy5!sj3E*qm*yhwI-?@5^{dVTU
z+NniAx8hg`RD<Q>ImJiIt(akKu(ut?y?sODk<<g(mYIYl=($Tm#4?7_&Cw4a3v_>y
zw<<I4kHaycza6anwEK%;LFBTYN8HBs&$Ta{+!<+DVsKyXRgkGOfjw~S{?Ap{PkgAl
ze%hEY)8HfWG{+Ar{{=MKmXLZ8ogOqB$Jcgh=PJu+WF1O0@!hEX+0f7~6RCQh>S}Jf
zEzex^kt~y0af!N^@y=JS@}=l<!<f4&By5wU7wgm;^6c{*vy8fdg`mA*HokSH*WJgX
z$?VF(((9bw^v?VkyN!uY^fQ~#rEg4b;0uvQe}Puyyh2ITOl!hOX^^zN>6M_UwV>+1
zY7uk3EPAwdSSYb|_~}5eN90_bLJm(k=HcI&=_D4gE+pvGcJMh_y#OR}>$kUij}S@j
z`{6}{;I_r`eMe6SQnSD)-;_dIjh2Rkn}HPIz-i+xkV#Mch1GjJ1VV_1OQkh_4-2LH
z{OiUN3m+Eb&{H85$B-B(S}KC93LQ9;4W5&L{rD04ul@K(Jq!DR<ja%kZPh3H;reGk
z#?57lVvzl4{4e`Kfb2(+i3|-s<H>%MPij}Rf_kI61peSIU`JBp33YNp$46cS)3r2N
zKU>1Qw{P&%BE|o!!eRZ9D>G)2HTld3_iatl=J9CRD>F9Gl(yX-35?BHtw@{^4lEy@
zvJcP+Zp431pFcr;NOGvh9B?0R6@i`Hzg~py;d}q!YKr7gt*ZMvsG|md4oAiaFk^EX
zXSpe;A+@e!B~%gH+iSFonwUwhx=xlBV#tO+visW2n&hBHebn)1r2wyM_>z!qui5Y)
zv@;4p7&w_s0FGP}`M&dOx)&Va<6#6;MF!@?j6Xa-tuZqNJVtY2W}Vu~hbRGc=_rQD
zw;5joBp`qKdLO);UVj!6_MIC>pLRp`;VQBZ%>vn55sf8Gj6WcohSN{s-HP<dNel8$
zM(wAfkvb-sbnB$rS%SwRbFEX55UED~8~fxr=EIYh`YbjLGEMA-p5(c)i0i{rQ@YrS
z?PAdM;(gm2D8N|Jsx^HILgKm<LN^6H<nA1}i>a*gFI{$#nTUHJJ%gro8jJhNd9O%_
zSnT!6Z5LkzMCeVun6(yk?s_a6%dxe4VaB8S<6%aZy^f5KMmCyXr)8O?ypy&FcDr>d
zY4|UtgvmCLMQ;k<BoJe*XR@PXRAmT_Vv5r|3SZUe^Y7y$W7n<Te^Yv`O!A1&t0!EY
z#@F=r_}ts|b<5eQh&CGO@^M~&wD8S+848xCcWwGB<k(j@d1hk_c5T~k@MtlQu!@=X
z+f6N<o!{(omEABwX5+nW;Gv(h={3@ZIIV&Sc6QXD?Y<#y=Lt|?I@q-JHh6i>_oVq1
zUiBQTvNM$eNL5K+!8wF!CH-koqg}5&2Nn<-@abTAKVgD0%G)(E`pX}!k(c^^=lCK4
z2)QctGFl%}0UD%x3mKIsLCfMP<WeA!>OVk3KM9dbv2J2PgO5=1j_F2AM0w}c%6u}E
zaIm*J%4%kOr8++GU1aJl0dFz@Z(Kd=<mjx1)3jtOVQ~(A8AfNgRDGqrzo%D(X*p*l
zOwsj##b#;R&}o9Mj+}*_XQr~9<0%8_m)MI~A(AbxJB;U6y_l2<%qB^D;#x|pbaTdZ
zM5tzGlWOeM&)phhdWu#rylrx_a;KG{&5<V^unU*G7QfPcw-~V&!OvEchgk*!N%Vx*
zaUB?P%0m|;!$S#1`$+{|OQCf)Arn~uty0I}ptDGI^kW@4!l$|6yBoMmD-9|huQ?T$
z{@7`Y6eCNP$JOx+L5H66!kTH3p1i_wCox5?Gdl59T_$8eQc!mP2aq-u0clUrC6P+F
z2uF!esuF*<|Ao^g8XWyzI_u&BU|I36EF1ZkrBFQT<ZDZ;>m+kqI^6$uIQji%ug8yT
z>h*Ixs+$t`oY@jqHAu&c6QWwZlFf9xwJK@q8q{yyk9}&l#7rgFurKLSB`Dwh9_K@B
zUqNf?@e$K4S)Iwiu%%&Ag4d;MK|&&~M^kOr6w)L=z7HXFpbpsl0<FcAs_ROxIs}k!
z>pK#>YZ^ay!WWc;cb(VXAjFCo7WLGDQ}HfvHWkV()p=p=12n#+e{?4==6b-x*jxE*
z8ebDP|0?h<DvP52XO-R{eE4fwVr}A+ypK%+4*#^fB~@UlBOOe5*^uDDPQND6_ASD$
zk8|q8i-%B>9e$ZnfSDDU32B3QL?aj!C<f!Kky@7$K-$FaSbuvLT?pvMcV8n=ANSa_
z8Yvd9u{r>8qNwXnYajQDGua3Bo=;7A4FF2~eud<J#*~n1?Li8!CXKfZ0vt%%8q{bg
z)a=M+I`bUw`6;amb7}p2x!`RxR<|;yuv>ne#&B-~HOlL<^5UUd#e={>-1q^j03T#g
zmsXHx5P3Fbz|1)cz&9pw6}B+0nhgZ|v@xXw+Rw|cl;>u#N^T7;MGTo1JkwUa?PFS}
zxGJQjqY-t>pzVw79OdU#en;u?l!_i@OdB36cIaJ+m|uq2(2w#8L1Cs!>46h>Z@ScF
z-N;a(y>}O`-+5^JqZ8;iTH#2i1=<K6DHVT{qD)m+G9RmLd9~Cass3G?H*ao&EJQ3Y
zb|j;jlxK=s$RmM|APF?K&r;4pAmJ8toktEvoT*{{sE0~{gjGXnZM+o9jZywtSagO>
zlGS(+yjbc8lpN9vgJH2{QxlJ(+)Ag_@m?56*CjSQG;9eO@c9iw5*@&)VnswvShOl1
zAq<^t5FruR_;SA&gx0H$_7+(ad0~nYl#1>lOfOI==^M|!32=XG1IW)P<f(B4e>&Wr
zc&ZueMJ>UBaO!PgP^E3YFRM&)0z5x`^K|Motm}78tO`I$F1_ru&Jn?<;0w`o_rx?W
zA6fh^8`9}p=nDIIc<d)y#WUfqm~A$2WWcI-0E!wzgNSi_=^lsE_BXrGW}JsBmD6}q
zg*(TJ!M1OUZuwoh83Ub$R^Szsxiq7yek53=gS4Uo7z>yU=Edi3*;DN3G($#HR$U|k
zOTMUie>iejejrzMXe%;yRqxFW_aQ@A#!=VLP9huU*VVF|COj$x)?2Bs@egn{FYcz6
zGCbr|U%q7Y*;A+7GTxCbU#m)if|N=>2n0)uld~e}<=f%EoR-g=cdN6gK6CBH)udp@
zIqxy&-rnBQ`tJ&XAl7uYvqab{>EmAfTvJPkaJzx*bn1oD*LFCjk?p&$<}Ni0L9G}s
z?B)4kB!qhe>UJyHoUOjQcs7W%h!N1FMs<1{Ud}XlJ=-F_rK9DBmrav<n8<CR@^L%v
z+=Z(YAm9?+{K1P=FK@VJ_@(4<pEFATW~G!fajC_SoF>1dc)*pOW>qkn<{7}8*<=!c
zISSFQbCpVG-A)_!Z=bo=ID*l9w_)wzgV)#O*iU|!Vu5=xI@AJouI(WDL<6eS1<n3k
z*MA32+zhj+>Odr3iz=iY_pvP2;k5L2OD_F_330~F7G~Y(_2vowCB9c250mabPQ)s?
zl5gruDDWrz!`u14sg;+YGWwf!c@-iz7}DDA&fj8Ndti1@bnO;LUPS!b!E(if^zvv~
z8I`m54!HDL+-g?RrE0d(^~o@U9RA}~TUy{ONwmCx5XcBpCUyZ2Z~K9>FcKOCPC_HT
zD#U&)xd;E`$XsI6k&Bd%v5M9y84`CtN{FtLMdiEueN+F=*Onxz?f$BEZ*Jr_r|BY^
z)21y{y7&En9u=)XS{*@tj?5<oU)i)Ph8O;%@JA%^^0cD(UPTg8+3wJbWGxvbaXwrW
zpRpEb8Beb1QC2cN=1XN1NYifpbt8A2ahrq{KVm0Ha<C>eDXhrCvF%NUaE_kI36_8U
zHH33NpJ0S4rbWme>9oJONDrmumNt`B)QbaZk=#3`bWNfgzWBp1^w+Wd5&sjp=Z?P2
zWEHG`Lej<X0o^t3E!=x-%^OQ4XukGmMfj*+*4Gty$3qioy|BlQ?OLbeZuz%V-(I*)
zA6CU@|MN{*EbunIE}7h;7wdpASDnlFU4(ZHyVpX;>FK@O^T?+YL>nl74dki!nfYg?
z_$fNS>csKB%vsY=X*$3mnN#Pw*=yvP)iIjyBh@cx)QS(6K6F{*fl%o-46>R>)F?JO
zalxTwd0Mv_L+6!4<DYpI&06s-HA1^14fpV-`<gf12p{dz2t4{`Zi7<pTdwtYH$N}f
zl^lA@;7~Z!Z}iH9Og9NNRr}}#V#X6j9K(yDUsDiv3_=jSY|@X2o^hb@UEQ{pvbFQR
zM^W%kX?^xR1A_iVUDmV$a4`m?=SoJLZ5$DyUxYmO6=jA0;+g)r`b+5v_<H()fsKB?
z;E29hC)7Hj=iNb-RBvZg?uE2IwDG-HS>u!YoHOr9OOS8K$g5okbjC@5>!2lgP8Gg!
z_hlbj(bwT=0go)2$w>?m8eQW3Xgh_%PxH06B^dC;@mf5c=aQW;*r-3Q5{7;53VoB+
z^)D^};T9d+lMlPxpy1S!@rrw_-otH~%*M6*RlDk{kMyq1d32Af%H2leb0-NqAO|+^
z%3gNI7&+;5*+@MnKxtzbIwHU7k4U^Q3M8_9hO$g{pPEC=G&;7}1juKY0=qMjr#GjK
z33PgS4>@?{w*0cJf%N2qw$sN!qX>#I;7-yY1b{aBxKB_VoR5b~;}`UYI`<(;1VOmd
zjI9T#geDT`{j?%!BaOPEdgyWPYY001ia79G$fMZX;1NDzQM_{$ReZY;5YT^h`vNAw
zLry1Ic{-ZXgYRY}G;|O!b$5A2eZ=*rngaW$YuR3=l#DT}8f__DHAZc5qQ%BILBvkb
zb>+32OkEP*inZL?^$*gm(7Z!-H=)?V_YBeo#Y0%Hc&6nr`an?(I@D%hU2`iZJ-*X@
zstKDlfy#$PkTl~7eT>OHvZhopF&t5Fl<*O&-4;*jY+6DpCjxc?4n=Ot7u3KX?7lk+
zGy}flwNIBgo!9g_^skDu_&*5qS~hw{Pltc@m)tm00CoFyUpX<2&uEI&L|Sd>SMEP=
z8eehQlpK^h>=ovbwt6e^tA7zUb}eX?%YJdNOF^#U+RfowfK(GOP3f}}up{wlDL-mp
z|1z;M&=H0)uxyR899mmRcYl$fRfTu%Jf9k<&bO)6#F5ayzOKNSn)EgQS+3mF3pfAl
zCuk-YL0zSxbdHqxHTUlyP9t6R2+6>`59Ie)(hUSTtJ}4HE@E5p{_?X7DLdt&l3{)P
zEGH@6<e-N<-E!bJ>9qG3!~{k3%5M@KXHn3^&AZL5wip)(^Pzk{Cf@76tlgOk)VN~D
zgMIj^zD1nCksu>@e>>h3MWEIBtI+GPBchIt_I~<i=#$m0POeDjF;tDoLnR*#!zvgy
z8PN%vY$9)DPi3WED0}8%C}_w=QN5qCYCW?CYAfah$-mq%BgQhw)FyW7PVe_GA3Y{m
zzdyN%md~hFo#fR(<77E}L?itY1hpiEyk72=`pi}2E0N?=)B?FB2688d8zHLz;Un+C
z!LJ)$A6hQPT3c`5zKqJ(QJD&Dt0X7`RQp#cN#i&PJO&RIa_)f@uMGzm<YG8TDexk)
zn-Z!g27n4oX&?S!z9EV8$p@n-K<0`D3A~bo{<s+UOB7uT*ckXU)UDb_*sE%k<_m4Q
z5-OzrkNsaZB~d@Xn{b<i6zu#$?Ltm&>AQE}m(b5vGDt0(h>iA7uZD45-QW1>*KR*~
zQ-@m*d}`ugg3YUuY*#hJgB=J_k9kt}kHF<kW58eGu#0D6C&6r_IOChkTyp(!b3>Z0
zE9YS7y9WQ-fEnR)k`F;KFv_B^PNg~Y<}J`Q*c$LIn#5QGCkm_!m!@tmFsm1*gSdw%
z2-2|tv=Rg9Ok%(+T5vmwPF(}rUE*zDLm~bnaF{QA{q85$RT2AFV!=XbI<4NWw%5(B
zY>Jc)MQXN<-(7`HKn!?^KHldzBB6SNV#o91sDvJGf-`tH{&;Tg2~x|axzg5x50TM{
zY!EFgccS>r3YtxtFPnYdm3=p$6}&#nttEMibg-cz&Qz>gLb!;$=+gUwAWrgBOD~YR
zd-3j@Nt@*j(v}yt;PDp=+8~(ul#(yh5<z{kLf~9p+}l%wm$y8iEU=2_ZQIpBlSFfM
zvV=bK`l(P|e=FEDVn9ID=)sI$QqdlnK?sMJURui8=GC6({ONi%s|{M!_EM4DNR)p=
z`CPbO^M!*?i2?8tN8`<cTn~V|{$!%0E%sIlBNMfkn3%-a(Y>5!0wCJFmh9%5p(a1P
z4Yrx*!={S{5ec8=2U@+}b(TwM?YjmJ)&^aM!ISYX?%0&Sd^xfmgYQ5V?wyai*?buI
z3~{CdTlVyeiZ^kOw7n!IU7(;c9I0_Cq;sFHm9V`&KyF2`$2UwhS!M3-=u<qX?4xn)
zn!hEm^+|C1TN?KFz16FF!Tvw@xV|)y^J9~Se>t4Bsc4AQ^L!TnqJDzn<08qWo$b?)
zb0}34wFlq0;_+(>O~w>H!&7d*FM`EIvvebgcQ2udcYW9}bl3CJhu=ToU5mYLUHNfH
z={Ab!ZJun*pURBFTMQXxAutX0I3eVJ!Der?`?;hBB-iJd24DjZQ(Y5#ziqFlKXw*^
z#{brfFAI)(2aj*^=<PQch0IF$i+u$QNL0B_4ulT#;v(UuawENjE|<A?aCQ3u11h;j
z{{w+V{{HYbqJ!Rfn?McOb?hb*Z<x!$5})LjARKhF48jafS&WEjf7j__9QASo(`5<h
zrUU76f|QN)K99~TNUF?k@{POymff8bJ|?AjJK4C5^g{pwP8GkuC4an&_9|saF$sU1
z4D<-+;MiyU;{BxorC?uT#d^)W$l$ZPtcc11Fh^%H6RmqOO=#YW_CIR<;`-w1IrT;8
zZt)PA5fEIyw+=UpnBteHKnjQ9E4sx5+O#j(bKX}&w+?(6Itj!oG_9+|JIXx|U0r^@
z^}|SKqUZg8ynP2a*6sUui6RZ7$leh}$jZzvJA2FCg~%R7kzFBKMI}4kW@eJTHzkC~
z$jZok&r9R!`Td^ffBfI$J&yM{9FE)V9^dbEo!2=(=jT+-p}E(;)GEpWTx1>IV_iy7
zJd;Va{Mk?-uptbby%)Q7dl}!4nuu|IGU??u?OE}g0oPwJKe+Hs7uU?ZUf!T{=>kc$
zcN+AKwF&LO$SR#<+M_y!1E~>bbLU`M<K*=I9(6$d->gZVyAZNZrGLvz_1ZMxSFU=k
zhvV5`y&8szE?Bn@wi7hbO_8q-LtZ>@onUZK^vR*8x|qosI5maSGObthMbFmq->)Pk
z6`aJgF^2Mo&NpLy>vt9e9i02?V)`0syw{$btP#HnC+=(ui5}JnN5Ol2Y~Pi$7TfeZ
zQywtpm#3!aab?9}&)d#^%R%24t{&JR^!PaAh+VSOKV|!h{OC#~yRrDTl-l(z`F3p-
z+XrhLjx|(Lw7~XPir3x#z^ODGWU;r}amRn|Ebt7-66jOKTw7zlRw+fd9y_xW+r9nn
z9l=9Q+1KZtnvT}o^&Y0tGuXpP6(coSZ|_6ao0VQY?NXM3o9~`LF!AQVhi|6&G@Gk9
z-nEs9oQZ}S(-biWg%+xb%jJ|l)QwCsv<L{5EStJ0($geA7(KM41SgptVqVHTpO?Jt
z7M_O9K0Cu6))67W$4(LX4;?nd6g%ls(~HtujB*`g#LzNq&q|JJkGLpsiT;EIGIBw}
zC`ia-Mi4zw#9WEn`ZYFpw~AK15B_y5G2hv8`JR+>B>B$TT~YEIGkq!>Cb{ZWhX-7>
z=XS6(^2;RaWFUQkk^ILTMs*mtSG8fnx<BynKM#$cD;=De1uo*512EM{?Ayh_I26=Z
zi^>)gND~I-!-O8ZTSq?GA(!)VnEpAb6yLiS(j~jPu4v&o5QNdXYD<+kvmULtf)Qyt
zYVzD}WaA6lCpwlOikpg~D9No6chv}bka6SViGHGXGG_UQw9hU@IqG=;DJ6mAW9e!G
zVLzH=k~k*rC#uW(CH;?biKasBC~!P7>PEx%71R;jowq%GEB`>+;f@3q^5v@}HZ)kY
z^wHCJRb1t*$}f%HJf_ycyC3)}cyij@?AxU#`V1YZ1>S>S2<{T<^Fg*FT5DnO$hAt(
zNjw#YS|E;)TIW3}StV`Yvv8E2aXa^JMDTgqbrXI1Bp>aJ;De=!93wh1FJ#8$xxQX0
z^OB8(VcmHl4-c~k5`5*D5M7^48Mt8xc@(m9hf?|i4;7vGbnB?dyi^|#h2Up5Fe5q!
z!a6)PJkaR$=G~zh(|1EwLJ^5{_t#}^_Dt7EVfFE8OpgIEekXk;!vH3cHdWk+_hjC_
z3*_|ai7yQzX{J9;_4}0)Xc1m}v$G+?f=?2tW-G6;K{~DrEUoWJrnoLOrbh-hGkyf-
z{KkDpg$Sk=SE8I4y2w-1QiyQBJ!BJXsau6O%6hR_)FuNnWLui<Sj0)>0@?d~rz~i{
z-J5w$VDCyy`GI-5^;^!SE(<pePHY_b#QW5B=arUB(aWkGN74rlZ+mUI2BUAV3CVzf
zx<w*CeOJNC(xNE9@>QQYdIA^8<Sz(6?f%^iXHRrOKTXh9FN49&;b&HmI?9XFb+~`p
zU~#gZMQD`ir~%e+^l(na_JE_6E<^<TJ?Lf>F1sEPysUW~Qmg^NyCtCrC<q*jzkVbZ
zn!=9*Ec)m3$T|Dzu%$t*)#<VV^MU*R*t~WFPlKKiBxDg)TS(HKfgTq<;<NVyXvr^}
z<!<0#7YnIMRh3lm*#j=Qq<(i`g@!oQm}u#$GtqMx8>^|Hrc>&go1?1{$zVb!lyO(b
zoKS3*zLcDl?lYBL+3H7+@^FGL$9EpqTbfs4((^+0)3%9U@4TPR7%uGpBG28i^Lay!
z6Pr++l!T_f$K+9ZQ~LBApDyDJ&e(bzmz&QVN~L2-su=xEQy$-`YuQ$LuAZ!ya)z74
z-1u4S{71;Oe1`^yeaY0i&7|e4e$GL%>F(mD>XHJy(wocf<u5YHFhb@VZ@*AV2?GUe
z*H|xjFl*@n%7@yCjZ5^o8t!OR2|SgPJXrJn1#?$uSrl{<)zYO1^{^zX3C<d3EcLU#
zF7g>g<+2D#JD%1WMHA16OUu~(WE)%nu->a3O3E@Up^iuD)|^OVKc=p3MP^54rq^Bf
zu7+p)xkBo@Vd)nM(4SXHu2C%e#$H$}!H|VGozh;<$-LDNGZZ<)o_?0=+<Q@ti6V+}
zbEolwQr!`O4)Cp)Yk&LJ{m8d2ARBi@dsmqV=$LE*Bt5e7yPmZT$J_a5drIqgpC6#{
zR4%ci;IRBg5!r0;;6#8$Sz=z^xi^QqLGy4OPg=mD=F>+bJ2`JeDC8h`>ZN}!%=Muf
zMBk#t^lwiaJs^-nT@G^VwsF`qcw#OHlS$Ive~RuCw`9C)bRwZML8KK7(i@=&yB}{(
zQmmq0z~j+6ebPEd+pSPMMx{KTDAyPSMEaBD)60_Q-g#WhycnoRa6&y<aGKce!MQB@
zJq#6`S0m}A^6H+a&yu}kJzl?69Yp+yG1MhOt$}~`;^>4uP+n_INPBj^=k2+2JhoTf
z029YB+6R2-$x=*%`8F{tk?QB17;g0%cqkA!2t!$-fS+!WZU}nw#tPos-zQ?aWJ%g*
zlDK(arODap!cBwl{@8s0HGb){=w4o=FRI~jQC*Q|8anKpc}4^vZ&onKnK-DiWOK4b
zN6NCmx=riA=W9Q+@-o}$6BzAd)|niVrY?7JZ6DZVtpurrxUh{FzpM$y&Mb<XHS8`6
zaCuolfYT>>x3D4Lgft1OnObid#S<6EhYh@=N8?<*HdgT1iz`cCoE&w3`6*;Qj9NHd
z-}I*Cm@Ml?tE*L@daSay3|CnDj3X{4fA)3`nM~pM;K9~NZ;cn(hkw@tO%Z5W+x*_>
zuA2E#THz$!#XKM)$YY7O3=U2PA1kC74PRRuedXiGAT&LW>NJDv!Et(_7sp)&UElMw
zZataL6hGzY21VhpGd5u4WD(Z%z6rGe5LNiqA+e}iG@JdqScIbnJ!*>pFM7WUJhd0f
zZ+Kr}B9GGxgF|%Qo{)U}Oa{x7?R{s1*)v)5W6LJ;gHPLHTWA=l7xYYXT@uk)w9gBe
z08;Y?VBdPGm20e`&0*kgQDPZqo^a>!V3sZ)N~X_^!er~dgV&^0c!}n|4dS+%V<o3Y
zd}*yd;n!`w1)4&!X=WgoXg{#3odo#6(s8M8?J!G7jrQaDpy#ak-6x*CZ^k*`zGY5W
z-<-VW4pDiHNpx*dm|BZUmkBSs`l+5Gk<8De6F^40j6D={;qvkHty-%dB}~3$%@FIr
ztH>1XJezKOTD2DIkP67VKa$U8=G(Z)Vbar@Q@(x)l6WJ&bg{ET9OY}`1cQ1!S=tzB
za^Q@X@{mpkswbE4l0f1mpK(?sPNG<^GW)81fKIo=qj4t&(0iOW24tio`!9L(N&C4x
z+Lyse(Ru+B%wD?atot!*p7M)DK+JO0tDkEgfq0OZKV{~&?jA`5Q+a|hxrp>ZAQSun
z{h>f7qm$^f?MB99tyXy+X5HWVHv3-l6>#3<iAVrO5O1I=xa-ZW0|&aSZctdB=DX8d
za7xI7X);R)Xp=Uz9zl@s9R&&70-w<cjZy?uoI9kwOI-1z$O*q0$sv!Um`7&fqZ(8h
zNXYOMG78yuLISJ6*>Toq@4#vL*m?e<xO-9v5rOyDV`SuRO&^yYNm=KUmuP<ncC<^d
zOHmH7L05@S%0~e8D^AxF`u6XnzMJbXhi~0`>>Rtvmh)6syJ^9PVkfrncXHAkRhT&0
zx9&8dkutN@n%sU0W?sePBhs!K`h1FB{D1Y?er`->m4B^0oi3iq>CSOJ0RIalDKDJ4
z$u%|zNmhZwjo+4_mGdIh684oREwszf;4x|y))#p<tpxWjy|YXDwprsmCOyTQNRjTT
zHqpzX^T~t1h>;#Vy6oF9XV21Yv7Xt>OfA4kJ*|(QI)kOdJV8Ss1IjXI!bP`>hUsw_
z`Ia{dtz1TApu>G1Jtsqj?)NTRy&9I)Ium9;3Wg_lh~5m?tN5G9rw6x$&WD^2B{k&_
zs;_A@$YfV^n=4x#0|LN8{Z&>Iu9emqL2HBWug}bBxQ2~k!~^J-er|x+D=4AG+j2K7
zpeKBV{rE0@j*gFuCM4eVgvD;6e}958C`U`QDvHRc_%c56nqZCrVD`25#w?%P`j=EQ
z0E6sRk1Laue_3{{OCMqAFZ?rxUIfBKC(ovR9zQr6eq~?UQ3GQ8<0f8N+Ur<>_W*gW
z*L$2c`7?~<(u)bc?gwM&E<r4B2Dw-r)d))HZfA!nZr5!G<UOrApNAeIKUeon+@?$%
zvs#+2BBXLB*OP@*H%^KT$N+BOyejm$967nU<Z(Q4Q1)IJv1#O9tU*}d7g?jgB8KnJ
zEZl#wN&1PZddTYBZEI7L2dKd1QoN^!;huH1UZLIxP!W;(4!Ikld><#%`(Zq%^=5AQ
zH`9{=<l{UuN;6S+kF&Anr?OUyv)}RDf@Hf+PcUv95HR(&?3bi(g|RA^X`1|-l`CR$
z%^FEAsC9vSTiM0UdgV+6cRZHoDlQhP)cP=6{3J#*))0d^vNZYmCT^K<JVjCU1ts%|
zDAEx<Ga-$o6h{q;ikPu+dCp5?<`$_ALK=}D5_;lu&y-+#Sr*b8mK42S>Nu@I#jSC?
zeB-NXRuwOe=^_B3>Odpr;j59b>rFP!eaqx4iKv>o5e&6H8$T`>Mf9TkDpN&V{p=Gi
z9uMym{l&+lL(l&3Hy5=Dqj9P){qpnPAm`yFp2-cAsjFnzEA6_18}SbPK`RZKB(-SX
zc5|O_&y()i*md${dZCmfACixJ=@*SHs}5^V#MQ+B$l*5E)4R~P`J=3{3)VXJ$o2?8
zPb!JQw_NcL4?v&y`-k=k@-4*`NCo3=ViUNn0@*{hjQ5~(uT?ws=$YlYAwz&uewv2i
zl=;{yLUcR)Oo7)y?Gn|wyNee?El?I#<4kZ|*@QELJXvE(D-7+pU+T7W-dSMSDaF*1
z<cX?O@={J$R?%dEclTcLj+<vg0-r-2u?Kqs+F0v>N7)RB^`pCiyP?@tV7b+{Sxd@C
zA7dVR&bCjPwud2rUQ2NKKvD@|VyNT8OIzt~YV_6T@qB6uKHxqLM9g0|$j%!1n#4FH
z@>sTs1qT{DyL7xhlHraSSs_-U{wOP<VOb(a*sa34;c@Xjg%h_w;zWJt86})L5jw8M
zV4GoJ`Gku)qdX)2S&fkSmZn(>eUS%QxQ}QE>3uCJcHT>+PY(9D9IDDq>g?#iCZ!wJ
zzCKn*IQ!kd@zA{I0L*kSTVkC^TJSK^XgzGgXq(X}U5dM@IKe5RbZ&&qE36A6?mWi{
z(`<IpCXB?O^6SsVQR1T~_=%1DEi9y)c8}=-@Buj;#BVj{Ye?T58O)fsUE%8ZieoQi
zPZ=em;;QlFSg{b~hH9gE5IE%k`W()1N*M1ab{!q}>-?DGKVz0%mjYQANZd1d6)0d?
zW5Ya>)aZ6bQu@(%s$9}-hc$;_L@3=)&YXuDU({vKMITkyxz)Edc3)|I{o!O4clBpr
zu&jqk5T?NRMyFr_ZQ0ik5(-lASThFJ1o;P+1+vG(us_43T+NSe6I5?>#2ty<KFn60
z653HXRK=K_Zp^VJvt3|vxA+YuAT}FO)tWOga)f|!d8N60^vbdQuq*bx;fX`d=eY_Z
zXR0t$0VD?4BU-Qot#llt3x*5?@`M+?VI)d9O`Oz4zOUgB_tZgRh7pQSPuQm)J3jep
zF(W5wM=jv^bmb84_@@)cNBUXr=t403Ra58|(0!Ah`-mF)AAW?65SVnk*v)oL_oE%)
z3KC?6*s2O-wF!Ui3u?H;JEuLf+eGH?2l?<jd+9UKO8T8Zh~TQKrC4r@r_iT8ASW+8
z!aEF99kMXY=X6Yp<_H<zZF(sFtWDL=m{}iMk<f~JI!~z)pUsx?i~8p23Hba|yd#wC
zp!}Jk(_{vZ1|Q*#fB%SyLtE*D|HG3mT<`E0RsAheT*<T3obM9Ns>>dVnHPfbMOH)Q
zWmo%xScN2RZ-4@P3Tz|UjIV$;8v3?zOo3%xWIqt$geD6o--Nj{at+aErl-;5EgV8V
zK%Zx9Wdb?X!0Lg;v3lZcv#E^Q&=-%I2oty7OgYl`9B;wXS1i00{-Tw}>s#c>IkTpR
z?J?b1z81Z=BP;r))(u)?lt~}PIh)$vs#@9!bS3mRIRmZ`?g|(6D0%D;^mX;>D9piY
z1Sl^DtW|+fV+fbs3!C>9@$a`nEL))4{2ss_^n0@PGd|$NGXpzbB={cnKyIDq$O_Dl
z4Ll8G<JU&Xu4rBbSu^?OROlHhHV9(r+VlSK1V-G=;YZUNtQjLAEa$~@v6KhV3bMck
zW$p>mMjDOw2=au0Qf=mFHfv;b*}+zW#@Z1Pu2~|0=(QY!Uq7OxKdEjA3{lTn`R7|@
zKeM<$QmcR9&IkmsAfK>w@u9tr-%A0D7D_%T`bVo^13EG^Is7b1gZx=?fpC*IYak=S
z^7PR+?m+g9+VeFfI4|7}ViAJzf|6*^ZhfT--lr4T6M!hUGIjvYLmY9kBA8{=p?QwH
ziaVKa6?E-%VIpmV_;BZam72NUoTQcF7~E+M&kup=(gN~Ql}`CAhhxhQ-ZyAW5hR%6
zgG4}r2P!JSx9c3iYfq%RD7w8iS}tUdBP3|Rx(P?@NlOgcHim&JCq_vvFL{GC5AB>>
zlEmf~MUGL<?8q{Y)l0w&X$ChqAX5WzFiRWEfr)?-*K#9O0TTKh&&NN^OdzCD1f9gW
z{{F8W{y+kXV31GnT3kg&e6E%*NW++va5TMx_v**YWvS&-=X-BMRUQKpMd}NTGctuc
zwQPzcNepdo9h~Qhqb(eVjoJ2^U(dvy7B@5d<to!6SGi6yH*VKe{(;xK*Nx7I7;1lo
z;pcFnOJ@WDx%%D@>|8C^gGXTJ#6!(s>n{gE4n^&eT^QIoT1vE|l*hCW{BgnnPR@)}
zVyoMg9N9`)qv}9H1orGTx_y*}{f8HP87E}%*npqum<G9YF>*n~J915GJqPW6xIePU
zCqIYHAx_EaM4Zrmm&Sh%$QkQA=E^1-u{iPP(zfp3m7GW#+O>2$1+@1e&0aH-M9VE<
zoDD@tS}c3e7>qDC=s8QsJ#)Y}T}Q~R#Wsnf-|27gT6f95lxfpU2?@t-11k0i5HA^1
z!6HI+a8<MFOD9E%uni?A0TW~_x&B%G;WwR+{48}-eG@P*>mJOovBt{VDVOoyDX#WO
z;&%{F^j^1u(HD1f&YM)LtGm`X?-OtYj!Val*}{$Vc}TRZynEORnm5|n!+vsH{qrig
z?E~Xjw+<X#;elG}Tm#D4Pbu3qx{c(3d8|=yxjeegdFida?{LR$cKz3DchNA%Yn&1g
zU3-ZxaZ#HXvg;MmLU5JtqV1o!$wA`4u<)8Jl{1Ip2}xwa!s!l%b1Ce4d3znujJyYe
zbET@;72E5+EZy6)X?oX;c%z}%)<~=*L<emGZ6dr4L#{o;@?4`Yss>#z++kHRsc3?P
z&Dx{<=>kvw0E?Y%|HVw@@jB06vLlaVr_<IEk-swzqgf%`44-%PH)4JRdN=vBCndkI
z0MPF?EC280V!JI-4gcK|<$&5{N3R?AogsQjyT=WpAf5(KTI(qlvM78h{uEBEWYV;6
z(AU}VGYHP<by;>J{M6<Po^nT8tUsQ^vflw}T-qEIIbCC8f1Nc{aVBN>cB_Hw2o4k`
zZ=564WO{eWRXVRVj7o5wUe%-w@k|GQ#S_Q|PwAD>8!S6q@WSo^^_QfZJnQI6>!`Gv
zFgVgZG=b&rcTMy3a;l#8E-IcTyiy8dRu;ccByGGRUrmX9L6&|EzkWc}{rau_hx2%_
zRScA%hLOst>-E)~ifsDYt_(~-9fP(ZJ5ddbz`x-QzEAhU!)1l>)9QqZ>NQ!ck>C_*
zi=r7aY5FAva}1JYwa-XF&-bv0%*VJkCt9$YI5niOc9AnErhEhmQbN<7q?|M^J~`NG
zPWosv=JbgTb!g##^YM)GV_&lvwK12dh2dn}rX=dwftjlPm?4D_mB+IKTm}t(r7q(P
zOhObJFJO<a)a?Q<VOJw;(H_4Kaol~v2$3Xe{vZN6iLC2sAqqotW4ySshg|G-7*Zw_
zjc%U+Fpqt0g0JX{PZ~;FQl}0M;{c|6-&lG^w&Up}7BU3veg4nD-Zchj30dG{p84e@
z{4mz8sQuRf@0AdqWK1QDIC=_Va3NBm+1=0bHdLq|fU6}_c%L^bHb~asSqCgo17h@Q
zW2Ha3I@7Pp1&Ez_C!|5R*shW!^dBiwAt|5EF3Ahb2iV$Smgj$pMxY4{saV~5AstHo
zXcJn?E<DS>jT-{4cgJ?SY99Fh!YPkaWty??1Aoi+r!RpVG2ocGKHkm1l`Skux^7j7
zXKy8&PbbCJmq4B)c{i(ITAw`dR4jh~G++iUfzObS`!3LRmL}~)l-u!>>J6hC<VNLJ
z-tI@7WVDX_klfxgDcS}!*HRlF-vLr@Vk1`ti!ZxdxrKr|sNHa_i(=KioJA>_E(IBx
zw?E^nF`mTMn!rL<lNz_Z3r^s0I5wUs#HFuic;yBlKCBj3j!pGilODq;OxH}wMf)Tw
z%MBPPhWd5^NEDRAX0Fm$7E)^)QqxHL4Yxb+UPzg**X>^b<zf1uYs~5+ZqRwMIGGaO
z07D7$*oWz{CEZp@3li2Lx5#DZ^`(AFz)s_+_ndg1n?5@XNI_-}2;oh^g6ME&45@BT
zpS$~+^<JKKQpg1D+VJ;i<jcBXwlLSfTt<iyV)j|56t&Q#06EvN8jQGA#BV6WMM-ef
zi4;!W*(C!#2r|hgW=yi1(2;=To2r&8Z^>33OcflUDUpj}H7nKwV+YWHUIX$^&wLY`
z3<BSccsq16A>2fOFm&3yWH1W=t#uJ`fZkJJxg!CoTORNcDS5hV1OJHApwUnf6wod~
zgqS3*nEUOQmZV?EA`|F}s_)MnIw4JnC%VKMMJ{!^%ZdpH6pwO__V+)0=CGx#U=Ik3
zm&fy2kqY9Q)X=-GTk$bEebef|z79L!ur~!tqH^1Im{AZ3AR6N%gr_Kp$capiaHG=+
zFHkWLG2YTu={`Y4<rNQP`?EJFLYnE<Sxx!$YnH;V%_j9`oK3q--NDn>njolYLvH*D
zG7EW@nXL7;qT~S<qc^zE8k|*)T5CHx>i(v+DcQyYrS)v1YiG|dI@(W~sedMQqjLin
zLJ!DOXKhw6ts<RRSWo3*6s_vORAdw=0<LYQF~{)ou%RF6cKRsi^vfx2>yVxefD_pe
zHqaf;I611P8&qUaW70z^R03g1f-gSVwdiB4YIXV58Rs=jMARWdz?*hl0YFG*S6;u<
zK<K=!Q17=PnU~(n`0UquM!w&{x1G7i$yL<dZ=zWX0}ICip3}}w(?;xV2s1uCDybp+
zk07>r_JcY>aSmQ7%0>PYmO^X~_Sf$+1tCQxi4;}4sUr(AMvsAi34_IGj+{$BeuNJI
z@|R2*jOgq0WooPRLUCITlv`&#=!oxZyi<>10SM*sl_v0R!X6xl&VEKR<m&Q1p*#X!
z5UT$>btz-${vPS^Be#a%m{$yokm;;(HLgOP|C%RIHC~Xk2yB6&UQxi`EuA;3%(}`-
zsGV=MybY>x7nAO5Y~atYW~xE`i71(ZB1JxK`359^9&M(|xJpuEygyB(6D539%=5Rv
z$i9VXG!cMyz6kc7DrWy@{JGI2ua$n^QV^EbBvB5?z;|L>FxJ@?(GCyoY!|u+oWIP{
zexK^V)w*E~wZX!xNhh1G*iGWi5RPk06^S~I1-fY8tgs9VOd(Kqf`GbB5KPx4H*U9W
z$fotu`&d8#LYS&@Q(~`)<Yml@MS?XEaBLrQ#_<Hc=nN#;0T*Rmq~Px5nPpE0%9m36
z3#aQU$RW5l&@l6kB=tCK%+zu~*l!5v<nrZk;)!~2%jIrur85U+fNO>Oy6x=`^QO6+
z2_>nF70((cF(v!4tK19kr{WjE-w}8-M~=Eoa9k<}F67y59>R;kGcuU>kY&7=K5rYs
z3eaTj*?ap;j(KE=d@_y(IxQiiM95*)3|*WT#41AlrUR>;5d?ePVfe{CHcfV2HLV@~
zqi=MSc&xkHyeD9K`eO!j%tThVP;5eNIwX+ugIa#*nkvGR+&@fBgxEcl7BnSp7(q|?
z?)#NBjQo4nklR26+BjNo+~VxX>7)rPgH#$v?%+qYcoq-nf3dZ=`-`~AZw)~Yw<iLH
zq0u1F5b2l-r?-iyb>sJsFDpos2Gcv_S&h&ek#uXSKX;hAr9EbesjCNZ`SvDL$oq)j
zQKGqo8y~Iu)p0LC|K<bm`x;_u*{f31B~vxckCEOvM3!8irBQFUwY^*lHAO5=b)|?m
zfUcTjIr|*QE<W|+bCf>5xW0`?Xm|h+)tdUw>dG-?xCt7geA03-y|(aV(lpoWsKBpz
zB!iyzl)d1NJ`b{k8C_?y_Ain%&c+zHpBi2&{OB%6ERHd2H18LxRt-$qUNUE%_5efH
z$N|Xu`#awPJ12>I*xC2L4Bm5yWcW^9h5M_!GhxDaJl!ekhSF4-&lqPVupH?p^V-Y3
z^NOIU<oXw)bI3K~v#Ax$iP=J3Hd}gy_6a2Fl<n5bUBGz?08gcA_9^zyCsZ69{O<Vl
zgZ;5~LO`oMrUq*n*_s%E%ffQaV^{Umu#ObDR@#@1zQPKjNX|&CJmPefgmAFYHM8f*
z0$G@sN2dbcclI#lC|3#niS()^fZacZLepLkk`FMVbmW9O<>m8<aoaJM9#;^QFe^M_
ze0TS_>QuSP5g#D67{5c~Juii+#n}TMj>}AV_qHE@xEuup4Ho@0Gc<bkewYTI9ig4g
zr2!@~^H%&(?VCKeSy~~BkFh;qF$xeLdT2nfQl{ps!gPPn(j*VH2<RB&`XN}LcJPqe
z;fkc(zZ+o_L89sh2d@~G1F}<h@mpcUC((!D5~DXaeV+x93L?+{OY<A>_KJj0aF9ds
zma+|{3{AgHwhx4+);~b57VkgTFV~omt0l<yqwDv#EWSV;-s+ck8Qa*gq_uID#U=f}
zP4mP`Iq(D<Pf7(!&}8{rcxvQXs+_rNm2ln}>m#ScB3QVrlNsOT)Fy`@lR1uHNL1w@
zmZZ{SvVjWj;?-7x*ELjV2Ycp*E@F!PjrJ3wAw(C5NKe?lFHjKKoR^p_E7Hc|tJ47V
z>XG%n;5MuLf+N&IFYvC(MW4oNxSV#GiHy%aJ}cylb8~VjCsEp6{;MR<$eb?_e1$Hw
zlVi{@B6(Z-fiZ(e$z+>Bz{T!GW~Ia~Y}cLg3s_#Gp?s+H@?7O%?hsw?h0Cj?kvmIC
zJC>2fp`*3s?^e$NW$&@GcF-sNm_)eSwi!VfAE^c7mTu#19Bi*_mXyeukb7a%(6}9V
zd)^kl*5sWOo>9nL<B4&zCeFUqFP+T803U)c{vnHwPSG&NNbdd|%#51NICxKVg+7rG
z=w%LDVp%XhRgqnl##@C}lHP`gQWC3N)FsY=`7ry@qw{o;@?N5=Pm#7;p*kB_q9sQV
z5NdOHOW1Sm<;j(sX$xFOyo6j9vmKiFW3Igks(73|@do1q#^GjNCu_a+&IhMfVVyCz
z%_ET_UhBT;l?P)?hF7t6xI0e0UWY6Z&v%20HHo2ASC`&QZSKcZna!YWEiG7*HRG5B
z%_CkWxqiaa&NU9uIJU88=-J!WkkL02?XDUxnhyL`Q9*gR)7<qAySfp|Tdo?T)q2R{
z5?lrQ3z_l|-7^a^2Bo<3N^2&p+x#J-Jrsrn{`5kl0f(mEiq`m(hP%G8bXIjLwIlD|
z2ChG*<FTHQVtpHxj|?vwFImqWLkWY7fg}VBuWb=#7X|VozY4!`4}hvXazLsF&B(dE
z$DgW{IDjA{^_t;9yO6$Mmyo$DWw;J6TD8hy7@Q7tCDh#}NTPF)_z+;}A6&RK&f8_O
z%C0jpDxEArmLe3Kbb&HJ?IDm9r0EsEA_?FGxPlI`{q$-QSS8t=QfU0aVt-v!YEpYd
z8%Ae{JzZ3Sn6Nup(SE30e}l7UAEQai{GHZPP@&VuY>1Lc{FO``@8->z&%PhmukbEI
z?L}A@lmJu@)x4hMt0bo`OKx~y0BT#T^u3tWM0Zw8eLCVX{g>D~5XRI&7=v^3HRlE8
z+Sg^vJ-P}dlj;wztA0F<^|IWXE-8k?&_K8F#8LWG@d)lqK`!)MQ|ll_`E^;_ggwiC
z8{BUb1D+nRuNh;+huW>UzNW#74CFgP8d>@Gh};UfJDCk|)LsJGhOf{^bnEkp@3i@R
zwta2KqDZ&xdopqz3N23c^Z=<D8G$pYv+*2*4tI}GkKj{}8U4}zpi-BZ<HJrIqq(|w
z9#Y7@T`$g>c&Cj3NT<}()m}9A<;FwVY|a-1N<I->l*i{6JZn}Tev@r}u~afwPAnO)
zq7)HCxHY&VeyB+o6e86|n#Bnb&my6%Z=Dyu)0K{L5j7QBzj`lvg4lEnMnG9vFt*5K
zUi4|#9<oquf3%<{`ON}%#e9M$@ysM&(a4QBNWQJV8acFI7n`L0e!@Hg#)ls53H7OR
z3mMus?|~1+T^m;N)`Q&wS44h#ZnD=AuE=YDzYluinhbOo3@#Yk??w@%D2mW?tr4#d
zt%0E;hh7sNPuiL`-0862SoX=C(r}gDTDvt5y~?+GgbfCM_9PQ>c+lH!L7N`$13Gm8
z!~GyCju1yDq0Q9k1)K{)#Bo6Qrs2}xt3Rv=4#4`bA3|wg8b64WtO+f*-VGrsCdnne
zN&m4OidtIJp`>&znL^1#5|y_en2^-B+Y*Y2OEbz#lJW*OrNua~wfC1TNhe6xH`r~}
z<cd{EU>lfpsiX#kFEDUjy?;{STtA*M!(~VWnn8!22w`&tHHUbrzE2aqy2R&4&~JS*
zLh3w~sJGtzYYSd`G1z#=OgNMui?!7T9@_v7*a%q_{TsxuxRa?QF6UhWrOG!>#{zmp
za~m+K`!7Cby=Bw&BGRQ^*k7x}f3h!P0mg7Wxojs!EqSY^Jm2Iqc!`5)I~>DiJ}@Xj
zT66X$Hh(CU6aMx?8x!oSWtl(3F6T-7GXh(}P)|U6C`ue%qnv<ze?p$~Av(Ux8@=fU
zO>`{@04c>WMdeVJ=Y7ld-C_z{83$X}b%45}1V%npI%6P{*i9cEPAR}8wi$35M#$TY
zuV;>3tLt-hd97d+j@xzOzKXd^IzAR3y~wCIQL4Wa(OwBPpVSLphsn>iG1}v{tbI*L
zobk7a-<a%_<0v!jhcQvquETqWCQyS4=cfsz;@~IZ5Q;M1o_0hWK8%kQA<8(<_8E1M
z`ndvGPiz%tI6T=E&p)Qr9FNb~1(1^^Y3B^fJslBK?}uW{A*%CTF;pBFl{?<5&N@1e
zPF@bD64cnZg;}_LrHt0T7Di&rbx7T)Sp&np@*?FQyLW-Q4e*seIW&G=7hH@HBjiZ*
zh~bG`TBBo&Km)=7Zg(4f;cYf;sTnZTh;>5CrY4vGa3wx6A(rt3k>K8&=n~>}FeV>l
zxcMY1FLmh}Tf0Dzn6z7u-In>%b!f+M1LKL@iLg#wm7^*Obe9Jqub{(tN!EivecAuU
zG{KwDda%~Lxu_pvh|t$IgIO~1_GA>Z5_K(S0K>jyq0tiD8kXS~?@w8?Vi&cFY{+kj
zU?rOgGSz*4$RR1>OhjCU`*>1@Wt{LZC273`zvjgVBRZe-5Sg4g5Wt(3^MSlOZ;y)d
z^WK`fOZ#cnBwYBp)L?-e_=~kITBh@_L{Zz>#u{}q6lL<{juL2VGF{+$R~+QTaCuWg
zJLgGax4S^#v!b&m033}TVqT=-NTlhbiO@($q%o)&J;z+Le0x*FqEJ=eUgQRYP>R8f
z%*$D5?;<i3YBh8lKc1$M^$4my2IejCrvh~{ne`O?CZak89#8NI9{u`ynmv}^A$eF2
zkEwW2lz(X<{v8=)pA^SWE?Wc>3kvlT+EHGRY>RfV>%(iEJX9`oXWI+<1)h7o-SjKr
zlp583awwSWr)HqNfT}2MK_-Q}g=aLhPy)Nw7;P|t2~u2Y@`qPnzoD^Rkm&<p%6pB_
z7<0gQmdt!C6C8d<XMdoU#@NA$-+rpvuC1thfPG{|Tefq%J>{)%wpI_ob)&yY{Y%b`
z899PKcus#uBzy2q080aRFTjGu^6(x?`B!^%KwvHh$@Hp$1b`^W!pz-qU|=eD-(pa5
zKTZ@nzQytG_O(NJ7Ep&lS<?&HN{n^QU>(~1FP5qn1W(NzWl*Cdi*dY_<wK*Kqmp<y
z-+R>mLM2e0#GOn_R+K#5&9SexkIvPv6Ra8sl&!qP90F4VzhyE9hDOw&cr~IGB(1RW
ztOY?I?Um^?pdlC-pHLLP!cWRxEh9S$#+xwn%Fbj$4Q`D2tZ}&C7SNnOyI2~>Yq!|F
zSKt+7@lo>1db+Ky$C}b0pkBGrMA7^EIO%tQ)1Z3EzzufP#_de(s1Nnme9K`1!nu+;
z5gCW~)}DQh5lwGMFDJIhUrm8h(oNaEK8yjj4D$#F<#Zi5h6VGP;j3_ZUwLh~%zna;
zk6H^0WuKMUe+DQlYH04w0mR}(9!**N+>>n0klJZ&h*7U7U?uTgO|<=3+F$r_VURIC
z3B{=`%Gdk?Gwzj!)TV9!tP^>~oHGg!p%FDWNXkB=*42q1er25Vuno~G^3~eOKH-f)
zIS%n!_YqDArih*jf4deUp>`Gf9Uvrv9jVO8a(&K{==EIj{Q9X0=FFBngQP-sizmjV
z@`%TEfVd0{R8t4^g#@;{J|A5rzt#mabpkB%g5+4PeY|z1#dN1p@LE51%L|zGg6+%!
zT}?{F@2Y-k9)eoXh_wr2FKu=ubuUON!=5O0aONV!jsdRX`TUHodBx6lN$GqQK9}3#
zSRjY+X1mW=UF1OUBU1%nOcez9#6MB)8Y3o}>W9oLyPX`rp%c3dU`KxbIeZlh2v>+A
zYwP1s((Yc6uL|~&nCdwBkZ-xr_zx3oFCk=v!wbq^K~}^M*pCr$%~y1dzA|vZ!dRvb
zyIue)aUM(D7fo%iRCBlm#!CeC-K$o}gZzDmeY@gkgm=ft{Uhj|aX+9u6hqvMIsRto
z|4x&)ee&AonuAqub^=KPISeJx0F+dcWu$-XkGr{vcY?_visf5c9bsf}iSzVHm`Bxh
z+Ys{i7qG`*yya8+BVhnm{(&K(ZGpuN!{A-Hfu5kP`pt+nqM;?cn_AfAb^sQLyyq>6
z!fXKZL{UK<)YhtAe}<-5TGXsM2D7g+mupPS&_1yJ^FDmq-G^AVl|PBQX`MkmI8C((
z%j~jJF&ZfvLFKaJg$jd!^VEarD}67oJlfN|%fZe_AV2)=OXnhTDPE|g<Q=u?PG<sZ
z5e@2{IL5ECCrkGk828<Sb02DZvM{%i2`<-kcX-4PG2=uzl+UV_M^M}$riNO4)EqFJ
zXwz_iU^mZ@c0$L<Q>fda`5!}aEy|h(VD!eKGz}uHyVEa1l2sz><P#uN3Y)zSzy5)b
zDo6roF@CeT!Hgo4ImOEO7jvRvKCk+lzuxqfEuQj|Atc1fRaY*RYD>1vioP_hzF=<?
zc`5QP>8EabVFI~eAKf&p0C$_<<l}4~SzJd|B-4}$2?$r@7k#|7UHWaaV@*%SUi|2m
zD58Gj^NrH-ihA~1W<zWOu|0kr#~+tCed$R|VDR08h0Ay_Z?;^(vW>PmmRrVSCij3I
zE-gHN5gQTeCO9^n(!rQ3%w&MG)D02$aa`<qYbJtFX&#t}4X!c!ilGASCKSHfpzs$$
zjJ5k-T-ss0y;!Y@Ec=3CG79$MNwGT<r7GNn@ZaGVd;Bc0TFH)fStNOpY&1^*pui9!
z?_5gnufR_*Lo6s0O8yK$XN9IWTbdtZCvsDt(d3|=pJC;=#o#MOh)dwgev|$c4mQF3
zgFX6*!s<szZoy5UD~4ZU@p}dn1PFfoS&e_%pi~RKrT{V($2YJB`2^_DYpnNn4!Uiw
zEj;2hk)BD0y|jjjMpL~xG`4ow@|qJ6+G;{U#dwBoBHk(Nxl&R#@Ew>0{Q44Cao<V|
zb3-LPTFFwpwh@7=UGo0$Uw&4x;n+${);@JZLk?nJ0|?_pb~c?5=Kjno%*D8Hatvn6
z#`H0ZLZ3_SQ_=bd@Zyy5wRhNwRLo{#*aR$Gux3&k$^~#kIHZpoo!-b*V!^lUMVkh{
zWKW6RusCvH_de4OK8*MX*sA6H9503O@o}yldy&b_+Z6U~fTS}%jjBk1vi_EJ)hk0S
zwvP|q(?&lJnOjYe)Q%&9Gmm8%laGpSo?S_F1AJehp2=%cC13-Kii%=l)6~>V1?XLE
zd4<0?U2KKY3ba^FLrJZpk)wjriPI<C78Ln;$dt+T1cmO35UsE8DmJc=N6yJiT)5C{
zs<1?HuR}s%Rf?E*SH&YWqgv+Vyk6fvvEkyE+f8A6us)K)u5OX%io;eW!B%c*a%+VA
zT4h@0O|0uT*tf|s61yVXg@tG`Cfif;E!&eHkSYKj0!+2gKTXh#4`0aQyqDM(hC*Ug
zxgLIP0X(TXK+fbIS;Ko?&%9~`>_X;)BNhqH(cU{7L$&1Q^Wwi2wUl0LxSuV`hzbdY
zH{8QSZLBt0?o=FHA!1jTb$QN@ys#|+!UwSv@uyp7K!UyC?~gfzY?6ES?%=hJ_xtde
zy!#<>NDAL^h~1d@l*1_q|Asv~^Af0ALV+{C=Z!#r<Mlx(nmm7fW9|q8+D6}}cc|T2
z%w&u^bJvdgtletiA3cPF<#%s|iz)#ciynirVsbtKV?B2LKVN>1pOPR}DTjrfSf+Nk
z=O$WW4!8#|6LTaa9|&mY<%j!2o!&hZ5!TGl&ri?JF3_t;PiO7!>e4C~Ng^e3ZAs;I
zqVx5ofjivDwaz5li@ixJ7Goi*7Q{pU;~2q*KDY}fN$y$43^Ck=7;TjAcgzRS=r!Wa
zc81`Pb%LJlF1c}a4<Y*20=&~O?dT4bi`LCMc!EQ&qoJ=SlY`X^48mfb0zPzlDkDen
z<q3fCA-0iNX~NRqQk{cG;JTtCrF9gRV>laB!woMoDkMU3AVNyF((I{lYn&B*kF>UE
zn{yNSP1tr79oUNvF>a;PDL~88`-+L+OfJSN*?s=Kr3B|6R<HcR0*KL;A&r2}J=~P+
z>X1*x_FAoIgZ~-=vw(}w+mrBZV`!&*pz!KZq@R4n*!=Qsdaa>`;tq6@`Ot?9*cmx!
zo9cpsQz}r+#CTdv^xPRQo=EJx>b*oE3<qHVn!iDFbSjDJBPjzgP($(W>k&Ey+jHOS
ziVE6r8`wAf6f+}O(@J%aBPL2$-lWf^kn8E&%bg@ielS2!*!CGOzu!caMtr5ym91D<
zjZ(D18$8AVuttV{9tQ4)NuCQ)K75r8yQ|NLqr9sh=8G${1q*tQ9(iu^thKRjard9U
zSNezOP!a@6r*Q3Xb8u`q(e+k;_%1;F-379uZ=IVsYoB^r;|dAR->yOm)HfVX!^R7(
z)5l>uESccgN@BHsCJlj^$~{o?POhc)kpHa8+nj1s;0=@=JV^xzifsc75K}!_6i4Nb
zz@0i(It^=JwH^W9Z!Fwn`Fs2(lG!=pBz_&f|G2il{ydF8CKzH$5^Xo!GHr8aZ2h62
z=tvo$zt8OSjDGg`+Ot%gFIqBp;JcPgy1?S>NSDDGAMRms?7Lz{P>p^K{p6Rw`d2y;
z9*aGLww)I{d!Qot{27m$8Y;cJNW%={AWci{%3CqA030aO(Yp5Sa4=6N33M)e&+Am|
zJg%wt1d>ZKetY!oPLJA^L{wt*IW2OS=VvmRQ|Nrf7C>y-%OAaB_;n!v`<6Z&tK)rW
z-ZS>OOcCe8K><a&iGs^X;@KQSPQ<LibDNXV6;ZU>)0o&f=OP}%^ocVxOJ<MWHqX)Q
zYRz3~9g`b|i|d6(jLV8L!vcm;M$|}bD6p=(u1Hg9;lPa;X~IDuJdjujEjRD&OB5>B
zyL0l~mdr~m7+lKmL{KaDYak`)1|yeQ!_pmxD@0nb_)m)LKksAnWW*`S5%a!gEuQeX
zDCax_RrZ~k-mNvO!2mrQk<SH0tZZZF7ddZ~4>26H`uaR|sGK9$;O`&e*Fi)(ucxzY
zUWiWB$-U==EQmf$Bgx_e%|mK6>;_dua!9qCw<gZk)e9~)aq~D&sX5OMbVn{+x<UX-
zH;W6-!8*4f7Otxy=z~?8>Vtm>HH(IyXTV^H;j7{ghkT2Q`6{wY`lj04MK!FASKZl6
zbjoU`oYSnwx_oZfP}zst+!%V6cZAL3OwXMD3^E$KXXfv}N;$3pF!~+2*E9F_5jb(C
zaN>M@@56$-T)v+R@9^718(&q4`7$Hqz3JNgkn?T({p(2Xf%FIsWlI?)nKpt2*42i~
zU$$@o+g1)*ZLT-)IMS`pP8%`Fj*RgrhM7%$Ht##rz4q?qogR;+Y46R~_M-hB@A|Sp
z9I^*jPZsv7Gz^XQ!L9Gg!ax7=6~uzxs;q3_X?k!AtW|A-3R<6>gP>Hc&9q42%@GD5
zA(1$V&4$vOgLB@gNn0}(+3{NdDQoatS9fzOW*a5#1UCy`SK+TK7URTDTsnTzNpL1I
zd|`?6j=-{gkH4^_$<@>L<f-H|9;>ru#_{3KeX(yhh7`2v{>N)X=MCvzS1H&*P1ZN(
z(ZXK^H|}pT3|9!vYEK~lW(9cjQ>|4%cayEEB&-hZVa0iunHTHp?mXwt)qU6i8iMNW
zn5k}zg;FzQ2xTZ8{0A`_j=jXL?6&ozSgq~PBksD{1!v@K$3*YE-nv<o@u<AqEAVM=
zt&hOhp#sApF8Pd{+HKDIWd~LJ4fjSEN)fIf=NU4q-X|jcevoJO{U>srXbK-Mr)y08
zo^~$_70m3)7KFYG8t<>4dQhEZrB`Aj>*FPN_;SywdI4KO)2017p@gFG?|7cRP@NVQ
zTtA&hnNy94-S}?r@w!tK#-YQ2o~uhhA9YzMd_lsL|BhASmx9*X6^Hk)nR@zr$TLOo
z{d`@%_Arqo+PX6)hizkYM+@|Au2j4~`#<mgua*5N8@{C2@;5Aq-4&OU-(pD9tI2=K
zYfBOjHyvz&B8UZhawP_j1W*GC5Ut9&OT#3*=ciPkPgd?bs-~^2-C@&o0$RR9oE@P3
zq2#{7QfJb(8kAu(SlQSTgo35`NZ>$5Uw>P}(Pfs}5}=Qb`RV`t$F=;=fBH@%E68k)
zh18iQJoZv&Ckcl4<RR}qxpJu{l)@SC9bcfjr^Nq>)X+26zIdF(d9v+I$3jk8T2R2F
z`(6#*fb=>Cd_T)yI%_+$p1Bhuo9R7D8keU!L4{v}YcBKOO8K7-{okMK#mO#TpVI2g
zQR6r-RAPX25@`q_rfc2X`dwE2+=)8G!57fy)}0267>Lyh(An;b%n45+QB|$J4b|hG
zrdaN7NT)P0CwCu*K{oA>%IARq-+zM|8lOTb1v+|MpSf`U&A|Wt68=#&{p}@wo_O+)
z&Rs;AtA;P&vt|%Ap&-ar0vx4}FZN}iYo7VJ+h1O9Zq50Kem(RJ0BYbGcN5A`URkxR
z&W3K`UK5k@r4~VlRmeX^JrZ(zb@Ppp#%Oev0@4TL9K2p>ErKqrGxQIUONKu5|LT_>
zX%U!WW1kA)?%+zyG?UTpeybw$r6!<t2=s!HHZos0iC15LISDR3XZD&!qT6%(iY*tk
z*Naaqy{;%Igw}*nQA`sIt~(D{r2-3`_EcApm=&uQR;xCZK&}12w{i2?zO)pRQBH#Y
zl_L5>l@2reiO~}1W`EzgzGJ(5p3S7c>}2<cw3<_6I=VUa4d-VJu9`LJ<>We681~w&
zxHR+hKJ}Eu6WaJP7@UvPc`-hOe>Wtg(b62G1A&mmYF%DlegyC-zz&k!2y5aPXxe$?
z<M61Nt^^L0GmN#a9drjCjH!<vbBLqj71=AsuXGR>ae@BI6#ZWxr!@&$Ros>t`u8@O
z{^_Nzp`^X0&yHli*Lm4_*DT%Utjo{|kBQK&g~32|`e9^%RzqX=!uHMI+e7s`ZqR}N
zOpnoTLM`Fp&SFf(WG1HNl{fu8z3)}{9F~w)833}QMoqwPJcf#L8uTLwgz7i(AwBCn
zJvF7KuCCrE^}kS2%8Xd`H|-7ZHnp_xoRV0g&eTf^x5;c8iE`G}&LGWM`1DApx9h9u
zhO?-F;SIXpiKyb>G2P?ue~ViO2=ZrNv}@0pAdk?I&Q~I}$L|y2kBbLYX4UkNQwQ*q
zAHHlt9W9+^bNGNjLf>&V4NK@Q(E+NR5ByLJ$h>`s>|STzJ5O$O^m?xg=VWG{{vSo#
zkBI0RTjoel(erctE@>FYoTQ%F)ck@1i@}x~Th=y@u8w-PPApp#M^M*)W`A<k$`*Ob
zzb~6W>@hC#bFQfL6M<YtNc&cD*X~1mkP3b<J)TgGVTA~ZtwGVXD<3Xir=KsiA3Prl
ztm@E7ahD>l(rNupBKZG9$MA$mdy@M9aC?4!gP$6f-*dgNrk!^?PEJOtUu<%1u-<PM
zf8WvF;lm*+r!L897S7tym9tD~{?NS2n4Q&4N!Ht0`tdnV^-8a`a8|8ie2!J3c}&-|
z70;dV<#Tb^wwbA?_B>QhHO&vsP>2)I>1C*QkMn%=^`blby~HRD4S2iWEjuSIo$zAj
za_TcSa&9hh`|SmCVy!zJMeZd<bxov(s>e;xe~q=$mWZ2Rt?v&>;!qIQ={?m#zl5RK
zu|3(t(2aRDD6$~5zoJJ*{f_>uwN`$8R~N5{@?`#eY@DBd$$rn(e+v&vA-~d}5p_i$
z`DB7%H3l8b)L_9Y?S44(-kS@x&L13V$FGN74oJfjO0{5`kMGQF{Qmj&?fn7bPvUE@
z?D=zlSVF`WE87YP)Eg?*UR~>7xh6j_|FGZu^qvPNH12wx(r(KhuXuL!1Jlv@yo3DD
z4l6P3f6tX=vEdoJPcdlZ^|8G<yn9n;FI1n~oBH5NPS=&9u`7jZudlLouNdid3^3vq
zJUdJ?fLrh^&0~F_?=U&<Y~E~tW2q0tg+@DkMbdM&RObiFE1E3WT9O5t`WEA#9pOD0
zpY^~ZM<L?+`+d}_rDl(Ex`oqjMh884&p0PxHFSUGwpM4C@c2eV04B=l3g5Zd&&75;
z1>J3?SIRMG*&x1(6I(9ad9f&bhxC0x9}1|E+x1_+d{OE^|E8H1TDlZ?EUa@P8fCXk
zA$6cP<_Dt!>#-9t9yE0G@d<INH+Xn?)02~{(nd}|;0<rNg%vRK+?4-O3A39SRaWoV
zyXEosr@B4A7u2to3mjO`?_JctD!A>er2nLD;;P-WLF|cWJMs*>ha4gi6eTtJUmx^f
zr7i6J^lhZLy3KTI{?&ns*_Vue?aNU-><aIQfrQtU5^Tn&{`>NO-jV@I9xWft6+xGz
z4Fb^O&P9%n83#Ouq+uct<QRkQ8Hf!F2=b+VwLT(bLCn?gnHux;^Ke|0PfIOjCWFZQ
z*zu(v=Y@tx2aChRTV|tljEuy+)Arr{^06^~J@E}ce7hJOHu(-lnI~LdIWR4DHy85L
z3=ZM94fd}xneO|SYZUtkr+*?L9uJPL-_3)1`97r1IxF%QdMn3v-?0M!CU$Lcg^=5V
z@sj&qzloU-MSsL4ewPXR+=r|V`m8&q*~WPNm6@R(4iuZ8K9Dh(wnRl1eBn=x^EV59
zNd}DTm##+%=&jP=8N}~Hi<JahjwfUxR1B*`#8d{aak0<-xctpI`$+3BT$%3#PT1?3
zB{@?+bo-|u!BzYqz1iR|1r$&8xCEnmFXWH3>)+(;K{V*MA6j}Z_T+pM8B$oAkOYaX
zbgH>8H!Q%Q(t!^*J@I);pU?m4wObjgXZd04hPC~a4}59h5H1lSxecHAARUr|a`}rh
z6u;FfErs83@6pR8s-&h)Y^(nz+P^O2K@nU5k?0;<8b}qP<PnpmBQK;Dr6AddE+R$#
zZx;c_VchH*C}wz@ML22TI0X1<2jRZe*jmf~x^L+5E^*?u{_t=2O=|=R6Q?7vAAd%^
z%0fTayZLYH%_G1lH90?%&Iw6v-y0wb9)-M_hv0GzeA!!>e6aWYWk+7p5^S!&vA!|o
ziLdF1_YWuf=WPeU7YQ6kn`VWA)GT+5E@Vcjf1U^PKhA^Z97G5irH3F1OO4La0#>dN
zywq&(YoQp&<^I2-4qx@XuLX6ZKAr#TeE$8BpDFrM(*K7qbw*#hj%Tz3pYiiXzF^-p
z6qTJ-)Zc>)=dWk~<B>K#fosd7FbcFkhA7@O>(@F5FQ?r-{v!MTw%Ss6EHvH7D8Z3u
z2OBbqW9)`Yt2~8u-Hm;jxN`fiE$+Z;8;o7@`z1&J`P~s2AXj=g#UNG$^D&j`8#h!^
z&05!N+A~Rzh(dfTu+RI@^8(=#``h;hDN7zET@X#o7#Pq|*VNQyms3}dvbMH9ZD=@;
z4@SKd`!w<zS-@-bU6V(taxLP=jTj#0?BAI~0+usK-Ze^8Ct`Y@_n-N<)$kpKUHDN$
z5UcD&{`GY?ElPNG#AOhESP2p8zc0sdzq-MG2#>;H5m#LOw5ZX-wU5WD{;}9%$tZD(
z<msQ6LlI&%7x73=q<?mB+eZ3D^NpXutHo@7nBzoXgIiyjjq6MU(oD>A`PgKL<2dap
znQQE++pZs^1{x6vDxGvPV(d@O&pflX)omP$IA%3JT&1<^FCBo7aT7B?2n$Pl1B7J4
zz7F@Vi+RU@vg~P0g1f=kUx(%Qql0H)VcU)yd87ANz?!t=@L_hNV01VZD>14E-px8d
zb6_?aIs+t24j$>Hp-RJ0&FNu=<KfJfEeT#GG1x>;9y1w`mj_5*8eY9R_4^3;oyM8`
z8qY*oh(7nPYy0C7BP`NqWweC<>AOObo;VfCMa&J4e+q~}$d@L4RaKMk>o*N$hl_`q
zjP}B}ior7AR9jfXmjc<4e&Vc~1JCd85_34%HdxF*-H`n+R=`gj-R%RRC1e8v2p_+7
zy7D0wxv6#P;EsYYw?;6mLgL7ubK(M6EBU%TaPfcYBLYhXC=G-``u4XR5P?|W{PpL5
zJVHr^pUJYlb4a|rL$y$XC_)~6p@otV8F(Bjw?ECz3xNtGFZF9?MH<r+9p^&+CA5+P
zU~>eRMeL=2o3r_1kVP*8vFPxBS!Y*eID7|!tL~w7AO&?=64=;=ZY6kE`<A$tW*$`-
z?tdK>r`a*mub^LpKKmdF;VW^B-2Y{5K^YVmP1W1~%i8YcKuIj|<68UB?TNxZeaD!e
zM4xpL9%Cma2q|fYOAQOyoR_(JdcBgHDv?T<;}BRj5fG{w%*|Xqc&%7+BX{cW`xJY8
z3?xeLuDV6*e=O^-KYcggh6ee}_oMGYp4_h@(9dAbVw!29!4r?tv5wAtB!HCYxd)(t
zcwFa)J<q)D5t5>DEvI7)ps<a-#t9d#d=5S!`1u1{Y%n}BV0fOX+PLf{$am%WrM~ai
zmU#3{e_i08kNCjel{{&9h}QWA&aWJM3BIY>?$uu0b#Okn1G#b~M($<tAu#=<!e6EK
z?5#j@;en$q3P%0~GV*!zw!e)$iwfAB7|W4<-almS*X^V2kN86~R7xOir_8B!5dL8#
zu`6N}wE=Rour&oKQK>d~LKAq8tYEF~!>tNylaTCsx<pW1JxNy%_V4EhE6by5<8$j*
zr2lfamAYuH)sJc*>tA$6?2biY02#6a0PaL`eM6BX<t~We(Kr95#44#kiIt2tyy(YM
zqjJh%U7?B^Y<7FiR_Y#9AaeG+5{5I>5a)8jxaF2+x~|a0!)#xdrhBvs@v&wVi#6Zq
z-)0P<9DJq0qZFd<y6nW>)DVIDAQ6@5az4+TK-TBao%#7^9f~wPt+`PiSlJa4@TBGY
zGtuknP%sL`w3frGTHy8D?maaW1DTAWPfd6BU>QK#9`x}`68MzLXxJV*2Z!87xPATo
z=4+*R53GlaJqsC*pP%I7@h;U`=GcKsns-uQC1!}NZhea$YSa~Kcb)(6S9gW!*RSnk
z@ae6ubCE)$)rEu7-{3%lu1{<C(BOMu1-rUnwNjsE_p>-k62Jt$*hi4FiO@$u=pOm7
zCl_aKvznMQfM-v^^K~IvEDa}dpWrvO?ua1PA*Z6JT-_D!ccu*Dr`~nAIzl1~ZMwQC
z+TK?;Us3(@LZPSzUa55~8Wr-3z$W)g>D@*;LJuw%RQ0=2^#KWD*A=5JJRy&lET{2y
zK@G@Q?3VtE=s-#GsgZG0pc{JO5KfA7>C`;s`FZ1j{fRkZ&N0s~`%~Zwy%&0booIK{
zK>$k|w{1`W|D655SucW@P;l!fAtgoW)pF#cZ4kYJjf2C1REylV$&v~?tk$m-?-Q>$
z%HBQKR~||M%j1`fVz^^wro^zI1IS*@u6vlRg5+Mt>xs+|JB6KiOYo@1_?7#|W(Jfl
z8wry6S-h{kQ;D(TvPXM8m;if?G{djm`r9M1XjmkS|0pjKj8u5+rNbW$jtdi%C&8_*
z7sa;U3(8-P_m0Dt`S^X*h4j9@bD#t^%KdGT)IDzGwy4C1_mwH*^bGr>t6@73djHGU
zOd}~FbR{w%ME_r|<SZ=BXUb!@#9p^ucU>m3^D?^5X?tJTRnNC_i7MU0A-y(ZEMc6`
zbw%-v*2M*62iB|`!v67$flU5cZW#_kHEcaHWMi-C?y82uR;~IlrQ!Zt&r;w1v{AHI
zVHK97n<Zg=edFL_TvTInmte~U3_F_cEAGBcPzl%1LQqF9*jE@7;`^mf=*0ZPA%|U`
zfkn2sTAP4Ajpnx^4wtiNhG)bsg0=<ZMaNO!xHu?aKDUoHfw+?>>^*_lV-`4Cq*7%4
z$i5rF(P|QyFM%7V(0s|?qlC~b3HL$z!$^ARjgT?u$tog(8~ex004EKrwU~kn3qER?
z8I;XNnocHKhCcAQV=K8^Zy|G~f3w*PGPlK(4jfpxYx-arkk$YX+>Zp*qtl5^<e<+T
zKOMGA3>R%X!AJ~x9NL(*NFnIl^~wI_`|IGdoW9-;L6W<65!_zybP^k>Jpm5qKK{Z2
z08U|dDVmms@4-#R7&(!-pu;o9g|&**)%I~XXF~(Ij&O&xL7?~qNW4klsfnrMtpj)v
ztQR;v!V`Mm{z(Tr6J(<Q5wYgKI)3~Bd+8UDlSaNgpeyeDMpvLH4NqtZMi;>gnELc?
zKYQ!M?*0L)AKvJguO580>#-_=HKxQBMTXL8`L4w*deeIWh#_Y!9nN6&U{8LvZ4S!#
z=beDBj7Kg;JQD-DK5x(oGj~{XrO|Dv?8;UmehjkR*l~?7p-A683YFWR`yR23G!7m*
zAIfXsBVr8MH1(!WA|2Qg1REj~Z1aBzd+V?$*S8CFONfA=NC^n2fRuoObP6I!H_|F8
z4MWEODh8>f)PSV4bcaF5NHcVgFvP&nd7jt3_wW0C=j{ER>)O}8_yd7?-{-kwt#z-P
z0#F~G=<bKc|Jo<b)Pu*1R;8l=?+RNAzUI?^Qvv-cxqp@f2mfbDaQ!``Mj;P2BP5~o
zBG)O`_fFN;ox&&*)~&PadT6Dk8)`jau-55&O`q!+{S7mvGoRODYwVJmPG+>#%Q)fv
zg=_bu@rC4huAfCl$6_quC4%xo{smZe<mc0HR0zDHwV%GrW2G>paa!BEID>lhll4G#
z(361%(J;H`TU4}yBU$Ol4lYc6JRY^U;@6+MiT8$;D$F>JQLch;>__@3-#hWk&-RvB
zu*M~O9VEMAhL+-Mg{9<@Sx<D6a8`O(bO%@YjEjFQsZ*b>x(gN^#?wn5EA;vnT)U_C
z37;?S1@)$RFoVun9u7PgX|tt*B%hd#yDF9`dvB%9`@W>_7*V12KhaBj>5w^1CSkk5
zT|E%t{{znjq_8EKe~!%H%u@kb8tlg4Z-|N`8RY$MC;<U*72k(~=MOGiGV{7B$eE~@
zZRXIK_i}y;C9%uxFTuA;To-`E1n&c&HfuVW4o1uIU<UQdPj33~U8kq?!*e*;hj9IJ
zcp1Guuht~4>M=PrJA>3vBf}ITXDp|+YC@}V`~wLizv<WFok^w#y<berT3>sf6&<3T
zROxz!zApc44y9Uk&%|52wUBZ^U3jLn^qqQzM!l5rP_1GbyTUeh*!(N5L-o3_TeCMw
z`H^A%jCXB?JoV&Z4gJ(X(u!Hl<Zf-Ef(JL&xDo9fyeqA2dF!)jpOk=WKX-b_u)W57
zC%evGztip1y4?)gj^rI<QwNN><!wF*v)ti?p_*cY;=_?sD$-!#rVA2AqQu580|g;q
z12)I?FTw1Ob7*i>#hp6eLNPTSeC=B!RzL<Gm?G#ap=kEzKhaD~0-fI00D>MA0O7^g
zF%a(<1@NPUm$M2%e&-RYGlQ?*aPJ-zt);<?#%<Of4)N7#?-U<Paqp^tEz855a_m29
z`SR<zLPgU_&YWs&Q;%zIGE7C*NxL;4koe=C5$4rd2s$p{Javw?9|VsyS{<B%e>mV$
zwel-Tw{d(n?&aiWv;s8>8?=H*DQJzPH8R!51*;Ju%82uf4BQ4^ImJ(XI<ySv^zv@R
zN1FOHM$PIl!QEHAOOI|4m6W7E6@(u<4V~J_-gZ`-5uM)zaKdc}-Q=gH{kJmif(U4y
zP){{pLukDO@89$&lNnprZ~Wed`}bo~5B|R&lNR)tzRwB=`4J*iM^JezP8!bW`!wQ2
zM(^+4Xx9zZt;4oeij(G$N{pNwiRqH17Hyd-Skaw2nEL9>x;wQE(;m!8mz>$i4t_dC
zThXOYIX;s`3Pzr>-cU(V;xU=2*-fz1abGQ_UX{9z9(#Zh%7rgq9BMiwA|;%Fzl#n+
z|NJVg{VVV;iqNjGK+HJjaM3e&y(X#OGzHmQHi9S~zH{sC+9<AkO^zThqFN0q!}k@c
z3UqrY&8baKl4F%z$)NJb6CEXu_s758O+t`Yj5Xl5pmVXs3rS(41)!FA=I!sR8F7vE
z{3EC;{;GZd%|Fz<pf?8nFsRMr^JS#)kYAbb%_m=W)viWkT%#wLK{tTQGjQ`Rz~v9=
z*8*XR$2E*wLEzA=mKiNs{%i_ip1*B97m?YgpUt?eCzk8;TzRdRm~?xj`6}+|tu*6c
zTKCdtif&1)G{aQ|rVe}Z0xW9_f0k(nSzCt>#8kk`RuFaaWKO+m-kObgNK?PYL`{`Z
zlv||xx}QAtsl_`g4sGXnIm%UH2_qkI;L~9UQd9TSf2<I|$G^;x3F3|HiU6cY1@Q*l
z9T0CI0FE6bBmhLrUAz163b;p15;UN(d!h+pK?|xNv`u|ZfTX1yuRvZV{nVR`BB^Hd
z5=yboe{*7LX@*Jicz27H!Qc$>EtN-qp-lMMNSyl=w?`_m+xa}P9O4C`;Ollwb<M6h
z=$LaeNK^fv1r-H1pEmTp<_U2;eEcDtnS=hhMp|if_bFg~lPf9R+%V|8fO$R=M(0Mn
zSVo?3hU2kDlJbo>_(ZMh@BVB<@?*&{Cw@x>L((wSg-SY>y!Ivaq5WBN+v8ukkJ~1)
zukTZEP0x!EdP+SA`bo}-7ED245DoXqHja7b$GQpCK=peFSC4&TLY=SNsb78n1D`6{
zw7$i6q7eJzFN^)9jz{?g+I$8j!ou>yP5Frd#uNPwb!*87AQtT$*j9d~_etV9OpZFe
zmSlZaCQNNwSDwpcxmaQ?HG1+#tou3-;kZO7$R~mfR5l}mn+5>W%0h>SUIw+V=hC%*
zJF6F{48g<x#NB)hm5y-2U)W?bCoAYmCxf;SQ~|kVKSUhByYj^<g4*Q&enE6zmHm@L
z1St$tCv_{SO@a=OMcrmvO4i$hR(LzA#`2;|sYBh49DM`Itr}M1r?Xx!p2fR`FV->U
zf{<uW>dw9#buC=+!$Z!iAq-+$i?)LGo7O{3yE&Q4ywfnR0&X{y+O#6YZbaT}!yBEW
zeRTmx)G%CC)%2`v#dP1j(?#`qg)7#J(32^KznlKxc<nLKN;~aYhmU{XVR54DB=@?q
z_`}r#f9s+blZWh0yc21yv81{nk5Hko@YWpyPo{#-m%O2bc&dL0FXxx&x$~)(O?1?{
z{uD`oC10K)6aIWUB(AeJHyDPo^$&_q<tyv`83?1MmoN<q@!+)8V#l47dUS+y)|DBF
zS{D+5>c5cXM++M-2p4~m6aQGhVFN)}xI`7i2E8M15Ezmb?6aXd^4g!h=-2rl8ii0s
z!b<$}8-R52q-@A~55kFy68cOMdY_yU1FbFlEfdr_D%8)kX3rX*JV#|GISgqEbM5cE
zj5;5<=0;sf^T*V%=wU}{&P59$z6sl$$!yG8xRLvI56SMap6hp2=G>)On853h!ZrO6
z$AeT57a)+LN+&uG_`a6xadyakz4hvJ3HW78aLzxj@$e$|$7$&@`5A1L(G-OScVbNU
zbz9xb8`8FZa+`mh)UmF1S)|PKb_>gLNg<c;JviVRcTv|(b$-(cV%y*?!=vC8vs2a8
zmjng%c@fD4A^r+k1#-Bn1c*W#5Ou*JzWA?)7zkmS%Y;{&|AZC+1lk*a4+njBFF+>&
z+PwPiSvk$`D0A^Q14TsJ3L%F<5&$;p#pl<CzBQN{fJ8ulJ6RiO1d?q+h3)8u?y$>9
zX6QS=(+FFpvv*mJb%jJ1R)1;_$vHtmFzPkpdun55;a&=Y;yUaPFC&hF64HJTgrlTp
zYeZ#9=vt(mn4~_Vzjpun@GTKTkjp_qX)`8iJylFqc%<cyoKAl&e3sZ(yhEr9exiJ;
zc9e%tg#Br;Ev;!Dc5I~7Fm;F=xu%vfJs4iI#!dfx<;BSl%gWus!?~4tC;bEAhpduL
z*+$!OzHbca$2&j*OXVBblyGalnp<$Qgw+5cFt9e^=B-XG6)9<AG@Q5eu7so*-m9fr
zeIZB-8VSkOL-_xjo@)w1g$jMX(1VKUH{<R|Dj)?VF{t4MNJR4Oht%NMGT}kk4T-WW
zG#ddPCW9a)RQLx|froixCt*AfJWTXsh2azU`+Ed-wHHRD1Wm*95tX&w)tDD`*YZ|A
zX;Y8RVu<$?7%K0%bHL3##!3VT1IrA|)6yHpQceRqJ-|lJiN0bf-JW!pO8kVs(Xj}I
zBK=A30o}7LC)ilA>#rBrFiXD2Yg|WTjqmK<-unzk-5!$kgy~_*u6XfIn}1f3#2+%b
zvh;<%TCJ+uZGN11RE*ARcME5Ucxg*1voXf^Vd&w=hJ|1*lc*7hNQirWjzv}4kp)wJ
z)x<$JFW+G%e=5cL0z%fn$NQ)M1%uYn_!r3?0CHY`Zw3zA!H(5$<b)G^O@QD@$!cwV
z=no|3hP7_FaXd!Zi(T<{KoRS+R*~%}3f4xBK--OYO)Agik&v?{n33sw$By`H;(?M6
zl4<TYZgHgm4_q(mZ{YjCK?Ru`KoxJS=)Y9SHS=&rlqQhTC8twtBcb{tWuX5FGS7GS
ze7>@Zk4Y!?b7pHKLaT=boy{oiFoSX<gkwyHq)wE9u!u>@WU)bA32sI{uSb56C5sPd
zjg4)KWPGl3vUR3C8O4d((AV<n;m-5^b-Qb_U|cP(sLtJM(kQQHW4qiuC+|&cGi;^o
zvct;y@xHplNXZ)68uIbyQr;uhFmv`N7@Zju2B|MUJ)rBoUqSN9LlGO|SE3~tnc`Ge
z=8Kalxg;*)n25LL%UYP5;<Mn#arKXLsetZ4<pi6f{^@~335x>UZ{Ba|rEE~I8xBrp
z1T-!)qyGDmYa=FX?g1MXz2{(lr55Pk%+HcnVsB%Qe^9DJdZ!YjMqJ8<&D@W7V%TPU
zT|K=eWKPRc9%#G?y4n__<+L(yFbmxM{R1Vz51h&;=7z$uF+h_6zI_q%F_0k$9dc@&
zzxAG&M5kQ&v2{^U+tNAnPsr$XErXm%O7!Tl*${8N&TNG$^TqhpSjXwU{}dB8<D!;#
zP;RM(C$bgl>N@j8kz-84Za*#aj)x<&o^Z@gy{<yHh*{<MW=$5i6pb#os*enfgjn+i
zp`TQpBU$(P*jc-+34Y156omOCf+al9*Q1nvs-q$Lc^>(=9v-U4QH>DMaOx*GwB`HG
z<Rk3Dag~&>?ss@jNQ1(ywJng@ca2DeV?IKVa>0x}EvU8rsfhb8j{FZiasW?wT#cFi
z9Xi2nzlBZS{#<BMr&l%Ms(kwwaEV?DYrAVOa{%m|hQV$sM$Zklc;j?n_M-<(QhG$P
zr=h{FuOwIgv5+W-(wND)B_u0!gx%xkKhsa08nrq*W(x`jNG?i*SLFFVPh2rOk?Z`_
zzb0O~Dd%u9{>Z>xTy4Z#3|JN%Cg*jsu(PivCfUuPl!&aj1#`t*OSyKBK2h1=W_@n`
zI$Sc-t@Dan^BE<kGBPl)99g*1?uZ!3g#u@>M+TGb7Yv+uurG0^#Jb64G2OoFavh%u
z(Q`bcXrC*L7$CSc1R>ot|HB^t?+3;V9>}g{$zv|)00#a|7K_}uBKJFDEPxp)0Kr$|
ztZKy|HzVmI6tWx)n{-)&2L^^$J8BPC@+n)d@>{$G1L3(~jMji@IucCp>^>O+x&Qx*
zL3`Z?kdf(UJUc};@VshH=ty;kEmL1{W{?@_?j85odY$k^P_eqr$mdCT73i48-ZA}N
zN2*h2W8CT8zgD+7Z>@v-nP5<O*W`}qpOnPdrVETzLY!7so%TUbxhFebi@{gJ`i^%y
zqSR2J)NRDf=7Aa%Z>^54{UNAOd0?-NsMP%pYS-5RsI+7mc#{U?6bjn-!hb`_q#(d;
z;9H43bX%370RiGR2k06H{l<wLMPGbSd;rGhidAT>xJ^=wQYR~e(10xkV~1+MaX``M
z2N4;4ra_t2^QMz4b;VTs3C{LDc*qD%G2x$NH#jF|CcMHRKp)_N&1I>1Ert-zm<-Je
z3EaT1=9qDt5jB~)dpJZtByp-8EbZU%fiX6~{8e5=t!7XjBV{XTu|Zy8fVAME#PY5K
zn@qCpjIsFLIzh+I=AxH3w!yv26NA6*6kCt_60R~?m~w$i-1W28MJ++MJg!|)p4og4
zF%5A5A|K*(`_AljnfUL6x{z2eoI6#cC&T<2r}j=eifel9?GGP&gZxDHR<8O(QH7<a
zmkB~10j#w}M|BsP?)s<vLiGl8QFEVJ>YyuHKmy43JgO1sDMMKrh+8~vs#>GL&OtDQ
z#aqhgA+%RTCXv*<ofU7I2fTc9BsRgKzPG^$-|b(sg*o3o4S{QXp}*rP5Ts<@G6$n!
z)1bgVDtuM|MS*q$aIQot`GIZ>APNJ7a2AHC5l~`?RYsSe8~?K^qy}C1BeZhOhV#o6
zv!j<#^{t25HktA6vd%MctD%=Z@YFkU@YSx<jN}>>+{EAEcc0pM26&`mN}VZ;4WTbn
zW~o};Bg;kNhX<e(JUPg_b_@$XZ^}x@Jy)1cAy|K)Kg5ezXcx^FJG`a4EiR{Z-LYN;
z-Tk41j^G?KloZ~`n|J)<qXMUi7*ut<K{w0f!7)pV1p-)6a`3Hz(!UQ`M9PG|skeMi
zibo=t5{=zvO^WEgmF5TXvK<(V907K9#dYho`HN%V=z)X{O?yFxCQRN7A7a$sEaORn
z8DNkq4spavWwuc5!}*WARq+KdC<x6NIuIt4OXRFR<L0d26%?#@smhSQz__g)yH>66
z?PJ1=$mjUT;;B!yv^iE4YE6U)0e92N(4XruzSI+ziq9lhnzSb89@##2G3&oUO6Mi?
z^#HVq+@pm}Ai3w3s`+wgpfpKd^E*mU2U!<On>kAAJCJ_pyfFYvElD#Dl=%S31r$lU
zD*tz=`kzpr3IWdE=M13A?g=!-sd|SOf-NL|E2qjJ&jR4=rU4cosEfq0BfQe#o1tdE
zfW@&c7#>)~2nq_)i+dD+fQ54PA8&@?Xm5UaYwEwFYvMWS{^%tVSkeyyCt~*>6bfDo
zADkV579gd}&7Kvyv;084y${j&mqJ^kpo=;IUDUc8EFyg1qSmamcjS$3ywOTq-s2(J
z>7;t#&i!$@t{HdeSz+1h+}0VxUp?_f7u3)NuSd?J+<2?-k(S~+dIv=cp9{v}C4qhh
zM@(9l3WS~cnKE}UC)Q%1BU7W(6EbPxdA$nqqHbhz;nw7p0JFY`=-6kOKD_`b?3+*2
zqJ-*V-Q<&h1Hw*$4tX`zutW<y#AaydS&f7LJa~El+<?kFcd)sAxo-Nz@!RrIhdTTU
zrf~vTqE{EAh$V!ltHGq!Bs44uY#}I@4wlse@{F7sDt^^M!oV{I!&W+{3y)@BhsC>`
z13wW6PykzBT3}BC-AY!Oc(jVFU&Cd*2qHvejV|?31lJF}xutHZFK*mVFjdX4KZ&h!
z^PNs1#r21hF97lP`An2#-R~`)ZVD^mwQ*~lOhT^#hL{m>n#Tv7mEWr-zX95uyLBfE
zLin|cq<o=w|4(P79-yG8@tvpc&_x;me^&~PC6JKGT$cu4D@$PO!Iq8_FlRZOdUlMi
zru6J!i~k|p4*YJ_Mc$zZ5Ixteihu>VU27Pfi#hq#J2JqxRU0hI*@CU$?vVmO`u=#r
z>*g{>o+6<1eg<i9?5QoKq2C2*)Qqy{zDa=W8u6x*@`<Y6O*|wg-UeOmRdG5{*{>-E
zpAQF_`&7%ICSN@&1t&Y>9YwD<fmW0RfbO^U>ulGd{+ol}==Xf?Kh!vW5VG0UV!jN1
zp9~H7JY_QTzqxSyzdjrAkWF#xe*}%icez{x)45=Ptq_=pvP-6;&LeuB*o_nofkp4z
zzyYK+{AA|kC0pP@!Q{o7Y=j4Bk=u+6@0!cZV=jWsBn82K-_uJhpu<B4KAQM^J4Y|{
z(W=U!k46YavdBOS53k@ljL^DXF1LQ^(Y=Cp{WI<T0b0<E>J$3Hxu2?tql2MZV&W1T
z%XMI!yxzN8UrqndFAlt-Ch#$Em>zrR<9+&@EzJJe^;;_S$GN0GKcxh0NyWsiPu``?
z3XH-+w;C@<rhh(0c?NY6+QQu`i!3w~p7P2ls^3ydY8x4khv;$tTW>NIU;a}&Jq6n7
zGEY105N;OiH8XedbHf(kuZHC{!x~mh697vRkhFalmi`?!^fNba^eE<botOf(LxHXv
zMofOVSYtohdI*~u90DKO{`Bhap08lC*ZT&@cQsT^ZC}veAfDEFxBGu)$p2rR*%N>x
z{;jVIkp~oEi`JuEKw@Rx0y>~q_{dr3$H2;}UsZT=&~;r^A5BO~^Eu$`lruHh7#|7w
zmP4Kreg|7|qP+E3S!jl8XlO_tcM+<|IF~^on+^$?+mk?U{@d(89=ucE60WSe?0Q1P
zce1>I9rB?Dks-=aH3Iu+haUk!XmrI+(1y5P^a@3pPGg%3@w_F;d9-0<=U07>dssit
z#|hqi;T0kFU_q!VGxt_Y0F~)cBK_HspK`S*>JP*v3Ezc-ygysi50B4Y^W^z184A|p
zs=9YdoaOV-w`(m6Z4p=Vu`B4?Vc*y;P*s8YKD4fJmHc0h2gi93wpujgccEtoNY}rF
za9{=*@RiU*GOBVMvWz#$231!@at8&=LmSXct$?832}`t_@dg`)O30D79#0;Nt(6Zv
z!~7b|jgCh|CJ)Qb07;i8B=w8n&H%yyX{f}=P%D7=WM@rKV(`xKm-C#<!AEk|jVjrC
z`66sn>7DHS9Ziz$nuUgHk7>Lj?7hs4))`l>!aAQguhHiD`=kS!?B23*&0t3Th$jCY
z(U_-tO{C4*dBGyXg4R7IjrqJBhm8^L-?@-;TKyWih5q&lf-ACGFDPW!rGgb}e`Eub
z`~S;|K{ufYWI<>t4qch^piYH7fP!KG)NIaz6tsicdC7b%1B~_TMH}?Qc+m7-y*uzG
z30p{0!U?qLw2KAu{Y-7Qn1$(MrxLkO{L9O5i9eS<qO*!lq9TvRQu^i0lT!we8Fa<w
zIf|0x@7I-9eW<G{POX)I{a@bdmlNq?U!%ki-^P%<S=B%ull9nr7dq!ZfK(2w%%g^$
z8z2|87QdQjZRNe<5i0VN6^!(mA0etCupb!e^$bGu>F4jJ#zQowLWIRpU#L!{_QtD{
zn{-PZSxkTRfl4G`7KuJi)UyI<oK*P|mfhpa22}Rwd{-=zn%?UJRe_Bhz3z7-n9fl#
zw}7V6p-<k7aFDgKA236O1U7%wSpSqW(k_7aO?#gL1<3<|P>UQX^re~#ynF2X-}n(w
z2KZbj{*#ja-D$^DR96+s079e2<AmgSuj2C)-(D1>bA;QmQ&wk0M;j})e$sNryMI<+
zNqn8t|5R;hh95-N-dtLRo7mNsyf2Uv@A?KHvU|lvDQ{6Yi($!ON-}?I0gO+>De}Zt
z9unqxsC?upJvDJi5SXBJZ)baw<cJR^l~53HGp*Dv*fP1KKlF2)%4V8pwh~(_T;V?T
zHs-sfQRI`zpgekb9YhvbtbYEMdRw!{oR(wkOG~fbl_H0|$s?-OgPvDsOn~cj4`t7N
z+Tu^wrjANMJ=`+~by=<$A476$WU6M5oVAr(sq6jUK@VO2-<<&WcV{v@d0_d0PQRVc
zl4J<1(}8ct9fj6TtM|Yk3xn`o*No>iDXmHIC1HJHyZfE_^p4|h)9Fre$tCBdU>W{%
zxO3B&B(Q-rJcgOj21XFU-kJ=5k|SbtZay~#ns7HTFoPM?R@ar6AHUSWX)<8J9JZc)
z9$Au98%!sT+Vfs9{%AOwQgOLLV50|G)x5R1$qX0v-m*~rc;`Ma9r398uytqw+k|~G
zuQwq3vA0q%>gTClnhJSG8<5fQ|EOqSO)A^5_%9z7@y6HxHL=xSBH*vCvQzeP+)C{}
zYgtvlMO*JN*vtWe1%&{pVVm#(tI;rfO=k_jGsDqZv1lhqru3l_{`reP0lF}lqt48{
z8pw7CjTW16wTMBzxHXtLSPf&zfcmBbfWfIiG8ScZ#y<rgam_$&(Xr7(<yCa1WVlVb
zyhzvEBe|uv90<#jiBdDPqQOIIwV7nj!`XHubxB?)t`@r4f}k}ax-+K;luM7PhoO#g
z!30DsZ+7c%tWZ<b<uRc>#6Fn?dKmU@nnHtOK#J!B#Nfrq&GA5o?BF9F>}`gp<0`)D
z<No_6Z*<GBZJck~E18y;!W>VGE7T*jL7)=gkz5-YKrBlTK`>21n&r-yY0fWlNP>a(
z>_{CPwkE`HGc|eTsiPX53_CWtj0b<;45a$g;!j56po)mIcZtO8HN)=FhX*@?y^Gn*
zC%+V0ObYNeib_iE?tVHSZw9XsvLAi#A^2C;?!+q|H5vd}TTk-1I#KqO)yc1g7D3u{
zk6b*cP>&%dWP)|yjnJg<@(ANc2ko1t2z)9&j|FX;({y8^bU-7Bd+6##;6&!$|A_rL
zSeNmr4In@a2Lecm3{u|ew8igXhLI0PMTL!ur-AquL(PTQq445G7?x5(^-c)5_}ogy
zzXw;<0qI0cgb)Uv9KaM&2kg9Y6itAd)DvH@+(;WSY|&Km1Wl<qQfQO|b{@4!J+C{k
zl-j{oHJucKO;G5D9M&#s3ZQE0?HLDUUU3po--RnV^~9b^jV08mK7n=Pv_|d)T)#hK
z&?!`%yyoAl6R$Y|iF+#V%>UNdQGoKqNlCYy)Ljf~pEj1_TKuApb(x0ZKVWBsSKOYM
zl5X3nA<5eHAtz0!Fljv0x%G?wbel~Oz8@Kr;$KEKYGqixfrre-tpiB{?$<{Zqptm(
zmDvH^siVvv7<xNEFWm$D^P3oR3!})q3@0u5CIMD9_yv7H;{p}Lo<NcK0O4{i7rO3y
z88f0d=IH{I)7Nha)|Z^@NeQ6Xk9+{rpzA+zk8o^Hao;tj877oDq5S{~zv<6vzfj{L
zdDG&W-^0lz0>Jey;WZbq?j3~eEWp06SeX`*vKD8NYe%p}2!vZg)(JE$^`<ajZg<*2
zx-T`kBD~-WJxjG#>rPiV9;|5r-vFQPY0X67|K_u;cr55CB*=nI4o}4GVH?K0wTmNv
zwKSLA71-bd6(J(SZ%Hl!%)Ms2%(rG;i*JVQE-ZiRg(iBl)l%|+k+zleIZq$JFP7zW
zY3qan0gcY0g#eUOAA+KT?p-yMIIs{we1%+MWO%q0j&o=YTT9Ug%T5aaF2*j|7X_g$
zQe8znRv%CL;d)4dE8TDIEX+T1(gflt@5RLx23h`;OfON6_ncV!Y{MqmL;)?sGt+fS
zX190uw&Sv$@8{9}R>Y*1&s_^@hPe#Knn2JAkYp}Gl1vw!=J00N;wMTfu_s|VdLO<F
zV{5B)VAIQ-FWWtY^xX%kbhjCJ2s^JhhT40n8=+G{gOQ|CE2bcHre8tIr|Y$5=dnLu
zM7DeNbKY!Ke!=(NcNJgmz=wm-=8n28jUC^wkU1ek-Kmf}vFpjyAXh>8wVjWe9?Ozj
zAQ9BKnUSM|Yp)TFWt}_)>g5MrJ@Qb;a+4!H)Cio{n~oHh*4mnJfHBx^lRniHo^X~^
z&Y8s&HzAtKLebm6g{<#+PD(b2o|eg3a&O3ez%hUX;9(<i{e+-7%N<)^t+@Qjz|%v^
zCxa&4V;P)Qyl8gy;a9|zI;Mzyj87f-tL*i=Q!odnYZ3YJ3Olwl2g$JBKFIP1m`Atm
z0~p{cspfRR*$F4e!^w<$IHYuOC|~2_u6RE6t24lUE%IctKgDB=v+iq67cFQg{57-F
zAA?cjbPkpzsO7p1aE3~imf!XoIiNC#cb-8dKft|v38Iwu#Ha4<<<~U?2yOq)S%<30
zTa-S$+2EaB(u-~`!^xLi`A_U#V=j8XBira(Jh*YBb-g57uj}LQ;Wq!~kmHBI;l=bx
zB@i@U7=;FFB>9Rw^q^O(jb6-!L`HMZ{!@?(=7j&itNacxE{&+W6+s#PdA873_h_3>
zyZ3`bv(P~L$>^*>-xvC+dWDM6vJD@7?@?Wn9clLG?UPDG%l4QL>&HU%buty9TAsV@
zI?A}6a<ocRN;B+NbEu9t_Hk}K7u+@p`3j)QZkW*F0{yk7qL_m(k@lvR;_JC9l<OJ>
z@%~X$KzhTVyUq!8km9t;Y?I?}4zT0EreiGe`jvOTfKjFTg}da}M1Fb@<xF~vt3sv>
zVxH^vc4WrJkIzY;M?@300ZAViIZqV?PF`%!)9QgIxY)|)jKM0W=}|nDJ#?hhp<*@@
z7<b9>4c(e;2Vwr!)u|G5g*Hz&(4YST-US|h#^-M*%iNGH3m$iH=`nsgW=SFeEw;IJ
z%?tPgErhAgh+2P3CYtee?|s#HtvNr)8V8#N?s`MTX3}lOX8&av*LfH4Cg@wM_R|Gd
zH$eFurn|F*0Kjz_Zvli~{PuS;Dh-2NAF8=QYw-ggxpYZDO=R}fDD%EhQ-su^O%FoS
zNPHuC^hTR87#lnQ3kmw@tdlAB9{DXrSDbA|2G{_j)pf+KL?n20$QiwC7OcH|$CO|P
z9@0oETBSX$X5Qc)SA(Gb;gOB6c~wJg*Hk~<rAH=*_Gcl)+Er`2P-y^jWH0pE0Q%Va
z3QOxhJAP0y;)NSsqjUy8k49f|PAzuBr6BXiB}E_}d{}XonAWPRZru?1H|NUSv>W=^
zt7`M)jaEb|bX;EeL&)4D&snIuTyOTRl0eMeDz2Z6zIf{8>&k$H%%Go2jE>_S_sNuO
zvWv6DOsn_kl$M*Rd-gsP78=f5&w38}#7lJON8Um=!}H8e0V>4LHh;B;)(ZfmQk%T7
z8(;y~!#DcUIf)8w7DJ1YCSRsT!F}?3RsH1xG(Yl6ey2Gu>f|8WKrg%L_IglO8NnNH
z$;1f^etfr|tksP-@kr<<@~VO-s5t|wCbT<M9M^gKJs0*EEFPqpOhs7_&xF_$kWeJ{
zSfxxvb7=fN2T8gSSMT23otgrUipiaTl3|xpa9r}OU<qqv_yJA=enReMDC4S5xanBL
zhvktk@H2zk1`T|`jq~w*%F2qZcxH)R4^g&>&vKT}W(zIl`XbtIJh`zksJX1T3!UEr
zDiPh9$G0Kutn6>PWI=))|4+IEGCF!kgy_j3L%0tcfDH<o6CFvwvEml6N!!}mdPSMw
z0O0009(rkby;b+yjV)k2sWAzB1o>0P3|065GF!0p*Jy@ceWpbaT2MKw*H=lmP|Ui(
z>_!hAXJG)}Q4iDa9I<u|eaO>K%bn$T)Drm!IwZR}1=Sm@SmTKuOV9sxNGP0vC~?X+
z6H+M13`IN|p*uKc1atzWyimQ9FLUh6Vhx*_1FbLo%MIbol39rgy{h7lmFLD)LU#Zw
zwaT587_&BLG?YECJCayiS>Ye7&(}9h#2+xDSGjY6(Jg_iX~&%Ns?tYs!>UJ6BQe`R
zBoNE9JkoPeIO?g>38Sg&c<5=xj(iP*OF2e>u|;1ZZA?CXJWsg=F0He8qe&x;-~Ej^
z9<oxgrur?=W*I+d*&fvc7qQ?v;bZb3G+x?~eW%6jdw^freSJZ=GTqdK^w6$eJe0K1
z77?{_q~UmAL*s8vK%yxQ`slJcw*d8ym35%@ign!(Yt*l^#w>nlxZ=eGU@%Z!rtM8n
zpIZO%Uj;u}rP^iwA-KEVwP&YC=H>~P#_yh+27U~;9qsXu!@0!bMpazF+{6;`vat;%
z3V>Gq%fDm`GJEiQB=FlAt^g_uE=lAZF5Jg4xIzUiFL0Q9^%Eu(WKHQIG@o!*;jaog
zq8P)%Uwlm~%c*P50Hvn$xhPfjH`eITsN2~3EzaE9vO)8i-2)Z&CHbcJrV(e&_(r6M
z>kV2>8_GcrlVd7xeQ@;8_cG?*K&TRG_$$|^Q1FLol%OCI55Pyf$l>JbPVaM7cu1=(
z-L?}wYTw8hJj+M4OhB@`a_XN#FKue?Z4Pzfzl@~ynvW%77pN=1s7Ga`xuXL~S*5o;
zrpcD1qo^9Wm1t=8%i{A;hs(f__mPBV@L3h;1WqTkE=!J@rvZpU?W^Oh_V=05RV{`K
zT9I3luZ7&>i`JvcO;K)?kScnyPdN8kv6J%GkZpAHCjj@Z+2wg-i7IQEfRu*c=6RjD
z0{sy7^n0G~VSL=~HG1bFS)!BMi$omvXx}x%^qh<)bO@lT@x@IaH&Osp7=D!C4Q+-o
zmH3P4mR=sh&VIM-RRen=10L=z>zSB<2d(d|{0s~F1Z6y-OG&pMtQ7!hYcv7n6^W&f
z=B05!rzDtSF1*z+c2~)aY$3K87YX0ac>Kq~?zofFCI&2uTtcQ610HkH-gZh(U1XNP
zVX9tKXn3=Yxfsv`x&em+sdKGBVjg}7);A3T05!_PqY-j%1PqJ9i<M!$@4BXtls%X$
zrj;I)3=dOKY~Z#0ybaX#y}zdpAex^IxVp9NW<*N$14Bi*gy|T&qeEaK&xGlm0JeT{
zd?>(#ZL+H+e+KjzYuLtu6PM2#1uJ&7B193?Vt$%Y=Mkwf!%cktxEE_TUfSpwdc<%4
z^hOna`a5~tn{;aRwJREIOe$|H*)+cbvEGxM+^2N_-v|gt{hoCU2Glh`jpLFb1(x;7
zDZ|m?gA*&kqG~gDWowSZp~RS@cfoc<C&ph>h?FfCJBkHT%%`E^g5A;$*e)ct4}Nl7
zUbFQw_vj3jB{IE+DNW(Bxvd@QN%MI(cazTUDQ|lg6A%LN4NTr^`J@49FpT)+Z|Nj2
z>z14)K$^u)v9-8+KT^B`fQr0F6BHcX_sW0kYv!{#e>N&aJ^f~DeDKOPwMJ<7+Q*;n
zErHCOqYX{AOp?;eqo(l2?ic15almZ?v5mXKa6=ZP<UaUB6vXJ0XWRZDc0*LcKWEKB
z00?}%F4h2_lRuNr&o~fQK3^p?Bmr^d>yeG2tWGZ#u>KK<$dXV3MMc)?+QJm6W}U^2
z1sxMj1>oTiP{dq?NY<wKr{`WYml3kY^UeBJbTu<EPH_jJ0n>DcM)tJVvyqbBa!Oyt
zddb){J2#AYCb5;bhi@Q3>dLzx^bO^3(D&U|9ThFp<TSkxu;6^tkCMQoCQQ5x;%tJi
zfT7n<3kkoeOaK6bS-phm_X16uUP}qqw<J};zRo^~7lphVC+j@*9UKaQ-%viZ;?{V2
zH@d9HX(=JC?0~F(`gFGn_&mz>mjF9sJrm&OEk81evIl^Iqmx?E1ZZ%*hjoI`Bd6+z
zyKcF8xLwha2DTot2YF71nc7ChF9U8f`j%1IPP-Yw+I}_Wh|zjUVf~&@4`edoXKE<q
zru!USVX)0i-09EZQ|_`>-k5T-J0PY_npqY`#vj7<`JX!CHWp(yaYF%|g(kcWM}gai
zW&(9=`SRJiMSfbfzQ1<!4^ugTCwOkDF1ps;Nm*dbuYc{ALgp09oB}ed(KnSfJb8%f
zi4)x6Aa%RBrl07ny_~es5gPK)<t&0Eo=Qp}!m;T(BJGEJn#)iYU_3$x=>Pb?wzw7B
zn1|nFkkYW2WKjJ&sF<PCr}lN7h-0)OwNq@b_ost!9+q8M2RBL>C)T=j<%??7r2tPa
z$A&VS+g=Q&bt<zJ%G8CZG?^VWlHL7mKt+X|z)FjHRar*!gKJpgeSXxQkuw#`$k)YW
zy-O*s=z8{Pu2miJeEA9Es*udrVrDARej;DEcF<^XQnXvXP}4eD3|mSDD7hdkg)a3u
z<uk{<hBvyy-o=9>59uM^;QH057dPY^jSI&DOw+N}UC~qXL+E_(66&K9jo3HGvO%2@
zxqH_fH7w?jNY0SS71`F)R{@)%Y&p3b8Qe79*iFPn@Ise(gBVQCGs*FC$U<fLQ%%Wo
zHc(1%Jq(JIovD$*N+rf!jumJDkt58UGg9mAPkT;Z@tQr^L0*Fk-oOg;aJQ<7k(5}-
zQ^{c}hxoGEZ`_PiD)0^&817<MSN1U9Y!DL{Zkk@qYQ}*sodQFD4BU3Bz93+|qJU@i
zQJZ`r4YoXuP8S}drO)UXM%=OjW`c;yoo8iDfVra!8|c4IA2b0mli1z_EZW0ENY%aj
zY^+NCSkkabnjJVwfy@bJb!NjrdvAS>ju2|Gp^=ckXEZmV?u=f={qVXIs+4WG!y9mN
z&)=*M<;OQ}hCBUwc}eOgkW3jCG2XM$SiN*j>SS9R+8E;lJP$2Tuiu#L@=Hu7R;^lR
zTH9_U5oLnJ-8mn9{~9roGS$j!#!YM}qIu(dE>1Mo)beEonBL+!Dmx)d71Cc3wmtI4
z=qth=DzR8CFnKX;kv_*FGj}JmWaiBH)gKou^EPbaFQbfD2UN#;y9m@GQ*Q2@J@vaP
zF>uvLI^Hyu=LC~D_!SLiEXL3O>^X$xZhSFQ29n*4#J$su{swlds9B5$sC<9VYyt8=
z6O6bSYc1~nu$m;7NK)0?K8WDD)Sw0CTN=5GKYu$jDjH)Cn7U%ob5E7`+_;-L0PgHx
zG~en5hW2O`3z{a!CK{_5V~nQYwvg_&s5}V)Ptm}k<p&^wSP#z4yC$J#nc|^deYbvH
z)gfhMNe+lX%p{-1b?%AuZY}hU@j>GokmR$|w|?KR!qooudpLG+Sq>Q|&l-#w#7@a^
zO=AcMWyOC?KAJ(1vyK2~Yrxig3YnV153+U|VQ0<<;L9b;0Ob9*oe75kmuoGFQQ&-F
zQyJ5cGU0o}HlFL%F@y;~@~dyt+Wa?UDQTeOZ10o;4OX@}&G=T#XEt8_sT_nhJ6X}M
zUCvH=c4h}B`;&o<etsCW)^%*6i1TDjlFs~kj00sG9jO+jSDcFH8t@{{t*zCCy<;7H
zl~N`dl;y$<Y+gQ>(RzSCmQ?o%8B`aaf`^SpCMH1i2>V|m9(CZke6}5(di(&$20XXI
z*ciJAGB?0Ni=DOck)0Y6L129&+mVwH4>k9XsJ(<tUMfu70QncVyN-4d1jkRZNxC5#
z+&_}HXE4Wl96p>A=~xoJvGAtV0%0Bx908j?G{V1Bq`}51RQIaRByLFh1#K^`bmQD7
zt2omd#wk!e*2gMumWhzDDkF=P#n42KYOj~#O8?5Gm!*Zps5V07ECi9`w5EAs`M%FG
zYAxW8Sh|ng{D9u+Rr~>29sF$vV^-~t22G)dwOF4XlSeoZU3|zEaHb&#*nsqDmF=0}
zjo}ZYE8HW24td^wv=B@I#(dPMk5W^7e~ES(UhOi*JgxmBi8tq~$r};R#+%``#|x;q
zvZ3{u8j^KaMcL&HL5%LvpqF5gam?mEBy@>=qgKvz3gX$M@_^3S3UhzNS_L|P?HTv3
zxOVKzRxXW6whYZN(r;w8k2gJo&yLc$B9nmtNLY297Y~Y9_bDVBgZ&5)O~+fxC*VB{
zLt9Ni{VWGsCA$Ms3^4u=J0aZ9h!9f(TMp|zCf)U8V;zo896~)hcAbGkpdDD~<HF(x
zjO$7qp8y2ma~z-9)YgsfV3qVRaD2aZ6uP~HOaM+E+PxE_qgOF+EnD|TnohH5O}%b-
z6dquo;b;%RKeK_1<VylBJwEuN{-{G%<4;4jyVAIO4c%kR@cpNfVE<#uL6Y?K8q`K+
zv|7~i*eKYu9BMnx2jj=uORb^?M^=}?#^cCBT*Zi~T|xnXw!~%v@yh|VvHaybd1xU@
zsYjboUcf;JC93EdDXQE$B9b(g2rC7Pb4PiWk`txPCAZpH1no!ftN~AU1NX1?;sLBz
z`q)$bu3!wU#pV)zKDHBpNr5%Eo{YevPi%M8oH0o_MXprfLr3iaMz1HM?{jx*bdqEe
zx&KDvkz*ImqP4TlSI)%5WYzz9Z?b5~B^u!hAsiEYf3{=~A7_n8_{O4V((P_aEFn%8
z0(({<D}`_6hy(tR)n1#=67Fc=h`Kj`+i=OAJMLnl^=aBma@TUqQ_LO%4l#Y)(ARCZ
z#Q$kv!+*oqWm+9H<ErkhlRiyOl4hm2*p;KY1Ew0_g+6*@4I}G0s`Sq8L0-+y-{h@Z
zeRkkmHnc%)T4Yjtx%8msra?Rs;%Oo0M6kgwrK;-V0bsJ9O$#TP0hw|90lYse^kkZa
zV{q@S#n<~dSb=xsooT-A)5Io)s#6qlbtqZ5PWt<ptRt)>dOAP4q3ZDoe9GgJS{N|8
zXZBUUx4wkFJMc&5f%Wv#o)u0Y;Q;P{*Z@l&#kcH}8mQ*1zgm4CZnE;(8PcbOr#wZd
z<1k^v(i^@2o|UaylI-(C+Mjd|x{FT2E|b4;2mZSqZfQ<%)f6#Z^R|c~GT{lC8G~Uw
zB!H5^t9B?bh>VuT35{?uOIug?BcC&ro&lZ3vvb}!j3@5YVM?yMIKLZ72((GhCXna#
z;3kf{+n@u%04veRlf=Encr(X?o%Dl*P2hMM7Va>W3e06cc8&W@Gz9NUQr9s*ezNFW
zbujfDj%=y-Oq!;>5G)%(clZ7OqN|la$@G^2)&jsKAQ>lR6=dHBbcukv@K|F_oYuIP
z>NpAv(nprC-<qM3_h2DA(7gC*_7fuZn0fqtef-$x-JKqlkw2Qu5<!uSdr&$~J`x#k
z#tEyvSF4v>8RQm@R=KgT)TjE>ZO~tJ@-xqAOKB#l0ItwaE78+7efFII8|9o(@e{?B
z`gp2>QK>MOqq&Uzr}|CTC0%EC&xX%bmdC4<V;L*Ygei*<eS9vWEE_T8BEoR#o|LMQ
zChJK-lqeV(blbSl(JL}+eR~pW!X4L8qEN33iAibRSO2&RP3HBy_%DM6DiubWlG+*@
z(9hK4gojR!eDbTxjpHGb{uTV9#0U=}vpqXDPvV%Ku#Wk??h_Hc(2=u(T3Wb69nI2t
z_LZlyNZs!N#aojRF`b=dh6)3!n$UdiVOCZC?3y>vyw0Ce^B572d+D#M)^*ngn#%9x
z6gm!mIJ4eRnQ`FSJ3AANFkpLdBg@EUF1xSzk){%BR8)+yE15g@B!kj?(<5Oe;!oVQ
z`uLi!HSYb=qvv3im(P%``QRKee{7cOZip^u*faDc^ELEMMdEzCC&y8ClHZTKhszKZ
zdp^?Bk{pjW`BvdarX%&=-TYCyLQPM(bwp2u7Dg5Y91wQC&&Ws1NC{J=LMN5i0|(r*
zFCb4#R>D@<(th=##ywFwI2{tCU(Nmp1*W_3GyW`1VIiuAz<ctUVwPzZh45@b)A!Am
zePE>@7c#82lV;H>hTFT9|D`+L^Weocqvu2r*h|^!a6W4)2%m_aF9H{ATSc|eVoPEf
z_M@HlaPn&DglU5hU6ZKR`<=uq)QAqNlfxaa?LGznoeBeKK7WMmY=XH3(6FDk)Bfx-
z-&XRiq?2}W{?V4&0oL*W+#oWlkm?(OBZI))^CE-LMPSP0lIGeF6Rq-&t5zZ+K9*ab
zzE7T$-dEMM^^u$D=F<6H0ZY10618`Hb1(rH=@ISvT}7{m%;*lgz*2Ga7GUFi8&5@2
zVci;U3yi9YMm{mXu5}^6mII1&(Mf^TS@}h^o$nh&$OU*La?Bg6wwx&uClk{N#j#lo
zo^{Bw!i~8{?bndWFOT*%n9GFHABzCo&Me7?RoSa|Tk;#1#F4>EmwIQ(X1my^h7nb#
z7h!U^WQ(c%=o_3#J4&p4(V}RpV$9vRE>$I9=7hL>q*6erY6Ty2kg|4eK>BKqzCsro
zRn3KP7u^vmzZ*Sr;fKe__tx+Wj}_wo)&eku*Ca6*-nn~4Quxc9m2PrA)hx=yYw36b
z(97mz6e&QxE_IPfLSz|9wQOuHN{pO%5;{UOj^4i3gL#KU1ssh}U%a=8=sNLZZm#!A
zU92Aj)}F6~x;qv;xlDttqsa_feJewl0%B5&!J_L|bZoACLp96n*Jnjfx^>epybZiy
z5f%6ft<tZLx_KS%=H=wTSJuSHyBTk!t0S1Drv^w#;cuAC@@69Z-W_9?s=8Ge`DR76
z7v^Jp)h*LL-+4vdHOJLLiQv@Vg;R=rwlu&mEG-?e@(duh8O~k(#Vn0?tBRKb{_+Ac
z=0%+v{;qXQ)#o4W-=-T#3la$Zgz{Nk1y}^ij?MXq_M_Z_kK3cc3t9SLJxQD_MDY1>
z#i}u)>O}m>n#4@Ld~<wpO!@PsmqFi=_z`R5*Ia!TL?x2Uy_Pwi^l4bXfd8<x<AwK@
zQO1(O@oC7;NnOQMc&@Po6~k$~KE2tKaLW|NyMkzN>DPolr-m9aM<+~lZU3U6c*?gj
z)Ps=7*huEU<JURYi=PQj^YBfbaC@b!!kuD@Klcv5h`KR4rYm6-jQ)nSe(}wDQTu#G
z{|8!@;=H=R@4K|L$FC(BXfkHLw|&`*cW*rEm$;E|qPhI5D{NEuX+Woh)Q_>bo2!$D
z)Fjfr=YsRI)2p2o9{66BudyBf(4G&MQ4!xeb&~XG^b5B>uiINs8XLLJbQJbn&KX%N
zI$DD@MB0CM_HW6*eUz<*<JNpSeV*wk=c;^?e#XosQ3*OpIJ|&O!k39$?9Llvj#?PG
zr{mkiIFi`}&G(b0ydHAR)0O*I>S&*eAO_oi+4u#a)(M?*#IP;nfGMS;$l)ybBB`CN
za`!<MP)OG4Hf-ZUd227IOKiTPaOX93^(jeoW#~!<s$(?2_jQ(wTAAhqB7EOeZ+;U%
zX|v|c%)bdYc?lV0W6i?p1)oW&#bUeT%jhJrU>yuc_=OeHUFIk<mNAzJFI5Y7Okver
zNgET+N2%L|D*s_PeQZ=uQ8cHf0k;}z1{8GL<p+V3ADfOB`Kp2PUQ(XfJ&}QfNu$ww
zIV*l{3qA4m@I1xX8h;$t_#s>LNFXgra;kaeT37g8j#=LX^FnVzo+QgHlPItgbrv+T
zk)KU(eh@fOky%dMD0B~Dmw<MHfkhxP4<0<_TbniRsfm7(nR?pQ+6aq&&uQQaqnT<e
zd?$OOjXERDSJlzdyMz7W-UaRMV{iwZ*D6Nl0KK`XwbO6Tfz-U7sc2gA>}Jw-4Q*1l
zulvm?S2xv{OD!C|3meuwjyy@LoV}XERYuBU6g`86KMT~~TT>%^frT_3k}eEb##<xb
zF5T_qmsZy;W$<09_-<3!N0HE=6z#ENQtwrA*DMK17jNwPjv~0xsweRZ;_MF^bq?gK
z)oa`E=eteQ?s3uBdV3WAxT<9OhG(`6vPoPrJ{8ostzu!wJb0@2%kvy0-@*MaGU;5K
z$b3;+W+r51=FD=}5z%@9rBtIX^9{DXS7OM42d;u1-uI-D`#q8I7eb(D@Kz>{;yw2E
zEXt+b)mmtMK@zx!H=7ZBOdvuni**%ceOyuUB7b^`>l<uKoDG5P?!P*F@Z+3}X>e32
z=K)-T?&D;ekqLb=ftxSW9aDPrfp3r{WuP~{(y91c>;&uvHV4MSLRn^NH2eoC5xyC(
zTM@&!tar`BhIaE^nsZ&P_<Z8^m#3WQG|u)*VWZ@60w<;WeMLsw1if_UuJTzLMd=<^
z<`*)%pnMurajUO9T(?=c1tJ?a;@&P;k=$(YWl$AU`c~EoJ_qz4={b4t;me$t2slRd
z>^<&sF^$nvuiTR}tq^OUZ++k$7aqR5et#2@!ZB4s>>FVEQqrv9#kY!0Gbz4HtH+gn
zks96bTsM_VIuhylT!UiQykqTRc8HHMo-ph>nnX&P^$Oo)Pze1B$He8$b8_h0kI^0M
z7+x=K4QkKV;c>O6rcSHI3a?J?^6Z|L;|1GUrP&OG=;*TFnm&GMn68D(e?L7#wV<1C
z!oN#)zwT;Fa8riY(It+<QxmF!#1(_gN}4OXn9vVXoFnLwBLNNf0Og}|tEk{P6Ql4`
zj$`}mlCMEg;C<1ZllL69QKSvBAVqdZzPK(tr+N=Kn4gQ3KCRAITwxX53f>0<%RJ~i
za3<7^6J70Z09!$P>?%N-bz6>UC{NVs4~xBLJAg-bQk?gYy=6u)<<xZY#bI3Nwn^<^
zr9+I=zSb_NUDvtX2&oY%4>+g4@UTmODYb{MR%{-<?VIDU^vCrGxAu8Bpu0Bv`tN4V
zV2bK`E0ylcd(?IBUy*2AA!bjW89TOcHLnRv>vc)&{^4RxJwrlDN|yVRo-sU$0xR|@
z@VTm5H>dPP`g<3EAX~XQFUAYuxAjPFJu-fE>HI_Q@d`Wruo>EvNWm9RzEPjD;1eU+
zuJe17X~1ViF1nh1iNNnInfR3kcxOabt5^Sr_`9tKYn|N1_g-DGQ+k6<?v@h3tlyh9
z8ed|SNcCPyIPWV>@uct|ifzN+G`QJ*O`Pw_7Lc0Qj{sGWCYQvow3qc4)y)Uqck$NU
zmX$#tDJ>Z1s4<R6d~=OV6Y&CW?5y|Ho#7wvJjDa*u=TO^2jq%N4Pu`&?e)*Ddy9sD
zzwTFG(9q*2P(gFA%M9q}C{QsY>hFKJUq|0{oO`00K!rHyrsUx)bG*L}G;s3OYn9_+
z;mO@UDqZC`ZpnM`ewt&IC^o9<b(ufUw>NX$oSzbNTYNSFw=aEmpg!U*yS?<#14o+?
zi%UpK;v`Uw*{!)BG$JGa^gyje)obaRnCn9OtI>}tCq+Fkc7uDnB+d~d6E^UNWC$L}
z5cEHip+F2gGO~OJUHGjFie7QpRk1`NhcSZno5FU(x6O$}i|f)h-%L7A*3wDey5_&#
z_gQ21(rjPFj|mYFJuVz+LID8$CaGG6pDq#_Jy&w=aJEnS&su2xdlC^I0zVJc0{Pwv
ztc@Aoq%Ka>+nR<4qodEqUHB2Z_s(K-KxrnpE2t;6RzNIW(s}k(lw$XA`8@kf%TzNh
zj^%7>TQ+SVQ}K}xCcar9r9Rg4P-CWhyzA<ny_BXuU`P0=Vh;weEWAc`n7w+BOXqTo
zFX?<)B{}J$Z7wHoM%qmpFnJ0(pW38#!Y$isJ0{j^RxYrE<S!@$K0El&_yR{~+3UC9
zqw^E*Qy`)ZingwtqquyILQx9bB{eDSMH}_&3EyD#w5FT^8m|_d9=9=iCf`^EZ6&Q4
z{D5u0RNCx?S%s~x3R}a2N0oNphxE<JluGrUe&p*jb^b~y0?<&+d_5ky+KkIGB`CGT
zX9VoS6~M!uqG_OnF>LV1^z6sS8vx~>E3BFd$(0`OnLY&jK`|l(Yo-O+_B9|e_rvx}
z?_%W6b<kHt%xwTNA`o=ZKiaR@^RZ|{-QGI((oG3%nKWk^Vo|XY<A;gjU;!D~jiHzZ
zgAuSl4$rCaAiHncquZNTI9O`)2j%>*=vx)@p7(vnqZ?t(HhT=WkM{eO`2<%qi%mPM
z3Z~&lJXbYi(5M~3tZ1+_YyNXa#8yud`$tRQuXhb3cK$7_Yq-W`>mTzL@;55;>r>o*
zDOemW=cHa^;nGgd{)pUV2HX2uy=IMN58gG!YX7PX8MZp9$5tjWjHVvfp*C(OmG<3j
zyJn8f9__Jv(J#poe(>(fvPbF%tg&o6C)e}5;vg?s#$xeaBdX_MH=Of4BWbTeShA@9
zp<8!~xUf|bwyfkqy<H%VRTwP-DBhpdF6QRffUb)M0``)*3(a}nK3EdcD@?U^ZLF%Z
z;vm~Jp|_)Y+{;og4VGjd*!)a8_t<(W3FSn!vD&YzZxRf4NOeDs{Gp7R@HAyz@Y<8>
zd_`4X|A1>(D~=L<XOcxEqMGxy`vleyEfz7~HDU8|&E(6IAeZDxu@`pNr<&-6?0=zM
zm4A&sSRZ;<lwUx-mlQux_O?dKX1cMVWJ#!4m+1`by}UXiJYvyjN3cta!qMF{?OZT?
zzZ2rA{;XVkindn_bK9J;)16|39aIFB4lp-y(5T?(Uc*+wmS3fO8a|Q|ZRr#2_e<5F
zhme_#dB<>!75yY1O4Ayk?8SS-x^4K8$I{OIcgg+L@imjk%$Qtvz3t@hLR;Q1><Dl5
zj#~`Yz^u)gW~ZSQzw@FATID;Ula8lzH9#pJ`Vo_WZlN`^w-;%)CWIF^!wymc2hN|L
zxR3!PmPJV)f|tdqLB}{f?+m+y*`u+5x6)Z(MSA+hEsYKPd7{?56Q`*)FQPpKm1>ls
zf?~JO@D@~1co0y{8qCW=7p*k<cQ?`a(DDBLpv~YeAlfVe!#VRryaJn{^zaYgOf)Y6
zS@C>VLcPSpcbut@pvYqKS7fnz*Vh+*@&LPbdgW(<wn0ON<>RJruno568`ck8M^EwZ
z)HxM=z7|QtB3aeI{MoQ<l_Td@tH+hi!<ObjSpCi$6V}CayqS~$ykfE{<oDug`=z-(
ziO%U1c6>qd+xZG#$)9X$%$(d^U&DxziDdg4l&}4|FtslF>w+`^tHjZ#EOU7fEYc>f
zXvCmh$9HV2FR-S7Mk$j;_chc&yr;kmSx=@|U99|RL4t%3#T(=gLz(?hJ79T@fT@SO
zQl009P-ZDmOSirKLgf!(&Ta9fDPvFKTBiG<Eh8C@9?u5A7DX3O6+Z(#GsRFS+#oTE
zx@C?fLgeOBmB?@^pX?b}pAt#I2R7wc)xS?oO?4XgtE0^?B;rqJ&v?)lTB_*%k*`R*
zsPYyl$uG|yfDO_FEH(<tT~ryPW)`%3n^{;8OlCm=8$ris(Q^8B&o{>`;Z`lLL@J)V
z+X-i>K!`B_V;Q@0!W6$QGoS4%JYZFho4xfDMSMM`{-D%<q{w96(V_<vGD5*pYCyTT
z514DO5MD;Nbma%>b32dj!n;3?PV=>gj(F791dXt_oylF2q|){mo&wE#gyvS}HWQDZ
z#)~;YlM36Rw03)BH1?w0+qZFP{r?YLR{<4e+id{>36U0%5b2T*X;eZHK|&EErMqG1
z5RgU$qz42^3F#WTb3~*&2L+@X<UTKc|9}5`*S%}GW-S@xjq{w?XP<o>2<&N8UxmUM
zwS*a2)W3#QTQq~ajHr!tOcQc=du{ei>7=)KiM@EcF5AHP8O_@W8KnEL#h`J+7GiI(
zhPX%lo&-4QZgr+gQDh9Yw)34#u19WzV|To^(QEhuU%X#E(GzCsIc(hHxGOHVy?PO}
z1olb@7dw3#E6WtXTbFF5VFdVdJD9N~EDKRaXxcL-+<g?>-hWTxBXOCi@NhgyttM+-
zaaq-nou2EEUo9QAq=#UX#BO)ZF<AB}kmbw9qfI)Nx2#{Dv(oS(FO|RYma!0Qz+r~&
zVxvVg(#l9NM)~M-N#kiRY&N7ZT`MKQVj(F-+1|TLs3B>({Yib44*(@2pXKzon`J`q
z_dc9`p*B#dV!I!e^58AD;Vi<P?-f;Kf@);!J#jO~H1aL2<fV5=`nZ)Wbwk5XI|wtA
zg+rae0i2p4vh>nT%3XM8B||!CmWa`Or!ZKUTG^2PYW-Humd2uPb!t~}!=bjs!4LK9
z5!7Uu*KxyBgo{vmIh)^JlJsdkqwTA){MZ%m4(A!!7s`_$A5f;YBxiz7^C$WY9OS1p
z>=#!D+Kcq6HaN-{!kq<w>adf&NMS~uij70bDn?X4kz(N_fDa980U@1<P`~j9#?L#o
zqEPwvHF|;s{tt-?Oh%J&!mfB(m~iZ!g$d=sXR=>6`6$f4-z^LG{exg~EElRPO_SYV
zZ54RBcCHahsq>ZgcvE5Cm)cK=2klORw;{DEanoX%yW{)44W?bB*)?S6Tt1~sGDZRt
z`<Ec+m(!?c%)2L(qN2V(jX(R7zy$(ygTkN;r-~RHV(WFj5lzfi$XCJ6<b_C~h%b#v
zHZwtZk2ZzP-xwDe*Fs4vpv{(<CpR$^vAh0KU;c7)<cnrth+<D5&maEX99_rQ+g;Ij
z>|`m!Sigyl%yHQ@`=JMK(5NhDA3v-O;5pVN&BEgUo~r&i6UOs$ufwM?QLoxoSH#&q
z?ENv=9Y)DhC0;pKvTe=E_th~jfGx)(GGv{s0gQz@f>8qn-cR%-_sAw67%kg{p_6SC
z{Y<th=KCz$%cA!|ova+Td-p6k2Cs<6wcMAz0XFe%@`<8Rh0SVAL0pKe9B!!h?F;_;
z^jd$TpfSFn;We+tohT8J&y4)8&c2Uf!4A=nkA#|QlZNW-Y8HY+xDDg%!_-+*?E`Y3
zcKtkc3fHc6Tv22^rA6HkI0^mwm_2apepi#OV)l0G&l4&5wvyI`{^e=HUE11#bmhIu
zDQs#GHfcD0n2h-swL?{nv&MLJrrM{RJM6SZ&)Ynr=Y;^Y{c3-0i1D{6lXdv;X`hYl
zw&2e3eYP0SO8qCIdX?wP9VbbSsV*0tE|-A58xpG`HKD$~S!`E5V8CB{%Q$PJQ*taF
z!3rctOp2Xuk3A&YtaX2{+Y8Qz*hJ`t&u>k{rM-?lBZGjP9Roc$IT0?!p{(M2uWaW#
zaO(8n_at@wyt294*12-cEQ(1fiBnaB8**UCm5AF&s|wJUaVU2RqeGtc0tRVh-Zy1-
z>$&`PQuyX5<@uoh-1YX@w1FydZ;EC!ZK+tzVx0-A>Yp`Si~u+56H^)?2gc)%iObgN
z6oBSZcZH&bT!lIqnA_gzq_)cd%dX@DSMkXIyeAP5WOK@Xg^Rqq@k9=$?2`*4hDZO#
zcBUcJC`Tlgy9@FDyi~;0%XL2G%{bVU>5z6UnmfVgt(M--U+K6b-72src5y4?O?(*%
zDY3Om<1O<uwCA<*T0X8rUsCjNoRcws<Y~H2&oaz^(ex9&@Uxj?Zqg43=_{0bzn6Iy
zdh)y*?!&&_c74#MeCcT5b+hpq1y(F!KFuoILM@GJT{0JR%3DB2?8Iv;_7qc=9>9TS
zkvLE~-0cI!xC`$@n|qIgw>S1b2w;7GcgJVE9L=h%Jlv8@<#s5m(5@Uz9_bwZ#)MTT
z@nDC9c|zhNPm+2R*jMVN_%7hJtUxZs7xtxRa0#ji`d>m5a$lUnSynV*RaRLEY`LyB
zUp-Y%!R{xF^&c4zu63E7#aLtt82(@jTYl;AQLT>R6Q>xD_-w@Eju7vIUs^ut$4j!}
zqdwVjV3+FLv3F98<3RgkzlgwBV0(DUqmD6$2h1Gx-TGYPB{WE<REDBX>N#bnYgq`O
zX5%E!wB^iFZC=@v7fzAHX6`#Vg0$01s=?k^CYFw|D%z|HH=wjrk21z>Qy!Ah65;Wt
zbGA0|XpD9nuxE_nEUsyNal?2jO_(V(>Xzk7+@^O|6{sD0LdIB=WLbHv*riR9PTn2R
zwVb+Jl*h-WHUBb^B;B7>>h!W6V2p8zpPB4l&k=Mz`U%()NXl@RUq5yN-x(z!si(W!
zlfi<500^oUj#ethO6!!wlCEmmOzN=y25NY7*?_9>#dSBEpSTq(^!1+Z`{lRC@|T-H
z`|Z2Hg*66|_vJ!4wg{&f=(Q{E#ct~av@Z9Cv_@?7@6hs`&|29pF~+WQjSl?ON=_+u
zc!X6lrhz~lOCY-k^wRsPZKi^uT0LaV`q^}I=!Ho_>v5VQVvBF1=1e!XSifm@+<FB7
z2b^-PX7$iQ;1|7a)NQ8{4Io9zO{i9coT$GN>2m-eANloCK)*8pUnMv#h@Rs23k=?G
z8$W<cD^Fb<(^@CSBg>g2-hG`TUl2V>Ta<&D=3jh7KZxexoKwCLRjlibxf0pukE<%4
zaz+fNVm?7OFgz{o8|KW$W<{TYUq7_daqXOyZ1&9Z*ZiY*D599tFo$`w%v~#p<Y>$e
zBYhC<<+DBQvNMMV?Ix?Yml11^r~7m5S+`fynx_cG#^Q(SqiYK{_CQLwN!+|%U0<!D
zfC(LEN|7{X=m@O5_sf(>Xj9gVm7~XkZV#=Xj10OVLDJ{WD8Xx*4&V$xqvem@Yk_@V
z>_ZtCTmoM`AB_KIb}Q%26@E^+8r}o85GUrsp};-_N|QvRrG`Wr@9nLo%P24-G5g+Z
zpxs=V%uu`1(&RN^ZbAiEvuqG@!#0MqD2bnhP5-y4pt{v!_!gPN<_G<SiKy6gbLCO7
zbFhb1o|?+LPsujmel$ME>UbpiAz@$jx@WcjBiy4X@!!pWX-qLRS8h7+aH`72NV(c*
zAa7N9sBg2^OyeQ{lFLKGYLEre3&`{Akv!`wdghvb7Oy=Zv9Y`%yBZ8GZZ^?M&-cPb
zWogR7HiT6adtvE-1{cZ8B-}`0_$<E>d!yI5Vi^cg^d(>#jZP?A8<QyxVx63v5ZjcK
zHU`Mj80WZ-kzW4R<lYg}bJxj1dFEOL6LX{k^yWvwi#{hO%A%y9RWLoJ>7gu9%0G1$
z6}iuOS3LAUPZ($W^~d4Fn{AS<Tc<HLCt&a94fON}Jm-aG5`+eX*RdNkS<8yczI;p<
zN}ll+3)!dw<L^42o0VX>8gZ2L9~cd2cX!_IQC9BX*R}_%lGniYU&^FOrHJF++Egns
zX$!C!bq&mD^|Koi!8e5?x63+3s0BO^t$8zxlfbS;%LuR>`wingfF(JzeN+2*2^cI3
z9O8S&vT*N>PBWG<E#*Cm%fW+5uyA=D^}#VNS0Rc9SDx;0PiOG)<i*)hG}|SVT!)z3
zUGPTC<sXy7muv(L$B64)H~AaukHJ_&#K~z0V-zpkN=57^lTj+i3JXADY9~r8<Dsx|
zAAqferFS6%oOl8<6Hy9ynooq|l%b^fJ5LX*dkdls2+4TR=7aN|h;@J2_G+P=2XO1`
z9pdHiaQ|r5YvpI-6cqP{CPUjLhNHKw-8O!q{HSKUg+v$Un-BzoVnVy;<m!GDB~j;7
z6K>aX3D9>r#S>l}i*NUd&8J*q+))X=LpxE_2^|`$eT2Dtl6z6gb)|K4eRXlI^Rber
zr;GL3wnf9J2e^-6Hx-ppSBOz9QwAGyrEF@%ocJ*ZxSb~*`TY>ILca;}DbdaD7xeGR
z7LwjPIy>BCWLP~7oi`aHcICklf1<0bmNh5F^}6Z?plBG%F6+n#WZ#&|l*3n93gS6Z
z_o@dpZN{i*R=V-fyAO=1%VoqL*Z^@X`HEI^s%R8=x>OD}U`1<iUa92>2X6tzo*PNQ
zl7IQSI1hi|oup3f#<ce?VCTf!n1_V-=h0NWX}sUl+Lbi|&X|#2ShdkJcXzOr?~_ly
zc%E!WNI0O8M9JLJh%#eQR;6L5=hfC->5XZ+E;>L!rhbBqt{^hEmrY|@nDA+QvLZCk
zR5g$;eTS?+fJwY`$bzo4UV8G$YrS6!WaNVJtxkWf-if8eQ@W>kZZR}|V%E)=RQBO&
zL)y?Suw9nezZ#H6I!(E+(>lMucSvs%bU?S@tP(0RTV*$AEBkWAqg81jMWhR$u+d<{
z^>^fR0x#=to<jIHKn?PxFddQldss#CAg;9KNf4ed0!5OjxDjfogD+;2yfA8AA4vsM
zrVKP2HQHJ1n8zkZNcAtTz+XIn?gs=vZpYLQw4davq<0^s0^j^i6hK9#Zd-cx*Ql%(
zu{q<`si;zpoU9hhSC@NYB^9__Z-*!8nTZbQKS1f%pM<CQ0Mf{H)>I;Lw)Pc_N>BNS
zqZ=fhTN>whNF7f5@*2R>8@D<Aj#^0s<+0=Ri6zK~WNMi}N~YslfBPONm9PB`2}wMn
z_G9yF%VzJ6X=-k&eaFN_vbQ9^${~T1E}6MiX)c_s*~)b7$W%_6)LQEiD4j}YWNat|
zeJ518gI&=`3)+FpD&oTE<!2aHDB020=xqsmq31_yWqa(pJn1<77gqM%Gx`|Ga&nok
znkW-Jl#StJ&v;UBX-02#_c`%IyG&J*8({tt67Y||tU@ivsXuoRU2|t8fLaS0v$=)x
zJD96$e^#{8k=XEiQ&eNLAELIO$+VW+8H+uDz>Ho2&`sb*Hk4F_%Q!+M(oKaG2I$}>
zF@WxS5`%tY-!u6@GjT)ap*aL)ek8G55l{D7{Ot=oRaXqz_iRqCJacE{@QdEbQPqBm
zNuYn^4+ax*_mN&Fi6l|Ff=U40;4evZ#zzhfEslSL@z*3&`G8rKwb=%7w*Hh*Rath^
zAB9>14EW;t@axVS<3k~~9dn{!vU)9hyn>EYHnS|;bx^S@T=;jks_Tuk;$h={JH21a
z2b92<7ssweGOM{0j1Rnnmc*A1J$j2xfN<4qcZiVhdK?DGF840#V{gQ6oMojC$Y2O@
z2eD!$ZKV|_(W_E#?uN1ME<c>Zig{mKidjO~n==%;POTW?dc@vj5_CdGY)r83-1MXp
zX1SR~e`LAmCH<jnN^n7hr`{?uMzZ!+j;ykUKoWJFvEw~+*4+<g??od~SQF+wMVAOK
z+oKJGqm4JL#-8q-T1`F;r8}*jTMcF!T^7XvlAeBzus!Ai`VM9ifDi@h7V0-Q8&{L9
zZsa3C^H8sD<}gvIr2G|oy(kKR0s_BHptNd4xE>Ku3|+OzpV)@q25qDl%FMH3J;hsE
zELj5T8@9ta1b|h#8s2o1`wnE8(LOIZA*`nv*sPY()w_m(E0iObgt)iEmfO4I4Vu&Z
zlFC5T4}nA7fIzdkZXK0v4hexQpFY3St^|`>PwKDp9`j4NZ^C~<_3WpOt+~c)=Y6Jb
zd$NJ3e`g%R!#jS;v3?&_RSPx)%}l?}wKNBj(qw$e{TCO&PdW`>WBcU8^A<bim}qHA
zd@td?hQ<dSD;cN}3p^bu>b&8?LR(K)bh}xCa}j=z5C<9O%QElq;J6AEb{8mTi7Ilu
z1?4O>%)mj5PL^7J#Q7WwZbi`lB_FU<hW8jyMvsx#`<>KZHb^Fy;)v==*icSD4un&L
zEOy|unmT@XoHhwo1R)p8`ndOWu*9|A#V9C5(?$Lo`QkaqO4VAfuQjDztM7Y3UMRw-
zp29xc*ZcDpd_q40ufe#7gzMGU=8!+8<WuTY3lqQPc4HV7GNGQ#naF7~;6h{qO%r+X
ze&=8Lxo;*bEP_89JyRyu9x#HPV^Qd!;r%Mou6(#TSsBZ5f@xz+6lw`tLBX)AT=&ge
z`Pb#6Q{4Kwe5Bau1KMt^?y=qt0Jz55goo^5-=ql%=MHNcvU&f<xgd28@!gwX9CtLr
z^%&?wF*6<8M>PW49K%=HzH{Xx1$hGru<-rbL3LWqBD1z>M(YQAwamSSnPa?Dy~r6)
zF=Z!(;{7yz+=&fGz1(SIQwW{Q@_QrQMa`VEwp?AvJI5H%olDjMfqF~nLQg=q#sL(5
zZ&f2Uzk|mV$yO1;ZJR(D`f;5BFG5FwiQ}6Iu>ERz6=$3MX(uEV_10&nX?JqCyY4iw
zc4F@>lQd}{dE*T3yyA)ShrB?uGwMoqae+Ev5&ZEg8|vcv-(pH>VnVBd*Xdx*<L%6A
z!B3NP%#tE^IB#{mo*iH5AdsQ+J$JBY#FqY9j+ANxLj|{Yyn}~d-CTF$*eTYut!eh8
zD|uAkj(RmX<5jg%cLex;4}<W?&*^`ZiYkJ{Bg(Zf`6>^-3GB$q{Hhb|O7ebNCo0G+
z_YVKt4Wjh4HtEXBkK3u*jsA{Jx%|&HDAzR7mSviw|2r$ur@6I-_3;_nDnUs9`B!0?
zg#6}4f~S85rCy8Sy%H3xBAH;4NU6-FpZSG7@0aC=ck(BA6y)^5B8sv*1L<Pv)EYWB
zJ?N~y8`9W6Y%2S+GPkk31$?rDwC_!cVccW)B569Y$vYt9_yhZC4wa}x=%AH@?GGXj
z1GsM1hbpKmz{eCKLYST;puUrm!#DM&RmfGERCO*CSzODa64;D~HlKldR&%s!w54P@
z2~Z_m=KBsJd=M7HEM9<H{;JWd_HOxmC&aQ`;5N17`xViK$MCS7+bVj7GbJ^Pc)T1R
zUt5kAQgJsPzhRsd%`#yo@gFEEgh6R_QiA@;fYn8+FhzC!>}D}mFJ9aP8&$5jXanoK
zTd`ur?1R(c{C;}(?7iDx!<UoTq;vOj|5kqm`$o3Q&1;`kq%WL*&XJH$Z4`e|!bj$(
zZe7TmM2%JZFesrE=a~5k_k0F$zCeN+&5G_8fFT=PWBkmgZ`Ttswk|y6iP?uM#(JiE
z>24D{wiM9cWjgy{^o8=lZVE-Vs4as7Xw-UIT7oiR2n^pE+5D)mP}k2i%^3@jKHq*@
zitV)&DWW))yjCCJig+7ks6vZI8RO)6xIyJmGb}*5KGx`c7BUdII8*JqHER<Qq;SW(
zLhusgaxJe26<BeRrINSqiKi&oD{p~u7K<O&ds`Cl^@P!)*PfeyLYDYz2g5{s9j?*a
zTD}0?GyP4)9EnX-eb85*+{HtkM$p!Jjw6y)+V$3EW;{IQ?Fc9yH0Rdlr|*I?&Fxg}
z9x?R3c`$=TO5P^l;n{J_i@N#=F%B`X2oiXy{X%QV-JTh?lWc=DF_(ve8s#M%B=AS$
zW?|c@=gK>b5_e_@*6!2gy{s>!llH+>vLzt1<~jR!UYu}l)1}fJ-r&!(+BGUo9b{Hc
zPN0mT{Riv*qJ%+{{GUvz%L)?Mf}`0_(YzX(eTk}kcS1ik&li0OEmKX`O?9yfeFtiN
z)7p=gqhGGDflzYQY?P%(19`(uv}d{COi@CIi~gU(1<J~ejcT2Im(14k*yaR0sr;zL
zU#RS5*ybK5;`2j|ZS{A?fH#$l>Q6EX`RuhY5#GPv{%y0v5{}Z$SrAn=k|Z~ZcNgnn
ziE!;Q_aeLNd$K}->`?!az1+qY5@!S)5$a@Gmhc9H>V%w-Q`q<4xjs-?!Y~v`D=xe!
zSkhXau(OgY<r5VTdT^X+V}3K|tGUQ?8fw}W!KAWcEyDV|8sOQ7Mdw&!KiydnO<FN6
z+aiexb5Gkn{nSmCR?|r-=4>8l{;Rt1E12_{X_{HRR$|2?{?gVxCB3Wu9agPo$_ciX
zrs#<1W4#CtAE?g(5J5dv$}ZK6b)e03w~e2q;`KgCsRZjnk5$iLT=%IV3B1~R;!Qsz
z{w*#x8)XDR_GpQ|opqG~UfjkQP;<bBD(p=i=d;afh4YzBIOWI6#DvC5ahBP-&L^(I
zV|vTuO}>{nC#mT`0l<R3x-*jctoS9U<9;SL1z7ua)}5T#-YuNVhq>DVQ5#%HMWilY
zYkz*x=^kUP)3tCYnGKOI){p`T9`~J#!k3Uc@q^msLbt9e9C1(wad)+r4^!k#H&}xZ
z)=5#O*a8(tgWM)&iZEp<Z2;Llz3iKk$x`q8Rcgh#y<6E_U}54LKDd`s?bl;AIVMTG
zkd(RkZb`A1Fxi6}^QIkz`(Z)}+NBgyKkAHDhvCNMPo4SZRhi0biweHn+}MpCBekg+
zi;my+`bXuD>{QXW4{_#Q`l7`f642hI68qjb&dSYd0M2?NJ}&AC*Czf-^gPat;5tYJ
zGJwflXU^Tby83dcJ}W=(^ZZEy@00gSPx!DDNnO$!o}^G0XqKBd6pDkS7PdxXZlV$}
zZAVCKtxOi^H#ItpgU9%k@M!mq3sDlw@Q3jD>vCm)SC8X1?Qk7Pc&uZ+=;s4lEL*_b
z=1`eQUtAMCoouF}X<)@#r)NXwIZl+nSPSZjVT^Uaey+Ou+tv7DEBaUlWsc8U^%677
z-05vUcD9xJk9v&P%%mQh*0Ow4+Sko=WM=+BZHFKxWmS%#q(HbiO;-X0YJbb84|Hz=
z2=F+<JsL!wz^c+sDJK$QylRV?7@P84Hz$Mp-P;I6G5~X-D%Z^8-rAAaPHQtaGQ(*6
zsk=m{N}$>7)>gxj-n8qK4Se(zD3k~?Jr=mwwDWBPFF%CiMM1lU3NgIz{kHBf)H?%$
z5^i4H^T%fU8+=Bc?O%;$jU9_svx3JA4QLX2Ts~?I(IbPDBb8PgdXR4T6I>Fp%wr*5
z#lQg-l<XMMd4aB}UkZi2uIi_UK{D=ZpEn?H=6R|pbcG=s69MeZ__+q%Rd(hBND6(W
z+x-iN06545ZCF0WEkL}H2X*wYG=o!~ECeDe;Uass$Uw1B|ISanm?zwnhIZ!Y-?8?W
zwgwItEf&{pOMN%#*7BLsFX`7zoe2lC=lC=J?8Kma_nCObUG)xLJf6w&H%;@)s>3*l
z!s2)tDLJEjOC%bPgZsK9lItk-y5GyU{^yD&2bFm13KiGIRIn{+ULBrvt~3XddMymK
zphG5)qnr!j0e%pl0^(fB5P=(b`+QjR>Ta~4TvO{0$BgK&B29meLLO9^-9_WZ7UbfL
zy1s0`S&iXdNE%8&0(3U9;5soJh&>n~=A~fU2KuUDTKBPZ8zbh5b;T>phcLiY5k4HY
z82a5K{`$g%HJ2!V04-zjPI`I>vj44WIy*P-V(Dv*HrdsBdNnKlAatu4_Q7eFG2KpS
zE44tgElHZs(rJz=f+ZUI*OJH_9OaS+mc7!If=S>{ix(k8w#IiXU%hMibE)aTZN|x8
zVdR_)QCr_WU1NEc9QgL8Sod+gevCljQ3zPDLmDx5kRFISeq;xt=zP5sP3{BXAPbNV
z09jUIqe2CR8}Zv4Kg--CyQ8=J`(X`xx{&r{!m7<tp~aYV+&opw6gQMHF3PzIpQtKr
zPL$^ku?|_Z2I6+s7y2-4nmVSLjOw?RuQO9S_zjm=rMrC44FxJR9l7Z~c(!Go>t_H<
zJk;bvBinM<b-IS1J<SZo$FrCBu8zJqM%Z?t!~Z~(fEbU$($h?!h%FZ}f5qro8!LFh
zDX!p(V+jx@%rqk1AWx12oC?Jr#(7Pct;gMO79`KJ^t5Pw&T$f_Z?lcxl*KRqcZ_Qs
z^a(@_s-XM+CQnjfD6@GR>t~C+b|%8z7fT#Dcfz_oxxLIMRYPw7TiRbf7v0_6<fJqF
zIW3gh!L>NWKe_)q${v=ae1l7b^@c)Wkp>EjemIyAwuIjDQ;?Vq1%8~chn}$-4p_>}
zfSNPdYulHvn9I;wPGKmNM3gr9TrUU()V!%MHpWYh7<t8awmBu#?`Nm>v3b07xLc=k
zlF8rI3WV!JmZHROf@DQ7>i!}QsBNUO-(h=~igkx?_S|Oep{XB5;NjtIgK2OhFiO*D
z*BybjDI^Io+|mJbGqld$7=_PJ=>z7FYWBZB1=v%;_cHF<Ws*nlzi5@`l<XE=Kl|C0
zvBYV=qNCj30b(|l$3NA5#eq+d;$FLk&NbIP>a2I#jL&k)n@Ag^bHo+x0c3=W42@K7
zA}Q0gpt**K0&14MiP7xL1his5YL1TYaTXro2!G!?9l~S;{jK(miE@V_ISRRmgp9r4
zAg78v>jSDdP`3$oYmLoMIf)+&R9I1Y@6v^2-)N>^O!$<L^0?r52AWkFf+l|_O_@PU
zz;~95gA(5kKop=UHB@natt@$<p_C##K;*IcArM8v;mWlXnFg5p1MlG?KxxMzaoB~F
zr$ni^yx#-Mo!i+`onK^WyF5voe@t9&m7w0}!^w2}fa7Sfu9dZR3jx=x*WNSa2;Oa*
zX(x0s-X-t?%@Ft+P~JxTm=k1a79B;cSCPm9dbdssNsK>d+~zv`(67C5eYnECdaeny
z-aZr5?8e0eqKFQw7$jJm)B(~|*rGp+eLGG2C56@J#o4ERj|SP!tS{ae{1P2JIAoNl
zZlt~cw@G&HYRwVaiExeb6=v1Pukc38vrb}kv%7XnhtTU><B1QM5|ZX7?;EEgY4kH;
zw8L8%d)Vmd%13%E$gMvN&+DHoiz*L&ejC>J09ig-(4~{{<Xu_Qa|~Zk%<a=nP;G0Q
zm~RO>8$XBq+=pRz$j{MZppmU;wf#-@u09=8{4&K+u9v1y*VXPyh**5b1cKqbnMXN{
z%un^U_tYE_uI}V|{Sq5`g+74U{4HNSSLWE~^1@BW_u?S<m1Pb{OQ;z(0cnybHdp^I
ztD|FNE$>-`@O1W53LS?Ia(PiQ_zXiJ)?Xu5UTSCGb)i3gA{H7zl_kmZx~iYZi!G#`
zG^57{OkSA-NA<yyC<!n`M9I|%t>k)qzGXasftK@A&Yr{aJFhD($yIlN{W24wFK2G2
z7WgQS3K@tr5l{>FhtMSe+DSpN@NOmbm(bZezQ+sKjiOc%1}yb#Nj;XYa!_Z+crre#
zZ*|thqx_9^)5Tpknj=MKLx6ga{Ed6X_wN4zNzzh%q4CH)k3v?DXagVMbn~#|<J6gD
z)G`1gd+P>z@{yMa`YDikYWb0`-r``H12V$R)$Q^@vA%JQ3_ynP&-o3_hOF<ptdGXS
z?>OJ^Eb^PfK_YVJp!zu_a@+=uB#j3}Yx_&R*3o3D0QlFIMNQahrJ&>TW23gPwe|r#
z$~+VpyV@6|0c8k3C*^YOw|obXZj$N>c>I#-d*Czr#Aojhpo8oH)|Ow{#bQUeE|$4Z
z^bi=tXzExy=7Qq(Q!&MntELbW!0l6+-uKEw$l!M+022n98Ff31sWsX+^B5K713;|?
z2rEBu7kMG>kn7%`TXY{7|D(+YdR*kJmb)ZBU_<}DJ0G|1kLSDb+fJTCPn>W~f+}yP
zi(>Sp>S(@{PoP=-S{s+=1g60(TFE_bac=F671ky%bftB)b41nheY~781R;&R@r1kH
zq=_bVeU*<CHB8u1^Uczat5mV@K+>C1rWWyLq^wCz$9G*Sqff(NF4)9ZUxYAmbV%AW
zjW_}zLk82%A0f9J;`cyC6)r>s9nOF1mTIVowmq_|s^C^#nD}`v8>ri~;iw7&wG(1z
z*W-N!P>^O^`~^|fRH7;F4n%JwJdbRbMEW5|n`s$=1Ejj{rvmO>@rUw*_6+{l-zcUw
z+e1{d<f#}m6G`P_(tS>G0r12@=efb>wcT=^Z^R|`A&G@P5kQVo0W)d|j9hlbC}{gM
z?P&3dYjv|*+H&YmA8FYvfjzVE&CCky$ZeWvi*HuVw#AR-KRfVTAVWstLMdCmCzDJV
zw?w%W1fqERj~;9hyIw4rNp&AK3QVOd<->qlV9oh%pAiO->G{?&ZV<#j3{+)r?YJXq
z>4D2F!{-P;o!Rc}lUPJ00zgyXGGm>i{#2D845n+gD?C{F3ELE53~`Lj2J=VM)E%;{
z^{5!oxOGq42Sdf<l)Xx!ijiPrN@%DtEwEHtJD(u4)iW>ym<31y#Hv9>M?AKLR|5t4
zxqomR(I~gnfSzV;)W`?WyzMO6?sq!YY1xBC%Ca6pDJdnf_f$S9Fj~cq^LX|JQBb|y
zmo#!u<4ld3s&nC-^;o|bK`VX>2*5^&a_L$f$WDFQdkf5yCL(1JqP6;4L;1mLPvrvu
z4YF+kzYDm74cN!oGeC6B9wcAwUg&)Z*J@IXKy}#6BT6k|8-Yj8k>PRh3*PvOBbp)^
z^2Bw4sAV@rqJ`R@05*Ig6X{`cFi}I3aqclmM1L8rtLl4kJ8_8i*f!s(;Q|NQ&n(u;
zds%hy?gJZAzyD|u1$tIfXZ{63eS_@W*e|8=5#Jd3(|_}z4`gQQdz}eyZyn0CxNe!q
zD|DK=_pxF)+2{%4U!|gxVPenx>Gu3472+-q022r_v()XdIbLv{oWXUQQ{1jyHq41%
zqfMd&!6kHO_-crW9wThGknhSh|8>v+;TlSFsxT`5suhZdwB7}RX3?*0Ynd|-s%KoO
zf7dKU#VLF<{#45*!OBp=&#@gup+|X_OvimC9tH&LS{~O7^+a!<1F~cW$iw894Xxf(
z=*rW&gs;))R^4$Q@vwc)hXNf!fnRpdH1)lOfQrG}+d`Jc9pwFM@sEOU@L@VxDFzu^
zuHJz#dcnbSl9?MV%88%syZByxd0B+R;)B`c?Cg^mgvuah*_8dF5hZsRn+~jc5Ry>i
z1r%5sAco1l<tqPVW3I_pRy5SQj2CKRo@NSVgK>;DuXOf>B(L(?t^S=DCBQPfY2VO%
zAnTmzJPyKz5$)GE9bw<2mj<v58x+3oJ}?>5A%(vqpvJCZmzJK})9G<=9Mg59x-4WT
zZ!H({mA{$&_<YuLtM+3`Po;sXY7Fs8B2sxYu6r^h#S$7tV`u`kl$8QV*xB#%zgaKT
zDpw3I4XgX2St1Xp$;|`giHO3dAi8JVV362D05b2lkeiTWRdAL}r>H%-B~okeS=F~M
z7EVrhs$(uPE5IyK&TV+E%>1IYya%4LZm&yop{rKVcOW+Qv*rg8aJ09JdtX3qsTD#h
zUwO(OT-o(8ltLISoK(;GT=R8a9E=>TaNX2CU)R4>T(6A%b%%O*GEo{Yb~_dK7q<zh
z%zbh(iQ!t|=I(A?Ec3qW{=vC%BSTO3^_yDvo?P>Hl1KYr)bux>*YH`<&S+`eHwN|n
z60W1COHiL24clUC&uOuo?yC1TZbf0})sHkt&^cJU%?EV(TUPc(@r<K$(cvMaS%eA2
z9Yn=WD|CB$h~P<fV;0SPGyVv~2ensYU@Y#YXhWd}&%utof`U6ZpA($yw5(7r!7wia
z`C-BT@WUcke)!NzxQYXy?`eK6gQ6%u;@MwgC3kT8K$vm>W6Pg_SDnW%=|_QeLhT@c
zMSwb1b~z~=U_1%uY0Z<NggcnxP(Ao`KJ>?k^5EfafD`17d!K6fe9jmGv%;XYL<$I!
z9ft2{gTb8sEwGeAJJ4XzLGiBHN^g8dAk=z){v#W)Au{pATCe8hUYCSa2=G;#or^%7
zMugE)XWwNb+m<Xp5d}od$pBu6jZOmJ%x6(&s9ZP7btB?;AQ+b$W+IpWoDCyqa(Lt!
zwFIL1C&0uqSa!74L0ZA7Ctt^=WvKA&LpMOYKl5vmsrMiI?l1l2Ce^*#DA1+G54>ql
zDe}Y>Lgm}ytps~ZX~sYu05GUmB^hPr#kj}^Op3XpV^3hS6&9pmv4%-u<Z<+6ddjy7
z%aL9i#fX#XFM+_e={Y6;g6=|@e;$m<Poo)wtf(`GXOs+h?V#WBl|s_>LeuQIsR00Z
z>U<^c^e3aj77%S4K%T7qp!fy6OKQO*Pru2BPZx>GGE4UcV#gIB;#~X<gKt&>q$nnZ
zxMEwKD4WzBJQ}CV%?OWg^!tL4mwy+zCz^<a0g=oDv3SZZn~AcHaU~O=e`b-^<@G}C
zI%_<L%byoDOd)e1@&KvkqdB>P3@A#E{|k{%OA&A=dJG%HJQ(M9+nikCd?V6Cvq<Qu
z6$?=CrY`|_$W1y7<yyMD=cS*OBVI<lp)YyZ<LepuXI&F!Ix9(m;VV|H)G5?ASBCxS
zKIO6FCjnM@SE=>E?zJBh9(KfEe@O!%^bOLaed(zg+Cj2{PnD<m_Mfa~tG1=1r{&>(
zhtU{9grO3@QAmF?F?8vYNXaI5!uf#G7~(v3+9p^0?b)os73Qa~s9Sv#O&jV0b>-@U
z8_UtpSImL2Yb{+h*<|T)F~P7JZkW$8J+3G_{$D)*>eHW>2%w|9`vJcUNC54oKkvDu
z$9}W;7J?Mr6)Kd&*KK1yWqP*dD=YIfqB6IYSROD{feHA<f-l9bP(o6wCJX8O5ok?w
zpAdso0UQ%L+AT$=51I`j^$+x|-8|USRkJB0foqiZk>0s4_grQ<d}c^Be<SfY5<y#n
z=h`HaZr_R<oM;d23jsshZQz-@wq|Jhw-&nsj7$azmn0}{N!EicVCi`6bnxXKlhZVi
z;<_5YVvv}^Rf|{^@%`iVX!0O=BU5Ns2eN+`g7GInTiUzoZiwou!xYMq<d{X#x#WBx
zjtd0G<Rd7JyFv=xtw9v_B5&}y@L-&L<BT~>e}u<__0F&n`FT5az}xK5t_;|J-`W?N
zcP~pPEF@|hqRBcvF{em0ZVi|&B$e?oHaId~7f1tXSEqz~Gx~K{5`Toz#%Y|d8#s-t
zX7V4in}3Q-6cJnp9W&6JlgEK2>EM?=b9Hk6RvAOO1zbDl9zDP}kLPbkIduTJi||~y
zM{gd+R%`_3OqI=4%?l_^Va)?aKTWu7-i$)mnHV55xHPbNe}}WXfNS=WYt8ACZWceg
z#Je4F@?4Z-eI51;=D=_FNT+7b5(J6)9)8JQVpJau2-!(SB_E*PD?KSxDhAZ`US-$q
z09^XG9T9}T)a@M_&8H$-bJ0A{9{Zm5GY<0ZI?+oMyjMom5Qp<)#|u0T36Tb(H9_{5
zG{UOqr$W40;@dOdkL?3+kk_j@LS+z_hNbT6y%)cA4F<U8%e9)Tw)_C0%vUUzsn>DF
z2(+L|QMGa*Qr>L#tD7;Z@Fe_>xpe7kQt9dp+9ke4Oj~sd*WOiMOLoDy`A0hA&5yy@
zLevzapUCeDl+!EftDHJ7fsswi_e_8O=pHl5<`ffl)>0<%fmuSmaPxx2{`u3h)6H3I
zt{S@In&o#4*@95z!19gbs&arTMF~(T7P$OPRKVv-G$<YunxqJw!ej4BsGol3>P;9d
z3RSIVQGwrRY~NyS01$+H+#N-C)3BpO&tg2JPmqMXcIFR0kj&ytgZ!hC_+ANUk5|pi
zE5INsnh}$v{2pA`=Jv1AqV?RA$j3f$*z77=q*>WgK;7v+&?TYWd%E0lxQ64%6$sY2
zfp!!?7tSoNWofArEk2zaHKvdp<@O*qaaQbw1eHr1sh)iSs~CtkDDNe|Q3EYoFQCUh
zpV2D7jQ6Ga%||)((9D|Q#ELJZT}RglCH$sr>+hMu`LY0-!n!cg=!TDn7vUZd5GE9}
z2}ur|d!@`RXO=AnW2K7U*HmEr=cwrN6bf{aW(A#1toe~}b!7e=&h><wZLZRkki=+(
z3?HLV`6Tc*-Z?Wh7|RqG-*cM&{16u9w&|IM-@nufi#C`E=*<Lh>{O4HUbecG@1z_W
z3Z>jCFy;i%xae1VZohu?gfQrp=t=TCjr(r2Q9K4}QDle|T^hvXmrRAHoEJwfh88^N
z`l>nqAjqsfl8JJr4h(U120vSs$MdR66m#LIQI1LgNyxV+var<?uUqS&>(w93aUKy-
zYZ_7gvUgvMGX~YVwiEwkK@_mVKnd7$e6z>cuG;W5hsf@(7!u?+kU5A|^=g6qhEa07
z{$E@G%+|*SusyN7lbsK|FQ2*molgsSbX8~q-b^wB==S4SGfaR%T!#G(Bmtl!R9_Lk
zMx^E-X1w0A-W+|afRkjV<TnSJ=X=kO_!ns(7`j7~cXZ9GDodylcf4=sc%FunD)$!U
zVyjk%ucmH~MSah1Zu9irJn12UYgxW?XvV$h&2)WjwY#LyygVxNVB%?j57FAVA2_eq
zF>{Ci<|&zAn!da<@NXdreJ*1|a9yNNfA&kJb(0sY?(rn6OVQG{IN5q6jD9z73$m>S
zSv6u_jvGZAHeDQs_poRNnF;S-5#AAiu-Db0fFAnY?DEfrk*q`Y<4mX=#wmL6SvO2*
z0|kK0iK@P5KVE_5m3Q@CXT$2*d|Z=&BJ@%4=M%t|RPa)IkZu&u^UCvyvW-2-NE(AS
z>r=0N#xEWg=?96xPWKF)xfK^i8yDxmFp2UsQ1Z@uS(I$$!sf}p$jE4+fZ79OeZR5@
zT_$&2<e)iKY4l!@kUROM=tl~!R{*QRZtwonk9-0f{2|o5^C%`|F-1sqHEVz7izbiC
zcVM-0-cd7vfIm>K>~!-al9I1A#F&a_bEbZ5TqD3qXa5-G{*Ucy!v;q|PNU8dXtqkL
zQc)3Q!h=rP>KbiuF*8c9JZ2?~k4$}C!@bnWKDz7@xhN=|><%O#Trs$iJN0)v06RDY
zG-TxqIhSy!lC<%YC4kk!(r-(^cndW0Nsb!7f_gmSwovaYiivFC*j*kRcA-P2?9!Z{
z3&+D7q+vAOt;>Phe_W~yN$&R7a)3l;N1*_W)`0+M3?@t!SGwEgjA_*}NW_|qH>x`T
zlNbYYi<w}{5b<#E9Klwqu&Tq<*MW7N?4?$Szj*&luJS55*B}Os4UL!=jgVt|uMLtI
zW>9x34+*g>h{_p|_PqRSx6fo@QYiF{5d-=rHRIVJaetAmhzN-djXmfsS$W;EpJkhG
zUh+doTv{LgMsM^WPaG~`0!_nAYmcGW<``W+Jo1^Ss>@Up{@wkj-qR)v<y+Xw)ELW}
zuH5ShJbo6me(7T$X=0TYCfLK=cJvG&IK|K4hVx5%2*E2KPWtP2i7^3$ckOD#<*z|I
zzJ(w>dFfR(PT<M+Ohy2vK?cd4-=J$`J45>ZA+?#Rk^=+^>mAiKXdeNwO>97vkZS8v
z_(5)(>BQsjjxyZ%NNZEDC1V><P!?2tPr(>Zn8H$Lq=2lyIaSn<3;S9}tRg6uU5T}|
z*zx${f<Rh9d9*u!-yLvCA{};tAi{mRbr#F<Cpt7{WG<9<C+|(j5u-$DtN@!mnpkq;
ziLfCx8%7y{SIcZN8<tWcZ%c@dO~S^nfh9IONFd#OqAWX0sd;I9j7$e6sdbiokYn$@
z7|L2FB>4>mT73v9c42{Dm~*DYgW6os>xw#<M7Wpp1-lc-vg%Qjxus=>KZUvew#_y7
zvfUFpdmFRd6-8^sIak&W`(Tt)><{q43^OF0Z;WgG46MnRq0HLJ{82tbmNxNwZP^6w
zLK2GT-*O#8RNnlzxwD|GH^iO1J}bgZ|3u*QicUv#TZ&A9iHJ5}^y1UN=mj97w~p2Q
zVf06GwjtRn6EqsT$2H#kp014&p=Di2VjVAf)G#<8@%u|jBme3SQ1O(DLsrs^;eZLV
zvLel#(SX7mjJ1xos0LOsutaj!F7@}5o!<(kdcc#9$-Q~LkaW@nXr^n^wE~nQBis)6
z2FxDjqNLVE0WN9I^az$C>kB&7x2+Is%W`i9mc=*|0#RSRb4gU7aw4`=bPC|`jKL98
z<$><JT{I&53+ESoJWX+{EuwzN{d1oO>>?n0865C=@dYxa=1@Nd2`g&|VUzh6pXT@T
z9-ezm03M8po_%yvwEftB{3rfwgeI>H7I3P5AAhNXBtG39NNi+rxxsJzxX)XrH$-BK
z&}L<j5YWfVFq?b9aQafZnfi{r%j_=u_V2c41r-nRE)O#1<=E@!XLK976*tv#C|XXd
zBby8Lm=}fBp`vNb6W6|*Dkj3!8pHGmYjki2scgn14-V~{bLXln*DZqy2`2bm@oN6&
z;TtTFve1lgt@arhkXn9372D)Lfophv9w9Ku`x*Do$KgCx#%t{QaVz0laa;h4Dx&Py
z%T!#c+g4(W*j91{ZBPt-hT>PHnJrN<#67lv)4{iakYCwjl-Oh50GTfpadYPI6V^0i
z93Gp=icsRh35B>oQDMr^$$Jgp3Ec-gEn31eKn%C<AlICSfaUwai(7y!_WZW$Ihayz
zg|He%VgV8={k1z{&UrReaXIV?WjDc`O3mG0phY)e-WJEMPLoYn9TK>s_XhO}m4gO!
zhf42qkA2ZZ=25TL0np$PS?;*`*qH4~-Hz*-1ZdB@UjhM<;D`scY+|J|aZ;j<(UOO+
zj5fXw#QoIs+L2oda$A484Mq%?@j-x)q*k$B(4q;a6?YXBChA5ayTB#^T)l*dUfWH=
zL4nlaqA@jz9-#k&wvIF(j2NU#>lgiKb0~1LVX%DpnSbP1f$l_vn2Kl#=x=v=?QK@i
z-A1IXx>?cvJvQ$izy>=(lu+ORRII#+(bAuCu$xno0SCAo;imQ5G%Bx_X(qsG9SCZL
z#+)Htg`uSNZyn(j3QAUSEL0CpfV=0o@RV3c+WO`Y2>_-2#FLnHv10GqDgl%!UAxc(
z=q*}?l<H{8myp8f#;iae$rV9tYI)E6WyogZ@4EFH3*us>f52bln;dv%Si?uQ)drB;
z0tfxC_0d{;IAXywqi<IspQ5bVFn|#4tT_~)M>{$S`M@qwZ8?Gm=o;yon8jVHIx0b5
zlW;t%6{qABMUW0tGJahqmTB=%TLnaee~QP$S_RgBRX7BmaJArk8{P7`pRl912VVYd
zr|Ov_X@u)Bw-XpfpAIgRLaG4=US6@JVFMW|&SSvPxE3&MaTRS0!B?P-p1M$)u-EYv
z$dypcK{JXs3_Hqd6u~uGSEUivpNbj;d$;>jssTjMG??rG&=`r@xIwdTmeKo1rh|hy
zT}^nloD|F%Z_lV1j0nz$pM<(;@6k06C7=OTP{Ak$#|s|P3@y`zEwrFz**@!uT-mS1
zGnjcAjA~`v#{{qJCAWu~6enY?7bdvFjzsLxW7G6+KBXyiw1;S>Zy#O3?DU}^@Q8-;
z#$OEsT*2@-#(p65n4#5$IWxZpU|x{w{fuK_heg@3vYDG1IPOz6)eE5ICy1$^8yDWX
z2MS7?ZUQd{iN|j=1o4OgwCyB012vx6lAeD9DQbby>`M#3m^C&4Qy=nR{WOXa#fB2#
z<HQfWU<!Vv9+{=GHdW0d|Ciy}wcO#*mXObWmxWOmG@~FH9$o!GSI^KHJKgU)YOu^!
zfhk#0GXRtX^QM?D9-GRZfV57^_9Y2?DK<xA763^I@OS*{+T#Wjc=T<pi40godid1G
z(69!;78L_9FIilcNI~zjs<oOB;p&w24yY9{WFpm=B|wc!_;x$B$h*rUR}CgyYah8<
zvFen9UpS`q>Njzr`g1M8sUWk#7_x7tM_bY09z2t<jiUInNS!VQ*LqEXzRzBw?3Ti4
z7;J*8M?y;7>OuKAjJg%}6Rj$qdSzS6dWMK~jq*^9CHjd*9klzL4-gHe7HjSjqv8<=
ziLqZy>M+-P{V4SOIuX|1*&?c)&nDFbANeZ>dxaK~yvaUpX-o&2RYx*TcLL6mo(Rcq
zN=XzU1t?+{ei4pbm2+x<FYZ);%jpZ*<YNNx6$xBjpe*i7slYT?Rb=$oe$1&+hq$u>
z<WGrdzn4bJK`VKlVoNKQY(dnjPLX_o7-Oi_=`X?OV9M#Uzt-ptT%^{hVL>tEpaauE
zx%jRrTI11`cfM+?eq{t6`Zk3j^Ly)sq)lUZ#Ih$qbFpr*P)=-sQS>V+sHUrM-%|L_
z%hCYfqqDN6G|Ki6Coceim^H98wE`)*Nv@-u6kCj}<q0`NxMl(*z!VJg{Py6wdqJ80
zLQ<02w~YObYFpDFN`liXX}Phi3<t2b!vL&|&px09X-juZj7gdrf?{_7Xa}QcKu4JX
zfP0k#!8{g_gmBHu$JZG6?{jLN^>nS@|8)G=!jIqed49-B8DUKe+NCdtIzJd*i00XX
z(iYg-r*{~dW5{E`0FO#)`lOBqq?UrtT0Sy@!oVzI4va_e%X{rIh??%+K0aQG<}>al
zV&u2@)*iA9wl>Lt2C|n)3*VEZ@RWPU4TrV)7+9qstu<U|4a#20Q7quc?_T!0b8BHj
z$IAHiT(&lhg(9o~LAS9=#N*_DfPvR0$<Tc|Wc_yTAibogt*xb}M$<`G(B~Xxq`T}E
zwdfIua)`}bbE!?)?q7AF?H08KnF*b?l`oFSDd+)s-oUK^#9H`n52X7<(<O=JASHhC
zWVQ2o1jfJsZ{b+RoEg3J%I>&q_LNEA$AwmWo_4^lVhjVrc2ZDiNd*j4cmdNwEvEW=
z)C<<;9n6<ewSbQ0XVna2pII*iov#5Gg&cq?lESrAqNr@W_`Qzpm$Z1yclRPFDsH4v
z+yw5gZHTZS2r`+^;Q&l@j(Pc)lA{B@VwPZalVz`l{O6<ll+Z+;x<+;?3lVM08JPFU
zbQ7?wR67MJjIy*-{FOqKs4_sLoe*zFB>rV}{;81Qpy*Vk*4$SO9AscCj>`I2iOA<`
zjg*@&lJj)yg`vfq)Anry=F>H$tc|sl3lm-rX{vB&;LYCI)BhMsADE;R8($P??6gN+
z$0CNi(*?<uxCn4Q8ae?w1lOVQVUFu|Gj$_tPFBlvP%Ai36U;2oOttU{1BIGLQHI8k
z>z7xUZQu$EZ~7DYf|r*%C%E46T)6%G-oe%cL^9wELt5+R^4oMmmT`8q?;NSl#2(ke
zuLrL*8$Z@Cb+tYMc;dCHlnazlQXv?v(Lj>K2AbJ9Qw60_6Tzr1Q?sZpVyXGJFQ*3X
z>M?gI>0P6trtzJIHQy#5j-*6sg>|L2o%ffIdWpL6{gOnv@BkmzIAzyF3kWG4KtRbf
z#O5a%TtIa0-4mGbnDE38x1U6CQMh~4s<(ugdm4V`R%8-F;VU=Pk0FA;jQ~{l6zWU?
zk&;gv)^1w|3=%0$1E3ww2b!q*P6AsBafP)$u9M51Ha^I9?QeC$9zc>zMj7bAe|-I-
z<v>s**m5{0@@|+un3vXKmbe6UF!1XoI}zs3bk?EG(RnYveB``F2j+4xuhPLTGl5l$
z>e3Mi;QmMQC2UFohzzb2Y+Da@C7mmGtd|RpVQ^{|b9p!Joz4(Wbys4fTLI)}^MndH
zjQQ#9Lui=Zyk4ICY7R6GYSoYw_G&<(O-&hsDo}BRWxLwI0*MY0;KMbnAg&xU>hJI~
zr+ZKP<1gu)#PE22fNMCJ)_l=`46NG$wC_@vXbV^Wa?5}bFVjG1LM`_|J~rqFR6SgL
zIvl4-8FK_!EP{I7BH3zqcwRAwp!FfB=a{Wj1%?7n-lX=mmQ(S;fYOP4dpS_8pycX_
zZoS8t5i<n|`GO5rb@FXh*QhmM!h1k_C*`gg0<?qy^IP;I_q{2vSeMK~1^=YWWI+p(
zJq7{&_oievcyXi4LY6;8=DglQU_Y6h)badVO2dih8rjekcLX5fbY2$gHPmV3uxvqm
zBQnqX%Irx$K$bc7s!}|06Pgsf5$t7{KEz)%_Sm_086c+DPj3~H=9`1`yqs>-wdwTN
zmt#OrV(3%`eZh#e7B5^W^Au)@vvn7}J~9jUdrmjhHjEdA`mK_NC~H4G+^>m-KRQ>8
zN<v^$#@mAFfme`vT(ugJ(9}RxiW8-rXH^;(LY}}SOJ%{cEtnut($m%AWVIn?=@R(O
zE;$Na<9?CVhUVYe`(m1PBVB!{pYW|V7KxOIvh1zzcO_zRx6<OCQ7sH5pATY<ItjQ$
z8_=E?QWv@mk}phnq^}0^>-aYQIbHAcZkaX+b|M?gAk$!%Np>Pbg1TO~Z*^M9J!+j{
zZ~(~jtoK2}AYC+05qJMMMmAp-0SaY;MgKt}kjyOFTorrYe=@xfdgRn8@<~=LU?Bo6
zt(Y@!qB^%(S^XejqHEswrFm}-#%9_or9}wJasYIF+v$m(9z3@I5OQ1MK=5<lJzbw|
zsNc~cNrT2guHW(cHY}77<?<2~;x7-Aj+HyUY+DzxHdW7i4Y-F`S=#m*%%t{vbuM{z
zR=R+w@u8~fP|#s5BLawAr0#G($AaoN`XT^y<&sz9U77f2fYxy>h+4dTLL~=UF*L2j
zcNath(=s_$xcqM}`H`NYyOS#>zuKn$YL2trGM;MVxJC^CkA2<Gcvo`?!jPJb^@h+0
z1ovJ2%8-D@i+88bEbDP<NWXeM5hv#<t{;<C_?ry=AD=S2kaGqiRaeg0kcAiFPW$5=
z_9XF3UftglGY$?S*Mb&>d>AyQMm#QC7QdZ}BbmmDF$<>Qu-Ev>$Eyt|ME^yBKY~X`
zDRF~jrDr97`tM~aLU<nb+pD4w;)dWDUx!X$l#kjyIr>-#s{hc~3r!mRU%ws;fqyHX
zrSSX;*w+DaK%zPqpC+$vMe`~EdD3!rI)JHymdig!(uMj>bbvGzB6<XHk`72J8OX3@
zJ`9r1^f=|X>vr=ac+8E6t9g^wuzW>^_m41d+G17slI8g_yid^l{#CNxPh<=GkEicV
zAwm2Uh6Pc4@n<ifi}T;02VBC5A(evc|GpEpy1(*()Zij*1NUqH`<wsqtYw5DD;Q|V
ztpnu45I<kQs@zCV8iZ~bG^egqP+VwiCo6CP@1niJVz`-tw#v>edQhw0)h5;BSANZW
z0RA^HVnG9lE!_egB4Z>@OK1>%R{<1nhPio)TF)|$NHX5Mm`AzDI1oRt$F(=W&rmz)
zHnq{llJLdd-kU76wH?*eYjqlR`*iz?%wWg@ZpZL9?Z~r#S&s}R8W8_u16Q^E5cpdP
zMX{?q^dH;$Yhiys>0d*r#3)?7;p(kWqJzumd}ekv&YJ_i0=d?GY!T~mzxu64YCu=p
zW)kh9?*9)`w<Dtzq>8(5s^0jICm+Eq*2bh674{(oTXMPRu)Xece@RIujZfH4C7sEW
zhaRWQD@bPbUmg3u#^O%`c)^7+P^&u#NKNp>IY<0|@1P2tvD6~TTo33S$nXN&dFwE&
zbyee8Tp3bVL`#~u+ne*_of?pOZ37&J$tD80*~S`Bv+@El-|b73#|L0QbdVJ(tHzfl
z%b;@oq>mhJTW1*zMtuD!mS|!}(8>yTK{Z*}rlETkfux`a<n5hv`|908R~JVJE?P(G
zRI7;0-x!K>H^O@VmT>?1kpFqqh2Vg^t2EQ#f&Q_pG9Z7UK<DbYkASa0SCrByMZ~TX
z^vMH3;Vouj1cd%<VW1=#`fAt~7f&1<0m`9I6q&#5{}I_$kC&x?imStmecu7?z2@C2
zbUCc%Uf*l@wrF@&W)|0+3MWb6PKA#~XrTWQ=MFS*qMx28{}<eZK4qY1?f^Y=&|Eev
z`Rttf|2=bI=$S`0<cM88^Z$fN-#@u}<_S?v&fSBihX2|4OQN}CzxTS;NpT0ePlx#L
zTbh^6=!kEoYV%8u^j945TYQsYfBmZo!S<|CbMFru_$Ost84&f+%*Euctht#7T=n5o
zrLKQJ^AH(Ji1}c7!r?~TEd>5=ry_+By8qkf5TQft(%Xd4y*_V*hBx#tibjFYUA_r^
zP6Ce{LKt=3gphtpaaZf<p$0GwxCnSZk3gJ#o~xXeXaw*$YA*donx_feE)9zR1P~%2
z%OCcV!qkM2uh)zHYeV34eTQJP$>(3_p{rp-LPb1HL%Q#FexB>^w(c(u&y|A0@O|2T
zi`YF{KnbY7oKWmKl#PFRDWf-Xk=Dn?A-jO$F&MLYr1p}~L`Mo&!)A}3gVtM1zSHM~
zsrV&ptjxjS@O$5Fi_#^xuITZTUy*F5ztfCw<sy1<fgk2(r7bcIllcI7cap~@j90u7
zaO3~$7m*@(W{)QB6~saB`ZL7#i~IDV6Z&KCq3oYoBCDo`5-G8F<DY`M>&AV>LFmW}
zjp@7vdOTEsbLmPOC+_yU9ZYrb0t})*S*VY14wsY#B0?3Pm@EN`$D?s8=!3lg5+nT|
z+EFWMpj<Wc!8BHb9kPgT$uHz04aZWdP3S&azIOr#0i~1h@e4NKYHH4;eeC3@bJFJh
zLPXrn4!!CE@e#`Z=T-Z&fj=O`1Ub13FJdr=d*||K%6<(ijWk$YM<XzErvASe`wFP4
zyX{-Llm_V#qy<z!x}-rGq(ed!kd|&xN;)MDA>D{{9gq+d4&B||oo^q%@817?@7;Uf
z8v_Rp83WJnoW0kYYpprw`s~wkSBu}`d5Sp-^G2L%xn*5wjE9Q^PYL*zr*Y1YP9#(+
z+;yPiX?o9JPu$d)9?8&T7r%C1aS+iRZqV!LdHJdn{{EN(%(OmH$#hO2;j}qkK@Wlo
z$9L-q0yj6e<}NDp&41CrP&nR#R!lg!&g7siew^kqt_(jj)bQW~;?DndX7=D`=7^qF
zy{%GfqRzO(Y}<S7_IH39wgcRbJ9T^1VTc(Mwvt;kG^zqTDgW4U@xx(qpHmL`AKCf;
z`Ei4f)fhe>{Pd^pi{t!<uW#*h0iCu=1Mv=zu>GyXt^JMcr{^01E(Tj!$lz5uM1pv5
z=f%2EmwFFI6hq^6tO~E&o_@6ets2s&CMcu}`~&Sj2LO$&6)e*;`36VCm&Y5iZ}Bq)
zjoKX7>rvp(`EMQ(4-u4&T!L879tvp*Y^rJSe_%c=9I?0Y9I%mf?%N(`#?1jlMBJv-
zhWBy%sSRE!Aqgz?j*R(5FaN{m{&)I#0unRj>a6Mv`~c8ipbSzi@EhI>#d2Pn(pp{k
zs(QxQlG&pQel!e@*wr32co+3uXTO=>gV&@$o=j_U40xI(0Px9&Zd<?k*SH|513&%X
z<c9!@mJ6EJhawVfju?Xfpc~|}{z;eQ@Oy`BM*K%(L1zqybiX4$0^L5uY)!a9Wq&|&
z+j}qbcbn=4Kip}#&&6JeSEbES$;Eb(B5%F#olGebJQw`uuNR<pa{MVMJMsr}J|e`B
z2(++@CEdB0k~*r`_7jYmRx<}=F}ye69^9No0Hr#I5E!CP0q03!Qoo6pKDS5d2fw*s
z^KchW(JcKw?aDXU94*%l#h?#={1dRt;W7uX?UXRa(2$V}y!&#OIOc5dN0rh9(>qx>
zHhL#f;{Vrcl!u!`txCTT_-N7H_hri>-WohsrL^?l2#KV$o~_=o_EMtiFEu8Ji2wEr
z10DR<g?03iKV!B*DcWcZ*vF}w*F)dVR5Y~b>+0jrCU8e^EqyJ~hoy47PEijJP5R8U
z*FL~rwkESEyY^|JgLrPn3x3P(mGU2}lyty5Io_KSHCg^l$xi?_%g`TM+mzl0X)lv?
zbacQIs{N26uaxH#(7m|^vIj-(E0r|hU#~HvszwF}V?>&A;Gnz*2PF@?3hS>SJM%+;
zvc6DYZV*7&c((jm^Kc{#!;z^d^z`>ButXri{%17Zlz8#_rSL2RB9~)0K-Z22$rWhe
z_;&Ue8UOpo|NA}h!-F5+#Xi-MKgSmpg#aJoVlxqFel{1;OXG!?|3;Csg<O$CPuVEx
zA6$SYf_ST48pm7PIj?}6nwh{IBOhXB+ce<M7iIA0jB|oBuI<Xf%na?%l)y%uc)1Z=
z-3a{L^c}hxw1ENLyE+iae%xL8+~ES-o?97vSs47%MnwOIo9p)s4uHN4MByw<q--}*
zq0&T%w;mpRLCjS9#Q@%aVEy8lI{>0|E*sY*9w-c$y@@&gul)0`=?KpRoK=Zw{=(!x
zxyV<H5sdrSps~slzfx?Zs_LU#AD!+_1aRiwQ4}$WB>2B#Qvg7_9F$wzEi!*-o^C({
zN6P&c58{2fF&OPQ>$bK4oRD!hhYOZ`xZQ-7*3x}C!4fmSkM~FtAOkhFMxUvL$pK;e
zEJ@)U7oNFA>i?&PfTut-@QQ{Ibfo@*{gQ<6Yy3Gx;~p#~jccfkbp(D1X8AVq?EoO6
z?DOuf)&-w$^-0nu+kbu1C=U4VmC{}LgMaxj%s_(?RoY$Tu+aNwnHmF9&)ulkr?ESC
zDL%1+2Ox&Cz}cUohJQZ6YQo67%dr-6{PZGVPyv4|Q?P-)7IYSXg|BiPBpPR6()1+8
zISKFsGhZ4?16{LFLJ#~1UxCBe*DRtQ{s(2sz;EoO;4|P$)A3_6({igJx<=Wn>l5+b
z>?i*-nM~u3<zrq$h6E4rQh(jx{j@{?=(qmrg{qsB3yu_TKEc1qE&uXqJ_IkBKbz{S
z<)3gEiV-O8TB_Y7(C6w(aB4gFI4i!vd)D6^cRE7x*Q|o-R(uJ(5CAXmbdC_7Rb?b~
zLA7y5=O6&0>!CIyI}nb`Nfdkr7vftKNPdsExi$aP$M7M%P+kj-I#_QIpXSmci1ac2
zTJQ#2D})XKp6JB?5rHD$5s2?nyz1Wuo`3!$1`|92ZIg`1!=ta7)ZYXfg0bz$yQ}wn
z$NqizT|3j?Fn{f~fE{F)&=b!ZWF=Au5*+?A)ArXM&bJZqrqpd?M@9~U2I&B%n(@P9
zk)R4(`~m(_Mxks09x{-n30=-riPJV7*x};APs5K9n!@luRRkCOUd`etf85tU=&w{c
z*+yTKah^R9-8Q^b?(W;zv59{m1%EQhzn?7e0{&#HS7JQ>e29Ph7Yuk*T-c=<QG!2N
zj3VuQe8@!ZS9e^oq1`5xOtqXr#$EQu?UNRH#=Xn-_`TJH^ex^OQjdMs@B9X`BJ1l#
zaYU|!Gzv8Dxe*_#0ubf@{_aNt@vnoC?TzungJ)Ww!v&qhe~hkWtS%_3R9uRHqnvY`
z3ZPc_b+z{`OS6|b*gUnjReIsZKkvJwB(`*#rkW?EpfS<n{O$`t9_4Fia}_sRW&0=o
z{Nqn-U%hkA6XnMF5Gh)B+;Vo^jLaq-3FR&83uAdUGHuBt*cmlyMdfzia@aQbx^RT9
zHnW!0tj1yu=lXP}2CWAB=neV%qnO2!Iid6}Zp$&mD_(wr_|u85b)Vd(kEr7}EhXpX
zCq1>MzDs%C=1nSh!(6IgTU&vF5P4K0`e+)`ZAAFoQp4DT#iB&WnZ}-VaizMES(eXN
zUd{OrZtE8$CUc!ZU)4^?pL!h)dDt!$CFYv0t)@5LXIUa`5Lv9M&UfL3xh0`L=ew~s
zy&kc2=4Rc~t2DcIyhi98S3Yfg2%Ht~!KP|#yMT6S9}JQcafWfI1W4};fGPM~z}OZh
zDEB(RHi;0oqaK@Q*L%SJl@hSe%Eqi;_@u(C>jB|-zpOUAS1;k|7+BE3D^9eZtbY+y
z0Hpw23m5B0dakFB1oy7V8s^*(K3YMPZA~?kF(g6EEPfc>3lf8c<DPl*E)d~fh^S!0
z)6IYWYJ*R%YOiFf{y+{hLJS2UO9o<aLvQ@98P|K3w*_(qW~9AVu+lW`qh)Q+gEnrx
zIC!QCa@p}Ktg;kb#kV)Hf9id8t0(b0*NN)Uh)8|iL(ehcUCHH@M2z}~kQJ{pY>_Wd
z$SLN$2>~IVaHaF-oQ%_spBL_XeO6MB6_r{qie{bTXChvQ_F8^?5FyWbxx1jRSB7CX
zk#AQgZKY5f4t9!;DlDge20x2jA0<@e+%x~VTI#ibQ**u?BJ)`!IPqIW-@s0{*<Es*
z<Ht~9;WM!C@$}cGoT*R<4E&ef!4%K8hsvY3&biJXkeE)587+j74I#lOeoF&1@c;sN
z)E`r9>M`MFNKJ22xZnAMu}cORBp9w4&vljP*Aal79GHL_spY;4^+SFNq-7c<x+5;$
zuZ6_{cJh$M|Mz7Oov|`N!b7~zz%V>W)48um1DQC*;B#nDwY2w)sUIs(v{nJRZ;S!5
zp9vAkE{gMAXjX5!F>K7%1h&Sds$Wb0o(dkmHTW;kG=#om9~O=w9v=EF(h-qtLuZXf
zGxsYunDuX7ww)A!E)pwG1L;M0pV-IbAYW^hOcSyo&>l0mV_PVHZLikjROG~X#iZHh
z_%7vk80VrYW&AV0jh2q@V%E$*cVJK$c0oAT@;*%0orEXC*m#;SE`lK8SXC)mZOtcs
z*M+KMV>Dx=T@R)jZ_tU}>xKv$C4T2kW3gUJ9}~XN6e^|yt1|Ap#h(JFxx1(AH}*HS
zHv}HdNkMHRLnfV{z%VKTGR>+0^f*PLBjo=bEBznh4_YlcA)lij{b8Xjosexxrkm59
zG1s$Eii0T^BN<AmjoOZk7K(w8-r!>I%g?IWyzF|Rk<znUa^JX*M84lh@d4+k8HTC+
z9-`fI?DV@eBfkCiSD3((HIU-#Lz=rtL1=z;X_m84Qb))5HwS(Z?`1^#&?<F@m>{1U
zOR9+N5m<2k7SB`5)0(hV(+Ol3(!B4D8_y{ROC=-&?Q0rIrIK!LhOTVaG}3xI12cVC
zgnUjGkKQ^=#5Z4#5rt|A3AFKKdC;{vW_?aw_Nb|_@aFgvCn!;XMCJ(^E;Z#-L%7^r
z#Ou<PN}@S}4LCm>?TwWgwYz~@r43MD*4fI+X+ZlqP#DgG{<q+@oC$!}4GOYu0*G*T
zpP+Z!uFI~>!dm6vpjc0MymBKBF-46y^%$>+ntiMDwoVI1!yJ|0;@tB~KCVL*rp)U7
zp|$<H2|?T^Ooe6PEmJNVO(gFIqv@znP*f%HBvFW1ZVS{j#0t4pll|3*vROUpLg@_L
zRRsoh>IXe6{7NP~VPAMA%$}|EJ<^GpaZF#k^GshJJxi_Qn)tof{4@n!K3+n;T2i}L
zoc3I`N&B^z1BHFiD%~#hz6Vd$EqkKNeD}ASofH2<jTdi6aPjmr9^er%)>5+-54hyG
zcI<rFw#OSwTrv&!2poU^1K-E8Amf$Xi2vc3*xMBdMvoB{Vk^oB1slqQUg>VEiFMt#
zDeWbqVk-S(^T32XU#L)gOop2}14|1*-+Qs`*)l}~i`(S8!4vHQbbqWd!ahQ)W2$zl
z)E%cAu6=9PgIat9XO~rOV^nR#SH8sm{o66I@w|=^eQ7IJIa0Bhr5aR=62f*u@yQ~&
z+Q~n>yWnI$4zsrQbye)Lyzk-?TVfQ4tv3n^TH&Xz<fP!EqJ+i%v8B|_<<eIa#G>T>
zPIXnYuL*+dM?yXFB24cBRg>pSOY-P?XXA|5qV9Lj?tQC@^db7l#9JmPK5Ln*N};?d
zondh@U___2VwsCzUzbdj@+)O*VXq=RAFG7h$3MI>zQ9;sLc#j65=YX<&mY4wjK4B|
znkjgfg?q~%-t-l{0pYk>>U=5n-EK31k%O6x37!AY7Z2sF?r<x<;}O0eDgu6I6oNIQ
zjqAcW_^oGS)wUhCAzz&g`c?={=J~^sVW5LA7R#uJM@(!E3rE8yHwNSOpXs74P=jCE
zpp8^I50`m^sT<I8fqp+{({W5ipJ(H9WS9skt+VTCYsP*Yc$;ZEh>Yj?JF-{({(jGg
ziH#H2;<|B^5nw2SK_fI_(9fGb6tr*o-b`UUO4Zvt`H1gJrt3)iLcZ$ReSwB+eEwBK
z0<-HgZuO&4WQashm5g?HBt9NQD*vuMpDkEt>|Vkadd?*s)IvKP5u>2L8=JjYw$X>W
zLa@{Qk@qbg%~vQ_2}|)V2BskidJoSFn-Em0YrIO+Vir7Y<}WWbusAf7>#O}8;*=a7
zo~c9|&)#t1xxB5;q`rB2mm|b-lTf!cU@6|#(SlcLULHqsUx-4Ur1Ux+X?yo~0Tg~V
z5ZM0vw_p}1v@%IaKMa20{D!l9%lL%GQl8^Z^oXoIvihAiH4er$HoJ36$9wX41*RWs
z(a-E(&&cbao1$jf8pt%)b7mQT)xwYF;>CGM`LUrZa<VA4^62P?h|FbU>7GyY)#l94
zueH@)_eyQ%==gap9_tgiSvC10VH=6(T7=fW`1Xp;hI%|qm>&IYTNsLt)~4vo<k-gk
z>$06o!o=srttA0h*#gU1`ujOr(K(VVioE7up_Mu_J!jV#QhL==#&+dH$M2I3=1!P)
zDTM<)^5}A4ay8p4Y;?-u0zY*Q+@ySYI5z!IY6;JaHR2YPGQQI<mu6G(^OAU1p$@Kf
zuOV4<O*aKZ9h2b^C75N4E6tq2QschYbO_De!)BlraH-P8!20lfI^zVQ|AZosq*s~8
zG56!Syyr?*%xs)i;`7qhk(TEQG*hLvJ)CZy>qmlJIHqg!vxQ?{b%NF=30uTVSw{>U
zLLGwjyUMJmg7vP><i~t>$ous25WmH<?>$qlTdG~Z`RKRCjU!~#;0&x7oS=Z8k>;=p
zM5A5pA%xK)O?9zIaRmDZOI@UVcC^hmr+Rsjfc_9k<8zG<%t;A*^;+_~pq03vx{dtS
z(9)5r&=JKmyR$l_5v|-2WgfUouVd{Ym3bov@~38)h9En?M5z9}|C_X+iX`WJS^&pW
z?|D^e(%4EF)|371vgjqZO8BXXDkOs61kAh!Xk}Sz-!eI+;T~7_YZ8q8&%r2UhU1@-
zJ~UxBP+B3vP+X)N1t|P<LNQ>m({^9ATq?%KkVL?DDeW5=zFHO;-uku7hXq6`R)RS4
z21lVq<!h7Ny5UZvMN<@4MvL9xPeye4#_wS(so4p{AvMahZ?;qf=WTRdak=l8oIZTP
z=KK_WGxW1<h}87LMa;d%w+DIGOw1zJD;pRb`Y=W9X{=x2KY}N?5f(T<ORFIE2(FoU
zc~T8fi1@D@Dyg#68Ir5{&ZG^gl)DxDKW{@hk`4Sm=pEXd{t3a?N{B`XknV{({WDyB
zefK?S>dtDBy}GxBs6vEso_2$4$q}UyRID$;Ua@6y01}KX(^6-%LE%&O!E431-C?TN
zI^Xz59=aY(htAo#Zla7-%P*-lr1<P5tSWD*uv(f<ZYHBb!q;=rLSlcN4tz0Mn2dOz
zJcif`i@(H{I&~mUGgkFagq56YYVwxUZoQMSY2@jADcvWpsk!#R<Bmhg^Dd1WU1~>@
zvd)%jJZuNqowfR)i~q{RQ%=9>x<{{py65a1W8;X#DOAej1=8(}*VBl;b9p{#cN|!b
zptDSfGnD~!B}uTx&hz%AzINqH9@y|I8O;fvjD>gjJ)g%U&5=*`Q#<`Eibg^K{Spfp
z5WC1FgLR%+K!u2Lwa1OVP^^|uN<k3@2py(i1w2PO4vh#Uy-b9pO4ORZ+4Ps9yvb+#
z+bj&i<G9GMO%wQb;{j6F=VeqiFLh8!U6pj2+_?WvcABW;D-&1U3po3akpyGu?YBoW
zWoyevw!$!b)G-&ZJL#WzuyB4h|G09<A!)m9@zY-G9&iLwjGSLo{P<AE>e>_&EnGC&
zGb#ST{+IK0ffw@Cc7&x)3C|7|7}|^XOtGLRdFR1MFq8@cp>?-$ex0-SsV@AcY7&CI
zyRDCn54K|Te@d`@IQi^R1**YX=JwrG-*ZzF^y&+RTMrTC;&S~-L=+VNQ#-fU^E!{q
zhD_L<U6jI9*LeP1m;WAQlQez(EGm2K1^+r!>VA3HLWI}4AI`PSICO!I!Ji1maUm+%
z2U(>nCYMacBMe`1%w0;)3H#`;dVDg<JXON!Pd~^%!sIgQW%|iKL#8Qfb(qp)59d^w
zv;@PD4RIFbYWCWL<kVY0dx$J~RYD22;^7%wUOZutYk}tayijkV)JvuDg9np~#n*Di
zj;78ext}T-QA6Z5RM!+!S1wv{2E^14oJKgt_lbqZT<~5zd*iSHma7xdStxha#rKp~
zmWaoTe`UC#GRVPfpW$Cd7OM9Y=hIsF;dokijx>mfK!;A2Zx9*DVE`?CO@rAbd(ZV1
za+}netlSGPjm<O6XT<Egaa`^;!-puQfAV)69+-l?NO}I=?tK4sB6e^pL?)d4V=BMn
z=FBe*(UuQe3By6t6JN;VI1I?cNI0c|^x@^%UJ*7K$lX6ZA7E)OdfAx>sG9^?%teqP
z!<wmRAO_!%?h83S|2ye{&wBfC`$jAY6~ZuHx67>PvKSYb*3HkCb-<0X%3Q)iju;=}
zqV&Vz=m~)>b^IpU8+OJjdRksup&~TjR1`AlvJJ-V<6<;30tD)~n$U+e)9oI2bnYj2
zsZgz|Tg;a@FVhvsR1{x|Eq*O37pV<yXX1_J{;7cn0wEo(zmo2VLuWPJ!nP$Qlt=BF
zT_g+Ys9hC_r<;=}fbmi{wy%c%b)6R-Spx2TX_eALjNhsc{3g;z`jCrI54yZ1e^)O^
zKRVN%i5j$hSlq{if$^(&HjjZi1ecTUV3zFBtZ*23G2>fkg92P6(kc<pFfvl__lvYY
zIp~v<yc<x**gs9Y#<k_vIon5kXIt-fKVj=C_V_Dhw>qYH9UI*fD@<DEb{(H0)Y0_@
z4B53y7i~0%aKddtDLpvTA0DotR{Xz#gR3ec8l;Jt!=)DvGkkGb_qqMnrZugyuCH#Q
zDpc^9@{D3O23Eqxg&SMwoQlz!XGI;ce7uIGNJy(cE_`HhjB4^^$jgtC*z)7-#1NEJ
zH+3*zl;wWp^=Y}e_?7<bH=HXGidBM>WYauakjs_sZuhT6#|7QX@yryle>rcYp-=_!
zwP?NmbG*|yEod!PER~L0nmq)av2UO%pR}&+Fd%L_*i5e4_qa8~1o^pWl7${<)z}VD
zwY0j3(kWnmDF~J>UZ#3rb11-iOg)d%-8EeoezPPqzj`aov2?YR>c!rC5J@A4PLt1X
z`%8-RR9Y`?W00bC0sl1ev3&(!o$+%e&SXv^`Vt`xls4?XXvR#ZeeoB^gpz&9p+1S6
z1HI;jP$(BNj5cu%SnezW_L#=Y<hGCZz5-RBE7(^NCFo*<g7UWMQOYptiQ7^q_BCKh
zB>~|ihrXR~uivI(Dxa!Dw9@<CzVFY`ARkXN;!4_aV4LohKPoTu&E_W>ou{9fgF=Ac
z{A6|DZve{;0IbCxV=&pg4u*5H4^D+oxIDgv1^$_aKjxD1?WbA;LY~(v{XN1AOjEuX
zW3i&Kx52N)uS1ORE8Q|wjE8++K59p3UHB$aT;3_H<ZnVqh=&JZz<hK6K3Z`|aHM`<
zQc0iBwbqW&Yv*k7Hc=~lmwiES9>4P4Ma9_Cepll~S<MiA%XX*Ob-Dc9<oO9nM2Tpp
z3`q21LRlO7C~w#|3v))0e#Hw1&j_oIVW%8pmhj^}%$b!G%m0BZg3~m(fcSLFedClN
zE)I=VuEMI0RJ=|Dhvs`b{lc_B`|gb{Pab8^y&fav*P5aTb-(4^N;C*e@EXq8+NqKX
z1l4_h-u15u(0}g~LkZ+d_qR4mWq9CJEAu;|qHy;BMI>m1G&tjW<b$96yA=0-f&>HX
zx)+nGaFC!>?3Qzh^@x#i5zwcB#aGaA3$O!}4Y~#kEwwCDfofU<SgUu05=#J~N%q~q
z9StDZ?+wz#SY;!kE-o(a!$j|lK?|MFZXQ7?T_ha1{qfzb%C%%BFsmZPvuegT6a&XV
zW-t^)LPj3)ebFfSpDVp<WpKf6INu2U1x2P+X&L-198P+5>mNJB6$Nh-vp)1{M#8{w
z5*CBZw{wZ!_!5eG=GAgb1rJ+xe5_)#n-5d|xJ=75&Ke*sQ1>YI*Wrbxg8llI{R=O?
zS8D!A@%rZ76IGl0_OgkBWlePtJ{LTXo_E?%IV<A-sg{+kPW%H}DL(V|iGK+u5iTs6
z5MUVs%AlY8bMtfNq7;KAl2Z}PGZljVZEf2-zY$>4##`j+iqf;XC#kD)yTd_wO3)ld
zs}F1Gm1>$AWQQzt4$I%hc#8QEj3AYwU1JZ3a(>*m`3tnvw`;!>9q5r@csnaAbaER0
zOK;S4F6m)?oh&h_>#6z^P3C$uRv>l71%J9?7$&oYZj6;_Z705&t9~<+V^WzgsgSi9
zh>ngPud*V+p%%*L4;VA@p_=nLOG8o3>udwqSUTR%;w0k<fFG~P{yqFqgu{;pOcNs4
z2mWb$Y4K3~U-2`uY1ckYTXoxv<f*uI_(5YdVHuv9#F<7`+sN%UclQsNyax=@nWNr?
zHwV-PRN}iBe^O6JBs--0?tqIzQXE$6rH<$0UN^L6a-;XU?gM4Tlqhj1rg-T>IOD_G
zx2EQ`G4ordcTC%S+mQ$Oh!BaQ;}_ahxJ1)wC9-@iWKXv)F$De8RsK|!z2YDVEGKIv
zxU%q?oV^_E<yU^CVnmOiP#jlT(a`cl$V|KBAuS%Mc(>><pL2|z4RV&2%fQYUm>&2q
zru-H&cAf~Sy({@m?nq;Q)hs>zdQ4qC{sE1l?fn-$-&I|RYEFJZD{XeOJFvEQQy+ZY
zy|c;<aM-BHU-=aafpadczzkb!J;NYpWUkSTu;u#j9@rH8@vBTEIYGYX#UXqV24Kqv
zZBJ<8bUq!X*YZxr!dsusmLrsul9C%DeP)IMTxv|uq2~T;ku;(aKvMg9V=$)>UBId&
z-l53r>J*rUu>%_*Gu;Q5y;Z;%grG4!#vIJx2x?1St(RNrrt5AD7ZBzur$+!2IuFlD
zTxuaPxY!Yx6SAJ(_}rcO3?j~2uJok|4z6vDmZ%rM2%^6R?7p|-<(XY;N>o4-hID*Y
ztL6rDra1xay}L3swsW0;x)oMs6>F#s_Z=NJYWBQn{FX30SfD1dWq@2@zQa<5GZ`lH
znRJ1h(kfML6WPgV3OH-LD%HmvyaBV%QxArwJ@rOcdyP_kvhOY=?1gscdOeCK1)tk5
zE;P)Pqs7m5XTpF_$2@p*l$NRhmjie*Ndb>Ebu((A;62ui<6^}*O~AutlazcU_~>F6
z>)Q6saauSIjY7yFu<l*}W5$Hp_ljE|9feCQ5Xy`cK~Ui5{S)kE^{_nEL<FDFT&lYa
z#G~JrS<up3A9bwp=Na%<ptytH9?)wxs38%tS5y$tdc8ToeWA`glF26Dd_;{T@nYws
zn<Z&Ft!!^_OhQC<;K6&9YCRezEgj6jv|Ch;QiFG8r6?C7b>5l|F3m2iYHxR|<yYK2
z0~bpsAJQ_Dk@I5Bk&pEDqkLecKvj(2)MI~;db7hsnQ{@|KOA?+x4GzM?R97H&|<%s
z;aAq=&jZ;+ri&WXugO=JsUHK0@|Q!-8QUf49Y(o-Jn`D-6F#t>icrE?w;cNt?Y3HQ
zfyC+~L<s`U1hL&3xmt2ww5Qqp7e5ORy1|e$K)_4z-rsjFFl-W7g{4P=NbTW~pEoLs
z3{zHgKk*}TeKq#U&*BIZ9B8+lTJs)ODAciMv-j)Ypk|#FxU<$1d?_^sS;=K||1zEa
ziIsi$L9uM@IZ2jQqm{doG<nR?JJ+Rm7{4`CDs&ci2!jphWPW;$Rv;sV(~~PotN0~{
z_E)ZK8Vc!5>bextiRG%&PT8y6+?JH<&jclvNMOIqSlhO$tEN%*_K<M7xh1E;pXblo
zphO{Y>y|=xnNIq#>FM8(mF5e@OZ1j{Xgrfanjf_r&LC&=`|w5o)dq3!#J->3Y-*8o
z;~}<x7_Vky^H0C)gDmNKuPgW4pb_%V{h<a#OYg$Hg2xRAvwt9ejw>gpQQ$bXEayAv
zz=25<DN=lxnBjll&R<pXOhAAftR9RLlVyb)?f{G(#iUIAL!k+PY!n4pD*w>#qIGTT
z9eMw~eiRd+YHclL?feH90P8glQ*p{p<diTKD#pw8v9fe8JO53<CgczzD;i;?S{4Fh
ziDKO<`P#HsQvMb|PvZp`5;TAS&pn;jf&@FB;-o5946^X32Ub!vaJiUYqV-bkJt1|`
zA3?Z@hqONzP+66W11Nu><5E8s=a%m<nAa)o1@h-W(D0z==Dsk0u3SuCb<zs0JP0Nb
zYXBy2LT-nWZ_~{)hHE<^mi7y6sQrdw==^U%el0Wx`$!0`Q%AAZEH*#$36BljBtTl1
zJEnqC`dyC=%m=chK$RNsx_t8VrF;+}Tsl_3C1$7}kH15PbYeoQQU7M1e}yLxpMNYp
zVpphej}3GK_PT;Ce*OWuOvH9w3lp(N8O;D&Rk)CxdOepU6wyQ?S^|617RtIX>y8Jt
zHPOPQGZT#*b=z*Zm@0A_E;BQAw=|Jn$3tJ`*xUY9Z}C31xOER#shCp<eDmAd){3Z^
zr!Zwf&XW5HQnDJa{0U)J`uxvhPWuI13oh)c*YUm<UoUx+TTqr?*|16ToSpD{uy#U%
z%O-vx#6iuf^m(5t=L8S{$#N&S^LCCc_I(-~<Pl-+z7nxri#^rh^_j-->y!5=Q`F&l
z&*u*LCc0biqY`Q!A(#3Ui~D^c54M2m^|MM9D8ZgIx^h7QfeHl$YVigtrOx^nxu^Tq
zhB9c7UaKy4^;aPLsK2tM4*#n^EvAL&>^2(z(+O2|Ugy5^tu46?UkkZG@UpA!E|nCM
z$<`7@@PrT^CP_8<&Z|PscSCOpke=1n>+3Ah&^xd3wbP8;LUPc}Vt!bGibs~OvTd01
z!<CCfK-A%$f2u2{cS2%&P$GWish75mrQpvCD^E$^wd0*p7J)NH{p%s_eNMzhPM-ki
z=L-X#mNQ()_2BO)tM3Kbmu0GU(b5^$SiyE5DxA<at9#{&d{m9c5|4}oxP5&OQnB_U
zM*PoBMX?&Pe=U&?cs}&MH=-HXQBa!iy}OKFbHa%_J~Pp>*inhP-f;uB4Fq;TQTnKu
z+}gEvnT10I$!}0}Zh(C&L8DlUh1<6%u1(OXQ$!qxq@tgh>m&An*yJw9FV&HsG6pb9
z&bSW{q&kl&ELE!5iXQ=)v8@VxU9-*VEtXI$fm0IV&Fpq}7AogN=K;$8G*CDL+86>z
zktfT?_Pf*k-KvH=^Y$)&Yk)+mG-)bTwZFc2UCV_A=?IV+n+L5IhA2oWe$!U(C>L_h
zBDu%6)rm?AhOsx%XWr|0<n=+hQeAp|%F1uj_Ku}W`%oxsr}M7|&E}fDax6rZX`GMG
z4@`j;g%iG89OAZ0)-TNH;XKyTTEOz*U{MM#t#eX_>o;;Fv}=P|jHNw3iFJHGF=Gcc
zV?nN~QR>#Bu6^(K90X+8JGR+I-9L!dvt!}QbUf&4;OF^V8uBvB`<fi@q>PK&px97N
z=S@Lahc6t^99}t#+`l_aI2q@C6RhHqBY%?HD^dLo#_#21Kf}769<}|-z+%wlr_P33
zo&d?{;~##UU8K~!5`Jy@JT_Q^Ys$}mKlzh}_8(D^qvR<B>BGh(`$A##dS4EIs+2#0
z_e?&JeONHE-Y&N8%p`K5w`0c6#t@?U(<;N_>@}F57+ZV%IIuX2chTw;4=Klr{;3Fc
z9?6|A;gJPj?=;4~e!<)mp`q^#czDKp5^NSH-!%l7TZ2uxkzx0R8kWaTw#J@Xk3adH
z!mG@Aci%^e=}RQAzZjNV{e<q(gH7t{-Git?4xxgK1OO0v5orGJYqmhpYPERy<-QgH
zB7~5}fY{F2nR=4k3;n6yCtu?aw(-!%QuqT-97vUn=|t~n*q&0`%QVYupTefeMT*^h
z0$=T7qOkDoLgqfNYo_sBjiEgQ@Uek>z{=wb<m;Qe<u-~sjHh}*Up9>fYbHEV-gtNB
zG!Ys3cf$(OGuB}T_(+hs*<h$OgZ_6pzDk0DE%%hcRC2i|R?U{>j6!}V2keie4L<Q~
zViA<?2CQ7?5+OrI9OsgkcN?3_Hh3n<3&oeRprxE^KC&C57)63uh?!*iPV$XDI0IJo
z<PHWd+mcpux$aq)byQ-zP);Ia7mF4QLS?mFC~$1;2dj1j*JTipntrFVs58DUF)=lL
z3e<KRXKYCNBUn_`3OJQuno}*9qILsJ6e!@v<W7N%^Fh#IP)FZ`g|Ufvp)PaUx$Mmv
zIGlt92luB<1HX`dAPWXE55r2EYU_zdQI|l0<$|Pz@sb~uWnhs+$ppl~@^HBu?pN^y
znAt$N#GRK{AtFs%;tKWRss(m?v%;8JqYkVi$WQjB>jJ>qB=oJ(0-wB-grg+@HyD`Q
zsupqfDXFN~>uL#Tr8=08ZJ#q1WKto+OjI9r{RTDz{8aZdU79A`-(+aKG*uqWvS4&K
z%t{I&at{jH{1g-vrlG9)flpLjlW!+%qWm>7Ij>D|J)iO%(}oJ8Lo?i0;Mgl5H2O-Q
zwf<Om4fsb_`e<DMVh5S)16gI4?-m#uV#c^yujeO#OPF(49S1X!mVzVb->0s==)3m@
z405<jq3VvV)w<$JHMp=lHvTUVqb7}^k%^}-AqbF_e%9hwsYW+~I7+wEPSV5Y*3S+{
z3)hBk6Yl56M;4*I4pqupNKKC}cx&%*0$P<*8Dozb<R2P9Jby-&A-t$GlhPiw#=yDf
zf+Q6``j*J`IhOSm(*E|A<7iJ)TtwORC4Lh}KID3`9MsX%p+@$g^POk&83Cpl6e;0{
zw&|nA^-Q#?UiuUT4FbU0Z~7@3)Q!k|ApcIBa+CA<O$AtYV3|Y^6e?D8brggHYQR?J
ziXWBkDxh<mUVdLFWh_E3w-_}0@``7L$p0E}`9^_u-{VKI3-a~Qqume;_N(fw*pgXe
z_&~wki)KhA-SF-qu>f81sxiDXU?;aTjf(Z&)JnXCj-y6bESKjN)5IlQ<~o|@IHB;e
z4qO}RrVsfjnknoXx{u+2kE2)0*Vrc_^5^AX5dthhFFM$8QlJBBR#`D(MH08gk3w7=
z0NrN5PpXSrK7L#$%*06Rd_~%xuAMdZad!>;k{2p|levds<QERNOuakoS|o3Hap}rK
zh-Y0{UqIfMt{?se1=waz0lWaqZA)W)puoA0OUb9=pPi3)egVoH`T6{}a<Z~fVyUeG
z=xom){`^Ae4TwGAU|&@cQqC}LT@k_X%ZMTQ72tiDm?6Fy6z0PW{N}KjYWBgM@XxSl
zqC-DeIbkWwPY2XhWFU_pzV?beOhrh%I9gK>D<z0iu2Z%%-<d3h$;&%xx<5@bo(9JB
zSgq6UTUuA1!16p_Ehp5{kwHwu@Cy!=HrQv;lfZ7}OwO@$`d5|Xt4}c!O6|*nB&0sK
zoX-c4Aqpz0o7pW2&7cV1hfR%<+t0qLhe{m!wZ6balL8R^<3#AX+$YzVjcVV*p_cY1
z=8f!sxi^CXPXcCvQ+{5Amk;fjjld&N|A|0S{pXH5E;e{G2_X4{!ik5s;Iw1}*jSqb
zmhqrCD}-4dMk2ktC6C7pE@ZU53_M|RpPI4f=+{s^CMH;P=PMC3wdjsP1?D&s16f$;
zTlGT~G}SBmXFoCy;?yPK7)SCAb>JV2ldcM-n&%hUd<cf1$O@jR`|epBsS?NGR<%B-
z+!!`I>Wzp(6n^(dR*TVx2p;R+O?~{S%{&p<I%B`0G@^_QXvBGlq1--%I`Rq9Z$G`P
zSE+pUo6OZ&Q--rlvh*Q+>G0t#<c;SF&IX-T)rhKty_SPIFOm@3$vZQp7ay=GoBwnW
zlli6Qj{qJ%0PFe`L|+ckh)dpU!Vu$31WGV8^PV_4u8c}W8Z?L$7X#`RxGMGI;Fc&$
z{s+*e4M5w21o_%5#{E=0Y^!4flsCW22GfaUza#jy_N42v2a8tSe@bzM3F`uja1-<B
zYbCipGQMi3xs!6BJ(X0YVtR;H;nDJ9+V4x@kGF_lKwFyr&CXT%w2-xXa`Mfba5$5K
zZY}EUC*I`1C^g&q-^wT$r)%fI#mc>DqzT%9@(?~gW0VwOrTZ>Fp@r@;R=PR$Zey}o
zTBXmSbS%PGuwM4MiZUAnJJl!ehhMv_eaBxLuI^d+m8h_vhK<0mRtb!d>G`x*+|(I;
zM=ay#m#Kx_YsIorj)3<SlCZ}q_wpl!7B4tJTc2xbS==*}UQ3!Z0W-G>?Z*VE@|>0G
z-6_1Bpo%ZR+*WG>2E?*YxK{ES`xVZ>5>d?h4_j}T%kpV@#=ycx`SNYByiNa-F~|aG
z({c580bAu@*apP;(QsPe3sn61FUf8MkLV1#t5QmJO0v+6Ly=&>F{I5O^<K8;p(MY{
zE+yPVz@<)Wa|nZ!!IsnI+aq3k2hY{4GJuyZ$yKodigP-Q*9BfGd1jLH+Uxz{0QhAB
z5As6vALl*S>JHQ(&Sg)saJ@1!%RYvXQ(2p`InnllBcCIYX#Ek=4Hlw=0Z-g|;ioPc
zdgQ5~@_lFrKk~g>Hq|+5r42{-t(~dgl~Ap@>xS`_X=#AdEfv4${%01Nje|Ok2niP$
zahHGUg9>qcRT#=DAtEIiq3_IcP3GEXP{r8ev=RQeLxDUfwd7U1sN3g^ayt}(YsMaE
z<(B5*cV7#m;4*@Y$Muot?ECk{oM!F|aqT~L7nT(r6!kpEJN~*2(s#Bf-cT4y^@}|H
zp^zC6Yh<G7q@q~VbGdI!F_B?y1@R8V@S=S(7V6ksDAo}nc55k?Kqn^aPWy<?&ANzK
zHW+O&%U96CH?{@&%7W>pMpiX>tF;aq%7K8GhAoY}GKsX|faoV4qfVT)hp<+I=mSIO
zvmK#c`Em(8687`2LU5@6);lu_t7C^kSfDF2ayP*UF!>L=>F5cKsQNQZ-z+nxylHLC
z*XqY`5Gse$9;C9W>+9YjnaPRpV5(~Q&Iux{J-33j&Y8*=v5gS8?i%=x5=3KTM9AJA
z;?tzX(G!LJxEJg2jOPJ}9Hts>ihsSiSn!tO(d%I(NR>=wo5MTwMOeuneux+hSN`KX
z?1t;nlkp)bY`&8~$8Sse5^QH;U`FISZ<ol>o_<aIlw!mvdkpCrpeZeMf2+3X&KmJ)
ztHTa9oatb@2}3BjrKE{IDL|e<01a@TMUwVi%Do46yRY`IGb(uf*7CmGwW4m(H4d1e
zWGmmk?t2i?F35Cno|iZi5s`2ugbcI!m0~Gi?6Rgo*u!0kcgY@NyvD9qEo+rK7xjSC
z5N^qhVO9Qq>N(i_h`(&Ssf{uH<*6x}q1y`aP>U3>9=Z>-EB1bPnAwc9S-{Mz?Pxhv
zS;h~rP!;&AfyL1~q&QrlRh5jO<kw(&<U-|AEc-P5nK_nxJnPD+KgK>&-ZBVi4aZyn
z?y>A4At7;h?K92S0T=LEkF(NPM-wm~gj-Sn=&y}o23@pRebR2>6fA?U`1ts>Zb!N5
zA1a|-h<)Jlem|4C8PWzSC?_^t>u}H$3<Vy%6IVU)Y$RD<zoy~XKe^}8zjv1Ib6baU
zJM}#inMyLPUHF@r=x(5&Ma6s_vQSt!v`VPJSFQ2WeKI6}58)O4z3vMKqhensV%p#@
zkV@+u$q<4;$D3Tml>RcKLqrIqajXG(W>5(9lG71)Ep5HcY~D$Oi`ARo`5yG<pn7mF
z1-p5j(O<=oI`~oS;v$FZORsgm0x~Qj=@NVOC=nlVk7u=lyF(7H90S$#^3_{`QJ=pA
z5>63Nge6V!+Mk`P>Q8=u@cK*Td>t;@Lp}p{4;oN)^b{m&(IGzB+}~z9P>jU^g-BSo
zlLM6xih6iSl$9tp>h;qy6BCnc5)?@VSJ&F(fW|o@G{}g=3g&#>;GHiy4HCaRlV7K_
z{BPm;pLd8UpEd$ume3<VeyA0fpFZ)h@<52ked`O9@tdlxZLDDg^3RAEuCE*`E2~1?
z5GZw`xaTHmkpJSCc+LKyk6g5*zenRvN9JoLRdk%-CLI?vFe;$qL}DjFa3E=x83XqR
z>CUEq@2jqD#)2y?_eCaIqBGKwWl_z^S<+z7TkH>Lh&o|mX{C~4w>wMF(lP|34R!U`
z^ZYzgv14t27;c5?z6KCs3N=TlBf&OhQ*H{1Bf&U{+Ba0ZjB9IKFDLA$dT8_u{(XKg
z&=PV{i3VK<wyxS7K9u=Uj})Pi*oN+3ngUi<M^4Pjq_@z;4)#c=u9(C1LHdky=BvHA
z(DR?^fIt<jM;<sv5fJ6D<L4{3=Kc5}SnvsQg$E<c>E0f^KVK#0n!*gI0otWBl5O2a
z_+X_wet*e)wtwd?U+~bO7tfsps}mp_wqjt&tr0LN`;J#D(I<0C+GUdHcY-k;u|6fW
zHd2G2hnVCpofI)GhAmK&obUg@-B}`tlIg3-D(?kGSdT~yDVE=E1sku?c%Qd+H1+>R
z8u7Js_(acSs9f9p<qJs-zPd3ms(cGp0D-07jDV_Nk5SKb_9kVk_I>uB!H4HjxKnv0
zQz*GM`zH0aGV};+tzf!omc5+7qOvPD>n4a}-`dv;k$a69@pdpfMEDI#H87gQtF(YQ
zZw@@(e;Fm%I$_b5`n1WYfIZw>hM~rMQrUV17%z1KEtfqQ*k84LL50u=xzhR>JRj=;
z|EOTo2&hP01<sR}odCu)g4{jStz~K7=*d7J6dCtQ%v%gA{0=leOaMm9SStiA@$q!B
zjY?oAX##w;G49=Sv?|<vQ3GI0;+(-wZ{%<GUS<FY$wq3i0l;@v05_;9=keLEKYch`
zcuQAJzc?g>uVxW29|c;SGAxy^(dr;&3&Xdoa}5=y#++9z-L2|K5ZHAiLf4v?mF4s&
zM{nuetd5=cgb<<51r|a(`yZBxw?Ye_71JKPJ7w%~-U*Mo-y{9-RT!#0?%q=OiB)jD
zCn9cNYFpaZAiF({?4i`SRWKB@BG2aD-&Qzi`e4jijl2GCb5f;#REhHzk!cn7BrQmO
z?0kiNtHSoP#~y34E-iwy>zg2p%DNvrPJnATfv&#%r_B~XIP@hZM$pchs63a%&Mebn
zbO~qwCEm)GV=Sf9iP-Cs34*uo2g=4LR|tqdET!I+29Y4Ru_D!dNEHVQRlO~7rulcK
zJ$C%W(ptq_U)c*2vO`6!{H)LsEz3}a)XN-#-stZ6O+X8GFHP0%PNew-&DGf)7-yej
zjXzawiClH`3r&m7G;%0r%-cLc)}PA#lvLmFDKO~~hL3@$7lx6M*IYj6xZ1Y+S^S}x
zBCuuTBQywkJb(Hb8TRUt{`I}5KF8&PV`y*K{DM1&@MO`+3y6MIwcs~p1$0f@NXR$)
zmAjkb;U%}4x3@P~?7OPcM`99#GAXd-kR6GO*p6ze)lIT&`*Zu%9cx4$mKnhlPdi2v
z7oUw$A<)7=h>=3aE}EsujQ2nif7qw!<ZIy*8Huq^i5EkS&rU)gj5A2UJR0~`doMi}
zdUvj_Anum@K*6qp5rbnILENo~?up~Wju!0Tb>+zmOVBST_3f7z)-R*5qFao-p)sju
z;NPs%YoquqAv@uWkV-`Z0J+x0@Bm~cO<jg^Xpj|5k%GXWc48(Lxa5QZ-=s_+(0uU>
z?}?%vl3zi@$LaAmeW@%bC2LV^&!Rd^qQ!#1-s}3$OL9S4(RFrIOIUc4lV4?UJ$?Yp
z`W4{0LaJ_quJ^><nXb3jfWubD^b0gUvh_cp=oQ=n_ZV#3#!T;@1S1Db^IvcnX))@+
z|EP30!1J>2_byRgz|#{cj+~H?kU>POhz@W)BYuLBQp|MgdnuxqV|pn<5n!nXB`{QD
z@0}rp9s#G4M@hfXtO)fL`xY}Zae|IF6`x`I^nq)+PnJRazg<h?z1b!X6)s_}`xMXb
z<I;#Y3D5?vt`_{znf_=4&uuLKAGys>xyM3gRl;XeNeQ$5bF&xWRaZC^&^Dc>Jihre
zK@IA-p7p{E>?;0{E<)6l6q?H4kl}dpci!6m-S2xhia#RX{d95)>S$N*l+U#xXvVJ7
z>6ZM5R0rrsEsJfF#Iml3k}hO90!~uhF0&HGvWIh9;8Tld2SahAu)Ud|KgLJbglF_)
zw0`P-dt4Y!8R_H}gWIAQ93s73{O;Uk0;V()?t2zm!FSGEeMT~$cE9XJ3nILYw`GAj
z_5|L&%r!f*p?3yAbPT-52<9m{V$q*WcI({ns~M5Hypa;Bd7qCpSojr78Bq05dSo`A
z!H{14QWB3gUdt47ebY2i=dIL#RMp}q6o2fumdoYKI+MDVzgBKdQPFV*b)<@4b2&#h
z)?vD-`{lQ_-T*D6h8{5Vy;qlbDO`mXx>sNh{nl52?EYq6=T2aO{BNqFvL>k2PE)Tt
zeJ!-cC!WXtlMtlt_NrPJ&&QI?<oW5*&t*+<eKCjOaM5}w{nEGqX2~Q8;33r>7oa~V
zME+tQeB&Xoc&cb`1rfRDS?D+DO+ceMpD<na!xHC(m^Nd8{r5I<FOmeSt0+5X#X={a
z+yEpb$M{6weQ2>1XlTgBC`F+A7UVM)wv?^5)U27uD!!)~o=UvW$fk8B0dJM#1|)Rm
zz4ToKZr~qpZ?2}|X=!6_!}X1&u=r1%*ZUCxN^>flt+LW8NC(r~FeiGAwd;!`xPEvY
z)w*kuWsEmd&W_;+*3UFY%ttq%T?ZfRk2HhQhOLz&;v?-Pa9zy=4!EDzGRDNP^GUkd
zeUKsDRY~80AzYC`f>N4*TwnLDe%b$wbjRdxCUKO37?lbb-7gR!RYH$Wcc$#O$DcB}
zLJr#qgQ$PDN`wlF&wp=mLvl?xE6ejU?8~i!fy5QYhy2|{;6k1~BC$-;e=Rx&tDlR6
ziur|ar72h$E^O$&!LX6OE-2aJ-RIK<0rK?qay%CtH3yP<(BJjUtu1Q0h_1^a1N6q{
zU&(7;13JGgoK#v9x8kh>XZG(aq7D}Ra0Glh)%BIJRDXxN-*VO)w-E(mEQks6yS;z4
z$->h*Yt0Fgeybm$YTUZ=wS6<IJX?D|U-v(>i<mLvs>M~mrg1I@U%Yi|EA4tmWud8R
zdL)S$Y!*^iXT_E{6bRrjMV9He^*yASA6U8fsr1+N33;YTOD0Bd(zTMs_Duudeg~=k
z+=?eO0Dr#jx5wUB$rKzVek<pQPY<%yz|aVE`J-i~gZx3e?aEm9$$l|4Q6<!INlvUv
zq!dSHIl{w@#Lkkz&R)>&q$M`oApsF<1W}$F74i-t(A+7F(64sc(962w&@+B@Q~Rj<
z{e1r!H1h%)pdhxv>x;F6=wPEaOIF(`Sm%7eD0ajz*10^Sn2p=*b#Z__zfL#*+j4&T
zjqj@AH-44d_p}!L@1ZG_g)Qt`LU*BQzFN5m5uCj;28iB!QNt~-qJQ*6d?=$rc%wOB
zb7x(d{UH|?ufKOGZtN|r6*UHGH#X}Q9OlaytIt12E<5q(&&!}K6tRafhyK&u3N<KL
z!zlhg(yAV<U#v$$pApYbqt^(Gusq%0RvuWa`}T(u%?Kw~!9E&Ly1oiZdtZrJ#*6lF
zWfl?1pNQ%0U!&MZLFz6FgW5Y)xX&D1@cd5QjdRJ*^T(8o-xh?{w8~fFz}~Tlyyg*+
zew$M&7WD&{Gz;;%48n8!lwdDax@}?vFszoJGC2hdJ<(5_o(Z_u><iUj$wo#*8d{5P
z?n3K#VTkt<Gs7F$`k{3snsxC4g~YwpydZ0fU?6xipT7Xq-i5}SC0)CH1-#V-AIu$G
zEOhgNzm|NZ1~wy7z<VV15naDMn2~}l5@W&>RP6Bx?>FH#u(aq4;6gSG3D=y&AdztB
zLxHheIGD7{)8i3c_43a7ifQIJz}w>748bE6c&-BPC%Mgf%#NWH6(6OfrL%zfB5jR@
zUgjMm9-kIRP|L)x`Xk=?i!jb+C~$GezMBB5-O{{8P(VFeThGsDE?~X6Ln)2*hM+)L
zY$?-X-aSrX%htiz1SR46OapK8-L1#fvv-`2^{OAk*SS64rz9Z>0&PzP1d@0H2Kjja
zbK!2-$7U<AJ9R*=t<W&`ef5&bsG#HEFkW_FNKFm#!tr6aL5sIn|49_5(O?6uJk0DL
zof%}!-w#|e^+3JGiu?b0T`OZ6%E|Fq2<o!fCo`;TL@;oaMaH%LBlCRx!mP9=8&v!~
z)Ad_IfwX05wm>#Ill$uM`F=4<XxD5*zzw1FQm5fsK<vqbufrsDlD2>uA!Rp7*f-09
zi92+8^S-b7!CY-VIv=;&<5Fu!5d?kCnx2z$j1Ijj2{Dz!t5nd!4|v}sSht)2?x^t(
zt^MgB3A`AQ052@e0cAXXE7n^B4lbSHlt&1gWiRIZVt&;zTw=n<!C!xo*uUP9&)f=u
zUu4w5&qYb9;N9XXbgRP#d~Bo?M995i1YLHd<%qY-)s5Z=c9wO{Vwkkv?-^CfBe;^}
zm6DjNd;Y-%_zn6hqsOmo0Kyb)ykESHmU<)c_=|w`ONTt?$8@F5oSsO-lo`8NB`87`
zJ`Y?yc%n*N58oaPK&w@u5h0sL40bguxMFDWG!v2;8=Xc-BccdM0kNGAGg5r_iSHb#
zd3oBu-!f;M(rVf9Z?cw~quAM%){in6q06#$k*e+_dC2j5*+H}oFD$|A(TnGq5a-|Q
zIdpiVg=>QrMM#ewaK^C#gV4%93Gv|u7@OhbD_6yw?-Q7eo<-{?&9D?^d=*Sfep`h8
zM1*}_$JQ`Qfd^Ztg%AIs`D8H8QkA>AfGaDwp=h+(FPKpJ9lxh^B;$p*TV3J1I3cdv
zp&PM@Us%z82R~aW3-PrRxZ#2>sSk&xycbjMP=1g2=19~aObUZlWhzH+;6n%&Xp`lu
zy(^+_5X>ua%}W^RF>qQb_nQ@1>1(RKbQdc!Y|%~M1H=f4ARH+`DcoNS<IK{pbNHFc
zf1zPb??o8sy59dV3VCODrh%c>pY}d0AiW`rX8@ZZG0?nL1zXM4cWdN1!(tWQ?@V@|
z6}7iZQi^)(BY76wg3`+FY}eo&HMcidsC2MdRF(yHvA!rTi(#7ui<9AUQWXB{3>!(d
zIuFjO62PVbP6tdxENYUKMWC4?OFr4N+#SmhBvM6(ZczZq)Ay`ml<#M&qXC_Bsbt$6
z`S?JGUvs_|J-9vMT}e?KyL>oobv%=)q@q{bT#<lik9oE^au2kCz6Tuv?Q!(2piBg%
z^x&kS7S5{)hm=zcps~^_GwcAan5HfxjK&_IG7P<7nB!W0BEUa0G~48UVo53Fsxnyw
zOlui-@XMM`$+3u-X_vj_uFfFB9%2#6-|wQss@~5%JUanXxvya6t_Bgh+OEXFx!rSN
zQ^ykIbwA&K2>b#$L4OntETD<%%u}nid;M-MN8+*W_qPj9{l0*=uyJelVEfF>=zCva
zAWEu;m+qi9SYH}*8gWKlKi4n3xB@mJ1Ii~pIeB4W{opM<0bR~k8V*hrjS_8(QOL34
z2eqvmO{W|AlQ?71f!pP}Ejhb>LazZoZm`fa>XgqtSHSHs`=&mZXYeMj)YhFd2*Pu6
zs4GXJyvjrcSZd|{)K16isvA1}&e}nFr=k~-obPvl(Q@@<XU(Z(VpSX_l#MUEHAwyP
zEIT-OnIul?s8ef;g(~5ZkCLW9aNwxvu`#3N*Em7<$^Kh~r`a#gY<sNeSAeA<YKh!;
zzFCqLez`H+6$S&r<3{yzV{sHTv`}aE5%v1W4`6C45Zv*!<}LO2SM|@=XC+}|5oc>a
zO(%QCm>l_&a(X}nxaOVmu+D77t#Z<Z{`6ica}O&v`W{h+0D}mTwi9}V2+j3#aeLY3
zSv_~zoqhpZsjfZOD`gXwVBYcJwViDTUAPMSCFyrYnIF_Lt-??F21A>F1$#&0_nMX_
zy2I6XC+#s>f0{hL!AS2$z-Hm9$^2(;AGLjB^db2MCWzOCIef<-4qJ2R$ko#R!3H0R
zPc@3~tZMIAT=t*fYkJ$fZl~5nlKWfy^e{ek(v`xs^l0V&a<P8zyWb7bK@}x1k58-A
zsT@?B-aP0!@^Ma<BwkmHr;!wXdgymtgihzfbMvp0>sxv!#ulwslQn*o{LMk-aI-{l
z?NiBkV(EQL{EX=F_Au;ZW~#z!9sN@&Vxj&?EKD_{^_Nm{Qgi?bsq35h3A;gi<fj1o
zJs4t<vdUqc=Vu74Y)2Gid?cPFpEKZ^AuVbFG#JLr;;{FLh!gUBk#^qq_J4%+EEOWc
z>?8epCH#cdJBAA`Mf%?&;U#+_VScW?5Sc@M%ZGyjTNHp0?7ys<<WP19aef@+gOKn+
zE!`3H0Gd?Af_rcyzV9jq_%#iUTtx-U*hl(?2l#r;3Do>AqP{Y!%B^dglu}w6L_q0~
zj!k!Wr-F2Mr+`w@NP~2Dry$)e4bt6R-{PF}eD4?x{=tU5?={z)*SzB7>PI2;yqJLN
zZq_Uvsk)nq6~h74W7)SN@>-4B8?on7N{$La`LC4zLxW{O`iTNT=@_n^g~KiVe*9t$
zBf}XH?O6ijx6QV=U|;chg-g;-AD(aGljpN`R^tL1v0r`hrm1X4dJ1`<MB?{B($_Dc
z^D<X|3xa+H?!K+Q6-VsX<h${ulXqg+6UTmeIdyNcGAsS<*CH}?`{A$<v)#KlN&<bN
zF1?X&F|kRJy9)i(0C^g&<;!7FAStrF{kuAY`GHz@hj;mrq`F$_pjKQI&hyW}r0oQH
zxUVJ%tb=&7ol8lFYmj6<TB6M%dp$>3>MDh94`?nQfRoslz`%mx)ah$5Tg+#o`MW+|
zuc1+{%vXHytAhK=*@!l=#a_1p%Eaz&?D*?@4A<PfWsLTBLxcY0lb2wFLyBPS{w)=`
zfssMG_9X}vP=*UXUomH68hxQ3o#pYJgm*u4dVsbiCGyRq!GM*2C@(Z*f^QViUwIVr
zHY7@3KyYc*Wc<qjQf7h>v$01;jL4LCQxpf*w;wO}XMeC+#1mIeO$0+NfW;4Jt*0+a
zXV4B>SPwyzfMZW?P4jbr1XyKNZ{@yNNt~7+UUI~Pg0>>d0o2<Jk24TkpPAoM20c>n
z@JNHnQo~(<;abbvGukQ~T>BNrKb`-21$J=@zvmGG{;V{)nPs2=H(36Dv1#f2Gy&ei
zjO3wY4rzqz(YGA#fxb-KVHx1vv_AOdzLx!aONz8=j#mdMjQy<mFRG(-k}H;LZpuBd
z`^^#7whWHg?(*|BY^0pX1fAuFi#!C#{Nacredu678p1Oq;THl8y&L?YF{>rd&t$F)
zYy<p(otp##`!^h)acxIgvI{V&sP*--6dOw4qDn6>UzEPqCInXlb>g0TJB!y5i0lQs
zVp_-LoTHfPIta*TAT#TC8jpNl|7RMUDg852K}D(&r%_)^wc8G+YClgjZ(UTiVP2a(
zKBIbT>$g&y772C5ovkB54eN=RL=P+J6VU@_Sp$*?aC>o|Y^p^Q8Cp__atekfpwj-O
z=n>v|oz6O^DijKH!1a0WeOj4iJEh){0<8`I(6yA0VS^j&9c7rxmzB8Em`YvuEZ4nD
z#Ft1FKE)dymT<1QsI)#1m_U8ALpbUle(x{6vlF(~U1jbGr4MJGP6Qdx1K2?HP{B}Y
z7Q)FEg6@}FfOf}0jhy%F0UD(5V;MirZ2)y@EaPpkvX2J&vugR>i?ncfQ8ZUIdvU%h
z0*&*e3cI9Xs6t0r7$UhaRoI{c6SkGah4sxDECi)xJ@{DDU7+I?_tRg31AK1n&Y}rF
z8eZRa(yd$bGn-?x=sVlU&Vrribys2j_p8lM8N~sia{-|O8k<HgX7(0*OoSRar=+Q6
zG^4k;q@kT}Z2AH{6qRS3ZarjP`yC*(=f0IJUHzG3ufCG=ZJ4QvfP!Hzk8#p)Wfm7G
zo&M9VqbmT=9$sX;;3pD<@Jp(daNyK6US?b)j3CXZ-#Ze!VB@#WGa?fx5}^Vm(HgDD
zp1B*>Z~#yQ+*S;o{6n9WHV@QZxTkL7r11I{{p41?1uHx?7dRoQIzivHw+6;MHCmC^
z+b`(UuSl7c1>rWlmU1Y;I`#MGV3QB2<4Vs0J?SwyUJF$i4Ui2Gu{rb^LC~B|^Ke00
z70Hp}Dvd^!^)dc7)+jr{r%%@5c;%On@3(PQB<{C(zAG$I%l(tSCiZ*eOeqGq$we9Y
zbnBNKU>_Ar^_C~&ybX0y=cHHmJSRa9lek+?zOO}n(oXa6Wyuy<={kihRz(B6h2ZNT
z+|A4vkBE>ML%J3jInt1lA5LvKm!}WE$9`!bIosj7s`=m6kbWvrpw(rpZBd}EWuFlj
zM;9;h;X0RhswowH{zMOuPN(M>i3A;PQcOwK=YYr98{{|Vq(=c6uj~UTG{#KdQkuY2
zn0bjIa}i^$$PMR#qyiW7!VSKCNp)kFJF)|6@9dVbi+P;l+!k<an{7OrQ@Ct6DH9$O
z5fai8>j{jAh@~4^&Y{ad07GLkUr|_H%f;8EF=!`I8hcYE=_PcUzUK)%Sli|Bs266u
zfW&xg0A)<h6FxnlBbJ^+jMi6(vkYCETE`!>4-}FCzdaK;l1HP|mP9)2M!%C*sI}Bw
zEocPmGSVfWrCroMLdR->AI~JMX6HtiEHi~30UwSS03vftu*d^jQ4~lfsw8b~%Rec*
zR9{g5Oq_#woB~;whJLf<YAFWQu0MwId#Q}Pw9Vo-{CuznsMM&o8T*K0oW8oPl_|`(
zZG#S?+h)T2o7x?uwi<Dcn=NJ-&i7}P?L#?~wmrH*Pn!cf;(5~47FA=bJMxUoY*nwP
z{llm~MtDYMSh{kz_evCNwe{h|vmKD90(?dcmVab^Z9FcsM(nF-P<s0KwKs9Tz*=V#
zmMD!GI)S6iVP`EdmtL{-4gs+*a_lkj{}}kP-z8tQyn@X9$FXAVq2wr)9By#=9!ktv
zkc;mG1_X<1WlY!0nrkU@jgopd1IEiRBHu5R@~maw{TW+API>tp?XB}(xI_?l5C3X;
zJx>Uq7N;|pv(VxMi$Lkc;HV+g-65*@kCt@zKG{J?hBaa&F~TRFVR{iE9|m@$0M6BU
zEBIxm%!&QGEdIalZrmh0X?`Z(d$(uh07R9WM{>MiOQ(>}5c%OvBp<GUEF3%9gfoF!
zl3#Q{h2lzL#<K94-|HtVZSsa|Oo^1))eR(RNVx}dxm~%&8x`q-H@7Y|H~jN2P5CG#
z8MWO!u{6fuIa`$67Oe05n&V>lc^Y2!Dv#kCv{qyNP}gqos#y}~JOOI2Hx2S7P&jRe
zTFdERtsZ3|`egi~BhOh%z;k`1z09iXtaHX3XooE@R&V~3bmmh5=Kf(`5Y90jGBM2%
z8Nfe~uHj-o%Kx7n=R?ic>X(z;$Y2TOe7n6-UUz}ge?cVF-yE26UxN;~S#BolU-156
zpz!J^HR>&xVK13t0{+!U`$IJitWO4B_9#FV`pe#o=kJp)!d{lec<h3c(53TzN8yz?
zaQ!HaSadY)^uzzU$Reqz<udb*fbp46A;oj?XOZjRH+N033X(yIY7q|Da2X*Qk>l#Y
zI1@_)LzQfA&sW`~<-SA@y*aaVKMb`gTdHLkeorRMp*+?$f)Ng654>)2If~bFXV>0_
zH}Hp=a#XYLV%c4>N}dP4a=Ht?)Waw^bdgXlB|?$<RmMXU2nICSv)3nUg6$<rQ8Rhm
zhmBWBQySH&{4gV6NzM?-T%wL9qBbtl7niK3l&7J5@&X_d)NQBzw9&$!z)dV7OgTkf
z#}88tGyvoeb!&RtNMQdUbTpn5y1q|)q;K|9s@-9xRQKKJVF$dG_$#&<C;3AkM~zNk
zNpHBV5DJnultE|4JA_RT({9pF+w&K#oyD(x-BWsa!Tn@nLTslz6y$5D+xxcqJhe`y
zq?`kB*7oG<2IT&u*KqwUkBw(X-;Mp0<Tn%e7xXXwW?$a4cw8~rt_bKfyN>CJvg&D8
z8bR|qyr$PuF)unFj+(u49XuEPKzTf!x`mgF2*-|bEjd%K_8~bP&SwnLH@mbb6ucBE
z{eZ&QoG~AVc?@wm)OqIS`ZEmSz$havWZ%`dy`Dqc84Uv<HIMA#N#d^)*qkHU^M)~D
zTYyakfwo6dx2LBE@QkuTKR|anZcBd+n2^g93ZzIa;MA)FE0Wl>RbrHlz<W}3nDx8!
zoh&>u|BnSnAP%{KxC8tOsVk$F0EtN`9rxDlsiFid6>B4<YLf}JvD%VT5Bx*ZbT$jM
zJCy}Lft@933!N7Cny4J==^U?+Qp1o2H8(ByrWS!`#-kkO8eo)z%&@8*i{BsKlik&|
zZ8Q=b%UtzE5c4E?U%vKYKQ<b6&~GgPi;HLRGx0w($TT%ldf1Wd&j;U@N!8<a<*ZBK
zG{kfv!CA|4-QiwMD(@&ww#HK%n>edO+6UP9pOD?mC!Giq=5*(8C?jO<g#0Pvq?YC`
z6k2=4CDPk6{%$S3NbBp<vLA8c{3>nKQrm|EY%`=>_BUa^&hvo=o~@V~338|9dGK7i
z9QM&s<w%O$(f#Hw@cz$4K)u;G3}$2!sR4rPe!~XWEAw`ZWTB@(^vuq=@2TLBaJD~I
zJaFj^xkVrI@Ez*Iw5{zMXSAa{6z6$vlV{a}JU`Zt-I0B?ceyX5GsQsH7da{x1gNiH
zwR}mAEsHrW=A<Hq#Qe6dO%3L$41zO~Q!wF(oZzopU~R3dD&qQpF0^L)et8}myEqOa
zCPY{6MOS{HJY%=vTU}yZk&t@l>WY7TV72mBztoF09Ytk7<-XlOh1SrVV#=G~gQ|bu
zqGnMv;nhJ*`@Z>KhQG$uXU2qJ8C}3|^^fuyigxAWGyKFW*a8nh*eLjN+&w{opk5R0
zdI$G&=1TDIVs=7fIJWAFCV~hna}edOJ7Mp+^OLcke8hK)OSmuD;<kMp=sNWz{`TrZ
z0wQXpq!OG%#otTG+zov+7>=;5*Q`Z#>zsR_F+W9KufH!wV<0<uY%AB{|N4ctY2UVL
zTxS7308kHNP(5wEZvGHF;Z^uVFnf&#=m@_^TO2x`oO<4BszWCF%@O$yMl)gAtrtx7
znH##tP|nW(HcEiKVw5jDSGC}wI2_A0=2t=AE_wck?A-`0VC2~+7~#!GqrtR;e+$w=
z{xrFioh#kx`%vHW;zqXXjtii;tGR2Yu#!<NU^M6AZuPMx>19)HQNnN+(EAOTN0LIn
zfK#I3dNVfg00K(boeyQay+6R##Ng1X7286uiG0NJuk5=SQaksZ1MHkx_#`GCn<&Vo
z2r=NI0YfjAnKGiUlv{TG9<(a?q<wp@RuXS>EMiWV)du}fPKlf#c<os2RYdvI6B?~1
zqm8EgIlSMH4_;4-gY_ujm%ZOv?u&~;btI)~1fkz@W+in`7_k@A+BU0T2ox{Xkqm~?
zZ*)Xj?-#qv9e)&uIt2*0pj@)MC<^vl7j=J$-ka3!_g|w0Smzu}EyYP?&4_%U((0_{
zbCEq`8TIw-OWspuYgBwfeEG7vC+75STTvYsv(W`hK&(QY#=bXtcl)5J)_V{t$NeUb
z;Hb=<%Vyo7>Vz4Vx@c%XM&I!(G;~C5r;Mf7{9oC-K9&&(`9^hb+IXXr#SS{Ic)KE>
zKFf{dh!R{84Rm*P9r*hU)r&P08*?$Suuh&;#=v29oL6io62`yWw5>#aWR_QgPv6aQ
z$DJ}izEr$N=#E;jXx(16o!;++^${c-VE5#?P4LE}3LMwj-(6abUWM0K%*^GLZvAOC
zn|R;1F>#Wpc|J<~&TDYJ$&g=a<l;Fx+z~xmVvL(kfR=J6DQiKok#ThOwd2G3@WKAl
z2Irf1Pzk3Its_z05hF_HM1sM+e=_vkaen?#<joPzB49NoCFYFb_wklRS{PfRgL4N0
z_nkzu8~d4ruSkx0M?TePv9#)P?$bZ7M;!)$;deG0HE-}&2<(mj64_0@YOtS)0Chxm
zBE1W+fc9tp-2W%NFaHskk7UE*#HJbU5acySzu{cNL+}}J%Gpc0w7m8HyZu!*X)I0H
zLmCJ<1rW4Gy~VV+4h18$wt3GQlh+oO+MX#G&$^z^&n8>#y71%N$1=yO|JfBSR{&Xm
zJ)s*mu$%fdb)Zdw$Ys%2*{hjDfZTk|%*?Ai3GbTi`jt~;t61grv&>6`Vdf|vPF8RJ
zf|9{XgB&_?og}qP4tBj{0&2EDUR=TShxs4S{Ke=!uk;!hh7r!R=yA-3d$zob5VObc
zLu<@~EkDzcj;=zn-Uxj<TJ8c?2rIXx{JPWmaJlBt<5FZf6cCxUyZL7;f>S)qW|?3Z
zo!MIG85f&6eZuwq6yS&i2glVGBgXXD<y2(5f&!1a(>6Q=HS%j`Jo9#&F{;R(4cC-w
zr~S*B66%B>lH)7o2Tv#c*vEvbzh{VYTLzQY%aAZmk$K0D&!yHS9U3zJZ*S;Kc^+&+
z&WI8i94ze<c9tBS<h%6+Z)IO=#4iC(2*(E=EI?56_@PHWp!u>L!HVO4%UQEnhK_6E
zUiPigiLP;H+f{hjbzHeg6A&2xUZJGufEs^g<q+U~$+mUV!~25$m9x`KZoTs?{5oAS
zBwn*zTV^&}3y2fKyQ`NFGlmsN|Ex-*a)muCVO+ZeM9or8(N7S#TD?b)qHnM|*ELW@
zTe-$ud$<uo0r=HhWC+54_>~~2#hMCKapp@fX=e5R-wPmH8GkY4OH~mtd`Vap&G)Zg
zkF~ou`E5yJS}$+IddwcVbMQC>GeY#{!4i2;4F<0G1pvO<i@SVDFW!8NcCR9PJVypD
zGDXV`2<D^BRBbThMQz2V;SHf*Q060T-SIIQEZHFD=J09C$k%Ap4>%oqc<}x^#Gw}_
zP9iqfRHCb{oxiqTkf;Nb0bT}bmme_5qXT*8{-tZud%*rNPKQhs1sWfYzIU-QtofXg
zJHF|{v;|}ZN1$*DF`j-0tY(mUEu`_|^bG13<2mLbFz}@FbBz#%<5qOnT&O^Uc?8kg
z4(N0laB>qiJm(#(;fy{zdMMWl4qZTbwoFm3RsFlWTdI5%1Uwzc_D&@V_ySPf%$eLr
zQU|qzhD{A7DXioAr^T5H?{n$3|6)MrKYWq0M^r=3G=R|@-7}!w?g045aD^B@KGCvH
zgV+dcG-!nojsw5-Tk&JR5_HJ@xHO~W(+<*#CTh!23YW6K&1A@QTITw(#a^%u?andm
z4d2h&Cq?~rJn6$Wso?htZCZEfh+hQC-*R~kq^cshXL!Iv@0oira^a%kF$vg<tI@Us
zORNTa(MEae*~%_6q-t;f38ZSJyO#87@)_kqN`%2`DB~(9dfCyJ3Q0IeWoeBr3jD@v
z8>pOA>1|`JR2+Rdr*}tfbuOYK!sm3ta);(Ktg5*t$lS$&mu7BySdo94-5gZi!Z#@w
z*R~0Mn&7#;ZMQq$9!t7=Q{n*B>f?Iv(u)vR3b3_%+(wDSB8a{LFnj)&ZO=%bq-%dH
zb%=++e;m~vUm_nGspHC}x1{G@Yd%kR=>sk8+h@hHlvy_nGc&p?ZeZm)?nE|1-S}qI
z<pVdBK*{Ju72x54>@`Y;WwRhuVw#QWxwkUrE4fy5jX}>UT9D&{kg?_Imf1-^oW>gp
za6#&&gss6u@-E4u`m7dw#Mo(>8Gc>bQ=t*2?1cEGI_t&glzJe{m+rdM3jTiWzjS!`
z>Lo<@meBXU9p2qLx1~X)Kq(0`9pa|cn6~W^Xy|q}#348%y*Zfpdj1Yz_PR3>C|M?Y
zb5*833f#rPtrq+9b@JRI!x)jde5U$8kDyOGQx=A-tdG*4*{$hzYnuO99~r3mRii>;
ztQ(4^#Jau*)eAHSvDIkB7;KFM6*P`=UutI0CR5GVa0oGfB@UrWm*0QzKtx0qG3{R8
zTD<Q0ZhO6FS6@wibIhX6Bi~*(sXoil5~Gv$y8Sau`WnfS_+U%Zc)4I^CgMmM7O-An
zV3<Q=ybo<(K{ujO=xwe3vQ5d!{2(pLp(j6>?zcrzP}A<JAHt|Bzhl4g!6_Vl-_z~P
zo#@uRIM=Wec%s~*s@78#dY^JJ<I=Ap^AavV{=N}?Yk%pOqJ2}88UM^)x1{>mtO|;u
z4YftFJ^<Zz;y%^Hr5GD#`Leb6lJ_x6*R#EOCorWARAUKU1hAq6Q1IqZv(hk`_$hfd
zldH2-zvsPm$BBqS2^@<li!t$v*#6@xd_>L5UJilVoOA>=>TrzXI}8gAls)=f65>5C
ztj5H=TdkA>PaTrRsDzDbH{Vs@4;sOJExS?ekAM`vr|(OL$`94JhF8L&C*?T|u#Nmd
z!~f=DR|P;4VrR_d@H0tLy_Qx+Kecq&680p@vjCoHzh=KKJ_Kfp0-2y+x%nHeH^0yI
z(U7*-N!^_)0k18BS7&x!ptA0c0|nyX8TL*|AgCF!WE3A*vIqz?5<$UW^GJBwl8n*v
zNTc>IF|3B?z1MA$a;VS}n9&?e&tfG`x!C-n<+oK%0E@0wfPjuo7SsMOomP}+ePUeU
z;MQnGlkrSxc?-yzZa*cdsZqGF<~Z&7y|=fT)U$p!!{h#KK^+;I9Av%z0;-9WT#-Ya
zbyM9Osn?kWX<QCsFrMp-^jImvOzA`Ts!NGD%@T)ltO&f~tZWFfyyDt93DrOpNZZm}
z22a6Nx(c*Vn6Y&xBP4hHp)kGn31%(0%~Lv^NIxhz4n{{6GBt7>-{s&=v~VHta`v*<
zVp_hlDT*ckX<g2W3jKNATw<lTq&E}S26{wll(-=XqdA2ffp=yPx8bhx9vvDeaO>+i
zUy?vH>|64L(8oYIV7nr&;SJCc?VE~8Q5Ho3j=Byo*+kM{XwI<+A%%<jxj9~(KwGUp
z?sSCcJocmdsK#tU3|A)48BQI;F<-?K-!<1O88zgY1e}^HXP3fZljhhPNMLGhk+BI^
z==3L0(>K1G6*)F{<F2UJiCn42vUvd-VdIQ>E@8aYbbVHczU<=gLhLN&(#1a08G;=!
zZEf_FwfDoI2t-geK;^>aVaniBIGB$$BFyvcfoPHUGCxRv>1c_{(y|1{d|l7t*y}wC
zuB|w4{~Q9nD2x>3Kxdq3^IWxl6R+I-%h75<ndC5^q@qrUb^jXjKQ4gCE5=%@`5-Wj
za&J)Z>Z8j%fMiLkAVrk+90&#OIe_f7F<M>r6^1@&uFp;q9}Ed(NR5zI<B+>gut^|H
zp(|LgA7ki`&M)q9mB5ZvUyTZqL>wOc4ECO_R(JQFzGGOiX()w<5V~!u-j#klXQFjE
zOPjwdeP$t9Te=$mLyvP6e&*Gy3@KJ1_h0RQr|U{*gJ;Ct_HcNzCQ3PPScjw63TYX`
z$>DZ(@Q)Xr>t1LGWZ`ENIjNf?ClhDzz@qOll!xkMMQ@l=ySJr6ZvLY&wyI_GxSrx=
zb*DyjXN?S8C}04m_@j*wciKPh7K!IqUOj)m-LM>|Was1jj|6;3!tejLWLY$_|LvUA
zO+4P4@6v=bE4>{0<{iV6(|}B9h2&m3>H%f1=DsvpBj7dajlpHoxV~*&&2$Vze`6^p
zQeI%9Br95#1Efnj&3@Kn@s-&Z?VBYjaAC%@FR}-aB15sce!)|<ZCl|^&I3L-kFS4C
z%9@KULwN_Bfa1Dn#s&<j;p6&a^=A5m=E}j8K<(eYi>Z_V4-#nmIQdeOfXaU<`q}c{
zzCAbe?P!eLWd$b*0vBk$?JL<n@-k4DGSJg#OleK6B+6Vw)}Dd4^vzC)zK^|W1%L-^
z;<&E&4?=1HuO#zNv8ky8W<(X?rjS>DTvA1H<G0%cpK4#SucH&{G*fr%eHRsQ#YynS
zI!gJymafOn7=obfPAGBtJolPJd~cv2JLdWclqs+C3lIsk*M6cmIpO`4XXN1UH7Lq}
z#GyZ)av(b11M;aw;Kn$rNZHN{;P3Z~B(5mOq6S;gO&6NcVuAd(HDLsx{<7U&P2H<I
z=;d|+S~Auc?UIu;hxXdBSPZ(-u53apP>+6X_MSp&;<;0;cddM)&VVEO%6ZfeuMz;c
zfU1v0{en3v$+(R9L$^$iS9l^5*AD41j)&KxA*tY3C|psbgTv1F?#OOLj56#%^}?q(
zos%#jW+|rluc3rPx`+2gR~?lEL@S}Cs}U6^CM!2!S2zw0|Ex-3l!Vi&P>mNph#i5O
z0Zz2jOX+pKCIki`zBO99ykwgghO})3=_gs_kQ$;nX^$*n`zFJ&=f$pSpRPABuo*`u
z6R4d_6J)~n!!<t)mrBMbHWEa@<`m)vCLVDq8$iEoX!@jVe>%l|EQlnL2aQzy-d5I;
z7#KJwd};>ViG-V5=Hp^VD@E_|G}$W*-aoUoe}Hd?l<y`Gb*7x^44n+q2}q|B!b?q@
zK1_{vWhu`7TU;hSfZ{^(`FvJ>hv_xU4KBUb7vLE}Lt{+Azr%!TZ<E6w5r~?UDePmv
zHB3(=<iA8ESDnC(lXI84wq*{yN3IOm@?8D7^Z`$D)d_MCBK4sRM1f*ExHHVq0Fx%u
z#l{+}W@~In^@R=VGALyzkRoD#!8;z_P&}tSHT#fhckc-p+a2!APEC=c0Vz_uhue#s
zJqlpXQ<&)}?vaI_KgWXDpfT}FUhMq+c4C??(Jyta?CDkV){MHQnda$C-NQ@S1JSma
zq!xLap)m=4sSm8Tt;KVGue1<;-)NTt&r(Ok*IsCtZkz8)w(lfC=aeEu={}l7X&-%e
zYz3D0mj~C!`~PkPs$tS&M<ND;A14pgkg%uxEu459vjDAGJ!ci^oNloo7)~I_*x8Fi
z+OA)XwzuGVzbw@D;}ElQ5!s(cU>XWMYbCu*$IZK1Vxlm6@qtY)%_ZS3oOB-Cv8A@r
zjoyReb=)SRBqPhEvZ9~C=3Sor%I_-{JHus$0<{K`H<_+Q&WN_0D~H2l0oP}+>c(AI
zb|x!a)SO!FRg7H;zFWZZ^QXiahay;0elZy+^$OQ4arHbvmZMK%^yF=Xh0MP{vuu*1
zy1mr~D+}p#*t@c*ckf@!I9tcG`;Z4ya3X^R3ITN5&i97%Pm%@(83oB<b%d{JR3vU;
z4S-8g3$IGZ1;CnR_m^7cKR1KDPNrD+x-z$tr4S74hhsX@@rnGujO`8u@jz(j>p23u
zsu^fw91`tL=<3an*{k{9%@jt{#T(l4@we+2B(j4>++IbRhLW@@+xZOvN&BhF#j)4j
z3Bp52744T5?X~hT1Wxm4;BKJ38iAEwf+;<eeyl?!^dox}^K^GIM3H}YE>+=lt1Po&
z@mf2xLQ9R-)_}ddu&KhiRVf+pz`wzb1pI4E((Vi7;Fbb(!G!?9G|LB)9t<@HbeIn2
z0=aSfGsK@hHncAlUKmq{d^JA9_;G`uz*jQnf?nlDiaGqQjmbui_)6RDKb#KKtDA)=
zR3uj?c9<$`yMy1D-|Jyn(3?!6Z=m6{%()!3(-*`Bb033HsT2s5Sj6v%T;JC%e(hEk
zglcMb9naBzjpy_Kt>*kp9bWr9Pk|4-{$?(ZaKFYwvm%av^h)pAcTJMY>%%Ys2ZVP$
ztoE~xbn76qC~Q0Qp|;kX@vk`-Zcz&zpV>G?WO8ONN_*`|TuQ@*fdSX3q}MlpNvU$*
z?SUP8bvg`m)q_B@7B;?vqS<smlXp}ytlil>f<NFQXG5abeu5V99ikLkaox}bU{hjB
zNSF4hP>FA#7c{lA=sw9Nu~xwLkU1#NvxFU%NM*bziJG3Tw=?Wacm>T57Z6iaP(o#9
zd?!QA4zQ#6%XE`8y9aIm=`ZpBYVI0TPCl<_dRvXIBmK$pTG;l$u`f@*1ra{vOy6}B
z$OWjE^<U`=g}fT<9w4?=;KRuBqCQa6I*uRj*wMvx=L44BKOB#x!Kn1^x&(};g{3`W
zr^-mBP9{&vJdVbo;Eh+MC@I3GT@ZzDc#U}_2@C+;A(t1$E!|e4j8hwKj+cyg@7Fmh
z^W#EOj~?&v%_dJdW52CuN#b(bbg8bzQM7$}`vy=p;^AmeQ`V4rvO(1W!QnWxjA#he
zJ+aDXwbM?W{G{HhES@>SK!N|fyL?gKRJ?XMSL1Q@MUBtw)L-wyJ!el*seJJgA}$Te
zR^4pK(k~*Zq=}|{u)YujU94h*{by`iwU@w=NQw7+aM@n>;LZbZ%5nk<!XRfPT=sD1
zE^&a~@+XHCIk74EYGKI**f`F*7<RH4{sT-ABA0ho6&1qq;aZ;4oKBhf4t7bAQV5H|
z1$}w1TCMf*TR1!y-7>E?b^?1zS9l8#1e1WE%wR1SXdhk*v3#!JnbX3e|LSut;U5Ux
zgI@#J%cyP^WGKQC#!4#x($Ll=f`+0QvD@Yw3~UnX;99dKI48_Ox4UELXJ7Bb)DXvA
zcApk8hEZqFNns$b5*(kdaQ*c75<*E$1GN1udBpmJ$1)C?dzI_+M{o5PaWSH3zYkns
zIIR*fl<Ehb75U&w>Ut0Z7Dw$cGP@`I%`Tf%V-Sn6T+*74*blh-m_s<D<$Di~N8Q%n
ztlO4^-+}!#_`U7m+6x6NeQ6)OtO?QZ=6!?hkd^-Ef0*ZFVMy?pYyWt@o;s~G{237F
ztPn_(>V9hiBVqXFeFGmtAO*?^tM3*umM=Z|;6^Y}1~6pbdkMW84&~n&)5B8(qmgnC
z^Znc{HB-P=Ly4HHn9hauJ@Je;6wxhTfwvz4)KuTTt@}w0Y&)HL9({u7ouP~}jO+W%
zL2kd~0lJY|Kmn<8x3~FEm_hv$$f1&BxZe4iBqhRa#(QHSA}Xcj16LvC0raX5v$Ik*
zqW_sblult{p}V+}d~)$Doe3P>`W3iG-V7*9rY}ejKNQPbzdadKMZY#{Xv9T&y3a&5
z_xN9<gvrj<6P&T0(X5IQ+e8Lag(GUv4Fmcrilva>G0Qso{<u=C!BQBE^5VANj7x`<
z5M5Dim{8w-o~Le<(PlX6i1}<4&d`_B8jn)n#siVQDcjT|2I~Qyp0|(NLUpW@3r#>l
zBdzUYF?cxqi1Q(>Ox6!b0Dl6V#h*5xHeGD{WO4)&a>ndJ!89!>Eh%xqLq6%XWT!7x
zL%K*mGx^j@b-vN4>OAEZombYYdC`ZjNA#&AK#AO@&)IyF8qLIvEo3cKng={oDFpK1
zF|;)TQPh|afeXL_nDQkjZ^++refD12$ygt!fl;Dv=%Q`#@xTI{hzt>(ShK}y1dzKv
zh}ucTb<uzV89*aK_lu9a3_INzSWD3N7A#OgMY!Fis2E=mtx??5GDBmASxWMoFlCUk
z?mItfMTXGo60+(??MX>`gnR}BEedsOf1_T(7NfpsOjh#cyCIlr)>4d&*rPJBL$Y0`
zp?_-<;XgI$oJjI;10EUsIE)Fo1+Py&2Nq$lT|Ht#f!kGZubnXGPY5@`Kth@0O1NV0
zf`uQ(HYW`YGT^I5w_5xO=@sO3zfJaYdM-b1pxwQJrkX5UI?G9`mU7E<PFIArMcSsg
z`a3%Mq!&9wh#iS)7uBP@&ecVH5H6G<g~0dCJ6JJVV5mS{mS2BVn>02_6)>dW(<IA}
z+Me{do7=VDi#^8gnJ+Kbpu#CiLvlK?2L(0{H~H$WWS?7?q59+Ee_NM$n4hR4O9bh6
zY-PTbg0g2-EAimrJ+6Mnz~z5UKA9ZJtXLM_@joVd1?|5TmB<${l2oaLjK8lQE2K0N
zHdD+O`4$_Jt&}G96sr>dGdf!3@?9nF3&rVT7Q^^EAFq*mma8<bt+Zt>mqdpn&*8m|
zsk6cR>8m4;c#o}SAAPU(A6|brP>R8cUK{2)`m(R321vtyL)v~ECUjJxB!F)>KUo`z
z>vvy~=E!`+?{l|&2VF-3rE_ZJEoL%B&I8MNid*i0QEtqvusiwF3pe<9pS81UaK7|d
zr-^vf2~%M+OvoJJ5$e7EtNrEJCzh7cW|LRPxg&DT9j!xE0!mumZt^7M`?3bk>Gw}I
z=XV=ezM5AEdIhGs%E%+rCZvHT7GiXQFcbnX6f06NgQu20nq==FJ@4Ygv}WWZ6}ljX
zU$%Z@7(LvQRr;u7K`U$Vep~Nu@pVmq6(vX3yV!(-ufLF=@_`$@VR9gtyx<qTfR1O>
zCpQH4)$2(H?6g-<-MhCEbP0_+(!bU7m03wpk3=>(&`4cIY%b-+(g}>oc)-gIynx?Y
zwB!ulcT%o?x%-R6Z@5SUJxy5)P3#o@3H~ouXh#3+B`kAmBQ4bzzhOqSz^#lwV#=S5
zT_SR){%TynQ4B;HUBxC$wFJlD9u85+=LxBEs_lYY&}Vb)`y^C~y!^0VGDa*|@?m;7
z>QO%7-Jd4or{?gHGX0p_Z!Rq;l=2ai$AkEUV?q=#q&up<Lc$3-#I)x6+EDW05<VpY
z9)kg(MoGwC%;ryJltx*8#%2^_E>P8xbaB5~rfbsY3^c(H=CjBS5&29}iubMQ7vhNf
zTM1Q17VeQ$BMuhKN04osFIj?w^gcOL^LNHr5Wd)qO4VNW?u9V_%c_@o*00lmmuG$-
z6QDHE{Ez~QuA2QhIc7H+J%Z{_a`0LO7UEPJ4^zX1KN6f8^?rQ~xE=drH6Y=f!2XV>
z%CdbNXrRj%?o`&4OOF7He(LdtL~fmS`Bv-<mBNjnq(ldI$#w131l_c8)Tha3;JSKn
zWXu-t^L!0$Msio`;s~+xn(oL<!LcR9O8<N4)8>tHrdt^k^>!!b7c7Z0<`YE*f99hE
zo=r=GN!$@Dk-i^&j1Sk0YxB+`DlmC(nDilp!e+`XF^+bpqssOJx6Ot_P)Z%y!%VRz
z%Q{lnPx|R0zBV4WO%8MEcOIVQ6~nh<g1#LNj>TBwlFyG`CRs)jNQiu>NFBxC+=}o}
zc$$8~F;0JBYj9xmWBIBrbMFg%6ckc@JXGKo9!8pjz5TQ0sd-atAFer|v@4)CT9h`t
z0RD8Nt*lx1#bHC0;JKC-iDDF3olR-LT5Xv161c<A>Za`^XGG}zgLdCUum5+k1@q4%
zOSwgXk6l)4Cq3lp3O7TH_Q?)-o@p15*bdNT{GAOv(ERKgAAS45hy=!%Ka$Ze^v9gV
zTZIO}kI5E!Q0X^D3xzK9qc^%4)g!sDwc8NhJN1?s#zD`Sj{S6#wYkH|pBVE|b2pA^
z+pRAWM;<xULW>B3lkkO;5J-fog3xBsP~lOpYI4g!_K+nu|I}AY<-%SNEe3`)d?$0+
zDvOWl8)=3R*j0XrengDIOS09%%69f7KKX2ceb5;byk8R&b+cwLQ4z!g>rg+B_P*1B
zu1tPz>WHl3(&y!01{6){?v>xC9CT|O3VxHQSy7I|uS7%FjmRweia0C&mSoJ;6<X5h
z!1gX3<1=sZnNNwdFC#$8jJ}s*iAS+LnHkyKlNs=g5j>@ZJ@lq<pxckw+y74BM=?`x
z{dO(YJ$pd~;mIUOb^5W$#XS?AHB4zhLCqDl62^Qiap&9>U|TwnEg4fS8|Z~r6xK_#
z_Q8Bc2J!){Y6^}kKX$%YaGa0@8L6_rm{fnOWEFo=V_eu{VU-RR!?zuB-Ps_Dt7q!R
zqdY($zra=xGT~@WDWQRc60v(;+w$8<eZIe~XAbSMMHvzM<f}RT(WW^h(*0`D#SjRk
z=DrV7>==A8@zRZRQ0@L}WC)veHYzcW$LsDGW%0A0Z6qPth#<arWvNqQoxRaJ-^k}B
zYdmmb$?$MVZ)<7b>~vtVGc@51>*s!LnnSOesnUP?NU$%pG2`;@$%~q|l4d7GPt^F^
zwVkg~K}NB5rmC(RIKa>(Gay3}BJ-p0#OyuKa5}#};rEHoVqKQL2~(?7h|Pd&c-mrs
z8@qjfyCiRx$xE}IQeSHv7hu)%8_2^H8Xe7YR%gRqyt3rJm{iFM<jSRn0CbTid_I2^
zAScx(_{$^k#)LeHNcD2rB2}{R3uKU^#U|Xk1$z{r5!Lpk|6_t~)|;<ewe^|f*qKO!
zl9r&WQD^z{u%x6YM|;h@J}@Wv@aZ>Viqo_vbP;!pe9&KfvxB*s=cvmH?+5qJ4^r`N
z4a<GqBAAAmf42K0j`xE!I2v=g7uS2X`HD^xC$0A@JRLDJj)_wDx)o!fkRpARVX)Iu
zqaD=B>|-|TkH48jc}n3Mk^BiR4nR0L%Dwo*Y)BJvC{s6OdX^WI<vvgATVL<*&a*Z2
zYkcd)oOy7dq%Yjhn12toSEqdIx>mDm#U;X9%G0xwx}`!dJEctj@UW6G{HKq^Y=K+l
zf1Rd5rFF1&4m81_Ji3SGY)SM<CP6Frfi2&sH`wrK|JGqilABuR_M#pH#{W>#pjhar
ziXv*Vdw<;LbZIz1rCk+r>^a4mG5ZjRzafII4|(z93@;-g(V@~eT_0)_?d*;I(S|?`
z3X0F>5T=)ri{eN3C$W{a;XwTC3zqOpw%0dJeT$D(bL?o>8L1C!;Zp?IX@*Xpxfn4?
zp)sj2;C^h(E-{NMxyDE3=s=+Q!aZP-j2UxWBb(N6jRp@(VpF*ecE3p!FOmqX;!f+q
zq-8ZlaUDsnXFr*ug~;zz%h`~g>=%X4kKE%Xo~4ToZzkp*1Z(%cbh!f;3d}pTnVwh_
z--X42$PZ_zIoE8+MD&F2bp;Z~i<7%Xg0is>AEE4Paax(SIO}0mUV^OeL%oQ#bQhGu
zd4a9C>GGd{4sNg8HhC$gnIBWAlevuscQ>fylo;HQ`@-D`jO%bcvC{&ONAe2>_lFg$
zREYAC{1iTaV;R^qd^av;|1srjJk1&lru?8$w?bbiGXoE!z};{s6>VCaqg8D9h!oky
zWdSaaioQF03~UOr5Hkt#NZn1KV+LmnAlp4Kw;@DNR#jFq8MH%%5%bh$JOBa*5up2d
zsX5t`B<bph84+aFKigaL-5(35xEKC<p$m0>Vij-3jAy5s{!#N{&eU<pb;S>Xw6aoF
z&8EshVWV-gIxI@WsQPV?CXUFIcX)k0Tc-n6*<GP$^M*f0fJ&g!&Oo1A33GyJQ(NI6
zwE}+L-s|G(9iI1;R#r>RnggF0j<_|6qoI}8O`<xC0^stF*bBN%>X}87B-4zq`W1L>
zQ1mIIOQsq{cCP>{v76XsVU;pH&Ev2EeYQ>zHvQGc>lYM6=OZ*g4ABWR-N0EFB@hRS
z^5k&7Jf?<e4o<|Z92I$Xh|}^Nb@D603(uFAT5BaXTPzApBi4SA^RwseW;pu_r_Shn
z-l}1rbeH7gEVOf^HVJoM%$K#cPTXaWaqic?Bc$)$e0iEEzVWOTr-)MQi6B-4I4yAK
zFH||B#j>g(>Gvoda}!>jbw=#6ghw9v&{Um*hZG5gg!g<{XWR1?+Mu;w;(t@hOVt~-
zyNv3*w%QN<bMnB^A5h?`C(KF+A<8LARaBs4yn|!+KIgamaUNmMLni@4f!+c~1?kB+
z-Y1LzuR771AJ6#PWGE>o<wnyvC7x0CIh<X``(}emyVmkt%<V&XN{7P-KeS1Da7wwI
z1GlDOec4A~B+UY?LiXwKgYq1eu~J*u=22m__@OB%aBE*%uWB+9N{=+DLt!P&p)I_B
zg69s0j+R#>Y^zgQOS`#H@+>V|dO76&zULB(#REFI_54jbUuYGFw4*h<?U1&BR#(TV
z+H6FYm+U+*TDAKEe?&7(y%FA4R63XE%ExGMDxbgn@+sQsMlR#&$2+RtJLs`NBe7`_
zO8uVOKC86*baB^vd*e3HM=^a6HpxEnmZ7Hk<7~6nUet#?##2<wT1#nt(wvPo32S?_
z$d#(qX!*<(D}K_oq9&MHLhH9I3kL|ake=OJIBuKB<UlOiWx?s636?!6Du~!F4IVx{
zC)8vs!B-O2kLXHC7bj1Hj+N5#)0y}C1Ifmm1Yc6hBthrvu4fibV{H0;_xfs`%wx^b
zXxQsFPEYokjc|XLJn8V&9kebAcAC}O@VSn`L!uh#9=~)7bQ<W+eNpQ=qZAe_Wh2<X
zz^((~H5y=jjq4f3hy7ma9O|f2RD)oMw1RawtCiMylk-Q)*^eM>e$4-77DMX1GE1|T
z>w)Y?iDsEz3zP96UM!7j|7ByJJ81Lt8q$pVwHoYm^xcXEo5dJ&n^re0r(TXQ&jOcv
zv(q__W$f19*rrQByCVwbN@OFBKHE!RHJ1epLqs)rFvt%ON5(U~OO5f9bBYx`LuD4)
ze0+Z27^O#6%H`mLdCB|MjNT!Y$Zwc~mx+=&uJu*bUe)P0gQ^U@ANKw93Uy-mhgp)Y
z@z(@g_7u@EUXR`8F;!A_P0ZhIk_%x7kIHChGu66x+~H}hZX<@xw8`=SPNm%>H^;TU
zm<w7%y_k+dpsQSOw3v=J_ayk-%-SzvV=}l@X)pVR)ihFH6?w!@()*+fcg{WdEb{Ik
z@~v=WWTe4t_=M{f`~EtK)uoIfmj(j-zA;gR=<4jYo8U#21nGKWV&a?$`NJ0S)R1fT
zfyBQDmkP4A7U|*!G%LkkI5iHv#&1qj11w*>OOH{h;-AMzTSVW(PCF2RA9x`QdS0e;
zjfCe8@aK1M?e&&v&e!n^-861tJMW*Yl-)Q~82yRq4=`ijc2lu&0cjd*D7!>YCUx<P
zH|>pV{RicwNRa~P&!>8=qsqqAeFH@N`n=Kix`<mw!SWN*cyy)p!P^EaPC#4?2YOI(
zo8RKL>)2`Jh}t!@$RnHC`^Y1@O9%}06KySqPf!y_Fpt+-PNJs+_%$^)u*Y&JhIVtN
z#o~`Jxr@-24}v8}I9|S&<bFo!dx?oEmQz2YASKbO4NAN#@tOi%RVW_UE`&@M?^i1G
zd~f*}dk!>Vn`iU4CC+CnS~P!RDL+(+o0veIrNevUf96Xz)t?BegHjJiP#SnWeDYh-
z;1H`_=>m5InMT?}y*2Hr!~0TEpwc*$vt`%o?t}Gf;iy~Qz%7X=WB%viQ~>59892kc
z2ITaRf!=C9o-nZm>OdaHUGe=s?y%xOMzXuBr>Yd0U7v2!0zDJb(&JEzWhpm=Cy{CY
z%-&hP<w|-Rzjc%&Vf!k+uQ6L>0i9Wqi<*Uj|K6_1BLo3>y6KY<hG%QTR8804$T=%4
ziR6l1rMWh<4FnQZ*<my;7%IJq&5C3x1F<YzMn7>d3~_1)#q9sCoE4Os$B-y=Q1yLz
zL@7>4-cdF9-hh3tnaIfI5mS>NRjvO##q*agC;{a=FVL21a*y{ZcC(65t9Zm`-0iBn
z;e_5nf`#Q?Z6pP9SH&Ow*>C<z)Gw3nd^=V39NzF&)6(mm73d-*ylcpjw^@!0jDzAo
zv^`wT*8}m%eb5_AV+|g=&*GSuwS*`gRVgFDQ+|d*_XtB`ZW=`f?2S6}6|*9vl^i<@
zB@jo%EDUg5z_a=rWJ87n(*rZ*R}^PFuU#&-B2~N1mhqg(ly?jvIDsZvHMT{^M1^<7
zsn4=8Wk0puBIKOrHf~M)3lnA*zv;R_gPb9Sni}AaC(D}3Y8FjSm7wQN{cN6HZ83v-
z!pHcVfGw9GU9N1N2CLk`JMfenJ><^6|6WM<JuR*DKp404{vQR7gBiUHS@m;b1dl^b
zO%&vjn<Dg=9z5xOLN~M0Z}<ymG8L97RqtMHd=p!3h|JX)c;<NZc3V4h+(fF;s1@gX
zwR!<L5CEGud<U?tD&$tVVoiwWkwttvSk?lb51JGXH1bH4{!up?nn;O7<4s?G17&6T
zMyUMp-Iw<Coe3Is(X7_*Sz1K{=(zgz7>SgOXgsaS;Tw#o-|oNsAqVHJ0zKnc$N!7T
zG9r)XH4_e<M#h1RmAH32<IJ)X{QtNBGNnz2+zWO+#HLHlTpL?ZdGvu>pBi%E4`Vy3
zxZX5%oDn<t%^f_<&!5m9^Y?8Vuc42vl-{(P$==h#WBw3+3nA?5PJ<ZSz!^+G>|!wy
zwuKziwQRGnrgOfyXF@ccBg|jOuq85{P2AH(arDNe(#<K~?@kmbQ1&AzLH5~vUEo5`
zmaa`#Z1z1-Am{ecVF7~j!!!Ig$z58HAK!G8_;{#!MJ1$e*~8Ub!4(xvy;kZU%m@7g
zz<i1$8acn1;Op8?AViMezZGC#`M{-8#nPi*912v<<d2gMvh$}Yoc<}{6OlEZ5Qw--
za=bKW36c_n`%5H~YPi2x;RBM!)64uqoA({KCqo1;>mrE`FTp<r?nR0;CwQg#2&Opy
zsDyxpHk`K>Cj-3?^Zdf;e0m};<XT;!qK6xx?fUZYf|QpYU-e$!dapk;i9W!g{WL1D
zR(`9L(<k{)&wDXoJJ2BckR4P<-f+VEmC+^F&Df<a0@AB@_knrq)l1HR{BKWkYZ&?!
zX<P#>7IvRjsU~Fc|H>5vO-=`pelUr*TdK1|9ub7|j#c^WyhWu+g~~$er+|z{R=?nu
z_Y{;25Y>iDj*<y^j%5Q%Olj<0+##<X%UIdIA^fz@`|f?p-tS?~fg(XMQwYB8-G}Hb
z5U!&ui>b;*!E;Xj9ipV*2o$;3BR-E=O0T0oRDu<?_vZtn9@Sk6fy`(c2nuFPJ_A1s
z=26%%05d2862~x*R%cCRjH4^O(~P(L2D=J4>pXAlbvzDg{<QEsogIO|Yqms{cX!!X
ziBcn}9V`aF+s!tenydm(`7ExIjKzB=>q=-;P9Du&R0jumbju8W3qF|oThTHyqJb?!
zCuO?l&qbE;i3uxMchEOiTVyw{89C>7=S}R@_Gg5rc_adKWC1Me?bgCMK7m3(;~GGd
zb(=z939M+um0gGR=M4Td3Bs_=?}y^21#mmj)AF>)6R;6kYM(wH?$1`CEG^d%HJ1v2
zA3d3~C|eDoEfq*PYLeyK0EiDYWIh}w-bKo{qe|JnSp4~E8<PYV;QCAn7#2;rZSltC
zE=~xI@JP@iESH1Ya38;T6{)WriSk<nVD&XforNmaXi7^guG?>fzPufcAJVv8ASIIg
zK6u9#mDf!ICdYHon<gl%&T^oy55znNfgZJ7k*t4)UDo{ro66jJTD5PT7$4uwG8haM
zXo4FA8)$MSiWGR<Y(;V`$m4<H(-+hJc?3$baZkJ+z;`)OqwYqyOjQZ)FO~<=;kEb`
zRkjsgr)i@>%|Wy|eWFTdypbn&5VKS7*DU04j?Ylq?>}~19_hbQG1Iti@~PWwqtor)
z>S`-lhK=Ed(<3Bb<(iwWJ`>3}(S3}z-+a4r=ht4zF=cp`QrgaDa}&C*<Z%bH(lRhP
zZ>KGZJYubl5fS_kG}I0aZS1BNpr+GNb@|RQk0a9catazw0?odbQJ{Npo}MyB<IFz0
zmKX^G7M3_I<>fl@HLRP*%z9{&8<&6*HHC^**q8oS==zzZqL-Rz_AdASVY4l9%#}g^
zTy>e~X`gvNIeGmz_Q9M;ZhlNi2xunfdD~%s$X|pK=<%(ICY%+6y;X_#>$#=#pFUS>
zw@NCU(MqH`BH#LKrOmbIvz;T=-q~AXd*}Nv&9KQel}M@#0v$U;HQ}(Bg>0fag@`hC
zqH`xHoQ0o_Id;iS&JudBYvtR=#H8;L1{V%eKU%Y&qFRfAmqRX<J$b!LnyJO|TiwhQ
z?@eWtu#S*<lm)nex=6o|e**1p4oc5Q+_&gb7}AUvk|t4+%F6)Dq|<H02L7ue;5i4f
zqj0iB=pj9+bl}a(4cvxRjQ;cy8Rr#GW$ZUYSxv_4i7+VVzI+aF3jg@a>H`PvGe-KG
z{OF@NW^9`--}OqsGxw}!7x0TafHYDOpi)IUW^K5NyYbrAzslMkch&V&jTHYm$8-?F
z&zJ0&Hy3uMHQiC*&6K>bFu`6HODxt0uCO5XWtoX4CB)%YX2ZFZN6v`mVWpEgyD-I2
zt^wY>iIW8aeC-jzmZZZZ-7D8omsLynTLmzyBje)IgwI@K{?!X96rw&Vi%1aPNu(Rc
zPYXY6A~TevQ!P~Kn>GTtXF1R*&_@MGd_3D4Q7>g2KLdBW3>H^@_ZCQJtJ(V6;IM^r
zZsnA&M4X*^r%9DwCv2y$kZMRZu#?UsJX_mZB9NSL9sf$KG0r)qSd=PIAn&E{gHDD!
zUr<O(>WF|Cq4J__=c~vOOkkJd5He~@{@N_8qslN_>UDry>6nXwqpVWH7QrhOcC46@
zBIP`}siig5#eHEuEFVDC6=lCZy~KE)NB5E!Uj@a&d=2~*1lrsmno@FKo}cYlL<Lv1
z4(wB*t;Gk@;nov13W_|qMd$7%pK|+i+xx2ZgpB!5u+NnOw8|`jCT*s1jPlQ6zcPu;
zUA6ZMz)sNdXo;9!s~Yj{&D76T?5yMc4%MH>maSA^*gRzMy=_dWL(PFrcS3MxP}07y
zDuXLcg^DfFtM;V$V!~43G{bCgo9Zhu;1?mVrv5t3q7JrE4yHdv@lT9Vu5=L`MT#fa
zFB%YKDP7%5Y<F_=wBE}LeSb>Ebd=bB_AQVtj&0|ZM4_?pNeAr+u64iUeWf}dTK%J8
zhB^5Yd-H4^?swHt8;<y&$NSx?t3RG`USBv%OoMn6HiIr|h()3gtTA7l$s}Eh#tmvE
zPKyy&X7;Ua)VMk6-}4dIcs5lO*T}mOo8J4!3YdzCLgV_tLEE}FG^?7NU4L)CX456A
zVVQr;P*&Nl2|>$Qp65h}ovTu$s_CSaC|rjBDqXH9U1GYFY1OO6N<dBjM%1%GOO$5g
z@PW7|uQv#BQ=GM{?R{jOE1Vtm=P4-7xc(ub%iKhXVu#OB#VgopB`cy=M~RDEjWge~
z7%9-DpgScZW7A-ezWgff0)ms#Joz`z4h-<JeW9JvUf16!(~j?&5{N%^4cE8c9rDQ5
zUh;}e;`6$*Dk&>Jcptt8B;biOCIX17tzdisAkC}-(b1IN!X*i~EhZ?kW3l%Vbk)lA
zUPU3Z+YB#<b<v&G;2#wi)4jxFf->y6^=ShkWAb5y9K$MKDyzKx`s0mOXC+>+M=xH4
z6^qDGZe?n+i;x%mZb}pL_vLH7n4v3rMUe8f$$e*eI0%{<7YvnrxkJC-^UO{MaOH)S
zl0V<>?-kO#8%N%^vvZH>YiFBhxFY7TepKH?+;CccK~`eQ^eK|qZ|Gmds0AQKL-}D7
z7szUJ=ft5gdnvn~pR0TUE1j7C{y5-LE+!6&*wg`Q>blyawQ%?y2L~>YjIll&?Wv10
zYcG4A4T&ZpZJaq%9ikMTy;S9LOP5U>%9BmroLL`WYzqT2$rn@Mmi6a<48PyhyiHD#
z&<CP*v_H<}TM+;|+N}aUMRV^7!U=Yk91ugo+%}>qAi;Y{xjwB}bC@otmQ$6VrTMR|
zSBNdC@Vgo(hM4^ELjho<n7-=n95uw}e7I16G>VC)2IClvO3VZ(Ggeug4<E%{orQ~o
zPtyqdO|A?+MK~>j8@Y;SCNPWa86~mBa4r1!E>`<OxjxCH2^x{LK96zNanBDfuniWU
zeeGn$_7>-!{Ze#ayC$U=T)<LzptZ(MlNDZ~<sZQ1t<)~ueaq<UxHtWZOn}2|91h6p
z3RMjYiNNj4K>2+8R?tFR@Q|uKfAg;?=rNRUW_fD<*)_H2Af1g0H!>+8L|WCnjSBmJ
zw7qpymuu5DtbkI2G$J95G=h|LBPAf6N-HTyw{%KKBcOD5r*sHE0qK(N?tae;-1}Y6
z`tJMrzJI>8SS)ob`*)q!%$zgF9CJ(){v>kidELt9?w#70ySm40F_nfF=6<Kf3P)c{
zb9NYPug+2BQ=_<x`Aj`H;V%?&dfA$&ygWoBW1924-&w+n>oQn7=FdHnCGMc=J~P&?
zr?Og9OVMDvX6RScUw_y{qwRq$k@tO~zA<xWW4e5{^RJ^~0!KB~0*>0VqQsi)#Ck+w
zGcD-fYGEr$J^Mx4m$X56^A}21k{08{$E7+NRJH>RwEKd$J}Nb<ZUy=1L88;5rX=tF
zz^`_qJ6#@$^=Etrj85ug<HEX>48rz0=-|fGvSNx_+e606(bY288F##HxwBx#_|rd)
z8OMA#`Qz25WObInoWXp_o6t!!(1IbVc4riqN?3l%i`c^2y~bK=^M&z|A7PkBe@_qC
zX!e`+jGA(E!bD`@p)1FmRbCwRvY*s%DL19mDZ0j2?yZuE_zGUNwQ<j|h$Yu6pFQC)
zi=}Q}`79r3X{Y1CF=tE}w60qPIOfz){nEH*S2DO3*9Zm#;}6W1Pk<7a)#c&HxpFvu
zbLOk$0`n!JzEnz7ShH<c1l7twIx1KsqwO)qxgV$SK>cNz+yWd8z}p#SvR~6MPMl<C
zCgt=>c=L3@(GCO>0X8%brG6a}y|^ECxS(!F&ci5E-?2lgU)66~`4STpPD0zBQdk#R
zj{rAPq9oKw^EOTmF7_0)OWlY7lA7gsL4cBnv`odI*UYwE_)T3rlU%iEAY*$aWR6`+
zF`;gx?9C6jU;tf7Z>0#p@d@Ohz>L0s>l}{Jwu_thVT&M6e^$CK{}Mmx@L2OL?nXQQ
zbX%nH7oM%E7UrPJ3pqotq&F~P@bT4U#vL2%h^lDlYlNE6d~a^W%jagJ7+@BVwvT(6
zsN-n}hs`r*JS%w14s1U4XT#pcW1}w{D#P3x{01MuJWm4e)Pq~v{I%CJ{q{c~hZhS>
z_<G%s2T&j{zCktJD3FaRD>B?<#Kl4$1W7^m#JL=ekCu<>9xkqb4dNnnXa31={b*vj
zue~O~l63iGckU~KGYx1k;DL4vEie7OX(0Vo#%w2Wb)rA9w{KoKm9fe1v{k(v{q!`h
zxm3|atO5f7u+vbRM1-o;syhD4T*_`%Ru@iy*D#Sxy@amP1t|<axt<_6L>yOgfc>#4
zsn84gFL8$ZQ7_abJQ!^}*q^|!3&3JZ4KL6v;0&&I?{e4_QSj=rqnp^dzZEySbQqEA
zf+d>U{jSAZxnPp73H%U1+z59a`O^STA}=pnvYylZ?r#fSB@Mhd*X#Rl=3|<YXO^UY
zDqc-moL;{#zo+;>N3bN4Y&_8B!&OqV&4&TwnOl^H3_@y$S%NejqA_y}!h0twNeFDt
z=nPdtdD0&%YCaU6Q&+u%JWTK>@2R64k<iiEtMKwSvC94AjSnls2td)v$cvhyrk(sn
zI$3C-nv+Z}Dcy^ox09!gm?pX|a3ka@&g7HDxDkB(FW0tpN#0V~Kd_4E&7{P^J<hOt
z3n`)1FT-BpE<fCZDmJ@Bj>fw>FFq8#>*;@V4GLY+ZmL%l->l({Ke@<J%jKOTs0*M8
zDPCIOBr(o2RiLoc+L+}Vzf2A>5UV{JH@);0uT^(&CFt(?rK;m-=qdQL@pW&Bkv-#r
z*dj1!+|T@HS-Q%l4Ef@pw2D|O)bHz+K5)0dO1c-XLw%%8pIUvdTQFT0+P(XyizfN^
zuq1l!MzztpC3Q(>4Vb~#q8%)o8r_R#1}%35aq*dTFSnMvxQ?`nz=U1gf!6|#F?%qF
zNfFlGS0vb(Ec}L4JIF^<9iRmSHTK8X3Cd}#=?R9yXyRKD>)oJ$o#Rxl1;yIxxFix9
zB`+cJz>iK9>+Q&W$?3ab`wTi5W=*Mwedrg%$L+lEEO?v4`Z-6Q&UN={?SN@sml@z3
z-G`$lzr{@nm}bW&8hI@{%Rl@|5pokj`KXiKo*DPs=vX7qxE#=%8N*xW+JaEssNw<_
zPq#NpVN9PYKvXdzB=2e$ELjbkLW@4I>K%WGBQVYJDcm9*E%Vr<>p{WiWBFbA>c`bL
zZb^&`R<_MMtbFq$C75^p{3b~#jW|`-^Ct(}2YAb<DkP<LfPW9|*Y?-_oD)7I=%R%|
zS2xJK_iF7WHH{Jo#5D6>JJ>xY$TN-s4q4l(mIP&M-xKbz@z&~J(*3sdh*!01&&JpC
zFY8tg4~q}vO7^y`R#)%jcEj<znhe$!fVG5Hm(sqd)nDa8!FtMZ<C_!c{Nc%l#=jr(
zp9oGAYaE9Cgi*b8%h<_#On}KCF*`<7p*^dv@%|dlYB#lguEZZkaPFCUy>J{C8=s8E
zon6-GO$WEuV0&t=l&8tnfGy!^Ndl*GRk#u>Ldz??Zmf%}#ogZM&hwVmCd!eL5_h<S
zf|qj(`Qu~wv@eX~f9p~gROI=qA;BnLv32J?4?|11((|m|rz*or*kd&C%8zlQl7=L4
zqPs8fPuxgB-?Wx}N;Rpu^nBQx0?GpJVp+*}s3*av+IZ~{>0U9U**d}RW+oZ)tG!|>
z;gpkD(z=cbg71}UZ3bwnL^A<C8Pl^~y9bUX=<n3iOYs`@=yOXgC1&whTc+Fqs>GbE
zrj_hH7*7f?8%eHxR^z!+8hl5@N>b-b9FPWwZpNY+`=sq#4EJT8?3Hf>)J$`mpAe0i
zQvZl_{b?B+amZ7CY0?_i4Rcs_2DlI|LFx9hsTI=rd0jPX_B!I-y-P%;5@)BdT6;!0
z1|pxt&3$$moTL_l-7V9Ev9D8=)I<ta2$0a3>?iEqF=vEG#{rPn+EML~gEm2Ei6L6U
zD_brfK=*s+-WY*j<ob{;fSHev80xCK0Vn!2_rnN<14{=kpxy#IDJ0gxCkUiz_~DKl
zosZ6b8lP>{<n%C=3;;2pT~I=vD`RqNvKrcMj4G7VPAp2*?3k5hdq_r=^n<=gIV8ch
zneMv%ErOUoqPevHl-udR1X>gd^g$|h#U*i1hYEJ`@Lj2l>YK~ov%HAPKG-BQw;+qy
z028b1Z-=|cn$NZEr+lmzyT1XQM-7=ourMcHz5o_s{v}(R>O7;rYQsn+ibI%|aFxpL
zoj{mc=FG(pp3K`ut8Hn>^i%zcIFnVSlqjKWlyt@DW7UZs6SW_!PCkk(tm86GG7iys
zQP#v^?htwehSAGRe?ZW2eEfUVZb(TJjM2^~ST^kD8drpSem!~q!noFMGiRzL(MTQe
z1n!NCW&$r3Kc17f^fjPNmLd&UfH)Zm?RqQ&`Onu5NX&5}?Kp5*<HPFUM=LFLA)TdK
zny%$=utoWsSP-5V2O`QMU;ox+Gtbct-zZ%mbv=F%uo_@ygN%+1L^r*l$jj(Xv^sO`
zFf<epUO6It;Cshbs>Qk!kYRfbn_$Sl2+VzKCSPm=Ui3GhAz#rye(FL16{X9Vsr%xF
z6sQxZ%_s^G6s5#Oh$ssIQ=!N#Ev5P&^FYT_5UveCR-pFmxcgMM;S%hcI@zk-HtT_b
zHHRZQB-KDklkA#@pMQFnYf>W|Q211RAqpQCW|CM^-I8pIkbRS}AMs)otP6od2C>1J
zXoVLuYkr?KT(UK&<Bkmp{kT?^kMZia;YAOy=9z#(ZZUl!vL)g}H_>~o2Vb~Tt@7-=
zjb}$q$x;kI4G?>>YM<jpnN6TsFlh+P-`N>*ZK}zB--9J>N;gpIzYtMA_R<CAS=R<b
z$^P2)>-ps)MVzFm24SKnQ1>^?0}Ued!K(!0MLj4FIIJZrHX>|Q&G`&?M*fi@(Q{ZR
z#)&cTujKCYb{vsfUOO$A#d&SWvgv*Cug$>CU5`}SenFWugb4r|@~tm}7&K$;R9h~e
z6J;E%Ln*d*&=)TE2}?%YeJA-Et4%#WreX?l7wh0tD!ucr8c@cJ`Lg8Bw3y^Ub8e#4
zo6Y4ZO?{STQ$zNx`<~$Y(+kr`RgdVCPTs6G(Ia2xGO#A{ix%9^EiLO2eidV;PbEmy
zH%vG^AlJ;|xM$GmTP251fh^j3zpH_AY-eu6CB~HXPPo~Ek2R*%p<kS+5ul^7=AjY4
zzyU=^MTRN$-2DiXNooMrwCx9WMJy^@k*x4-bkCld&Cs#tjlsosZIaJ4fYrAec7s+#
zb~=*BN0tYAHH12ScC7euk>R40%dtpS&^#uXo-UNtR@Mnj+cEpt=M014(yItwRd$wI
zXBqHR(DSa}<EJ4|{o1W`w*NyiB8pY6(A|0yBU>T)o8lIYW!<hgCa7B-BKK8FzHz+&
z-8sB!N?mcUd&ex=aNL4gp!D2-h}+`hRd#D4=A+cpgb(&~zoZ)NycMT)&3%gAz0F)V
zEJOB)M8o^Zk^c+?An#{1BiwiOC^@=_z|6A#(=THIuh?dXS2JjvSErq~hd6?F`05DX
zKjK@4>SdEHInE!?cpQIjeNfSJ23{E4w#qjyQ<2E>_ESZ8@DxnSNH~A2w2}~E5;$Cx
zV8~X2T(XeCR4;!_tm}PS`0Y2_iWZnji+rk=Myl3)c-ef6vaXTZsoV74;bS)6!k=$b
zmX?j#&U+$~cua~MB>0Fzmnu4Fp1OjCz)8w4i-{mZY`qWa#Ny)6#CxEd(nTRS8`o?j
z`5h`6(IfWSdcnl&mkIOHA~KM{l5C=?^5_!L+RMw_qvgfNJ`tQ!bl}LG7j-zQuxrZ|
zoUirxZT&uoDPm`9v8j%%DVauhWAlX7s$WE~4i@xg0{_mown|yA6=**Pb*60k!G~<g
zCj@nNqUzqYW4u=*XU)VbJ!90om&D$bfWmB2_LnI%2Q%H)_%jvciG9jE{`5(4Ne1ie
z@T)tFBnaI~sxf1wpWz9OmwHaf`8v9VX1P^VBKB{Y1zaz|RY`zyo4SJU{ZbxdUsG)t
zxwZrmf>~eP@0QOK9+W}A^gQ!=$S(-Us7mqA7%AK*A|8~2$f)a?x5DTG!1p<zPr%VT
zN;3v?p_0g6{;0J+QaFOJHRHBFLrE33GsiKjzEo&TzRrvlb`cp<TkjY3{JSp8WzZ~}
z(NDqUzOTfE_5R^t?t-agDI#18Ojdnr81vvT8}$HeP07f7mxsAAYx=Sd99VxS>5~j(
zzWm&rK2hzJdGvBteOQ+Hxunah1P<AWN0as~2)PVl1JNKyNmd?X*w;;oY|fT3{GGv1
zVb3xm(6vA}CL=b!$M=)#v3UXMq390AhnkP815zWtK)~T0LCuWg@>lXjcx0I3sH93U
zJ@GZT6UUc!0C(nSeMZWG1Dgq8Cn8$*Q?^nzF~a0{fb*&rsi#!M1DJN^Ji^)i^^~zE
zfl=qVd(BA<D<A8BmiYIuTL3v!Ny(itgI@;nduCLZLA-?mKLMH-0NWCW8%R!kZT%{a
z$KFEd*eI@NvpsX{P@<2|(Fm;L&e>Cp`%p5DlaYo3v+g~bB5Ff@@vF3CPj3a~@zi`C
zhdwfZLASn6r}pX%tcvsR1v4%e*kj`pXr@%x-sG6t$iErTH~GDgR7+(m4K^3{A-Y_U
zcM+ij4_{&3At2zbC+8cwkQesc$3c7ih%$#I_5@^^`Sr_dwX+G8v16Cgjo~q0VqoQa
zyc%GbThk$=K$*F22hP9??h@dna<gM9Pm8QDKpz9~?q_9ys7R8cAj><1HjB-NLBG2L
zGE)>7nE<BC;xbegJOB8x#&SOItIWrF4N^8#RX^zNL7?E3&Na-1?|pJ-4TxSAqESXM
z{me;-K{a^NgD)}Y4iZRaU92O%zFN$xB}USGz1df64?$9R$CUa!1}$c1V^eI&T0tOX
zYe=w~ZGV|OPMDYEPpiV89rQQ*3~zQvFSv&+@auLao$mZvQWq>SIVFL(ueO!^acj4|
z$jyDKs$lye=cb>v`RjyH{^{&dr|fn2xVG>AIaoUV5dD;(N6u&;4_P;D$4k>gVEK-8
zM@RI^hPEyWtgr9h@ZYSH5wtz*HV3TU{Oe742(eRfWLUWf)Uxs3Kq7n4(uq7~8X#+-
znP${JLxsA!JqHW}rBfS*tdCALqdNmX^X^Gv_RASyJHD(2kYmO?KWNyV9c`k<m2Q(U
zX9qo}w+`tp2HVqR|NQDH0(+*0+o2`YEk=`1LA{gDT6qKSfvWlkuadXC<y+T^m-Gbk
zKGmIY@b@1;BL0Cm_*+1OJHEnAZmzphXjOu>jPLOGYcV~U?x^Rp>V4dpVi3UwPA1}>
z4#F}}T*AGOQ3hR9K&+1E2U#c}yrMOnE)kkv<b3xAUMBitEcS;pSNI)pe~2&t)^)#3
zM^*sqc2L*OXd?DO%uo}>c{fqPDEfzQKA7#nt5qO^m+E}Hg%LM&jL{`huu4l?)9Y!Q
z@4s>0O<#!9MYqXlL<KIrdpKk=TQ5r-eTK?KC<)_^LAe$56Snr)_#??n3cD>nqXbuk
znr(opi|H3En^O+0&={&ZEL|!u%HdT7zXlb`kgnx|aSmaPq*>SD4j>_L)@|Q&qUn4p
zouU11P=*}6jLdWD<azCt<R?GSzWA;umTXTf6HWVzEo+j(@8J^*Ct~_+00kwQQ2Fuj
zMFwy64@8>{kQ16dn$lk&P}>S>nA0ci*Ab`yY2UjsBUdr~Qknd-8lz|L>?xj<I!Bt3
z?f`M!RI{-Xv9MbhMjKh_PNy-)SXP^T{=zbh|2*&LRG6wOsIzbHCyU9i6!Oi^WsCs!
zRAXd7es$7uT(1^$m@YTPx2Qi<X@q65iYq%_0xZS>h>f6e_J!O1Od1N|XO?!XcJDkC
zc-BuNBzb34yu>!87x|`liRheB+&~Aet(-KqP^d7}^8$^GQ8kIaPY9AJ3z#W8GPB%y
z3foc1ix1sPCK}@W;=+~x;R0Mh|E(uw(!t0mV*tNs`u^+JugP?C#CtLGn{9sISrP=k
zttAVP*W4ICFN)imOB`GzUq9$G-tg5RFm{fjX``>LBZ<#1*&{rdl;G}Sa?~9oju(z(
zS|ycTsBLE=I%~MI@+HfU>%A?9GYpoUfRqV)m*RCv&2I(QSv8%}X#DiUxR+uNoUTrn
z$<)91w$|3w@origvAVe$8OM!yBVz}48aC{INZnzWXX|!Z<FQaC`>iQ)ZCw#XnfG|{
z`eOexHIXzmPij>}%oJ6DXhyMf=S*+I*&2P_!GN?{jZ)=l?kjy2n}Bja3_GNfj!uCV
zY?O>U6-#bF00hSFz?y&A4nstqKLfZQ*ootZ;^W)OFSOYAv!gIfIjZDQU;};%c#@~p
z-G{HHpM;&awLBEaGb}&gQz~$e=vVYMRDXC$wp<`70TyU-h!)##>ues+cYNm5mY@gi
z1+TpB$;D}k5xZ&u|EKyNv+@e#te6NrJ01JMlAYIY#pl%8Nyok(9t)gS8_+$k(IssA
z-1(8*Oc`(||E#`OeHzv7;9gz>cmK+jrFkE;!ONHNFMZ*#60CJi0nu`9em>}dU>~ES
zPNYrG${5`*TRAsKRKx!{ug#K7V@3YiAzt)o60ajGiSx#@#(n7mT2I!Li4d6J$P-Kz
ziQq?`AZo?(AK!PnkNZQj9a54Tyl3aNWA^+$LD5k8;5(%R^8?a>Y_spL3$(a1WeJDb
zYG)9geA%%TKooM7_2zhC?83w8ycFqRGFtSu=@j{iBS*M%&Xqj3`vO}4uXBJI?feek
zLK4%u`H|Mr{kxoxj^rqR2@=wDWzdwI29EhEH>(sxk@o>PoF1`Z>8+WPRqrvjU_u8(
z=Mvx?96g~-`;q7Vy!wyxy%L#A$Pzl&Jo%UJjVE?)Jbs*|tGfj_o&D`k6Dm(hthb|2
zvgN-ChZ_b~sM+5<!`<97N9Iu*pF;|W)n>p|ihlzgpaPe%P)82Eq26rVxXGAKE7wy-
z-SM6nl|7%(4U=~Vh`cB;^d{eLC_c+9f`*Y$y#E71n^PrNM2I()iy#s*OOFlG+Y;J;
z{lOpOjoq^U8zY)F(KkTPd|di++ANbbSiTc!&z{<GbyP+^GG<rIIx^;`SRHF*E>&uT
z|8w#MNfnh;5%%23-|RgRtLCZ*KDFBonmxkWsp4px9pnf82i%}2XHY9@vU)(*T(0S3
z%%;%I(=Rx-8~moz2^8D5_ZqRP<#brl+sFn|4Aj{plG0tn(?m^T+#}MN<=ZO&I`?KU
zLv}o9xun7gSw~<R45El*cR)DtVD_W>SSJ1VlU2p}y+L1#u+&=^q^tX-&vKaE5)9e*
zLyJNGG=9T&Ejv>Kod+;eA9Oy2&)ybPW@Db{+F>aE*uz)}?RXigi7?GG2#--Kvhmh%
zKrbWk36?H!ubi?YWUAd@1<-&*WId%#p?_b=%NjG3Hy^XKuG5lQ6~P&?V@kVs@(C5x
zt7_XuQ4<Fw?ky*~3v-yLatq|A)eBn31&gJLg?obc9FUeqDTyca&ISmvyN%qZrHM19
zA}lzT$4pA!00Jl<mcQCR)-h$OKTzCsb2wd$Nd;2K8iQb1`PW+KMaY)B$}xWzDc$it
zkWfFL#p|j80?1iP*%3$k(oEH9{*R1-v)KV>v$4ZxBn;kRD9J?H-ODrh=)c2o%);G)
zjfGWaBdsTWO%gU)%iM6H*NA4CH;v3XAt}^OcBunmiAKV-F&%d$5FYdeCH*Ye!pPxL
zP!ecPv-|E)i_T@?;>&rKS!b#>nqu#FZ7d%!KS6CU9|QzW5$#K2e@CDDAZ?t9^r5c4
zzE&4m=mLXb8rajW535zBH#jV(wRm(XNfxRV8x<neE7Su}t7E#{7co}P0HuX-pix=I
zX&bd+`SXx57C|kqea~g<MrqVd!Ne0TXh@vLA&2!rz3MA~V#e8OdD*(Cu?90=jU?HG
zOvfp8nQY&m{*hz_YA8eCL}A8+sW9_tn{9#X)6dixJ>1lvN?l;Z4!d)353G|O5Wi4z
zZchNI><<{cD|MK$+Udq~AV=L|tmaq{u|G{HpeUL2Z6sea(_D`9n8)>j`jp6jyQ`XB
zqtSnx%t_YMXN8KcJIG&1Z{SHks>wyv$Bqi~Kt_7dE{k)bZD^u;=m}>tYbOgQoh}};
zMp>35GuT)1`>JWR-K)8jP}N8OrZ<QIe3^(m%<Oi2smUA~bug!I-i2_6o3B{RH+IJD
z;(6E@Lhnv?X2(&lr$tJByB>`hn*g9kelRA+iK=HGWO3L<#NmaMrAex-;mw2`X1_|F
z7HU<qzR|`+7BgV!$uU$?*Lj2Q_-X+$x6w!aZWw-s*hcgNQ_lsBcOCi7rF9RcK_^;L
zoF{c2U^5axw<1s8q;UlGL4^4!=&L@*{Sftd2sffC!U65e=8&e?RRztblESn)*K<2r
z60P$0Q3Gc{LlW~9-ysgbb@+SOIh>wNbrF62jZEospMtkX4bJc`Cw4&ak~~p!sSf`U
zag0A5IqB~L?Fp=EVn9~c=EkiP1{73EvAS>upNMDNIDQ7oG0u{!1L;xmsFXz6&q+tL
zs!8(Iikn}-RbVBhZ@xJ{3Zw<7L_FmNpOTMp^-SPLijuOS`=}}mloWp;6cQ}gj>lsk
z|J9nc1Bl3^*Di}(G{V_sz2~_@-Er6ak3c?#`)nID==LIOHiFyfx4NpS2%k7S%(efN
z$DvrncAu7hpxmJL0nUWn{OiMNCBvV48<6!1ox}#DPuocWf&}FgzqNgzv9s8h!^AH>
zHi5w`Qdrtq3xjE{CHJYz-fCtt-q=pNjKR=xf37NXFag_rFs4d3;Qmx!aqIJ*3q!Ky
z0>7bMl~|K)_jEP_9q+<4J-BF$9{wPS5YjC9sOvE6b2q^nb^FcII`u3!&N0zgmIOy(
zxGWQ3Z%n{FS4kkf=bU_~4IlMSeq|F~TtEzLa`Ts-Iuzsm-Qa%i5M7#f5<u1}b8D>(
zdxnUTW5se;v8EODrRv5i?vqGkq!gnVKFdhWDw}K72BgfI<<EP-Y+#t68DycF>s@2F
z_K1T4+PU}P|5|65U>S6On5ff+vOR)}sILO43Wh?Hvs(R5?s#%%*`8u?-O7_Zt-S5m
z25UfPZmch!Zqwi42qYvX-%I?5<j}*lliaTyWg8-8vQ+%QML|%hj;GbdK_}b{4$p#}
zZ}ZscN1xo0ZaTUgUW8dRx!yytPI8acfeYOcQ(MP7=23WXDV;<f3M^t)wFLW2hr-*h
zQzYWF0A#>)4v?`K>_waUNKTx!$Q@ZL&#E^Dc(;Zd*kM0w-@-(nRc|LP@$b~MzWhix
z-CZXd5hatSlEAmJY;HUVH&de%C(XL@vO;#ov=Y!p&0kjNkNtTI$cPouB56il15-K5
z^o%x@*q|-*4JjSVxuii3{VmgSgjqYrN>O*!M!qTKJU6bCqq}X%_fy+XxJl_~W39U0
zPPH)h2wc^qK?56%OD?5p<OtJ1GYg6YuSbSL%ISME*H@|&g0p3Va;R!lu_rjE@*FbZ
z8%c7>(o}>8(#-Wrq8{voA75f8EfUCbz+(~oKF0WAHCgqco6sPuJ)|<RpIUBuCeD01
zwFcAIK(ukAq$lvfOE)H1pyyUPxz?*B-m#B18S<@ZW}Ul0R_jkjf(6P`o*slr@;@&O
zp9CY&?Z=pYF*pRnCo!A8YC?`frcntFJV|+78QOHmzHiPRD5*B+hB?Hz+o3y#esPuo
z;`%_?P+?PjAasYpwaaD!biDM=FEwU9<hc1^+!W(-&r8f$pkiF)v+YeWeul)-VsnsB
zARz+*!@{??FzEhu(d!a*hsKvU*Be+@ICnmQIUTHgb3LWtsn4lveOvi?j_%cLbyX+g
z3N+)5?;`>ts6WJ&z!d2Xy(JrvzXJI_1uT&-;-C0A3g!Dp4#gi2u(&?1V2C<I<n>L@
zLzo_dP%U`^rVmKoWts4*EkL9Z@_IjASAF!!2H@|Xt_Y**V(H^Ts9tB}oX{u=fs4QV
zr#vLPk@PtMT!jo(5wvn??O?A{y5m~gs3$z)d%dLhuIX|vD4SEYsyzs_#gBsx7AFdU
z_y?<BP}tKr_k32)v*8eBF!=D8Y=3LKoSs@P(H}^LJZ04WfCOVLieoc=uP$=v_jO}2
zr@1sKYDyN~rQ;FQ_VYv9AWyCb{DJP}&z<BYTjb8qA<*-5mWOlORfvO8t>}%X;l+v0
zu^317=~4nQGHf8+pB0;0<`=@N-1UZ^$G(IXzZN<lv%fc)|6LgL-q(2MW%r}HW0E0R
zLMZ)!cAr#ixsreHKI^_5k=c445&a2Ms(ydra)>qW62)8CA^QGe!_zoF<v!Eu8NhLO
z9d9Sm@th9WVYuvemdsTx^ff3;<}i<Ej8oIw0VE~}l>3sj{wd@Y!MZOR;q!OuNv_Im
z#@vX8iR48YX=zWHt$+#lCizSQvhSqp@l^L@BF^*i)|cbR2i2gP@Srp#g)*9QdNIkw
zPC%hrLXro}#0+sLmS9It8xQ+EaZ_`&A*n=yl`zSTBGP+GHSRqiPl$0SRLv<~a=i-g
zeWf%L`4JloPp|b<)3%h_o7p8(ES_D0rYNLAF<zP*o|s!Ro50)ha`*&D&SRSn=P9()
zE$14%iCI?wBt6FUn?`1ZmLeV!CNMEFxGp2e#7Lze+x^0T<99P*{o*E|PQ0M4izvN#
zrfO%3E^fUwU0+)vZEtN|lK)Ip`Yj5KNsAWxhJzje{=FV>1c?XiJh$kKKQBOMLV0bD
zZdxQz#{TB>ol>5O`L=<6SwWCRCe?RTFDC_oM)7wLKFJbd>vk?dn)W{K5GoiaHoJqm
z$(Q`A)WDeQ#38(LSn<nYS<F&MR}0AK;Q6DHbD{cyF<0TV#P$$Eo->}f1fS)}NiEqG
z?0`;Zt}Yq+z2dy^<Ji9p&w+B*_Zwh=n}$rlpLFt9kIVZTT%9G`E9eXB6xII>RbuCY
zA&l2))|gH%pZilKOqsWGrDodzGJ<?rl6?%6WAk2a()j@{QFAg<3Hfd(nG0o4ePm<4
zpQgi6pJprCSg>#bY)^M)Jy{EJA%3Ljtw<-k8b^GVVtOD7YdS0WLV<3hP7e+f08mu0
zPk&7m8YqIt%q^E4#m2@i=Li`VW^XCoPFbj79zm^^>pIjexU}wM*F|9Ol4p+70#>+%
zKM$C=_48p57C?(*fc3%SDQ5#B^5xEY30Alq6qSbkgYK^pZ<QWNyElBw3=hW7lh0Y!
zir6lcq7n9I<vl;5Iu;Aa9V<09O@PnSs#ZKNZ&iq-mbKa}A9+7laOGl02mVeyC|4Pi
z7VEyaEjq2xdv))BF^G)b*K8+EKpqX)lyA?p&uf-yWuMs>;X-B}s}e3iwCh+kR2=X&
zzBEReO(DO`Nuu}c?keB>f%v2qnDUD1&o!HnNkm9sb<>UXS?-L#FvoJ3tjEt#&=j4I
ze8u0-FVX3$)QJ4~tMhN5t;9~CS}+1Me+K(>EuG0_P<lY8?;RmRr)xDpb}D{V%Qltw
z@Z(@*;m1xs0h`mZ`B1@}_88KSa|Tkwx)Hi+#2(ueAcE8ltDEMZY@bjmN@Um&pB>|Q
zh?#to2lJB7A0)X?FH3U2S}4uZg4M&q6*(V)FjUFg(rwdtK)<{JMyN05I?jndnT=EI
zQe2ns6N{I%z3?%nF~(o#A_MvW$n3*p5Z<yvB{<9NG!jYy5pBkR@xOMc;Dvw6tx3+x
z0R9?5Wzq2W9U_%FvqS{RrbY1IYh66KcyG#6&o@xDT7;mU3?W4~-lqkiOKvTLu>wtG
z$qX1^8Nq{+i1Cpg7h7>x&&8f_NgWXb^ZxdDd9Gl4_}vpJI{-%dUf9sZS4BDIg-Blu
zyQeMS>4xqdD})PE4wkh>wZ=-S==Aq+R`hYrKahQd88>^e50e06FM8%LxsvrrQ}8F0
zO51Q-<j1IPD`;fOKe+9=%DEXY8^9z2qbClR;q8Z2qy9r5j97YCnSm{ExLUy>bU13h
z3cBu^2tfSkt18BWegF<9_zlqS(En9?p0e?ViARxV?1z?}_eOszq8?vl>DWsdaK$lO
z-oFGairMd0k{~8vg4PbVUe*}q?I+j;8QAjNY@ZJF!*9OwkXdR=*(sv^8R7OgGXSe}
z4s7yaid6$rp=9dgbV^wmnvsiUf6wstAEkI>hAm%$&EoH`1NHUev3wH-scsOnr_^u}
z#g`x}Xa+rB7PC<n-phl*w&z%OgzrJD6OsM031x*IwE$1IJ3h1ecoVb4p(}j((`aW;
zPYaOkWC6R3WUY0;(D^1?W}0+gL(@~<3H0P0Rgg)#wQqQWGl2$RYwgXsSZ@z643Gqu
z!@e1aKtJy0i;NuH+fdZ~9@+$q&1%BDau#q+``ek+qXFI17bw3-Olfhwx%u}upZsPb
zRTvIzg8TsxUVpHM3<Ju^A?7*I%p$)HxTK!3<U8R0`Ua4*@kGE;hlhrXp0PJZ@DiZF
zs2mTu?SauoNCNQXrIu}99EjL>fL+Zr+gvKyw>QuQ_~pK!l>QW9g9LjIt_5#`oQ$Wu
zeu0z|z}o%A!4ai?7~(Rz{Zt81HAl^Y2TcFAuZ6^O*T^27N6+q$BAmE-()r=KAE4Qj
zQBk#}r!E!!j~C>DkcX6~U5m##?LaYHsQVcd-TDii;gWK46*=uQkV*iQod3q??Cov)
z^-W}V(gp@Cl0Vt&VeT(`H<(2rfcQ8T^f|<lpuhQr0emv)VJDjkaJt11w1kY4BxXHb
z^&$0$?Be=r`?|=e_wfW~uT0-?=Tlxywm=7DkXK3cp+lG|nEEOW!t)Y=9QrO65)HVs
z&}p<QZhHJT@81u)v%a;J6ifaAwEkw$b_)fY>E8|Js+VR|+_eNfXT1+KiP$%v`JcqB
zmDdxzgSSl{cOMD3@dC|jFEBRu4a~tl6R?yEyRqyP1I$<fcmA!qeYv#E%v57o(j!g_
zBB>ZA?Gn<~9Wz?qj-5H{_cxzF5LjRfz>xn6A7GL5E-x%V5RwL-ad@JbLxr=vb%g}r
z-1F&R(#p<qlgExB12dL}%y_lgO6}I@zEJ~~+&Aj(qkmsTA{HpWGZ)mDqmdv%ZD@fo
z*lzP(cD-NXNgKE+;}*H2@#1W5Pv-U|1HFQyt*IbjDJqd)7gq)|ABu-O_9-hX!+Pc}
z7jYZMj*ul;v%yhmoL@;z7f1J=^PjCylvzx@IHJ-0WF$FTjljr}c{56OC+WQ!@}kP6
zDBkTyQ~<ta^8gh_&RwY6NCJ^3fue;}7k4ua4XEDWRaZlsOHrtzA!~$s#s8`(kluv{
ztc9ONx1SMwe`)~{119O4uBL<qS(^?4VFpr#=b!}bqMmLmqTB!7-|ufeH-X|6Vyc)X
z5Oldqd?pEj1j|jm)BgscrZxnyfXCl?yI$3KMx&v{X)#Lm7r={K+uEufeiOxOgU_;w
zrSAwCkiP|7Tlj3XT;*2QbOCmEqB?3ox)d1`zmV5TSC5#53SP+-pZ}&!L$mKgYEjNf
zQVI929?S20K8(SdyADM*)0Jvx7k@~dcMF-;`;8{gpK9iYzO7v?&D3w}fyEuYd%jdp
zAN^;!W9#86m&lkJSBJ~{`3#ebUxHt<)oVM$Y+f0>?ZHp}bBwGn!cAO(D+FhJwftUd
z=NGmlgK|yE2*s)<%W(lA1uF`e_^0L#TBx&Ll<JK68PUbf$2Vn1`Yf{4+`tzXIK8jF
zq~^CBBE>|rU=Sv{L6`_Fxi@chdm^96fGPd>rLX<h^}$FSR4L2L8xR%?zd2kRT7lBU
zmxV-ob1S)p4c#(U#U)LNw=nHr=MEqx-A)T3;>H3+hcGOO@03@UT;*#dCGbMe2noT4
zj>o_oKpQRyB-qiscc|X910Qc9w+Vz^+8+_p2!bT~q!Ek|5|;<KeedodjIvNM^AILv
z$~4zphcNe>;>tVRxB4jY0uO!bb1Kg!&j+J7u#Jz5Oe_0U$xW-8(aoh!k}Br?k3-IQ
zeP8D!tEWZyU%Ja&Vys^sXE=E=tU}4}WejEC{H_Cey;T3xj5PvcJ|`Rs@!%$jC$2bO
z7F=4;IRAVPVw#H?AVX?43C{ibQ%2J?IX*rxI5-&NSD-vL7@<app8^4@DgwkRRVGwK
zv77jh!4Hg;h`+)N1mN2Xi<AX~P9yIy4}rU<gUJ_{QYB4Y3K=p4E&vHz87-zzfUc$g
z_9fT~y_`qy?L?uUUO<YG{Q#D!R2YCKBK}?Lt5fuTUl&*C(7cZL*$5j$PrLP9dN3T4
z|CI$#(?|x5z!&w2`+asETMpa-THf4m=#)ImL;}9OudfSRjagJFR<{5<=yC+~j9VP0
z8r=n;)|d<Mlx{``HtyHfD&j`Dq&~))RphT@|3&oe!L0d4pY3ec?b9_wZK@XEXtcj^
z{SLAkFcxE9BxD>atBuiT(D255$MUwy3~;GB0R3bEXz#=S`oyBp6C)Yf3Ew<183yfa
z^$$MR4@V)DWc<(it+(RNkKdkl8W$O#>h+WqYc)wZ_;z@G-EU|cSVr_i5Yfw4G-2Cv
z{gSWTQN{-p&9cPoEGFGarIhnk(yWe(Q%6twuZvD2ALbG2A&+2#hiMeWNh0>m%h{3X
z8Y6pyI0QtDJn**Y0B2A(b`v~SRUCN1oQ3Ee;<qoD8hXLhMP||eyx^uU;1Xful59$B
z0*#swz{GnCRAM#0jp#U|f!P)jEiIzr|MiJSK`e*v)|>R9fE;?_-t}vdw)^-eu&MbU
z0WE$z+fN9?9p5A$)&365B<Ey<?-Az>Nu2M{ZlfFUD#l5=OEeW8>f?1jztVS*;%7f>
z%lTbs+f7o2-f$exXSpdBzN?QWl&HDJzC(XKCiz~yidhpe$Yah6!3?C|hj&!G_NxQ;
zfLxX!s6E1>qnC<ec9erYe}>XY((8~Yfcv7V0*2y$hAt*WfDKHQ;t_#+>Bf$~;6MRd
zQ6+vK^7#M$<quJ0Q_b7BbZ&0vP(d)0)kWlTdh$?V>8FO9*Ho=FEgs*cWB`6buZp~w
z17z#v*zI~f!P!6V!FepbLX$YGP`z`PLoZ{;ER;KJBk^$iZl0?#(LiGo=D~~NyT@Qq
zd!rE#yx;bWsGW_XF@NRSxj_0Ia0Px-Oqar3XPYEea6spLi?|7r);l4}z^MP{fSP0z
ziRZlTr2nM0^oFwDyZ`$%R)T=oLoBF|`+xuFKR&5H21uJR;d(6~e`%sZASdHhwHP3h
zMMM`F&sIh;`WhMMS&psK5xx3aOSsv!g#{}~EK!z8ljOs+XgtY7yS{Tg>EtIwzY*lJ
zb+GXty5na4&Qh-uC+G+=a0X8utZuMO{TW?cV=+ez#b9eDXoe=ZEp*!mv0OZFgC7cW
zBi_KhAEobXU}!zc1rJN1l_?7zRx0!mdRR};#8y)i>jRlB^!cDk)4%5c{oVtC2m<jX
zM%4fSa2w73%Xe~sh1rO*lww22VTI<I`B+7`T3NuymRB68=hf2dIH9{UKd_v6`&BKL
zkP|EEda8^BNX&Squ^OT}3Di&!qaq%6^nYCqmH7DAc<+5GMOy8!-xeMlx4jro+tI@L
zP(EyKCDcgsz;HRZ{`ya4F|*z=^FDDQ8hs7K*uF0RwzI*aLd(m8fZzTexG&PkyFo~e
zN}ms0J8Gn$6v0yox4C^wR+7MGi9hMH|NHt3fz9Rx2j2U)Hu=vFoT3IL)p#2lU!W5S
z3=#Zb=+HM>iP$5->s3!5gEuibHC-IO!%!C?M;o+fyN>@rfM$M^kyGm*E&$iC+@l~y
zwu~1^ivMb0apj|hXHl{3W``u4mcN`Gd-8#F?DRySws>s_^TfD|IdsGacM9SMT%Uog
z?@=Kz=ufEn>?p9e-s@%WP+)Rh#Uvo7W&q|+gqqvQed~r?(BuUAkK)JwTrVE1;3iTL
zZ~OE&H@b;8|KZp23?Me(quSZs<Z*4{DOIdc@vCDkGV9oeMKo1STwbv1TXk}BA<j#Z
z=FWyquk+YwDoZHfb<Rj@TZ?WRE&F@CvPzKeNYd(JMlqYss(LlskcA4Cl;-uXePDiV
zXNMEs|2hCXkA?zxp4$I!&x0TY-kFz2N6x<=`7S6Q+2C&`N&oZPy0O;DSK#E%NBk*n
z@(+0|Bne+*3S@cPpDM!nm0vYf?b`^<U!-V#t&wD~MZr>3M5*cQ?QRZext{yeZaezS
z7|-sK_f@%beq#Sg{k#-|G71>XdDY{asQ3+TZVvlTSMSi-#PI}*SsXX+xe8<DP_p`4
zwps4Wb%Lst8Ei?q*A*#iv*zx4ZpletRAAC>#@M!+0L|%GOz(ZX%1FYqvjgb~1qY6A
z>hl7svlr8sVH(-i;aCy_m3l+5m}EilmWY|U19<E&bV5)LVItOX)~C<hSt%+t_4t2P
z&wrH|TEb5qE*~-72WNeE_hr|3A+KY##s$e(1Dbnpa*%mxgB1?!>|Vh0%D?%cd9t4z
z-N#F^=Od44TnUGO(}XPmXyzHv;WC?Pn0kxarn&i(4b}>n-k`(pGe<0m<ea%fBpmzg
zePFXqmShPra3>Q=;7*TKL-eU{D=RXnvLch=xJ`nfEKUN%M*@t0U)6v9#UD2iqNW<f
ziO3)u^^OqaV&(OF?NBh1O0%Zrj;ctn#1O^@zqjLHbx=q=deDW4@lJh_C6c4EHz~CJ
z67k!lx~=SaU)Rs3vw->Cj^#+@8Wgp583&z%pQ%*`L-q;{7MGvSxH~@2R(@c9d(Npl
zzEycTNFiTu-toyv!ahxZ_Rp2c#ZDxEHB9bDe0^wn3Qw==NP;+=UtEuFg!^7+=-y$?
zOu*~DtZzj7euI^NKEtRbc$Wa55f^6aZ?j$VB#Mv|%%^{%O3rhC_~&!>ledE%+nUkr
zbn;q&LB^xOv>f$aLB^I}b2u2s^EQo4OJCg9`fi<;kwH#DA(EcO)6IpO^h6C+i67>U
zexn#^)lt{{EU==*0}xJ;Pk91OI;L%#PVUdHgqTO;zy;Hv(@!l}q>Q;XeOL8l7jokQ
zJmh2n(_5c!C4yktYN|!=zYdWy3Vg(-nE5W^fAjW#{#*(;J^fofGAJVY6TtOH%&Hr7
zi9h7`MOl6>AIWb^PpKYz%%JNXmFh>Kzb$Zg1SUB?$-hkhg{Ri7VT0(;PIt3}eFh>i
zg;Dc9?DN#;zxpbmJmPPB@)>b;=4WeQjPKam@s~CB{`akqzTm~FjD#v&s}NMdhl~w3
zk*kgN#hU6k5<i|CM5JA$Kl}V^zMnOa*?m{tf9G+aNJ)V1$q9y}uO6<D+Klm50K>+~
zmT8`&O_^cFf_z(D+)kbQ=6ih$yp)Bysb|Ng4DKKNSqJ8p#AnqVvxQ%?JZ}5_>t|xV
zcckF+c3dChNdJ_b#D~7kHeAdERH^1Du(frqda7SX+$qD0f_C>=&V}9_`mG&?NGUp?
zhdwdAO>bKfpwkU7T%i1Wx=qRe)*Dmq`SIiS`2Xv74_c50JV1xe_Nh8TGZ}C4@E+Fh
z#(aD%#n<ey<crY-D#XaNwD+uRN=vNFj^!No_AWfOHFn@RTv&c=8!Ky;iO*lbE9AT{
z|AQx?AjKZR@%geGK0W8F!qZ2D!J<ngo4L$IYvq_fLiP<d1<5T!j=n-yWtdWinUJnu
z<z>LBtCL1|euEcJZcnetuUclT*ZYyeXGVb;J1&hgFM#nMUpJrt<0nfi*%uSh6^}yI
z<ifhBhd*9e)H{X==$&R6!?%lN=}l*SZCrx+n5l)SROh7b3%8ng9n@LO5k?=1*)dr_
zlW#G!W%q#5qcPsCz(a@Yf20Zqg2_FbZTKIdz-FtIzT|O$S|q>HK-Kc_5s)yx*i)By
z%9|Um{|~cA9+GvEidFOfC*=_I1o<_+pdRMGclCe%g%vV5V<vSZ^c$`bxhN^<p}MU^
zB|MiH^lS;Udmr^l*KlLPYmp-9$8?l5e4mnNmywrW>c?`nz%SUCdrjl)qnG=<8)uG5
z(bd&5y`}Q`Lxx$?4cOhoFHk@1B3~Wyl5D6KBLw;J8qs+?ewok^lE`(yJkWTFwW<pf
zZJ%!ZaF)I)pW31v|EpT>cPUqqg{-6sbJz}p-g0_j!%Z@9ut-mzi=MK$Uznni%W<#m
zda4c+_3U_H6yy{*NgzviR%LA23m0Kv^UEy=_+rYp2Gm5H6fnosSxKIZmk^L+ZfiFn
zAaWKA&B>Jrs@A44(s7vhdoMJH+)HB&Pbz2VqUbH9^rKq}84vG8bK&F(<H9v6srI~x
zREsnf4Y}6xd5^E_3n{3iyr?*5W+3Dr-J7(A=R|?ok@F!Krh?b@qijqx7FkqZ3*1S_
z07*D~0|0~590j=e%$>Z7&NEYVA7ge4hJONmAV3P$F>!+aEpXHz+(}xL;h(Vbp94Rr
zWg)DcUHXS7<THYzNLAVn95|2oNQlN7a827AKY?A{&li<y_tCg2m+QgZ)?XNml%i?u
z7CO`3V4}tuc0sFn*xlc<3yaWgZs7OyHeS_Y<?SJ&@~W_N;FY{KqE@MUF7wU|ca2jp
zh>SvEt!!{Vcb!!UsMz2qnd&;j_pKk^iRk%4-+Z_eEwHoR<VF_t6;{8m^Lz7yG3Ezu
zn`@ms=WWI`Sdv9e!;nU$P|NvQ>Hz`!MYq8d?qGJydf$UWV{i5)_`pGuhe!G90iZC6
z59}%VAvAyGCBKzd<5ps?W5>bax_5WD@i=33a;iC`xpZKW=A#ef98()(id9IrryOwq
z?<ucaD_D=GgHs;HNgCoR#`&AK{Xg@b#}J}G`d@?1Etnwq5#m2+uebF>EoH$Xgh>Y8
zWZS6_;Qddq2Bl7AU$ag!xmsX|GL*f@sW1rvWyEmFpEDA5#AE%_AQ7}SL$-HpG;Hwe
z)5LxPjVG-+6JZzAQV#s7vWU;!48oHAogF5UV`nlm-yOZ5@f_N8qY=eAG@8MsRiWG=
zo^embP2F*E7aUW6^fG!^Aj8eu7JCM%D?Y&xBqHO|gt8=#D2A{X6P7aP?$>CI{9_#~
z9KyK!8O^1_Mz-5;78C1tb~D2@vP))8?5|%u2}*Km4J%Ek0tHN0cFMHm^T6!2QvI$u
za*U7<K#z}71BmW<ecRE}iisRu1x@yG4Mhi)`Bufh)`tNFZge=3+<&|4CLnOfdH%l{
zX}=gKYd+@VOGrbt{|~Hx<PeH_5%P4aU)nq;=6dihQ-$_vOtIy|=+6%KomH9G^PR_H
z?G>`$VmHo6M^E4~^BL2(lo$wESiRDvjN(r(V*HHAv+Np_4tvNn&=^;Cy1f*l_Ny@_
zLWT$>MWwm)p#Ny|vwXuA_r0k_fY!{k*L!cyxV+=`j)paUdY?az^JNfLn1%{-@4Y5L
z+gkNXd)<v-;5gSwcEPbEwsA45DZQj}k5UfSu*=a(Lyd<6v>$LxXB>%v?-(i9|2Zx%
z*b3a=cI~v>;3nuO-{^*rBF&!(UwPE8Nt$w0b(#v2muY*y-R4OWjU}F<dYpg~oLAQR
zjUFdN@BNAqEV@~Y^znRPbLq?H@(z87BD36DF+mZ7Og_ftK+G<$Js|z>{YWYafhClJ
zDt{XZo>(#<EtKAwvswFZapCsQ6{Nwz|7Idal!FdF6bT5qmKr#kpS5D7DP<|KFu&3v
zsc=2Kr&7<+e5A$f_V$6Csbhd{eJamDv6YoO(Nx^lv5m;P!Fi!GPPXE?^tTn~b{axr
z4h<k9$N0trr6h!HAEd;!!lk5@^P_bcnO|YQr)nuZe%b|iH<--Ii_fLU#(KKthIa*b
zQ}9RAH)ev|+M9(Qj-gYe#?EkCqDS*{76A}7(2;F+Y>TL`e%<Z&x*N@0WCTjPlfgO1
z#YxkqFCltW0HKfZ#J9&<J|;M?y%L_8N__g8esum#KPZ`qi>DC5!5sp}|NrOU_POmR
zz0NAH@IiHOZ$E;eIAZE_+vkW)F?ndw==sVNen9rSdz*2FA{-GMApI!1@~SKb02$E-
zJH?k`|M)uw6f0Neb}9s(HNJv*#AU5dS>t<u5Vj6fc~ycVeSaEc1&k&)^PcOAUA=96
zPZ~w4U)U;Cs1P@jC)H_f#GbQ9e<qqU$S!}8sqI3*<Z|&Y?MtEBFL9O+#uAUwm46bJ
zN5*^&>3vQ8cIvvFzVzAq4hH(V3E=T6I!nKBg2kV)KW8(hpQZfk@#ME2Pt+$Z-er;i
zJpRS&|Kag)kmbSow~dOvuT_wIv2E0Xxv+I{uOo9tF*sF!q)m);@w-I0%j~6~@GMj8
zxSDn82fen0P7U6gD&KZDfCA`wgS4ZgQ`q|P|8A;Op#Vd{s$5890Z|s2=Fjj0m~8T+
zK(@lFc1zo<sk*!STlGWzfy!AkcNl$xMY9Lxn6sGS)}JK12hYqn4He2#jIo}NDqI|m
zBW;{q5lUS0WT%nyO4pukR2Y$0Za+F0a}OC+bYP;S9@skKE(O9!Hd(HlXeL&#lF7%C
z#pI@0-obIWUNl554kLW5Ur4TYe-i!K<-T(?bH!9f_ybGI3h5SLr>S;j4n+Ip74DqU
zj_6>dv=Z{^x}V=@9|4~>G3rxke)-p@MgRJAVoY#EKL;>P+yCj)2;j=+M!*Ov2935h
zq?tBt04MFvS(#8`P=KJyEshs42$c{>o`9fqad|wWR$)dE&7l5CIalQ^XzZ$io~F>X
zVfPoI`SuWnOgZA`meWM|+8;3=bC~%rEHvlc$0T6jA@gLx-~zjk@Qwkw02Pl(%m3iu
zW8(>E9r5?^pPM~59ex1Rxd!??0Sn3dEu~om!UeK{fx#QFDM78y;Tc%Il+(5Z_N)5>
zNfaEMgZ50VRLwGzZ<o0m<#cm%bN-PC-{U=alQ%|-MQS;E8o@B{d0XRlX%|(D7ThTf
z4B}@h@L8hYXaM{Z%#s1hBn}7N%AhL=HZ+77Onlv6db4T+Ps}WzFmmk|h{m*&+ju!q
zSaczFomXdPAl>1mkd;NptNUvk`S+F!JkW35*F1CxyXuw6b9fi~mKmxemtK7q%+49w
z3rg^x!AaWRT}<i}KHDIGG@0-5Ch&9Pu)oj>s5(IqXeV&K{_!{G!BH7)l#G<PpE^9G
zeK&bK#w*-7gwnk<xc}x=CrVD(wa;@{{K$)&N?AQ?v948a*5HTw9nIrjLi3>GAs&x;
zUi#|xjd#7^!Av^TGu}e}6NB&R2*a@(Sg}j#H*b1}f9GA4$4R6+X;05x*6Vv$I`{G}
zanyE0FYo*w$+G4Cwc(Z`o_o;{F|a2ssSA_~^7lR3^gX@??YWakJv#)#cK0gbGdU5#
zDw}|JJ`d@}2u?4zHBc^&$co;^*H@#o<WApF?8X_^<i?C0qrbkUiMgvGfc|^?fl?Mc
zWmMm`rUg4ULDMX8Q}niHy5$1u)Hkj`n+Dda7pix9w|bYgb6CS-t*!HIJh3G_TnRmF
zUwbZYFI->035|HH`fbVVV}p8rPk5upJNrAD?nQL!C(hhY?vto6!y^8Aw7f79;AtE%
zJ@;8&(XO=$1!C1)t0m^3ya4&t11P`pc<c%E`2P&P5@@@=oVk9-`gJN*uPtx|kRG6J
ziL6KHaYL(KCD|Pq-!V~=WBalt8~S*W<|3f!wSneXU|CtDdr&`nk^8lqxy6jzdH&l$
zat%PgKm+`k%87EtQ^1^?t648#V{mT6RsECtnO_+2Y-J`QW@hXr%rhFdfpjn6KluUa
zUiy#@IY4_r{|0o7PXHb4LGH_+$eNjcZWfmHhj<2jn?&5)+!~XwVp)VVJ18W6U*l1j
zLS2_hz<CquX!!c9)U1p2Pe^Y7HjN%w<m_f$O6Li&GI_2+JNL|E1UdJJ!(!4O%=4>9
z%1}6706Ga!46&8bnfZP>jr9~A@D}_gEE|buTq2X-u7Ga+$d+<uvQBKVBvoh~jK<nk
zr_?A!D(Vw!34t`=R+cYE3LZ;phLl(-VTRe5ogJ*k^4J$tI0k5SEbQPe@Yt?-14Sz`
zT3XtiCA^I{AqA@{u$CW3U1yXSv~V^{KYyTcPNwh;f{h`s&xyzraCf1YVgWS_oG-bY
zpbD@^x(SR=`5k9`2gMpn=KSuPiM)4er`=YyuNj`s?47F0=PMI7o;j;B`~Y~w_Awhz
zN&5Vcr%Fn|b6HPGrjKZrj#mRf7bWOGX*;^M%&Zx?>l9sVxDl~vE`53DxH<WvThXB}
z=4++PF0xxP%kbxrR-^q)UBT+owX^+0ha%%PEC!OjSh91$9wscO!wO<t@)k$TE~Wa_
z$cIH&^_(qO``=#JZ<GN?1UPg@W@<ceych9L>BYYM%>pox%Xb)r-=E`q491^q%Il5R
z`d43D+{FvNBf3!TtYnE?=8BH-Y=~Ydp*-9XNsSNr#LcRKa`FtB^aqXCrzvpLY~5#a
zgv+w3IOH{yVI6V3hD|#PzEm)?yqKs%rMAWpm8X2EA}B8;_{kqZ>@KFl4+(L>a8hzQ
zA=eX8L7`unCs#;VRFZO1=;WlIJm3OJg(#5{?5~!OUevPj4NVR4O_lQ1R`$&jO=Z_8
z9}aD<B~4uo9a9t?3hdbo!6zZ_Se(#VX@AfkQY^eHz<6J!5kZUpotCZ=Mdq@pQU1FN
zRT!M+OfSN2FCLrgVnrP*o4xv1u+25WQ!Dv)C7}mykuq4FTA|z*c3=Mwcbc`ItnmY0
zzh1>%L3Pg}C9o5)#d^%)7g*7RA$$)gn)@MRQ~v=Z8J2sFYiPTGEYuEM5z=RrF8Dr+
z#2ZOMArT58ml(K<?oMt+J#i5cpMcZQ307o{-sM%zQ9=f^D>2ZP{zPmccs^BYpK7~4
z%;kDk#$w0PRH~7HMJ4qHNLgiR*H--I?R@qTIDi9~T-yc25UKM1+~2^TsE&cZ4X5aT
z@Tkl8;N%3ujRjTQ6ACMD0)eWJ*j>SVA`6UyrqYCCPKrNco5y(dbxPZp;1+5H@-&!)
z>-$Vza(A5g9f3Bfpr6`FO{``$Ugv2dx+2+%0Mk~oS4Td=9LL+#JH&WG9+xRXUJv$-
z!6o&G<7W-F>Zt$s$4}v5vo+RRk2p8`0Xqeb?;P)`d=s(P9)Xyi9{EZGoB8<j#mtXZ
zIG6UtdTm;qDkK>6O(>o8O--eBDq4}yse4YQ2v*&JX{2Ptv1c95Nq%w6%GZsYyu&-_
zO%i+LUX|_AmRFswT^m>*IrF2C9`rds!bblOQ|}#5b=>}sbF4`A$ev{rvUg_L64_*r
z?7a_(%xuCj5@pNYN3zF}S%_nAPTBkS(tY2b@9&@E@hCmwocHT`UDxw^j(6;)`}1YF
zW};P;G#64dT6~olh?Td2o_oNJiZn#@Czoz!+H0-4^}A6`*Vr<<0%K#S!Ee8Q-sGR|
zs~L)Mjm@eBt4>aeio)^D%ceCozPn{P<4dYq7X+n#BUY<(Ii~IB!1~5tdEa-c4_gAd
z9_xvObzaAR7kH|o(PzYlwyiucF;X^ee#doRi*f!t-y8ctnMyQ`{eCAGEZ>Kezl0eu
zRy>_B4PtDP(=c}~7^>XgS7TIDpaNGti$+Y@9H}N-;YZU!&}AAcL#rvbuD}o=LZHG~
z;p3EXE_q6@aB)NUazee`#B_wK*wY4SAbr#L5WH{yvc!?^L2TyE`znG+YveBrXxC#&
z41`6qtE7`2BR-kA&-$a6tn|#_n^!U*m|;Z7XRbxlyCWl=?>*)LJC&?zgxi>`0`37A
zJ^Vl?NaX2gD+c-+U_-px`WgqNwwwbftRO-{Od^sXq4#6xOgIkVR4ffXDsXFpqnpmZ
z>+MhnaiZ!nP-?}0i?~ZO5-uADHp}YJt6lCD<`a-1#%^x%NMqeQ{+(q+It18&XF(p0
z2QMkuPnA*K(zK(=Ys`dGHrb4!wj%p30^rHq=kfQGH>VO?2#DD+)vUS!>2awSZ+|1$
zRz$&u-sPjO+q<r5YhJvH<!4yHEjU#we(%r9o`bFp0#@EQy<YbX?qO==7O)loa(M#*
z-QIV*ydec6c*7j}Bm(aj>D(GSNFSF6`4Da>q}&66px#3H!h(+<U+=x^s)m%weg;`<
zV!F>fWd5ZFHs0n{cP%Vt3ViXuN!LD3>?I<m`ofyV#)8qC&I>=;k2_GUWft03+O?=;
z$#;GDT+VgxN7GFgbTb)3BEW}%z>H<cypW+i-(mTEw{A+caw%<%kcD~d#&v$8qY?=}
z>=&)Kf*nTClekjC-&ImBwuE#nid1Byr5?TE7o>04?19&K4$MC1r|r}Yuo=<L)@7tU
zh=$t6#a7+*m41ZN_)&-`B)M)vwfRmN9rn%#hdEQIP9(-35$fp|(~KKuhZSl#4mEJW
z&JIm(v;pBn53ea|y12z+MR<YjVZJRfD$t!qs}-AwE|)IHd~GELzPhl9D>8m0W=8zy
zK3@)6+A$=aS7d7B-ocd{uhtel@~uwKGWjGUJUbgU9rwsNH~nDXX-<D!>y(GAmD(Eh
zoF^C2x(Bt6KPPG$A?yF_m^xAZWM&c0LCW^$c{@RiQ|pNFfAd~L$@{|bss=rLdBs7|
zXmZv__yDynCSN-(lbwG7N-JRU9}F3w>&cp-`)lXS?QG;~f6=dmLAuYv+%OY%X6jVk
z=!D!&SbcQPZF|%og9j+Al@fH~K*xH&-(l+j>n;U5s!Xr?uJ7JdIn6^&n3F?!?8Co6
zqR2rcv{~@Nib+)a0w@4K0*Ri1Ea~b<6fp~aPB$T)_#1Q#Y;zEj@?p8)s7(7Kc8Okj
zw2^bugHrlBP?Pz9=<wrgn(RK%<$cJ<`^dV8xFY-6?8OCQNk6S*m@-wtfFaaW8ScOf
zm&AU}5~DVMuRvh>o_E4xr7DVnaQ?ku6>2N8U{vd_yj%A4OxSjW_Wk%*)2A95J%Ha+
zxU=k^$Jx!az6WTB)?!M7(pj|*lX%aI)X+`jjophOhq33tgmV!%d%q<u+>U`^ZTN{)
zl4HXTW}<OK%*10~fm>nVMEsmUMR83+#kCE4`U8PXo;<JJknV<mu80f$29lEZ_R50c
zJwnmKwtGI>{0lHG0im7i1jw_sx9ZxKqpsggbBBA3CI6DP)BU8gpRTa}r2NU&e=$B@
zlSSa+LC)n$$MyHU2Bqbp_+fO8lys<V;z7d(b<9XQ;fY4ZTL?KPPGI0>I?K<q9#<kX
zuIF>udk<kUF~65hyPKr`a=)e0GSDOybek)p!2mHG&Xzf@n9pBlOGmn^KwXEpuwg<;
zodn2CMx!u_kIP>rn!M>nn<1jLcSJ>H4%SoYu}Zul145)Nvbi%@fm1ITAnk1rP<{Bw
z^Q3E1=8cmrD)Sp|WS;=?r+A)<b>24?*~t;vM2RD&RSy{mgYJ)xSIuRZ15)?zM`pi4
z>|g|E<6M4yNIJ$Un{J%PRX2#~9fWSZ6_ID#9dv5HE3F6@w5Ln-^Cz?$p$zHA^!F*z
zm@IDJ4|zbk6CZ`zmshfNSbTeybkR_m?Ddx%dFqcuh7E;q=!brdgF2W9KraFzY+*6I
zp=UYhVcm7lOcAA8n`-B!C$xs+l^19i(u8~Q-Ea@rb2mr5(0@0eaU}>Uja(oR1hig9
zifCf+6)&}kP>*q+1HL!n=Q8Zq13o{-VzkZ?`M3B#JnKGDcH)k)QYwT=Sz1~OVI->_
zE_Oxohg=5+UjFI*V2;~nCv7*%sGxY3`>>MyC8i|l+zM;9E*NgHs-hE4#4;vrLj|;`
zUoeQdA2OQ>{f`U43iy=|r1U?&U^$JM+6JTYbQvcG&HD>OKos?CVeySM?!YamEd}>K
z2GD+VZ;lrW$zx+klIp58t@2><NCBhj<}SAM23we)(l~$CTrF^dxHo{?b}18ZTn0fA
zi?X>heR8zln)nhe;q&I)#n2mf$jv<!0+QHukIi^g_JQJ2sU!F*g_cjREW_FKhXn?h
zQFw|pbK^g$rsH{)7&TNZMVYn-`f<=Hx@Mo4hXT>im-~90KNe|$N@L22dgQ{W5u7uM
z2=_+NSpEEJ9itNg<FKndj5wHly_xyy+Z){Ii=xk?9<sU8L4fUx^ZT+`6l4$UpI-nk
z(Kpo~o(;o2uV_0;bx#KlI{2Xs{(*yx=1~tTVB$^o5{4)Dw>N>I_f`}MT+UW`nOJSQ
z;5Bblj320?|G0Vvj-Q6A%Cu+@PJw*V=h1x1TmKu7mF(efHzxzG<d<5ZP}{o&ZlwP{
z_irzuuIRA->eHV~em^g+jnXEaQO9ItZG$(A6*VpO76<R8gZYQyX=yktKeRC2x1f#p
z*n+~sUX<7>YP7i<{Mc@&<N6w06o0mUEp=)m>$7Ng-*@jp^)uQ200u4dgT^ep^XU^`
z)J2cb&p*wlKCWBHf=kJ!U~=zvel-Vhy?o_b_h62n^#qM>hFa~YF!-ap<0}uD^!|I}
zuvTOE9^X?(?Q6zB$uU9)*}8M|>@aEr85;R7QhYwW`nx*hy%188^TTA?N}T)Rh+{EB
z4Ye=D=g{(m1|1P+ib35<dxzId467`J{O3XSafW^S@bx{i*54f(V`lIYN~lzB)NFL9
zt&4hC4({axb3;{Ih%Cc3(dRMZs6#)oXyhSXa_Ecjy560hv+Hx7sx|A^u1&g9$i*+R
zKOC;#i-_U41ao=h)Z(qbC=y7~*Vs}E0`+Z4LfEJGKb;aB&x#Yyi&#v@1wH)2hlp3-
z!Ph6&d*S4L0eDp{9`i;XFn;i{+~+dKB=y>zA^)hH&;x?q9B9>EEYCH0;C(&>$`Z^*
z0TS~kvWDT<_X?N5j~k|~fQr_YrQz&<Vh={=?x9RU?E(c{8h<Rn*e^n>!E~`Fr}`<2
zeW&mg6YV`q4P8VY3nU42QD-gp&`RNg&}Wz(#8po}1z(+K2sx(l-Va$DqPpBRyZLtO
zy7@@aJpi1C?4=AxY^|cX5FnU&6M2V7S191vNw-vQnBb|H60e>t;di#tgGqg}7uoa6
zOMwJiX`}KheH*i7E(`6^7y0;A++oXBIROveIy5~K-FI0GcKKmg3%pq_#O#0>cFTQH
z*St0A@vAW83UF%0mXs%ARX@`lSwA~iPah#2xLO@d=Q?5hkx#aA`V=Ag<HwIN{-%Np
z@<#7nYHa-bd6I}zih<mQd4E;PW}lm#9Xciq#PL?<?8gW$l-dd93?WdWo|Qc)2-n?^
z<@nXr=zcNt_YyqpI-$^YS3@3pu79m;Cl^u=X@EJ4eK$n<(e91w^RwvSf}B%-;`WOZ
znqZllyEFLSUPBOs&oOP(EW_@bs1Snh4^%Es$FNSX+uR1Uq9<Bu>LT0X-+WU{y+Np`
zcKx>vdm|45&Ix=U{wX=8j&k1>XVy8os60%uqCHy@q`h?dU?D!?@5VYHIWbi^-z3#i
zWsmB>pT#!ntjI>Z6LlVsgRYy`{II$J*OR(NV(EG)WHCb@U(%5^ItN}8>sak`-{%PO
zA@%_0eSe$pbR>K_s5`u^MP&4lhw-?12LU2j4|1hl@M0@*BuP~+Hn@>RU$1xXJU0C+
z>FU|^=Y<NtVF)H!`Igy9!!Z2DFxsU2DAd+H1-5{j=eA8A2Frc<9{1e^EV^>tf2aP8
z<V}DR&!ghD_6w2CywwRcqU#;uZ2iJeJyWsr8vP(_Slr3oiq8K%?V^R?8j)ZvleYkw
z(JDV55e~vVGWZw?bID!vAv=3J-xL>_Zadz1t~?x4;`f)g&H5t~E33lQY7%1wX9OP*
zFow!=bWQ-J);EB6nBSMl4z-131JEN`hz(?0|7*5%Bj>0`42>;7hWbFWuPO+YOU%gF
zz<FbU1?&HP1ZDzJcVKIA3D6V^5C>nBCs6(+v5UKNyby<c>}}pNcco(<J*chFXL&|J
z^eIDR5v;`zM07n0zVs#I{_HOC-dMza>0w?1nz$E#@<q~bw>smZQ2Hp%kw#2boNa|u
zAdX{qO*jcft9;{WGLj{NAk>&l63Gr-cm1_c3tpQVqTCgCt5n65kk#|~Y=(Plhk<{=
zWL^M-%tZtRD7tANXbIi>-GW0)`|9C|$FH>H(qfk8qP}j}v6yrqt3IEU=7<H>H*(*H
z1{%52_x$Nkf^9zr+}*_mZX4-cC2qc;cvl&LDPdPHWf{DkXk18VtT{Z%>0rOd&i2<;
zxxSPMwBOD{r0k!LLdZ+<Mn3=P({HuXyNI38rOgy1l?na}HB)QJQ&+_}vbUn}<qbw$
zeLfmfkttq`Q4aprfYHLTJ0qmK`Gc^*XCnz~_5XEI5)rqLHob2==(&#Oef6OOlMnvP
zQSD?t!nXBxno@r^fH+n<V{2v>c=)h%esUqCE6oA}b<$7TS(NM?(Bdx5w$!LDf(@fn
zpM)b9?-ad#$-Flb_TT55yEX@Y5CKHK^*uxtKI{l*>}S&_4ELesf?3fxvwj>}?Ws4m
zwkH^7(%6@NFL{$0KHyL6ZSMQ3@f?M&7cE2m!$Uvs;1IEzK&A)=F>9?=%jx0aE*O8L
zfh2z+v!Yw#ilILBy`j0o5*Bh>8ctZV>d!lS-%f<I@p^6NeC2QJu8H~ZDEz*2*8XG-
zd<+K9AEFGxAGEB2a%!IHj~khdl&Y0yPyK3P#+RfidZZ7K1N48_XtqwM2`oovDzgz^
zQsSdyqp~$pY)=1uF9NSo6W!a#d*K5r3VbW$IyvL#RJd5(^G9u{E8tiRB4_BU@JY``
zH13uOV?ZKFQIf!G1&io7;%)ysf8Yrr*YcM5r&@f*c#(DPSItLfzy_hpXMdjiS2+Z6
zYw!fY(aaQhz(pZOg2<7e1Q8gx;`Smg^RYlCYSZ}(vn-X@gc^8loW#5)_XQd*{+Zf-
zX)^EI?kAu|kEz(&-G{~2Zegs&#|4XBXqlFCOaAWXqB*v{37;~#m_c;+ILPMaCLXuu
zqp!_BcU4NT^aklYi8SCIjXN%e;C0SOfpQ7-MP%>{a{<mcY9UCIYdUx4)l!8^3ut~g
z0vb+S@%L@P)BnphHy?O?$(ijN%6b92{gePFC<H5PFC3#$9_sp5;CDV_LKry_7t!AI
z)UEb(E~~YBBad|i1H8Yk#b8Qk@i6h#{Wl(qehM{$u1A{$<b6v<fagruajTz_582`~
z4`(py;O9|ps@KdLIJg4y<@-tD-I*6VIhO>%Ce`2hU?!tQn{p-88Cb9{eqB#<^UPDX
z3=(xJjeF)#7QVSS>Lah;32J?I7%oGM1T|n}nNGa*r}9n1fjymg<wc#hh9p*D*Wotb
zyt%aCtH<9h4P7_<<e5zI`TE<V?<)vlZ#C7{w-^j<31;md9L00kP3x}ZQykf){!k}f
zOD1$^UPqGx=m<XTy&Cvmv?ohr5=s4rC4+h{w#kc~>`rg%E2II7;0^@VpW}C;3>P3J
z#7P)JbC6z_saoEKBB*f)(u^Z<t=g*9j&1!%E2bhCqvgGhLVvFNU_5X*Q<nUSpRpS9
z_x(g2TbLu^5c--IzEcI3q|s&;ZAq!;p^xA7krA@v1m~D$?;r^2mDf2at9zWznk|E=
zR!O=+H0+n!c5QPo5IfRt05E_Fz0TDCKHx9Wi^<H>ZHrUtWO%_v>-d3&SpbzyqnEf2
z$7@u+eZ&!lC>8L{a4EdPPoFr@_JG~*6}BSdL%D(+aj&mf-pAN%sAO3ap`SA#<`9cJ
zE+%H2oDGzA_wsDef#f=Jig^Gx2GR%5Qtoi#PPkeCRSiGz{<Y@WMT2cBhJ?MJt#e!#
z8KDxiVILXlB&+LU>n`wmK<P@|;ZhBeBxQYOeg@3vQnU(>k{+mPwRpc?1RI7cAFNZe
zj^MV1bxbDG9lf%b$U$knH9p7_j(Bb@hfkW$%h`~EquQiOf0nrh7X}cw76q^-@ef(^
zNZ{`)4`vRs26>nKZCh}|e12XWUy-qhDH%C<U4#W*+Z0nCrHm&?9lxBCV?30+y*<A7
zoL|OW<i7Q+%nKl(Q(83I36{<ez?KTZXONc6V}O)rJ+@z(Q~H)DUn1KmpUFfXBVmc9
z#tt6(w3*dRSkT6!!i#;%RwJ7ylobQuO3i-(+Lh6*;kJJ@{{1*!Be)WLYmWy%66xn@
zr76H-@oCMP(H1(=&srP=S7Ftp+O6|#!Robh9=l<}!?BvW(|D*rTLiT6>W4;DRJ6{M
zB%59eL>I%X@Y&|my{SmskhkGClfrwJ$X}S_hX{yXa0Q|BK8m&A)zyvgdcR6<C0F`V
zi;wBFTFmyd)J`eO>+tPo7gZXeU%ibY2_NNX_;d(hf1<bB>%F$4qc+zcSy!O_haYX~
z-m8!k4g1nLoM9CC<iF#!Ko>7EGdo%zAMH4hIX8^iT<5N2#W<y%0DaVOF5S(*KKxq+
zyxQw;kdC?_0o<n+`_9ma+2fXk2?g_INBCO9@R4fWuxtS{Qf!s8%A38)@@rGO;<<@5
z?$ACy(diD4bC!!gUezj{`ERbj2jrVYsa)Cmna$$mHTzrqDDK?B|A4)>CEHaNK<wVK
zx{*%=0NJ<7f<6tp{>V(0UYZLjfc5kDFYCMCx-39-*JpP{!qMNy<Wqm%o^7avpd}7g
z{$2$OA#Fx!v(FWv!MF{8(5iPb_`UZH*LS|wo;3J$5$AB{m!PAD&0N{h8%AJ^LE07(
z>IO*!P81F_D7E+>E8g{>7M2C?gnA%y0chL9rB|l4<PQZhJpComQ6FX1lZWck<-%XP
zYx5d6R%z+fm1g;<P53PLgki++m=a;`x7}P{07DtdOXfbvCyJI;8v+jGW0V@vL%LV8
z0WcRPTchS{n#XqL7efjoe<u2YkzGnSCo?ni;|qKdq!{h72oF-sc#o&RN_(N@@Yq)<
zbS-WgL<e22%iPFaUrO(5VfqQt$^;uX6+6$j{^Cp8vEY{TMhwdr{(Ekd-nCd@sofT9
z3aO9dLDnc_XMONGKM$+OEtsBni)ON)HQvrORS=6}P_AuBbaqwLWK_=&vX}k_+C<*$
z(;dFyyXdq0JoYg+*MnD*KHIp1L}G!Kcyxl-PVlEy6-1J<V?B7n-=1~JnKNVZs+%Q}
z%<lL<NRZOmS)E{U^+B9PBWgK#*g^p1^q-EzOZkRI(Q)#R2o>i0^AmE~3~8#~w4w2d
zRsez)sKqE>&J>wm(clzm>cG8z>HM9?(E^}Thh$!T$lgM@3$~{Z_xf(L_fU^pGJxU-
zZN8`<-9qv9TeLlL|ApdJ?zDv3a-piJkY|hbvgP#GgoGWh(WR?VI6EM_^lc?Nzck-N
zd+5|XGps#20)>n0iN7cI=8NNSzS3bc>r@FeLQO1^uM7%_zUVJ!NKf4=LL`YB1OA9U
zI&nizXm|uIxHHJ|B1BRKoY}%AehV+GaDnbVG$?lvF^!KJ@=N_vIxc=uL63p>+cYEu
zfXaX14SM8lagKabC&rC|5Wf3C^&3AbTa_ZIxD&4tJm@~&oy;hr+v4nn7iGFvB&S!R
z`_!!tAdDYaEFY%ZLP}(SF1+j%gokKl)9$xTFI3A)g}Q=emg})sP2t%nEr{%~_$6jk
zxB!gKw1G>j<e%@P_2`IOjp<U-j@PrI>g43ZY%venZR<K9bk=i=Z+BmfCJDRF%3JPF
zW}Agbg7qg2(C?2xwj%G*6xq=qCCI<oZ@f9*=EEXmqtsk@LQsV%{qx=>gM_!5n>z4k
z3DENaV5%+w<((<9*)4xp2}eB9dy+Bk5lYW^YcZcFD=Qn(G|F)u9X$>BS%Z*{0YM@g
zkfKFw$;htjA{_AArkLXZR;l;JJwJbK(BUv^ms}gpDn2W#1{o-l=Nlh`EwbHzPx#XV
z^8|8`H1rl6>R*Z5ap7<EpstYu@}(3fJzz=no~EK4=$OxfY?m?KtxV@SeuX$|0|cd^
z3^U;$rVJfrN7^8pjI#(n{|yXLU|TsLT^{38t~!WYWULzf{QkM;a3{-qt-=hm$rP^i
zDxW1I(I#hW**i5=pMKmg3wFlxEo!IMILE{))kSQj83$G7NJl_Y_lC|hV;vu1AlAUq
zX;c|*(Y{#yp~AkE>v8yNby>!;cO(Bp6PLSivG=?65p%>Fex1CVBkO_1y+lhl%cj4<
zU((Z3q8Mz|dJuPQB+?GNO(bl9fdS~d%lCXKW+PZK9NbFgxZmKtu{7mJl_fHIqa)S{
z(Z{#ZI5PRHRLBM49_gv@*Ta`x^cfgK+w7C6W)Fxx%djM4LSL3Q3$SqbiHt6)B8!Y_
z9391!V&K2W7tk3d(>6J(;{D6}Cw&f#;QkDeXbr>Dc*f4yzH6-mL!`xC$HPkC2QcZF
zi~he&r}pipQ(Fy3ew7SGYfvTminCT(ck_bCZ<!}5s-IB7-;DMc4)#~9CJN_G4DY=I
zd2|Q^7xVD;@af!4aKF#}!2V&K#r(JDtE^M7f<i9_s#(GhnZNJXpDy4fUQKK-!xPGO
z<S<H#r;N;WXK#KIppb4K^_kfR;wHlUL@ad3ugwapSSYUa08%jcSV^J%iL3${@@XX7
zDyIN1t=v|AQ<1xr<H+g<rV?|%&TRC;Vt6tPA2C_ecY44$7GJ46gBX0>Px{Nxqph_Z
ztwmye_h|iB4WumeB|6W^ci{T^3iR+#WG8reaJ@|%oayNOH<;!?Tf%PIDhjj?_qrm%
zeK6_^fIzZH3m1NOT455o-GN=N-iH}O8pzDN$L=Jrr-p7^Ug#O8w|S5>P5J+}1hHA_
zMgRgD4>YOk0KPGjfBRUT81xU<+^!T*bq!-+MnHjuFf~uaCc-{$W2V3{!LuBYgn(I@
zHLmFc*?oW(MPrCGS>ERC&DOsjq<%j9L5TcO8?Z0P%jPm^uALBCli(nR{{{ZC;m3Ct
z+>dt*xDCEyg8f1#z-nYsT;yB1J#v<LjyY>3>m+KgHL^vaU=wW#Y)lMHMw#$I>;EDz
zxKweMTh+}fJErd3gW8uuEMYc3JTW*WYsFP1NC*&|0Qo5ua?^Fsmmu;=oe;bR#E4I8
z<cLRrP)?m+HJ?W<7(GY(<3#4~zfUuCUcR-XD^kmZK-zbCJzrwevCHCAK6akF?=$aL
z?CK<NHD-DP;9{--!QDa6g`&1NJUVx|$jFIOea$AqTT`O6xsu6%4;!*cwY85|L*u}^
z#c|?6O-d*H<a`pOEM#)UL18Rb1gCl1Hn)*1B$yXNzSoz2qs<jBtC_hfPMV5<A&Y>!
z;}r4k#aHyn6LD?sQiQzs_oROQncjWilebE5T=tT`>!Pw!xIpUnmc^r#NU#aVmNI%h
z*GLTFgope!P;1P;9XG_FuoyftcE%PpgCqAPzwSdukI$U|@48*7F9BS)2lzwO9HUcf
zJVHbl>{}^4FI>kPxB-hqsuwS$Plj^<Gm9%y{A}-2ZHTblsH~r9z7&tJupVc{=cEnU
zFW?_@MUr%PpJhb6+=J451yAERSi0=gexv;l!@!3F6s*h?hg+bx++=zIHL~29YL;Me
zgR{RhG*bQ{EUCuhlk*LpYtxZc;c|n}V0FMHyM+-bSkJi}$bH*u)@A3e?bsB233-lu
zWjk)-PuE?PHaYr5AA5)sAkHiIMyY)3u_co~r^2o&2YmdgwW~|^W$6h=_U7Y<<<IRL
zG@ZK}9*5i~PSg(e(&07vu32*(^U>p`_*cF8YRTuRlLus<wZz~S=X0=}pTh1$V-+TM
z?OZ@>X+_(x0mhQ&5(P?g02-6$CibzF`)_MWF~f`eg@@^c?HlYZsMXwl!+IQu&%Ng?
z9X=?GJu{41aUvG<((y837`sH$qMc9>*hTZ#Q9@gKryzn^AKt-C8n&tPoJfQ2{T4Rk
z8d<anwH+xP3UoEJT_^*+@yyX+`TtEqS>)id&BF3x%r=47=w%-!XSu!BUN1{jCa-+=
zvjvB-SvQFB+3_cGLdkcO@6X4d6gq|I<4wf!j^FvC{*aN`Y!PizfHzlREP7r=re#`K
z@t`jEt~>t=O3l9ySY2q}O>~((Js~%#5PRi#H1Njz<Ez41ZN;#h@sKMDj({TOs|2Z{
znRC5*8sCSHMA|no71ABkaL@3b^>kaRTptS+=JW;P>emxVG6d+$C@jp}C$3&IZ9ghB
zmYb;-Qkg%eko1&I?k-5Z`64qSH`dJbY$5}1m&gY($hcw9;)?`3FOG$f*pz+gi82Z=
zkG3J&rY2tI0^BWO{S--nTWqpioj8ehf%W|`fULn{sE69^9%w@Pvhd1G+r-gqyjrqe
zG@>I8N!qF54;n4c{;m$(J=cq-eskC!8*+K-HVd_Nmp#Usuxr(WQaY5r-3dBSudTPS
zU>0I6URP1)Q%iI#r=EZCD9St*%yEVW^u4pAjr|dO022R3Y1YTkr}(JvHX9JwBs>d8
zD53oTX`W<ur-1)*C$vZLi6<EM-PqcX^m65sUXUwlNxoouG*)uA{?030?@A96--nFa
zL0S-dqz&pJEx5$@UHsu5qH{(Q0ioxixHAX9l9=huBgHd?Kfs8iWG;Np(5^Bc!vt)1
zQ?NX#<5JN!fHfTF0jEx3l=e{>0-#Q@vG}W6RFFns&18Q+nV{JBXi}eC`L$G)W)Kn&
z%Xy}j!)7EKcjkFHSzT#sPJoFsCXA!^)*wVJTa^FqWy+5>V8ZST3^vH%lN(6lXj>dq
zc6N3wz*IO9TmpV7S7U;I0oAUYwl)8xEDz?PAz9hC0ZaDU2<Qa<9L2QFo{QhGzT{{4
z>3Nd)7D<9tE9`IV)A#jiOm@M*64`E{z0F$I;EIlm3}?7+C~h0jH$t-koHDriz3$%=
z-N%F1Nay=7^q_*+@_ok}!kRw`r<i-DqE!sVBIUhb)VIkg=WZcHm0fZ6$wo2f#yVrP
z$IsN`XbZrE89q#03itREu%}#}u7n*rKmW&A!h>33;r`pZGFu_afG*JfciZjqXdG(3
zsx`xwBN2lk_;8kFkmh^Ih*Ef??NEqT>wFM8l7Dp<Am@<`GB9R*;v7DPt@vw^=&2IY
zQ3|}B5ZVx@D<$X^?yI8n2m6j?ms~o_yb9}s3`Em?Ofvv6?-zCc5eK#X$knxm=mPD!
zK^QIW$;m=U`1NN$;AVIU_#3`pCF%CtYkIZsW%kD7|BxBf?oc8PRy11wjf-o~vrTLF
zsxjD&`CG~<P~2f}hEF~(d|O6u&m!P{aV4jNxn~&BYL5xHBC~7TkD~_tQZutOws!)s
z9-*902)_OKzv`e0=mj+oem$9CPl0<Fw^0&5FVC`P7cwFu6yT*-{<~JKB2hn}#l~}o
zYO`Nuo}ZYbrMrVn7{9C}4@W=w;BTE+dEpZhe0u#8MWUe(yWTvLkMok)c!=#IJD&V)
zs8mH?+CuZjK>Sf(P*jVkum3iJsogs-h{;dRaF4lIz;^HHITxwGyik7<QVH%clw(4q
zzyCnt`vNPu%<{m#is3IdG{a)C-rpxB5Y0h``+4Km{sMNn8Z(l@>_{UA9b*nn&f9IT
z`!BySfVG+Dg;MN1mBnn^dEIToLsk&Ug5rj}Ih5s-qEkc`O?SS;EbbHi@-CE-Z$eGO
z9`o~<Z<`_^v+w}A(9e}^lsyirY)Te!Pxmh81C-O{=^weK?^K)SM;;KpK6?#{M-)c~
zfV0V&C(=|xLoO9+2N0Ky$<p+tdfTEDj_26>?H9aul|RBN6wr(xjZ_pkRsq+*3`xJQ
z4{Tq2s7!{}Q0$0-Z8O%2n<cI?pp{)@5u<jF17ciY(J-;T-6OCzBP*(zf^x60ngKXy
z+63}g#+?DUNPT|%^9u}9uGgjgtjIl1`Fh#3P(+&t*`)hL=e~q@*<XZHeoz{kJP*q7
z5dTTgM){}t`wBS5M5t!h#V3#L3&bbpcdXD>QAxZe%`fe68-FZjR`~f#f9%32BOCJu
z7D&%vYY2u@=4zKAb5QI^i!rS4_}1%eFrpcT1IS;svsCP;^_8`=Mgx<+)j*KDWja>Q
zq~+G{mWIDOn%j9Vve97ne81ZQL`;1bu?GD#5a}yQnusi$yL9Xd3+lkIC~5L>28Pej
zG2#f!e~hA7_k9U8WSob=s9h%8$bs}mW*&0u6{9=g)4m0?H9DfSx`*_d4|gut?hvJX
zez9|VY9q{L=iVku%XM;2M#zh{S+mr)3G})8f+^rFUFK&brW$qZh`)SCyPDWUfF-1Y
z%S%<Zu22TCc#E-S(k(5hYR$HoQH9eCrV_q^sc$I0cz)9!hB3~z{4>cfe!bOXI(|~r
zd35+4?WAF&_qKP|5VmcpvhPL%F%vJq{AZYmb7Uo~`iL$;gK4lxAAD+YxIK-CM}6Sm
zAiLnN{XZ^14*{L{X%0QdL9d50&Z;cxKO_wjrflf?unYROSzX%Vmft!nN==XmI@Cn&
zD=cVMEN4uN%BDORu+Isq)B)qhuEGBsH~xqE(8#}q_-YtWVYtQhFct-F=gFCVXlDDP
z$?8ItFcEIH#Yt?`Lb$YrPZXAK-#g9ZwkcT8PBm?g`KnZgMzkSTw=hsfo~l&)UDE-F
zbs6#5iivgXD+qmI?gepwmZ`|@k|P_+>5jyW7ui!9r3`A8Ch=a8z~^qF*vzq?-hWwq
zsyy>94jjel$E+1NlEr~Cbhw1-$}8&D)=K#kPbKQ9-wifjOtRT-ym@$|q-*caT5CV@
zjF8?cTPCXTURSs~RyvGUe@O6ldkiyhfUyIVB|{ep=+JlyvU>*Y-JDU5UjqgwyNG=$
z(4T-DXf4`lJW5`tXLpZcPC@I~TV_y`l760p1_PerjA5F?;(X(*xK7Nc*1&kyz~dn*
z70r^<8x0Xd$)^O2Ki~YBd@{~74}!#X;dl$TI6gEx>bD+f`STmEd5!A1H}VrDxapi^
zLvu`jK%(|KZi2g^i9J@IwAx8xKd!oK>`T4R5|ne9*8!Xf>yBH{VUu@~cbkmyWoY+P
zq(M&a<9Txe>PI{o5!isP+7c8Z2!Z_$cKM^uawHhJ-=>=4?3v5x^J?LV#y8In!R0U~
zrc`wirpEVsG@g7P!ykunU>1%ACB0HqR^EO~f{v)I3<2I=1Nry$Q~`>R8ZNYPPLz5}
z%hDKak1&3*Nl#BsZe219eB~Fx0_HpDBfidHZQtLOKockG(T|y!&|Q5`V0I@ZRAa`>
z^3|)$o{HdU4NOfVKZ*j4vhU))C`Opy8r7{WKuS-+fRdxy>oN%9FnS}JoGS%GyG)wp
zG(HpUwRmCoozamm>#2E^U2AASIWcf_0o<ms=$gRPlObRo{ZLd*<JdIM<anleHO(mH
z&pU#tg1P*4X-Hl8a1OA5<cORZt@)VAd>%|Y1eLHkUuy>Sn5Y%p+&Gm?%KsKmjM-#&
zaBdH~@ke2>XO7Ks?&u;x+j^?m`|I~?mF>6t-MEI8FC%z%O;5ma=af2MJSH$9`f{G5
z;_og%5G;I8*(StB)EF==sT{0y7weDdw3`CV6nb&4oj1l1k=2%mT9-jOmp<rWKBY%N
z#EK5{4?H<Vwz8y)joFlU7W6{slCm4NXK){CXl#SFT+L?YB^%&xNEB82yvnEhQ}=3g
zGZCbRRW9^`rP!CoKqo-3s>QQK{wnDbeA)^3uzB)XZd*%KGUohK?mILWaB?N+U|nLB
zV8P~dVSzNb?&<#jAzZgjcO({pKp;D4x;kz~y-w-`B4P(SnCxH3n8+s6CNF1c%B>uU
z;CL=Jo#T+`6Dpkd^XN67I41AEvyS3)w>4*&9$~THH3uDXFjQ;8z$u54$mk;u^Q+3c
z{@ev<D-_wS7LALdfwMDMlD#51p&1f04q<y#6y~<pCJgrz`ks)}jShw%CL1L-+PvrJ
zwUk1NF$>ESxUi+Wh^hQa_9yQ*>S6M_SvNRfv9)y{sAKRRNqe+>nV@3)1BZ&TPLthR
z)|DVLh8)mijawTfg+SMQ=NUTB&+%;vT3uHh8gRL)xf8un2yJR0ny3vtb5}<`SA8PX
zxOtaK3`-I~1cvCL|F+=Vl^C@YL=UJtsGW2$$~qwkhrAh;Wx$4lhTaj%s|UHd<J(K-
z16=utY_X4x;$uf&crhKlwkB{fZGVICDZD}@H2O~M7z!R8xPM}o8DQ@H0Y24~fB`;6
z+rDvy;XWb(xKg<X^l?!5eaM+=8uu6Y?#0QzzU+O-nqq`|w0S0tbEDj;SknRV;gV4x
zx}U+YA_{N#Du%CJ_Se_Hk#b)yK&wsB32S%W`07Du-R6n~``y|>g^qCUXgFQVJ+C5Y
z3%Jv^EZ~bmqS;Lk1odt@hN4N$sQMr8aJ<R6ou#cqihO<obHH-^$YTVkJ?sdqTd9a^
z6o@v_e>@rvn;dllWJz+@bVlQaCl_?P2(T}Do@Xkq6-HYNv>T30x+*6c;>KZn2?b9W
zUxECuYHQWwef~H!?{Y69x;<Hfam+wvg}$r3fg$Rc3XWesvt_aahL(afF^qa+Z(!&_
zF(=mxAU|9e1dnV7KRXHmsQxOw+S!+2>^z|&TZ{#aS5|v&0@Fw!MG+!Y$fteK`&#_*
zYW=M9^Xwr$v-UDtN%=@@yC%8h*++6;{YO1Cq$@M}<=D2zY8XspR*>bqwQGmjDmKDG
z4ij4L4_H_*WNt16Z*z~V7xLDHz4B@ETftT!V{bIPslF9DyhgD5X@yra;EUak5+btz
z_XkWGY{yn}8!$bP@yXl6)wjKX7GMO3as8*QH^O^!7o}E80=j8VuWj5gCQFyrr_TA3
z^v2SE?#7Q~MUa~DB{`7dJ8?`xtv;E9MO=5q)nE4S<m@Zcv;W=OKV~vo7yuQAP%;_j
zb1#JS6Sau4f>q&yp|8N$pV<$qKc#|GoR#K&e4!br^P|O8jDFU2z4`PhBilYhk@{oI
z%P$@+?(^zy?#}@5_&`LorGy{VDxf+5*svyI=o9_l;0JoE2^h}YfZ9^_4Dj#&v+&P-
z(?@Rb@C@%ax~C~NS*uh_eOlK)II={>obDeSIsV+$O92gqC=I<4YM^MCmd>-A+^L7*
zoB!j&)P$_npNZdFayPZEXai`u4FJWbD$Ja_Z#yJ3a9jS5qjJw82OEZ~7C&UByYV$P
z_OaDv5FSjmgL!$zz*;y_Bt>aOeU&ItYZQ&>h;esA0BdxD>(B7K@(^7DX3sNC#Q`4%
zX1|L!dAhMpAMY(uw?m^EX)u=h8dU~6N9cV?_!zH-PE=&tB<_BX<2EGMj^Jeax{(tV
z)n@8L59`omvA@!?-}ps<6l?!u6!itgPPDo|wM6%s{AXr%4(SK=Csr1(MMIYYbSfB^
z5nTW&pro=DG^35-cFh`#58aG9g@~W6%ALDpC7vD~jA<YYr-?iFv>!OoR3z*#SzjLA
zB}tcB$D!|?q>B1E)AoIuSg6SHUBySh84A8O7cQftiu+vfg4m!Dj~QslWC%3HbwW;M
zSYLY{`+Sfyah(%&y5;d#8F13y_7^Lmkb;i=h_1W87x5$+L|tW}_;;QhQPeokV6G)~
z@yT({KjAx<0o==h8*wa2($;5u*W<oz-O;n$=z8sFsWbdMFOg?pcaWE{!maWN;Ed5R
z%;=2;0)U0wz4)u)WoG(#Iv-RO5KFbQw<aRY2@JE0T-siE+y@8uxom#O&BemEWB<`=
z3fIkh3RRQ9!n&Z5KL&0cX73FwkwXJ&*H0<JT>APepC&*2RvlA86;_VxL@b8kfGcZm
z0U%B}^U?;8B5U@8m6vzq?_W?)9igxF=yZ8BN3FX4Q=R(J5YFulr$z6nQQ6{NCGJt-
zpNZb91~knKOOx$OXQ~BQo3TW`s-v-wz#~U(l~!_U^ch~@?<ROMeg~4V!bXD96J@E5
zjIHQAgN6H;BJtEhcy$hwI&d8?rcuwO@9%s;J`BWs7i<hUPJo3NXR{6aF?dIAO}XuH
zlarHE17<;tX6|H>!aI@yIXY0U6XPw5UOmz(GEgyaF}AVPhK*h!PC_aCVFgE}FST!o
zdVul`YMiv`_oLxHU4cdS;#prJ%RpVJcmA<7mn4X$`Y<X0Eebf^LI1ryVfo1|<nV{`
zK&PA`W8nX4(Fjb6rm*^)5dgs!fKl6wi1w5MIIg~5(er)Vu1G?Cn9vb{`JJ^HX7ymO
zOMl=4%xm}Ah*t@AwIAb@%coVCw&95rDc7bOxLHureK&s)pX~Ur63LdKVIZ=p{++fJ
z9xSjKJ+8oqB0gnFzezfg4yAG7O32u*P_|NYTI;1J?CTcYM0nh;@-FMWu|{@KHzp-N
z@Mf`IQy*C#`aOaRcDIKt;ggjB>T*`{w!XzWO;>EQo;2tr=#iPUk^A&w6e4L{40!=t
zlO1m6=d83wk_cDzXCn5{q<dKfvJnGOZP!W)KlYbz#8%>IhsG;Rx+ljwmUk*%pRSfR
z>EzWkJO_2^wVPdV{uBTbx^(LsAQ*^zESO>lewq~4_<X95u%YN0IR=7SjF=Z*bLab&
zJ^!PYxHBuhf9xTY7wI?z#*m3V5jfJ3@Q<gFwLk52ml)+#=R&-alZNj^eb*^aD!|G)
zt8$!?cm2iMj6`{j%bp<0$(s#%{_kzN1!g|pVh11pVw2C=BMlylhG~xnn?h5h#^1Xw
zM!8<{lh3f!uN_2jtMR>m_ghhtgkXUGMeY;hD+|U*W3Q(&br_h-IM0`8;^fp)M`Yh@
zJ?J^KV%^V(O8Nc#ahGzsJ?5AUb4Qv{xH9zR^Y+}m_2W_k6_1!U+dHo~rAzMCesDwG
zqwtM)ImLd(DU*-8VRPxHf?hR#^OJ?ABX*E~Rc|Z*Oy$>kJ#QJUQ7Pps$75NAup5p#
zgK(zP6Qm?-ISsRr^zb%24{lVvem4r9>@Gz5Y(zy~;Zi-SPoH%hYc?4=fl>3m)<Qv#
zMjVYXf3i1<TZ`9~8|GPf(@xhvJ1;yTAfZV=?n%lff+|c6!nb_-?px~LTEH62E>6#7
zLxA#)jTd#@ZabOJm+~qm?<Lzsy4yu6bZo==5=1XA0hnx<nzw0@q}905)#3}sXF35e
zzeEE$X&4$4rv^HL%m{Nf!|%*&w_T0jb|@o1^y@qcH%g5F`fo95)PFD0SisWq#?MXd
z)7Lwq%<8toA0n6)qgBeEvU3uj%9|%qoBIJSEZIx79_uXQn>&wK?G)*hUFeR`3IwbN
zkigr$6N6CFeB=Amyc<D;Bq^HQrrb?|V-GjJG8(y<BB|@HQ6M2+Rr6c$vlvztllwT9
zden?*6h@|setm6~t;CVGA+F!xwS7&ocS5mmc$=&c|Iw32Cx6@bzXCkhrn)tCf-29x
za3ehQF5sm3e6OoLf;P2SEweyO=I9ic<kLEa(8^gKOM*F&m$HMXBG&-`aaNkPp{9&n
zpaf~7T!!&gN|7N^6;w4iPieun6?z{e;T_4bdslk>x7h61y!sL31bFAcN^Ca!5d{HN
zKfLTHe2a{k;V`)2!V?4@L+~lnMy7sGT^50Ten9>HqBI$Toz3H~U-zD59mzhB^+gBN
zpSw~CBBCcM`>ZqFBg`G#*QPi-jg3;4JyD%^`C5tIOMFFS8U|#Zw<N{ZZ|xgR^lV5(
zs2NvufzNs5R-AW0?LqkOXf!^BYJuwffiEHAKVrQC*%q%B6QR^QgjY7q-yhr#9jqi*
zkp9i*8g;?;{qPz(F6uyhFm61?r$nCR_rp!vF3t30uINySrwJj=RL>6#^12mnxv~y`
z#kMR0J?BH$mgWmJ7R4KKz`FDkJ^elr!w7+uADc8`pisK9Zk89ttfoMigc+QjbtMS)
zn)XEL2b><eIa9wx5P<3Gy@1^KzhjJr^>g)Cm>rWHi6Qb=_@Rdkv)}l(%1uz#;?AAX
zSQl~hAFCgL@DzH4OK`Oq(i~rl$2uF#TM(^_YsZJ7wtX+@)%teTokXntJD=0sHbqI&
zC70L7wvC=r>9_of;xtsBi(sR(?<jjR_Mvl(w`;|icEN5(X0I2bGKcvcsN~c_{hW?$
zog#9cva+OG)&RN*g}T3DYOB$m$$%t2Ccp746{*ut78P42%I$Gd??RZZ5uV*9b9@XH
zT+p9}Qp6_b(uF(OQD6^Y9Yp(IAY@|`y(x;Ve1|QEh3xU#2yax&y0yah-$bH(S=oqt
zSj%`#hJa@<GC4SyYzb&HG-O1b6pN^n+s@y<zb`XwZnHH)wh-*NK7!;69j<3pKY@<(
z>lQ$>bfBsQ1p*x(MR6sQC3vAWV9D6OWkw-dh`mBF-Q;0=n{R1v_C@DG_tT1JsXP11
zVBi75L3!YzT+?HVX%%zsF)PFv7Hl5|!hQ8<qZRkE4Eq&ArrFT4|MoO&bTay*5qGMr
zIB{Y;`oYD5*_eEg=fWR0OTqmzPqgryLU3fmVcI3y5-XV4ev1rM_Nhr*TJ<$Ef!^()
zT7Hp*kJ?!ah!Ob>_L84=$UzpDR<M=l&qzJqrd?_xxeiPpl8oOV`r^4h4f+?%o01&Z
zBu@A#_9O&mROj$%r=dHBGMzd+t)^tYlgRTC2HnfAu8ZHwq+>&MkIWU8r?V<2tnfLp
zFswg7Q{s9Y9_qb%RRA#>6O3=j2-5o@eks3CgkA%$>3a=>d2CjZ54_H_#oE6FA*qC`
zeg?aLD}wiqCdG<#^<$^z`Vn=Zi}X_d>^SOR8BpjJom~(II|IXm!ihOo69}y{7|rg_
z|7FBZu=`y7!{bC#$dObG^;r8eL>m?!+57k|^4kgG@K&DXKis7Jk67Za0|p8YouECi
z@&cdgHuy_?z<pS!gKbjU0NI;X@DCxUpH-pBd2!dG_EOy*71sZ`CeKKC;T}2dvdMpB
z%Mvc80hxAb55D+djTf7Er36b-e-~f!5S=K*(wl)J1jEr3+dJ62!mLF7kD<muAd~o!
z3vysn&Gr036Z&!RYWZF0`pO_Y8a?x9IK}1$dyAmYv*<4cvE=Umx_p~((B-djVjyfP
zj4KBJ*X6%a0NetCcSA)IxOAu;x$`eMDl?Ca$0rRQ`?5R|y*hL@&RQ__ng75$L}jn?
z>jF1PB!R!2MrhPK;h6CV^@_E8tCC8el)Z=o^WjbHou`gp^YMP6lYGw_bsCcy^)Tmx
zI44gLVwfcOcXOg>0m|MjW^l&K$*z~3!x0Pd7;klE=ZO!vW+c%+5==EBV_#|}QL#Jc
zEqL~u<qgc*1a2K_TO(WVQ-E7;!`QqNj($MU6DPJeuTNxDPD4aRN+h!T(}9(D5IfAu
zBVoH5Vjy-gsvN~TosD%%O9P+Is1_rwx#o}2E-El-Gk!&_-O_MRE$3VPqPYtwPjfI~
z{rO=1@gwh#e+*UQJ7l=<>lv%M-T$Z(cW^8)F(gZ&jG5K%!1-Bd^x63d1<4FHO?t_p
zm@(rE8>KJCMQfkgPLD%oZ^LQcXMJwALs!_Hse6tVEQpNk<44??(Fy<iSeoVKm=BQ3
zYwv$xvHcEKQX;G;_sYY!yReRRW5`C}HAW{Q2H+%}P`Jd8dQE<{9xU^Tj@=6|Y_f_J
zY%!q((K-Te@u><Llg`tCevIp6V8br=gp^f$Qk-n5x|jYra+SvA(9dN#*Vhn!@ULNR
zyWVFDic~n$WZVRd&r2l-%#2b(i~Js?*R6C>1PAY%TvS!Q7(KYpK8|1L^jJf%S^+Ld
zH^=Wd#ZAVc^Zj#a83JS^7Ky$C>`rRXqJMkc@cS*u$KXVe6}-asor08X=4FF9V(~2v
zCJr5SY<7juzzvsvh1F!4Sx38oi}*|UYXsNW-9=f&AB()5l3dYq+ln~kXBrPO996IA
zpiABxqjJ#8VM$vZiK9iGL&1*$@%~TQ*TH?fye?o`=kQM;{P|^d0R~qe6F78mUI5|O
zt3Hjtz4^*yec9n~58pD<grt9mt`0zc>&!@l?UYqll>l*zB#o%}xFffydy&xX*hHu;
zB`=buT2`s#3>tA4WxPR<cOa05-2LToXoc1y14%nd%Jtbff(z0>mxQin?a&buF2&ul
z`FPhy3EUVqQrrub4Qn53LOJQ6cNp~fcwCJzB?Bbk&|mP&k-tHX!)I_>b5rFT(f?~)
z&u<&oG{B||*>zX{S0upG$oS?~Q+ua7ZHH4AUbDVM1-ikoTGGrB0h?&ym-tr3((ek%
z+Y*DlD{cBVzN*U|#Hz59X(?l8OySWk>;!$EljAGCXyFLZxmE%@E~X=-vBh(#X?utF
z!R|4TY=51~4~ZP5&Yqb{Q6c87kOOQA!m)T>`K4&-=IU}Z#M`PGMfOZ=>IKxj?{jB`
zqni>La|J)OKW(w(gY~$vzJM&LVfYHq^84l+50kdoP+QG0#qM{KyDIgw5j#t3`?VyA
z5fHPTXBS5_th~X`gk#O5uozvWu@a{jXh)@&D|6j;2dB~w5M!Z@+7;SBB<wM)^H%}<
z+Hg6n9U#F7{SAVqT(Mz_p$XirO#vs)x7iPO1Ie2v#Bn8`b4DYzsTw&RJir3u+6&{g
zR>hU2`}n9bY)fmnwRVUSfj5C^(zj9Oyt0s+>*y8_<Lpycwl;nTxu=F9v(pgX1(2!L
z?KY-HEVf1^XT(hOVyawoq8<#1FKg|H0g<1D*54Tmw)<104()a2d+#}np_SyFI!?%4
z#n-<ZJNh9A=e7pl>E~?;RM0`gCXgob@Uh(}T<wFXp1$Y(@E_N&9%q2)4i1e*AVEZw
zJmXIXwnQN47w0kMYCSzs$+QeK0@%GAmKLmEtCB0`wQW2i)$S;$@0^>Btvw2+$4{}x
z))A&o$b89+%v=gN-z?6v&FSk?`MAd&1ejlX3<Jx%quhb}5?YBdqCw8bp=nxDl`lzt
z6)A1WKVUE0k<xLzQy?X`pSiGA_uxEU)-vfVn2B0qm1oy_#Z^-_1fY_@au*8=mF`6~
zO`V*zCnx-gAkqTMJ?HjZG(T9y{YQ<?E4>{!8UM9lZ*JuR;K*d^pg<$KG4|>12Bq$z
zZrSzU{$UR*<gbX*WWYj5GRC}XCYHMSx6E64%CJDG68)YjcER$MBRC_v+ldKKYQ(Ah
zz|-{5x~@<Yh5<JDZf|(kbLgew)$HXt4{G58V>v3+w#2$O`^dQD_u~bv(8g?_(Hb=C
z{(mgvzPj6i?jlW?6<))J@dd|*$-e83DnwykmjJi?CZ><Z<(^%oC<UoXkW-xVPT=l(
zn5?9iw(P{cF7i7+)qc=CKX1ZNF&>p8y+4W3Xib+9*Hg02xC=NMPaSc#zpdLq*BhJp
zY{Vr>J21E~ju_>kITu;p=0#J{Q{~c1R#127_=t+)K)YsQ>!v!(Q%{-cajxA{aaP_S
zrG_?X6gN($qu2Y61M2r44b6pg24UY*gLn`Ui-`zgldlAl_S48Fe*!9fjc*>SK7~0m
z?+%w80kNmdY2G>Gc1J9M?t2X0Z^lou&Je)R_fDM6N+H5d?TM%3!i)>d0?Nb*rCmuS
z)Ee+o@M7X6F64CQ0z5G3p!3Il<u1b?cJ?I&WFx4T#vrrd=rG{m69?8F);uxz$X6#q
z=F077J^$1CH3pnh@cdM+4-t~o+^vOMmd^=8C=@M;yOYFLWHAepXg2@Zn~;AJdrfTq
z8uObVsw_uQFVuEs!C~Sv6zFq*9i~($kDddqJrP6Y`9%X~7&mIC=Za@;pQo10p;(QI
zQH0ibqyvDjkDY#w^35L3WmuCK-TgWaNl<KhUE6CPCr>uG@qrS0ti)S2f35iFQ|d)@
zY=siS5lH!QhC8(K;awa-v;UenYvHsF6zPoFCRJ><-6qvbinYhoStlNX@B?7Ff;wGJ
zu)MjxGFY^(@cLLCwWfX&d|nKzq=D(jt}RcjGw0m$PF5l(pTr{%jd!=J%54KJ0p-6c
z=eXQ@@aqlz?GkM$ON0<$-!<Zwd9zptIjkEaVz;Q~OFht<4UW0V|EtR1Zd<qi8O6Zy
zfQ-NxTk^RE=$dH|GB>RYd`WZJ+k}!RH0hg5kGVjPJi(Z(-q(+Xft-(Q0E#$(pvtiD
zfzX}XpPJFmpmJ`W#O{jNRsIXy&>t>X=Xhbp7{Y(!ovsN{g{6m=$LM0yUqRT5&J{fe
zUYP%OHNUukT8=?rU)C~%;FlUx>zA`O$PgZA3K#UM*}gp=4JKCU4kk)R^Ia|T-W8;<
zhPwcRw^-x-xex!NhW0Jro}CsFMlyKKi_bsxzXZOvWht4ELSN!=9-vWXJ}Sm){?;?K
z7v3n25z3xQL)rOL=2FvbLFhT)O(_`XK%t?P3axMS1c^+J<>Y+m3+YW`3twCf<14<@
zpH}?9xVUf2=r`y)96W@EsP;0Ib&8`0*sO+VvAMY|-u>sh_TPUCk`eQtL6&TdCEtFp
z40s9o5){6PzC)!StUNX#t#V_TP-N;qoyc}XozxJoyw#52E-$4(Iblv2ZN<`5L*&@O
zDUplnRln`4H8~i=)UEOf6Sce>0|wv;1(c(4d9=lpQ}`(l$DU&HUBnygx))<TmAEHK
z>MLb4bf3s+i$}n*{=ZyhgV(i;_NA^+e`5h6v3osxj_^H637<V$ai6^%oDBVPLvpj=
zvj?1dWp+&x+Q#KxWUKWrh!inh2UimMR;vNt3T!mLT0K>#qu?`*1&!co5U)GMNjVZ5
zh+iGX?}v-IN*B|@tf!{knN@{Tz_zBbBjTbhB26X8EJ=*D>z$*lfJtHLFwo~{%3Rf&
z1@RC)7iUKU|C24Pw)1U>t7qFb9X1nn!X0FqU%k%40!KZu9fvzck-UiA-iqDs|Hs!`
zhE=(?U86LJQc{AHl!SzY2uw;k1SO?WK)QR<q7u^G-7VcA0+N%SG)PEy_r9iUJ@0<@
z_wM8SGmo`ghl|O5U)LGq9OoED-?5EH;`?9qLXS8f%dZ6NJZ=R*s3~Axuk)v-5^ldV
z`te}@S#KKPM;HRE^j93mzDs_Ex*t^VKU@HdXF)~P{&m0IB%Sq~Cq@OB-tW%Vr@Qb0
zrAn_0bRadBO#QReCEst{qHoiMwCE_(>S=?NF|JBv2<!xrs>wi&<t(KUirA_t_R%Ib
z05vIV$MOe6F)gZ(Xmg(Idt4kE10sUb1$x=c{S*q>r@@6v+)h_@#$DZD?N;tdX%wlC
zNo4e_e~}1|#AUMb?g3Js3v%BT>eJf5FJ%`L`)yloC*<DSaMaXlFw@i*#Iq#FwDi<)
zZ<ilO)oQ_uuIpQ1NYtXx#_YFd#RvoJF@rzJ$E4+(k_ND+(}pA)$%IEZ=z!%<thHGG
z&roAFOXt>(opUd6@x^P5RKL+MZNM%@-1c%5Bmgq#n&D$F@-P2*uah+Mf~kA*JeLq(
z<={KgfsyZ8{t_?2f(JHj86Jdb5c_4~{=PSXr^RyH!K1pTTMcT2O*cChGD0umvGjo!
zoG9PN)6i+NP?vJjW`ixh30R!c-G>mKQ@U&DOAI;Kr<05E3%YkO@T3K=AC`PMQ1QN2
zwmYxMGChro{k%eXC)Vqt!rM6982Vqt#u3gV7p=fwuTaWqg~$MNk<M*KFm}nFPQA?<
zv0-B^$xnrtEU2SPaM<^We%69=?qAcZm+EXhg#kJ~N^%mgi4?;%Lz{uhgIQ&CM>eqB
zeirQlG;~a7YS@1HM$3VH8lt>)6m;zK@mg%0Unss~WXKt8+BL_5`Pz6t6@at}u!`Ql
zy_)i09843T2KsV%@kMHjnm9^5f4<MbedgT1jGyq3`^P#|G3f(Z`?oHdUQbOL^%cGR
zH4K(s!GxAyH&EcyE%lHkZPcUX?$*84A_F8TCP4pFz5450iYcL6MWaC%@TN=!Ben6o
zCkws5-pZRB71DSx|0=Txk1D*!@M$S>!M70!hN3gpEcSk+RCS$%>Pu3}QyNrA+Dp4v
z{PBNz(-|Q;x#KXNu_2(8LN{k5Qj2Y@>`@tBF(2O#hD}PC?XkxY29-n}Gn9tzaPd7#
zms#KS^|pm>^3La|keP&UEz1x2r7b82S7O(Q>2!B)bsT+s<~&`OhuhJBDAq&awRg=m
z67FIB?T-`(h9AQVlAWQmfDrr((bOlyxqa3U_^R7$a1(&0%aI2nXw@P*5q3aW<agPb
zMi*cU^LoEMVShH~VhG4wz0AG@1O(QUxvO^1_boLA*2w=5FKYHNCq%z)LAkELp1ec1
zwe`x4;;uhaccY+Lf49%&Hu32ZpoT`C6W;f!c#V=_jPl|Rf_7bncJ5}(_*3D8wHkKq
zvV|&|yvX1{=Wi;sI3;LX=`bjRd3vXxOIM?3*XJRQ4p}>C!j2h?b^wYGdeSWnbZn#5
z*R#I;X^eVE&R$4hFW3uaw5X}aTH1rS>WI6t&IA+~c4fl=h-lWWPG_fimd;qhtEjn{
zAOJ_v5(2vdx|w+147^ghu9suDZ@esG!0ex<bokK|8cWPayE8V9Q@J#Av=6b-A9Q2=
zh(S&@+}-|E=dLFuF^ltQ0~+r2kt1@V<x$mD1<C&7Vh1qoO)Q;5QpOt)2MAr_R-reg
zRcEy%WsbePs_bKMMd>5Mq5JZkB{NabaeM07aAGWV0G{Ehi&_7D2eQZns)WJr&-i^K
zDi;BOlw(BH{)tQUL!BTb>g8tK`Uyo~k5jY(tlh6N4f&};e7fl9W^1<=y3qg<Yz>x%
z*{%Qv$G@~pggAeLCCv4)0>T=bh0?L7q(6Y!AqwEV<NP8lLYe}Qd30+QkYAAql${;|
zG-pFYCx^Fan!i@kRX@$Vkwv>)uixr%PKb`qUKz!t_E?1f8~{+%oU%>IxSKTL&O)om
zyw{Z#Foo-p%h6VHY(`)HvAv^kUk0yB$ay4N7TDMp95-#G5GAGoz>QiGTK5}@y#^96
zD81b8izOmt=@1*6s<dbG4Dzw*&O6f$^-P~?3w+<=$~|=rC*Ej*{2g)IT!3-gK7h2*
zh&{n94?7#W+DqKwUYsJDs<xOg5$~9ORa?7WQn{Y+b(2G@3JXDXkP}SkQqNErK9Llu
zY*<SI`@UtsQTD^n0D>TkiiB6W4AF6pLl1=fdKyHMkW(cGSDfn*;29C18lX{1EwMcS
zMo=vio<9^qe@%=x=4t=EJKGLT+jS*lN!%u(k?tQ;{fdt}tPr_H57EDCD1~<Y%`G#C
z+_J;^a~T1Is{fsY)f2hq>JtG9fqySGQk9&!ISay65X3VhwSqSpDC?y0iqU}WoC#z!
zd(D=Qr31kt;N#ohlYw@hSs1}2N_kpB#0`kL$yx*X!(=R?XEv{B(hAWu9%fSzN8O0i
zEhd<=yxD%1Zdxt6KD8|q$&O;-kK9<5k`FNKQ5Ej13BTIvu2&AX#=QHhh99KOoCNv$
z`J!p8^#+UHY+er;EtcwuN@Sm^Ih;HzuM@O=cO^_c6z^+N*@zD}`3G$>Ku5IrmF^Ea
z+yJCKhN@3vc2l0tNC|pwe1pgP9&xEbfHN*}czW;afnlNWm%7Uie1a%_H8?7_V*50w
zEmx$Af<*5wbE4s}rTlgwjzL|6;p=xVSFt1cf`>jgg#b=Zk;y9*LFu>_JOsB@;x-dG
z5qc{e@>A_5!pq$M_jI3XKua?*m~Kgei4yYosM20n^@Xps>wFLt{&q9W;gfyF1nT$Z
z8^pjIaO7nKd6dX!<Ay&2#&ZRr%2gIC>xBKSEk$txp}bRdc611;S~o)lcK0eMgLqun
zTsi=8<UT^R6o^dZ7Y9;3BEnuwFvtO$WlZ=<wd=_E_eq|$0X@=0H4Qw0YHUbR`x8wh
zSu~<iQE?}nN4iFe4fBCj&fNP7=^B7~b^tu@KlbEe!oC@B2^vT&@H$|-s*7jN3pcv#
zmHJo%W<0lCO1emG1x#yh)Q_rI42c_%i%<Cg-aE6)wu#E6Rb}c6YIPmAR+%zW<j#yH
z44~PIZuIT+P346Z2A31+zQU1qqca=m*4Hr2EGYYF1$a6`{UAn8HHh578eTC`SKK8B
z(jb7%h|l#snZ5*L04AISXaW#+@mRXhkX<`1%-MIIUfkrB2`9ZpxUk>2keT{0vel>r
z*c|qUqy$vBG5TR9qM`0oTjK}Cho~OtGuE1#E+H<U(ZDhRXMZpUu)mapydqX(RTPc8
z_D%)By1wm~!-x9{hYQnKG`1DC#t!?>+XwwgkaPBsOI!Kfp8)u?@HM{ex4?wbfrz&g
z4}FtU?m5q&iQJrm#=UKfdJVN2iuC;6;Z0s1lE@%x>}j-LN^`^<=fW3VZrw`bZ?V=~
zc8B(QJ?qPj80t=cRUAg|E>+5`K>va0DU7DnV2*0k&YCU4aHId6;djWd5u+K(p5HI`
zEm349RDNQ$fpC7$aS;BVfcE}8!B7mD%PMFzo)oJ3wLOYcqBWDW$xfRq;wOIH%smob
z0f0#xDnpA8^pHdTOdA3p?hR&{d0sMR!A@KTC^U~s_tQF8$QOfhF4zC?Df+)YFIHs9
zedYd>*8`Q{JN(|nHdg7NxjC<+Xk&inIKjdas>J*omhv+B?ufb-G9#U9$;q@o4a&EB
z2|+7XcQp{j-kHzNl?0OhnX*JmTC@<-uqQh|F|u!~XiWXB0@>^3s5B6|b*?D;#|SL3
z3I$Lm#1VWa-~roO<NUIs=Rp?>g$er4C^98Zr~X%FwD08Zqh*oqY6C+?^oInh#|Pci
zS)#ru$l|!@`wA(A-4Y^%VqKYT5A(&@{W3qG$IHsfwga6iVp2;MFu>_<ayxocl6b&f
zG{MX1w5f_e7B$+*yr_Af(Gi?{i;S;p8zz|^58H=<E)hEan-O1RXW~2`6Y{iQY?c4Y
z4H^2+U2T3aPOEOno2jPKvI8BM^~}PLoB3Cz#b6ZM1w_4KR>@vi#xJ2!5QEvS?>f2Z
zSbhXbO5}5KMUtn!Cc|PBTHysL>FH|zCR&&$?az+6OH7CfD1;2V9;_O~y&>X8DuF`O
zpBN7}eF|{`RlNrgEbTohcid1O9f`2(1u!+Hy?iLmGP2ln)CFxfuPCyrb^Z{_)SDQ?
zMNP)2e)#*vsiu#!pNMV!z|n7fPwe{OA81Jt)JNiZyk|fEhWZpxJN375gQ7JdM&$uy
z#>EYkLTRxzlD!sf-1nq?45OZlR+oxihjb}NT5M0a_^Jq^`pN@1d`8|c&R4+GXG^{Z
z6<z@MNh3im(nrC9zF>ihZoD@3z50sG7c;t-rjv3ffNcrA52UN!hE1of9sj=t5YAVm
zGhFl~RQ*>fk>N-;Y<oysdJ!;9jMUZYrNn<*gkm3GgidEDQIC8ZH#60O>4Oi2j<~i#
zr`dXvhp~1tb9eQ-Kfp%h=W|9v{0OC;;AOGR-<>jA<NCGDj6TW#Q`%d-eo*97&A7ka
zgS75bmj7<f4jReM(&jK7(J3|{4I@aR2vIqMJz!~Ru-l%yA&_10Q|oj%!J{vw3IKe`
z&o^eMuuFsb?Zuus$Zl#J0fCF3Wt8`(+oEif&+({Wg!8e5v-gVvDvdZ5aFkaam$d$V
ze^=r+ewp)WE?W~gQBRemZm%yWiTnN@Wd}!EByf^>sCc*S!J9>F83C!eD^a?LnS82n
zDi*NUF~o#YF}D01<wXe+cb;O(4=|h$#CiOJsn4>DS$T{6(MwyJt}yDb$#*w<<myM8
zK-ka~$aW%jn-M#OfRu-?XG6O`(0^UkWr8u`5M@ZaA?43wITd-MD--R4iAy00L`-bK
z-uttARTN_-rBA8><x$-)bBub2VC8r|TQ<tzE8F-)uds!aQXKgk{1~7jWj&2J<Xk|G
zw)6Fbc7I**w^C$*@_)s3o$<*s794?f@?^ckQ!tLmoeh(ZV-EpPYO9o%zf^oe+5}TS
zh2w9M#e_ev{6im*^DtRxWtx!Nmr)Brt5BNV<q?V*fKY|+^EhaUWKpw(bU@e9HaM6O
zEqT<_4s^@CmH2b4#&QW<#{2)OvHaKsdv3$X=*|@BMt|xjUW1LJ%Ne!qK{rfp#v*vE
zcKe6vIG~BK{`6j}_hW^xw%Y16Jct1yc}hL$UzBM!6Tq(4n)j$M|F)!mz~X;`uL(x#
zfG+n)HBe58WpN}vqTSzCi_=NnJjLv-MX_!|59mp|hfKJgbjW9W!3P8r0a%of{{7ec
z2>{CNG@6^F9~z4GH2Qx2B<lF;g^vP9Ql1X-?IrrT8*%v&D*w>b<C$M#aY(R6chvE;
zUbM94P%zV20w%Y6K69hK;`ev@I+Byiwx0z%MukLa9LcGt891@4;nmql<jz26U;Bn`
zxQn2YI7Lz72txW;|BGVG403~}9HTVrD(@nHgQdz)GTlS$0>>EC*yT+;Nz0I3QKMje
zcTEGWKFaB+L%vA9P*n=IooCQCTTp(4S6K&?3pP7fEX#{D-w*uO_^Il<*3#n^<WHUi
zzwa`O#$DBC$V$Q{oQ%c>`(2kdn?Fw`3i;%}<Q_md<UtVp1REXzfEc~{MJ#4JjzbKB
z^->fY^%OA^dZH@260ZeD;OL1ok4Aew5QkQxEb;T>1zDlX5f13%7#Fq+(oOa~c(ha$
zs7dn%4RB)^oOI~IC$aaNpeN?&3P&WB21Ta3QU=cD7xeaZ^o#3YQ})aux27RgYPr#V
z6&uW^{9qz5VWV%0Re+$L0SsHc26oWQB+{$NCSQ}ps))xr&OQ65wtgPjHMXXR0;BC#
zC7%V1v*+exOw9f|Dp7CMYW}HZ$nq3sZ08q#T+C<AoAm2!Rh822Cwp0AN)*KK(Ym;P
zde!3@k6W-Wgx1k}Bgd#ee#T=y3KWPG;yIBI4i3DY%%TC;BAyK4^v{f=Kr9B~ksI)>
z^7e{>Ot;yo_*KG~@H>r<3w8oAB4%c0FWnBaQA9_D=CFJa+jdy$xmVc@@yMw={gMlI
z+cULF{J;KyZ>cF_Jzf}dc&(i7bNA%P8mjj^Dyn^gc%Tu0KJj2_Zh_ml{N_=0Ez~N3
zf2G~@fwR}Hs^Sd4R49Tq>Whf;ntwIUMCTC-a!u9q5N&V+ehK}P0Z?JMx-h_Ur?=#M
z50shOk9(|A)M(zNk4+8RIjXnmVwZPRHct3S|DEsw9{(v~UN*>)G1^wC+m^@Zlk`93
z^T>xk79yl*8wK)RIm<FShYz^eZ!^hQ7qaf-tI-aPGv3T1)3dJ|V1@s2|3H=*G)soB
zZznO&se}A>=NbP!2pt_U^OdmVVJ;t=k}QV^8)o&;5L1{qGGw%Td9yhn-h2*UY(f6h
zv(zdHqvZLJ=o@Y^_vfVQZ9(9?SxWL+|MLYAHpPyhivlB|vnand@=q#s>mb(#sl6P;
z4p<u^s`#PhI}#Zc^&IT*pxd9slCWtp;Y<Ji@g;`3pnWng6^gwgENfes{|@w|0F^ld
zqJ%>=R@%wJo&pw3ExVs-ZG#t+Kuh_?fxjO`ly-kQAJ1uxR_5F0<nJ-&#C)vj94PHC
zt!NE!6g%Vey#m2*g_fcPfnzQ@T>}oy!}cYhY5B%Sf(1-)U!CpF5f621Z`*dpHqRFe
zI2JCeSVy6*FSh&L2m`TB_LlRd`_XLL6Ni_5JXL?sr*RjdLW=F&g6ZII51Y4Ze#1BN
z&0SjVIVy1OlQJg}e6fklrpykkb$w2l;wI{2Qys0+OlK4~+V!E}q=NO-vae>O|DvdD
z)hcINB|;r6(Xt0Xv#FZm$4-Oy`CtmtIv`K7o`IzDbP0&jA&;1r_>wJZMvFLd<O(0H
z_L;cdC#6K%FY~^UuNPsYr$6s>nmGzul})l_5>V*sS$o%5RsyDlzd>czjo1Z$Wg)i1
zm~*eNL8Ww63@7L@LSY~S6ef@5eiDNo6P!aea1Q?`UI5-6Kykny_#Z-~b9>tyLVuMm
zFr&iQb)sqff3`V<qu3uoazNyEGp~26__tQbcnc)V#Iq{ZIot?bCqDTI6~M!roFK|)
zo{Z6~m#69hHasK*2YeUayFG7a?W0}<bL#Y8r@(0{3@&`2sYTv2b|<3QX|o?5bts^(
z`I_L#)CN3)MzG`eAJE?hW}kFCL#+l(@=|BjgT#QME80xC3G~eCl&*mlDQ8_gtVyM*
z%NHYFbEa%)wY^b)SDSR4Q+YsEnauVfc{7$ShKlcM2Sl}}$5ad+71<^e<HNqoUK37Y
zM@zjP$DKjr0-Sq_>{e4NFhIdZ5vr170^-1^B^;kKF}lfl&9mEH0gPdcqE!>Xf3;e6
zA16>G8=SVj0&x;L@7wE|+^~r!sI+HW7Vy&dV1<t+9bSueKOoTH#MVxMXf<U|EAY>s
z`w}zt@rnYF(U;YWePFRg$xadnqmXB${cn=zBbP5v(?(96D{mcc>Hwkdt9HptZbFGE
zutX59T~#~eX%8-lDF8;w3x31(w!)F^TXZXo?wP|+MQrzh_=s4?SDc#<X!i}j;l9ei
z1;XMXPg7lK4O+l-tvc1L$$}v<DYdb@<l7cuzxCqOdc6V<UWt1nVOyomoyhgRpFK^$
zWjGyM`5XY!KdF~MqAo%;wh{rs?-^KeBB8O;Lr;|ruv_kW4**TC#9%<Y;`e#N!qQvN
z*TYddZm|WlM%QQSxq(f06fsA9hbNQGI@#ByIbgYwubhL;qE?6|z}e-V9vw{cu{G+~
z&lLf1ZWa;yNP=b0&;H87`y_$%KlQq(-S96(MN|)*LQn(1NC<gVl$E2WzkmPEsI^*J
zh-eOl0B7!ZCK^9I0JVw2lh&9KZKO)yK$P)6y&6^sC|p6EQ4Q_+*2ayTbuf@eb9kAA
zYV_}Z-2K%c8y<}3Y67SP5ngGmZ}u9V$Oy*XM25%C*WY(NrCGifqu)#XbndP%_S1>s
z^QhuKNtwJs{~I|N4>pV7zoZ7sKbbc5d(=CA4xi{rx1#s*=#0DF`5T$-_JG+yqu@W5
z3!taU!3M(L_t7?zU)ue!`Fy-|4Gb0tC@FV-{VMrkr!ESSW?Up3V~J+`q~E+d?R|ZK
z+2FjBh>jG^p|5AR4(!7d|2V$72l*ySq<=OQCUW}@*`n#oB9WG$(Eg8SI^{MC!VE59
z6#SWvS*w>ORk!B*$(w|#(ojB&iT?ch<F#xuerW6dkNaL$$4J8V&nb9J!}UEs>4J3_
z1okF-wo{Uijvwe_VZJQoBU5tE_Zd*%kX-GJLDh+9P9xa=;9^Ja$|!0Jps#eH_S&Ij
zET6HN`=3&ZVmv1{Q>nJ!g-ri*YRz=LkB>mp#*Arb8k5{oJLF4=Hbd}vZBD7#-|bOJ
z+};|<;=+0E%b>Qk?ZDxSzn7<d_V0iltKyVR%U^$pn@|j@U1LN|B4c|D#4g_ewP*R|
zhb2iQwy!Dca9l~umcFoi-<)WG<kEvzgTotM;Pe_uf#?iZfFXaw*2lPc+LU}lw#AF>
zsmh=$Hld>-kxLta@ElN&<R10+dzm5_1)|H8lm9a3N0+lPwBW51864Igdf$2M?uSV1
zZnRS|l{@cffk_S?;6zD*vRKT0Sl4APXWJ%oO{u7jyZM`d<NBM`i-FpA1C(#sM|_Eo
zXC1G0YUK-xcp0<Y%T#Sk5&c~#3m~Cp9opOeEp>JBK)@{3e)!isq8vr}Z{n}~m^noK
zQ85*APhcwk7D>6u)vh2rQ*M4#rGWBgDb;HO7%aWrIIYj1MMPgoj8xq^CNbPV-!&Ta
zl|OEBh%h~7e`Ew2cZ819t2~4Rw<72mKnP~Ud6=jq<#;iH{KG`&f>M_C3IS_*rAhu=
zk$!bmAHeeP{D<>p0cAQxEc3<mpp<HXJ4zvF%V^CyY_lnK840BtLLG2nU@LX%TXr=E
z?eY%-FwsQ&&YMaqh?|#3yLuHC-gIKMVdu`uwg2p?rt0u!K*9CPOzYAxz)L`;WkXRm
zla-m`2F&a`Hu!URZ{^>ptP4$jiHOSiJ|wEr_>BB;g7l@&QTcx>1CuPBA8&xJYQ*w;
zFq#o9-v>53n3pzXl{vx7wjdrqe6^A*?}KtfSq~(F&yPoyh?|Wl`3JxPn{9W{C~b(+
z{X(<RTpWZL6F@?zB3}W~6xGr^!0Wv1JSueqdXXrX1ZETu_6yNZ%zPQF(E=m}yU~J#
zj)%luck9=P&NOW9Vc_y<Ip+DAZ0c``VmCUd2x+ih6$8^2nKc6uKGgGz)0ok##~}<l
zdu#wI?^aO~`a;aA&J_8=d`t#e4y?MFLHVn3v&)Yz=-JqC5X)vXR!+mykE>_nOQ=Y)
zoM^tz-?XzM#r;v(^P2-Og4ZbA{&rkT0%dL^onl&~DF-OLo&%d8vCzpd;Y-sAnBSL<
z*H1@8a1dpy2RMqz{^{@dO(mm8iS+Ngl2CelIB}D%8bnZ6aVG-8!p@?`0XN7gUXIhB
z5PwqPDxj>Raj#uqlfriSC}<qQJEM3%^KLyWEV{$L-z!XBK>>~OvpB4+MRIRH)1Z`X
zSda0jSvosveg3r@k8v;2>Q#Wkchtn#4gbHw4>#b@TDIm=MEnIp`4A-QBn|)We{ky>
z5X*t+dpIQ-8QCO&N9O_Y=$USV>IK6fP!v)~InTdEC+tkupC1YIs-3l$PpxA_1aU6z
zeIV+4&t3*(cV-$sm<pm2$%{>T+J_A^*IiId9GdU!uDP#*c|TXZW}X_!#g8EwK2%yA
zk&+E;U^a_zWGJW#kCw2Lf2ga)B*z-_Z0bvNK!rb7zp%q?7))s`zVdGHR&cG!E6l&U
zdEnxvrSIrU$YiiT`E1V*2?a|+$%h3tT<N^StceOE8?cToo(K!TNJt5-pk>yA$OXl<
zEFG@~cj?k^U3W7uR0j23Feo=|t1)Y>=-l+o1gSUtp_CPzXr>KUhqEQ}$;eH#Qcnzo
zkN<hw1G#yDhx{%nx;TOMi@wOxcM*LtGlHaOtE+Y_Q;SJWJUaaO12)+rmzwTVOH3E4
z!cWa^)}lv84I;b}ls@jFDHwJYwRTeGx}G?1;k3=~pUT3lSk0VZ(b37|w9IC?CdPv0
zud!QPMSE{w>OAi`?o5v7K~Z&4Ow9S1iA96|9=RmEh#Et`$}~bnslLAc5X?Y=Fw{|Q
z%A3lbSh)TM;PML;DPp`5np2=yF9&LeW<4qee2yO;V*VqJI;HKTILKxODt^4g9U9;m
z^ZiI}2{pw`6+-{QYf!j67XD3W(xRqUhsM_gzG=}?XppGst<+B5tmr4I^%}U*#ihaf
zd9;2frwF=`duzkJSg0=fTyXr%e8xu+4kmf~bvMZdU8`(kq##r%F_QNWn=f420FM*g
zUISLNM9q0qv{YWv{IrAhh1S-Y@ZGv)+-2S>FgyMD)nbzZr^Kvtx^b^%hAn=N<2v5d
zckNwE%M>Z=WSc{AYc1l?l{Rd$bUO@dqq}h@DZ|C|1&M0t_5Z^Ks19=P6kkXK3My}>
zYpg7C$;#H&Q`5jw^(<I#Rcx1QP6x<|)FNzjfa%z-#7FC(i~h3OpnG-a>mzarH4&w~
z_F@v(Q=`m<q&%fG>+7O-?c*BZg4yUN4t7i#3UJR7-3=TWN1GQ`n_en%O5#)B4gCp3
zQ+WGs7^CJ<V5dv;#D1+{P!tXJ7=;#T8KB~nJFQC@d)>UhHlb(_T_-Gr=7Q-blb*Uc
z96Aw38R0_Q8>W0p16x|%mqYl1J;{Il-VplUnne2LaQ0AjZhqm)Q0eHvnUI59L|n}h
zrs8wm%UZ@WU*>-0g&)ANM(ZQGf`$c=9*m785Jf^}#m0B7St{gYM8g+tYen)kL4heY
zN_Z2TgQ(C87UlHw%9K(B)Jj`#p?89tY22XtV}$Z4!^OSdm$f1_?>21F3km%SV#}2#
zm?69e>7H0H+(S!ct7CeC8`$O$;G>hOih`CJWnQ1~3<ZnEN7NBL|Byo0#@(ie+;!rO
zX%0KJocVyt?4rE)l^nM4?(gH1?NbifE@n#J>8)cY8{&-Sl8=Ki%M3R^<>#o`*`hN$
z>2h4BCd;iq8=!W~qr!IIn_T&fK6Z@Wt`TQoxWWo^7(%jYFHZfy%5Mq{ZCB5KF(($U
zSf#u;<s?_Y`&i~HfltdDWs79`$k<hri5JWlZn;EW&NlImri{aA`Mb%cti2xiqBIjI
zm$e<BOsGfq;!)u9IJd3e$W5YW?OQmrS*RRJ;;~r$a5IxLpFDLU19sK(@Ge9^)~EpJ
zh_sVJrEenQYN#Ds%r#V-wrO@moQnI+03q1eDxL+t)_e5(O;$56Hm0kK`cvH3rKC<-
zBBiO^jNT|Q;0&ARE1*i0`p+X>7v!aK>5>heTmrk6KExgn8iNiVPKh~Izec)N8Y}rq
z$s^|xXDaV&J36eSsHkO^L2~DbARtre1$Wl-IFKC1kH`~EQ3L-sKI|b+c&6n2;rdCu
zN-_6VS3Si5sYT7kiJL56SWrySj`+w>I+$ye>5EL7y)2z+^=>ugmGRtd*d-8Ro<f0r
z_wcOb01eREJD|BAUXBs4<6ldYH}(1oNQ=tWTF)h0e@usO4!k{f0&5iQv(b|9cWD|x
zFsIeDtz;=fOqcSklT?4(DgH8fcn=Rad&Q3OlXH!kX9BVtAFpKdh5}edpPeVIHXH4F
z)N$`=^5S<&V4fm0$0WLTx*R4V(4xO(gn~np_=(3(8F8V`C{K6$=Y5QT1|*k38&cRF
zcNLorj8G{T+A_cYYC${xCOI_^|9cs}MY^wvXIjSI$?=gNahRx0>dKB#R83a<MJ4Ri
z!%x1bMgRGCwhsGVkdIcrKDNehc6-$+?$>Y$&G!0;i*vg5T1m&|(`uPk?XF6*n|dc_
zYYWCEEDp%YRlrJ{)v)kZs>fyqSg)Hyq2%#-NsrG9t*yFZ9Kh)j07pJ+S%Cu_cpo)z
z*|d!r9$|vxk01Z{f`w9Mwfc++nr=~2>*6&NmfsFf^k=xBj)_QEtWV8kv{sFYpnbj+
z!{lBL9Op+%HDuV!(no=j=tSfDODytBh_WlOAfv!!t7jERkioU3X+TkR3oWf}t~J7e
zY+YA!c*t=j&eKOmZhp2AtP{r;-doGJO@c~z_SD;d%AT;j<8;`_g@xdkO?4ExxD^nU
zU6&CJc=uYW<PIKby=|b^4<dokF%&UEExRp-CY6kh4ZT7H-AkE@;iayx!GAL`--asx
z(R7hh|K>%zF=Y)$Gq-6NrBST1U#wqpR4yEg>K}#5NM$KTD$9y94<dVA)vtQrF``qS
zgdeuPz{qH42xJ#^T9eR!|HmH*Wte0r{ge8`G0v}f%XhbXcfm9rnkFd>1qHyI@+L&h
zZxFB#I6t{*$qvs%1pAtlNabdX$LT?PioSiN<R#kTm?Yn6tHgB$mL&EAyzn~o;q9nl
z&H?a}XvtshyMXxVT;iG!HxCz%A4>AnA%D0roaL|mOPm^huywSY>@fJe`Ht7)JJ3hS
zvd_%(X`SkIM@L8Rq9Ow*_~YL|EztDyw<h$ZV-y@bvlw&F+j=VU_;K40U~@r;Jq+Bg
z)WW@b#Wu^K`GuCMAImMLx%1IWB6k;I2le4V2^1ZogJWg>=x#MmpTraAz?oiqkw3P~
z=9CwUVV|OCkL?nglO{d_Cl>1W0PYE9p3Ne|d(WX!;ebqmQFV91NLto%SqKN;inG9r
zV<gF@4kMRW=X@nfiCitS+4HO1!eg@-Q=oJl!<zEQQ4flWg2M?FECv>D-#4#_Z?@xW
zJGNizT<?4{pu%I=-!2I%4P6F0Mm#M_?w5LnTg4fk&A)#E?|?<8$rkGei)JM^C)(e$
zLLT#x8#*@IeJx}8<BO46ErPRWu|CfGtmx_jY*0is?l$yo;pkmAn~a=2ZzyY;gLGr>
zX*4x|O!B0r?oX}J&+@vg{0S5i*Cv8naDfcLkv;Hyb@jw6x>8*zvj~_!Y4&MJ-Jm)?
z(VIW1UJwr8Tksg2B)s)<o^5a{vYQ8KaEHgM`TLiaZunWz{EwT<8%2GNlQksx<C=AN
zb)g%L2(>G;1zk&(DY|+^3k@Gj&!^2{-&rRR)NentkbKu%b$Gji!M_CHyE5)uLh!&+
zx7qB53$Wy*N+Z3L#<*ed{P;0<VePe-SIYq=N^@UsS;scV&5g6^Fd?|Cw(k)iZ(mV7
z^j*V8u7?zW8~k+Ko?8>#H6-8l`)^kHFaPvVxGIWV$n~`Bo1F#yWY3rn9&M1Hs7Iw|
z$ch|AG0R!X-1C?1F+U4>Kd;Vpl-v^O<V+pGbGCr<MX=36^?S>1mbjXyF0S@(edU)f
z>Cx8lt9JSAfGul#Zd@j_8SxrYV<A|si`_~^laqv}_W8)gRhN!{dEqFEB6*};pYW|4
zL=Oe#`~XLIt;$_&Bvc4)xIU$x{8Z9`Iwmqr=q{9-CExLbEOMcsw!P*fA`~ooA0ICR
z9h=1Bj0ZN-VAiw(_BUe?=A=BPYdfebm+fvPjCt>=N)kcix^PJ}f>&kn5=fLQ3X2f!
zF%uQR6H9bRwz<qs-rdl6I@es`*uJ`#v*D^!G~N(8Z0Xeh4lT7=jj9UvB)qrJ7cEs6
zRgIi!LWpehxy?LeW4^gj+jU^R6ob9$Fh!l{ltdJ6n~I_86s#38q*ARjS4Q|$k%uKJ
zk>BHHN8e9EfE`C94Z2y-Rkv+;zTbft!davq>nPNZR$78jbGWk1>YBS|;iR6@C~QNx
zI)q(7K5d5+IklEo)~%&>1|*>3falZ2HAMhMR8`Qm;9k?Rleem4<4(;?Fw2@Wr4z8#
zQ7RK@;Dfzfe$dKif#np!bG%oQSemyOIkm4@thcaDb#UI?dGZmB++cJwvw`##KTmqG
z?+D;+xLQi-t5lDaw%}5^Z-3WYNLo(>BJ-nlzF&kfA(S6kPdBUQ3xlR-tR{*R?6^@p
zOGH4W8qBdbUENow`3s0<H?{C_9Zj(oVWa}-6+^UUUDko=*SpPUddrLy-<fg=<Q6r%
zTa9`iZp_wKcAsH2V3yS_VJ|n!04oHEvto>7AgrvC2e<NMe?Z110Wdu_eUF1LhW9U_
zWqDWX`n90d{ZnIHy#N4ma?hCs)OA@X_cvf8MgBy+c`AJC3dws<{L?*%MQh8vv{~Av
z$D6DuMS&jI*0SZXpoYuTb&zdwFZs>BoUwj?0AWpTh4&o6L>62_#$dvUo<buxV0ia#
zSwKJIz0DWLV<?W+7LFfJaVg)&;Uo#y2C;8t&4$!hUr6<s0A=d=Y0qBko%J@KC)>#}
zl<Jp7z2`D9Dd}UryOFc~Q#uO1U98nBYB;Trj~mEke@b2wf>b7h2L3s)Jg6TT87=ay
z4=@%o8OEbqr=C5DJD$hAVJOCk2ehl*GdFvnqJFXGeP?k^tGf4Qb0e%`I^Z@_?#=)T
z^b)VExT%zA_AT4h=)i~MJv$k+RB^PG!|aDN`^l<(14f9t7cG-RG-ho*>8Zn_W;Lr*
zA|v^tgHZ2nJjVs~ypacgUWciyjcv>f|LWBXx%T22)WT0&+-;3#COv7+A*SI16#lT5
zYt?<67yHG_gD+1$_L$$4XRo`P)OIi$x5DB+O*JbFjAFgfKlz%iF4Eqd^GQhgT+N$#
z&hgwY=(p|t+z{9MkSz)oix=U^ta@={bc&8?6=`zw?T_qUf53V($NzjHLK74Zyy!ek
z1y9IJGXF5sYub&G^GwKjYMo4ZMH$}+W~x_@w|%qd6u04S0h&+2<&`%bXDvpMFCVg=
zY_QUhnllY(RBPenn7>*5Nhs8$VV~go3<V~Draw+`vNsN;v*wN&@5e}W3Nu@OcgJ4h
z|22%ajx2Pj@hhAsIa>|Ob4TAjyb_E|l#A5a3@N~+6gO+@fjdZd7x0k`g;$V~UU6-?
zI?h*qbZgO6rGcTAjO4{2oz!n_mA~@&HFJLa5!-xyqw@Zx6Hc4y`m+e!In+xJ`;w6j
z0qLvTvGP?gjM#HM2mXT(Z)$-Op31XJauL^k1G=X5l8VgcC*X?a3f$|E1^Y@V@ZP8l
zPzD`bsz=aKwbpvvUaXC@c)EA<99Js>`A}R>-Eo@4d5KPV#?@&*H;icR?$o@zVXr!<
zqDADu$y8Lk)yAy|y0@X_xsd(-bxK|>OgSN?RoI1hvP4b2c}}%sK2{BgF*IzC=7pIW
zYbbRCB*tWRmXSm{-`7NcylY@kx?>x<Ow(N73q;8&fH^{a2*xk><MrPsxuhSj;79$R
z-}&J{K>jm!ZX(wjKl6vG;1#v%2RbckdiJ*{CLX{&xd5YJxq4v19mxu&&2Sjxxv_P4
zcY7wX-1V$f`sG<@i;Wu~4CDYB`^VhPmurBi#l!M?3x;wko>m4tyiT=JAnCX-$&k@#
zSwt84Gn;rqL4tdy=+joY8v7afi4PJ=E=W^^t}k|@cXz_vq3|YPilxL3tliUZ+wxpA
z#u{0`=;~V+wS_L;&2s)N?dfO2S&?fqb4=}XdPH$a@qeo@U;M2w-77Fto6+5^FkNKY
z*KQ<&@Z?U74e6_=zYpFwWQ~jD-IMt}D~+nfnT9;SgsObJv#eY(*laH&nsV)6?>v(c
zX=lzC5v^L=$BNM)R80wD-OMNpjJ%FS*3PD-SSBd0m0j|F8Xs?o9K4Cnth>feq_#;l
zNaI!`=#PkTA*J`zoHaeBskjHEJ;fi28`->Tqp2quYpz9dEN4D2-<%MZJ=u{#PvwnH
zR^OwLEpq>wJ(15*pcY2lBmAn@r7Yt5>~_^g&Wrkt<PRwd%xfp|0Ig8)>2e{r2^xMt
zjgQx&j<^`P<B_@*8-?MB`7kJp%HLkRDQZgKfB1ZVyC5-sjy7B_;5#<N!nmcVv_THA
zV0WuN`>IdzK6Z6Z$(u4S`FI+4LxJ<TusN^QA9Eyht6qyiu_(n{Mlig*vY8U6;=$t=
z3SaT4sg61ORky17tXJDA%QhReLf)K7(o&H#dH^wq;3SFjMrO(KBe+2VM)KH@<v#aY
zRM@bWXHnh942;yQn`X%cK~n}>oL?Im@A=sZ)6<?{gJZ9cng!hau=#bi1sHKu@F-q#
z=9`ZRoPp}G7mLNBTp(WDZ~C%ng0no~OhO$r39}2{cb*sU)>&_j>~T$}S>V~}$SLte
zpU`|!6EofYO;PVwt0=Cg@cK~}*rdvol&<Qv0@%*Z78svCx0WI5b$6(|dN$csIcL-v
zMBL?m4a|d{;WKvEUwW<wK1~e=vfhb>7}g+>sAEA`4aCHJgUPtN^69CzUq3qj18L|k
zYupu3d2h1g0;pwOS6_h*%ejt2Nv}scI;pFW2^ey9x!bb+BJ6+vf_|^SYDYyxNqcx;
zVF2qB866y;2u275qbo)o5y^`X2*zeFCp!Ycs0HO8Js!lu$)tHXZ$6@aDWetQxt`%4
ztnUSARs1V!5x4fz-t{@&3bsNh^(rpEPy;`pf@Uqis)iK>_AzH^4T1tYs8kQ;IR!nO
zVJk*3NEM*P`i+1s<`@)D0ZZwviJY^)27UGB^TGX`FUnP}*^_ej>9=N)BO~ATY}oxz
zNr%<c8wU?Bm<&F}7H$??c^w}4R@1%;>a>%)Yn6b7?N9FdLT9G+S;2Vd(M!@9Un}MP
zPE_U7?lZ5o#Br#PQR=V?J#WX(Y{i^K_TmkDZlUDW7Sq*_Jo>DS*HoxWt6qgR(zN@<
z7VkD1!sx5JWUa+H7Vi9sye@5+XNYGpp*uKG&s##*(^}ogLpyNNS1Y*9z)6qrxt)6}
zP7Gd{<vYK$yEE@Qli74lzrl|zU)7SmXhA~o;x-wKE-Gu8w#q>*`247-?FgI}D>P+b
zTDi+}D_XRjhy!WVl4O}}0L!UaBY9t?KH&sUiXGdWkhcwSCQl(1)aH4*xz&3B0d=Vu
zQ|}lr)|!m1j5Ok*^6CW66g`VCND*amB1+?()tlE~gEZSr1y97UYrUH1;oL1y*X_C;
z>#fMxRk}LoS<HS9DWNUodxo_lGuGx%rN6+E2%cN$=^2Sek<1tE6W%u+o#cwR-LILe
zOh0>_ZUFa`7dO-EJM|eBxiKNt75VB$d=BeJyTJbWdExCup;})*Ll?g~F1P|V4I7Nh
z?rNRl9s5vjO7~T((I}kJK%g19^=8}s4Qf3?AZ}qe1^E6Qf;GE7Nm5;-+3laBFZ;i;
z@Y>#h^0@~*BNkAPoX*tVd=|AaA^?}?YW(}AyZDWsv2(G&px|y@hbznxhy^F5rV-DV
zy<^xfHiZ&?1yeCHzPCfXoL9nPH`PDB-zJkNWIV=MwfN$qFXqDSu6q`GQ>}WInIN8K
zg75<xDqA(r-D5pzZn!a>s>o4D;F45?97u4BFH(RP`W+nkM;u6VbRY-3C861;Ft%+F
z2dDN2JxCNis4$>_5dpMie}VaQa!oWw)JYtdua{{Xyk!iJf1O(S1n(Yk7$D1&wHn|Y
zhZ($w;z|^gp}+{Ta9Z92>dA~91L&#%Ft~QEFa8ny2!LXcBC^3e7lZqvW{yKC1;@ry
zz0I$B4I2&-Bd`8eO@>+r8fRvl(o&U43`Hk&nZD}q4)t<r>{c%co^$Hn-@N6;)GfVt
zD%h<aqH=t&lW9qiXsupWUbHf7CAR=ff$H+R>_)6jj=`S+uROQ>!$U0D0>|vD`uEs>
zE+3dC<^Q1!1hKQ3*;o=>zi1*3&(5;6GQx_PknmuslsTn~r~CR%mX<O*4t7h3YRYaR
zG66&lk9*OvvAtlB7+jc1Ue0rsaS>EG7Gw_;3%OA^FOi{@zVaFc5;zcdY)tgaCs*&#
zQ=LK}=t`)~vNZ{94w{+?Z-rEgD7G@Z!)^gOn(cP<hrmtQ>|6}~5scxv-v2-wMSgK*
zu-#5j;tB^Vr0InM?n&{XJ77l3`9nR;`?mfX45>0>3U;Ky6gOD7I<f%+52zGBw!Krw
z!oSdZr<4{FAkn!OU|Gt^+h74_zW_RlCQZUgIMHg`S{>(cI`)ACjuS!O+UFr6h^*k@
z=`i;Y99Zci_dbAOWrInr6tNZ{E7W-@GIMitD>Tm4la+h`TYWmqxkT-%P2;w+DK7J7
zvzGLeo_kgw1YVRV7q(k^>Q2B}+>x|UfkrQeXUGJ6jeGDV9TOcZX6lK1wR+t6-qtDI
zL*iD{xpzRWZRTox@d$9lL(aOi$kb8Kfgvt$;peS}%@9w}9xwr;+}u%<6EL8jxtcip
zbD2Eq#rv+5uRZ^qs;vdaGgP*ReC&-nwT94U4Lt!o1%i^@qBN5#|92;zBVY<)4WXU}
z)>->6cRb(yz6pv!ff1)Y27>_Z(|~CX&5ED-ugb1e@fz;{<<M$d!&A3dy+kWri7oGl
zx43w3pnBv8tOrwCSc%Bm5T);}4p4M<c4m(Ms3vd7gDcMoLg|n_*~Zg&{cbLlyE7GQ
z7rYur2R?uBjKbiSN(S>Ee{?KJ<zqyfzk+BMbhZX82}Eh^5CDMN;>DcHLJW|qY&Vl$
zu<v8*qLdK7$%)R(>jCXnu%I(6aaBiBW^ck}BWIU-C}b^<92n!4Qh^V4Z1Gn}zWMj?
zU4e`nC8ActqOV0#OU<giMN5@zLJ#jdvmy8%BX52}m`s}8BoU+$9GAJ5cF#P+LtUTb
zc_9^P|0bE^_y;-~k74de<xU6V{E>y0$WLMJ=M?TQQDD<ziruo2Ue_12kxTRXni1eW
zjOS<X?EOOEeQB<N-OiSnHS)jVs2&(~H#Oc<tdU-aF{v|f^Z6qT*$ocv<`;wo;5scr
zr%oA^Iget~&jQOlP0{vRH#u^WTK1*8mt|=IU`MZMnhP+NSdt<9u&;tVNG0u|`rbF=
z=QPkx{3;B=P%xmR+Xl{N_lNO@sh!5&KJmnw8)16DZjJ(CXx>}5U7)DNd)c&I(HG2l
z7$FU<V}2Vio%5S^ZVyr9A1HNL5R21Ir+j6r`cZzXTMwdSmjg|>&)ax^1qWIB^mf6U
z?XO=Gz{;s*6T8i1c{{VI5#9?ke8mnMsLB_hsV}3bsQ6I!`MAuzIt<$Z)q@A{cRcyI
zd$K%Zc`#p~XG$g}c=aAH{RJRGxLQqkcT>7h?8*PAMydiYUJ->DbvdHt0`>Se!TDhy
z$0|_Qb-Gq4+KZ_YsJW%+^F=~Acmm?7@I>HSTJQpYgX2NL-m^GucGksHkf<$E_YJkF
zuqj)@D^=j_MDu~UkA{7N8v16cLn%+gD(=)p2g=1p3A_aGv521_z*`E}iN=>(x?Csv
zorb`S-5wPG(x!cZO~8li5D>Zg3T{Aaq($Yn5)p7h_U$W}th}02O8Yxmd2kPOuZ-Z!
zn2$MrMg(LZ#96Au#DaXwrh2Q-^Cn5F1L$DZ9EyJnkKC+!wCbjvV>>vfT(i9I{}NS~
zO<zp?dyH%kj9PYh))(1d^KTb=$|1iQ5h8@!qh6&S)Z(orH=%IAzXJ4@LiSHOVX|_o
z4XND@21!I72R;Ocr%u^?>r7Kt{&I4s1XN0_cMd5;9?oIWdqS$iOyP7z<KmHSe<*3q
zG^sea3Gu^J7x{N4Cn6m-WMm6GqOT@-I9slrn6=J5Xhxlqw=Cw38@;2iU3ifOf#U3V
z$`m@{H~RZ){nv*&$vKD)D-BzvQNZ+9o`yK21d=B~m6TU_j1UEeoXt8!#;jhP9S3f=
z&*BGYfM<NrU?2P#RBu?IOLDNa{@oA+EGFrric*rNomR)@J#*&WNaf_@Sj0|Ow7j}4
z??Im0Eo$!Nm%3n30Jma8aUEcYmg%BA-k2}FneS@y%AFd_)nkDR598RJ&pn8=Ip&uk
zb^sCX*AXG3Eg?Y2S&k=!2y6jAfEu|FB5hKF1);VdY}se-F@;8H*G-MG#NTVX)E3HW
zu<ov<FG-CvN3Y>xo;xggTmO#rQ4;@1sDMfZhsR}Fd8L0)?}w-noXYIQ(Z-t>q@`_@
z&e2ZR-aU}4I2o<4{$5tf4IA>z@lK5k&yDp3<hT?!)}2Cn&u&;G7jPp+*p9Aij76Rt
z6$>&PZ3{<c?i~q*3`#he7~bhzCmZq9&zGQUA0RXTGoiMDDwIC`SnItp)Fh<ZQ$xHa
zdk{IS4DzSwj!E+L>xx0ocghuo_4=eA(2smDe)QNq0l}c|9U}$;aQ#1C6@5wDTADXI
z*p21;BLcif>j5$mIYCRs*aDP7>JPde8-oDRWxF+A^c0A`WmK6C(^7qfJt7cYqy}%y
z5*<-%*Mo4|?c9tQCW=tG!<v`K%_dJ?6zJ?dGCWt^4eeHY0E_?ovfRL0l>N~FUmuVl
zr)VrNfu(W*@^AqV-gkA5_gBBYy|NkhLR77;RUkET0Yk+CFw*;Z8RFalN`5x$Oc4|q
zWB5ySWVDCs;O+fjXCp)uPWa`y;6l;0f>l*Kvb;2p90HP4b+O5ohc%HN%v$@w(Zz|6
z4wuf0yq<lCI&7_#IS7L17@g3U-I%!w9Bnr;WNm~*M(U1kx%sM!;R4w&o4UrTH$h&H
z7&_&<&v{NXnzcA4kZd<D@v==f9;J!d;rN?Wd%PJ5HDicO+^yFysk&J<Vov7!wWlQw
zOfmIZ{fzW^6qN1F-KeD93ZkuanIyk`y($r~-i~p&l9$RFjEvOmG<08e_lo?av>ORV
zESD*KPw88b##0ErLlU0+uXjX)4?`g?LbtFz1<PlVz{RZlDEc9E`$f_K)A~^xD&EUK
z1$TE~{FvF`@Tv?FB`*QqOkiDMKL|Vc`|Kd@)HM3FIauB-e?L3=+R9Q56Yx-B2G^}F
z=r<WRXuohFdsCItfV9Tq@TcEyJJ(cI>r_><0WeUdj5oBZn>>)XAXh?#B_JKL)YcDT
zL6RJno&oSr0r7L{OZ&0G&k=2_KoptRBlxbumO2HaJj1pC2k}^t$|AR``;nf&wj0&Y
zy}M5?EW|^SG>LsGH#JDCT;o|l_bW*&FClx}{8@pM;)zzqq%dj8!;<k#c|I6#b-@oZ
z*9qKzhsDhpy?tkQ>|fI6?1k-(|HB1X<}0`O{F4JJ1;$uM{CL`CzR`PgG^>po)aI2<
zs`Ocr+2cauBj$WDW8^qZs>#j)vQv9UIR?%g-(x6_El7Y?pmVCzyf^((mA0g^kXp~J
ziR)^Pa_X7CsX@KvT49H;O^Qp(D1G%G=18qAqT!*++73j(C4sOqh_vju3?_$DAXpor
zP9J0c&#MIr*d}Zs)}hqkvd5mpXA_H%w3P~@VvAvQzoRs58svW4)$FvTo>eylMkrPa
z1%RE&1vaYqcdAAo8VB{G`eqY=nubSyKmyS=i-Qwy$@eL*?P$KbXtiyE?LBpTNwGo=
zERA($4R+sshUbwGcp<};HBcqA)O!G03LUBh9weU)#LZ&c9GIXw-#|`89EW+0@B>8i
zf%P*3ac>#{R|D_rV>J~`g30kQ0ZX-P&-9(_Z;hLd2Btr+VC^nuf^{lEE!2z*TOyou
zGIgjhK;?v!yO7Ar9&zF(MO$BvU`1UCa*^w?YL-w(ME5$~q>!&HzhIuJR|XBko8GhA
z4}~Ezn5iKwQyNW-X5G~U5kxA6*yq`Rqoa4*K*Z)DyeMB!^Bf(F^W(RAi$OVuI8p!g
z;VXg9#K94#Bz*BRpQy>}2HG<)z>D2kWtP<{zFMB?qwYc_;?;D#Io|iY36X{`&-OUX
z%u;Wy4LUUuOo{UQ_^uy05TUP!0dXI;c9{2@+)n)~EYF7~af9dCSVb47D&z@Id<k3?
zm)efNI4`(y5V)fCeLn!zxaGBdF7#xRU?%wD;$o!MMn8-CO6<#8fE^|<s(qiava%B2
z!YC-HUG>3bmI5tx`D4K&CO}OR1#b-PjoHJ0ewhWt{n*9@G82=0JNeuUjUB9!S|^Xf
z)@w3bb4=%H0JA}Al9XY2476azBC<UN#9cL3$~EpKM4o4#K@@rPrZEBwk~d*V!@r{+
zJ4W}|{gkaIZs`k5_Mn&VI#i8S$RvgD9big4nnx-xQIM<kuu$!oD4hHl1|Gc=1vzxE
z#iKZz+S|)@t(f=cH$0i1LcVO%-?o)1f12p=(;FNfKI2l6_^*fl?>ksh6&0zR0{a;W
zFRT&?oV|eQ=n5#dVxpt5<~X?&Q;K;k;hTe`x{eVFi9E8Q{9Q*2M#VhNU7wIpb^#<a
zMGWqRs6C>V-hqNMS>g>jRb1q!(qQ8<WZX~w>1%>hcnU-}fR6gwq%K(hiv@;m-RH(y
z*`c5Wl!ce-8iU=XQ1;_ESFkbufXVW4!NHdzT^5MHJXi1zzAG*r<Lf6rqO*t}(8yre
z_V*SBJ^{0WUnO+DCMGk<ckbdNN_CTTu2Z|z40m&^bTyC2>jup)(z7W?<>+~Zvb`)j
z-INbU-Ibdb`9WBCP*(y55Kj0~<&7sDAQcx0QLK7fMShxci7zZpR?bN^swHFEI1voU
zHb#d?!|y++bh?Ymv`NFND*-oY$FEo5?s<=?M~X&A8-~q431=w+Sq%xV7;i9C8cg`Q
zmkXbN{Qn+s#Ipu@6!b<606dO$->sJhJ7BRoQDK!+=1yy1el81WUc!be0N>~RF{q*M
z>6AS1`8Sxx=z^(>I~c#<%dmsn|6SJxe4K=J&x35dTqyamA8>hI?iXvD12ZTt+eOj5
zxs!SKJz204M_aipu>!2Lcy{ID<KouV4Z$*7=H0^$#V=^7bTwZ-*~I^6NPEBm?#wMb
zoa}Gy@G}PyG-@r;Qd6wB6U}JO2(l7ON8jqvcCvQK51+v#te@YS@n^k%+Vk>8UAAak
zf}V}e=hpHA>$SEVpC@}~Uv~mFj=Kd!8NibqYZV&wH%2(2?uAwF|5H7^hH0ZNC78@H
zm~~f0M;}v+!l#tRdgm_5RjSs%_!J6gY%?W-zpK%4kYF<V#H?E%UuS<xoe?qk>1lAu
zalP0=PO=kQx({%K19&_cqWb?|p&6ge*J*14VJ^r7r2Tjr#$eO9U_v)CD&BZuQPrg+
zf<?0kLS=$3=p(S`;{cj6S;_}KF1rzIsb#xyTJDdaDWGyq<#))Pnfqg1ze@6V9-vii
z{(iwJwNTFDDHz7a05?D3?VHXDG4NK#&PMqD|C>viMHbN*lT(R`d)eT6li`%cRvBu`
zg6>k1i&b6@dipyZTxiRjFCY3A*j7R9Jra@z+5zs!ke=uII0bop+w>dtRt(kZ;vi}`
z{bFU<DVKFC02t>_XWO{wJ<T}$k7Ri|Y#S#t^5!8Jk;<z~vcu-^F5v>(bz$a#xWk)P
z`HR6A<qGSD=pBLMb}n2r9T1B9Q~x^x{~HVtown^gB4XkmP>Qh7Nd<Ezjuj3rcZEbD
zwkeA4Pm3Y>KEeW55m`VN@!+2mfwoKBgW$<Nh4!86c1oKAq*TkOOW`lj1=N|jc13GJ
zNts9a!fb>dFsv*lS=lYh>j8$XIG*yD&-xjVdW`~8QmKz|-T7)Lu;-u7#@YU_4-xI3
zmmNV0BHBvG2I6M?uiwMhh~OMpPa*J^E)!h`2M1#yIy3_GFO3p}k0fZ`15n}V_6XN=
z@K@Uq{DevC;CGm*qpWI?vEpClQ!2Frva>0=%BPIBnmCLO2vHSKd3Wcv$!8~ZN4P4J
zr>Ckdzds|tPLKYxCfKq|KFxOS5ciDoG=5v}*3V9H&S5Zyu=Fx+%p96%q%QbMc?6eO
zqC5sPc)pMeiw@|Nz?s8^)dXg4_uAG^(D~%tj_YO*PL*WaHyhG7E3bBV#>SBuiq!4k
zO;sCZSSCMzjqSYmAYWhqXuST#=Dc<x+=j+%YKrznQZxz4qnn$9aSG<aEA7O&DcGZ^
zsVttm^kRyWV#^DUSh##BB&$Bz_JUWrQkJ`?&?!6H#O!yVxO<zr9;5M>FI|v3oqB~A
z-drlDCHjUn?t|YV!NUM({nr1(+gHa`*)HuW2&hOHh|&s(fPhLj1}&g;qXHr=-K7E+
zT~Z4b5NV`q0g6fqOG)WwNq5(o2XKG;+gsiH{Bh1d>-WA$JkQ*7Ph4}&jLuKQ!*ZY(
zf}6sd^uCp@jip+*8%B7uEeP%L2h3n=baE!t<<tPl<*%ttnfXYY$Rx&40-7M*P1kvD
zhKhkI0u1+6&=&b8h6cxW?&vHl815AvK?$T<Mt9o{;X`=%o%&5o{OeY~M-bF;vR596
z1uaH+T;r1+T4}f#0aP(tM|C8L5DUdxx??Dn5S^*6hXzl5%oKTz7A0PZeQl!=f#Ach
zOFpG|o+KOZKcA9~9Z9G1($9?TG_$@vK0P<z{4^Va1ox_+qX%N=lhNxk=vH$QZ$CLI
zigP_89;md_q8+v~?VYaNx7`$zC>iS|RcHEIdl^$D>a(D(b-O#kh&I_>5m#DT_yohU
z!~Jg?nQ?73#k^Iu4YVg$1Y@h>_YmJ=41gpT+O%!_kiA02oh@k^N>F!V=k2ytlDacA
zH)riMjb0vl?Mq`#C<xQOR76(#aEpe5W#8lu$F~uAVSj2wPXWMc%p!k%pVtAnb1|OJ
zYeRPwe?S@t(zCgmFUWR$Ob>7-sVrH)IPTQc|NUyfH$5qXQ~nv<L=PuGDvLOkX7cw}
z`!w)`RE<ssZI*AcJoh~kSXxhszUq{EOh;9jK@`0hwIMVtS70iP=c$HXh~XORse5NT
zp={oPIc6+cUcV)#ir#P)d+yYgs3z5fa_f57^Oj<j2~v_ma=)3J^2#RH0+02)Azddw
z8*#G;pITX2!!{Daq3R~gVcEW-4@BLG{eA^95-lcMJ`=N6L@wKxC(T!l2TZT%V8r{$
z=w@#8Y2?<<@7bi8!jcBjZP(GZ%($a68LIUg$G4#!aIT)Ww;E7h17LvAbWTC(_*OSR
z6BARZU?)!cQOEp>vxUT76*+k1-e*89n~{Q*UzGo@gl5hm_8EWIwDcpUWkt^T=<iI+
z&ey!IBS-KgwKfL%@oA(mRU()M{o&p`81z}Yr~=!=l%2J826Zt3hVAnTj^2&mW@>bb
zrA;J!2C1~1iZ<ZRYRoNo2<c)e_OW*qNant=4b!)YZD_0gxLRT_WqhS9!SJ4pd~CcB
z{^P~c-tv`C@6MTz@Xe+)G%omy1m8+biEKinlnRXsHdZ=h=hm?~Zux@(y#uA?;$~+f
zTT<3b**&oK?JHZ#{Ke*il6YeiQIayPN4aCt<{ivJ8b9JxwbdRbG~DbQG_3kKO{GWJ
z79&eMR1?)ES@$+zLv4%rP{i|1DB}TJ%PbfisdNC1umJ!CBD``ip=Y>sz99tgz3T6h
zw3GHzh+PHx9ZbHy8gFf@tBYz|(#$twwk)0XfnNI`(-MAFJ1nl29TwNAV-PvZLFC;0
zZuH=;ge@U*zGt}$T#&khN$~9rJ^5$-mHJKoTPfp(?Jz>o*VEFR<+mFXczvJGVfdYq
z-@cr2s|6?Wn$yc9p=tCc0>TT+U-X82U*n=xee8+$nTiQ8GJZaq&(6wznT2V4p><r|
ze5g5tHL1%n^}gN6i&~YKwFQ-1#RE@^BJ|kmr%EyUVS@wZ+rhEV%Oj|tmxWLhmWMTN
zxZaA;@5m2{*No-h&l@jPiDUnPRtgVrl(A<=P1|u6H53dyrC{kTxff;@#joSOJ*SlX
zJu8{Y#ma?$ZR5G$`!q+Zmr5lW=3x<`)fgqF+G1Ya#g20P8m%Ml5-D-E&vG=Zf+low
zH0>{}NGzr4o>_0ryhPK+Punsk#Q%0Okp}D9w_Kmy_ubvPFQERt8xErr1Tlw%1W`Ei
zBlRILbgyotKFC9HWMrfkdRE18o1Z#EiMFry7<|0B$N(lG3QWSgTTMrIZ$R0BcXMC2
zcHJtSE9Ae#ag1y`F!{frOS}IY_!DlxaX0TsTKMt43rGSc_)*5oDD2xst9UJEX-(De
z<6lv<YPHL_Yp3?#UJueu4$gJIm)k6`Xq#jsJGWFf_MW45MB$#}=S8x8t%pC3_fWBx
zmx;EFFD`cE!ki2@b*_CnvwSmECpGwf?$5kpYWYErh}v>{i&A58ZvM^wdkVe{%ef`A
zp@SVd{mS*4W<2eAwn1N~Jsc|LTipF>n=vm_Z=_QnIPPfi;F^PR7rLOS*rZS^RcT$5
zDb{9wD?k=K&mJ1u%vH6)6_{~nA+FbnK40M#!*nddc&bvWxR+A{X(x+Ax1HMcYZ#a_
z5&B-gisFvZAUoNi9Z`P#e<?~6FQXQ|<>Vj<ZvEsy!zPxGlyA{&re6Z=J&pTTKKXB3
zAO2|K>Cp6b*<e@}v-WuLq+=`Hq>pik;pAhr4fd~3)$GFGIxGaeU{RsQu8sG0zQjF5
z9oI3go7_8!FI^`@b(n+gLSWvYqzjR;OLS%uT~YZ{g5!=NM&B;-FTGjvjjgRU7<^%z
zZL%=GReYRCIF9#jO!NI8pGLbTdody2+eN+8i!r49;7AUjy6xMbncDysY|jCo$kz?K
z*}J_Nhb*gU1k;>VV4jD6NJz_E{QBcRSz;!z#2#0LcH#q1PY<vL>%0R;eiP*Vvd(;<
zaFN|exUd(;7bI2KFi0rRVA(YUCN8z3Hwen>)#}qZ(tZ2`S*mEJZ7n`z#AYwFe~_Tu
zyVVMpQ&epPBx$wye!3^gd_xH=oZ6S)Zcz2n((pt4=7#@w_p9^1zSUL{x@VY^ll@JJ
zH_F=Y4_44d$98-lYoHj(u-3M?sh{B(lr%f7fz~b??Q9zSAskC<7B+Mn2b0kjIzk&7
z7LN*Ji^E(tbgPvUPWaB|z~@~$pTBX=X4@^NH;<=fKNB867m*b?Renm!$_58eac)Z*
ztNjGq9BW(ldNAT|C`ghf2-XxxtG4zF3a^x~qiA6b<%gCXPTd_T@a5ca1h?Yf((eju
z<Iv6F_Bhgj9VCdZPtS9#VatyX_m|lIEok0<fEYnhc#%83hoYE&uE=>7(>#iA+?W%P
zV1%W@Ut<a^U27pv{M=`huqK_v*H1Rl!(1mR#3)FPS$;moyiYs0p!eO8&Gm@5HFS^r
z)$Wt#39FTlAY@{z;!ea{%^t~%cyag(SpYt-ZN!&u7g^Z4t9#MSB)Uq+I`V$r&u3oK
zZ_1Zyv$4Ll{&>BGe!WeuxQAY=nO%f!&RYA(P<hh=^;)$WB%&6|P2f1yo^5w9fi>i2
z(JaBQljERa@<FtrA_pU<tf|RT7r_w$>tSkv(_jG0RNnIHSXgA(PW<?OkiWpJUD4W_
z#g+47x6B4gs;~W_sU&<n{}>2irU?>eO2|p>rj;*`zN*bR&Dd~d2Ro_A0UWpe^uJ}C
zzv5Lmhj(C1#Xj@Vf~cQZBJq`uC=g6lh9|rfYTuJPJV-%lThDk-Xf4M#i;PoAOUv2x
z8Rre?y}c1?cf2YRXYs_oo|1csUrVq&N^NRF$&+>4VNr4lq8p6Sa_-TKUKc$2nUnGx
zPIKMxN54ZiwQa2pGp!aow0W-I$W)f@@OM{AiG2_}*fAG;Y<0N1z+lp&sxS|0T=rp!
z`bJ-SQy2Fu^i;i7w(-aU428sNB}JmE_67OecC&1cZTI$E;}Em%Vi-fN#bjt`DBl)m
zUCXQpjlSQv(^XQ#6W+d!vJm|)&j6}N25Wymi2p|Qz8(bCdvOfA>mK?vK=mg6c7+cr
zLp<>6ZcG4N;!g-;=to9e7%oP|W}GrChSkSe*kn^_)d{ZR4gF#td-LI8i5XQ^Ap++Z
zr|o#_k6~@k5)$~>yV0&sU6;iylisqQ-p)`?^~i2&%m2En9xfut_l@0eUQ2lOVY6v>
z>uAR>vz0Ea;DyzZJz2Z4<v`HoaM25^WHENe&!XrWlIE>7$zaxbEkk3UkJN=@?n(EI
z`0WqyCA`aoD45NJz23uRklRB-<m0z$Lk4Q2!Q8M@!4T`qLez%nZ))=bUi@}*AwHru
zdOK<(2+xN$RjY#qMjr*Tf;ZO<!44xnet(|J-6|v&bYY0MCd+J_+@4g^5Hb)GyQ9(e
z%|hc=&@Jk<h~(qA6mh7oLbzJBMBR+s^p#fW_zvYizki<X&5Md<#%^GlYm#-^X1UX=
zH;`Ot%1;e<N{NGSv&UE5O_Alql||LOcBMY2`3Vxk_L<lFmqCAuYTp+%&dp-GaC*rE
z3&J9#XL~(rdhA4N8$S2eP3wIPZ`|1Mtg6EMejHINe&-Ul8t1d1CU}Ug`mL@U23<)z
zf!z%+9;hR_g8vtDe^3H+<)Vo;)QAs$%G}YFSTOimSAIy<C#-s&vDPlBV`Sv6Pk$ba
z>y$mR_&s|xessUxG^FYEqUTv9g;}*k;>d0aI?v5U=rSiQyQw#_s@Q95{gb;GG?k2X
z@rqgeMpq;7=3d1lfwh<Ew+?r6B<b`&p4m9#Ok49#v86e*R`waAN0oi$gLx)S!`?OS
zw;6*8KLMzO<QLr?ZXgU<N8iKae+aU?;F9o@vNeb!W<|;#939g-@$IcYYsdivrgId8
zec4dmyaiG)Q)8pSL2gMM|07iR*dit0o%Y(g;Q$WNFvd!Jyca(y_SqAiCun>H1~JJ7
zXS8yn2(QGUmPA2>#YL6Fq4?8bFC*$e)dUM$S&o_vjk<}@7xTGmHi&n$>mi>TUsn>x
z)*N-&aqRl#gvV_UAm#3?Kh(U%=pc+IzW&^1Cvp$gfrG3m!+j~b8&dx6wLXBqA@1>K
zg&**8M1Z7DmFK<{HudM~hbAQBf@>H3tttfy-BK?&jooT5v9}WoJOs?n<O}AK?-5>p
z2!zaBlkJlJCT?B^${EML6_3`4drXGKCEW-mYU&|g(PTde3IYw3{oS{6N5IaReW-^z
zz;AEs0YK<}8ErAT(@3e}MQTILpR<sIaiDuM8jX<7(2<htq<}KBP#l9e&00BiUR+VJ
z>pjGw?`O0$X6~1J(!qcTGt}nm3zQ;UbiBgtEgApBO%O9rR7+u~$}73*Z|sr`v6XtG
z4U3bmK1q5w8F|fKB!@E=5>g3u3@|>94$7qI=DlS)NXJ2}fL$<5KsoUJAvebvcttzp
z<99a=2Mp3LD@zyn0x4^KQ88-x-O22J?XdzV*~$fDq|W$+V8<dcrV^ki9j4icEGF2E
zBQ<nib8W2Eg+9%;7~pGqnen<i_z{7@@bk){We!p9ZlBFh5(Qmnh)R0SC7+J;H$`&K
z&)4v)%N9*vWKuiS7+SBn1@Nf=hyHUNik3z2r0ekqabCK&6Axp0zUUT+Q}T~Ie4co(
zsVPCRJD!nEuoKJmPszf8+7GLDfb{`Z=7v^{4JS{u+|qO}M`B{4|MA0z4;w(sboSlP
zhgX3sI1Dg&DC`=q9;yjR1PuNtwE`C<(+riId5GIlN1j8UjStcV1|ms}wMHcL=0HIM
zXllcIG&L|ZkCEu9pEptb&xR=&$!W9G`R2v<c^Qx+%1nKGxWZ<V6B9H-LJ|Z!Cu0L;
z=1F9n6Km!5SbyXxfnQR$=h*&m;+^Tf!dyB%RX-XUWyhsv!S$M&?c!UCjg3@p)4hdP
z$|ju9gXMljLUWIYs5Gi(^KPzjan~GX`g{#qXyqmK9z3P$9=V#_U7Bdn_HOk3+PuGa
zZ_Pt5H{9XAjf*?RXd*uh8fM5MU?<+0@wUtenDXdrT;gQuQ(^Pzshr(Sb@lW7{QPel
zT0*jOa>Q{btZV?_jZVD#3N(8^9Q#qRMCsiZ{|M3fFQo?=9YP@Z*3i>`P7O)T;pe}f
zOcI5kr^`of$+%@pg^hEfOrv%6gJc_3!lYNJs^VKtbClLHZ8X7l*Xt`Cd7>vYi={@E
z%R3R-{7_XrN7)m7R@9%`cx>$%Q=ao2UUFM;dXrsec^nhrwoUn<sX!cI)s5hks<^%t
z&GT;f0@aR3gw$g6D=mZ*KP>VJs#LdD>986TjD1{vy(^I)S17ch#XC)Cbtoj;VY=s=
zy{f0$aYtg(3s;k%h(&s#9znNN5u4$ya{I{ceONj1s+N=T0Ne%$o*lqo=dk`3VB1Ta
zso-@hUrSLU>a>Ai=98Kl+%sK6OC`fY>^h8WoOSyRSUE!EID13FID}b~de^gJe$46x
z2Rj$0SVP8d_L4`VRo?M4M@rJZnJL5P!zBdMb1vqekEc*Yv!?s{1=m(lIXK4Xerz&F
zxrmR&bk`<t3a3WY@w&O-Hz#cl*~HN>0u~jRJ#d#9Fs5g~*<;bpbg@a(a)E9l+NS(i
zs+KjJWM{Js16mEO%c<07O2XM~4?m9Hf8SnBG}&X4C^!l=#fR?0=w#de(0|M75<Ti`
z^A(IXz(Nz@&_68$t1b*-kKl-d!YK^H43)%J(x|gDR5&ydJJN$&V75qh*L7EV2USq*
z2X_Cn5NbEOrI+902_X4?kUSN!c!wC{!FRWQaRCs&CyDsIw%TW2*k9-`AT9#^6I$h}
zWiGD@0Fr->@AnHgY2rux-P}=ynugEwZXcZw?=!vcIGS^9Ls#p4|6Z(D$^0^1<z0lZ
zs4}1{UyP5)O&3baPc_yvXK6tybMV3T)^O5(^8?%_TQj$7Qf}3a7K@f`-Ecp(_C!ky
z#K>DT+1uE+ro{l^)ujvsJe^<bGD%BpA6${m3~gE$xiA{WBFK&)^s1#Zxn{A;-tG7J
zI9`Bm91@Z>HcozV>e5yGBd18eq?4&D9YEpEKWlD>0v)4l3CPM54`52I+ec5~?0V5H
zV7*WDKH=O^yI@Jg#pbkr1-8>4^w3KRPRl>{;v|v;wjva{&c!UvK7O<!qcOihVKaZK
zx0K8?T**ftQme|&M~g=GuXOwqTtGtKVU}&$epiEQ<x1QgO@;lp{|UUj!S<}+TV6p*
zeI7<gGD>vUIwiZM*Oor`wz=QiLrZAs#r@%-XA*m9AA%uR3<em2=XUzCiOZN{Vpmlt
zLr7>ax$+%}HRLqLr57`cbfcDvB^z^WsM_vVZ2B_Ml5t}Zz#6=G_Ff+zlpgrU1TXUP
zsv-PV-X6#JTu9~=m6YN|e5CA<NinrBl&w2TFYC5fu1O1vVz*B-(82;&3QvKcvhv}V
znP!s2@5SW^F)W8{SAGB`*csX5?#g;sRsTgnAIpON6*3qTATE()$0ahQ;-R9KCT!jv
zw)Qru$e+91ei^H;v9xjALfweORnZ?^_$WcSx(LDce2gvYqGsdK1E7iHHHM$Xv+r~m
zXbAR`l@m6(R*>>%bPopyM~(2c=>;ywYzr|~={A30ewQ}CzWH3iV3hOG=29{fSJ9>g
zf03+OQcC06qOfR2=rqB??5P-kSbX$9hH}(ygWNCrf+W}dp@o%RN{=8~AL39z^yTW;
zyOJboZ>2XlWOx1}$FxygK`LTSeMM<$H(*)YSzrb&7MD=)n4PD%(HftSa2;ATn1NSp
z7o#_8wRso1N7B^aU!9wqE6wMJDKteJb9ISO7!6kJZ#dQ8JMi8#2oj_T^pX&lQbOsc
zhq(HHPI!u)Kp^-1k?gy&J{SOH=E^}S)*U&|vH^u{MORb;Ndf;w=6nC^*@D~!bMaFo
z@ExQYAp4-j#kOJ~B-?4_2IbIDFtq2H2tw(=vb=WS46YEqe(FeTgz7a(sWANjhvK&n
zV36t3#?AI3$1+JAOD`kgA}E#Ja&m<9*s+u+$zF$l66KqqJ5Ct$vEfZiVg!5e>gM!P
zt%b6ayW^9g=Hzi(e;<?shu+y#36ka0`z)+G%ZC%>d8H7C-WKsc<tGd0vS=fZ^L0Kh
z7-r|_>(N-<AORym@>7ds`K|U8%Z0M?MXyVg0(H<(bYo4G5an=b%UCjbt5S~nxnBMH
zwo6C<11Bfr(bu=v$2rzJM}%m0=&3Z0J&>5+(s2S5US2B=TG<<cvQY#*Wqw&X!lByQ
z+81!~j|2wMlk`&eBdx^W<0WN>Ykl>a&3K}ZK0}FmltIUj#K_mk!De?T%nm_Uarkk0
zzB5;~1efZMIr?u>6#7B)+HcDv2UVcP^!l*1Fmg6xXAuVfQPm*+-=7n5fVZjPcv>|G
zC4eMy@H)b5;tEAC_cPC=^*7;a6<jyCtx4p#zEPCzwk4Fb#NFJn7WrLqytyg1Y>J~g
zl_}18Fun0-xNdNG*s!hnE3$xC_WX7R%R0-sDz&)UR?D)>SF0&hMc$eV@~=I7&N%)g
z%0gt%(mN4pbaMQ2p+t+4BEJWH_o*46#qEogINy0pxy)Fqa?#iQ1(w14Z5)y(^UvGQ
zN%p2vqklDxfIxbRtc@>VlQy%QCyn=ulCCmCv}o+;!OP=os6F7bIc6~duzn+D+g1+*
z@2_8<*7M+bI76`YgDMCu6Th{4n^vJ$H)3hBvt%+9Nca*oj(z*36%-s?%;tt_8q>9S
ztJ+}Mj;#~y+Yo0Jb`f~??AZ`B0|dd`%x}q;ZD9t-P!Sdt5)(tUEyzubL9!R|f5Q5_
z(o)XiE=i^f3S8a6rS^G!1Js@>idJox`c(jU!=TINshu<X{iRZcoXhO9%AFg?sqW68
zX}wX0g0`y+!lf7lK5`nnuCe?jRbGy(BsH$Nd<kDV{UtC~e$~Y{>>Z|*=>2Y|0shj^
za=3z0^d(EGp0M_f=B-yr`yK!ep>@B*vT8f#X3TOy8nvhd54_XIY*RrwixOG+DQlZ=
zoBSP3DUEd}>)w1Ikui2JYwP#noayTGnQAO_v5eD4fNAbb^aV;$*S1UN()-*-a08yH
zRueB?q+}<_qs>@mgOnIrY9Q7R@xmn#m<%?}M5~^=@sbWEJkJ!mAbcAbafIv;`aHV=
z7a#$4;psX=tGc2t?~W4+j)*XdNm5B*l8a(MST|Qi6ab)ppn-hc`uAFu*8?Okpn8^z
z@J8rJGa-rcZwu4!`h(4aZ&Fu-TV>bg6NCiGK}cA0NM+t!6KG<4bt&7_DA&FKw(JMr
za3$;vJ$zO5EVgDyYmx`4-m<NkW~Mpz8@DfYx5>*n9^PlUI-#HJ-pH(ocE>F+`B=6l
z3h8kggmfEQt7aaRw$2+v?A;0da>5KbYe)C%7>R168jhES3yXd-`k{d#G5QC5lc2d7
zg7(*{%I`-r&kd?$vsMvsTh3+a5c5LX9W3{tB*6sriOb}xNTKtS7xv@FtMSj~UtZFW
z;i>7gEN9axv0p5<vY+l@gBo^->+EN8e;eS&A3ZH#`z|3d(FvWAo2xAG;?x;je5x{B
zAD;whOt4D^#>_`y6-2bsf0Qqv*8JMRNDBbSR{@c4Qp?@6Kg;)10Oa8-8mtn};3pm<
zriSmtBTM>LFYJO$*ih-w!32rQbMhhIOzM7Bv717g=+b?jC~;nn$!ITdPu~FzZ)O$j
zn<;*pS+hd}fND_g*9&oO5o9N8m!`2b0fQ1xeI7jGH{mfpM`St0Oz~r>EiT?IT8`Lk
zv-@2r8dwBxi;0Z-Gz7VhYxf}je{JG+!7LuP(NShVoq)f|`LW@HbM(gnh9-#ofUy-5
z7suhYz7#e)SS84F1A+Hpwo@FJNpto1IhOP117VbY4Qx-L!ruREWS#`%L=oOzfMkb9
zqAkOLbL2|!9GOrmU^iyq-1MdHDT2u;{t1{VL<|GFG;nv2sD%j<m#QX{OR9eg8uy~?
zjWjZP1e^MZv{_I+`%rLnykdTIS>ytf{YFGWg6jDR<D8m|#ODU?+a<VBFMX1<h}=FL
zBvBJG8#V-R*(G}}(TG}aiF}{P+=~(GM((z{Afp(z(<&)c6mj+E620*hvrK&_)K_9R
z)}Z|y&l2(0gPnTJkg|jV^b-Wf$M{)yqWRu&B2RMcCL2Z1)@F`|1zGCmmDT9X@^N8d
z6aoMhZT2cI$Jj9DbR6}qTF2CsL~vYsIMPGh#-C(pBx@q^c5q=X_%3~dJHV#d=37Fm
zgLMj(Bh+7Ai@pH0&gK{aYI&kvU)dT3TVM2zY+<+SIR(@O*c&^y>Wb_rI6}qu%}O5X
zsD`kN(|WF^r|wIq0_x(OhYJ(!O;(fX1wAYXqYoy&Fg4qRf&+~R`I8q-LL#1IX{{Y)
zRt1(pPK(-;*}o*u<)ABr4wI+&k-M-551&)^k~-jdjrw>SJ*nkR79og-g4m1=*ps2)
zEaF@=u6>6NI-3|JSVg;v(4c@*=pKiQWd1-`(}h^hIHnT@?na8Z701|j7`tcsCG8aG
z^Zrnv_j&*rEq_c@t~YWiZ`*+aX}NiyLaM!B(|-2>G@m|x_cs3^#20;l+?QCw&-5zr
zIH*$S+UV)7A|TdHFYM|(whl(RVHc<x4`V0{mK=tLhijmziw)QX)-h|)dP>>y_V_5`
z`dW~J!|zb7k8A)*Jb2zngFJ{Q831>`T-xs+?I43c7>rLEviFeK<Rvh$%_N@Mz;Ef}
z?|PFPv749?kinFq#xw=gUeC|ba!h=7_T|$sn-zM-zgkTJ-$Z;nF(F|EEut%rJYN_B
za{qsNJ}@5>$&J=9Zp(0}Itb=;56-V<!34ATv-2?Jmm0gd-n$)slwLy0-afw<+X%Zo
zhG9;SO^i{#UOAE0&jT=6K=Xx%&*fa5;?RgN4a?gv5F*ACA2H|fYNsvB8^`=!e?z8Y
zA%}Z_{bs(4-;Za@1ny_lc8{Bs0EaoUg(2O(iSHEiwcjv`|B;NQ9?muIRAn!nxtduC
zp-W4czBc^Ue>97`A@xRII?`e|5NJ{$KXF2;+<nuatI*0n)^<2dtrjNHCcx|%5r;M)
zY!4x(Pb23Ze*?@i&UryAs+$6%M;wfb@lk@dKj~nFL^$Lit7x8`13`z?-cSfSHrLRq
zbxLBdDeJeY`aAVS9`Ya#w22lTp%MCFN(kdhDuxvi2#Xj(yQ=;-j1_^<@KSO1viD1P
z2l;`DoGFpEJ}c-Q9Y#X9t}gAk=q4rgj)^?f8Vr`2Z^O~w=k4w7p6a4Tm8?47iYDbC
zfU!s#%wK-fvh$})_3KIh*MJs$2()XmYFnd$3q78Gi2j)<d;xf??hsZE!wj_t9Ys(`
zj?{J@O>}q2F;y>y%;8O2zDY|hHV?{s$P^63$_+x7y?ar-H4e}~E;b$dzme3hYxWo^
zNu0W5jC6}g?SrQ{l=Z}l4-W@pRe1kMyEWqWgl0-%MN83GN~!^Lkowmntn<_S)(O2b
z<;d~DtPOtiF<@3a)14k)S)VRXL^_Dv)@-Yc)pVT_5&AhSe7#e|B%OsT0R3f#FmT%<
zf<ry4!t%V(50PtV{|cLT54HcCLjHQx-$?j7LNheCT0>BQG%yD39pfSI;^}8%-m%t|
zzG2=CGvT2_Ao|fk=y_FTO)~Wq+GWr}#Rwdtarfqn$b3vj4i1b}LtdLM>+y3pW=?Bw
zf!4lAY5edwEq-<AU>*k!<WWA`YhM>mFf$W+l;l47%QO8t@PGT3_fTA;5_EBvKpyry
zJnXa7;1zus-@Ib_^$0yuCW4q4+K1kaY#{AJlRQMiqDw#Yt=n`Qq+xQj@YJV&rruX=
zdM7%-L@s4d)Fpi%rDy>5PHtjZZ&R-(U2}ASzod(KkQ*L^I+tiT+<*TW&p)pBp96@0
z`<9BcI9oUv4zyjNnEDN$G`niEvv;?9UmkfVf9=@8<&L7Mr?<B|dlO3AD&`#+1L#TH
zaOCnc!%7u;1IGC;uB;&M>tXukNjel?k#{R*gmF#uI>oj@Ku#-~{04(Zr89NW2;o<;
zz(OZU%-YK8{gG1_D-PaJ6SB6k8LnZ9<LdZYTWc+pt6xPt3PVdln`+W&iNqX$4&aOl
zB$a?keg{eq>O=)au_W?r%I*JU0e_<`e>|9j49`5vD{S6;4RPA}v6shJfW(MYM{JmE
zP9nF;(-2UM<Bs9rO`M?9;>E@6f@`=KnBW>Q){-h1I|qa5Y7y;ma&oeW$Ol{tIMYm;
ziyGHjAVv&^en)2^OIRXfJ@AxCbL8n!e=qQ@zIPovkz-QO+{Ls&EB@^O{`j?u!*Dax
zvkt#NnABO|8;!=(@aU0H5eIxD6t{nHz$V(z9<<AIRm~;ajeqhjO_eafITRK?S3}$Q
z>FP2Y0S6%dT!x`R(z3FI#YC*Sz^iahy$T<uAD{g9kMxgc1>E5uH<}?Faod^-_aHN<
zq2D2S=of55XRw4LjE@wtalsdw{dD}z2LpvFShaVZRm>7#!agH^RWB5maWL#er|UCK
zci)d#Nzz?YxtBNp-w2CN8SL5~5E7^kf;(<197zPsBS=Da2DQ4ux{5cV6MisUWpv!K
zd^6v6vZG~?vl0koq0CA{4T965rKLhm7!<gONIt7K;2&ETxMyw8B9C<P|J8F@l%PvX
zwN=y9&%jWcU=op)ii4Lh(7484*oTveIYJ~nUWAMXO_tfAwvp?(-hu(@bj-}mhgM&J
z-$olwI?q`t0=)P}tKppQmKw?Cw)6GD@$XX{jq_=IwCllDAGxFEUwjcc)aJ?}Lc}^X
z@2FF7JvyMk&i~t5!doM$%8UYy>Ccfz_hMZ!DIAXsbrzl~5<VVjtBqHKTZ8^+`nGwJ
z?dAU0eC6o{<yQR7?&q;3Tlf{LRhQb99>MP_P?{R_UveehKIQDlM!d~8Rwgnfny<R?
zzOK`CqdZn<Lt^1u0B^(g<fM54%BWW_LsXX!Q{~i+W8AAzjTXg-+7}1(iilY1E;&U<
z2Nc=aY+Chh3pl!z;ceOJVOaG<C009pYcA4S&M)VBoPCd)EwsoeKlZ-Fac$D#Gp&=(
z@|{V}DBlHi!1F&Bg`x7JfTQ_%9JcQrKp<lHf5=EVDUrMNL;-Z?0LpEu3-Z1iK&TU?
zL(aV+i`=9x-&~4DloxSz0WBz)MZtoA7F9GZH{|*H4|vpnUL1rwTHwmrQ4l)Us5nma
zVn>p<NTq=Ald-nad4AzK`JPcCmGd(1bz^%@<8m=dr;&WulV2gsb-i0i<<=v9MGxcs
zXY=(xw(eJ8jJ(45E~OzzC;7@~7^_+>Oms*rEL6YD$vF(uy+UD%fDZWOqd_9+tuO^X
zHT}uKUyAI7up#_W+`_vDPipRExwaGvMJmY>tsPejAr+mMW_?2_p@bjFPYQX&qx~mC
zL3auS`Qe=v+&enb|IJQFF93Ezdcip)pnZe*3g^~^-QZ6w7a16|=io_!L^9T}_bdL8
z!u%4F*hRSJk8TZ8z;n?dGl&ie)jh-`G=QcxWWK2sKu?U^+}uloockrf6FWagY~GH!
zqjQGujyIB1*!@+Xt^$bq{$`TG&I{%Mvcbdp+3DAx`8OZ`KmYKLS{TR%hUD!yz<)Bl
z#uH@#H0nWr&b`NR<FE~OlR8pN82G3?cae{%i)TUiP+&ZNYBU|+RfqSazy!F^Jc^9?
z6RSuAq_AGbYA1pE%VWR;iSdE+^1T*>ga9yi6C&Taiur-h&;1_E1I`(1uLo-Zh9Dyj
z2E0PtL+zYY0CF0Ee>##h1eaiSTH9g3%rHXjW$C|gx^R+EN6?DdLLOoU=`MkX$huQ$
z&{yGQQDQvt<2@kb(=!v<4CdYyIg2Mz@X`7hE{xa$Qh>ljp>`GQNf@N#`iu(YJCX*<
z^VBcO^IvuF3Dl3U_n$W**)J&#Xd+CGXux11AVaBzZ5grlN{{e~3}E4wb%$}1w^0P6
zF)Q-^Gk+4o1w=1p-PU$A{{h*nVon#F|Kbju1l`hqEgy%R3PQhzoUqggL$fo)I+C->
zm<gs#%J(>7sgnV!`zd(gsTO3Z=+D5yLUZ4_4v)UuOZ=8;_qyF>hAdirtce#<5YJSE
zt9E=xE`R@ckKmbvc56rycooXM5H7IIbiKYt5Gmx8-JGJrx<DQNi4C~sJp(|4`n@im
z1YJy1^W568s>*OLCvbP4>2G$xOC2#RVM3cop&6;^%MVo8KwS!1?~Zhu=04XfurOD4
zY^bj%GIwBHZ9&ApAMg!Jxl=%Lhphq?HVbK^2(tyWNF5m=yFzxqYml9P*k>v@D(I*y
z{T1D!y$)i!%#m!V!&DxlVb;#QzKm!^LrGqzXx>_K;o|Groq?ULar>jX+zY6ztgn5i
zJi{>8KUzddVj5a)`o<={*yc^<#rgjnm!$z>(q3#@S&h84-tK{>QR;tue5f%eX`O%u
zG9fmY^~R4uks^1{ll-+TOuRT4eh<dsu)|lXK3~2hwJW!e(9DAEJAX)s02}^O?X!L#
z|MD|a1Qg!sFte;XSLd5^t@|1zwPZ)O7lGVg&di30@$W}{@Fh~t?|H_FoD(Y|akzHK
zpq+|&nD!8}WC+W;Uesa$ch72c_e*_`9J)0S;NZ~TY#U9>SH!bmwgP$&d68bbUO|_7
z<*-j;^*NlY;wgCKTj`YPPo8O#y+12>I6GA3`U5>NwzBi*Po}<6KA`4$C6Mj<^($=0
z!XxJNHnV?m0T|C+IK$u-xT=TwapcD3I({4I`Y~#7$f!$2f91!_6~#q<&0VecgQ_<?
zYJNjhBqCP`a3Af{LD8o|np@p9u*uGZhf-?9g!B?qoh5Q#cF{|+=Cu_7c0;+*neNz0
zwd1P)896AF27Es{rCt{f0tE+AA~cgmJP*yi2njwoHg*ygNcdB;j>R=MH^VrJ@JTRA
zhRsn-Fextb86o`84y17BRy#r@gHyP3;OAW0_7hUaR?86|!I%?{_KlkLLV^cS_03!&
zC%a?)C@=4k0seMgo|nKfec32P#;t&%LxR(uoC3gzPG2^70>#^p(-)RTn<SBG1uUh}
zp`j;IZN{pdU`A9X^!@I=_`gW*IY@3%?Tvc@jbeG{aPZB<XCu+3hWX*S^U*6hdNC2K
z1ia}@LO#V8c}>w1Yd>*ymLI(5#*sgKRM0s6o@zivPGr=Vsy8J$R@WqUj^ghh7h#@G
z`%78}1VDVh!}4l(SUOBdIGh-v8w`a685kAa8~L@mIuN*B+Td>(VS0T8GIZ$f-8V6I
zts23>!N*ePO~-d$O6X576#_36uje<FqZ_(F9I+y&9GG)Wq$tO7fyoG8_CU;uvxxLB
zz|1I4Za+V92P1v=F3_pJNgnyv+ejK^Lv<PknUH2_f_^;5#efooz;b3-RVca9&<(>Z
z=>SvCS-1e3!nI%@H8P_2y=~8Zd5J#$sNx_$xjz2Oyi+ixY6UT|v~~u2@!e0dS9LkM
zmp0puBMOoW3bONvLIgVs!h<Nts@0Aa|NCQl@`H`3w=y^ZHlxA>l!cvG2s=D1jWBv(
z-KuI_iOl>_M}Mb7rG)3%*;`xQmw=2Fp)n!gv#B~BH1$hq1Ac0S;Jg;HbzgRG*FjdX
z2yE7MA9k=5PcTP$1(~l;B;-YDxI#(ML;}%X`hr3E|7_j&+klIY9<R9%g7#E_2i=)=
z(U{{$jJfrY`{D&JBV5NWsQ37TNt#$Wmc3)eVG`@fS9La@Y@;j~X40P4d?)AM{qs)e
z;g=<!URAes@bB^${A={S`)pqUsW!`CibtYWRj?oD5R#z`h6ma!{_^LK9QrogPO4WO
z4Tc%h%1;05?O{~_yghO>^frUYP|Ux^7f57<kj4C2;}N5b;8HgbxI7eP2+`yL)7$De
z82p(+==P+7Sg?JT(ExW>9Zwwt%{WsQ$SVmI)K@MF#$Dr@$zkm>DTytiEHyP}4;|`R
zjA3C8Ixb04_j<CoW{Pb4q>8?9fY^s0F(kbqayf5Td~UV9Tg4dnTuo_xcjnMjl~Pew
zL8q%S#I(U1>Altd%j9(&(Y*LEnrH&^t%&`K+l_S$c~|JAQ&4l`Yr-_lYy#}w>r1sL
zOn99?<92A{U{xi%Xjjy=`lQll)+E&*<i-uET((0W#!%nEvX)n*ky@pHZMAlv7Ew$U
zH<eEBy_uq{u$uRFm?WS|tF7_PRsMXVf`Ye#+ne(0(PBmyyGE>e(?e9~3A{_buFu~`
zRo*7#E6|ZNUaKloELjv+oSd+6l}9Nhm10GBO&@JwV|f|3OW1nO927ETF;wPwQ(~j+
zbed{USN?JeVo7j&V{dCCEl5AE@7`V*&>jqoin<J)lDqCyFbvo=5(Wbx5zdI|CDZ+I
zm4dk}VPfxy#$~YMd=Osb%$t09wySFwwLIA|_nvPF=yKs;PPb_@k@pCRnQ)m+5c2tI
zqpRp8ZS1QREX&WZ<faWDWf5}KP0DB-n<5yBxfgi{D{A-Vy1hr~jr>x+PL7wIY18S}
zcgpt%yUI^~tDAKk`q9%pH`$eB&^s1(CX*;{gmB9-B)4!=H(9qt{>Nk-3wyYmUV5k{
zR%DXsim%HlBB4}go|4&<(slO-2>pDcb2hL{g#NV|3L>G?TeGg;`fPG>Fsd<?&xrH1
z@@9P^nmTnZ98z&7`a$3lb3y1_iRy*Lj(4}w&J7Hz>0>3@-W{SvIhm>^8Ju)t4ttmK
zKN&|-GO#wc?ik*Ln?N@Pms^L*(SL2!dKK8^lMee1f)6^E2KieWfPQAN+?>0q<88J5
z#X^!vm+#@DT3qsp<Igz@j`<L};CIMn%~-uMedO&MFtKt=^GIP^=Pk|gN+o+4dyniJ
zJPj{@q7yft%v!r=nr?nf>i?qnVKA+|X{e&O<q8K;{>Z?FUV=t%=SB&mZ}d$0nb`>=
z8p4j(#ja1Ja4{0&qn!_{=*wMmq<KWS>2DjE<iHo((y>%1jkhDbB!OSu9i&=vb1GG)
zx#ZfJ`Litis8)@mc^9BrnKL$-ys^7oSswpChkm3NuYkH7dU6Ee6nuiAGBC606qTUE
z4e7picRM5+iOPZ;E(@ln<=C};-3>`RWAT2@nYr72JuH;EY#Pf{A6!h$eyCH0znNWm
z#j!Vq!(i}hIl<(?v?-Dk11fh0JJQ*MQ@9&#uf>81FwaCY7@sDW8f49E?XFRma~u*C
ztuB@(se3QXzaLAIJF+pIFI9h!Tqew<CTSY9Z0J1>^BSX?5bCmJ8+Irv52u($G)sX;
zboJfTE4#v}5BFMfT$NDSm)qytt=liUX){JIRni7&rgu1QyWLe(kv9J(bu1*<wJnT>
zBlP3uHrAo{HT;Fv<4r7nhONbUsLDA!$I6Dy>EM)PZreCv?zaBPUfKG4kuauZ{B*h3
zRKpjd)4!ec@V{?!g)6AmNc82ih!6JOiF#CHIMo`ifFKPa_4yBibckLpZJzv5q<Fr~
zW-RnwMxBeA``IGOwwl+IT8xaFLYCw|*ScL~CR@!nD$C9)z8|!Hr9*I-B=x5G)w^$|
z=Z;Gf*OhVW)lLP7hIXYb_Bg3vM_!m_gfu%7v83nmiDD?#(}OomSJcxtX97A)ms4|+
ziurg8Ck|WEt(D5Mr{`L<oObLq<)|aEETzc0d8ch+R;s@YJ=ZjVm#trOxBOD5+K9PB
z9e@30*7WsnOVb7|PrR<+4g}dcT5#Tf$%Wp!SKjwtGq6K{tCcO=qdp?Iz%Zu(H5L^9
zU3o0mm}#-tHe$h^T((goIqZo1396upB-(lWxJSZgHO3?LHQb0>y0AKpasn?XI9kRN
z)-CG>g9fgxW@DQSvpptRrcW!&FLI!7Q0Q6Z#khU2yz0&`R(P{`H4}e|Drl-<DowED
zKVvr&Fmk1RitMf%*a)%Pa8?IUP8Qc8fVK&}O>{)IuYDg)!C6nwLGb>9WslAF{A3eE
z#AXzU4s@J27&d|<p(C64%7^!nV~1ggcta5QF2Z6s7iX#HbsuSy$}jG1F&Fq)v_^y&
z`Rl?Dl=9xZn7?tA?4?NIopghG)Cdu?n(MpeaGke8XGSLl@pRN``dwGIJVt6U)8*7N
zQG{E)a*b;!0v5ZFhVEQ>RX?f(m->NOsi_g7Y^yJyJWW+`Q|t;T1ym_E%Odm$#<Cj~
z+h?cFw$BZOL}7`@-llh__-SucbAdbiW?A|XU1C@l!(@n9XOh>47xQUFRL3^*WA$FC
z;Wfg<0nPwCbgxt&|4Hn{A2w35zCXWeO^1(o6tUe`yAPUN)`~~rH`VZ+_uk(+qW0nm
zuWAI|82O@u2saHx2t4-o(?a$WFNdJ-8>nUs0E>jfDjbX=pujKFt0tZv5{hOsafQ;~
z?)8d6a?qkY{o!2`{*p~B!pKJ44_3H(>#KZI$MYvi)1tQjVDe(AvZ|(CE_zibdf7bR
zb#VrFzaTjQjOi7!e$XVa>O4^WiaMZ$!7sB=rbA>UP+3l}q;XB2hBYvqLzGF>gOqPV
z+x=RV?>Xr~QSHGt6;1Dl<rJLP{U>XZxh&oV=8ru@yVt7UI@5NOWU;JpK0!mVppw!3
z&fS*N&Xu{64B*qFMnKZ$ZkXG~t;?$AQw{w%KgpO>USrvn?#cQ|VAnU#LjX`fX_qCL
z086rBE8lvIv8GSE-3(T2m@}7_^_`o0{s<4-f;lAQ>o$7Pb$@p((ecZFkC%7;RV)+{
z@uBcL;Mcv*1O8=i)wc@e^~=r+_uKPKBc@+2LQ;iKLPOP}UjS3gMmq~KhlXOdDqEm@
zQinjk(7$^twzKvF(4Ugi&Dmj8Y)l~(Ob3+U4!E|4>FnX8$w-zOmg*Y!y~3vAfm#2h
z_O-d3D<$*<nr<PcnGuk$4zyh=%Kcer<~Dk|{k?%zqSixiUlpT~dW9S0WG_gTfATxL
zj%h0O-ZrEOm{@f!N{d^QD(qw&922b(CwZ|zSaK4(d6v@XM&KjL^48Lk<9Lz`>!Vc$
z7^Aw%hGh1evB`q|kr$Y;t&2LD#X{S;j$Uo(foqme_`XH{D7OzwJ1uy<X|0ySgxB^3
z*Qm`SJvt%nCu-B0b=YSu{1XD3#>s7oC2@7%sYPqrJZPyeF3=pji8sx$GU(^7Q_^DB
zpJq$;=sD}5Z{3mCc6f9U=l&&XxNQWM!ajr86mg{wccsqH!J&YRsHuyH+tU9ATb}Cx
zdy8L-(Lg)r*-lSw`@Vo+Sti42W?<m1w6ru#W`^CIiSOP?%`d($3xqX<tQ8vog6&Q}
zF|ybV$|smW;pa{6a?Mbk=c1D4Q=L42)6o`#Wi{9gy%fenbzvsWXo_7aE^sYkX@*kE
z?s9|s$ffqpQOP?&xOf4w0p!k=btNKO4$7_CBKMQ5ni6hddfatT^Z5o(*DiZLBSD`T
zXmt>!V>zO^*<~v&k`(@=ie5|4N!Pw~SyerRljFYo{61s;+pm<R<(D?cADq!eo3hdJ
zoy6X>3xhfDxA@0P*UC1%52FYw3Ay}>d1Asc?mn<$++QqcbN(;k?5}KN0y?AH8(Qv|
zBlh9Jj`4*=Wt8mU-;=5v<d9bICrIF67K?SyQA_wm4B42hpR`FyxLy_~s^+fErMHZu
z)x$5vUh3qz6BQQRZj)af*y>DJSg|Uf#lh1TPLi)t*mhPbi=9t+TO{~`iUWdRtb=*h
zDh`o4Yfhf_Wh<W;l}xYcDra0wBCz5Y6OHVApGNC0lN-BD-<D2sUwhsQ8HA&tQfl9P
zWPsGQPf?`Tv`^HO26Y`zwK<iAUrrV37a$)9UblEf`9bk5jP91vB_m%~Ya>0K!3Xbo
zKt`rE11$eV?)3xGJ^xr?P6T(5fZGnRrO?a<GF%?U%S<>}jLw)18JqQ>Z5??q!9j@F
zOAvrbOI#hHrNa-s-;o#h#bb^ZR_nsS?c40qX`1x~F9%BXpabHN^@G8-V)7qZRbh=5
zE9%MNs!a~lPao1}uRa~kb{J-IAH+NSCHy3Hge58(@n83&1aJ27gEdQ|xOB5=bL<VP
zp3Vj1({gTY<3vKVz1tqy7LAqG_t#bHZBhtTZY$Y2$xDk)FLYg5qLqH4-gft)Hr1Y#
zez>pDHx6}-Zx)+jE$hR>9!)palS}iOb%#87q~JN#K=og^nEfSi`|@49)px*e^__h{
zvT$UHiI-utEW#)2eF{9SfQa+3M)MtRD&m>mAoM|sloz{17teVlY%TQxb@IZA!^1Cb
z#2<S}m8Ng|>b6<d1KW#<o?ci;6I-gUMBM?$L$X~cCqnx}Z{F2NSKVqVJA`T2SA2eW
zT}^L+k|UVZih)pZx_w-?OQjL=6v$YDDEv4Xn1?Tvzy25-Ak|8dYV*T<WL|N3fZct)
zbAI)(hN$5Si!!l0+YeNWnfmcF_jSkeAN-5y`DK{o0fd=&%<^6WH0V|KPGA{+03}Ql
zFoea{!TfpyXo6C_d9%-f!m?zN#|ie~FtV|+S)h^td=CR8UbO=Qh+u*&`L^gtZrxkJ
z*%FVXA)^>B>OoMy`u4>x4Ma@nAV3I?WPS%wf_@&#6pIsvFj`IG+wlRmwc%H0uC@<{
z?xnZoo&z+{#tuIX)205ZW7u^CWfsN#*>XGR@smgU!+oQW1im4q4KmF83@r`Gj5T>C
zvWnJr<rY<5Uk;KF_!*?X7w{EE#WLk-6)bk%qw`?%y@YmS8s)o}oe-nd=HSG~$~I^%
zXNVXUmjA-A#L$6ZIizye4>2r;U|6h9II4DsAl-(*YjEh&pap`F!(>Z}icMF?Bes_n
zIh!&?32H{ntZaF`AJvhWGq=4!ri6i!dzBC0!ypFrez4)V-KJ*T`=^#14Y1|CVQEZn
zS2r)-yx0mGwxw{h4_BsC*gTpN9)2j=!+<r&4>jR#Af_}Sw!{DE?MM#;633|U$FepD
zwK8iQ=5@bqa@~w`VE5>^lNKz0I^DCY0`QLb|HV5}0pp<kQsUj$@Q$8H=Pw`RUA0gH
zRI)%Xsv=$|-hU0sf(ahmTmG6hg`bZ(5;GZWLWH4Z+3*5h5SgaDl#Iu^H$=h$$JU8Q
z^=KynO8mNM+g*qQ(sLghW5%0RG`lxFW121~c~rL#28C1+CdydbN#Br}qi%dK-@!9O
z)klSYSZF)zewpj`*mKwi%IZ(v!16{~+x>WVtd9G>A0mHqr91lEM+eVB%yLo|sk=2G
z@S>5IkxD0Qwyr@h_prjYMSs2tVV5Oj|E&%*0sVZkXrjXan;sQgbznw@fayJ8POKwV
zPfu>=^x(JJ1NdY!)IvAo8cVA?_0}%iWQeZ*71_0v#418YB(i9$+RosTcy%#TM^K2j
z+{x2mspMM9+t~E>0cGi@C$_C`&XsqbEfL+63N}U2oBIkeNE;=tA?j-9+{fQYj5++I
zB4;H2<ackoj@*bklsn@am!jS3&HqjTe|fuqoM|dV!dh$%uE@b5nCIX{VBY_{by~<V
zt8}TubR;JDynk>|(3yEGCW1!T`MOp3k0SuS@=U$H2jeY%n=u{-8%E{S0_s^(@2Wed
z+x6LUPB<<za8c>Y^)fm?2mHfd1m0K9e;_h7T>M#X+?BU&s8%``9@Cv0Ij!ehTRVhZ
z?hOZ|VZP+1<&%|%+YTYQW!U}6n)Gaa#X{gKpT_$RuWY#~W-&?nZxrgx<(q|H>nxy%
zQtM80NFPH%Kud#Z2<b@gT$|)?+~?vaaoTz`UO^UepR(DH-j(zMT<m2}ryAJVMYJUU
zO&<U11E~=Hn3uSI4I&^YJNQ>t<OsA=@|XoiTNF!MTU(pCd=s=AW8l2^m9(Xw|IC>W
z`QtBa??8)U*z+DYq`+NQ4f^eY=y&E()?GR%tn0OP@=wD7GfyqyAMwJJ#a61JgDzF#
z$P2Se0#nZ-Dr_ts>w4j28!{<WFL}=!K-#-5J$C45>T-9t5hhGCGJPYpe^v@c5zNgK
zSjsj?PRowY4&F&^8CY^*$l2BjR?mx??tl<U7(yg2RqVx^-<*5ZSPdr&`|p^u=mz$b
zh1gPYYA80Or?$&1Gq9HdhVWm?1J-%9zxLK|KS<Z-j2JTtM-jWR&{EjdH3zFY4+sx2
zDWF0pCd_Ium}Q-i$VldtaceFbA_}g+lWdyR(6Yk{jm4qRc_N)>lC*uBlfveU^$ldm
z9*Gc)2VHM@Om@c8(+)H-J?GxJ-P4F+0o=Tr#fV|1Yb;|uaaGSE;!}!>TL}47H=CZ{
z>9bW!*Oz4{27pPLtX4^*CR5U{pQx!R7q|J(k%R9Qg(rop^|%Sco?eQK{HqJ6XacCZ
za^=fVe2PlniZ;H4I}isGTGI46h`a%0zlGmyz3-a!2U^SE$Dgjpwe-^#)L?Pf$K5aA
z)81q%`so!E9BLiDYbLE`iEJJ0t!CQ={<Cqpi5M3)syaXe+8|o+Z>q=%4~dP4VA@<?
zMmo>e+2f!#jL>)MCbf^lFpr!~f8dM!UM?3PAm41^gjmY;vr41=&r``!nw7j(#^0^l
z*BUtyRHXa)twEiIzWbJveM;h0iG0)KfEq^(xh{B7sTXuumO361iU81Q0!tI2LyNcF
zSS}V=j<$TsRq{TLK{<I~-70G)7gb>c*=WrwdVNJs!}`1ej|fk)BGE@OzHjO&vJ2$P
zy(hW&(v8xG=fR=ypn>$X*7G4)S@^T$Eo(kK9$^vjrOegKQ~|vhT<@eqLgQzTIA7|H
z-K(wqr~7A$m?zk3g4{lhy-Ekqc)Ri)0F(b#Y`Aa~!gBX}^j=sBzsm{ZxlD-?us!qX
za;6TW=@;Vf=$G<?Z>8f=Xz^pmT$k|)1HZ`jX$aA@&#zi*h%O$j*BwL8PNQ7K8;nt-
z4j2?ht$K}FpT#ep={nU<p`pF_uvEnEP8RpcP3xOHA_7;eb6N?g6fl}~1W#J{mFIQW
zel`VnWWBfKKybgJ1?u*V=J+V|Lye=3tv621T`d-j9;`kOul3j54*s<jSr^tC-Hur7
zoSldpAkWsE<d1xRuXm`Q1LbgQ1`}iQ7Vu!HFDf1jgv@b(Qft9(F?aWE^ZMt;-_S|9
zlp?CZnB++Zg`V&Wx46?+o<vvVJUM|Kb}qqDo48ATBoi)7V`=ZcWJR`i2kCbW9nM!j
zc{N<KXvNpyzWZuKGlh_2!ohb3R1bD__8aS!q@SUxvpID>!BRrF1qJ)51g!USy!~%0
zn@{{8eE3nw$%@21F!%^2%;8`p;>52^!9)s1b==Bhjy#@K>0o{w;VSuaoT1zWcq%Rf
z0zMwk1NLY-*rIXv_0wSORR7*j;za}urb}FMXL@p9ylphzuuX2yE$Tf5dDL7tE2hSy
zM?=|IX%<4VXu<43Nr-^cc!wy)UYB+TI9^w5#LSmmnC>)N9qT7R@-;ct4<B=N-in-A
z<e2Q|{W6QYUK$%eU*7p<^Z}k$d@+A;g-t(H99mwUu{KkX|F)`cNBjtHk;Hsk9-#J9
zR>Cb3u+3D!;tW(<{>wt|zlsB?l}eD1Ec-&?%dGr3kVoCQ?}&q_Vj)BvPmA26Y*^l#
zIr)nV;7AI0VTpwTiW`y25!iW`<)Fu(Slo}t@mZ_qm(Qkb<T_sYhSy@zBEjx_JG8mh
zY`Be4vpX<4Pq)>dHiQ2YCHoC6tE0P)Dr0`Vtth&+CMD<*7q0H=*xg%RlXH@zvsY`}
z*Dkd$&X%+K9X+2gSyDZClq`62q^Rl<{__@1t!}3{<;;*N-ldv{@!5tAO2Ij1(iL{&
zqcdG=s1J=$v<ZG0YU^xlm}IcbR9Cvc=K)FHmvG#v^b^(W<s906fU1V&QVU3=ylv$(
zr#QRfKH_}Wiue)A19$O4?{P2egCw>E%;=Qa*MO4$qI`K5smkVWeMJgb*TJ1Z^)SJx
zs(D>TR+jP7rSIzsr9w!RCHN0TaTxVdEM%Jb=Cqg5qHg*g{uXw%4p2~JG!P)m*0(n)
zET3rT^!7QFFK>nPOx=|hV9jfMSZVjfzknDsypQFg+k;BY_Dxu$Q_dl}i0`XH-tQB!
zeMOm1==C#?cU|#kUZgS2q>cRCgEic_XJZ5y*IQkgdp8HYJ$0V9P<YtrizauEorUa~
z`{(`({4pHSG;ssMBwjQC+e%km3Sk`XDs~rDWq0be2oI+BMQ?BY#B2xAlGiVnoCC~+
zjZWRydXW>oC}aavm%MU+yy9S}BkWQlc5l=lZS&rw63hgcFR_H(5|!$iLBNudZw=a3
zMM#k-CM-J%(bUw|1{c|2YZpeF5M(ci*sx?BB%8tTC#`ivtaiRk)KN!OPN!>eV4~~a
z*0cJe^4LB*wAAS1WiJk6hp;3)#Lh;g2T*k~vW#cWyjBd<I38I=Pr&QWrnl{;T*3Zh
zSpVlCp;Q;m-pz_k9Te_?2$h-DPv0vQUC;Llx`vswlu{>!q$I7hbq@x`^9wRuN=|9M
zZ(C+5m?&N~9F4t)Hz|?MKCz1sic3A*NN>-!p3w=*c)vy!&e2yYTfBo2cBa0i*jhe2
zoU-Ncg;UTJ_y+VpJ=p&1628kpFzb@k>Ud!0ULifOaIb;Xe%YX!v<P-8iF@0$(E4hw
zo@ZeR=C>KCZcbK8g!WKmh%6K!PtsLrv+Rot4u0cA2HS<mRV0%@Wd#!z(}1%0zYzs7
zSHf?@8d`4kkdU=2|4{Z8Y3kIwaOmk)?3287iR*UuJ!`c#+3SNO*Lsd|cwyh@5!b?|
z%}B1ICFkVen6xnIT6t6`TF;#&Ixgnq<%u->W=coZX>=zMHLIU{FY6lt*dKLvUDfSO
zZ-bkrz>6q4JC3P=&WOWHTML=I!=3yJ_}!*;f&?|*yd($PnTsDxTdbOVCFRd&N^0a&
zMOAV#i6S*>B-FB(P||trVfaQj)1;uRIx~mc%cx$#R|e{Y{MLd~yMgl`RTBA!EMG8z
zF3nylh*NcjzyY{MUF9Agt7EB|1~B2gr!NPUmD*EjKe7l@dP8o-OVl(b%A&VnKS|F4
zl%HpmEr*v;2;FI9!d%vC_+6AIy%~2A=e~V#=Y2VXBfYS5?d;(0xHg;O7#YzSBYoK9
zfxGR?Ei#r;s@|wd;Q`pJG5h_Hwd$r%uv1-Zux9z4`<wxi{*&Eoq8Fk#_-V_?{cF~3
z(C~|{bZ(eGHqc&_gsAOWlMqXx%Zl&x@cMQnJwa>cY|nfr6l`)R*H`Vb8ajg+$BIr*
zyzW_dIK_R7!gl&}?!28<4+qp-4p{n*-zH&gxbwNFr9*3T(GqI8B2is3#MnB8=j64o
z&9xE-U)CHzamVI<Bmdsre&D*#f4JU=ET0sC7k(MxiHFGY=nlu@0*s4Xra>V+lF(1W
z2<y)LeqJ;#N{0CtJ&uD!NTU+KJ?j={X41qt*4iMpExW{t{5Zl-0v_wg4v$r=q8)M{
zV!d+M6;Wqoj>N66s=^dmbHbEmQ&&w11|J`o4p8CCUvPYVVyY8JQ)yw;vL{Us3q%A^
zTY6yu!hF*+pYqBPY9sW?i<GMZBCM$hWdn)N8xp6TOZJaDU8itVwMj^bW0Ko$R)co9
zdde>uis<6Z@tmZI(08wj55L<EMoU|Vqd9!@Ef=*k_i_gGB8LZ38*9fvC0}&&{1DIO
zjG~6|Tk*H_fx|ICet-iQ9KAxkvew~jhh&e|+nh499$t#3v<P?pKeW9EG}rwbKO8A5
zQAo0*WTnh(r4ll;X9HhFX7<RI?1XG3WzUT4l|3S(Z+2Gp{MzcdKIs14|NnEZ^PKbC
z$LVx)`ucp{?`vPL>vicA%SaD6yzgCDdZt;z)dxKQz*pQD{1rD?4(=&Ctjh+PDER^w
z<o}u=xVn~s<o1<_rXPoEzyp{Uu+wn@W=_h`@ZIfLK^4=Yk)u+#Z}VUx^Q<Q2D1oG@
zbaNs)tPZkqR_j_1NL^k9QXyWj4=F}E`>NCL|IPvg8TD9ZwAd^1M5hhS8LFG&K2YDY
zzH7nqA08?Ld8lGx_PvMN^O<>i!yk9vCLeEkQ0E^4Q#e#ETQ9@H8exAuBeD+xHfrHg
z1FG$w;hm!psSq%FAfix!P%1R%W?$G(?5G(aMwNWoNbbk9IQyI{ukkd&=K%rljobeP
zM3xE~dPjir7^1DW_fP?@ZYXQ)66=}s@*<d*z+<X-^5n_0f-gUQ{J6Yc@k>V+_M-)_
z>p9NlFK^9c$z*8epN*;y<hP#trqYwCqxJCN!RFDg7LDMb!VT^zB3TCu_&3>m`Xu{L
ztPKz@RB!m8D)6J3q*-Rax<Z;zViiOFat?XDe-7s_B5<tb!<B0%#4B_5y{0I%ACWD~
zy7uE?3ApOkc5#sW^*gT*yoiZO-%Lo$Y>Ri<O7TC#D-G0R5`(rdq1K3OpoO(yQT4Xf
zu&QKEE$LjDpp3@Dl-*Nx&ygg75RwEf*A}w<hpEkhUfF0-(eZDv-VvAFFG)}is@Z{T
zl!X7@k0!x`;&kLF(u2AJ9XP~AO*e$Ud}(*M9j5n!cXoCX-WTarxGaBt5F`bcUK96(
zx2+|TyG$pRkbuBlcc|d8mOX5f9e>W28C+Xin>{KxhYfX*w9|i+(0*3|B+!zk#^kJD
zWuDcePtvjH05^l`_Urx+09;QvUh2VTdz@oJkTlr<Vb6|&>0qfwT@W3oOvFVZ2RT?W
zYJhAn#B{6S9=qTuObY|E0mo<rm3y-E=$^!U_`riWUxV6bCtk$GnbuEwu^7YPGC^XB
z2Z<eDSZEc10q?=h(WqVuhiy`bcn~24`<gdJCQ^XZQ8h0xjs&IRNHhdQ%D<XfKjfxx
zT6vJqX@rl(7v2t3hJ{MVu*+a(^J_uwAZh`nI)6$=sFC+Lo59u^efYo5Je0iDty`Bi
zuk;ziHqEtEA5EgaJ5UYeK(%$Y7XqM=U|dTO%M1rP+3ioh$Aa?DGW&lY==sH2+l%=&
z59fO_qJ=SCcRB&ct~l^o<&XpH)V|tYG(9a0>+C2+or-`3_uJCPyIWC!ER=g>Cyw3j
zjZX<~g&OIxRegFewJ)1rZ^6^?v3e6mfWCR%Ks75;{NS=L=vr{G@0x^j{`V8aNk<Gy
z5}y^~rYH8?6!jv{I5hM@3=MkHHLl7>PcQbv9IZyV(?@VmxMi90@uFO?ZRH^><LkGz
z7A1wH2lkxGh5^@y(d<%nP=IDegy(*@I`_UluaSY_6q)(%t#ZQ7Mj{NI7>51Cwm>QR
z<NvqETjdG+Kh5Nfs5L7|#(ddA`8LzJb2rv^<tpHjc3ELi<}Xt_JLsTUm8yL$--g+<
zp>Nq-C@`AnNb3jP`=|PkmpTG=+=u@=0i11R;GW3&h(rMJVeM!HVi>b~!}ZK}F};o@
z68w+9DX)ku*)V`;sQ_`q?xSijFO`CjboXd*A56-aBB>=+s!0VhPy^-e5UV3U=4boA
z<yijv<+&}w@2{b<KqBC~lJK!ejMCiMsiaYAmjmOcEc%6$FtvlQQ;!w&nTC#y>B~N-
zpSgk9JPpK(;Il!YeFo7XRO{{icQ3jM5@njq<p_H~>`sAY7FLTW+YjOX(L9sbO&mkc
ztqsHp;AvkPRdnJ|Z80oJ)!W_OJ$PN;(69+&aAag0EZKG0(fZN72Z5bDfBhggkOxWO
z*+w1&aS^2|FL(d(X7*n|86Kb|60saHFjNhhX+x-%lG52<TfG(fz>CBnB!qir42U1r
zkH*7#pxG9bPn8O(9UtCluRiEjV}&t*Z@G+Es5`8u5XKj<EDr!5H2X>^2LTjT?zJ-U
z{=?<iA=Y$(-0&&D0J8Q31y{I_Hb^ozHy4SebDvCgK5MU-UK55j(#W37b3l>#-6F;g
z=Ipms5M%R~;(}ms4~79~aemBG6RYSG{}-u+OZFlCK8d`J2uLugUFH)m!hN*+vJDF6
zi-;X8RbJV{<ZrJ1ilO})8kLlgK>s5@wQmI>Nq|YhHK+5}@2Nn^%=MA^Er1APeWU?<
zgiNw`j6Yb~e}=EBaA^e`M1%;E0vog{(&IniFNP70f!{K>CqHQ0Ock|BUgMqy0KXUj
zvpPfjwvO$<Ib@j=z{Guv7ZB<zA3B#SW6ogN5!QxAj>7)$&rLxG=GI>zz{bBQ>R=Tj
z2<pe=C>Y5aR)+SYP`+h4GQsdKjz%0|`7OJ@H~9Y*nYVO9El9PrwuYheP|bv#r6jp`
z^1yaOiy8ssfDPJ=z0X^R`0>{h3PH&H@@SCQMGnmm+4Xt!(rF+U|5S!6fbhQH@k8~~
zAbi7!w>f^WlIoxC*;NZ}xFo4_7V_8cz!aO;e7GIx7nWg#ZWAm<U=k7%8rz(Qg*Lb2
zMRlTh9-dqk<Q>d4Xx0Z$oeT4#3B4(%?MT>(Y+M4W!_hhW@d08G#jvoy6Gg<TQ>C@V
zfx<|ELAO!LS4T>Y|7a#a{1oA`Zyx!$M_vz=N|evhJFQnP!t}z*(b&nBNK(ry<Z`My
z(Ppp^O-fF#hX?a>cY7RJO#s{1fy>|85gE7sU^b)+u_;f$+)F>C)CAN-e>qfQo(RF)
zfV%aXKWuzHnNX4!n)*reKUBE<!GrO!HtfRdh?^iM3xrDlYcb%8II{)|12N&l-ptr}
zQcc(d@M~vM%m79Mfh>h|^(TZuW<NH6ZZGswC-;=SI^7TTvZFuly#wsRUpD--AO@Q4
zY05K$NW28XpE+HI9-n`UhQa6~RFio(D2b5>9y%R>>xBrO{2Ia@bnFbzEx1Rb*gEJI
zHNaHJIPl&O3Cba)E<kI65bjG@p{EQ1`IK6hWel9o|35z&EYJb|jmA6jBIl6)m8U1p
z9XWtb`Ias>u?><~ySTU{T?V{hwTWHu0j5e=?mu9v`WYdgSnrC81zaaLSj1U>!5ry-
z)#yWiF#)oT-JY7Z0>nGxBR+JeFhhw;UrX)xcwoXW>P+(s!&Iix3>}a!xp2?#`Ioa`
z`&h#6$L`Dg6$FIemPwZ;2{Sc_!`kV7N%J2(=U)VgM8dYyT~daCZ-6d2Pez(m_F+BX
zQ06S(cyl{WSOdtXOCw<Nt-ks2|H?HTNS~cZrS722<0cMPWI`L!#x)X<dy(nN{{pp0
z2d8CHCcuqk3x-f&JHN>O_}hmNgzM44p_mdS{RNtSl6<E97l6Vu3@T?(wR?MN7j#f^
z=oKZ{(cg2EEBpDzuqc#$O0p(qff;p{1#FpPkC)1SIF#!ETd`UVh}@fN_Xa7az>R|y
zQLIBxf0Y%MSlz#WpQlaj>tz)mRJVo$c!k`%RoDpo+gHaJ%d5!BPuLd~7G7f5dS<FK
z$&aHD0NQ_@j*@)70P+){pzRIn21w!gHbr*>_5UD0l5jdU*CW-Lz~m_)BYKeA+1QIF
zo3^9T=+1Nv?i9sDZ=G-hI~trgBGM?)$``t^N$DTDRxKz=KwZt<;iZ_BeR=oKi9`&p
zD==R3W^U*~7nR;uCGDnV$!kYTQ73u6)~K_8s!EqHJ$DhfKt&fz|50Oxt<~ZuU1nH;
z1nrBrc}mV=Yz=6J_<nTnXNMw6xD6-=s4lnc=UV@0QSKrH>^8~R#)xdMLh0lDV*0Nx
zbD#6@?WkAmWq^-J>c0)g#D~mH<hf^4@1K%3JscnfnLxuSeBqM9wb_1JB)UH+%>Iir
z5T6CxTHh8IiI`-@y&#E7K&+qY8l2~DAnxDfWF#9%mOC$k59@%m9BoXA!ISqT9#8)-
zW6d`3OCF%wTR>-BPr#c&*)oQ1o(R%(v`f9m-?%!w7#RG@6OspYv+ar1>5~94`3*n&
zm)PN12GWe%;uZoiAxVKd2Ugj**k=nu7T-O|$jCMZ;$OYG0D%q}CVpud&b2~%KMw%V
zk~;7U9DrevlXm+8B4RW7FnK7$r)$CHBXpPFP$sAv8Y5_<A<QTHj|e!r^+7VBRbEeX
zW{_y`>RRkus9e6VG}Rf}y|E8rT-3d@^nc0=>%P3G_3mi<1FRcj${B)7|FBfyw}*G7
zfxGK_O^}Z0=x2mBKn@!TIV{H-PALAepJ3$RkWE!dd-qHK`cfcR3K?K7*){wzP*TXQ
zU8De9cch;pGiahJ^YRBZ4H=>2>1z!0I=T>?)*Bn34|pZk5U}H}vAz$vO+4n!Tic>q
zGZxjOfq-utA<I6R2$FW-rT<a1K?;|^SPST~Hi|+B{}^Ool<Ch7eLHppIP9bfO$SV@
zGx3G9LN4kQe@#6<?#FWRB9v>$T8-0m%Zg!Z>BDl-9tt3{wLbC+0vH%-9~KDTYS}Ng
zxt1Vae16M#Km@@uY5M>O7SLaU{~{rKu_s7$B{ws2lIHLBgak|86nvw;fq^eUW!&b*
zoY_;h$vDT%*I(q2a^~y9`orM4|8h+4iY6#W;KlfJeo$b;=MIJX)j%xNuS|&G)AxSa
z-~S;FyKzth`u&L7eg!2&0t(#AQ8bBr^O4<Ybbxkm<mQAQA`?Y4sg-B;KhZo33A3~u
zGlZb&rt*A-1dknTz1+dR!`~9~*LU<DZrp0`b5Ia;!Df@iF=$Z#^ql{epO&suQqU=G
z+ylW_CKVbFFhP3*j(fz|ONx><`^klCD01*)zz8u1>4{m(F<=L@o!rRa5c&_ZQO}_A
zX{B%4k5CWR!YD2B%BbMVyruW&-B6?6hC!f&CW>|NNz70k`BRX)!~&R_jr25)5-y>|
znE&B{X0+~iaN-(_CXzDo8#^Z%#Q*X!PMQn$VVX);9%?7Bv}s_rUF$CUbQC6iMs<6o
zGg?%YI5OS=Ci|4st;IkD_bf*(!tPV>l(|H0N2_ba9T)p1sHKT=@b-U*T#?}aAXnbY
z31H`dT@Su6)l+=1F8!ZpPe4V9qsb0Z3nnmsa;c|`_qz)8E^vs{75Tiyd3JchC~d1)
zZM&62utqS4bC|mG(_W)*?0w--FwDbnOE>IZB!{R%hOoWhKRnYlxY!BT|CI*>h8v;S
zzJo72j4pbOW`x_-dbtFODFuodtIW}}pLdJz07_H9#n|`>O71uep?#ydDNELk7I-a9
zK0x+f?q$P#P())m1quOf6)#L4wIzfq&Hx6r{?GEZ_ft1S=(rH!{Xc~ckRg9@Mpm{I
zZ+>F|{y494<h+J30Z(k@02i>1J{Xbf=;;=~3x+nytQN-0Bxck0T6+lXS&BhPTNK5q
zSlDF!*IMwWXEuUvPr9$KAM$j42NgL5L?yK0(Ej4-{`!gjm^J%~hFLRe^itxi&Zh5}
zsWvmTY{Fiyu9WH|%-All07f^2wfn_xk@63FfDQnJ+&6^FGQJrJrE}f$bpF+59Kvda
zYva9Rf$Z45%eLoJ57R=0pf3JBcvq?@5$xYdN(`*4MN4(x2kZ%Q47C+VWxb-#!S+XO
zrByT5;UxPCtT6XxaAnH_X;w^=HS6wNT?;)D-}?D)D^c;Sl}zYYDFXXFOQS}?bxwxf
zlH1^jFS1?f+ZSPoTmIg_eTL<C8Wa>1$ORPa5swKB--CQ?M?m^tq!8M_0zZw@ZGGej
zg?kZ<hl_sH+a9wYh23xZ55M4bi~=B-AiLX{#N9;;OPeo-F1-`Z(HwdrJFYKj#b|Ag
zT?Fem5Q{`A`#T5!Q2<DEL4EBQGZ7m-3&=?*U{zW{Ib-7s<lz<)d6J;T1s6sE;TH<-
zS13g@eQzx<tP4Sq-89xc#WL@Y#Ayw2t4CqVRFIiSEnyA2+Xdx2={q|=`!owVWHZZj
zU4+b^DQARCl)V4nh)>RhMbT#+B#~DuLv#FxMBW9idaJbvL74;V1rpg{wL`6QR9+sL
z_iCku8Z7#aKcum8AkedKALSen0{0_5E88xOnTxTf?B);Vh=$hI4v=NV21C|5@5MM^
zoN&Fw>|i*ixc>W5c*6mBL&Jza>B`?#8$u9d;QlPWoBDxU*?~<wKR>{g|AP)yiXcIj
zd`8V&mz3M2gAQlt@x(frWr2L>K0y_wvqSCiY+@Iq=uFeXTn3lhu3C<?9%jmUqyl!`
zX6_~De1)uZ5M8ebEk7)o1E*V3K8v*b{@&m2No*xG9OT3mE&@b<@(&IY))9`%-}sk<
z+)sT4`{iQE{?B2Mj6V&7j1*{2O%ge3v&?x1g*|%6>ntXuXnvhf?6S^U=FatXyMB@(
zxb(n<tz4mheN;%BLFx{0cdp%__do(8Yz1mAMh-rPpT`lnUk&}Zfa>WAk`w)}+F~Ao
zL<lMicT235hCl8uO1^E(@AYG1+9<elz=(1EQ+gO$^J3%`379RYw5<pU>^Q;U9t|B%
zTj+55-uMp<tgkTJoJaJ$@+2I&8y-9<;`~e?H(fMIH7ERMcRr6)R!gR7ObX=)gl*6K
zO6HkHv!!|gKY@~qI*(rYQ_QF2wwmEKq==L^{przvaC^umKLMMy7N|Ypojx!*n^0&O
zdx)7S`)AWwB#&e`Y4)FJgjg1-jWXV;eI5;o{pF&T8~4g^>+`W6w~^{)o@x4-f0r+M
zKDeg`qiyM+H+$8$<3#UN*-Oz_26^>DR%K~wS>dq3cl+^cKye`O7|5doJGdW5Z$p&*
zj;!444;=pgm5^ZVg0A$Koiei}vCGu%IN*4w)k;Cjdm<7=64>`|q7U+~4KXx2%mDQT
z@*~^>fvQ6Us^nh;>OCS*8tcnCP|MbSirqf%A{X<~5d!X0fAw3x%ML>gY#EkPG(k&b
z4EQ8NfXlkEtu1_K$7x8Fn~Uo$)E6y*F;g0-7!Vf!PT*ch0hD#AMK;}^72_FX?voc1
zxRAtuZ>$9DUJyXpM%zXIMO`6ag#5A0sTYwjWE8xt$WO0o&8hrak38k*{Mp>?WF%dG
z^x>Glct7FoNUj^OXn<tJ8TMbR4m!+mG`6((Z)&Q$y13o_761CoLl}5L);#QbSOTlr
zm78~=tY<lsQK+i9Q0aCUVI+MC|0OS+`+DY$Xigx$ok*A4_;;DPMJLP46{7{FJ}bEn
zm`78CWhEJNr`Xl$0YAN-nFAw?-b7UAf#@{j!vuF0VFAP7y5^l}pdfDotdGH9sXbCt
zzRB}YQ!@%+X_E6rqsKP~7F<w`+ab?RHOc<jIFnd`WxLFYMTl6yv)~1Q<z%>uVUfCH
zrE%f>znm6=6IkyV1T%b`v0kQ*w|A4AZGTSfbJ-F7Nkyc@)@4L_8cPkS#N3EXESGYp
zQ)D)DzjC4hO7pO`N5ZsX*CquLyO>$&+T(6QfVH=Z(XKclBpiQV(X+bx+JbT8!qZ=l
z524M;9zNV4FlVldCmb&-=#=BN^u7b3ghm!*0D5f0+MG`?ck6yel?Se4O?yg8%1xh9
z7o0b&!%KEdY-}=oW`pd=?q=o5`{sYhiklu1WFofcKLwe;!??;a#08nnAvWE*ceitg
zvry)KcWSvi`5$2Yv7*9fv%;*1k`b_ZoE<&`$DGMxt;{byERb4yj20Ig7=xFyp6<#Z
zRP(Y?JWJkNNWzUPv~5jjgVzrbm*~;4t$?p1oB9SPnqXiA2u?$R7rDuolo{dIJ4t?f
z?6qv=@kuXW<ktkiwT<avouu;B;RAA{s54J8uS7UFBMhUfbA4*Sg>+W$lI>ZNb6A)o
zQhBKMa0*?!tH7*%4hFP>I&UBQU;Sw>zimTSAPwJ!$A$^4IBk%G_|N}h5RvNYe0oJF
zcKhDg_i-9Vl{>~8&s8*^Xcv^P`p>H8s@r09+fj&I(L9-^8v6gn>T4pd4$5q{P-e>|
z#kmjhAXLVdcmcDwk;+_hSw5juD4L*$U8lH9$Q-KVP>YtFrSh$`B2%E`I*)CGm#vw<
z1hknB9zI!RYf-jMpq3}?dVI}%R;~?iKIs6Mws7UIGABUB4c<j<Iwm#T?68fPjuLwO
zjVDDdE3TaC;~BRL_%WMDr1}cY*4y5cv%`nKtfV~k+k_{>oLeO0Z?FiGh%!e&BFfD3
zFP>|!I5o5Iw$chFbvmuTg(@Xlc<bf-4ybf0{|wa+r*4ao4U!MpH7@T$0+k>nP<bCd
zi*r!9b#iKz85Ss?ZZvH<(EzlAu-@3AB##_QRYxKH2sHbo2>!G_OfiEZTf&w1Tgk<J
ze5LaA7_7{!D5Ys5besHQ4fULbV`-e;2xtI7hcq3I)08wCDldEc{{8kvH64q*+}sa?
z^JSS}XO?$X+GU{sV?9v+3GWq0HuN9OvOnJc5fAZWClWvIn-m-<O}9)TmgJzZTNw(3
z4jZIULHeh-80p=Xa}U3*%uAK?Q6yy=O%>bndMSU3HWbJL*u*M+6-;3B{TRISjfjq|
zak6WzgxO3cJb4m}cpU)+l;xu)@m?V=>bC59;EmzEZ&V#|KT-*&Fxdq`CeSEpa^Ceh
zWc<N4!#&{XX0smj+WO2`AvPZ0w<noYZm2nwx>xLw^STZK?id>z^DMm#<+047?*vH4
zP}>G=Fqqi_bL3(3EF2tRV2PW7i;u;ueDmGAzz#}EN-ncO;VE?9d#0GB8=>E9tHbb4
zQ{b&tG(+la+Wg0WsjDhzlx9*95_D!q9Mq9uux?ad`Jas~0z%ksbNzzdzQ@oZ%A@$g
zxO<6BRdF7xF-k=CIM6Z_JcFs=q%Rv&N_h{hDvV|P{KiVu;zECw=3SUQtxZyaDAv%o
zTy6lA;H~-O<H)6_yRS^1U7X!HlY4jtETJDPx=@I7T86ijx=eS4fYF;kmh`?K_7#bv
zWP75N1ffx!d1*|9!s{s@GP?y#VX_i;XzF|p+!4>a@Q-#w#-(JMHEq!JUF(?!3zycF
zrKi_h`zSE}=}+{wsQ%BLO_-A<S#=?IQ0-m)(IL@Gt?j=z_|p19|Da+GWkyRlsE|Gj
zW{|DD8GQ>{JKX}4$S2TtS*HZH|KNVxA!o$VZ4FRO*`bc-0@=dHt?9Jgx`2AK5~z%C
zf4#AL6eehx4FP47odai0Md6cl(Af|dN4g#}cdu|Prl%$r3&#(Y1TitGF0u9aGWr;7
zQ)M*)uoQ>m{+#OLu{XFkrJw}lqVuMk2&(x8a~mT{4m(99$FWK<2!TNj<g12I1>lqE
z92h0pF}K}CJO5@jp(pVbFU3)pn2Uvl^vsNlB1@LK>Boerx+k5p6t9LLRJyKMhp*Y#
z{HgJK-{=~-QybvQFF&Yjcuvmd9@zcS3S|#?t|WuzaPNsxL!7e^t5W-(S^p!snn1dt
zy`de5D@%j0y54>mo$8_ggcj-KwtlPtuHN24bG0m8oi6^jg5%I^*>8)H5Wz8bC>Y#6
z@QKgfLCltzVqB3Edm1p%IYRPL=d6Xal~_Ko(*}{veq-70bz(uFyWX$5F$iDcg4opD
z#c*hAu#sPr_c7X52_cU}0Bw0pl)!!r-t}k;3H`=B@3RkTqxZyp$>tq8h!=b5<Xxg9
z(ALsCGEVSedG@I4dB(Yc5;aqpp56p|bNC`J=-)fp=<~4UCX0!gIl6$Rdqm?%G7J@;
zc+vt8-NC+1ArX;*5B8LJSR4?S=3m(31Kum&kmw-qg)w*n??tb<7qm*wm`t4kYke^@
zSqGvVB!)vO9NLi0x5!jtAB3dz*dQ2^t`ncNV&DL}_B5*KA!yV}(`!_0{uTRlamdtT
zc6O@z1Q*7Z?I78(5>^Q<+ovJU$+{|>4j(A293leTDM%q0je5S+&p^V7)uYf2IJ-fO
zzzFb0=u+-`lOfX5(ql;996$=V+J-l8dhhX<yVv96<G(Yz(VA69CuC#tIR`=M)c6ob
zu&SiK5*0gqmIQ%<qL8c}v9%=6%8&*|uPZqs4Iz6iLF&ejmv`ar--X`fp?8~n_O%F^
znKAc-8~9VACVm7_g2`J~co)-I-(G8{5V@q(xrdw&*{zD1o-M_MP3o`L(-pt}G0T$x
zUP$0H1k_0No11q)2&l@}d*Who!RP}!-6Ssj>RK6B-mXCa=~!}M50U=9*EYrsSQThm
zJat}iAe>E988r#<8nWi&%}JY!@mbl~3J3=!C#(n&XC*Io*`6yL|L(_Gz<tixj~i~$
zLRFz)Zxs~0T_9jjW*c&IjNsM9_m~+zeSw6Mqlq5S1&X+;h-cp4HK$Cu$7~Gela829
z@}`{lnnt%t%QC|40S4J$t0*a{thA4PD+kE-O`G{IZ?3fq_j7>k+V;i|Fdk@z_0TKm
zVX0Z!P(#@V9RA{aPd1Wd_PzVe2X{Nz81(x-dqZHqFw<e&<1at?QQ&HShh1X`cH;$f
zComk;SIjF*!q3Ts<HXFs3@KUn6o82Yri}pt_!17Z4eP*%{+1&wHso#*JqKBRVfA23
z1R%Q10Y-o`JdtC8K+}ft;BW#jkox`ZCm%wh()yfdus_>A2^lBi`C*X+qfY0#J0D4t
zV{o4U{U4p9T<nX^Q7@$;vn=ZVY*ia}joY8mG?Ln1E>6wwmoxz~4LiZrwa*}fX+DSD
zn)bA3rBs~$q?;Xnw)Zz7n7jv$O2_ZxrC@XxZhy6Yt1u9oRq@jyQph$F!UwGaC+i7>
zng!l4Y$9E+qP!Gz*s}3lgn9w(c}t{&v3JCKnhIzqFcZI<gP+;<lxC2dG;9Xs$V+8C
z0SLfCE%u^Og!DU*g`Du9mI#*(4Rm$Df0RO9SzwT@<tNF(C|R0@N<>~(`IfZk&Y}c5
z*T4@JhVC`x37cA4UWRjMNkgl8EVub$sNcon3)2H_Kr4!XnBG+58TV#}9gOy7z6V6@
zpxB1mJ<qmSAsK1u=X5dTZm6Jo##E|k%WtrZ(Kh4JAooDmn@7Ew9I#I#Y^6;|{$kmx
z(t;QuuTwuv7^(gI(UG`Z@oT#RYo@Q<+0iV3vtq~c<3yD4TbRvwu{gw8fnS(O$}HCa
zIdAExYI<mhMNKQ!k4b3L<vOmLD!PTh4jT2JEAhWJx)~W6C8GGvF2cI+<$0A|6oT%K
zY+4j)l8p_6v0BH00@HzM!lIY)l-6zCg56K!oY&CalOMPisb?l7OL6Ai=Ok#n54vc_
z;{EO(zB08r=G|MkESOR}3p@scPk%XUe6&n*`n9|Mrn(sQvs^KYvdjUqTwa&-?%f$t
z#ts+JPg|I4SQ<2T7l-n;c2iu4lLg~8Mc?nNyE-f+{k$kuHmSJ4_~EB1pVI2`<WEUD
zXC^(TC9mG|Y;4U!Z?{BJ2ft`rcdK;1mydT{pm^h~>pv^o4uuvr<COJ@>9S^2yz_Zj
z6PLfU`bECHH-3425=o89Z54m*Oz>CuZcQE$;T2iY%#XmOwG;n2`Sj=ePLhWYA1(u0
z&{0n^!|7#scvGz}sR3|rW4)&lb17Y)R7q97Y^~9BE!MVMr}#Unl6$B*ZPFQVRLC4$
z*QN*2yFs=zOmaKt^m;6Tdql;l6vwV_UB)lAoVH4acIDuf86k_$(|<f3n%wS!#Cz+d
z$-LoW#9hW~@xOLZ3zJ>91U=L85=84w`inZYWA0SY!Mc&%wX6xH_@S3qVP{9_0)F@8
zZH&Oh7Xob6$l3jx&2m{V`I`N@b4EtEr(|a(;Tu91%o@A2ypEk(H&9}`-lJBQsp;It
z;jDC-^1w7~K#VaVju?3a!b#2ShC<O+Z0K48n4;|0dxuXxTXd>jzb+frLswo%)r%Ji
zbtb?a@YRPX{`PKspfZ60LU{niv4(hXj}_TG)b$Q_`;8knxJ<rYGMmhL9t5$SU5OQ?
z{b{1vqgr|ioGU6-DF~2mXLH3FYtr0U?Wuw&Y}5F-yVfU*e<xO82skmvyeM>oAtXjq
zm><ul#rO;&@Odh;QboX9l|`Jv5b_Q|YCaA}_C8rKEnAX_7*pR|IA9Zv1f7~6j7})J
zvM40n@(M=kC<}BI5n^3*^<n_Q=0dem%f?NMs`LWH(U%|t$1zz<TJ33nR$Cweqmlg{
z-M2sGC=t`EO@&Ek0Fx}k!ZQbn6Pk3apWj?t0mr%Sy%B)Z=zgS>)mtennxHp1Kir3`
zCZlkt6cQgEec3i@<}PF9qRqCQkAn&svGH_|^ri*Y$3o>be|jBuXlXCy17BAn8vlJ0
zBilUVL(q_#(2D?N0j|~k<ZYL6y~uh$BeFds=IPJPsCoP+iXX7kbDYZX)?DroD={ss
zD?gL<`rfr*fN_**mpJ0zy}(93?{{-J8^(%6H^)frXLIU#eqAJ5bz#t)yb%;e*{exu
z$G7$IY}A5dw{bgAjk323ulSe#+?t6Rl1V*tePzYH<IK~QR({JbviyyRfp~ZUGq_0i
zYeVvOIhmw0mK*@T%k<n8mR?uqv1&V39ii~-cgbE+SP}_#kl*T*;cSGB4K&$lPZFe*
zvo~J*(7s^HD(zhV*lfhXI#^*gB#K0vWyXYEbAn%;Sjij~%GMp8E>KRj(9CYyy^D9y
zP3BBfSZI|fi&tyM!k{MsGK@EqRBrOUWT8A1s&vSROrr4uDk@nZ%wwxx1)yxQ<M?Yy
zbWx^17jA^`5GgsTx|v!Z?RJb_=G6Mam6f^<;Lx+Muqyub`W4mQz6Rq|OT}%~Nh#-h
z9u6FrmN!-WxO)$xU2zl0RqA`<(NN%Zy9Iv-mb=wib3Md8tO=c!aj*^tO1^Cle_&ZS
zgnbG$px{s>%Wn-xMKpkw(}U4f{1^hIkZ^i;lfDf;<(KB&#Hu44fX*|4Nfouydn3zj
zjSjGshc@M!(DW}ouC4llN3E<c)2g*Y(j|{PL!vT3zj4n}G%&mvHoqcIUZisM1K_Y6
zg|5!ndlCw#iP7sp#)543w#r^>+3cpfBZf>ph?{@<mMV5zO=n`&JHKn=Q9SFcBitp`
zag!90)QcWa*u>a;4P1RaZC+1fYqw7JQ2ea+UBz9Ck}da#<o?f|WHgxqOxh9-%ZEr-
zoyN|ldPO3Z81`YrpR;)?8cS&xSO~Fx7TxrEK7otKasy*=wfK0qY8DNcqUEC=O%<oF
zx6dS@u~dahb>W|1Bu-D8PY>Cd7tYg7HU5zpB+3eNZ{2t5f_u+wmSqld3jf>~9psEt
zylRry?)NadO8}PVEZMwN(KAUbe|AfrB<E~C!QE8)GJ@+j8Pl`!J`6T@N5m&{lHjc>
z#0UKZRQFA;^byCdnX)bI1ndbnTFr8!t1`p(P=!GzK{ayj1OJU`JaN|oIk%oq)8Vfn
zE~nUso;k=W>Gi)AHNL(q#HoF#05HXaoUgX}hJL=A;x-0XJq|=Qq}>>S&x|iPc{-GM
zFh|s<q1g7>XJ&&MCTq4jdGpJAbsc4!ZQOMe2S<}NkqM(NlwcV#+f0V8BxDgIyX+=o
zyee}!J5MDQ;xUZ7JE^_TXAZgsl?d@iPc%0p7~i`HybZc0F^qM12+@&Cz5Cz6kgiKY
z5v;?)KzT1LABcURcUTqT6Ufef=I3@C7gkGCZKaA)dTL_4dV=+Inzadevcl8!G{JgG
z`_<`gnK;LJ0kdRPYqV-?A*<-BO}YxOIh3f<IxSW`^GF?8TYWFKD;4+jr~SjP+4qWM
zoW+RA?S8hJxr`D8rO{`Hh`mz{d_;}$S)P9VlU~pIopS7l4>rulKQLf$zdh3KwrB55
zDBT);LSC)NV9V$zY~{?VunD2I!{HqVy4To#k4g~&Ab?l$9$~C*&6rO8DJ3gQ<d72P
zhI%h7Ms~_=vQn}Rzy19(Bsk^eUJkoP0ImhJ^1yabhtHNf5k|DxpY!Xp)lL0~UP89R
zfYr=P%^^^`x$Je(ZQ8WXF>85fCN=GyW^|lfDtYf)m#j4I)Iqvr`eCSNYCc5EIqq6J
z4%-7(aB#i6sVdzke4W#<l^kBET=59gVdr@xbZ!#6ZmBx5*Wr^AcH7cYJ)PgyZPFi?
z5yy2`$vK>G4UfXE_$6P_Z!CbSp7|y4RM17>`yff4tfJ?X<USix9?z-g>aeTmcqdEJ
zP^2s3UP?QGLrua6c6KUT5y2hY>1rCIxzx}if)3a*me(27z2F0$O#`+{puR2mJP%Fo
zHN>2uW8HZSo9vw*229`0STh-0(f{O}I_bZ4{|KgBs=X%FVPT-8*Rig9#xF7mfz`rZ
z#s^zb1`{r)2@f*29k4o9XA#4+4{?}n37v6;QHKJ(5Ck#(7Zk|I0u#VAh-lrsdC2W6
zbX$R>sO9+{isYB(l+n>Rm55uI`pQ(TXbocWQu~8csz+oDHY1lZ)I$34qn!UU5v#B;
zlswH`zxh|@5o{cXC20yIx0>tbKTgt`6@FA;=;5x%KV_%->t_wI*sAxp7CdZN4u__~
zdpKD7k$=`C+>794GRfi%3Z1uJ^$Tb6vwUbB8mQ#>wqLg`m)ftqg+&#`)l?u*thKMN
z-8Jk=bgE#z`^NoN74~`6BXO<5q^}=6tXq)?Solz>lWmo>cCs;oABW4PX+`Qt+$aXa
zp&nbY@ttsi!c=%?vF2S_*Y4A`HXM`pk9L0*+P~`+Hf&&D3BGXBiy+DK=ouy{ftP2c
zUQT&FCrcu=c41(Bside#De>+CSMzB~3A?AgwnFWF+_U%QrtITKO0t?UzjhsabZaJc
z9p}o1^2BE9@^LiUz<sFH4XSH$zBwn%s5SF7=Rv+KYpEzb<*3H-$dXw7)=%fpgIKQ2
z{v`X_g2P31y=IYduxCi0Jnk}Q`Q`+b{jeLZll_{Pk>5HAbEJA%v3JXuT#Js$wU=?b
z+q>&S`2o@9#CT$vCG&!Fy6UBN782Bk)!g$~!DSJgdh_^l!u+H5r4X1G#6^2h4Q1c+
zcrw$|nkRascTWEf3pduO$ixu0J~`F<iT)$Ou?+f-_Cfpvtb^m1-un+y!6QFt(KTr@
z;E}tvv7sdidsAOd+VXs#!wd;k<5KLDX9}C2V!}E}zxVZw5SuJ+^z7|fCj8(-_}57i
z7H^M4MtI5NToB&I>A8_l7N=7=n!+TGZ6WVkZPo7+e0cxg!Ji~24DBa);4im$uv_M1
zj82T(-J&nRQ?M&1Bd+kAM7C}tX%Ix#fHfAKh23i(wXIrtqdcR|$(#}Z{-p43JE;AI
z)4jHx>b%i>I+ou|cU;PAYpJ6s(kXy=Dug8)%n!B{pVdT*<<5r1RU0n-TU(F_FkbJ3
zaN<rV&#LG|AtU}-4~kk`Zk9v0k@<rFW7t&irH*(61?6dX!Akh8_TiBwKVe1PBf7*i
z>YLTvTYf=HHRIpJosbMQ1=2G8rk7Im%7`HuH43-3gdH8&?;N<?70Xy3cxU?^Yc-^-
zuhvLzp<Knkj9;5%K2oj8i_qp`W_p`pF?&KQRpMbQQ)t$7et38%Ou@*&-&0(l(?dTY
zLV%R+{R7?J3@tS!C8aUodR>MYm}a0*;~=EC$V2`%!*j_9uYJkJnO<(`eE%kgaQzny
z&WM0wFMr~d(LtY9@#<cybVb<Bmhfc0U*jSqiX2~<nchIH+4#-b`A<c*R5Xx{b@mKW
z3s_#ru+YV8&neKoE@7t{hQ?lJ>nD{CV@>)r2KE~n5UQDs5g*2Wae2ah;8+?fw&0fc
zwiw@ssxq>6K4jV0+N>+;RR0|h;ucw9nfs%`zXe)l;syfg2kf{zYPd2_AWn#0_C8T{
z|E1>BfhH^J9?JZiikitLWL(V&{`vQYE8PmRn5kN|PKj7e_!n4&4sM^w+=LWY2}t9N
z)Pp)4nfO+!c>PcN&=vKZpy+Z$8$>)D6|-M{yQckoQeBvW>T$ov4DC9;epJ3Az05M`
zirH4C*H%v5$j6-~wydss9d*bDF=`scPg3SJ$@+<JwKuqGy}flJX`WI<LgG{wbi%`F
zv@poy1;hi(foTKpqNoRNzU6q+H^=jX9u^5fT7{<FgFZ`g9_vSey<GueF~h5JkmT`1
zsh}y>t6^=Gapy9*n)d6*KklKNbGB(GZA+J2d;;Vb`H9OI7cT}}W^6M6ndz&=dAd@`
z`>aCmuE)tdYjWeGe9n0*x*38&8i6>sT=%}{cR?ZK#MVqAk^k-A(=T-E&dA5}pkeic
zs|kzLj+gHR!Oz}Ef`YB8=bbyyOnR)R_?MdVDx5Tg@Poxy@h+9@|MZqBXe{TALCz5k
zGhK#eMlKl!D>Lz9M4UcBqY>;HZw!d)B-b2pMSFHPMuT`eZj6f`70@rO_8Gj18L4q}
zD%DC>V5nhbAZN5XD~hcupzMXxH}G6^ZP9Y558xk6c2@Q$9#Y|A`|dGbIOL+G>$G%3
z;^Ri3(ER6DW6crmmKO}q>Mu4vBfNqvDS=}2JFSJ+5<Waj`cSzqt{APcXds4)iR?BF
z#+{hy2lf$yA900{KRhnId95T8CqeD?bK#>)wAK8gr}zagtwS4^5q696k;Ts$HYo+*
zZh4BX!MY~W_JD~t$RTkwVnQ%|KU|*$UKZC?6Yw9)fKtt)H_sKRG1Gy)GmL|tLSkN%
z*Nfs#8aU_WxzNwB&WnJ};4PhH$YmUzfJrp#8?5FMM{9wJC54^Fn@TWQS3aj)_mF$Y
ztid_o$N^M(F0;g?+Fb&wcdr_o?Aod{&0Z>Y`I;2jF6cf!jMb3hAPJ$R!N%f<yR4G`
z{GDfPe7CFWXbn%QCb)EOzTsg>RJcItg<{(z%N?_77h#4YD@jrI;l(rdM{{$=3BmT)
zw4Gz41`P}qYtY#!l)jY=lQgTj0+QB|m=nvF<IV*f&s%R;v?94vi=Ug+kxy%9N_*+9
z>xkp<<HCV=QlAVLOhz7S8I*kzV(p<QWs7QDQujk@SaWec`PvKFc^y6>hqN*@c?E3e
zV+tZw7O}kTetkO1v_Z4bqAo`NjOy{cr!`G<R6@+)66CLsdVM-dv=#-PwS$o7Z~JOY
zJiJxtytS$r7j*M!1Qx2)<Eu#aYtlOZb&167<8fy66S%deWtOZnezEIUDn`pJMsGg$
zz)Q~>!M*O%Z`?Ylb$C);r0-s4OPq+-86Km%{5<UP*x)RXFOnKW!p+UU@*)C#i9viM
zvYMNdBp&o(4L{xQ5sPq%YyBN5SDf#0hrA|>h#7GA{~i3fD2Cm#+>!!+3%-E!YCu5o
zY@<c6`0UAqMJE=%j+1c=bAb@M!`9*?9+u!Er!7~t%*v9qXz3J@(eY-!ht{tQUR2}G
zHz?p!%G@|#=hH;OjKs)khWW>&SScHjrI&?+4r%j}Xd7;?BZ6MUoeUIL)xU&#EWTc5
zj(2)vu#+#Sj3eOWDhBg`9E?#a6Ov)Kv)n|{DHJ(XKK;QgtH)P~PrhuzVmU;ajrn?Y
z*n|3*#n3F{=CDorL;#JzwR5z=ot4;VE|o1j2683BOt!Lghk!G@FQ6u%r;yb&lC(&j
z)hk(wZ|g$PzT_f;*6_LS?0iQ<%sgO}*5{ZTYh<KF#Z|s~D3>YQh7$GThIpym$hC+b
z5$Ne9%*JTi6*L(=W)A-m-@RKIX!6V?)XuQl@UjDLEqlMb%*zX}EsaAUczC#C0>1Fw
z_~_$ZS-ug<_Ur0;5ajR<Tp4s)Y2|_u)^XHv<@<d6G0(Q7Nju+=Oo-{X!iMncFO?hP
zW^d3bPm7<j7Ot3~QLRo%T|CA@c3H)KL0pXf0<t#v+f7^tT)t_yVPj02n^i{c;I}!d
z>Dq0qo;Q%}Pt~bfClWJa@s%p&5gp3zAm@58O6zCLyivJJY0S=t=@bGRI;1>E?%wXR
zL6m|vdR&;7b7~r+WLI#qG(?1Iv~K8-GV!z8qO8+Wqp1rPE_HqEJsU#xMHZtbk59U8
zwea2X^*V-0U3wt!hpprVqu6~VhYX2gYR_O1w3Ed>cr17R_EV5S-3u)6$L+6-C+XdO
zoB)d}(qrQL+dx-71zDRFdr$z<Hq41-XuB4{Y0Qjuvgo>N^~$;99V9UNLHHvYjoS=7
z6tUs>xpqt6-`zPw_mPp1{5y?{IG-zypT8t4T?=lTTBGXoT-NG5jbFiYC+N;F3gv{>
zu;%gN@Z2g>^$DZL4Z+Dx4`F|;_6th8JdN6%hgA|vLmgVVq*A4qC4*?YMIZHsvd6Sh
z`ZwM_o*kO6eU7%F=o>(qv{t^ca<g(}4w=^wseZY!J-{|15SJBXM>@ZAGLCe_LHaGG
z_0eu6rVlt0WR-ZEEh)=M)P12@=-<JooL$^od?If2UfCgjPA^oNbze%0&C>T9k3$Jt
zj+Dt>LeM5ifHA&X$=>(&3z1Gb*WtuFO9emAM=uIPp^amvaIoJ}kdtJch>GtPOj|th
z+Sr~ikta=`OE(TD?SE4__2qpMz|h=h&h3~-bqqX?_m692NcKjNs(uVt_CztauNd{Q
zvvD&Slrn~KX!V>#%2O9k#}A&rnLG)L(JfXpy=;%)YJb%}BI$RE<HNBARgF*;9ChYv
zHVI-4=KjY92v4y;`@vYN*YH^`3;$V9a3_XKq%U?z7r$LK+T@CTIle1*$@;K}Dfkk1
zqQREqVx!KpUoSu2l@N0$=10F;;tzs4qj9v67oAO#O*3KGkptK4iqjbjmgfmEme8K%
z2?WdZZgJM=`y!JuNK*#ojS?KMNh4B}Bbc)s5C<#v8y?OKzawXJ8CQ+Cfs<pK@Y5+-
zS9I(bA#9U)oEILh9X|OTr^`}}o|nIJK5hJ-Hqin?yHQ51)fJ#%vGrU7L|c#OAd#3d
zH<{T`xejR*CKu~$VmIuoWYQ`yDcH98@tHw3c@x|>QM<~CxTn8%ZDj?2LP?u(T9LfA
zMbJ#OLZzobyiC=dI`f?VXhA5=?Qst?JX@Fv4g&}pOFeEU&Hd;84kNg=j{4#<(7VrQ
zzg=s=)iYezG7*sKKPoAV?aUMUCAg8O`c2bF*dlfNzA?gYZ?qHnw7g}Q?@|;()S6q6
z+N5XdIccZPneF1hA{_1OWNN@yxWv=UhmUnqNflyGWmA?D2JhtKFG7jBm)ubC3X%n#
zjBDkqGB4>MzHi0@*&+L=k<xUG{#synu7S+s(RUCN*mcK@IzV$G_3<@NcJzo(SeMZ`
z;S%%(Jm;!>w{<2F4Gn|QBYCn6{pE=c9b##dM;2>J2<Kbn0+AMuTD3J})zS5uWS5tU
z_Qc1##={m()&_&t#akkQc3bdjN^G7{$Jr^EMVWW<R2OyGv1^Rx&$_n@X5@#Us`D@2
z=~akZ#(=+_ykVtE5~AC24`R{arZ6_4iAGXzNc-|}-w%Z2>vLR33B;YHf-}2gry9$~
zjtV{Gg^I3xqYzVVMMj^84@TwHIq?wzLy4NekVA<5fv+n}mF(@utCo1+r%#wGXWOP|
zv29#&UU<W2%Di6FvU357nEcxza&Pai5=XiQ<OZhb$MC**AIwXsbR9eeb^9jt42|8z
zH*hQ62vkH@>wI=Hv<d@*Z9a?XpUzaxh#LJlQp=5M!>gT#d_v?6EvF|+h_a_z=n~L;
zOm!ss%y0B}8LuvtNvkfxkR`GS9nu_@6T6Rk(j(8)yu6@Tt#aERq`~R#dG_KF@2d=b
zXLGyn#f;Pg{NTgwDOT)awQ4J2Jvt;YVGw3s!I1gQP-Q;MyOL&jfk*3*MSGp|iAa)~
zAlO=`@U76ytk`AW%v27S!Tob-IxhJ?=uz7-ByWd;yRnATWP3c_i4|0w{Y~#<l=qDj
zUk*#Bzs<4JdKxNar(0zHy5FYnadGYzpgdC^Tk41551O2~3^Y?iQnR8vk&mlMWUBM1
zF84{>sfLKR7DXAn`XpRkr*23^8Toi@)Q09-MJpFU+G_hF4y&4W6VeOavf=3yrIGAe
z%e2(VLu?~xqeLFgIJ5Nc8M!I$@k@LT1(nWr)t>n7KLrG`rH=F%=&wS!SpfUEZ$eX@
zX}l@y%b+imN2lLv#_iqzUPR-8L^O95zV3zS3iL=slUL1i?*ub^U+DG3f&f|E?s)Sz
z5TX0gi0}jwJ4*fzhwjE7$`kHG)p3J#-S;?@Cscov`YYkdTSxbzd9>LGC0Yb3pfh&T
z42<&#aROrYl51FN@`pM-B3pdgQk{MTh|Yh_kpM0qM&M#5ak_F`@mOvr^b^lX22uy2
zkI3;v&7@unj-;QwNqf1O!25PJqh#Qd)e{&X4ftSU7;Ni7tbaV_@&oU+kb;hT=f65K
zd{5BfPEkLXR;VG|Z!R%JFDZVlreaaTx_nqT3-rERbav6wImA)Hjjr0NR8QWwp!{Cc
zt50Y`ECMu7yu&mLCBLb5ZMyQv)J?lU9}KoaDO|n2H0O2ERdTMOgsUD%ySrfYAyYz{
z>HNL?^hn=4#IyJ<n()P!ep5cWQukxCbMD<GRV%I?+eBlUae9(+K<vEWM#+_QJLn=Q
z&W%GVl>Um|8e#dhMk^<>_`Jk_1;kk^Z=udr%Eeq!Y*t?wtK{AVBvDG4)^`!og@~En
ztlmMPoR6ejs1lKuiNJ102lGD2jG-LxB{L0JCvF^T_M40rPcqbaYf(WBl;0y03Sb(H
z%6!*~CvD`GoPNg?ZmW!-*1KQx43O`bOSo2Zd!hJOd?uqHi=%e@;!$li`^i(}EY`_q
zv=DG~JZKI7!?Qp#rxBM}AH@L{!A=64L)l1!Cj7JuDUZ}P%EOCY%75N-;5)hOm&cyH
zQKeH_fTn9`5@c~FEi$ETjMB<SY{i(ahQXkM1!J&AfCUs8+-D!i9p|p9mn}2Z=VH8E
z?e~$or6HL{pSxbWJ%TfyK`BoFWi6XrI|a?CFp%k+WciH2_UDAtR(l^)R%8+jmG%c#
z`K|1#H-F!Cu*^ID={uPFI4y%&l~@!Y``~wdyFq5wl^jwEq)<x5PagKoXr*h*oAweO
z>u(idJslI$h2OY{aCt!WGtIl@LyV%3>XOsuctjF7vlG@Ot$dWppV>X0q_<FXA7X;L
zK0h^DW&(wQ%v>Q{yXgE;8^-h9ucCBjHaNm3lF=mz-78t>kF1n|Dt1x6>Co&TH(5)x
zef9X#*tw^9HZ5B|qS5%-B=5waFP5TJm`U*N3Cn|?r$P%%aH)pTvtOtD$ZzEZP-*5r
zItN{;I5Kv1g~BspqmRQMQ$;@$f&L0`Ng4uixa}YJ%5)y?)z_o_btuG&ynC_Yr~`Jr
z<s=d&K0yM;zljy;9dKTa3pk{sPvQr=@<V2VrV32=!nq|LfSC1|9kMu;G1)YukDI{`
zOW!|1Iq1u?^NS-J2o<Q)=CW$5)=$$li+pKsRP#~c3&X0PxdiEy$ZMAXye&ALq%k4H
zKO57e4{f9l(W6k~=w+-M4ND+TwnISU?_{aag=TuUv|ubZ0s=O1R<%r%hOP^=y1pk&
zCkX}RMg$7<&xa9q8IK4IX1%|zHH2+dd7({JDY}W8;N@ja50&fMoIP|x&iPtBem@1u
zB)!(jH6lV91$K(&i)TKxsz0aTG_|!W9&hCf()eH$!UJyPyKr9XBX455i-(UMGaumR
zo@$Ya?68(3XT5c5#2!EtmZJgu1UMawHE99Q6c&K=6xnouJtev$zNWI%9L(TxO~0S~
zvSV97lE8Gt;%r!L;O$Sa8vRRds}G$abt{ene`F}e(5vGpv1(*#>!6j^2cy@IPbxuu
z!j`n7K5wA(J+5st;fOs#{{<5c&pI47ok)}WjTlMNY%jwNpC26EOam&mZ|9jlT8=MU
z+{}LT#~s@tcMKX>4tLDNy?4i<_CL1;Z3x=SzgsH{c%dY}+QBu7yMr>v(7TOY`a^fm
z;$3RyW<CNxD<!#WGim2bwk?J23+lg;q`t)>ygrM*gV2*f|JzpBqUEg@ls~3jc4&=S
zWKZdIAEy&>P^PbGi{|J2PR8Gv`DmMUyx(vZBNMzLy+wc;{xArYw1<>c6zu9GZryBG
z*%Wf3XGd=?ElQ8^H?@6(2ub!0n?i!?_y<4c5p>oJW4QkLX4lQ&4AH?X=0bP>FV+pL
zs-`o%nqsB>T~+K<?^SwCI6d#Pd7HY}mgputF1&o5#_(DWMZ0d5f6MoWOhh-?YBEEH
zPnpe_<>JAn!XUMF&Z?E3g=3?vGq=v-wX1e!pdH^voGj_cOVRk{&sPro1-`no0b8hi
zEU(BzRKzo{&}9vBss<2WBnp;(b>$GAFys>9Q-Pv*;{3@G+tPrZGI5l2#6^zveygCK
z3UHs7M)R7kwQ(UX6G`E(gLh<B%ob3FSjh7%;xw(U$^KA3bIfZ}{drwdxq@9_*(>P(
zi(_{n26tb*_5&&LCr|>jdSai4)DR{Sg%c2plY&Ie-?c+B$MhUCs_e38b6ShkO^yu~
zVJ-Z|)bto9oB)%%65hu;etOYCX(I%d>q<jvpi**Iy^Y=Pp#y&V>23HPj)te$?&_T?
z$qf>d8scu^sQ?@vHvDRWOp4A3T`IX5UeY_WP9My8O?T$6(}P-(5er?*)u1a24l7HR
z@JbBj>L?af&f;W#usBMc#VuUFjd^J_6WzTvV$gGkpDL_2YH=3h;CAIondA|@b;EX3
zT@lA4-?;slxp+q@Iqtl>&CuLA(HcG8CbaMp3due@wrR5wHobg_{#6Gqa#eVtdqrk(
z$AUHCTdarSIO$cpg69Dpn#yGk)|<l~#BZUZ8Aty~2zs-R^`X7bfQc+C+|IujI_T=Z
zrS|&Tkw{+S@UCP<i3kt0V|m27fxP=-g>+lI7^mafYzPpq$%$_LVAO!n%e2QvXhBYH
z5aUs=U)S+Nbllq)?Ts>w*59OAfNGV|HST#9RopQY-X@LUlVBJTkOJgWH_J}!?;M)G
z3LlD`i-)Z2;1cm=5Jx%b8QVW4a*3dQipH&B?H}@Q`Ypt!ZndD#f=v~HMuJHYg$yVy
zcumN^1ASSAqn#CI^N;ECM7&`Teh~AFflKx(yJoJwFLdH=lBT8~S7B)2tm+}0k}>l~
zt}T#0kl*C%XM&XF1(I=GqBBejr7Lo^0;(%tvg~jKzk8+6-k@o^Bunsa67$-@+5WN|
z=k%hco!r_?<CLI2xlAGE%qKMKO0G&p^;7<0PE8$(DHdVL2W5wt3orAe9j+)tKo=&n
z<Ckq_)!k~zT~2@c@?&e=;SK|1xavxa44-aS*s<z7w3R*VKG%n)><dQ8<DBM?tkbnW
zDnXx;MlrtE>T`MxvOKnAzE}pPct{jmq?y)^d0xBml?;EkR_OYM1$6(Lfz=6hLE=oA
zWbptGQz@O{GGK|qJDl-eOd{-F;%?XkPr9C)^cBeNHG^eo%?r@m325tk=i5>-O*$+l
z8y3(r*@Oc6k<l7nFWK{pzwSTaNMnqu&Wm*Q-=^RbT3I${Ro|vDt2F^kQ)i+iu@3%#
zf0c@w&N<*QjnFdvc!O0ny<xUpSMGGrje{`O;u@sEF*wf$kOq?5C5T;Fgje1y`a((+
zUs;?SUY$p1Cme=zi+&y?B5<3;^3Z`55XU|b_Z<1f1Af+n54G%c$!EKaZnaF}Bb!Rv
zq$+lN+l%Nn+^?*8L$^LJOWWPF`Bdh({zz#BoTL6gv9)I11OUicoy|>oj{fh$P2j?t
zjwW;Ama;z}r>OXT;<N`vZZ3s&GFL$NObu&b>ZqTpK$h2}ennhNriAK}jaHUo3}DSJ
zf>?ieq`PUGXft27t{VyDOJj|pl7$y5ye3z5IWHL@X7ncddf4GH#jA`Dt$uqT)zWa`
zfVg`|MeXRwCbR{SVJ?oo#}0J5BJo-QHM2Uk#tdImISnhQ4D{;-RC`}^@EV<y?wiX(
zZ%-rR0sIhV-wHE~oSG|`D=0-fzK=XKgV|2)xFQ!PoKp6{6v_dP?b@(9Xm+SVFEo09
z+GcBkuO<(@ll_HQXw*74M~3Hnz_EUWeik^k`Y7dYY3M&S?JLMQFD~J^X30%ZAN$?+
zk0UQ9O<0^2y3e;~h|BkT4-vkfq;EisGsylAtp~SN2FPEhYGN~99Ku&8y=#LqzR7+S
z>@G2Q(ugAPf{qxsoY0_z#LbaBx@vB&M+|9~a#OE*3IIpSB!zoeEs4=*POMf?O5*2V
z99li)+ef06jrOkS$BVrydUaXedxkN}>iA=O4q@hP%-9o}+~*~I4!cEqQyBDA-7%MR
zvz!mQu3>$r-yG+@+VkkR{bor><t)!no9&w;=!=(66jsN6fey~fY*6(8B~*pmV+0uG
zg_k_TtwnZ{f%^k661YkqEe&ivmWF2`o0fr&tu^(IG5>EYfbkiK{)hD-AARyPsE5~R
zy{ye7&e63+l|1LXFeBk5zP4)>do~jIQZts4PZg}#9%5eDs@lEOtR#wn&;~JE&s0Vz
zM}-~N(h-&?1g)VKo%lKvaB+luX;t3jP=*Zy*snvt3(Oz2fm%E`#LFk~Y(N)WaZI49
zeR~8w@$#O93aFEc)Qv1^ugM=dhB9!CIY>|FAjyCn1ilu-A&qow4$Vgr3rQrgc+eOH
zA2SPeO()+780i=9%eeCTBrh}F)xO<Jx=D054NA!x5Fj!%f?ptbJPdA=>Vw>q#)l`j
z84_Nba=zs(yDXT&k3Q0RNQP72fOGn@-(xNvI#o!dD(=`FKbKl8dgfCVVjR^9F%I?>
zTr+8RM!K7~W!W<+bO*LC2uF+66qmVVWJ+W`mG?f@-grUyK^y{dinE`f9AH6;F%LW6
z`u2Wlv~dQvllcU=hu2FXpo5@6=h)2KMUwy_RZrBjwwa!cJ%VzGo708nbFcGwp%PJy
z$Evc3j^Ed?JML);IULuM9}0+LV~&15VA;%p0ESW^aUVI4ckw-uJk)LXkLi{qoaL7j
zqS5D(?@;U&Z)`3*+oe*%+2`<!{Zo*?3-O=TRt~BeE&!a<KqNO}v_!z5m0ypq&fC`b
z;U(@J^ts~amN5ruKd?aS*rqBaNh8d#9Esh4>KOk-q354`M+I+LNt|%rQ_G?q@5L9W
zpXr^Q-t#_HOR{w{HJGz^^RNlpUso#kN@n#kq-6`}3}Is&EGlxR8C~S%M+Cy5)^9;2
zk!P)Wj!!rIo;S+ONXc%^JMj^R3x2S}+$d!!Y?3S1p#nB+WU4d2ul%JDbb(eEhV=T3
zSauphnS-2Q>s)7%#e<pxh>m>Do$^GPD0tr<4OWqlxpHsdq2I~YX#R*kvkINX0Tkfx
z^{d>EoIokqwVkM9Iz#no%4;5a$6StUCLC$&FkCaq`cs_($p=InR)c{mu_V=x89eoV
z#jExda<b&aYu<=dE5m8MTTB;JQ*S#D-{*6?@%oC_<OOj<QaIGh5FuE#cze?@{=Tlo
zU_`3%LRAo{z6Me)Qq>?0A483_eDJBgP(Gv=e()4xQ;wxmavL@PByIXaniYTtl3aGT
zI<xOdJcNX`3;OJfLm&|em`<sqsdrh);s)AJ2=S`>hP!-X{4BXA+jy(Y8d=0bR&{>4
zWN{KKnmRtJEoH%)m~ZA<S$*4K-9?8hsPe52{4X;L3rn%eN1dVj&SjMe19c?Umum7h
zxzqb*)U4wiRRXK?)R?-becvsbgxm9-nwT1?@fja<TnO7-nUt)o5&E>fP{Elb6Tyc=
zMOeLpWPy;~GC1>)%5)q0+9q{s?QM}l;I6vP7!^s6>hqcnKm7h5*4_fB>n-dWJr<o(
zA}u8)(j5ZQp>%gi3P?AS(%pYlx*MdGZbV92TDrTw{lhrtojdQHZ|>Zg|1g|6jHB$m
zpIB=>>pAuEgrWu^*ORV#V^6WbtQ|Li2iL3@)xBw32ZIu(1YXYWoX$<MjB{KkscT?v
zc!J6tAb~M{$aviMe~|H^0~^7~67O&2V=TdcqK@Dhhz}_q8{^E-aecZW<I0myTkK>2
zAQ9HE`fQHtNWYmF{Q5=0K`;9|{QOEJs}RsV6fd9wUe`$gK(6fA!Jm=A0nRo4QAAGj
zK9D$8qMxkvCF3!TWXYz|YSq5kwLq%H=>*HKD8L<!1tXYAM~EbN9Hw6dv69Ky;u$oQ
z&*M3)XIN|0if>bs7{ZS&C=0SYjDc$JY1wuAq{!j&pbXVWom~Xh?VC5_eMPZ;_<@s%
zmkyU<Zu~KMfytdxwMd<@XwVSlor8wxA!AWFl}wT<Tk!=?nIE7xIX>3M{#aorhUnfv
zD2b%ARRItG2{>Ae;cxqbB;59E*n@uHuV{MP=!1|S@2Yl%p_ClAxx7Rf@_<9Dc$2}L
z{bgrXFluPYpzsMM@)+H+Hi%I6nPT~e6B~SGKt+(Gn+%{TxrlipM4|O4N~uR{Ls;MY
zT0tZTFe@{rC-gur0kCI0U}8+w>I=~VH&wLkV~>i|(nV2ebQ<fi8Pr*2tVVeRK=dm<
z(`$RYa}Lh1f{RtZH84=Y<!B?9X#mFO+G_35p;3A5ZQHri2xi)3+^&G0`R>4rJosf~
z*my7V=lsz&+h0cozLq`%kISOfyWu7f*4;$DDjpM6?MP6gJ%90%UUZJ@UCB5g$^?5M
zF3gThF95u>Umv$eAno&e8&!fos0vhR<(7uL-gm`68$clCEg8i*t|XG$CcytBl@NO5
z7)DGsc{T|meZ?w&NkU;VjOttmhVP@;%#|lqZUaE1v!FFf6|eKQw<FzMbs5r|_a`V2
zeR@{%E_$rV^;g>W8_vTsxGB2L=f3`Id~4waKugy8_!_m@dA$Zf>FM^b1!xypdmD3y
zXPDYL6-A!u1u)u=S+gh`!)YkZcC2YVn<e3*<B(Xvb##p9H=(7=T8lXdm(fQi?2j;N
z-XOfxsCz)Xdbms_x^7&m-v+pF{y}yhB!Q{MYZ3Ct;GjAeT?hD%JHZ6`n}D}{OeNY?
zz^@btXedFzsbI4x7BNUy&Fx;jfe}j~81cXpc!#P}=`*5duP4s#I`D{EJP44vSMmz+
ziroenstwU@*KPAVZ&r>`s1#{hB`-rn(XS7aAUGqGyWUziX9WK%J3X~@dWQtw6m85$
z4gcofO|SQN&@gjV&3~Es6=lKK6DiV$a^4jhEm=u6YnprJqtR28QW9B&T$FZI8;|gr
zH6J>e<!N%dD)(R8OUiDIGX#uXJ+8y^AftA$+Q&mN4|l%(d`*WTwvLwqUinFE-*zL;
z{-qIoC^&8w6L@ef=s}_AMuQ|;hTx$PY5Z|_t_T!*3JMC9;U~L{7f*ZY7Y(f*9?JQp
zCEoR>*8tPv@j#24X#Bcrt$Y&wY8t`R*nMB<>|6fqtA>2#!YFpj2@&yI#>{>=t^yv%
zbj03`E@X7BFX7Lkd&dGu>^cf$|1HpM9H9ce0o$h+hFmaVQ{LFt;^KW$8%nnJw8w!P
zlw7YssXL;$cX$jER}WwhbB)4K_)CIrMsEQ@d63~w9)}GJ-K}-Ea;$xMpE>AJ4}Dov
z0EsdInG>_xx6-Ik0uIP6!VuJgm<^vR&j1S~0ZbUuXIenhBel|>Y6LiQOv?o#L(mk-
zS6QeHdEoa2fS&i&_*CS*(cqS!HQx$|dMm*_5oPrRdTa!deSqyb53;G)e`+A^dnf>e
zZIE&|OFHqR#s;LM*K6^Y00D+DoD95kmM{ROea*N6JRKLY$w`p$0t^h~Q$?;50SvnQ
z%VqIL@C)tfAn`0P%13L8ho8uM@xg~Gk(`_p{dZ|Y>U}~D?!=WrCnmTJ6x86#V8f#A
zmc{gjv0>@|j2mNN!z^zb*!)IYxIQ#@effKNR58Sh=?CIxUzf+!6{auNef((H8`Y+q
zBG0HDfyLw4c~rw<G?|iE)o(O+LLHI*0<&gsZbp71X`MpYKV)zHAh&Eam1EvhSAzGp
zv7_w#54)cf7{cKG>LB??()-LR`c9j=eOecB61@&(W}F2&gFTQBA-G^2-}&s|)he(O
zJeqSm(Q89ji-{OPCBX_H!MF?M41i+TkBAvs6R?^_fZFlVM`F<Bg?br}Iz#p8S?Z`c
zRfG)iLFRVY_(_Yaow}jId;4-<Ncrevm|+uteG+5dq0_&8MEs;n==?_JYtyR8cd;Ss
zZ`_0951Th?@<IGpSAK3dRl(g`riMnsuSx2ZxuVAoUtgd01jzr6xg-uxsrc~kRi~oh
z!<e+mZ;kY<?5PI5dM(fo<D-K%Y=jVS$X2`3CdM=n<ISL7U2c2!unh2~Ya9Z*=%Py)
zg9iIhR%okSF$XfUOEyb%Jh)MGz}PhSr$E^=10q@w^~Y2#$5vKc|8S5aeOo9`pwj%&
z^LO93>~;{FuCMB#Z5WX8iRiF_i8hb>7`heqcCE|)IQo}qQw8*cjt#Dk0jD~ojIJ)%
z<E>aoCPENDGq=PyT?_)(pL=ra|Fb57lpR5YrH7?#PkD1x>It3BY8rjNEK^2ScUMCl
z+noK1NqOpM^ZtCG;A*l}tRz@pUh^*YtBd>Yb`t=4US1s3c5n3C5DPk_-Mt?OA>&^(
z!Pvu;(?X~%a8VEZsg793N%oLri~AB`!qh_Bn4J0i>*Tw%mvD-&{W5r+c67(ab1+og
zGElzoT)e>XNkKAI?(qXvbdE>%VN&THIZpx(N&4*NJF)kCH#tT8XuW0i;<XEGtl6$i
zM@|+L(<e2BV~_j6JlwY}EHJ!OvF5Y1Ejm`z-7vO#hRe8q??(+3Qzm;Cp>hooiwdEP
zCVR|gxfkDgK8Skpa7-usr%(K`0)i_jbzfh&MQgh?Ps-rVo#q!eew*QHMvtu|Q`TRq
zAc@!qdr&jz(ty3LDrrC*jY9jY!>R#L8l;YrG!zU)!~*Fv1K0yvQZSZ}(qxR9Xavup
zA@@|Yvj{N#K{>>U%hq7EP0#MB*5`p^({xn0*VdnnH9&B4<l|*&AH;=)-gV?j4$8UB
zWY%NGu+K7+=k6%qZ_)68`%TV~BQo*dl(GxSb1ym~$U#aGm%YJds8@Q!n{xnob*rq8
z>M`%+-S&C#<A~3}4+{B3zx^v@ctMI$NIx(_;x5!(VMU#2w=DlN)+G{orN4Wtm_Po|
zIwW#GTP{DXGs;HnN7A6R_y_K5H$$;Gt}USlgL=k}Vb>3A??OVY`_G(?9tEepY^QG>
zm;C-b5sCzgPVCW<k&*c0vD3s6iU9xG6n`K_CDsY&(gC=7$fMq>j@npFnT9uDUPeYH
z7R1U`j|%r``!Ph3`uVotx-dlA@7306ygQ%o_!M}8*}X_WW}01FgP5JE-}WwwQHQ_9
zAqB}hkLdVMFi>1rQu+=0)Pu@if4xdOozpcjbVyuCsjKYgO)$94WKYoYx&EoX5vAt1
z77LIEQg$98z1o46wg4f3`7zB@@s1}b-gKZ2^JU>6uj$}NzaQgRcKv;NM;EzgZVOrN
zj050GjGJz~T>>M9_iXmpoPfTbZMO-m34px1H&0Rb@#Jl~q9`eWf?1m_Q!#wp?fE_6
z4?WCGyVAY1IXiI39e}c-i@K_`U1E2cfH?3UygK#)J`S^M8cR=>>&cGZ7dQ;CmLWW6
zq9+vZ#VJq{FO_x>M*{>{0^IS6Q{`Ow%nE0d*3ZSqrsXo!C2$~!oI^y4ip`fU(RoIX
z;3~ksxLMw7lm1*i!P{Q!ls*8htOcr#9^i>$2qILO?UE#y=T&`N*W~7T{y3E=P{0_Y
zc>eCFy-z^Pz~!Oe1_OXlgy7Sgq~rYj1gK)l<#e?arRK)(@xmm5>hBn2gAR8Du7Cc9
zA^5u&y?Y@m`L%L<jBQ0Vvdo6PhTrRuP5L`^0?<2G@0>67Mz*Ne^m5rsZrzWU>VNdS
z@>=5!f66dYR#x|%eg0CmifDIG>Lek?51L)5a^e+{XbA7!D(HVl38?l9Y79mIxWe~z
z{uAQq@-YbMS!dkeNbOXQJYbBE0c~LSMu`Nw<9sBmQPxP94Fh$btP&|-gcwpXvb=OZ
z7!<PwXSaSNr)2t?_Pv_-As}n40mq7g`;_B=8RmI17w`>C>*njxTw3Ac(LMVNVaZ!J
zok^F6`(qgjbYuWyrJHxT@ztmW^fs?mX_s?`<Fkm8=5V=Z2{wW0;yg>LkJR6Chqr%S
zz>ul-r%weSTW$cVf9?C4*Q{=Ts#>b=H|4TkSey82_k!j155Z|J1FAG<8v!<grnUW0
z+B;1>l)*bF&W*shff-1U5i+!S<Kb!9Gy)2IMFb2kK@u{<4=BZtoVmsIAaU?ArgV3U
z0|3>|%ZVm$&i{Gl28a;(Vr@Q3Y~$vA2<Zi7&=C8fjokx27L5_>Bbc#$QR-DvgPcag
z&?)6-Q3GLw{3deMhJm{*4z)FEuZfK1tsq?c;$$}psP%b1j4;%>pF6Bh)s)lgH1f_n
z(nZb`I|0qT+7<xa)tWrq#R1XZ+g>mW<_P3yl>ZjM0Nl6h_4OgT2|^mbniMuYRaGF!
zGO<@4q?MTCQqLGigQ`&g0T2ei)Jyx@q+C$_P>D048|OM~zISTZV?55ZZLpa({_-Np
zSiVXNp5{eEoMCqoLz$^r-bLvgwHk{9XQoLLwKVq#098^ypAXZQ^PJy}8v!?&^AWf^
zEN~{nDzO$p#;<z~F;h$^H$H~wkZi#nPxgs@6>Zi6+o>wh&#dcrMaA~vsS}Oz6y{7Q
zOqGCh2LH*AI}COolkYUcXQTnsoi*ij+PhphUklE*Vh?a;!jfGgo$*Q97*$J1Kb}L=
z;6QXL)nm50{4_ZDo!1<wU<~$zl(j3(VgcesB@|%}#+l-1>HxGKhY)>vR&saI=9Yl(
zieV#cszBKIgS|$vS_)WzQeb2#0;ibW0kFJWMEAL5(Huz**|BOTHO9|4WebpWL@JIR
z@rwmzyI)7EG;yUy{L6Z>if5+&f<zvqS&-8c9O}cLFso64>OFiBQNtFtD0ZbZ>y&=G
z;sGh#IsB=T@fVQL4|4*OqEFweps8~szuGx8KKa)7RQS%ScSq;P+r_T5Gcv!Spq=L~
zRZ*#fH1&2LJoCO)Chmlsfz@6(48rw8p{!vf0!2o$qRJ;b!jx|-#@p$1LqnUGjhJV0
zlqlPcAMvz@Rnls(r(Vmnbb!!xY^Ui7Bwl%oy@?7&PX0y(cRhH<_Intc=z+E^)U>Ba
z+PtFCX?G^PKXKaiD;&EkG=N0L(sX5I3rYU8+Gg3C(HTeUhqC3QpsfjDEE4n3;LbhG
zVBWviXFd&-Bq!lyVh}f(a&z@$`x8n0UMe_+-8M{?dEr!r7}5p;%l&uY2Yr#K;r|MN
zbl-ZBsKM3`TxFkFJ4+cFuM8`4)i_pLBr(4{a=Y{k)qaPa>shJzt7xKOUa&LEBh;+a
z^l(~xsbOFC+6~f;kl4ttQkL`NztG1A6}svDK<YVs5@*NF3AWqeu@?I6L(=+)De(wE
zMH*KP#87JIud6}osmdnKxMw)l-mxbdfFr=Ni}LL|D#w&=osKsa*cee+YQH@*U4o{h
zH<RmSJmOqE4|{oOwRVrgkoS75=_*sBVRFM&+2)yX6UeXl?=XMH0AK~>J2C)L;2>4&
z=ru^S8ZziDHuJ59pZMu-Ict7KpW{j_rsi0WP&1t#HODonFjsLv-dYE&1KXT27^=Y(
zeu48Mwa9wcR^Z+P@@FMNen?3LiU8SI%Ab_p*Z{^yR1-i(B~{b7<0E>Tr<<(q0Hk}c
zXr-;hBfYnPa}JDaMuX|#$w-8tjX~gG@U71(v{q>)+2AR0T7q=*ei#wYA^_AeL%1$J
zXE6d2FHA94zKv-PqJ_%#IcD9kUIt8Ww1x5L+0KJ$wu;jrbCqo`RzyU+fwx??i#r9#
z5ly@OJ}au}=>UV8VABFaB1;a100M3Fq92gee`kVFuo-A_;_dhY$Z(`U2`+tZR@H^E
z+yd<#eKnNTy@1MSN@YkMr@Hca<mV2sGRYt7FaPwisjr=gvm*?_{0(|_cLL#zK>0?G
z)y(;wO?veT@$=*m?Kp>6`a_R4f-$5S5<oy=W0rxw<JEqNNYMSTl^T4N>TSa?p0<CL
zhP<Gk8@J0Sh=5R-kAPtV^l6TCg`tHn0RNw-{$&9EpE4u;V6QPXSWi`ra30^VRM}!b
z7W3EPU_8Uw<Js;!PXWdixip8nw+1ZOh~4ESD?dpv8+YB8@9?uTF<S4vj!v6}{m8L`
zVM8EMhn2;~QOJ1e=6wTakhz->TcdaCEykG{uu8SFn5l(&0D-au$V9x9d=(R~hS-u|
zTMze71caU(_^_AZV?r4GwCx?|ou8Ne)6_HznSBzqNPf-&8U$JcGk8hL^Sa2aN-O(z
z(g@pz@)S$ntYw2SGuombFt3vYEfSH0x-m(2zLo9|5EOYYv0<BrL^7bNUpw19a}@A3
zIbc!)>WB3Ro%Q6{osBo4!t6TYF3|Xv$e2<aYs&AM&VUu`t`e-y=T>r;+v64E=rGV!
zc7eY$BUs`&mKsLlaq<Byp_zF@vnR}6abI+h3~8WU&*pq;%i&7h(sxdz7Bytlf=CFT
zV_;N!#;&l4q2iviEf{N&FVf^vk$HOz92qKPlK6pYprw|SLYTCwe9mXY1=qCqm2@90
zaZ_`_c-z~G!Zkl+dX?t0&t)Jc`C_68%%<LhKnAyf?7Q^B=}H<$YO+Ha3s1vwfa7QS
zJ1pJv(@ub+_c0-=3mCD_j~blVJTo6Psc)*ZnpQuy#<GIl>s@wE7Y(gwqS4#`=l_>_
zzeM75KK($AR9qa~TQSK-I$HmOf6Df^?xppB1=p5G(oIg&SHc;J!K0Y?8Q0uSg36%}
z=>tX!4l;<8?-M{*MU-&r$Y(tDxAxfs1B9^Cow*iu2RiWrmw?o8)!h#az8L^{OX#jO
z6?<O4yWcYXS#=s|Kv^w^*mZ-6fi9}MV<4i9?Cy^g0QM!w%Fua0U2oj<4&Wj}bKrd*
z@TW8ap50Q$(=Fx#xZ{|Td|Ee3Ln;bsF6UP-53%LEv%k0$uQSSwZQK(l!^IEbOrQ%M
zenG6zwH4A{A9Xu~dn<b<gk9i#gJturTR!=CF*+T;_Vrf$PD;U;LOs$zf(nL8L?ULH
zGNAwHv%!eRi;>E{zMJN--IOY1{uCuSH|I6_-DO}bHM%7R26=`;aYQ-*-!R^N5V8V4
zReVrppN<*3(eK@#U=&oOm>X3Uguq;o7lJ?TdVa7<4j_5f)OZI&@g78CzuR3gbNHNn
zWnUH?4fTmQtWp@p$0VCW{5JZ(D0E3LqsH{42{f_KlDO?W`{L?7Z2&Ahn1OJa!MuDN
zPR5FvNW;hp?Ertf6e4~>q*X(^Gk^-R{NU(&z>15;R1-)A?)*3bVAg6C>MDN9+DLwO
z1=@%E_10uOtZO3%`BBsOhbO1Rmm?}VuL!bMiDb)(7N+qYu^GDMw;<1*PM8;u_)Y)>
zTb(Wbs7@`;5xZH_w?wPL^yLyaht*V$WyXj2MB-S`*gQ`b1`{4^g+l{Q2C(B_IM@Bb
zfX{VwWePyKrsh1QoLo<q&I)4!^=XDCH8RvM+<aP}fTp38fZ!+u3Uj?TV*6gYR%+0z
zy5C;^IN^i?%b&UqRNNS5qo^A@9m+P`%uhEQQ`y(VI9(a;vHqlHcI2VXywV$cM8bG=
zS$)Cno@e2*f+{y5J}h}KZ>PV)uJ+n9K!k|711B50rmuc{xBT0L$RRkzF4as3hfAEI
z9gDaBK&O^hSo*PsQ|7i49mc{FfPR0#N(B|4!Ec=v1M%#H{c!&I_~nR6(m^k|4fw0d
zhNv=w9X73z>7?Ew<QQ_zEd1<#x`?kje>&}TQA5f{f0EXP77I7cYsgFkPsi;8wk(|<
zUhK)~oYnhiD?gLrw*}Ij-TB9JydF(UaT=_M^+0LqPf|z1{YLy;l|(2ikrU%*HI9Jc
z2ccvz9pR0BMlGH2zUu>kpIJlkmph?>t+N?mN0@ms3W%=gY7o?{)n^#Luq*oselF?0
zAIN<2iHO7*a4VyV5rSHfF-`r+e6W9%8<@al`_cdq$GFmcfWP`TV8HtBCcG}an~YGP
z-JlMvjHo-YaXF(v$(e#cjJG4@Bs;oJqddluk22u>>K+$dlm#fJQ121P_I~&@GwZs|
zmkU72Zz1-zjcq+HdWTck9<%6ElR$Mt3aZ=@Bhy$T)}4^I=Sm=}OQ%so&{^jc%1czc
z??1;4K?jmVm}C3i^Lqlke$33R-2HKrv{>L~OdvV_NltL1-S`Qein_I}_}n-rvsK#U
zv@y8(y^<$>DB;dFphuGmBjiGX&nQEiTU60@e%`ZF@Mtu@*Sy;3>o)i&ixNBixYdLD
zu7r4j?_M$m8z7vJuphs7C0bB_D-9|~eM<!|z_F+|7Jr}qp`9ae>yH-Tk=s;WGfRI<
z`IDLSwXtIBPiVQkEW*ldJjFy^-)MQPb%}g5ttbm2{k7{AhwH5T^v{;fLsNG2Qbf2J
zC7|MYuhHAzemg-Up%@r`DXPC6)CJs4r4zt^ruS@_CJ*KS6#R{y3b}Jn7cP<D_&un&
z)<-QXF%gzY*DQc0Klh)IQ%-|+luejvMPx!M{hUWT0ckJ$gD=|&=qDQ9-RD&8SM3DW
zrUql@bz*%<Y;5eR`ObfFT?#nxxPman!h8T#eR}h@Qryb<CCRhE;<~TwNgsABP=cGw
zLxZrDW17qD9C%r3!4naS=g|1e!3Of(wkqdca9*2;r|vPr5fZTFJYT^)eYNr0TXA#g
zWyUju+#%}-Lq?1w{P_cK_p&{|TegcUhqZlvJ>{N6{EzcRdE?`7YQ-<}Go_O!GX*Oj
zMzL)hMa_Bg3}<V9BK1t;CG)hpe17C{9=nlnTuR;2^)e?g#(eA<(wO;U2tZ_?kTl<)
zzbA8F;P9N7>tum>oZTer!v3dMp+jGLaH;avS;-=u$}{G<g>+DNdoD8;Ht^|ql1(nV
z>C)zsFK4PxQ9OUh8dY!9BURNSPoYp{L6%tk1=SvQx=?a)+A~nLus?2llC6|w@YNN*
z`e8+C=)hjg6Y3?a1_h*%Jh~s@r6!$DG+4LFB0FgJeJa-^6Gy8hukuDHHN1VIxHOJf
z=SVIaUoV5Cf>x}>z$1`GC+0bB>c?Q#!~$%MUzC%9V3T}zUc#yUYsg~m550xACqUt6
zuI3H0Q=U6QSU~^~97gx1@~Eh2e-W#Nylr@a84Kp!LSL<vAIAZ8rXUoL(+XowAR^rB
z)Ktw^X4rh|==LKteDTXRpuFFme8>Oz&x1$#wkIE&t$`*Ah+?S=QPb~4-QhoKzplci
z=knrfg{ydb1PL3bX~9$mBbe;g@ZCVYh5Ah%OqEC%6GJ5!C4Rn;?)zDIG{UcC&V7k2
z9d8~7qF?C|$IvQC(glClbpu2G%*jTRRmo}t0Ox>P_1kd+;Jr*t3Ru)k#$)m1ALm;Z
zpONs5B54>;#Fzh+X?`XLY~w}Qs#>#(ek-3Zn$LVUzJLXa5kL`Ys{eI&q2nI@dgSys
z!13*G7Flw3fMkoley=@n1rtH+#N>@2>1_dv3Q#IPw6yePO#v+8w1Oi0>gS-&xF8cB
zAOq`$MzSRWocd1zK5GPgohB~LE?q90!19FXh)8^b&hMCFN=4To_nF%j0%Y-(JCD=O
z`vPD!7`;qp?N9RQcx$GCXa3Z`xqO$XdDE^=rMbM@&RLh}i<>a+4)8l-mhljsVp8dY
zdEUV$)fS5Ci(OA!gwviCP8|(E)gjd<FbZSK<JdvLIpHf&{=N$D``n_yx6gP(bKr~R
zcTCuwmATpmVED$SR|z}<d{9w$V7G$~0gBJ=ZtQ?1Mb|}q`Sv8__%j$`&T2rOZ!%z*
z0L$vAq81uWs&<v?95yQoCZ4~;E=38LTpujBrD@=V)?s{K6uvz;qysor#~)R{1>zF(
z=QdBa4xthDQSEfsJN?#xZ#d$`@OZD?%UG|YL8KeZrT6Cg8Ico52MYo_V^17IvKvTI
zK7DluXj5aai<_zNjZ2&B2)2U091{u5&Rv9XaOzG)>;|BdSOxO#9Df89?yKbYf*xoV
z1^&*BmXqw3cDxsxEk7@D>^YrrOOoLKqx*D+be}p`7B@=JlwV4w)u}mWTqys}V1uYb
zh=nxJze49{va;c1{00&QKC!M6{mp|?V7T6$GIsW_&?r1#h~g6K_k@d2qIGRN6;f{G
z{2G62%-tn0rgo013nON8YyNb<%zoL^XhCDaev<XRm)zDkWo6of$A0VV(pX`rJgG^v
z)W=(KI@xluscC%<tn7hRD*RI#@7T;-B-oWXNe{Qq^&%}&E|SVlELgP>MpY`d3lO_S
z;WKuhH*5u;;JX)(7RF;+4vUm6)*g_>tleTd{%)<YxYAYJ<sB<!FSjq{<o`ItaoC4#
zO}w36Ev6PF$Z0m}d70j~{MB|seU}ld28vl693ey7@>_ze*ZX5KWv{wLg!?7`;iYZe
z=qL6zWCQR!3@eahPE84=<PI$<`P{;sc~K43ydsX?*fzT3APo>%55GlU11s~~85UwB
z{f%+rFe2XJGUQ1t%mp9`2+TkH3C}ModPM8iil&Ct{Nm(R;@D$z!}nk#YF~(60h0F@
z_x6C9aQ?HL-ee9rBku1{7AVSLpxJ*jgFJ_CPe#Ybypv1pBBaBtLy|XDPrFjR6XHQN
zbcmh5p8PU}%;_1a&+uZ~oS~&~oz-YHnEShJL3aPyI|V+-GG0a#lpJYuG+@{W5+C+J
zL)Yqsg|<#vULqqvJ~|YVAPOaNmAW)_%SRt4Xada`8xEq5QtoikoxI<%kRH5wb1W_h
zT<9Muwyw4L;se47zfqR;!j-k;^%tAQwJbmHOgRURaMWgfLXSw;%oHa?`D6*$lhHfB
zhB~5`%Xn^7th2{`66;M9(3xE9$>;Ywz6MD#2J&?($&KlrJ94Qo6SikT$#*&MKu6gB
z#PNf;5_sfB=wqzz&V;Vq%3t^`+O#BpGSad}q^#tkY{^(p*zSkYi4ma0fwIFh^E160
zj$F2bkQ3eWS`gRn0yPM*;=ig`!Fgy}fU|T_9F(w8U_j6WX`qX3A)QklY2c-A;dg95
z6N4AnFdXVgStfAr$9H>wLhz;5j78{gg@rgNJL0FVJx$NA2sGmx?GfIW?(?4th&Uu2
z7*n=;ydsxswuatPYDjuVoP}Fx(ghQ{3RcVT-Ex}#<SDgOG+im(5k>aaR);Z#kB*1`
zTy-?mQ{t2m+cyJM@4<Y@=7MM0p}$12>0bHB2aXx{S$FbG>qaGJBeuBI&{i82cl0&-
z<#xuQtj@UANPN-@3|Kz@SB38B(ZQ<sL_hZu$vr!l9cN=U?Ft3z_oAHrWVJHoSI#M<
zVD-u$MlXq?F5(ieaW2ARR_%d+*Co=V4qRwju)Ef&Jz;0=35C$Xbm0T94R5IQj2r9}
zf31k=kPhSnD$Ohpgj}w6TF^7Wv__GDgV^Zfi)236s<_yv7cb`U0_hA<0wuMHfDS(X
zrlvy*ZF!u$2n`~h`da)D4FI>>g140RuKzK<+CBzB@pJ6##KVhsV!Bgc2$)uF^{pBK
zqPyyG)b81i8N-W0>W9GOC5nDbp`ggxc)bHS+cBI3pu}0BWt$UBys2w!J9EMO@s9o-
zI^RJ=+_RTUFK{|YtDlLu0E;^<0cyBoWPQBFdC-aSF<@m#HnsE!Zn&-GX20RkEma^{
zejri~rboNMbd}89I2>>Fz@|3Rz)yAxMP9K5XYBQ?czD?KPD)Xf-e`|Q>abrWvN<Dc
ztYH~rilqJ;*aF&3REEernIcx+zm@o@bnq6t;cGi5{2p*eU#V3BoycmI6y}Ohrzn+;
zg>R+{P!uRl><W9i13#gPKB}k6ykmNHg@Di$d9wmg(HH`&PtP)w{f|A<qk%9b-!e|u
z!Rd5{WD2d-!Qz2e!H`3kX)D3yvtAO=b&n0$%^4tDWce4&X%~<<d@rT{bF00sLAX92
zNML|&xfcoe4+W<q1MSj&&Qul9(?vq=G9A#=fRBJ`ThB94n9F0sQU~<^_Z!aHw-;iP
zn^i0mS#MN?fd74$sOc5H&p;ag=A>70*md|~u`)}>RTCOXj1oC4TG;%NfF@G>%K_p`
z;VXW7dE0oDK6WpZnr!r~_<hgRIBB(V$*|_<_V@H+CL4a(sj-=re0%U8&?{YWCg*FS
z09uznM@JXnELjSv-Rf6l39%$_x{}5?=`d({-F4Fn-1NLid+dJf<ZXMOJ@p@9Y#@bK
zzQADi|A<-~DS7-xs>dyr&+}QAgPNpzA^_2r4z6aGE)qSjmVKmqn=Ty7Hgt~M@1iqc
zIR%f7Ce|@L`bdlJ+BQWu168&Jvo11Kp85>tbxARB)GT%7B#LU(;Fqa#cmAbQ`}f_s
zEdWnYbS5nwygiw;s*`~Xj-esb*HqCvgb4yrVF+#ig|MYb1tgMlq3_R($uG7=MSbG(
zc0$ku-v}y;wvL2iM+kU&6r#BWrqHPrRg}7e8|xEZ-hyr@p2BXK6a!knh_@QI`f}z~
zac&=;ffG}ci2)5}&cgbWPcvXaVrMX8<Il%88R1|6OvPGNjR0MtY+W3f{vyWoF|6me
zV3d9$7Z|q_Vnx+*lNLtKyv<RYH%@jJh!D%a+>eQ-c{L&sYt5$b+<;(Q_IuF!+cI4z
z?z`T2HnZW9)^{td0Gd<B;%mS-vjUb(mIl~$`tdKtb}xGYPg?2vK6)h;%ntSdJBJKF
zyy{bEFt?QNvf9v1sUa2xhDQUDBoQ-E0XTySIW>jY7A-@2*su3+H_jV785dfyqVYDh
zab@)u=v`~$#vcr%ec6V$($naJw-_yXvnqcI5bq^ZJ|X&2WeDvXF-D$S@XZMBd!y}n
zX)ktM`{V5Pwl<*kOAh(S)?T~v<+Tsw_ZgRb7zGg45*W~z3+{kac#dn37seZSeJ+aE
zI9#Jv>}Ihtb+?K3%-oq=29!gEE9Qw3G_1BN=M024cHO>%?XgQCqj{RD|2qG$L(<An
z&JsL`q}PoHWCACS!X{kOQo5WMJ}iA;D(^V}AblV6YV;8=%+m1km~^nDL(<sa^G-L|
zh4**L9~3=ygZatgsUB}BCXG&o;i{~=9;qE!=;**}?C)l0T1@g#2)Bp4+7$GiK3R4e
zXLl79nVe&$T^vX0T)SxM?Z@j+9bu-n_BdwXDz(#?IPGlLEZuikcUY#cu?-jwmLBN5
zI$y7#$ueM=u1!_k-esJ&Xjlomq=8qakG$+@iofh>9HQ72s$7=}%3{bcZ84`5`_POz
zo_)>zMsPF)AD0XT-tm!J{g<Xp?Rru%_M)LOTykVb$xJi(y)%N$lgcr3_tVu_PkzO2
zh1wKWH@%R>C(h14A1Ej+6$`^x@{?#b2%%klVwHIv@JX{ol<3Pvr=4?R=f-$$?u4@x
ze3FR)0!d3-)AdbhC$n9z$5s8LFMN_xxV<FA@7d){o$0eALY;aEJDR~o7aIAdNUIdZ
zMLos@;m_7joic#NN&lxNkE3#!g(l;9{F69=(Q69OxaC}i)!d$;OqzMmqawRClWW6Y
zW0FnD!iWr1kMXQ(csG%7Udaiq<a0m|hGAr;^NMUG8!<`VBUkdp*dOTO(J!6g#Kavj
zV3CNsfX7fsGS&^)Fb8m)?D_XG<T<Xdocx{sipZs1vbffPC=n^p7!r-<r`N5j0Lsah
zKyw+dei{IrK#Wgz%kXX|5l?|P+eg(5`Xn|QyhO^e;%gH-(%7y$FNxJcBDLxP*B>T;
z3+d;WoG5wI`NanyBW0>u)gT^PP|s<R(AstPFb`tG?p(?)-k>+Jhj*Ik4K)oJTMzSX
zAp<dJlY`8=z6S@#__20xdvXrsD*IJ68Jo`=!BQcHp~s_I`7N?p(!=*kkp?(1sFl)e
ze9pc4Y|f50=`e7oTZc_G0|!8x^wG#w88hurgarV)<y3o)U=Abp$C{#(s9fohuJ9w?
zcu(MMPz8Hy!w>T-#m9tI*2lPPDmZw;%^!ef$ZXpgn4eYj-RAquHdLg=;|!)W?hz24
zl6mI$!`D~KCb3Frdz>0&t`mJ^VIF>DZVR}k={qC2w+3>+4wf<Y7caB|eDa?yyB_Q|
zo^M41GYQVMbkH!#rhE52-GRM2hIt*&0@W-MJN36or@MC8$00k8W1t7^1On_>_H~4*
zYSNFwQB^MJ?NbS0?#C9+F~k#S{wDkh9LDf^#p+FFfss?~_IRF6pmF^jccAZ&mrmtY
zn(;WdTx!C<7Y<)47UvMyPdi^riyZe~^aME0Fj8$`LRmEjl!?-FYDTb<kUIe1Bof>|
z#RhNP3q0~PH;_Ba7ZL{{&xqRk@rQr1Fz`2e9ko<5RY;Vne%wJ}cSQK=j_~qAA&G1>
zDwfWXeMzmQPJNC<y?ICIJJU&NQNP|=&|jQxrmy<DKN<a$EBMjxA0oGusdA|}RF_90
zQD|4w;<|os>aDx<)i|ywy+0Q?{!puIPnVO-#pRFy9n2L?R1z(`3PU^z%w@QwNp#_P
zox1T_q*-F_VXIzXM;)FU)jCy3Rqwpk2L+YfH^+A{%Q%_YiXVIvkeM1G2vnFG7k?V?
zzHxt_tITH4sl|i5@mD*>77~#uw>e2CfJX2j>XQ4@toBp*Vp=VdIBCFdE1}X5t2GQd
z0?WM2ErkK8Qu89@RSppk@r$O$>#*W61E;0lEkdO!3xE=n1|j}vQA1$qT!z<!8^5?c
zkHFvG7SIk@osqJ0VB$0wA3nW6B*5?Gp8{8k95n`ZLwsA5cI(wHH<sBXp9Ot3Qo<S8
zl`jjg%RvskTA9fD0OUOb?8#6u?*c!%9EtKVO}o&eZJ_$mwMgg0WHNEso+;F_I|<;z
zA|@B#O%OPl6(RfQ+1(JwTOBm|hL&$X-HG(vU+yad%vd39CuhVeuwS=YKbt?CW4&MM
zOkAA(v#l-c1>q+S572W@u!!#L-y7waBCydB$r5V_W&g-QyB`M)6M%f_g>=gt;CM+p
zh=a(<hO7XoboDnM$&0NTe21lnDL)H1O~bZ#IFJRNxuwN;8KsC@bGZz-shQdhE)z*1
zsJ6gV@Z)_U3%7qxx4<%}A@=)-Ata#-YVe6E)?l|n&C!IpI~GYg^9y!hWdb93inkNu
zw|*y?M{f%bjaNVywb>O7R=QIRSk;%#qSb5|=tv7%+be0FqlPjCExMRjy+JE77KCpd
z`7~KRh3#5J07M4eMOV`=Z-+34>~+I}Gll^|EB0<{?OV=Twe1wNBY;j!oV|5qylC-9
z@itK!4mx}=u&euUdt2Mv<cWokpR3T09auv`IEK3g|A1~w7=Rr%^)I8N6E;B8xD14n
z-+z*yG!P!s)~nRGAAPpFRYckPaC;nJ9jN08h{rI%D46<ytK95*1STvXd%7Y0uqL8Q
zkhC!eZhPC`9Uxf6ai!k(TltHNLQ|yW9E57f&?^%GrW~MQ8-&?O3=so+4rDezlZAvA
zP7eG!|7-T8(&gwWpRp6j_$H%3-4(a0;7(NV(qWZv){TZv=K;oGRb&Vf*_+>Ns_yt~
zndMf58Tw%qea(JGte;@L`&K}06dTeiVv9rO0$rWJlcfpAC~39pNZrW3Gk`7NOHPO`
zTpF~ft9@&V&&&yF6drAK`Q8&%r?~3*{c-2$9BID5H!;wn0q(Xx<!I=%Nw%kUOf*@t
zUe<4gs#q4+@BteCMl1fx$fkgZWThz)Enjw4H_W;*B9;+tIF(NOX|aH*Z;XY*g24WI
z^F0VsEpXkW>^Ti`Og3+qj*5{~67CiyUvPQ-7L4BF)8hV`pZfQ!ch3D1a$d1Y1L#G9
zIRVlwgzq@81uL(Or+*^XO>azr2oPwA`=)5o;AiH3A<(a*xs~BW=D3s}?aG0XD}oN+
z=JNeh$vJ3>8(#|q#`K=#LNIgk>$qCkZ`A>H1uvJl<m?;s&iDOk{FLT8M3i-nNh|BR
z$RkCJ_V6gNg!hmGD(p-Tkxs9=#2!tI#=cZgFuE?=HeCJ8-|b#_et<z^-y#!5X3$<?
z2j%q<8{2{Y2$wg0dYD`l#+md4!zvQoXo@*w*8-!(nBr?L0ou-W;~tFGBcf(UYmod=
z(tLg!tAnk^^|<-8yLV5e@Ck;q9WG6r16rBL^>5Ot#U%w{U}?fW!fcu(9w^tcMv`52
zKW?0!A)nfuly|$gy9%}7wIeO7?bsasBq=i+#^5TMhtK3Pj(0vu*wmK8*{;>0b##6#
z5++b;9^f{)udT5!gC8UN%-GREIDF+VY4-afdqJu?*MS%|=zj>70ZbvKAado?U#}u?
zF!v!33X$N)9YC!TL4VW3$;Q<XAKJ(}g>l>Vz|rs$kbm5<;BrSn1~<Mx%tREbNPq4b
zn?vR4&8JlEvuvCR-`wf80=3a59k~T!scdeAlS$Shjn<RJH$2VdSkBerZ`<XMSHDCt
zWD2U+bt>2I_SvXY&HP9g$}W=6TyKe@nG{G&OsxpwmSVHcGz#vIqPAAYGP&k`l&|CR
z?1)i|(EEj))7ARUGc~q(E}2itT<_SdXaRPgVdT{R@_mG1H%vIpa+)3jK9&w5nER+M
z(=?So*iLu?{H^GE<0;K}L8iDr$}7uM{t~&KGj~8iRGk%eT>J+fbk`#$u2v+#hYfyc
zwl9w2mJ=v8A`@^u2uwNO&mp{AG66l@rwD6ge|;cdRleJ%|MxfW`x}hs(A`8xmVxf3
zpbwa$P5E(ac?Y`we?B)(P8o=7@Idnzq5JFCT|z^|)AC`5kf#*k0NivN>3VFK0<BVX
zkH*X+n%I1aB7MpV^tQJGF7d;PHLs5fQm)ryK)%+9P_yi)507Dax_Ss{tBnqS;@uT8
zi9L5QC2=>Al88jXXF=po;&;4|<EnS4v`FSrK)bH!eq)PUprq1Wxk-AOKN3;k%qX?)
z`W_~1?h!M~pA>g=tLEKKfyy#yVvX7&u51tDLnZN~PQ>;(R|go9my!dvmAOf5UO@#i
zy*h7UecoM+ZKgJ_V4Wdp$!VwZ48=TJwf)69?6K>hTM6L_<R_dS$MO3WT#143;F^Wd
zc8L7JaDVuFrP2?$G?%aKS0wYp_1n|&l8k>EjjW%1bKyzhS($W=LdK@#&<nt-5F%V=
zdPBA+5b$+vyy8xIBEnnE?r!;Rk-_fygr`4iy-|DJ-oB@H_~%?|ULMQIU{$WJ(T&BF
z<!BS}8QUe{+WM7bMz*=}E=){w!=F`=&!?W)HE`W2bsvh6D0APUrxp%e<N8}a{5h*%
zhZX~3l5%sa(62K_0!KzCsE_%VAcnpPgo?EVQt=JoMeq;e&zmAH5zqfWX2_v)&Sbpt
zjjPHup66Oo*>CV=SgY(v_s5FvrF9Kh%avfKh5D`#rS-+ByTc;gmS}wqKOt>(oU(X!
zWmW6M7m_rR3^rUS<s-K)3RSL2_)9NeicwW->3HNFRtt3$$(JgqbR_^5J{8RBMBh7h
z6hFVlYp4wxUgZ!k(~1+=Z2G`RFiXfdX*^IOqLky{p39?~v-nt*Gh+k5A3(k_(Dg;B
zhhhD(qq$h_=p|EE3ZIy|8r>?{jKMo^eUw1Ct^1iME6)+?0e{wLmu!)G*eIAGmjR57
z9zBmX_$vtCdFN#ox8j5gQ()8{zWc3mL3F%bxo%sK#%|w|Y~b7XyKCpv#Avxva*fl6
zZP_5W=}ubC`}#A6#p;J9`KQ$VYENea&klT7-U>X4hU&|glU8!Jj~FNU;KxJPk4R|8
zk+G%MpRn6bu-ZpiB=V08Vnis_{pn;^(kALO;!u7PMKX;eFaM5};><>u)O4H=+l>vJ
zsCh~*;KvV_c)y(6gK3;CLUT+#&%bZ;_cuT8gN&4d`0Kp2!K~W`j>PC+E#FPV1EY=2
zc4vDg>(CGVh`A{bWLMAz%7M2R*HIP%R3Ci8gps`*UFA^id2ddMl-&Og{POCpz($&G
z@gel2!!x8)PMuT9ctexug~hGZvNyVEf#H>OZ`NpIxo4@>_7;_3pK|dSJH2tFmcEbn
zB^_8$@uq(ZMCpS|96aFmx>8y_du^I!fytmywFbXwf5u0hsZl9|g|mZsI*Oj%<sZZ-
z#+r!0WY6ROA4Ps?%uXz|3^!iJPD=kR-TmKT3ryBZ2(tB*!*)VKBSssUg&<6M{X413
zsHa21;$uaM45E>ZI)@p8W_TZ&4+ZQ_-_}W(ng66TF^hvmjdYY|mpuk{Ms>ckdt#Q8
zd^DNzY>=Q98-$uC`7hL>Bi?e{Qqa>=aT;}H6m?jHP&Cj~SMl5}te>G$6HT^o{=)0n
z_04u<|2nKwo@mOs-Z72mD4TG+;u%_j>Lbup;9_l?^lojKzT2hvACC{bza2>aJBwL<
zGZrlV5In|cs$s&%|EV+nM=-u0cXMEfJT*hEx?tR!=XT_#f)jgKkis1P?Ik5)QK>m@
z)Zwxfb!lWn;>rC(yuF1T6I7|SLmahiTU<KKs36+<+rteDL*a|rPq4z-{%8SIM13zZ
zy86Lfi<<3H72m7f-NUpdi*JVJbaOSMO>p9g8WRBrgS0KS?}hEth#$pQ*FUfJ%u^d`
z_FK234(pOL2<Vx4jx4k$<D1~Z2G{ENTa8D8MKUr8d#^_g@~>3H%HPRp3ffz-^aCnK
zKgpM@w+YL-hW?{{aeJN~--m~~b-FC2cG~O}Sk+>r$Z<LBdVTgggF-dUsKbBYA@?OO
z=_cf@ZCx&HY~X9+KkTN~MjIuc`_52`hLkH9H-6(3839(xV%%lMZ3U!te3+L2g)v+T
zzOxj5XinJnL1YM5K7X4ficPMuVn1{_46m8xA<g76TJF&c#;z!yaQQUGD|cDh?b+<U
zhSaoy*z_g~MeczfA#kRYGJlOi{$9x$P^@e)$c^4`eqGYwy)lcW8T=j82d6Gh0S+3A
zCW#Of>4MDQok|1g>9VV0hvXbr-ly1y={pu7UAa4&my!JPTUD}AOm@o=-H}Aco$ZI}
z#<>8_i8;cwg?(G9=v8@6piS;)HWJ;-W%v9I(xWFCw*m)1fv5AfhwOZ)pOJr%-{Wfi
zs|St8=AnFEn~XYp#mKdXcGyKdf1y((TxccwemUM_><LA)yc8}Mkt6yG%_GKh?em0(
z?u?^3K`~}C>U(hw&5Apt=NQ*}%9a_%#U5k^Xn6NR9$`lkap$~$8CU+J_Ta9sK=s9=
z*~3*_vb}a!X{N3AYZz&S5E!Au$mle4x0uA0+2gM;^J8aZVvI<w|IXG2(&Sncjd|5q
zyC>Qa)5!esO62f#g)F6-fxvWBEWrIQ%pfjzn^p~0`cgi&2wf`6;+mH+FR(y6B>&C2
zgM9i}G{m7+td)EiI;tr*-2C(sQ^vO^CvTf5Z(C>XP^}@|`O81{4-gx)yQ;rjwz@Xs
zZ1PByipo%$4C=?pmybVCS`L~<8K6z9-n3;@tZN)5da7uPO`qbBHw*3n?c@<!?o?(S
zRm3D$$+Hn1cy!Zk$0m1Htz?oxPGHjfFlHR?KQHyBs^<iTPUH9-uwV8u7!p9*Y66V4
z{#5+noBtc(N>9*EN|Tv=F30PHMOx)Y8Ty7A+5)>q<Hs$JhJFf?7EK@H)4Vn+#9$iP
zXmM9GJ!0bHf33}8oIt3$`}SM!*Yl^^l=;?~uC)?=VaE@R{YJ51cJcHKp`!J)HIgxs
z>yv)vCPqWU5@kkbiY-?UjqH$9-ald#F-(_|iA~*K>0XQX_peAsVDj^o|6n(UU~}3_
z60o@9NM)8d6B4E$nk;!p>@KA5TjrU@`DC5C@explq_Y<5S7HEvt5B5lKh`-E7|3@E
z%6JKkm}YLWW=vD^!cU`kcGq}u)j2Nu@;ACOPBgp=i8X6)!@HL_X7o=QP1tqFI1KKT
z0Gu`(2Rs9JKjEjpS?X5I<8n{pNS$n%*g}?T^XY`(```iICYGc8<Eikwh4X$4${50X
zofbeD1UO4yQ1nwEvIYPXt(y!8eT~=w#SpbyAeSIeq~8U&6maPBk=;BX!3AzO;m@4V
z?@Ks^!D;1(m*w*vB*bjLmJ@d{sndU)vEt7d5;;Z5h>rj51Diniu^ev!k2Lw*=|y4_
z#GAiLOjQiC-Q*Up+zJmcl28~})~M<ulHdHG9YyH~3tT?%HNiGJnO9{HR^8E~6#r=y
zh0IT~F!Kw;_Ma<hx#tDZKcUuP(Ux|`{wY}7Ol<e7HUoEZ(j8_^D6MPu$SkhDmng>_
zY~w^?Y8Vx@N#X;E;eEuT8)p@$3?RitR5fyLFIOBI%MzbqR`GpvJ;2b^>F8d2dX|zy
zeE?~L3Lc)O0GU4TZE8Q0L3VB^^=T-eZf93sjp!ha=3tf4#H&$$f4aUMO}(_ZMYMO0
z9&9+!d<wdtmCP5ZaIdn}8Z^>%nN6I~mV?&|6oZp9M1}|`GUa2Xw%75QAGcwCJRMEE
zrWR=8dzM=t4aG^{hXa3$tv|=~Ka8pGL&n4+;ZHmM+J{2N1a)=PQW;#b;Fv_f%L2#r
zv2Qg2AKh;y0r9!I7?|-34jK^~cOjYH);VgRch9OW^f$XS_ll{>Y%uIhVA|v**L!d;
z#Mh+m<Cl>ZomjwpA+E0P0)P3_eYvhKKAzb27rz(<2wj&ZTM9kXS-KL-(9bT4Vv>rh
z*z>d+r=R6@tJZTpgb$5~>a&tQzdZPIsotJn5u%ryTV^oRQD(gHbedZpv^|izxc_6@
zkcSxV7Udlmu@IISgFHHmG;SJ#*@ujiEB*5CXpmxSPT8RuV>0HtNQxrOS18ySJA`+h
z>EivZ=Jb=il;T#-2GF9T`qcdq622e20+FrEb<RzvT^UlPoge6S#mSRqT#*xub${$2
z{%Xl`QFHaXhb25LwT%zR_%@9Wi8ijCsmHJD3LbpZ3A-wLd2XGVR2`Lc{SYH65f9%{
zrRl(DL#OsT-jgq2{t+0aci#v?$!OORt8#OF`m*{jY{UViK$-0^OCB&*90_XM(1_{H
z?w$w1({=|l3~~W&ZBLvUI3M;AF(%|6pzk+{M6XkN0wG+vxJ_3KU)|~}2M#@jRpIw2
z<vf!g+RFZYg1_Ek1V1?8Y;?ye&{PWE4iX2+@{wq)0dN@oeM>;H6oslHrZqwqkSrhF
zbSPYy%Qn(s-e63?&#$?>Awd;!V5Ii&zS~Z6gDx4MGvA4u<yoZgeqkLd(MCUDelQ6N
zdj+GFi<+{t%ny=pMR0WjTe*U<rRY3&FPz7@_nVRk^Id+9ytM<Tf3QX+pFK+o{*<c<
zssa%&<LE0nPzQ1qUNmku#kwug`6KiimkHNs@JND3F0s79b(L#y!%ocx>b2WRGbOxM
zKCzaia7)RiZr%P>#r!j}ifOk51VWnP-yzN1&Byxzxh@dsYSy9211(T|SwCFZu;O@~
zsdvGHiP`wUpOzR#hg<ibr0eg)?2AL=K;cn3-9hJi5V=?F7>AO0HNJ>Q@SI^|l&P=?
zBC4iu2~&!FX?}IxCzI^Jv+C!j0T?g&8Y#yvK$42@N2DAxsudWap&`sDUz+nf`!6!6
zT%|eo{fA_7%F_Z1HprjmJ4)qytT!ZI+n>BIc(|KFs9w&z_=Yh-<4|n-cwQz+YPWq%
z;zO;9#8J#et!?d)qhS6OppbCLu4MlcVsFlO6`IjK>+s@({>|PZ&_w0UW`qBxLmtqx
zbn)Ne^Z-BM^eYtC_fIEmb3k|{rk{x`Q|W!!=7_PT`nIrTUGx5ba#yC+s>Eke5?c+W
z2QY8^rM|P--wpVy5$%ifl#4yj>qyA$VNmfCDbZt%jH?kcdq`ec;AZpwqSQl1gy>KA
zagH(dgf4$~{SrpZG!or49iqf&FcGB1vp5Mm76pQZ=Er>fA9hlRhsUY@By+)f%%?qI
z$}{&5sOWFM)ykxHzV>(=6y7j26|5joDsv&>A?A*)O;c)$$~3%wR^`Dd!!GSM?r{CE
zky9BcwI15eG?t*od|s*@n~mm;kC&iZ$&#r{JrHUfndDss0oZUsqqyrx&Uk!HKG@M%
z5>$OYGpF@<FyyfY$+h2YrFlo+IXZ3b>B*3M)h5vm6387o<6W<7g`6il2blh7B%*E9
zF8VNiD;BNw!(cGUAR7E{IsUsk5qtyDw`e9+LeMdGL35|fB8din7bk%;yzy?1ChI07
zhu-3i`i%>7#a=ojsM${1mn;J-h=Ny#(@`0%NR`&}+P)j>tD5a%R5NN3P&8b38RN(`
z@6Tq-c2B5?c5eCY$2tvf@x17sa?=oJe_T%AWgI)v_oDC^v*zXUkDb~(_;v5J+%3hg
zBke5~k{8I!%!PSP&nB?DG?w30kg0ZIO8&!u1h9$8&{s=O&?PeKNFqYi%CZi(K(S&p
z(z&wuBX6lHK<VQv<lK>C{H`QworxL@HIIipnms@8!U3k^6!P#~`p=EUu)M9T2BVv?
zZ~8ILax#tR&Rb;}T;Mi`Z@Ev-h{ju|#b#K6T{J|>8f<Bnw{g0RY}rDdE-WX?SK;F4
zcD+bmRU21l`05qwyi%9aN5u(e(MNU*o^lb*hY$US!%-7SBY3!qh_CXnizWr1r0`Ii
zgkO=FfCUI;DAq~LOdq+c_X=UW^hha(?U^cgGlkW?PAVQ(1MA&P3eF_x0~Dd%p)(9h
z`WTH|c>PRC99D@7?)j$aR8iT6ue!8LA`JUn6G1+gj^mqD$;^Lcl=oi%efGj>#mK=#
zLr+idPw4p-fc{^<1|ou}Yw!qV<1TdHzXNYs;QYNQ(Sil}IV&$<_sN5!P7`pK>>;gM
z&v%BA4r3`Q@s~J!<K&-W)Z8|sQOt)e7OQL7#?vw++AX5wc3b(I)RC~RyN7Haw7lkH
z8m+i4oDlERzNE10*L-Eb-?+e`HoV>#*!yK}JU!fck`}qnKISzhH#C;O?-&OJW3;~M
z+OsKvXv#KWehF(58^$&z4;29C{!MpByi2XZN&RIt{!qch!MP*SfCrZG<=<QrB=lJ#
z&IF+hUZpr<>1|qwQ;}nHrqwl>eBN{Ot_UWD&Q;oNeivrij5(!re*jwuJa!b5+d{Kw
zuJBx)k8sKRT>=I=ogD`)Oz^y0YghC}3T8E>2fZ-;xIB_?h2BB`vU1~yD+&9AVmA^j
z5~FFqe7wx%0NFn~ApF^qd;L{;ONL>29GTDsnEp0gLcX)uIRJ*9AMUy4E*Dmy3;@(b
z=6u^&t#+EZ_)JhP_iJe|WK*)&@ZEycrF7pVoz~c~iY-<(;a@!VZ@KWleWMRq?<XMM
z<GFdy(3&z}+hyKdQo;=)?pnb|rKsj|rdPm<#_JzfHNWnpr%82A5?wB8pl|MCOe{A-
z?|YFo#L(nt($K4<O!A&!knTSnl5%UowTs%f5dKcIf$umcpPPP;$YjViJ&W2byLe?u
ztiGS_cHP=SHeFvPMndE1YSI<48bzUZ^rd#h{=qsEYJ5v!`@*6l&gy50hC%LtX1faU
za-u)i)7yy}%8ZML44KB;*-O{sJm^np-dTpH8R_XgvxP&vPoZUtaSQSOU5O|P3G7=E
zNP$QQye~c?eGGh`LxJc+D)H&=JrP^cDAGVuI(i>3shR2H)T92VAB64Yj-`)#n;OHx
zUP1*e`!&ZrSL3Z@oz#Q_{D$~5fmP@0jKMcNS5agr7m0<^vty5S8cZ4tUl4}~gu}F%
z)57|bvbd{KLb?Mev`(&?DjcUYglu+?aryswY%tf+a7L$O8F#MTif5~|oJtP28ud9p
zBt_883G4auniY@N=eR_d2g2Ot`awDu-rdz@&HI}gZ=5Dv-&k!Tk!eo_*{LjCCdSmL
zN;H9qU+agMFl07s;Y=5_9d1EsYnT?$GP*L`pz$RqQII^gt7)lGWfovJe|TAI5&3jq
zXrMAvp6^_to3(Vy{A6YU>1Vx4pJ~pP1>EzHTenI7`M>>qq5cP0to<L#-U2KNu3H<n
zK~Pdql#-AJ0TDqOQ9`;KK_!)z4y8drx`&YNR6s&nsi8xWh9Q*h{MO*}_`Lr)|9jqZ
zu4^tN2AtV@?X~U|`@Y-RsWIJ8VMYs|s^<j$o-O?5cQ;Q>47+|mQ(s~w(~gH&@5>Yo
z<}#Iz@sCJ~e#BuU*<o_O(PW}xG)juz<+aEXo*L=BGJ96Cef`R%z&1gX0)7|rG^B2q
zOP<QJiR|Ua=_CCnS)I#FTy#HbAJNtBy^(1@d|a8|*<Rr5Wu^LQ6D9VIe^0vd)mrAh
zd5xfmYIbvasS(|b(h|mYWmuN2uocJQhvz(G(R?a787#U!D`st_l=OI~PXB(y5y<7N
zwcm_#E@C4x?hQ#XP;e@s<cT4eJW2|~uAzz4+AL+Jm69s9o}ElPsFj`1q|DnGzPCE6
z&09Uq#JjcI68q$<a`XU!!VDu09UX1YZlJ00n5TTKb++Z<B1*LMu*0=@vXp3zC7_`;
zPv;?q;5D`@+;$4ttsMpPL<lk5oNOLjV$1XT$aFfSza<@AZ{mHoRI&z))_g&^S*tPQ
zUGqU?d<wrgF7=TNOAmtdzL%A<OY5R*h2sX^i}^9_X{!!_Oi-f7BojJy?tlCE{NWrW
z-}k#*n3vGMqI6Nn-h1V(W+T-ZJOs;`8=njD5D^pSsrXR22Q(2syMsB17}9mEWMtC!
zKN8UM+Fq9o;YhpxSq>9f%N#4(QhU(Yg|kkSS>}mqxo`7+GC_AOlzQ@md2)jJWh*4I
z;OgQ~f<mf)$k8kl)4s|bR|}@sIMPm0uDTy37m&s0wk6^?KRUJ7$|fXHINJ7Uy(hH3
zeq};kNU6*7Mm-K5LKKri(c=`_#Ncr+$XzAEQ0HCjcx(N_SS@LWZWK%EdfY%(zMx^m
zf>NObcB^_?Y^_j->MgosrZ+X5TH@!F_s!bJZj`^t?|C9yG;FO9b0kV{R~CCEf@<%{
z2(#AmcQL{0QB_M8g$NY)Lq)TocpVot#~{H%)$HipTg8z-M}4q9nmuCZO<%7(a;+D%
z*(eE~``MaSY>mM<X2k_pzbU?|el~T^zv4W8SY%_oEt|a?7QN{t&D+Pw_{}4xMOmZX
zdQ*ZzF-TImdR9KHN}CM(j)_c~Q>|oHkKsz?-PI-SEZ)s-bqcsrHe23?OU~M{7-344
zB~tP2+u3KGb&+#h9Q>pC0d}<O(@(+7uaP0PShe|IJqWq)_K{(`Rc_K`5z!t-&bRXs
zYtN7(X02V=)==@bKHeW2|IeRbUObJ7yjd3{Aq+-DsRwp&lMkcH@3uQ1e1GdZ2^fU6
zzLGubw|s0O)3N^LmVT3A6ot-Mg_Zi=D#!7Ppl$PGrL73=D%U=8EZ}_cGStb@nNTM<
z?H!anWX3a+NpN&e_PbAw$g$Qvy2B1tqZ-TEJ1-Wx2Yw_DN6qyFq$nP1$_FvMD>`K{
z*WlOV7UiU!DiWh;jWwaZAA>!i9*6Jv+W^oX-RVx}F(OAk9*pCM*9se$8~0)uOeWm7
z*>9!YLme#JEGORR$zd?)?V)y9SNs-Vk$U`CUVd%mn9hAEZpTfem=iZ;claJp%*&&r
z8`?XrWQRqYPR3eCL7^X2_*)g%jU^Kp0(6-*&g3q>X^H)iuE4hWUMV<r`-aD}#T^eP
zhL`&)l~sm5H2u$6Rs7cmvZX2zxf1j1S2w086>fgJwEJNsM>I<>ED{XNOLc^FiVeFa
zm6IuQC(p<0332IqzxiWh4nukVQ$Oo*7mbnMMTyJ{u?hl{&L9O)hkM6s8{6|?jijM!
z`Z^zbWY4kQAHVFFtIi~793!^-eKjggvct`6SbP;@3(t*i-2V%6DJoY>dh1h_xR()U
zv^6u=Xre?+K*W`AOd|}c{nr%^4%qHG;RwzSx*dAG3U=*1tmb`3cRl+)h6wK9o4-%h
zKTG@Xz9TdTJ_5%D2k$ic6kY=dPf}kJ;^JX?4pYdU*!{{hD_F?u8~BKR*Dsx%mSQDN
z+(+q}Zq3~t*r>U^tgwy25D%MQ)A9vE)3YBm?DIFfuZE{7sTbTUWqirdJ!PWT-z=ME
z<FH=w$w=9LHmHW(qQ6*_;%yX-#nFUao^c6-XkGmhyNeqoQbfq>CUw<Q&z-|h#@|iZ
zqy~+Y-6J)F!>k!q@uj60G1sW+r+bC%aU8Jl{=6CY225Wihgn(@o;%{_#Ci_*H(Cay
zZ7f>Vni#9CroMd5bY2NxETjpk;NKtl<d6{55i4V;`c+kJi>%=3c2t&loQyZhI=DqS
zZF@Og`@N2$b}h-;i@h^mnXQ=@b;j!WrtB?HcYo|MAqbX#gdMfNv3~6R*p(<DW?fg1
zBGv3!tD?UW`_j$w`ppj@C)&MFu0}f;VsLM7{*le>_a;|1>u0~5H?K3OL=$)Lw75Bb
z<>TFIqRWnou8XG|UzpW;7CW;l6qgy!#*3SMKOj^@@IN2tzk?fIAszbswqvK9L5Dk*
zlXGqQ3%}~+yHlQq9^=eaz{Sp?Rb5}APn&+MoHPkKHZxfmA72_jFSSQSWvaC-=?Gr<
zs9_%&W-#^JqxZZW{OF-o_8fm=cZ^Xh?*~z%8lK=`hsd5o^Oda1n2h##n`82Xg9U6}
z_HE9B`Nj4Fe-0v+rK7kU$&G}ZRoD6rK@_q(XyPmplQs{3h80nFY3i9DUtjH<$3u`I
zggkNoMjv<tXQ-4BJcST_75!eaDQ;OsoQagF;8+%Pi_g#{?zkxsEnyxOChw+O8_PlL
z8qYz@n62SAWS8$*E;zUIlf=fg!LG8NzGwD4<3RF_rl;+xi@d6aN3M<ayGM49OkE4^
zNY!*`R}-&p-w|DVVa36o1MxmJ^^;xZ&m2((<;q6~H93dY6Y(JbCcfx*9QS|(i%5QY
zW;a1TUva+1BTMi7Y&gdYUeWett-}C-B;lR;eZ=28^gCbA<{UFeEOaBUZ6D`_KSRAQ
zNuM9hkW7;MG2BI5FZRF3SCo28uwgGEaL%1YFG@YcY*BmSQRcbKVAwsEsyE#?VE3NE
z?wM^e)GcV%$QNE&vH!#xPNmGaEv15utSM%fJ=o47+z>4Ys{E+``e8@b<Jv*`XJ#77
zJK70r+-8!`_AjsQu~+7Zl(2<;a??#aT&|IJTN+v*N;D@jRE?NR#lX!RNqTZ6OyK>e
zT9q<`3{>gUZ2_gy7ls8c(*&O*{_WPKC?mA%X0qfVcACh>NFDe9uv+wwVdB+2-n_c^
zqQh!Fu{)jSJ$d3f!(qlCQ~0NcFFJ+@mPc+1Tv<THDEd89pbn&6(NA#1qadJduXq3X
zBHcMIPfLvPaE-E6DK)zAn(Ol3*v7s$SLn4)AJ?KPtrrJMl2P1=H@<s(|2Q|7LD3UK
zBy~mJ-?z5**pBO0+&xLlxA75)sRY_iis>y<m;u|kFKOjQWHo7?sNETS&Vg#W_KoSm
zQ<_&H)o&L%slvN3{?};g)VZJ1r=YU9?lFf3U_ySdJDAAw`^;2;gQ*HnUxENeg3y_$
zn7r|x;h0mp>Oz&%FSc*omUpzj9fsvIe+{%<-ySovPLwH*sG8QVcA4S$bwe<8689&5
z?58g47g<Wkqi_!C+)uTkO1WG~@q8=_6Y5KkRz4n8iGO-jE78PTEnbn+Zzbz8)o%Q)
z?|q&weNBInHc@Sgfqqr?Bf8OQsTNdws$#$_J!{K6PlxKqWL+XI{4Al1n8B*gKQAiG
ztTw`$b?U0lC1Th1R@mm)EL-OBU&$Q=5D;)@_Viy+84r{D4D2iaV2AP>ioEadtdyMJ
z^q|<p0mdYSXZ7-h8-<1W31$eECkR29mvvh@2mu&Kd&G)w)Eg?VzQA&zrP->IPmty~
zjqfjsFRi7#SC{bJW`8Yl1J(IrW3gDQel16(Y99Ma%E9I9Qp`OwolInVSBS_Cm!BO^
zSuD@mqR~%+ZtcVDpSc;vd8$aY{2JzblxbIVaU`QCL$82HcNRtKqQdCQL(xGue%dBS
z*C{*KG3}fh&f&%TWWjA(c9)#QxJXC2X?1^!GPAZk{W49*(n2vL_5H$jkCwu$ep;r`
z`{&_#X6Li4ifx2cmHKMG5E9NG33B`4dROlB5?2-)CoE}c*2NpF|6uAi?hzmRs>L#Y
z#P4TIS!7x#YHP?MVo&yeFr0YEno2m=Jn7N6>vPm^!(sZbhNHgJ(xDnmp2BkYX|WX#
zJ6cvcMl8mkc)e0aTJNqXIOo8yu+X`W{cUsO7*k$@NmpupSK7mNRE(b2jSuA-11y$m
zl`AZ~kCV30js`14-fxroVv5W6t%yGKk#NdKDm?q)X6@HEC_}ESrFn09<7<)PT!{1e
zssVZ`qA$W(2Lo!3rJbBQ%0J#puGZgRIk-{w_bfo`WDw*G7v4^w-@wVEKGJ=Z5t)49
zNyOXh3${;O@h~r(R<gsn@T~p+l3NZCy04h=8?>J!qbhjrE0-pQ{aw7+$tnw$Mk+*$
zig1}mO=vl!WAE7PZ@tKS`N4q<(oIH4H^rEw9UiXQ@8>I3S<4ofs2^1J*K;Co(v^-v
zfZnRz|6p-`j6mB_zI#iXo_F?185_CgZ(zH6E#Kq`v-WY3!`YcktqA=d!*Dsv^EqL<
zF2Z!PuD9YoSC$YfdW(pS6wo#v3TsX9NDus6s%=^xlm4#bq`s*W9!3=(*^4FmP~o+A
zM-|P?{>9_2Zlpj;2~!l7U_g1SW`%G=)dQ=VodRkM-<*A(GGY5yCbM|lf2B6=?qrzP
zvernm%Wv}e`18%$^1`yca=d^q`1ce3;F*8_2_Fq0vrnhdTM;h-rX|{>Wcc#q7ys)-
zh+k3Wn8;S1ZYDle7ekJje3drS&zez_Mu0%)7mBDwz9QArcq1tI;`K{V8;cZ|Ce$TN
zs^Pr-in}ct8rdJKagb7Py|osRHQB{}C62|%Df_;gc5?J35(cqsrJk|UY^;41KN*?F
z;htleR~S+k!<NR-VLEwA*EOXQVey3ll?wik-gY)u$F!Qv%@3zaP(}QytZovXGC6w?
zMDGPI@0`7@f;Qrdzs&Eydb48Il1q8(57x=Rvrvctw<<HN?A!bD90b#sB*J|EmtYzX
zr)%n`1uc3xJws29TJnh76Mk+_!~L|Y=wihN7BWp|L00Trl@NnMx(#k@e7HD@J3r#-
zlRTC2VN~XWC`55Fr|ykQ6EUa>|IfRL#v<|zDHPuE@sFALDG*!vDODY-Q9lR1=lB<+
zV#T}+YtjcE>B_<JA`1C&&pP(*rcut0<7y|Hd{NL3Shj1!0vh$Diz$Vj4ER4j9dRdU
zt<WauaOj3SiwsSJP<<0!;3;72b{0N(&0lKWNa<*?7&)-@;nyO-g1GDFF8qy^#_`U`
zWUavwxp@QY$v6K~+4sLQl3`x>KRBW@x1g4z=Xveq2#biIYdNTQW>w<RSL#a+i@_Lp
z=p4NaiA5`lEg+++zA&>;Z+ZViPZ2VsO+|b*yVXmxEQ6W0t(@svn^@Yd`PQcFTjDvt
z9;BJnICpqxUbJyomf<}<gbL|^bZC6XDk@Jq(f&}rn>@4>N#p2=Dn2`!)Dm0vO7-bh
z;5Ot%X|2POEQ4WOS@rMm!+8~nhgtvff2O1FCPEb>9TTq|lFwwyHOXqELsGoEZ@z&F
z%Cj*>;z;Pv9yuT!HTI&awP+<{?j}$$@e`$R3HubD^+e&-xAgrnnJ2=Z8I&Q<^=W$5
zK=tPFss03yOQgN8|NY_muW6Ed6Cg6pOV=9E<q3DO6L|K)$n=RCI%PhKPMK3w?Po(I
zH|Ms^BT!0N2N#~N?v4i5&UcpO^q6+Dvny~-ha2lx)Aqqzuw1&7CLZHY?r^8pDIVja
zdH00za(6`;(?RoHMf=%0Yn3dSXM5GNT(yWFZKl#PGF#cN@*dfDjukRQ{NjF~QYu~P
z7A)hsc|IX}KgL8H0#W)UFaI6XO$NWtd@V^Ifq44TREw`!G{v=Ck?6z5#*ZChKZFi#
z^K*v1b3e1wcneB(?ck-iU8Fk67SKn}g8Vd{l7sik&ng!$Ur{MoBhfaBl0W5`|GFao
zeeK+d&}$@2WPnaK-8X-)k!zIL+$B8lNcL!t#7@Om!l%B}|72sX_cA@BmsOcTnmXa~
zkF@d^ar1x$!3ILY-zzBFsYd$Bx|O_hMX)z5Ta|bjA}BR0{2L=8Ij_*CO1+!8SJA2o
zM~(ijCXQXMqo?7ln&~$)ble-$=x5Y+dNphEJ$S%)-0rYCZg75l+H#{MGEjv3>Ajio
zG(+OylBj~GMs6Ip^M5+q4BX+Zow$F~_0{sZiQy#@5$8WPBSSgv$%q8a>%uU+_X(Iu
zl_d$quho7BzW>$7&!hn9J0^VP(a)wFI#~~#Ekr={E)?Y0`cLXC5A0tBs#GrPF?JP_
zq)Bf3%Jrxd9PLL<**I@qsw_41vFUJg#7(8j`=+S0A!@mI5>7I9?=pEBB0QdTg|cm>
z9bAsZ%=3CS-78g-VoDdJ(3EU&2^-1xF|^S6FNk<+5bAU(`edKbxv7qe`N0;`i}}F_
z@{!6k9&bY*i?0j?^Q&9hS#s+r1}GpriQ!-EB=i4XBNn=<?uZk0VS-1C!_W55{(Hu{
z|Hy&t1ZNYc?m&jheW6pPm017$-@x%tKKk!Jv4nMLQctEsd*`9w2>R8x;3q`bXzxsj
z_ReSU5e|JrKY8Fa-3?#&TL@>h)A5%HYt~lg=3xj?qkLw$(VBb+1#(GrEm-xogK7VJ
zD-Un^$9r#=YR76FJaAm@g>;&0p!x?HCx*E$7Nl9Uy$a7Ux$NH&5}=Uocc>aJEt3vM
zPhxS$8g+Nic9?2j(2z}+M{lCS=o70_iWHl_Ou{$QysNHivQDGq*$$jDj>Wf>uW`Il
zFl=2bG`HeN(EJ#>ReS&8x?=^-M7W}kZcXK72H8AEv>!K4l*r{Nc+g021pY7_ZyQ5<
z^ix=8t{Q;~AcR^x#N{7Y<(U6?>;uf9>Q*SIn`z+T(ZPC7=5`VO^V9!mYw-^t`iB*J
zsiAS%_)Vw+;2|sp)zIG+(>r@@AgkW`lvHyX+OWsL^^?S%V20m+ww+oJg*NhzHqM7w
z=4UdONJn>@7XIUZ`R55#ai=^{UpfU<AeSH4I;lC_x7+^4X9dnh3J>;k<{Un_*>jM^
zqQ;AwAOAUB6SAK>PdxJRQ3F=TO?cO{7LD}3f4oIw0eWo<19>-qj9CinHc&y|F(2hK
z0fN)C(?FY$V83*@zugO3EAw!X5w3Ic@H(ue4pz;yMP&fnrXuKZW@%MeH6B+3`OFnU
zLIY#xYozDmACChEQ6DIW&g;A?1Iq25o;0Z}xI`_Zb6p<;ZN6-0{QUetaRl?aplj9O
z*kFOKEbuU3u>1xlL}$K!{c1B&e|~T2)3OoVeCx3*0)DIV0(zC)Yee*NLVK*u$1%0k
zK*Pr}8RH7~(KqU`*~)p83Um@?k%_WERq)pS?x}xixyV3MWq6+C!2!JO5XqA`EdFCZ
zvH(*38DCr3Qxodk99r;vyocaBeB?QbuKa;Wns|lQ5r&$lMDQr}3}D;Lrlu+Hoc#2)
z`^AxROOL(nWm(X$wG9T5RJK@-XIM?m%@_D;P^a=Kdf7BTyd`GUnj_8)d<;Z)#3m!<
zMH|%X83@PL4K*02Wm_03qCVWNJ?0`nJR1aEy}+H84UJzJiLNQvxlD{~-<$MjcfUE0
zP2+bnI&fPEIGvO-`bu!AUsO|7=&Ppz@tF<C^w@0!Ia@C%6Vd}yAC+iQ@Wpe}Yo8m6
z@00@fm;|Xv3HNKbwifp=h;;tybTqpG%~xWUjdS^06(YcumSVd&P~ma0X+Q3pAx8K1
za#!zzbLdriFb+hT8$j;o;Br+-(+!x9=7E$j9T>zO5!@Os)A{r&aA&_2-qc9J_I$4o
zsNg(+=LFab{D0ORPQmckVbP!Hzxep~x{n@5=0jSMW`FrI&cra@U9EGl?zhnEE<sgv
z>Cd`<KUw$hLL!uudWz_E?`@=H{(Idm#2UYb(e=U$S|Sf<Qq4c530&mvEq0gRg@-?Y
z8@o}n8a2)p##5CS^q;&-`vr=`v0GDX;`oS|3E<Jl1SOSpl$h~ncxEq185z&E$7EY;
z0BC7*6bmf_PV!pDU(7E-sv$llCFOy)5CtI=Wbd9%xhT5w`i55C8VFSNRXf>fQ;XOS
zcrno=Z2$WH2rh`Gm6zY!K5!frZO7tO@InbKv6T$ncOYG{w%Up(+XrgdIiQc+;w61#
zDu(1+zhwDr3iiyzHqdtveE^)3x`2W|Z4-p;5zpJBg+Usm1@+5!=(HNm-?@rVIf45M
zX?{#oCAfw9gcGzqQPU#&K>+#;q6PQZ{|7e^B6><0K}|QPX1s$ZWQ$t)N~?0_1nG{;
zy5~$<5lw*&SLl-7HM<r_&L69g{+Q%8N7QF_lRkQlYDe48Q^uF&K-TU$_e)a#!;Ri>
zDlx3QrR;)7p)2ictDYAL>2KY-Mfp7CD$~WJn!&}8{s&Ba>!cHfIxGXu(Wg}rWN^+^
zRU<io<*vY%M~6EEUOVBF>!geV0*Fx{9*Pxk;k``4@xV^`IVSQ#9mXn$fY5up<xw5Z
zVvFWw>N>L!dOi$CxAUBWui~)}env-p@tp^D2vFbRM&FgrKfk!^oNub<^D6TGmn^uq
zT52&N!YH5ODXQvK_$P^mMWf_1z(vPbcg~ac+|TxeMfXe?U}2L$BPYTKXj|k1=jzdN
zo%0W9eSmXX$ss6S%OGt+6u?>oU6k6zlPz)Qehv3}C8)a4$AE1mxa2CE&SNchp03Du
z?<kdRT!BmC2T{QpGR2KOpbidG8=1w@e5gqSqjK5G@4-vqscz#uoXWOft6F(8tK{g3
z#b0lIFIN?NG=fPrQ!zuP^yLp;yFtAIO{QjXNX4z^I9#@uwD)G?%#GTk*&7|k!42Y`
z?dqlETTCW8sIe+mhcw<8{<spT4JBQ;A_0WX+p2PePv3w2jC}aj_p_Y+@nsMo;l1?0
zoO4Ye+YTpVbL^8{fzr(II$lwfj-ww4$cVr(w*F>?$M(0?q46iauzmC06P#fPCI~rE
z%>BXB@}l-1bVJb)i_RB<ETugxFGeue^FKh>m-$TQ8manjvq^$9EL=i@;Dg_l=%mN)
zOl-(lt^m-Rn(v?Yz!gY3efEN!l9H0p&e6^kZkB(3#$y&<%C-T-<S&Ro^IgfJHk%8|
z-8WNRnVr^xcYshlA1co%aRP)x2a`%}5|AB}=%=Bmc-W=q9dkTU7HS2=>%iYF4`lNU
ztgM|slyPPAR#z+ULC~JeZU5=8GA=A(2A7)41`8iPU;$O{=gkL2YPT1=lFz;vtNFoi
z52DdP^z8tndw19Alt|bWpbx)XkaX+It+3aL(ggT*Pe=uv)d*~46L{0aU*iUWHa01^
zeo2>?Bjv)NiI)o3g%yESggP2=+(W@<F9d|TDV9^s6}nGd6J7ThfV2T8_fL3-Pe{c9
zN1P4weup+#cVBc}x;an}A6_u=>&tqO|D68ZfW{U<=wP<W(0Jz0FPlli*?tjw&GY;j
z4oKojzMnY(HI!NxNd;Zw(Ug3Cjrl5|4k3Zl8pmxe)HX}exxF|TJW!xpz=IRe!90A|
zxW}_KicJ<gW7-u4g)cfSDL;WvMRLq#n~Vj57|@2M@wg@mxES@MN_aDxip8$9zuEhJ
z?jMGkQj&#F%T};&Vpcjb@~vyh=oYBpOlKP*quE8X7lMe<aK)qTJ6BvxXmN~zm-j<^
zWGnI*^o_c~<Nd$}EdqMW(UkY^Jt}v=M_bCC`;hmU?Y3(<uf-YgN@-{g8630%u_G$M
zE)#w&oa32ffq;aBMB}mZF}eMNP72|rb@Gy@!c}^8Z#s1;RsJRP<3BuAZ|zA-+6K58
z^Tr8o^67Z~;kyXfEerL7|LNG1qmO-6yg^eZHK3->HWI1}C)gX0+5(Q@bphvcrNf~T
z%k$af{Ep|8e#9`{hd`j3O`JP`a9m(;1flY|`HaL@s+S1^j>dEjM4j@3wn6%(8<aGZ
z*`7@}bR`K_rqOX3O9cDo7-~P?>SsEonxhg3yq@FA+TYvlwG3R)Qi=kE#*X(MeRJ&J
z0p*O~QRj7OKpFm@H2*p4V)z?#22@}ux9a=Y!(&8}V+`ii&QxL%j2n{GBAxCYk=A2m
zpZg?iPfo3@@<@@+-4N@Qc;-d=Uv;2sA~}{P8BUMCfJesdrmvr3(=|}|q@hY5m5l@g
zs9ai32|DZU{*Mo(Y=>BgQ{t)51Ez>CdJ@xphkw@T(0iiv2nP|k^(A+yH^5b3>cV0-
zaDP-I`P-2|zcZ=Tdd}J%Q%EQP$|<J!^^5RF;2}d>XuGIlGMG=+)YMcFB?oE|lajj3
zV%J{)5pWXlzDZ5dmx1z1_m>drED$OT=e0F?uUJuFu<8K7^$zg%=x530TN2HK@FA#T
zM%Z**HLmAuDAK@_5*L4$?FOu<n)Fvyu#jMx*FOhbWmABIX$;76=_Lqwg)^2L0;97p
zKX1$<ID!t$n{Ur0%i*%1wq~+Box89E$Fu^ys)37~&}|k3RKfkJKSfUL(uiDi$)&P@
zO6m1`7L*<FJiX>vNBh6eM056hgX0KzDFxtgHU?tlCEM$Vtk(yU3h(_vICvJPV9RjC
z2}I7J&5Gp2Hm8~keIYgme7<m-*%NhYd_n<!)B<(epF3CdPyb#m_;{Djoe_Cg7Oy_Q
z2%SoF&$mpXw}>+LBE7`dlII;MKnC2s@(tV$rSjX^>|)70VUG)UY!{Rmb*joI9Ku<F
zor1oIv`yJ{Mgv}*G3btkvug8BR5YmD&uUqxJsW#oHvQSbq&NLCgzgc3E~zaNj;HU4
z209234VKj=H)57O;2`1$xaZ#maa8>}MqXZx9de;Z)@Y>U^L|vG^2!43Z2OVbwrYk<
z!W?LXq(cUQ;UbS7f(S6<<x6}(pB$`wuFn$%@{)2u$DS9<`?EzkdyP-;^Qo-85)uqo
zFa8J2c>knlW>7#YktTS38}0Xm&^=f@gzZZbo2#^d-@}zw_MkycZrTnZTu%_z@~?3I
z+ZEvHoRgd3sao^v_q>WOoEJFjbNyaS+=~|Hb%|f|iG2YM&tAa1S>Qvw<%{&b5<No#
zNipP|vmI%S`tU06@c9gQF?YXO9O%eNiCnxMiHKy=l>z+Rf4%w8@g6o(>&Y%i{H+=&
z=V@?h6zZL}q~&0`E;RtWPb=4Bw+(=W7#-?X;(C}Tct{6xA`=-D)4>UNR8^{JSAX|E
zBw#fVvB#0Vb@Ph1w^N~^c7+w~s^KN%U5B-&S&qi^*U=p?s(l<J9lqq80nur$35UbI
zO(hUK$w?3LwT!^A*C;i@?*~>Ws%v^7e{6*k4+o8|MTv_p;GKFcjxJ``ThlD<C*{Jw
za<>;_&()Vej&cd0&vWz>)kTU({}`zzfe>Ur>F0&T?!meRtz34WK^Llot`U~M`W35m
z-T{<`baJL!+nFzjiHVVDR4eow4eHr0^xY(u0r4PnBWM1rtGHi*6>Of?=2QNlraDu+
z7Ye{KK6W9R8=pcar*X*B^Be%f#L}yL-`|pO&aKU~-LaY$uT^V5F*eEV^Ayj!Ae#i*
z6T~Me$T^KyC!3UV)!6wz67I#Kvu_DXTb#4Wa&)&bUZEsosUN;g0!E`x?@?!E2J*F~
zoTH`w|JLsNhZi@UTs6u@^@2grOxy8rbMUanaDG>Z@MknT#otZ~4HgB#)`p1C3p)Pz
z<dojGM)6XbpfhiR{|RzSp&>W-wpxYDMJQ^ySbKc=9kS0*e<u2b)}^973x*#D<fmdp
zOe%Ny%C*por^R638PB5xOuP+KixqOxHLL(+X`G^fDA(wV1u;VXdAZFzL77Q^Yj6|b
zB0sVCxjCboGARj=urU9s1>9J{#-LG3&!Sza6k3pX1$|&8M`orq@>331TPO$v+>QPk
zGIV%|>6JV|-mYE%x!{2_T0zX|8{p|)Rb3Y`hqdX8&*_&cPWt1+_VxmPav~HtaJ)bG
z|7@__EldLy(;b)Pqpk1eSs+<CfZ<N^2av!g|Ecyp?8+nj0G70l8fxVBx$K00FSyWC
zLZOJ0N{2ZcI`1oRro8^!9M>`Fl?TD!GlBU8LP^2Uh$C;I6ErkZz`4b+BD2pkZJ)JP
zJq8fd4V-iIr2`gM5$YFA$7=ZDx7q@Hl#7QzK{yMYS6EkUje+_)%T+|fi)zU@d*H?E
zt8%a|h>BrTwS0tq8FoB*0w?+Qoe>Y7MbKT$1Z|EhSFfTFk%_j1w;m@&M@Q><Pa?3<
z<C!6WKBzEE+KnQA@Z33H1W31G4SVc5p-SQ;eHYx><Z=rrPeYdWyYu$f$;5-XALQwY
zh9>macpN8&<zu$)wcm(#UQhlMjWNmLd36SXF@Kg9YC+k2k6s-WIXdAB*%~^{obCQy
zG>m?vx4raZW5gQiAR)}RhxTSbZ0Aj<5Eii5mn90G?0uEP`gQWLFJ6*@w^2o)p(&=A
z{y2GPNUbBeLE9b#1qiA^17Tyrhhnd2++-Z8GJ{WPzD@j9&oLIXzZ>PAPjfW)bx>CD
z@BvWoy}Rtk#OL$fafUcHI$Heiqo+HY-3h5t8u-+Jbg8nj<=SHzG?W9%$!_2nO3m(u
z1rN(E|M=p2q+}$s1cO5A?K@W~+ARQ<?;Y(g)8ZdPxM(ac%mH5-fkn(L&OEbT3&2US
zHJsscN7^=(Yrff7O#-lS-RHHM_Ps*4Y>M3EHXV2iD0cZOp(o1C7&05#$qx?e&%YHJ
zV4$NIQ*C=bh+Qdxch1)Q$jT3y#_5?sW09N7$DofU?!2}1WAWp0jO^<(=Tn`2Wn~~z
zz{q?aFDO@1+(X7U5YS4BTY?V{gO@-kN8Z1rT;Xx&%%N7pdWBWpvKJcii~8VnLm4Pd
zZilC_pQpkKMJUTRvF$*0t$y7D19xsv*Nv9AW*0HJGdVStW>{=ErL}^9x9+<^{&{-<
z*tSh11>QKg4L_-N9OEqA!w5@cA|8zc=CV}4H9bHrRkgLgjud43Wk|i8CDRtXSUp#J
zd{|)PM#1kGj%L(f{Sr!x9l@fhQi<HPzHmHtiG;%_t+#6tpt8-;fy1|QOX^jWQcZ?-
zA`743_H=7Rz|KccXWV0y;Bhi|-s`c*^%Om9WPaJ2;U)m1oxF%dMX><lG*G9L8reMt
zG9aFD+Xa$wAl%((z!mHQlp%a@7@4?xaA{^+M?icIX!ZNcu7-w%r7g$9JKn<b7^8?{
zRu=(G)<a0X_pifop6*bo#1m=>N1MOi8L)9Yg!ItEYa&5#iuQkD{}1*r1b(#$Lx^$%
zk{zB69wBtW3LXP^dKb~sDUDV@p+K?02I87~>>opJ(WyEHz0Zg(4%zL`UyVpo(=&oH
zoKp7=fOF#PY@(%l&HzHIe)ky2{Y29a3CPG^NSc5zeVnKHF&*tqTh<Ti+-^fosac?<
z|F2~;UE-{vywbszw+?DqVP5GVA|;tUv_x<`t*?h~J^dx5$K@84Knb1|P-03{8Q#=+
zVK%m_w&T1z6O{#r!8@lf=Y>VkFW{%#fI_F_n`_ve8iBM4mQ!g%b5H-C1!#826O@}6
zHhIm_wNrZSLaeSpxTw@UMGX&?8a`b&kXTFQM*vAyZCh?Aq52tvx)%$5SqenrA^fvg
zLFbEb$av*IMuEZY=)f^@p<h+PIc#bbSZ2FHC0rH+0`Y_Tv-g!)UlPym=sf%SO?>Js
z8PuO-5?!`px=MTn1y|Se0d6lLiBx_RRH;G<>(*C$=W$P|uD(w)osR%kry)hHoU~u^
z<t(#R@Q7CER6j1;)BGwsf3(SW+k31{Z3#^mT8$I)&PjF#=?1A}-$A@t#=(m9zSmWq
z6{)jQjR>FW72yj2mD7PjT0Fu7C3JB3T2Ph!zJNv3+gDS~m*=Zz0ab<I$aVMbydx68
zONGL8Pci}3n@R!-{`JZ|M_1M>6^LyV_!`<-!K0n`UuYF5bm%A;>k@f(9mt7@Xr1%S
zA60DbyJF0u=Aqtqn9aVwbRHBjs^9TDy-K#<k@E)@w>`3Z&CBokbSW1JuudiX@wMns
z1|vs2l+NYX&{$Li@=0_X_eG6G;x%aFazHWU--WVMm`SaiThJIsK(^x}fZ+?k6&|0j
zpJ9$G8mcE!<c~uvR9J<q1~;gGn9R<o3CwA<B&uj3cWRFnGk)9p=2xct`%DOKq4q|p
z1ZFgtNJ*Y6Ax}4HBAMcgO2>yQU;UDV`bOA;e!YvxiBcV&BP;#$*hCPY<nqqt(DIf1
zZ0DKg^TUiu*%nl#c{cBHUNpoqVv7{pnUGF~hWL&cjcG`A+vFx_WL;|6g0d2*+BeV`
z)5#?dF~06;y4kf8ZXX+=8!73Kc*mqqoV`{B%d1KNT4IqtyDt7r=@>m~UlzZAWJ6c(
zzL`75!GKHz+b$kqq|+v;<Twd7#RcG5q!5pg?08o-jX*gUUTDesc!rS^YS0(Zr1jnd
zEh036Y~=(w!`n=4*;*Ab1_lhHxy{=|%eA23Y$O?`GNy|#EpBk^(9hj`+s_%?^ZF7D
ze#hQ1PHhO>A-QR_3=D2@HJ~?;<zJyyV&GcpX&B@Hxm{>xr6H1>f9QwHE>ByrIVd6C
zd1jGzQH=oAfHms4$ns=DgJlS6*5Vi%a801h(1hO8P-k&_%#F<jhod!quyfh?z~xD!
zFMfS!5gRAVL`<SNTg?uY<)(2+`Xd!kM2D8)4kZp!1v1t7>cnO4-z|Z^T66nQBQqNt
zq2xg07C8VoB=ww2cmZhaglTrF0Wkh4a#pBESfB^NgJf?KpSUly6C1w`6|<j(o;ODd
z(fz(B5i-ob2M#{_oH{N?<lyezjdW6Ty!$OY`|P^?b5+c{*9|BK#onwqFOTx;vxhxX
z8T$Zo?b&cdC^6GaBV2z9x*a)Bw&snzq2=@Gb(|rX(j1Q8x3^M}Opjb?cX}-b!xrif
zibreR1TSB^Eu?lEf7q9KU_~%;?yRwvUwjVTTIE?|X5r#%4Wa4)n9+qeorYbNYkhGP
zOUs&dJ>pf*v60MCtHRfqhj#Oa#hslmOUXjvfcG*vU$nk~!J&7N#+N%lkfP~mT{nJK
zK<$7{H9wxU9d_~Y<4fpbWOf&c(NuO{7@ce+qum8Ye)FM1Vcx#R5=+EA6@G(~Xym>>
zQ#(e_Jz{}Ik0x<f^C|zz8tLW3Rv&4LGGk&Iv8T$^tJY$U0`F)b>+p}#baAV=^`Xr%
z&_d}`v!ts=J!fqql-hzo6Ogr(xlHuz&sYksi5NEzP`IOl+YYr1+wZ!OB+f>SXazP8
z><5BUel%|rxev@{NtwuY2+Ss8wrSk9CPXRQHJIFVCjE!vWviv7?VnFXZ+p+xQIEL&
zuD6OM-`-j80-|@?C*Mm}4X+D)ev@PS{!a}V-GD_;qV3OK2tj|-{rXA3v-J7XJtT(a
z;fe(V{Adbl)32PdicbB>!hG-{3-h79G~rEz&R%nD^x1!hJo0nl4=}3*hqk-`ebFD<
zJAp`MaaBj?GtroKaXds!L1_8Q$Wm5S+acABOOY$2GK-tl+|x^&(3F!@o&g~9-jONm
zA`%Kqbi{TbIEs<{U0!NZZ`Y$`(y~G%XZRO%EFx27sUu$U5J%PnaqL5ojNqUx6>!-T
zH=S8{9YV=x?YgN%*pD5P6HyWmg(=Absk|##4pb$chQt?k3-Kdo+X?FUgf03<0WgM}
z4CFGrRgt!zE!uINWc>M@D-y9=0c4Q04Wf(^PaOrFv?^^Rcz2!w2c@^cYE>b~K821n
zJKd3kUBHs^aHjX`4}4}GSsL-c5+aUHB1m-KHD}s|x1e<xh3k(O!Q=2l>t@1Ofpc8K
z`ukga>!`P*)lPAHRWp*C&UyiJbyzATQ^@-VxzFd-^UuQ!$aWd`sJ&=tSjWqQ(bcA_
zx^mDG^BlImWKzElO|>jg(y33jDl--s5<%ASTKQP7cs_SQ65x=N{ZkP;b<PIrGr2cN
zSWgP%n!o4jh;OiydVm|K%@u(vxeOjYB+V=6(%J74#os5M70&zHYDk07?dUqgAd=sI
z_st_AiDILn!eph6+aTKP{iXu4Z5rWx^^)f0y}z1*zHZS=)5Hw*oNbg)L4C(q?Dzrl
z4yjNf5Dz^gNv#?@yAXm1JkNRbDM!Kdwin7my}(>ydLT_Dt9#ust-4cLtYRU<{26cA
zWqI&o68V#&MAP2@WM)c5v1S)rLmMCq60MCsMOk@20$Q4ih%{DLL=kg`a!&tK<H)xa
z#YyBhy$py#D0n_x)v*^QkOZL0tgKY!yvy<H$EPUJQZt3tQQ25=g<xAJ=r^o7vXC<I
z-RNM$e$wuXtEX0Ok*rTEL;OkY+RL$;>Zp~cX>wmAIAItYq%`a&FPZ>x>mYe=$?<Nr
z-Cy}|vf;<j+h4bJfsb?^;OE8-pY_iX45odVH_0+V@r6-RTs#Sod>7d19o$QEAP}n*
zaYxkqk_h|Wt4h~H-p#pgQGyI@>19LQpr~hZTrqL#Ylm$}uECNy;0CvW+(V=Ish>^l
zzW*Xp@VIKjIrRP=VOcN3e91N{MCdgn4^+SX!hN6zsu)5gHiB{5LPj=GfTigx`^#^o
z5)mZ>C0`cy13XH49WTF-j0nh`&agSI+31l{N)|akDJt~f#<Jz;AVA@@PmP_s@&t~%
zoR{1NL-wmaBh_mPDiqA259x<%1#0uxT#gU7dk#}xQJ%NfC)e=#T26Vi(JR~Y`ix|v
zrWW!5n1#i+ixx}RYYh+9+SnQ~FwjMM*MrqZO-c}wwD-)E3iX6*1?mV8Vu-AK6tUo^
zokLpK1BEJ(myQykW^f6zVv$_lzoDB!!?$jz>1w{s|MuZ8S9xAMp<mZ?YVCYS$L26j
z`h&kMH7OCC*I~QGd32gS^cudvM7|pGc|vj)d|Db203oY4z5s6CKn2x$Ma|%!8ACh^
zwDbCT9uHFh?Yji5+c8i6F7(+X`<*}!NL-77^VY3qx!ZIy@syE{&gTWl--^uvY3`7P
zm$K+-bqo1aOeCbWVT|I%?mxe8sI+ZVBbw~<uarCQS`|dq?9Sehn6oa^!$M9J>*x+x
z5lu%g0<Ah1(*(mW+sN^wqkVHh2_IS_VlUA7VU!SNUM-qy4P**kTNle;OIF!@nlE!7
z%1v;pB4{+2xFSu;V<+w}SG}iiE?#L*aj~%}?dpBqlp_EBVK=bHKixO{8tSEn&_69N
zkvf}Q0&n15_SS0k`~FfI(ZNj<Bo+UeGzC6M4up|&z9lq=X!(yLvJC=exuMm0w0wLd
zEB|1oT+RRa{Nucsn+&G>@0;I}P{S}(fDJRUxvA-^H<bOJW%1rtE!fD@`{nTZ%aP&h
zATh!Gwfz2YPl<}f9f9#}Y=P3xwAs5`Qj(IPI9+;2lH0FUXzlZ&aUpAQPCC2Hqq@#7
zF8bn9jjFnqV)oCktE+VvwX%4i3g<6+2n8F3vX`U5RMH!+GnNhoh4c&{W%7$_emC8w
z&=bW~H2`3oKy_a^XX8Fp9_NQyzO`4?#uw_yv!)!jPl8Psm$oO`@h#&A1tvBye?<$H
zJ0>hEj~X86bSekzR`+fn4leC>Q*W$2!G_b94Bb&ljL0u#pfHnA?q4#3(pPnhe74K!
zF<^MfRR`UOdX7Z=;wGw%s@`Avq#zKkZ>imgR3<Lsc||L%`fELEdD;Z&(v1p7MN=u~
z#_$XJOsAR+s^uf)+f|lMkt}7P4g>ewe&M>lng#9Gz4qDR9sZzoc@WTK_;^rZiyk0B
z=^>&LH-q`NqUT<)ZcaJnfOuYy6C~i<F#rw-+7^s{$J5^*x1dHkjBpTELVuYgh6enG
zdHin%FQaiCUc6V7u>A)-R&VGA74kzE{~(Kh=A?v=j@OtltguF!?$=aoCILOq7cs==
z&Y#%39>M8+m1e;&6wnGE^95otH~W{RFMd=i7dg@#izgg^y%_uuh23NuLdD0aNz0CC
zl8&@0^~6a#_&Hh~2b#1zY<03S*Dyg_jdgMJY3CcvfY6$VXfZF_vg#!sT<jR;b0Fx@
z>uQA~70=vTM6UorwU(vM>awb00?0J{d}RztuZ`RrMC*#y<~v#fp1R1ECtt;sXeJCb
zAf`WkrbOIWo_$-joyZ|yyV_A1u0(;1W&fk7Ou3)uCUzm?57k1}M`-oN>5>PZe;g?>
zbNx7>iGhoKh&lNk3fG32<;lC8qP3ee%$j~5-aiT^kgAbXV3vBt%vPE0#TdxM$j5Ee
z<$g;hp59N7N_tTMb9$VJ=XKz+{Orx1KE@?0obtUpg1DtL4-f&n?1qsWWwE~EUy#<s
zEb1R3O{-*gxXP;8Ks1=Y=?st5IOr4hX?SlqCT)9~5@h76V)_TkJYOZOcI(H92hIs*
zgJR!Sq{$fa>Z`F}s(MbhTjNMJXSJ{|)eUFZ+*Bt~;aX*e-TrG<i5@X*6ho^ZkxF>{
zP)=6y{=IY%D^?&cdDfQBlR8{qqCsRz?BsyEDojH4Qw))lvTV*5`Q=FN{`8yeGZYIY
zcFQ_w0T*T5Y56nAv8EXT<Q%_|2xrmU-{Bo?KzfW|Bk7Zkdgw3k?wr*5FkY01eLHDB
zTK-RUj?ZCG=ra&RBZs8Qf86Tpal-M1O05^xK;^9ywjNxm?wPAmmOYWw=gKDlcps*U
z-cJ<$vU&oe23i%lP?b_h6GW<fdi6r$%9k*@;H6S}iN{t%=R8>Wd>S1239SzIY!|xU
zV?f2Y@6b@5#D@(JAF+v*Q*+}v0I{BRaT!t=3FCnlSD4jodzMq;;lVnx=i5vw3MOkL
z-;^ncJs*5(LUc_%RDa&@nIV5*%hNiiy^VSQD6g=UTpHWe$-r6Fz?*IQpdNZNUq8FA
zodC<>K&W}L;gOoJa_Gm3(A%~7WwJ8^S+{Ceo%jU_)s$3Bsp?a^_jRQ=?(QAcD;w_$
zoklL?XysY|g|<7shp*Qa^P&1avU3t=z#tEsP5J9Z$B>2e5!DCwvklZcLa~82xc8b`
z^!W<Df6<w!Q#1PoXx}W&up1S^lqs}y5$VPAtUC6Flp#jRj^9YFbJrybdMYco9oB}6
zyWIpDyDkS%cDPCWMhu6=lD&)fi|Di*NsVsB^-t!#ih;x18i*)d`jtbzG!&0f`bzs4
z_io`uSJ}IZF|AXXWIdme3sC-*^7(1Rm3R2$!Iy#9-fi17kF*_Y_OUq9v)BKwcB4~)
zLv&*J8zP><^gV%(>Q57sf=`2~Tox+XA%F>hv&}1DJ|EEYR0H$E7HCsCc|4JUT37>L
zof?7#JvhtK<4%b#F`PQbOk=UL@?;)k720oWv0h$2P{)%8)N%=#Ph+y;u5~^KMkc&l
zKD2pJ5g6SzAi_`Mo0>2&J*nt++cERJZnsVL?YhZ*5gQLO-xL}e`wt+o<gM|+Y<}}2
z)Cgg=N7mhHo41XQ`J0il-Hc`F7H^vfvrMb<OX3QQzNgnBf5PM&z4HA#-*y}zqR0yF
zK2q|P53>~n0egK3-C&v-4Gcji_2(gBMHRuBdPdt=NM_oikwbLpIqltg>n=3M-X?FX
zjm}DbV&2-jo9AEi{Gu11Ii*1Yzmp1q(ahHAfJQB4ZqP6C-YFlouCTgAF+KiWEq=zN
z33`=sq`FekPyK>$IVM_{9$r^vpAo(297oQBT>l3BUrC>Av(c*ILqq-u>Hf942(*%s
zY33V4XXh%K!`&K`Lo@4|K+The3+fUH4lbYos+1;4L~U%L`~HrA9iAVOOOA5e6LluO
zc+@?qP0Nf}FmFF<45MO!je_q<D$h+p4H2f2Sy6%qIoM8CywGYbGwyx3#V?-2yPR;}
zF4g2`4MyTQ$CdCNlxa{LuF(XlMuk(+Ugdt*ExfI&3uUP6x#YVV`_%7HHN=L4i;`ns
zH%Vn^c(3mLYdRVIyI0LlfHtLbVJC>JZp0j)>na+N!DQpVM`SCs&^;a^q}1aFMD_u}
z{`@3y!NX)ZO)K&GJX0hEhhe7~y-@y0TR;A8qE;_;`!y^|@vFuXNy1nS{DC_8FbIx#
zkhM1i)d1RIHfa40hR)C^q)}{PXxiiWY&zZ&m6`1d73MD({I;kS6S-_hU5tk;G48!(
z(w*XEW{ZMUP$~3gAZAlOSEk>6E21X;!A>M6G3leX*CsyJLaTzBp9#Wtj_I=LO!KB=
zIVRHNj4`D$b4-{9mrh=*FBk&pBT55?i?u04(PeiNHO?F3&%=KARr-`pDiT!h@KR%#
ztT<`H$W0Z3&T%v&eEOr%jI$0BaynY!#GJ|NyUwXwd;}5urE8X;GRe>kqe8p3{p&o`
z@0{wnioynRnv|CppRY{ulr8Coh~kFp+ey0id8*{U?C{Cv+Zr-(RHWHNF}A|5MF?Q;
z+|5G;DZB*DJVr7qgAGc&AcM(9zfw?@WYj9Z9RZCy-aFocIV(?8g89~bXEmUSPDntI
z2}PPaPFrSD0-{;`s$#gW<-Suq=iBCbLKre4Oy;bewD#%C$lTd~0pWW8>9)<k-|~0V
z=q<lWsYiz1@*tLnzeiaLaZl$j@fvke*1Chg7G+LOlL%{;D>M-dy-KumNt$#zpp>SI
z(&v578dXa=+`Cvvf>bjJMiZS_@?a*Yu^63fdPe2ae7!_^@7=A((L(I(>^cTsBKEKL
z9JNo)WGgS`#b=id28#RT@2E9Djj@LUh3kIHrQ{*5%%0P@8pXXZ>z2yczBrJ#9>6B2
z9)<CAHN>j~?YaZm_Nq#qybQTs+dz{!^evHi7@?K}-|y|`i|Q`Q*R5r}c*y*5_io8K
z6>?Q+Uevxh=$iZcs2E)g)2c3kKA^Ya7b4!>YOy8q@z!pwJayTyeKTq(nfXxrHfG*B
zUft9XVXw720=bJug*8xMO3f~YEGc=+bzeFZYqK~sK>e3P3LVA6<%`JZ_)?vagT;A+
z3xWvQ#iM@f9Y6B}nXkdIUD+EIig#2^eE2NI>|DuV2CNk)BQowUZ^>}od;Sm_+5I@B
zx2J?|=#<CWu&(h)zi=}mi`OT%FGHSjM0|SQ<YD>SZU6q*2oZsJ%2NmJ<Mp^t5X1Ku
zvfk2s=+0oP92SyD=Z5fkgbOPry`SB(g4N{N%ET4i9|wor{$sUqa6dFH4%K<?4lC;g
zxSmd1S^YTp<Dv+ILYkG`jLXL&%^P`A)LXVSV-%*}{9U#8h!V!8j~im;++VMdVl71f
z%S1dmy8Hk00$yPd6tFd0h^c>HQW^4ak3To>Zw5n%#r==PsC`zLC0j>aHB!Qt&EP^4
zExU56Napg0BtceaI599+%rDxqN99}}EPNuGb^XZ)vtiG>$+4$0Xd_Idla%^{3Z|+c
zap6l^=rkeKD#-f!PHUzV+O4ic>K&P%zpUpdf4PJ)q5iA-R#)q1M8&hBw@hzUGv!_%
zttTFf_w(-Rf|&Z1?O=70WYYrBY3?!qWLlk={Z;R|q>Kp$wh=#WuB&XV?ILmYS}W5X
z(*b5dS4WaBe&#9L__FynUFU0wztrI%uoc4=H)?`S`E>)U@*~apcW<vBaTMcpo3?8d
z%sHRrJs$+BeG+|}Up^P)TzmftIX~Phxfpxjj@5R)r)l{a%fv2p;LxK;Fw0swRtA%8
z88MuO3Q)S}Xk_4vRIleONQTa{>-vHNYrwg@JF;N%Y1@&|c4NeO|3UDUEaEY9lje`U
ztv3TG%p8M;oj)-iK50cyuG*oSy#IZ2Rpu&U#`o5bT3NS5Uf~S5!+*5OcO4V%DAbqS
z>xYZ}gzK$3`2@71)uKM>GO1(Sx%d>H>k)9l)H5LYpDLEnJAj(lZ$!B9x`ZKbvHdKf
z`n&wt0$e`cT!x*bS(BOF&&x<>{H1aha=UlR&YjOeabN$aBQII7(r(G1{Ix$DZ{gFe
z3jbV_h!oMAeD__dS!b<seQ?MazczMvcZ+<N{-rauPzo)zG}vFPz?tvHF#gb&Bz!hY
zB+;Bx>h`o{f1_sp%eG?}yFSpDBZz@<V`hmdSjM}hFOdP^OHA@NzjP8I`m?t$?CB8F
z+n0EX1U1a5K|oIjZ?wxIc4yo18vQR$mU;PK{Mrw0_4;>TO77-rfno8izlP<vOC1?)
zUjC?zfP8=Esukc{<L(qOrBreJycwgK{96@^E);|D?LrtOYfm;1xAClmqN2-2-QT>c
z7-DucS<F!_AZJl4bT2kUzI?Tl9%%T`HQZAr&oFV0D!lpXu0S%)!rBN35GTQ;UoaLC
zBh7Vx<3qSW+fSpzk9-y)#y}>cSZQOt2l6pl_N$Zr4Vf2_P~c0B<udj6#IvACbE<)o
zQe#ARZ>j|03hD2wgMYb(&*(YIjv2ph^w*SXU=naQXh0{R+yOm<J2OGg_lcg4?v}#H
z7Rw2%70_rMj#E=Tg6br5WKGL=yR;OfT}t%Zm{e@?5lQ%_H%|3GUWZa2x>eSVsn@YD
zQ=tnrRA!HVC$2D>VDg4kil6xs$qPJ0U?}!a-s+Ta{0a4}>uge-%-+R(DgEJ$-?Amj
zlL-(BRXVEV=qX~43zyd<12n`8Dh{&v+0NwgWzbxSW4M-GG(H&Kc%v-4Tx-?5-t!_{
z8hKwf!=)?evR6t|U0gUd{2)#BJ~mR)icZzrFu59bDb_i6K`&D7%PI+()I2Ze1AX0>
zlKXG>v#OVKsx#tl>SvKXOxoTphOzz0)=Y%7tXHQ{wx|D6FM2c^36jnSxE}}h1z34A
z5U$UaNA5bx&iVDft$Cg|_*RXt<#s^gw?ihX>DX%4>ZzCz3fkoUTqKDVl1vB1iR2UI
zvx9~|QB)qyLM7|?b}Mv@<X~Jb3+5)!<8sxQmp8dZ0k^F@g_|IaiTfQUV9@WO>DmQy
zYlq%_ynK(0-F#gyaGGA0O!Vo9Z0PX6I)y&Pe?RbGuKIV`_`l6n*Q?t{8Uc0HCTGUn
zc(|p_<`Diq?J0boNI-f;ef|4Z7t`2QR{C@>k@_QD*e8v}6W~u3h#n5nXfD!$S^Ub9
z2pjghC;y*&0%oS4>}xD;6Qb(&@R6Td6T=q?{9qd;PR&%Y>ej@=HK3e|j;YDXRIJGx
z+%Ijm`zr;M;J(vfo5c)c#SQ#>eM@CRIU9^KDi5yojF<_Kg{RV_ElirEM3DO_Fv-xK
zR=A`3iWuP`&#oF!(6fjJ>op-oeV!cE%0pt;GW|Z7ti@==J9<53vc9=pHM<=Nl^4mB
zB$cU#duV0K!}S->$t{w$NEo`1x*DcRj)!#~7$>)YU}(u0roO;u+?qdZ_}Gysm<Tme
z`a6LkP9jEnguLoSI)wFW*X!YYZZAW})b|)h(PGG4g{EQ>{@PH<#0;_KTBmK+laAAP
zPXCTWGZM;yvE3kIB?a~5_b~Nrr2Tx(7J9qTiju@I0-!^mWa#(h48RN-7~xBT8H_#K
zL}*>EZgZrimkPeoKK*0Cg|Qyr<)1@z>f8oC^a6sfI-Xo#)rL6ky<!$seeIPcI>(dz
zb&C&uX^s{6vHZ4|jqCj#XwEjxwu`aPbIL6|j$sgEoFV*Tj<_#{kr9-^gi$R^Y{NvL
z_xCKoVEj2UAt7|}wJ7iOg!)@T$A1z6p;YwPl=#?bwb~f;MefJ3*M1wt=gxC|uAVQw
zE#A|Z+$Q8~Mp?p?{)dLQm`J`U5I9G-mtDhxud#~|7{Um)@>zqBY)BP#I3iTM=HxCS
zWpvnDSnU#TWL+^-7PIF{_D<U=GH6jSGc(h>@ndQM%I$K1vJaXS2?F{D@*W-ZYPb{l
zSO898UWul?V&xk6HtiN=97Erytwi~{Zw&S8XuQ|@6@s7jn3bdt$(J8tb!r)=I%$Xc
z11Rb=Lma6e=$T(|={?~0eG~m)t|t6w8K!E=$dyx36^7G@u1i9OK6?48hDNbJIwmtd
z6m;Yl=Ty)q;?%VU_b-dS#zelI{+G)kte)qwuDqK@L2E++rjDfZj?BDQa<}*mHW&Jp
z%1o5C7nK*=F(g*D-?7ym^c2Jm)R9W=zlTYE^i{alc%Bkt%k0akHbReK<j$z>&H2XN
zhiCt4^r9nV;Z-Q_Y_(#foG^OGpMbT1C8CX9G*y6%p77jD?up&4h_+Jm195-B!bgJ(
zis{l)Ua^9*=i^FKZ7SCAzQ#{F9uDdnZO?JfCt{;)0s}iKDbd;8Pz^Zg=8OO0q)(mm
zm75{?1TzWfe$i{rM$F$Rq4Op>+c)`9$p!tZR;D?0f3Lp?J*d>B4^ul5)lS=yWCjrE
z6q*WzZg`?TV*+`BdiW$(Fx&^&!V;3wnC;Yh9hPHQ{AFsm^yxf6Oc`yq^@toJvj?fA
z*>4Ip#A-gj`goKx^Pr(OLsk(ghDKuo$7ho}hzV!5?C)VB(PLy_O1re4Oza@A8qt$3
zog=7zOZy3U?uq}$*_($$-G=YOm1JK^p=>RpWGh=3ODUcxq_SlRWnZ%IqiLbCr4nUJ
zl4Q@mj;u+H?E4tfVC?JIX5MR*XL-Nh_xK&Z_a8?N&%<Z#&vjq-d7amJoj2#fQ~Kx2
z1x1fxT5mnBR3)xDQdMZhNIR%IJ3CKDEnc_OWoub_TyeDT+NEp>Lh(FwUs#pU&G8Bv
zO5_a*;B68}8w*;O9GTyb<8G;HUbM)s_PJ<rV9JVJDSus0OY!2!D+0&FMz^t-r6uJw
zg^7WdV%IukaBW8dM4E;}Z{86SBL0|Kn6i#4Oc=Q&oTH-uHdJdg@LDPO1kc-~Jf*x$
z3C^(LJL=fR3rY-ccv$NG{f=9JciFZ}yT`LNI%Hq(e~|q&5W%L0sxNwZMBJGx{$p23
zv2HM_eDPx6_(olRo>SF{_2awt%2HHKIWN!s4ePfFoD0zZGv2Wa#is0zuRLhJ`#0#q
zvu3ZUb=WGo$Q(8L7V&77yBjGXgvibNet#Y!yL{KJ9&<=q&+jI->t7i8R!&LRo_mX%
zAHT<Qsl3n7JE=wJ%1T*VA`p5Jx^x)bf$9?W6?C)9yF{Z$_FdhUPo8AXRo(EC9vPH+
zt`%WNq2ol4S>$wgCaTc%R`{2i7Qsx-Ga=7610g?;aaO(Gi>$E!$!z!<l3z}qK<oXJ
zZ(~FduE|D&TEf|)(0O9!K707{?E(Kq+Rv6hub)(w&(xE-9Ln&0m?LDp%k!j1V|d4T
zySE48U-L=yff%oRZ6Zylc1igrXg5Dn;LeTd$P1?VvVfB!i)c!u%e(BJH0sW<{P4;k
zIf@>l`48{$m;z^-=@lr@wNi5n1QoO)8PoJBASezefGDMn;+frZz^g#^V*UH&`$(|3
zk&|>kO7RO_*4T5@Wk2j!^YQ!tV`q`JSP{CzP-TCT=dubq^&1?=UaFknyU}>V&@Ez^
zZwO5$t-k~q@a)UA^&{v`mJ#D~X$@E3Sv+LivG?GCd$PYkHm42PXQQg?Q+wHO#)EpG
z%)8Cs;a>wq7r9mW6|i59I+O!@l?dDA2ed5vPkg$OrpmxB{1)yP9qaXBjb~Y%yFYuG
zU&p;A#8<D&vg1b&aP29QXLQ9M`EalwTcr&Yw|SUg^=Z`ndz`#XhGG6wA>H&7d8cAe
zAaeP=AtAgu>wHwis7Tv74&_FlKWmR)+EaC7x>tFIt}FUWQGl^)kCc)wLk8>!(Lg#n
zx<ydLwSo+SNzZq=UkjjdQp7Pe9hf>j;ch#FMW#*}&|>&`UM1=t?49%-oS6A#Ssy(2
z@52v<mMl2__bLS^gpe2)tk+C5x{+}7>C^dLWUwd7v*<qc|NotsmU%@05crFd!l+>Q
zSK(60cesT0>kteAT=vraCWiNb8JU`QuR$|X4K0aRHOiPevOjSaw|{hj96mo)T@~om
zcI?>hqX>EDa+e_HG2CbaT;u)_U*OQ!ny50!eHfusrV;Y>32Qvn2!1u>dD<%xm*~JT
zSIzLTQ}bWyLg+CYvX(%cw61@`A31t-p>ZkoW??4(#d)XaP?VmBZMjJJ5cB|Wiik&P
zR@Vv7WCH97VO&r1s`K;N$zGQ43-x>f?|X8s+CaTVRZ+jXgIm#FeX$fcJQzt+Y7^)E
z2P-~227!x<zL8D`=zrf9vf!p$NATM{oaNp(`RPSq?rdy@@{vz#%j=6@lUB<)`^%R(
zz@zJp=oF~|Nq6o-7&r_X2;TrpfU}7l|NP#2l`0ew!oZGQYZSnUv@H5PK32K9JDnZ8
z2Tx=bPP{vi2ewrLh}^R89kKP{d^3^ppwDdsrT&xJlUM;kCRZ|`IUP7rZ?xIJ7fPME
z2dA*6(ag5oniD2kkK9Oa>!Z{xaQY&<I3w6bMn+PhH}Tng{LfIXkJF7Nl4^YN7gmAL
z*#)D?Kk&qaK)Fubtx^?Qu>u^^L+|M^M%1rZ=y5{XK<xU9m{uS_q<Cdb@AK_TPeVf^
z1J4Ajyp>%tiel@u<7ct$imnyoT*(?(7`cG!@g&dm<OTs0P=(9PC5={#DH4YtYJ69r
zPM0bKX&Wu(O$?x{Ehzo_t2XYS@7A~6Ta)8fL5{}jK};0YAH5aswM~>ys?q_Qg<W7Y
z(#oVH`?od7NBJY(w(W_2Ooy)hGCzJ&z+aVObYOf=4y;K_JV%`Vw%xk?A!RfR8~v?f
zeoyk{m}dW;;+YEYjN80<^6;JBJg|l1_-33TkleU%BKb1g`b>XALyapo41FOAj1wiW
z$SX3KxE~Vv=2-2_s0ZH+Rxb&4=S9}YyMOo&mUSWf?Ob$@DipZ~Q+zo$bM71i+uqlU
zSSp*IZ$Q#C^-_FvP6()c2ykt5JuaZ`9jcTag)}aBmdPuBbaMbMMT$M!==|eb<J74r
z8!Rzmf55G9UA;C_{Q9Ty3BD_Q98~|7^m0<BlInVJ-@YCPbP^Hco6b$COOC#-e8MI;
z23@w4V``*Lti5rD^^bRXaBaE}poU7XpAR*<-U8_U0Lt(^h@c-k0iX&XBs^umgpiw$
zE&`(8;Fl;SQL1VH2C+a+r@v7@_B{S~kloDcCo%tgivs6dAE76Q{F$wdJ%Q4t43jzM
z+hYsvG_rp($o<gCT9wr}<s3o(4S~9`>P($={fR!mVzO_cS~&0X^w-i3hGwdko{}0T
zeins<+<B!h^zGvK(4GoPHLkOSk+f+gGI21PwbCl-y#;pyNsmeW@$&lI@yNO?<5E>4
zYIkVGAsIoyp#lUy`^AhNv1jxc7GA_NmnG0u+BRQ%tGv@^_=(_07?o(nmmIzLR2!7U
z{VuuvJ1nOfm0czdMqTSnd(9Zf^L^NAX_N49<E*jQa=+ZRa=XO+6eYidD)NF}mdSxZ
z<3D2$HoUvdl4m!*OJG>>{C1t`Rn9kieoSm0IaJya6!D}Eh5R>*G`w3IXApb9XUjGw
z?ssjHvdO8NL#a*VK>y#Y^H~y%qvBhCeV?Kx!)+#GK_)zp*mS_BX>88Jcr2o9wf}zZ
zL@n4ejPxIrD|g$OlPBB?GhARZ^kF#8SLcmzFFDB@d}H?)`ZH`ITtD%j<FZLK5yOYp
z_DqfnU42=Z+TlsAP)~{SyNDK!&G0Su`9+~FH!F}*JnI`4mlW{|5frDzSm8HLlbxz(
z1MmKXal5mdW4}cszfsfM^ArKNg0eE@qz=wKueAl5P_swh@@LNp!A!vi182R5bzocZ
zB<%@B`JYw#aZKn;c!VM(a*lW?art-JkJN>locsuMPE(wGlzz6m)bSr-$s+97B;q~8
zNJ|=A&n1F_=XC6KSq~WDbNmG7`g>5LT?O{i4SPnHF5h+q2Rhb`-Kqm98ZVT=aDpzU
zuzeV>g5esz=fXqtX{Q@VqZ~L|PCdEc!5R+`>-CEW3T3S+j9fT!tfJ|u%DIq1CtZdY
zs;kA(4bJ<^JrTuRqe<=ojpfZV5ymBMbH0~;E_8t=RSV|Bcg^`eHyW8}D)(7>M=4|~
z{8mrERmLUfB?*+0=ov%%NkRv;!P{Jejr59uxMzH6k93b;mExRn1?5Z<Nb%L@7eI&c
z7@O5$2x_kPNRT1vOj&^m(SB4H#LL}M4P$-ttQ%>sOexbV8uz_`*uN0;A750|VW@4n
z`CFp5B&nAH6N6uE!%h0?VhE)ItO>0r?)8)3K4`x-mC^@S3H0D|_8f}c<Tv{>Nrq-$
zPYlzS9Uab?x%Fgax^<PwZu~3EZRj4>-4{P&RDS+qubYb=H+%0zGFxWx)1}A<q?CeF
zp=Lo#B=<$}NS{=`<mQ(bty42Kjh?J-LF?fYIKtBc!ui@>&*bal611k(9m!%-gXg-m
z5}v)cItIh^u7R_{Wsxm=|Di{el^;=h>c02xQ3l%Btscey-XXY)%9@YeGPExg?M;k$
z_DmTj>ymnMH(B~FCn%QldrQGp*|$%$t%HEi$jkD{YidwuAh__rI_8_SZoq6PWzZ%M
zhD|*6*x#Wkg5>%4y15T#C*Lo9yV1rA=hw?LYt#c5{{^#1?Z$V^dMD#xh!@L#YcFz}
z1yd0A-Gu(s;<r;-Enp0>@rx787EWK2cmoai>%DnGN}2g_XPxiI2avoeBQISp7l&`c
zgrJnuU_Z;X?b`H!wMqxXnk@2|MQS+<wF_Ek?s}h~gyr{V{07CA`QkP&6Hawj2eDj(
zvS=Kj-t=s-4h4f;-azvAzGUVdB6JAd3J#m;(Xv5X`;&CmjcP~GAYs$VCxGD#7oA2)
zohEQXopzi>RgYBTPkw=5dk(?Y2^C|!EZ(+H-edBa;x#yvjhr<;Da;siygq`M=_?F>
zBkeHZd6p`GWOc3-CV4Hf(t&lkJ^_#S+`&_sU+B8`uSCUa1E_wB!Gd|kyIZ!%u0@k~
z`ZB2C{tKP+pgD~K|L~^;Q<KL(2>i>3Q{Yi@yPzssO;+goUZAMO_c45S?vrk-xI|77
zQfQ>Z=!NqnhFs)L8H;XbQPTqBb#njvKE2_Sva**xt;Nb&SoUx_89z;WaILN_c|gg=
zQBgj9C0vB-$tlr#OHAs)EF&8B@5QYWdbc`lpHKGZ$Wv^jboJ0(vMB)x2+f1R``bWI
zA#atbYBJk{KOuBS{mIP|BkvP;dfglYFR}1ySJg7STpX%cck;|KPPWKZ{9q>-KGA&f
zEyrEOH<_f<EoOHW{)Ni!QNze;&#sT0XC^7)kM89C48M&PTFzApCtkGs2D4OG0h`}0
zbzZ_#It_66<OiPKbf?9j7%;IfDKnOOgbMG44|-wx?FV66z0@T)_Fh`bw?dwWRw?OG
z$qMX)s?czad;!CU+AuWCIG~|RpD+B2@sQ+Ku}n}zY8*&ane)3#%_h9_jH05+r>|2t
zMT{?;_@$rmoh2M}{dKqKv?$y=iz4x!%jma@2ZXM_g&8f4GobjWJqcou&o_QHLE?+}
zky97Wrv(MjLLXUp_OmsxXsRkJ>HQbghvimo_$or-{^~_di0)V2Pi6MM7XF$#Of7=*
z{xizfpzHbT0X0$LMssW(QgEnO!iDB#kx7hFK!uyvlfNW<w7kNm-Y-2B|It>@h(shr
zu4G-Cx5#g?<-qmiqo*N0HW**}I^h(L(wvPxf=J`Z%m?MSZC^n~{LQM*6_hfwa&;9H
z&`#)S3e5+TdO9!;eewS;_dVW?GtlJM)KsT9Q|{AK*0t>r^uH`Cq=*yT=XxARu+rSy
zSUk7xq48XRp~Q$<&NEaFyt-|i&b@xD|19zoi<Wk#V)}lo>reNS1EXUv(>VvUj2~YZ
z|Niv2iP=Jbt=!Vv%NL)n*7{EUJl!i@6lM9yMm|02;^J64%jh%>Tp&OiGASXBthA!R
zXZk7_N}e>`v^u%+Lh7V|g#dc-bW*uR5L@H0x(LP4-|Q=``il;{+ZZ_HxlRcgcjvt#
z^%sYNHE!Y}HT~6TTd9+#1FA1^IaQzL*PC3etSt(qKgsy;J6p0YRTv#T?IzVco|wuu
zzQSZ@Kd1a~);?JBFzs*Ng7Wp*RvpcG_$|frBmZC`%~nt=6ZdJeIbIM7c=Ej1Z}sza
z=Vxht-=x?N8N@$%`g9*p>q{i`Huh4^=8rHNg?wiT(Q@2C3fP>Q$OwaqJ2{x#@x85+
zP^y`ggI$i`t7Ap?OVXY^%~h9)iaehY%ZPE_FJgGzQ04WRJC}>`q0CiB9e}0U5CwcD
zox>0Ck~c}EH0wp0cB9zSUrpaO?sBU$UKp0ph;P52&d(fvNa`suq2D*e;QWz#OE=+U
z`tGCl`O(CROv$Lua?51#@uxpdC>0CM8Ff%B2Pke7<&UlZw6D-Ah&;~m*MCjLx9p3L
zWPUs(<$AFTD*6e!iv?M&3y<YN7U9wAL$zl{;@%J6DjDWeGo~p&>unP|9IGT|P*5gn
zlS!LTSe$1az!)fGarHPLxu<=dV2R11GSOF<wdh}2$iyRe-3J>I7AC#k`9mSTV_ik^
zaI!CQs1hY<R<---lP7z5=r{jex7>%xXxC@T7F*M^;;(^!0)q&rlZ<$ZgEG?h2#hYX
z-#*36qmt9(;&puJk8%E;7dqg{Ce&LIoKWP4&+?L69%YnU87O?(@Ev@;Bc~OfwT?F7
zJMg3?F1OotG?c4TdhA9YT6xz0ynSioj@g>W#_=A4LBH_Igt*V*_@4cd!#Hww^8t5^
zu;~Rg#n<mRb9uUaN5AfrTN@90I($a&zNpMy&O5@_+cnN>mINdq4J3oY0#=_4``DGO
z(A~j*8{ob2exGL!hmZD;g1G%@=k4e)5p}QUKvj=*<)c`h{b{D$djb8cEqAsDl)!s%
zg?r)m54%I}J%9Mk`4gSFYG>l2iVJUWawfF;bpQs1Fgg$q(PDnlV#>lPJ%~#T@lDsx
zzIajUf^WV8x%C|sru3sNqHqa)=PAK77%+LFNFrJcmvP!arWDkQfB9(x`UGFaBqdl(
zhrR$=>H369k2ES&DSXQH_z*aEwZc@VIw+YIgU?+eNn8`Z?xXrtJ<E2LF7N3(!$`Xb
zVtahP<t2ozPUM@nHb9le)Q#y|HY6La<i~POmeeiWqc3(&-!EXiaBx&`g|zRh*>2l2
zXY~4!FsqQSZCBCka+|`wzqhT>zT#Ekif5*b-M|Fjw2T@cMG9LDHvnSHM&JA(dTQ?*
zDYg=SWR;T9wZpbUBV>@R;(?(%S?(l-0#bcsg%n-*qMf3Z@Bxs_58zS8U@%eA@;fJG
zn5gomdaC)|?1xsQ4RK28({c>U{?*lxnUU`=iWAoGMCV4_d`>;@3z8GG;K!h2lR{~1
z&*#Aj&#yD3$OBL&=Wy8H-l`N`!+%T~j&2co>^*{G7BClTMb#o)!N2bz{(G(Z)yS)C
zCmcj)*uITG!RwG!KB*Rcu~)siUZ<qryWGU$2eES9V!U-aK{B_>;w9fA`<)ZlYtm-m
zmVkFtPRlPc3ep7-t{2BCm%Xstb{x{cXoVodU9!Iuu3kBxF_AOX854V%QcA-e1a0Z+
z8P`)~9oN}=S33_ITz~YYFi}T=M*CZL-gMVRp}VDWN}YL9TS(NWhO8!KTjl!R*#kJb
z^wlC|ntw9jj-3Z6jLOtb=-Y!GPbt+O`1$xHk`@atal@Y+a&mG$+=BVNLvzt^PDb!#
zQ6^m;M$2TakWIh5Y&ft7HESO>``t?h3O*=&XI~UjY_iM~p)C=rQ+pZHS>kSI?tS@+
z^6aD&9CUPqNH&?p>us+}@1(=M#{Y<m(5c?0Iqze^z2Y7p%?7u*K}6cyHEAi_wJRTZ
z#Bc2YtLG;=DWrhq>D`l^IfHyehi`~TGlth;0hexnxc>Or_{069W6!M57aYkAEeK`2
zbmls@3eCa7)HABbRV-t;7hZ}y`dT)%iJnegSz@;H>NLdUc1rklx)(PW`y}9b716#K
zo~bMA8<x{v{!PmX{My`tVZOnM%cnAn%-iLf&yIu`ax80_HuXQx%=$6&rbExPe7Q21
z^YX+mDVx^!*&Vt!AJOo7-Rc(4;7xy~`&L}DcU-@zSueehPB51B*BiS`M=#qIMbKlJ
z#W)nu752L{e#|j+@BJ<${`#kc;vVwEqtQLn863RR0)JsA&BSf|c%^)->nf3NsN>s+
zgRK*XYqN*Il7|PmF~ZkXo)?9}zndzIbSPDtS(Lc*6Jbyu5PUr$ohl=4cf6_ZRDO1!
zV`)bPL%5&&dYr?pGS&6z8$-oCp1+8>3}R*C_BpleQx#GZCILS#ay$|#?EjpE$w<ej
zJ!g_$=aq3DuF3APj5&U=PHHzT>tW@4J9h3;U}yE3K3y~p-#>ePuQIpNRb8z?{22A_
z_ytdaPpn>H6US7d=9GNRS_g1}#vYtEUfx0HXb{6Q3Ygkhyy@5avPh>cul3O3ax}#L
zGODSS8ASNm;|Bt%EmF!zh8cEqNo;E5tJjk%7j9;4Nc8(}MBv5=$QbKsAD#@0D0~n`
zsk9&{BCBj_r6~%^Hy&TF$6o^Zvu?1&va%ak5%~^DG4KWq&yG|_ODRg(k45qMpY1>6
zI5V4-lJM7=J49k8mPNz9U^H8^cdpKILAS|!ht81fj|i8&J9qv5jLK{~jpKP(b|q1}
zd3p$KeeBB8T`+QGp>LOwuj<UdszUT&vfvy_Gu^cH8h)}yR+z8jazb=_=MUucl_O!8
zIbZ2B)-J1jB2&<2sMLT~a-4H2Q=y5#MJ+kZ-E>~>SQl&Qr`lU4jM6OrwRiOy<?3cP
zz=J1nGku`B+?yVO_eL+36fsW~42-WfoSEhB+ON_iKV4)oFq+SnZC8Ni<b0Ai#3|0R
zdWaUUCG*qlt4fwlUPZTpxQDh0a&=TbIm4gXseGyPNPJpE)_c^jaOnvdpSS0X>{^$k
z==tkUYbq?Y8%}Zc6y6Wz;UBTOJn1Jt-!bupey$?NLQQE$$&IfCgKtXc1f}In=gNJ*
zP;-R%+`4e;MTAoLkv`7vSSR$R>cE@#*SvL64T-_y?O7&FrJLWa-F~%wRLc!p&Ook6
zr>Xd{@C0nknHBU5`rh&`#q*m`2KHrZR<0z?*iLnOtdoR#k@9nn9$JM+DhsP6l8BD~
zXXZSQeO_TpHrGcac{5q-l@`8BRCiS@=NAOX)N#G(kO>V?d+b`nI)gzj=NfYSax`h;
zO^^KfMaUo_oZmM%q1bfwn_krwnE}bkd0iIAOJ@q~wu5%}XEs)^=H||?@QT0cC=VG7
z$u=(B7|)@OU(h27(Zq8-y4EE-8f6xrtaK(_|AiAth1iBH%{Q*}!S$wOp@sTAx+{|E
zZ1;Oj-=4m~!pR4b_Q1?Ip6ommcVrK>YQ@tE?41I#E-~O{fXHQ~w#yH0`i?f#SI<;z
znkLCzs26QKeORIIJ_D0UpD^R#$DyUVCtcDE2)zdHk1>_5eSBPwzxNhA9p{m^jk(9p
z-O9T+-K<Hy6_Z|iLm)8Dq@HqwBHfI?QKZm_*pY=ipg13aBO3?EIG(<HK>k&2Gb+@?
zF0<H@_#+&-U)RKYG_2YwqmxJ+2@e7XpCa~M){E+gblJU-df8X&zY*2e1moOvGCrW!
z2WIvZgOT32;p42hXV&q%Fy4bRrsTD*-UrlyK3V0<y+;CN6^UP13VeJeH@rJhcw$Ar
zN1uZVBU7kfZT2^4!0?(E6}5GN_~aUcylKGZSP->XLAQDSLeV!N9(=OHwiX~wlDopd
zv){>Yge$+Lbut$-xI9k@*z88pQppX3`k@nHiC7h8+4sv@3TRCTT86!~-jHPty1MnT
zxy^E?YA}y9O`4iDCu0wm`pEc5(aYvfco{jBX;PIsr(5DrDKhd;TY$nZZnk~87#GN8
zn^@>`q?7+kN&2oADl+X)Ov|Uc<2H~7>H5;syk(Z6Vf1rqy&e{KF#A=UhnMDvYefhC
zs*#&~oK>lRfRwzacfV7;Q_CJ14}y%iz3vS5Xony{*)9cz_u-ka7-;n>I-iRzy@ae!
zKHwFG)iSRq&iK#xx@)cNRpY+ud#fQt<&90=g2~g7dTfOJRNl3LzTxhX`fwIlDc3$M
z^>fntEqQ{2+f<Vb*2uT0;cL+R`#FQCy0wsVluS!WN7|;!Lp==;9m=O6!&1iwd%G}X
z+oJFr2QCeHr<q+mxbOt?FmhQ={5t}(PD5Vs9Q2bw+^su_OLWCpm)@It=uDuFPbr)8
ziFTWy!~M``wzHc)j-DOOp42RTc+l##`)oy5LV#%h@`FtkMn#sg`2uXJ6PkxtZsz-4
zvW?P}GOUkd7J1VNw@e3jpU#vybf0xMK4jEqyK~CzRpy%oZFv-4SFiy?_vduj{hW4e
zKK4R`)Q!C@1I0p-1zMDk@z>E=OaLjj(?B9{Q?(+qWa86oF~(oCLHtRUH_W2gq3{W$
zyt#lN9`aXmAC}}X6^_Lx1dwVh3QJ}i1QXJmxa6Od&sH&95z|_uncjt&IW0Z$xEO}w
zLnk)4AEXGzdzp*0L@Y+8RJt0unylOuL*&9#RuJb2Z|wSD)JxxXcLSREo7Co{1toHB
zr-&b<G<#o`3`I|m0rxJ$l|D<eXk7w<Kbx&OJSTTiaecDwuGS^0TIYK(@SnB3BiJWP
zvy-D*=XAe*)~VS_?y2-F13?1uHJ@FJqji~qgYQ5`&Wp<FdltFO-{;;q`4ev9z8(v=
zCrG&7ngl3t*dF#R?$u4lRA1@mZj((y;<9e%;)orDSN)&TJX=|wS@PiIWwYia_!Yf|
zVz3*Q1ELF51ltZ-cc9srurFXz)vVh*Sme2*bGLg=-JHT%uxtFVQrNbM2z_P~vD~-U
z??p`OY;d*JDoI4$zxA{I^-N7FOaOTuYsX@-+T?_kupM)(SgHB)Mq%tIvZ8kSTa$?m
zb^Hi-6Yu4HWaFWqFSMBZ>~pWEv6R{tV2DG>8luJg=>@&ES61FEt^PIN(=sM+)EXdo
z<xJ2&8MT#z)k_oG{uo~G*rP2RAl7Vf91m&Mw1+xAh$dd?^MFiM5V}Mq8u?*`WIY@j
zu%Xb8tT~5sD(oWj>BQV{$K`i2oC+BbD|2ctrV;ZcL<R(whAAKruGLc2ju{;N{tok|
z)fqvPY$kqpb#4X;lIC0|=MjXBnN3{H%Pv!fm5rZcvo%?Q%wds*o0{HJ9d}pe)B**w
zDvr(}`{)o^XFbT-#JU(3g5PG<`_gKr(#dCD6)1+N!8JRx*I85CmYnmF9jZL>X}=Kr
zuMHJXC5`U+g8_D+<aU)H<k@!*vb7BK*!<4UGjugOG$(!Bhfp+O<~Aw8rt%-4x8o=B
zom<{V+P5|{El0wNaaRpFpk(}moz<J({L|W;Dkfl@Y+FPe4vq;7*|R6y_=APu=SW?{
z)KNbO<Gl*&leuQ4Qw3ZH95aKuKM`!pedpP|nh8jY6dXvFn2%lFXKnA^oy&}<Uxm`m
z$ahEe0ljuIw#eM7gE(m9(s2Byi03Rm<GkOc5_uNJ&U6ho#!2qGl{?J}2R!T!#qx5V
z3VZ@%sl?d_d>e1a@jCLTVILd@wc5+Hzc%Q6&EY=iBM!8e#4Xh-fZq`*|CwjG-sDT3
z2R$<ZWM7sJM``agWsgZ=zUs*)q^Bgktp8eD0yB9XP3B@ZxE)brl>u?z@+BrN%gcDf
zn3F1X%FQQU(nja8S*lN5+eq}B#QH-vO<LC`DZMs-%Y)I}(M4=%4Y6{`8ttJLJ#NAH
zT+IpzMxK-#g(Kas6_5ey^Fv`kW(B%p6bd_=$~&m^45{P6qh`X8qXjl$sg|HWx?Z-8
z7E!oB5^6=V1?cmaiDzI^gt6$0sal<=PKf)t6C1-E=#Ty@mqu2GD><8Z-S+Zba0Azw
zK+Zjm%jy!d++=@=&ch_$w3Vj`NptIjf<DKXG5PjNPH>#|B{q9jw9TS;VLj_}J}dHN
z49?Fb3eI6`M^-W;vN-pIqwCB6MiuiX56{M8O7ssvvNoinMKB-=<&Gf-6Yo{J2Hib4
zsQajLAtbkny^}O<VAjkcqIh|JRo;;)pnKtFH7VBz)vXGL)1B8V)s2B2%!)_9OaFTC
zAWCL}(^#x%y;UWEu4`taga2A+iRJAUNe?X&+7&~~5{qmd%J=Sb62#4W?84abx|{3Y
zij)d8Q1IWpfh(w9cIYcIi7x+PkP5l^G74W1{)+4G$2|nOdD5Dd_ME3x%I^avgxz|-
z?&APIJ}U3rCh40K_NR^5O5fBuHeNCzWR^O1yEKYa_u6Igu$%hLi4)RuJIh7(#9D|i
zfcFn`lqu5tvDrsRzDFDO=i4oVgID%U#kv4<zZ3&XAYL+w;l)EE3gmrm_$X6VRGySW
zIugXa&%8Um$L{r=S9Fl!k@_Azclvt02LW^R&z#CC(_ug|`bSr`YKR#VhALoAyl_-u
ze%NGp{77t*+)`~q0I?c?Q1A%;ToMN5Wz5*s?~RYc4LxAV^k<j!IZjv0T)t95$V{Vl
zF&`=)JhkVqk@dq!O(~yR%^+vv(|6k^Ivn46&6KZXKB^*=o4*}<`+%9AaxholvtvAa
zF^i8h8-#;!u*)xNX<i$?#iFpNEi&8wvF|RUOI-^_P-(S690pcB?W6M~EORlHoRwjy
zPUsP%3&O0~0!%YMe-pwz;PaaaDCE_@Er=S3;IenG&^31N=zY^^U==4m5+WhJ_#`gN
zqVOxHro*DVxJRWT{63Y9%Z^Lm&WQJ<nkL=rzM*G)KyjrQli#cJE}_2V)}-7a^tE<h
zhtCZH`72;7lG!bOi-&K1p9)E7b6kr}!7aICW>A-<&l>-Qbxyu*sms)q;jy4MTJ1si
zfZVjzD+5a#o9Mdrw7=FJ=-?7O8zPgR#}t^5!My~U2lMaqV-LiLl%7~>kZb9^GF04A
zEjo-R0VLv8T<^h1Z<ccs_VZ=loyzYD1SFV-!|S%bNjc~hW*3Z(8uoliH0T~XfL5uU
zQPU94xzHA*Taz<foj$B$aY*69QN>E3>HMCR9F+TQe?`OcCGsjJp!599-1=IJ9%Hg8
zynD0JNIgfXhY4)k_;>M=@=&qX?1$rkc%pXg!tBBr^ByUb*$2l8w&&ydA#c(FJ*HDL
zzShz171V(T-kJ<~cTfpFyx5sazK+ZLq$lRZU7yL(pHW!%o<MB>A^2!j&(NWJRLjWw
zQqjoR3Q1qkY2ANwCV-((>+A_cfE7XMJZHB)fS7#Evm#{yJ>fD^mnLq`7)jA1^qRas
zCRVz(vW?39eJbd}sZh9YenE(KY})9NP~R~MO3ORNTCM$0?a|43Hke1$f~H;AWR7Lr
z!dk#1J>*=Ihe-~VHDhm+BoqgQwf;5CDRE`JbZm?~rbFIE9*hR|;0bVru!<#bVsQ{>
zm!kcbd-{ZKz|L@<gY90<j5dX|rZV|aVWr>o_pYFhx6odVb#11@Zv2voiBBRH?|<+Z
zN>p89cfF!CClWc4no?E1GRTEoN;8#WlB?gW9WvS2u)z!0-Mh8X|CGhvdhuz3mb1fY
zM%nBFqPHhFCwFv8LD7G8XB?+uoadl-wG(n7#;m%4>qNy|9alA2>ko)O<Dm&B1SfvA
zE^OD)==kOPGeYhB<V!DjlbuXo>vlMA$Ull$cl7jk&eTm@_PR2;gNtb-HdwU8+95P&
zWZKws;D+z5^Ne?7*2;fqc1@!<M#XR%rfWJ=-`jm_pAqR#rE4>A)VB~o9?&oD_fmC8
zHa@Gn$r&J&oeycmy9a<!m6hVcYeOcBvxxW{;z3rb^wU2(Ce8SMjD<)ve~d%%XCd~)
zlSwP+uBrtUtb|Q~)1`Uu{g}V}vYhfebn|ofmENhBtIKxt8~b*V!&veY`wyi~aU^eu
zUp_seUaA(eNF2IvnQ-oh0J3m>Oik&~uV{;~-sHJKqM{EG7i5P<56h-{E{|B%e28&0
zls|UN)VUu#^G;v2B1jIJUhH*xzM>=F^&ZE8Z=Bz}`Eu<xkM)$mEVtc$zIIl0mel%6
zBJuU!0MB|%F>y8u6;x5PR=2X4Q8;M0US_kX9N~dpdC%h&wx9&>Ip48%&R08EZ{u<(
zxkiD-kLK&Vr%@+9sbif4)AgwxN{ag}$Ly+^WYQICtRaBx{%^KYv5DZI{*wqCmYC4N
zK$l_y^PW{BdPdyS2#<~IL;|%ZA>;J8&vW2)3b5%^n57aECAY!XI8xgor6>6jp5@KB
z&|R1tuDrGav$2ajgnu>(KWq}gt1edzfItzoyE7F`sbxwIco8xRZI2X;gvmR$Yuk9J
zrx!MJzI=>JKO4|GeScuub8Pc;J6maO-H^6?Vm<2g?)V_RB$FKux69472?SHJPv4#I
z7DKX5>!G+3&dWS&PaSi7@osHbx(ZfXxO|}ga;?k0#n@XL_Hs&*@yNJ<g|*n8D4LQL
zfpRgE70|N;T7)YUu&ZAgIY9N|^v^?78acVR7QIr`Fh6S2qYKeOM+}pn{^(vfZRroM
zx;OyX>L7q%(oCZTrOjUcrRN4!jSs{$@jNgD+OFrLMk_6XC6>&eh}ZPn6QU_M{+{*0
z^-D?|pBXPRPbL|oIa|UL`x6H`EuM)Yqii}i%Z&@WPfICo99KMiBm<LlEbQ!2V4DSM
zXzu-vJHiV$bqCJ`s;+uWEK1q6e20mA;(Q*&Q;$>1uh40YMlr&Q)82^}?r-z^DqY#p
zrE2kM#7%LJRg=44sM&ZrgdtH36UD=3{P6NPI3)hc3-b_rilrL%%p>mj@E$!v!7|G&
z)}*|O>cThJe+o7^!12@l&2bbQpR$GHl`y8MctD}1-f65}CO(9$M|Ua;T%*l?y{2Jc
zEfwx!C|Tx@?mcX8+*|3&wLYabv`TJPIEJ{pc8RB4RxQwb{g794#1{Gp3wssw|3V+v
z4U|3cLVDOiO2zm#vJ&W1o=)Hlu0}0Quw+>14}*6PlM=8Es_u#o`OBI}-UrQCVx5fb
z&JQg9V=8fsS4-7~yw4glo&k3(MQ*Pll>3!FCMRYd{%{l?F|`6&E45fFTHGw!-;hOQ
z?9Py^<KbQt0Lo^N-KL6MeUoH7{3_2go~nQAoRY=q6P93b`tslv+B2EB*=m+7_q?!t
zd4L6hBB?J{E^i=0336b$DL*hc<iEwO)QTam5(Esyp0zcs?V0dxyCTzZUMt7d<xm#t
zttI1a0Vdg6Z!YT4Lm4Ocq?&UE{7f6Dbj($#e>Mfyft<?#rLL>s=2kW$)OZaD7y6q|
z5h`yRX_O*Q7ja-^R7_8w;8nrcpj*?`HYT5tYcPcoEUb?Snm~$m9iMvH8&=CKKdHUW
z!&}`=DLgV%7Nxq3By1mhAhX^tk%Yb>`{p!V91bDNy~w}H6%A+%OPLV$xvVOfITX`e
zZg$Mffu-gfeqS4}s`1?LbrSTy1G{FS9uq}0p3H00)F_F)(3QNYa5y-~I;8)vB${xM
zyq5F%Shlo`0>gLZrH}tm63TCt;9%Z#mN~Vxty#U;y{4BLi_s`Ry`(VJ1G_M%!UmLt
zE9QP3Vf0>G|4*TI>S?8`v{jx-%7o9(8RYJG`Kej%<lXUPcE^Of+FqpQzC*m+D7^jP
zHB|rrTYnGYB?{sNUE<p{P+XE<=(@!x2(ZfYJ3@(*bYCP#UYD`73CwRA_Dsxw;{1fb
z$GX~{tMZ*GFKqmwimd>Cy`w}nU|hqLo<TTUxod<05iSst7!HKQf_wF$f!#FooL*t2
zm_{VQJ;pJG%_>bsC5s9}MeyF3tG6=_hwkQgT@Cxy1<u}33kNtr%<2wW*?aq^F_cVc
z^YN9jR_B^-PUzVYQa$VBnfBL1x%7dTsvY!C=hHIUH@*E1*MJ4jp3+-@ns}pU!J?I(
zhWoB^*HS1PBvcaY#%@0fR(c&hTRE%3^WT^S9t>>iaJ(hcaK25BJePhEPh59e?y=H3
zvy^S;vp3qlK5Nj&{OsiCMo|X9%PxKE8a~xqXKHp0_@2ctzWPEKu4Y<>20=D>GPcWU
z%h2BvfG|L{(KAtqmwXyAzvbKnrZz1le}|?}-gkQg3IOVNR=tp5@zzHkd-gl$^wA{M
zGT9K9hBsqY*7hHn<pzq^<n;VmCcowSAe9<SoQH|Qp7;^D;gy%fTj44!3X+jDeQpa1
zhY1TmzHep)Y}!rv4$UA}O3k8{*05jtWVtf8-zck>7I0dWn)-imTB-E^05O*8fPa<$
zt=VM1rX)>9hV;*M+2l#M*N484r@-k~Z5qc8QeT1-b3p&PJWo^kGzP=qDzeJ=U_~=V
zP72{MPK}%9<C7&Oj5?r!10uvd@-ds_A95A`tWW&;aHM*Y0liXB6Dqfb_7W;xpSS%y
z^f%h`*zM-+sMeT&hr(qEN9M6me)3iQFbU;9Da8KZtO~q%(F3Ki15hU6$`+a#fJL>@
z1MK(p^d4>*U{tb)!Eth*8@KpAFJAwn1ZpRmOgX6d?E>1d4X3hD__E0ra5iFzaV)WU
z%p%6=<<2I~XOM&hYDUi7vx8GIga<sKiNgi*e04NDHJAc1<XkY%BZZrBU7UXfqb6Ty
z%AK5haL8O7|4fCMj`E)l&D!jVr`Z#V>fzfQQ^Td_Hmh^{4c>adR0S6{ZDw~G+A#(4
zjd&UCjxWt{V%#0dBQITZm#nOvd!u|}6S1c>O!il_;p5yJR!!X}7i!qFxI2!N#Z@}W
z+lAnjNQxl?!F>wkxfN4J*?P^Y0mRQF4pf83Mm#h@3|Da7$Nh%#fpqO`UiDjO>SA{4
z>#BQZ!icEHBw+!KJ(<SdesRTiW`)Re`|Y~dArv09f^qLM)?;82Fve8A8z~unz|;w*
zP9<hvp?4x(v(o)_rRGedFk3S~-fXA^1$zSy&19jT46h%VS-<kIweI4;+V21bh;=2Y
zdZFdmUbNoj*t0py#n|a}xiE>@dns^M3-}FD@mY`Lm3!I%d5wBap`3eHf;}o!nEJW+
z`1hjwpt>=^9phX5I)t|?<!WB@le~%f=U}P{>h#+rm^yi$$`t&Ql~~7eAlrq1xuSM&
z2kPZ+3Nw36PCf7-g%3bWTwShQh}QBGNYCp+nf>w}RAMA=*ZoQKpIhT;0&At$s%GtG
zHe89-yz*8iWj-i8UOu+egwu1P&*>3OfVUyB@wF0HzsBHljQdcr$t<xa#*tQrPrP7i
ztr<Q1$!yf0g(<gr#e^eN6@~vK%Tl}F%W>Vh7`2_F52NmhV(6Y3=^WZQ5T|iXo$ClZ
zIVyx2#z3w!>`|Jnq^h=jdL(<m6I(u_bacj}=XC!Wa8JLBQCN8ht)^oXg8xXMn8gkk
zYg$2%)4>DU97cD)WPcM%YcE06DG|fTtICxh`|^ttluf8LWk+AK$`hIVYu{dGREQcs
zprSli&R2ag0LUU+7L_7kQE3X#&dRb<c~cJ2sf6yO!WeOT9W?Q6V9_f|^@0O{#ylq>
zl_^$XJ}Vi>W08F~DO8$1Sorf_T1rvoZ%<^N^k5C)xgCT=*16X%)#lf>SZ9?@6vO5S
z&c_frZ-<gSKbkR0XJl8=o24edqCfqTr;AHzrftKI;x}x|(o;0?Da{tmz~?vn<;z+v
zy6y2oooLa(!=M-ODZ1c4w2)8@?tF-JiC>58(yuc7P>UgI<*6sdc+1!|4-fk6%u$Ui
z%#BtGJBW^{S(pQkTDt#kHB)tSBa|b$>x!S7<p}M~m;D0nY#!X-#U}s36ezF!6{0C=
zKB^Z4FH*BToM=A(iE9b;l+$!#CPBB;rULym8>`DZMM7Lbd*B^c)`U7!jFb4g>S=T+
zMIH{St&B}Js1(8~RnRMpn**HKzPW^>$Yqw;*ThXx5?cTbV48ty-p+mKtQt%gas4b3
z9gL;31S@3zm&Ikj01MGkTG4=ps<M|dfJVbn%0)`0F52aA6_VLDu}O=NXM%4}eAohj
zr?nh+>WV-lAB|L)Evs5tW(s!td@<T_4Xk;v-h?K@u@%_Ri&v>v$C$Cq8Ufy=;+vNx
zm`IZyR`K*QQ>8bFU2QX@@HK<-3F>(HIN$&9yix0&r^t+yzOC1<K)XHU+_EVdLghle
z=O$9$JXefao<&SyG3xGS^OWap{mBLL!B^zzc(XbKE3QTR{a(;=Xjyq&>8hK+S+{+l
zT)znr!L|!+TM7`t_CTR2B~6cAKkOmW4=A8IJYZwQOeMW=&?kR7Ij1j|gHp3^w=3Wu
zEya~k8nBylOy6`YzSDj9^lpE!01i568J)QBW#H0wC#HpMCkh`g9`g{kSSkaB>q|j?
zWF%%%P><&M*H9qI&DPPEE%)bAXgDUZ{(zZ}_k~NQ192nRKAG>5>)a2NV&;K)ZcSBU
z4%bAWr*_rNy~4WBz}n3)%s}tkYlJU=h3vD|Fk9m!%77VIt2Z0zU_fYX5!+fw&iIX{
zI}S~oIZo5OB#TCeDmN99#!nyI`0oXLZc22~_ee3W3d#L0H@5mqL+qhE(n~*jBZaH)
zOeROX<JRJ7{{@_2SY+_u{hD6ZmeFD+;V5=pz~xo1nya>Oxz;YQ*K_EGJRan(85UP9
z3L7nZx`eQ7bAv+7Y-btDKSo)L#@+>2L3=~iUZPPr_wreY-D+@nG!47b`@4Y&K>sH|
z#^YkyQ+p3<8^Su?-*ZjF_oznRX-X$6_d5!&6`~{>VldKn<g(1b;f=i1Yt`5EOb<FL
z%@38-Mq5PGo-cC>yU4+{ywD&+KC)qh9_SSv`a~1L5}cm<C0k08zPpnF5d;NnTNhcH
z(#&GRD2&+WKGNb_SrEsx^6`3L+PX|{F!-?e{{{jhNF`YuQGC5Bm|31J+!T4Wh@i$+
zBK^-I)zp7`^mmvz(2vdc3q&_EoP1mAdZf|v2jwR%YACpKLdtKY6w5@0#*&2nfzs&`
zyR#D?ld34>MSaKXH~(I!-L(7l-u^(l&Vc5~3@l|xaivAc`7XbM@vS*5(1BA$Q~M5U
zAA$v9tqSCF<V3FbVZ($bK-5%i8RMY)#=P$@C+b&*wyCpYN#S8{7U9+bmbps--hSo4
z&X{!?U<rHrY(sOupV9Dwn?Y_GzluqSlA7JfiSuC6aLjR!#`#y1y~aDN66gYW(E$U~
z%~Tf+Y_c9ynR7Wx3%}o3P9xL5Kzf|_nrvm**ERhh!B8+=+ygDXfSBR)z$i_ZOVnuI
zK4$7pg*oz2hBW7wK{fX^tR<BF^Y<<fW(}u>d^2CC&SW{Q^(MJl?xu-HEp6gq7BWDp
zB;0879`d6{2$H_uJ+i(1lxct`@VxiyRTYIyi`FqXLHWtg%R`^iGU+K~SCG4R7@4ab
z_*K4F#vVPf8J#Zw-nSI|xZ4qA4zrKM{Ei?1S3O`b0X<6<HHQ(auECv1bNecxzvKLV
zWv&Y42sEdsMFg=HfLqR0Ev4{L#Olu%)y?uVM&G{~x?`zfl?zJDhTx8X)9?xr{-&no
z)U*gIy;qUT2yx+p9~^;KS}qN_kL}1Qotww@NC=k<s+z66wmc-gEbml)kXMl$g(I29
zcDr3;$SknbUtOmOdX3+#oXuAt8x<`?&po+c?<!hwTYssty1HLORKY&vq*EF1G<j94
z>KI3yye$(@Dsxo<ifh_MQ`wugs9wvgqU8fC-Ado($oH#7*X#W~gJPCf9i!)a?UEcZ
zZAkgl0EI%!5lQ-YXyVW&ZuuDVbobdRinah&{B0;b0{u_xw{f~UIW%I#s_sJ+xgu2F
zNPT-BKNAGS%ffXgyGu7A3FKW21yr+h*Nwei;QC)JS~|I}8EPQ2n~Fe$RRQzQ3%pm!
zC+%**?Kk;5%Ky4}3x>BTWIS^aT$7VM=jN&TsO#{bO@xZ4r@a=>l)J43K&-s%yA414
z>x%M_?89t$ACKtR1VYY-l9OwpUec+S)^^_wqJ(7enX-kj-n*0W8-{VLmL|7%$LFS9
zc!R=o&KlbWXH*|0*t1&31)$6ZUfan>ON(qwh25+cp_ws|S(JB)%Hf&`lJ4feEVId2
zja=+hGNorB&5Yx43m8Y*1yZ;u2OKUYW`isNZtZgM=!Js23n&XhW+KQ#GL`nY%&a!B
zaCxMJ;pV~}$xY&oG{5)?3JOiVW&7bizbZC#eD~cZ?;g1F1+zDC8d)mNd(VO0%kiue
zZDJSxlU}nkhCtdV6YqLXo*V3mk%sBxYAIZ&%1Xo5d!UZ*t8_W}X9*9dz+S-TiNETI
z1l@=6T=DCzHGN0{!W3q>x&%S#f^8q)yN~?$2vdH`gu3)X;88|FuMh&T-2+b%NCG(M
z$S^gEq5Zgy6~ElhL~pT_;yv@;D(sAw*Kl~z;wT+D$XtXIU*bG^PsX}@I~mtdlGeHl
z#&`lK4H~tB!o@$I<AAeh$Ox@;^+-<T@Cw8AqX&>0Uihg(Lrp~EI-VgvKvCQ@+9pg^
z_0bC&nrP{%0p<p!ZuFg%L6omKt;>>17&Ak9p5egk;wddfUIFrW-E6-@0R8D2?kw_9
ziqg$>E7Ad7I*DfK^^Z&F!FMWdZIugXwU~M6IXDGW$RgYaoT^Qo$_hFgWIcpUyrwXZ
zIP>Bi193a2Y_4JHw2T)BhWKI2tg^{My>tDlJVi-YOY2mOZ&E{dm)k+neeseRiKuvJ
z{>^f;eE!PM#&Gm{4|%y!EJ#f-o{zM&$R>hau`vbsUhr{lJIj=Fz}dL+%ksMkY>Uaz
zrmcf-l%0}W?FM4iNKS7k&E0P2dLDOgQP0(>u<b?O+E_aKz2NS@Dhk<R_fB%rz1uA4
z5s><V-$6lmvf1*)_A5lCgR&!+U0M-h;qzWrFXkhFw5jt2KZp8>fQkkO!c>gv;KyY4
zvuk5X)fCrT+wcU^VoG(W%wTk1IRm7pH&8#cTS16zj8v3X#nto1pPfWx2O6GOzBgw!
zeLRYn$H&n%(TPlm6d$)zq6;_nI9DxC5r$~K)>*Gki#Ye|LR)=iXCZc<<H16`cbT2u
z-A-A%Fh88WbaXN{26$q`>yPrO*T~Ml<Lm`zR>NrISGcQ2XqLPF<D0*cTj2PIgo?z4
z7_oBJX(G-nKZtg5?f7iJo%6!pKi!?rds{|jx%85j)nGkIMKO_-_4+2bRdfIW?xVkw
z>fUJ}7z<y_aE~}dv4K;6lTN>L)$PNxkY57D0#+MwluX3!-z3$=XqDjaq3<_WD-r$U
z+23cj0VH~>R6BcE=G%gYXSvUNO&8;<VEk=D2n+SPS$;>Lr``8VZdK9XN-L%Hp1bf?
z!+V=|(r~j5w~^-8_IWjDo&m~$0(fWJJJ8H4@otyjRXUZ&)!nsatc5B}C{+Hh!oyGF
z;siH01aYZpS@u46$}3$NcaWFr2vx&ePP~@U-oI*Uqt}-EaId+tm^^u8Dl*FPW@u<W
zee9HyA5V+8`P})}O|wx}s$y4>g4;I}mpMJcjDHTIC+E#NSJ^p@Y@b)UmKCWtFJu+j
z>U(2Vf@DL}r(VjKu)NSzID+tL@V?X--|M~$lX~^g_ENRzK{j9i<~08j_;<~N?RP0l
zqHTccMon~EgfjsoUJ<7<jz8tgZy)=gfbHFH-98|zczKc{Y|MxV`Wq-K=Rk!Ud_yb^
zpkL*^d$2E$1Cx+Z66Xn1G(+N!{bvOOmCZrK6QC%ZUFn-xnJ6;|%sh@Qfp}p!$HyPK
zXdc4`)bUv4lIicw>__UsJzAf+a)*lS6hmg#AmO8`(ig;@o#k9_=5l&y<y3N{-}Bqt
zk-+$y<!`7xt!`T5W^q~ee3quxxL*reG3=zJ@7Hi!w6pW*@r)1l#0BI`p&klP%Sx5S
zx4Gt771W94r+{veCUcqLAPo7A*PPf3Rch}Sz>f%|K3!-9rJkP`{B4OQSv_McXG_3X
znMCzw%4S}2KraU^B1Oa&@n?rBlYs}CPAGrf24WCd1&w1{@r*hl4h3^Iv@-%J%3#(D
z970wP{<EI6{A@+|c;70h<9K+1sQ$tQt>x0YRdj+6i+_D0dnz}mdQg-DTj@p_4Qsr9
z+{pyKcY}D{a6%soN1OqXoMMt>6u_UmQ0uv;OiruQaeMYVn@xgW?s6F*??Ax>zcXFz
zyD&YGDNsk{+2G=2u#DHR$Wvv@8?LxfHC}p3DKXogk8ZI45u90SMjSFEJn~lB_%T}{
zi$@j`=Fth)50;WB0-D*0Tul3}E+|iwuC{pJ%E^#BP=Cvk(O7J)IJ9b#LF7#LoNV)Y
zXQSAoDBd*<uZ{U&yEwR_W62eLX{F_x$oSdZX7nw6%DD{c+XI%yw=PFJzcRpPOy${_
zeIsOb*vhW{QhKN+Dp$Bn?jjiVI}Ul@a=399Nmh$u;3G|R*%{|py|D<p%F~vxF*OQs
zhDFmbvpti|OEa`p`OnryRWJY=U)iUxOuwr{KINV@?bZ^CpC5wGKLf0=gBK8{GyyKq
ztm~3f!iZNhmThPB?O!3c3#<+;y93CZ)u-#ZC5UZzUKt3)`c4;{0pzh#-3CV5+LT&_
zz4<d8>&G%m+=nYON{NAF+nuH6iC{cu#paZg9w~c=wFoO!HSlj<kYj%EI6~tTe|Nq?
zu#oqCxE?KHWiQ(NTFQa`49=LTePim-;69}&v)%D_5q#U|I<0Pg&H+$x|DIPGTZmPe
zYZr`}r*1kgw*(>9sivfN82^pdwoakQ{Y3!z?FQn46qP3_Efb19IC>h(bY6(gqHX(Q
z-iJDsqwL-*t$WIM;E)M=iG(gKGlWn5$Lj2_M}r?y#2c5vgF8~;gmnDnAqtHMV*NDV
zxkx>egNo3v7v%b=rALQYp{uyx1}*XIv4Vb2Sd)sV%3a%QnhM2HtWe4qA7Lntep6yQ
zcSY`d%rY9vf`jSL;h_qO#xyV~Oit*#93egObGE^sp#QU7TGWStyJjw({H$>nG`}=r
zp-j(*(P|hF;$=9lY%%_n69mN#9|vD__7%JzlmNgOmAlQd_}JOVEWc#YF7TMCMy^lc
z#rImDU{#p|GEeP1BU88^<zT^0q-8ai$5*;eYg~UrO^=BB@KRdmFno?(2O4#@O9VP<
zH*Zba(a%P7W7Dx|F8Yj=t@-jgsJ(b;;rc26qS043ZRLSLk*%gw7?h^f^Xcw_?Pd70
zjjeW3I@;B+)8)Mu;Lq{`GnAeU)2N=(Z%hCFZ(WW7(umQ~r-i@K3*rXeFH5g6egYfN
zw9X3VB(vTEk*_^NcWuXBisrluPI`UM>MkSKEBc-T?yliSW@Fd)MOvbEVUkZKhwlFi
zLQ1*z^YuTgs(c;}0JihR!g{bj@SnKZRkq`KD|G$@$9CeOyaRRMEsef??%lt0*8x^|
znyWu4`d>4k;6B}4g8=}Mjk)ju##8Cp=c&U2MGB5(S2N2$gbqUaw8&|GzUujlUUUeK
zMD<bqq)J!W32Tm8_(-C<<>{~fyx?cofuVV?liBzH(5?5!`%?d+U;fu;Le0agjzZUY
z{B=)R79y9>?hdUsCW_=};3;JT;y@zAMA>n&S`Ur$MZ}RCL{PoaNS4jO$}^;;e^&1s
zSsvMg!XrQGxoV>Dww(i+>!%LS_L~OeZg0KM`v8udG8nNu0)fl^sgD1DeUyK%lPX{e
zc{W%;3Q^`7SE4Zg^%(A2!rJv{gtyDUC(*LmDjaAKXervM(`k+!sdN<Lm?b67!?1de
zS!rXIB9QtZZQ4mihp-Z{-`KemWo3m@o`Fy3>*T1q{r<8j=S^gucL}BMz@@OV?+<?Z
zJAD4fKc;9w5|=b;2*2T{8brf65lXKo`^z3K$&B`sDY5%JdTmlRuD7ZC2+V@oGjOV)
zPn1GnGT}qa#uCUIc!itS4zbZ8QWC90_+e*VdTFF`*cdj5n?mSzHvIYmC}r<Yf%%gA
zKoi;Xe|)ChKF49bPuVQM%ij830`YKExez-GD>|<GQC=sM>{`U1f=aI)?}WV4K!N;D
zp4+)*Z+-xP-26a0Blubka;Ahv1<9K)D_wEwy4}<a2#x|bc9UC_&C7vl{d@Dml4;)i
z`Md%)Z@-KW+x9#DA1k$;8r9)d>V2M+zH(2Q4o=>i`<+4QPOrx2@7Ab8?h4q{<COQX
z*Ef4~z|e$nf0(Web}$YofpqFPU{`xSaOvBoCfl0fK|qpsm+OF+s8Iub=`GY2gT^2X
zyJ=|HK<-cIY;*CgGQh>Jt2&wgdC&g^3;)^@+Lz$dslOWdK*xH5(p9~GR5M!q$Rw-}
zT_&U1cybgoms6Q@w{=peEe;R#h|&*#(Qi@8lY<mbjfc<{JY(csdw`PkH1jI+I}Ov=
zLr5(_iZ933%8dDEV+)_Y^i<e|8O_Lu8`XpA8=aqXW;^r@p$<(sZWAj>!H(D3^#7M+
z@!us?-UrXhubv701!WQ}BwOcCJ@nd*ZpYH8D)1A2JsajJ<_jup4!AOejor{Xks>23
zaL)?%6`O_n^by#c1_9lYGu7Y;1=4J=>BhUM^?A1km9Fuh%tYYN|3!I3W38m^B4Yjo
z9GJXm=2uGO<*EM9kC#Q+TOy8%@B!aL((Ml%nJUDZFet#q9-9vRqpd=;A)q60=9oqJ
z!r@XZ*GV~Crs0E>F1X74d4X?Fz^ZU0>vZU+OgDFY%r*W$ZAerZh(YlgM>`?&qHe=}
zYKYyF^#(2g2PXdf+2O~>s?emREa@2!@1ETiKwmuswe7lDs8r{P+Roka8dp_gV0S6Q
z{<EUlhLqY&kkr$z;@?f`c;eq(|0cH``Tt)_pj-?eqVa}S=@I3(a_|T|ook>$`1mNE
zqM`ISd%y3ga5kJx4=E*SI-0b0aj@=VJ_91^Q>--Qsm<VjVcxyH02DQ}UP`C;U@GMA
z%&`3Jm+{|N_$Q!IE|mOTTH&E~Px=38@5;ksT*H3Ga4K6VWeLS8SxR;(j3rC96lt|a
zL`B=kK8I{kaU`USC8R}EGNq(r$ySMyHftuuh$3Y9es}ebq-naY?>gTf-*x`#z20e_
z=icx8xtC`=!GVPEH<V*fgGXP95M(PPLvZ_Eey}|cu2gTurFYjgJlyHbP=4uw&qJ-U
z`~g8-fS})62v>eUDvLojR1UF5=MY}J0xToCZKW}y8Os)&;9e6h0sku~)nxPqXZAj%
zAF*ZSn2YHc3<8wl{)!XBHy>T2iC}Ti7B3Ym?S8Q33s4tq+r=?v=BGRMqwa^jsv7E>
z?E1KP+Emcz4zL>3YiKh@$Kh)Y;cNY^cZt&c`wH4fH9$5bA~nvLcNN6m5jrA4uVwI}
zGMp-IF2E33P@89w!E%&_T!4|5*o|jCwO<{%Fzd}h8QvL~`#B+*ADQCB>N%VM`(8(n
zz~5JzJ<gXT5@P<K$}@-pDoTc`Mmw2LQ>Ws_REstUGI=&wTzIUt3NwRJ916QXvS{d|
z%diO^%92=KO9(8m0lcqFV7)}Klw%Lff&>C@PH)g_ilupA54;2)9*trA$E+EM4N9^M
z+ZkjR=(n34C2qkNFIexv#1XJMJbWyIb)@Rcs^#`YY*U&>2zsyysHg76ODE$68nt9V
z<`L%wGtsRXClg$Yd}U;$nP0zF*uZArl^#Gf;{q1TvvX26nTwb2v$NCiQoZbTYsOrk
z=?!x<q_ps?ZK)`~zh}90i5v$?1pNUZl+EC&54cA=unvOc_9+e~c==*1gj98LRL!9V
z#F|IJK&!PyT3H3SIVEm^ROqb%X#W1|DFpp2P`a50XTAykE+Csi%6Qw_+Ms7`tP_J6
zNZ{{x<P^m~op)YmxWb_KJGV0tAWk25uOp~ioCxB)y?b|(dK}m+(An(;yTD(L{Uhbt
zFu{~|48+Gvcb-V~Yy#qMJDxP@<eOqK|2}bdQ3~mSwc2;MZ!YUbv0DVcij2ZMy!@Jd
z{T2gTB9*XIZ)F9I;|fc%3)2{5A!#E{HxQxRqQi2MMS?baLZrG$Cp_HRmpcJRJwDKC
z;q7#r5g<kAwn2MK9MBGXof#l)c^RG)xsQ9U{UNwtaFov<y>^gNY7iEsa~>A0&Up_p
zZq8Uy`M!{$1yce_y-G!gOW#|l)suLJlw#(bK7qwFZ#&_jw~qrMVjIP@U@e0|f$LTy
zAwq|Z{>CBadejOFXEX`Xwl$u@2L*ylv9W`s36Y5%sL^9idM>DO9((zRIlpC3fRq}v
z2mIP+@yc`ga(jkAOruN$+l}j_`9fIxUNRa02-%LKPSXL1TYr!!@u;Hgz6n5Oa7YY(
znM!sSP7og242Rh*0l$YzCak`G8>2;%g_HS$B`MTVhpAY}4$B`-Jl9E9nU>cJIZs*^
zqk}z%9u$v8gmi!i#e}?hb={3C>K%8JSw|uwyrDp#Hi7T*Cw4f#&_ezf%jW?pe)Onw
zO(Gu&xbg7c-WIW6!!5bOuMbj;rEhT$D^!nw?xf?{op{S3ys6lA(1IaUDRa@luHFMJ
z7$2=Zj6BduWiOlUqBO65fksk2&@B*Y&dCQujW=;_V%;g`CV3heWITth5h2sZNg_3|
z?H<P1ii&Zt)EiE#Zk%2a<UMbkfmj?RJuHS%)?MrM##e@S9-by3_WVa9esf4x+l>6N
za1?(50M!#F`x!S@e3U%RRo#2D;tc{$AGm=3kt{;=k#QdoCNm{hr5P)P_FBKg(U0ST
z4PKZ>;xxy(2eiD+2b3};>xv1$pm2D=VwHlR+oqGNBmpjFdl81+b-C-2+2_3{4B{?$
zOyHCMxo|U|9B2w*aB`LklbosbmAYT;;tf=jwhaAEbL;Qe2VAt-F);7XW-Lu*?6qK~
z;||U#Fh1y(z!xZf23lzwcZd7cQX->2C3iLFg?qYl>z=SBZiN2HCxTm9aa<OAC!I@P
zlvo<@_JOlzEr*bu5esiErPb|8jKbCRZGA*(`Z~K7Ge%0XydDRYyUIhWn&!})kY`DA
zT;_64uOP^EW=rq-B>8=#_8@()EiRo1TrEg_^j{Wjc^mMuHrPouhq<$!cbt*<ow~56
z_xc`q^TnN;mi@3;%uY!UoOn?D)tb(`LFGhP9y)x>DwAcEF*-jUn7L`gUT7T7V_@q*
zZgcE4r$u4$264P1Q~XQgMoC>2I|#Hl8{yx?0>r<BmvjsRjp3j>i7ByE6A1J@>hRr$
zTb;R(=5jcDz))|Sk;{8=__wg;?8u}<rH&A-*OO?!I@?rhRHiZ#AJgtW^9@OGeYbq+
zAJTy5@1!-u<3}o~o3w=R51e(co9qd)dhJgv0OEe*@KR0X+FphT`n16|qgT)O!kqR@
z3Y7-G49^D$7@Tyad73X!>ijty{#Hs2fGtWYSwMGBOhT-rKW*9~MjCJFfI@=OKsRo@
zK`(!ouC>xD*<kn$1B{nuUtL4+PQ_ve{|=L#g<wn2e({CLEOs9<CHGSDhSshGJ7pVe
z!0G!`2ev5c^o8+~m#)#hK0UC&p0Rmautm6RK-YG%<93b^gbT_O{0+JCBu!rH6;4`B
z1otCNE}!i?_v{3&aISYC^IoDR*QEiUi-i-Thbr;%DJPmVUS$%jQ5_-@C-{4z%E)nS
z4$LkP<bH@-D@{sG+(4sab27fb>z@JbXZLk}%r1e6!ys$}UOya8In8+J<L4!!!*9P>
zNasO@704aCr!Qe|_b%rl(+e;W+Zvi*J0x3o`l}tR%GymU$gYrT!9}@w>ROIba+=5$
zZZXkioDzLjOVN4{E(jS9=Ve-S1icTzSQwxcp*8(&-4HVU$SRo*+qeiJ*yTeO;@dzP
zUK=1~B!NOzOGE@)H^5qM(qg%~ADyEPE2j)36e)Gn^7K|FtvHJV=OeG|nX2N>jWkbu
z!z4%RfnAMl^{)G7gIwYMvHejb)rD|<l+Wv4H|~&!EwqXH%Tl*iWYs5<=|0`Dldch0
z*m_l)X%YzYbB79?4+N+3wLnac10;&7h@5w1OKRLZrza&SnG^(vZ$>)d-#2?byj$2@
z^C?m{7!3dEF#;Q!Oa);UX5T#{X9qnwEEhHgHlgl8tC|X46oy5GOo>@p5TrCfz*ICP
zQb<92J7_#ce(^!Dl(Bbuu+^{i8nSz9aJPnQ|9w@%wu#cdMGqTPOE<|__z{o4oSZ<h
zM1PXUqbu67TA)U+JLzaYQtGZy4zw_-Fq(z5NxWe|IR_0*Y4N$@;+`ZS`WaixDwCg1
z&5HJ5^@c{DYV9x_Yra5cUw`7YC}l<o#z0DSmp--XXU!G5WM@VGWM#v8!0`t&!)(^C
zg=0tsgN-7CW0fXhEB(bL|G<_)WLW9?eYS=$UzhI(67js}*uR3=3k&v;S~`%u1T}m8
zCz1OA*18rRI9n-oJT_1=-;A?xUB6dNXZilv;OUg%E1LgH8yFLld*AAa*|UjV-~1(s
z^kTEo5l4a7vziS5zrB1<K!0vZXoBmxQu_}zrmqncIId98UeuC$jiSdT1>uJ{R2e%M
z)tNee1URdw8FlK%PywxlIE2_5BYw9X9VTvO7t^Bt!y$;G0p(Gu>m371mt?UALi*FT
zWbyGb71utv>fV~KSPLf46xlhOJ1!-n>cfh3?XrQG2>kn!CFn&iVD>*&vK6EU=`Hj|
z;_Cx}Q<}oV3}*eENaWbuNh$re)|W?N*(UrM)(>A}sUsEs>h>vnClRF)E<?|%Y@3xU
zrS1Vye0DDA4WR2NT8)i)Kt3#c404wiAI?*;hEcnQANYenm9+ePqq<2oKiFiUT8Dp3
z@f|-po8RJZ6+6R@2)qOxw~gTn7e2W8v-z>%<*}*=U?&V4%mS`o5H9Aob9Zf@8j3}y
zw_0RL7>ognT8tkjhVubg=#GCN)ABEEFqlHe$J?;Mx`#JPBmpkw4bKf2-2X7RGv0{n
zP1@*Dac#EKZU-<ZM(Rs2Zg?CMILiJ>RX;_u$GVvX0_O@XEy>QMRvW$`-<SV~n;BWN
z7NEVdLYp!Ceo(T-p(oaZFu1BIIogcLma6;3;t*@uiaRt(lt!7)Mzpu)qn}AkS#+{L
zsfpD7p)(cb0QaH%Q8#77nM+aE6wR$?iA34g&PXTGYpfL=J!BwAbe=TI^kELWZtRc?
zE8*1qc8FLUyHbP209&nL2<&P+b>l?h9i|}k5@$b6!?Xltd|EbRZN}fW&9Ub#I0&1<
z+*~@o5wPkrC`<|Mtbuge*)<kzYF|tC!Wmsa^}^wSUZB@9a*ouR(}GUsE8q1C!b=?@
z7G;g9#gsP}W+=b!tC!{2z;}cN(}Ui~22^G!3=G^u8^EKZ3fh{bNp60!h9gBk?zBOx
z|HBO^V?E`~J=a&Fx&>>=tp3&kht{)~AHkGnkoul-5%n|~3#{6!WbMb-%ADaRCW@zn
z_yJ%+rRmxV8x#y4K7K9cyrc)tMzU|E&K{tF8=cQW4_C?VBh}hs6sw7sUm>}VW{Uks
zF)2Z!%0})eR6L(MnKPu5$ZK_)Dg7Mk<qsC}y8>{J7@H||rv{ecopHe9m&{ReJ$eZR
z{a?e<F^z<hsROK1eL<BC4huB#yp$^BU3DI|*B0RA%~mVcvSnw{(Im4d&a)O$&SR*V
zNGDO0miS?47l~jqGUYb|;MHnl26gx#rLJ_^56gS)bPU7t)A6oO_MW3l%R6?$@i-#x
zQKOI(O?m5ZZV^W7nv*9m2c;smd_BzZf>idMwZdAH!*{{Ex>Y9^3L~72k0!V}AHFL!
zjVAj7w~Xok7;=6ruz!HVtA{SaX_yL+jsUH0<)hjTcv4&Junw@F!_t<}Tv_Afc~1%0
zADu0!@xKH)Y6I2A&S5?1{xC;>w<acT>UAbYN25a#>HFw^h-%`LT`h>(e?clAnNQ{l
zSFGPXT^ii7%2U9!T1VZQPQk4j1N;~PJ}m@*U(Y%%Nsq*&;;e!~d0QLwMPQx_=4C}*
zE>P--1Kn#5X+Zsq&CVX+=TQI50&1U?UShfibD-<=Cj77r&jAK=PQrJO8S@W-EY0Dj
zekaRbkmZ&66mH#a53wr$_ONdp$fD*R1ABM|)eW1(LwN&uTY*E{KfvMo)}B8JrskBj
zPC-zumu_6xP1-_{Tprwm0G<wp`h0M3y20*Dr*<{a^tZypF{%eozlL6g8~3!O6a6zZ
z=Jp&jgB~1yXCgY4@+VOqYx_+jTAs=(%y<t8kSVpcO}BLamzK{YGjF@hohwR9FOUws
z03vhSc3wiyI`XlUA_m8S!$@fKfRa~le=kNp`e>BB?!H0QY1C3*<JUvHr@@;yseE-F
zvK|n!_v4sNoGcJ5%zRaqW(JOQkKREReyboV&qi7!t<R|h*P2jREaDA6oIw+YO^{U{
zCLfBZOKi@_H(NP#0D0}ug8+U#&E2yf(@T$D7~~xKCNwdaI_(1%x;r?YLgYm1Yx5zd
zOWhuT7qIjC4~x#x>6)Si^hILE`)mc$xWXa_scwpYh{Ps5!4{d&b&%9&4>Q`zo}ipG
zni8;}p8;Ox`(OP_EahWmabpm-HR@vNx3i@4cnEkLX30&^;_yFVt7~VEcCA}3RHwY3
zpzn?r;&2v@H$Z4N5q{8adL&LF4O3kV4w*6Pnh_u9I;d7p`PjJL^^YHmyFbrfdT)FH
zJ&tAt&#!)%)6VSF)XDBsH)A_A!61JSy3d^vuQSxSdj1gdWp{4QmWr7SWs(Rw=Aat?
zjD-vC+gdv(y652Lp?_L-S>bnRLN`{rQm^T)Z>YIWy@*Q=QqUq4lw0_WgWQx$=Upfn
z!GWlT&)?6ao=zWt?Yi44GQss)p~3~`?Qb@VSY1sSH|k02a3qidCCUgV-3vUx;j*d+
z_b>>Pg6UtywrFaP!XUfHu=%n_kMM+a8={&^s>ajE1d;q7UENIH{FqUtSvx7BjXRKC
z_ncWAQA+_{2}PfmUd9{Dc{_5!4XJdgn<D-Omn%AzB#?8{P)IPOD^#DNd-crj^Z4z-
zZMa3-y~8>YZd-Zg-9OSR5S@ggTpG#pDXX|Y<5XdRzK>2>{xzN}uI2EeV%R!4M@AY<
z)<B6$11z}0%<k*EFbPR9QGR<cg%Z|I4iwXQ)|WqRoO&xBPd9zs4WBE*oOhq`JLc>1
z%svh*TSxas<NgOXy<)?toN(Nl_YRXyquS#}TmE=RjAoyc5zP9=eIPO4Qjj0WZ0>9=
zoit1E{e$;99{}eL!UfphTE3F?NPZXYfONZnYqld)))k@$u6_SwsyrbsvyM`6OG3Z>
zgowVkMQN)B@`P6>Ucn4!-7u9P;ZUP*x4E-07r#2yafdX8QAnjUh3i5kd{^8igZw*g
zZM6P`w%GiYDY&sM7O<>-zS~;2PZ;&!-%sIWe8wFdju(9{rh#*@Huk45RKV>cL?hZQ
z94KwJ_Q|Eg5h*7BD_mr|v*$!<MWf5n@hT0X6BJM-g({K^Q>)5c?ZC~Pt*pCPBxva5
zGW|E0K^ZZRW{7;!Dhs&t%)G)A3~t<P5j@Cn`&KQEdL`MA{(D+5icG<KfRL-<#{IaT
z(G!McKMjZYjT_8e{=q<MX3oH|Lbzd0@>jO0!qTiF&thfjyp4EXt`)dqpu@`;R{DVA
zb>BY=x1nD+d@%+J6fL=2pK_eAR_3&ZdszT$*UEZCD<%_#Fd2YMfi@TUBuQ<dYR2kK
zEf4Hv9YMzm3t0ObPfqsq^WtPePKCGt*?;*JA+CdVD;q5uT4t!==F>H(*R(_XR)-}P
zi_}zmmL0zMr|zE|{uUC`erI$Tn`+$8<b!TH(vzQ7=22du0UC&j0h3RMh?iHjNV1sJ
z8#;QEDihy&J@*}*^k&B#t7RcpWpu$}tDu&cxeBr>Emqr{M&osP%khIrHjZs9FEb-x
zp5*^r-NUIAS-tc86P&CB!&1@fk>V7CWGi0`Srxb%aKT#s5uo+$!LVUG<rWJAgy?BI
z{akVM$8)}3-zp!EUIA!!+&F--hZM92o3<J;8m-Jb-ga5@6r-L>!D~uFVS((wVF=19
z=Mxgqlrm1wIyaP|25${(_M7+N3DK2|6gMKaeDaE-ulXoQ@HT*Dl~W2Yu?H&T(B%cb
z8!C>so)9_u1)Bd<PfiQW&RUmP0`BxK%by*ZKXnu53dzV^&TX-bnDEco2BU=XWI^0w
z1{{f8QlE{jXX&@0uFhcEdROjv_IcG$H%5O~q#I1jaC5ueXiND)cW~vxQ(R>&!s+0m
z4YE8tmmvKjPq~YnOv~Yns%ElgmjZ3q5Z{EmOu1RId^9MZvB|i*trOj}5`Jp(;;&GR
zesp7N=q`Udh1ynHemHoyI3j#KBa*jtOk%4=vdS>833awEK)QayWW9>Eae#0x^l!1^
zYEk*q!N{<#r?f%e%Qu-rH9KD6v~hraK~Tc*Kb@GSi8``gU+`KZN8+Id{C}2r26w=a
z*~k>kxCbr|p?iVoims8i7z~&pm^NV(Ck+v$tGU8$@wstQKNEjKt=FjbP{u~2sOF==
z0NI06Z_CuRXiRbA`pjpg7Y@g1n{voCJrC8qK}qXYVG;Uno*kJfy=#U2o`y4ySXsz{
z>AAp=t^i`o|A`+F$ytffnIe~K+tgep4aY5EYm$cIPKdt){V<yX;X>_gzxc47ae?!Q
z^fv$Fo|R3}k)w2KrQ?~nDU%<PjC*jYar*xI___0F5W0HU>kdYi5`(R_34^_9XG7XK
zv~%{)op*37t9QP+Y-Kx%njxk%A6MB#-~VN2^*zbm*1lHs8Tm11{KJnW#~3~)9&*Yt
z9AQ}LRd50}fvP;cyrNn&N{IUNtz06rkM9=h-j&8ocW8E3`u5Mhksa*sALA~2QB&}}
z*u;z49^4K10m>8%-8}MsHQ}yCS-z^r-|c@Xv)1eq-(K6gT5w0Sb&{Uam3=@`r1y&y
z{i%{_1}XBWmZBZLyIY~X6g+X2^LCh0>jHT4^7|rgM~l)r7+$@ACSWPFwL{VNu>LtY
zY~!~B(_=rJ-#WR3S)}R7sG7Fis%EF*J;v{IrzMIr_qDNWb8vBs)ElGk$o$Y~dC<O-
z<rwpYCkC;sS#omeY9IVBJ}Z<FiOjNW=LOv`n?7gbhjk<8zc0~ADKTMvkhQV34Jjxn
za9h8AWT{DBUf!};tB`982jr=-(xdOK9e|M@1pRZO*geG@u~S?cQe>Inx<CC2(<Ivn
zD!d=a2z{K1Z>!VP%bmo7hmzHK!t^7j;+!A1U87a$J1=S7!MXTh<Eor7LOH{B&)zR$
zu{+~SPAiT^G6lq(`8;#0;8+9avWc%t>bX>n$+Ma>y#T)xl`|98?AWEYZ~&YEcw89M
zR?EaUIbSDzNQ?nDPhoYWFmNz*^*%BkGn_@c#fulQkHr}-=jv8E4rjXX8WBmPUb>JE
zr$Y6CwS$K6uCQD$7(UDp<@12k_Z5ctY#hvT8=d>JtE~z_YFe@1{!Q8_sS_Reu0Gav
zxiB5V$)WL%z42*UCVP^rD&OA6kJI1(@ZrM=vyrRI)bXqLx&@avZp^IQ@OCjccs5yv
zEqezgZNr{PbHpA;S#f*r`|tums_+T5Gm1Ie@I&%hz&(l@tFPjO;UgGkrqoe0%0~D8
z=g*%{+us;9Yf6Iau*=EiIXY|5=%xY-ZJl$>K)lwvz6!Arm*}??h)dbb1g%(t@U!=!
z{w9<M-52gEH%oh^ySu*!m)mQ(XJKLC^suZxIx60M2IHh_u~|b_;Q;RZ_0(E@R_t)J
z`zS<^eBMjH@>I%xU?^`M0A4SPf835fNi|pJlUjOK^EYlB9TXI_d&7p*&Aoc{3SB!-
zXCo^;c0~%(NOw@_GTatVK~jwl?^D^HP1LMVeY#E>f*jPAqMhk`Ghx1tJEi5bC&*9D
zdT(Q$;Ce%|%<%p3cW5}729+CjZbAUFgTf~t&c13<slJ>@{JO!1<iV`Y!KYSzq^^P&
z6b|Z?7_yqNKYsj}iJuMMIsFTMciYX)ZB|9saqU!BckA=A5)G}$V&!KJEub1nHOZb>
z^R#u+bA#5$HGzW5K*8TPG>$WTm?O%MFnXe&d}A5f;C8~(s2+FMW2Au7f)z^|7Wt&-
z04gCiE{9PAAcb9`9aNF8COy8cQZj=5MSkB3&kf(cxQ{S2d~I;c?y}PA;9a@i$o^4r
zAtj`v_<W`w$azBuMMvaP4|Ri-V~~eXSc5yxn(vzfs#(-NTV&PRh<}L_x?gCO7i2IK
zz5I{;ku@`+R_4Tn0ZefUy*xZCYuG8_qqAmGcR>V&ZK{kvQ%qKZbXMf@?k6?$LX%wt
zQ-wmI<zvg`39j79JYjg42c+T(`}@kbzOUZ?`0N}Ny8CCcf`6G4L8w)|&FJ3*zo}0l
zl=cWJOo?sY8O79s_cpExuG@6WZ0?@2S75Q1psW|%tboevHFB<{dYaU#{Kb(?s$$t^
zKKvcmy~)wOB<BNw;v&)A2ki55FIIR)3JL_xl~$*Hd^CK!3~v)y+ALs+9m9JV&DdB-
zM7K+0+euzoyjM1)dv0JxGw}0(5sZL}M}JKdUrywHVJ?`o3O}=wWbn6wx*p~X$}3(L
zijT&6z8oDQnDEi%CX4`)d7zw^*gW{nt~C!4?l0fHTQy?%aEsDine8zAq~d{FIEKKH
zJiOF3nhCyTUnrlh_Lpf9Bczve)iY2`5nT44M`tiC$oZiQ3shKx<ZR~7J-Mx2cb^Sa
zX=&5rFJ7GFeQjYyCo@CsGDENPTu5RVBrzgzz|UJ3<ah#-#DJVLgTy=(lkT^NZ$Pt3
z|1<R*Il%WuI`(|2q4#(42%U7r+#myUp77u(U6@$DJV5G`J)fB`k>TxNzjX4<78%|<
zmbmq&0!dQmG8E_Rk5EmO;?x`YfuVACsu$zqjnjiiPB}7qV>dV0DG#d%^Q<@nZUFAR
zjIkPxWTsf*8Mx!*?_|N|e$LMN#LQ4N5>W9^gE%wiZi55lqmJP^!0HdB?Y&m6T<JG`
z#*9DOY;bYe8}Csp3lpz+!47k@#;{ZYd)=ANxJ{f0mMh6+<Ogp(F@Yr_8`T?P9e93O
zC6WU}K(6eeC8OBc3ZE)ondKG~yeW9E7Fr-!q_}$S^XFw~#eWI~`R8m#3kEd|N=z*P
zZk|{78j6E*$&p7Y+OiLKda#O=0R3K`J(ix;P%wH@T&&%-d-uxD!%Sg9G?3+4{-Gzr
zDg>@hZs#IV2&~~ynHevXq@+r{2{KX+$xgrAoKjiycR{B*4oO#7YVia=N%ZR7^BZv&
z4Bx(feZt$@JI114%2M2uW$BMLJ#9)?$z<&j8(j_dNe2jw+A~)@w5^L@H~bLQjI4f?
z-5k)8;vTJsC8q3brQwE#Q*GClr+)|k?z;$V-Wn;5)+sX_aPc8-i1P=BXP4?#58#s(
zN#4RTBkl#S44o@RTrxpFA=r*iC-W#aT|4FTv)|{(8k{u`i;0b$?dFret(Bor#l2fI
z)*k3hPsiKgV(<km;P{qv(-H>=#TG!aExQhq`o@(LAzmW53O0M3AUCTYNoKr?@a;FF
z6<m=>qs->rANGUcK=IG2y*{!SqVaat>=n8C^8|4Zk3`Ql6h&mu=jNgmw(MXuQRb)i
zu-(jY;zbsH#;?Y|x*gE}vi=AM2#Xv_5c77F!{8w(P+Mlb<Y!=|VRD87s~wK|)V(N1
zx#0)Bx!CVgP%j9U9&QFNKm4U|8sCNY?b|mu508(5y^mjjhQrc)-8=|%G=dg*zaJsK
zQbv6foGpUmX8WC~gI+QP>7)5HLom#tKzB-=Ap3>N#Y>l_&ImUA99yLCg8C9Nc2%ub
zHN>+Jwo7%WAKP2l1kctG&)5o|E))rtuPw~WYo+(0a-#C;wh1FPE2;ZY2UPQ8J;ey5
zw0c42QtkWSE2UD9!yf~YZy8Wx%$8lT1tgph{zy}9m+Gh7_v_vkCdODrWM*e)fBI+1
z2Gb{=FI%ieN+udsHT~AxA~D4}$h{zKv$<!wy9vhR6Na1~385Q2>c$*Uxx|6>w|u{V
z+cDA6O7Yd6JGdhH9mc);kdIiQab28;IWSFq&h2Oq;Un|nYvRo%KJbnJ&ML~oC{Snq
z%fhMC=9}``bsu%&+QL2f0c(R-Yh5pgSG6mh7C)Jr8B{pY_uKl%!lduIC2!Y``+`=0
z*|J|hcmEN-u<g&@GKB-y^9x91T%9v$&`+|4*L*ZurTBf%^*=W36BY-STU3sW8#&r7
zDeq<4`a;L@D$KjcwpKi?Q|Tgn`SRtG)D~YG?zbMt&UPg4H$Km~30v0CgJTG;t~MBi
zvvqv}@m8`HkvZOG9m{L-ddYN~!4<o14PiY0baD>SWL(I%VCgDbOR}qAu20;chHgQS
z?X4F>CLnc4mwLBTHbU^$h|{`Fefc}OcynF0bu|GeJQ`G=7!CeEDUv#s_Adn;Y~Vo2
zvEU_9X!lQK&XXpuH$MuxHGg1(A*Y@>-<jn4TuQKX6U%7%JD8*AZ|+=e0D`ft;7Y;^
z^efi1nZY<=X_Crr3B%t`tGq!5mdaKy=-Jgay$QCf214FyX!kphB@HDWp0?QVfB^jJ
zVB2K)WE<mfc85<hiGuEnM>ZYfXNbUV<Q?46fD#PWD61loo)(^^$q>8&LjU?vV>+U~
zE%|Nrna#u<jfYZ~s^P+Rx)jdYUaNs@VaBfHhXml=@CQb}LeU@tYc^~Yz<`<s1>=L5
zd{{%e{wup0qo8I+tl^Ij^L{v$OxJDtin0}`3yYkK0vaNbvO=*{@Q;ds%N>S4`Wfpr
z^q;>N3Jx2#OV)Z|&n6STVa$0G5I=OW4PV}bn5$8qY)81NwqVMVCe#vsL8LxwGw8^1
z+}mBJsRGq7Ud>pF6kx8nxl1Z3>qAh_s{9uE*ZaIts?-5a{VyQ4Mdu)OyZ&!r;v9z!
z+O3-%+tiJJ74M(l8jnPUPn71MCR;xVa4)54rhC`oF%A5`!ApiG3&d*9kw1ZKR8}N#
zVmD%>Veu9Zo6wfi=<*5_gQo1iBhG(F*iRbPe9ELxP%X8R_K%V{#X#8u>CK5bdpk8?
z+?juA66FR{D{=8c+a_dH8b$d(Bh)B)py>Y@At^@TKO-c?DEw!Hq?oq<jL?5ZNQ#I^
zi@5$XLTvZIi9=*$WPVH@KW0|rF~t94gH$$p$W^PJn^G=#gA_88ji!%}96SH7{{iqG
Bv>*Tg

literal 0
HcmV?d00001

diff --git a/3rdparty/composable_kernel/doc/markdown/dockerhub.md b/3rdparty/composable_kernel/doc/markdown/dockerhub.md
new file mode 100644
index 0000000..91b6cb2
--- /dev/null
+++ b/3rdparty/composable_kernel/doc/markdown/dockerhub.md
@@ -0,0 +1,93 @@
+## CK docker hub
+
+[Docker hub](https://hub.docker.com/r/rocm/composable_kernel)
+
+## Why do I need this?
+
+To make our lives easier and bring Composable Kernel dependencies together, we recommend using docker images.
+
+## So what is Composable Kernel?
+
+Composable Kernel (CK) library aims to provide a programming model for writing performance critical kernels for machine learning workloads across multiple architectures including GPUs, CPUs, etc, through general purpose kernel languages, like HIP C++.
+
+To get the CK library
+
+```
+git clone https://github.com/ROCmSoftwarePlatform/composable_kernel.git
+```
+
+run a docker container 
+
+```
+docker run                                                            \
+-it                                                                   \
+--privileged                                                          \
+--group-add sudo                                                      \
+-w /root/workspace                                                    \
+-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace                         \
+rocm/composable_kernel:ck_ub20.04_rocm5.3_release                     \
+/bin/bash
+```
+
+and build the CK
+
+```
+mkdir build && cd build
+
+# Need to specify target ID, example below is for gfx908 and gfx90a
+cmake                                                                                             \
+-D CMAKE_PREFIX_PATH=/opt/rocm                                                                    \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
+-D CMAKE_CXX_FLAGS="-O3"                                                                          \
+-D CMAKE_BUILD_TYPE=Release                                                                       \
+-D GPU_TARGETS="gfx908;gfx90a"                                                                    \
+..
+```
+
+and 
+
+```
+make -j examples tests
+```
+
+To run all the test cases including tests and examples run
+
+```
+make test
+```
+
+We can also run specific examples or tests like
+
+```
+./bin/example_gemm_xdl_fp16
+./bin/test_gemm_fp16
+```
+
+For more details visit [CK github repo](https://github.com/ROCmSoftwarePlatform/composable_kernel), [CK examples](https://github.com/ROCmSoftwarePlatform/composable_kernel/tree/develop/example), [even more CK examples](https://github.com/ROCmSoftwarePlatform/composable_kernel/tree/develop/client_example).
+
+## And what is inside?
+
+The docker images have everything you need for running CK including:
+
+* [ROCm](https://www.amd.com/en/graphics/servers-solutions-rocm)
+* [CMake](https://cmake.org/)
+* [Compiler](https://github.com/RadeonOpenCompute/llvm-project)
+
+## Which image is right for me?
+
+Let's take a look at the image naming, for example "ck_ub20.04_rocm5.4_release". The image specs are:
+
+* "ck" - made for running Composable Kernel
+* "ub20.04" - based on Ubuntu 20.04
+* "rocm5.4" - ROCm platform version 5.4
+* "release" - compiler version is release
+
+So just pick the right image for your project dependencies and you're all set.
+
+## DIY starts here
+
+If you need to customize a docker image or just can't stop tinkering, feel free to adjust the [Dockerfile](https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/develop/Dockerfile) for your needs.
+
+## License
+
+CK is released under the MIT [license](https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/develop/LICENSE).
diff --git a/3rdparty/composable_kernel/example/01_gemm/CMakeLists.txt b/3rdparty/composable_kernel/example/01_gemm/CMakeLists.txt
new file mode 100644
index 0000000..c403e51
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/CMakeLists.txt
@@ -0,0 +1,37 @@
+add_custom_target(example_gemm_dl)
+
+add_example_executable(example_gemm_dl_fp32 gemm_dl_fp32.cpp)
+add_example_executable(example_gemm_dl_fp16 gemm_dl_fp16.cpp)
+add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp)
+
+add_dependencies(example_gemm_dl example_gemm_dl_fp32)
+add_dependencies(example_gemm_dl example_gemm_dl_fp16)
+add_dependencies(example_gemm_dl example_gemm_dl_int8)
+
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_gemm_dl_int4 gemm_dl_int4.cpp)
+  add_dependencies(example_gemm_dl example_gemm_dl_int4)
+endif(USE_BITINT_EXTENSION_INT4)
+
+
+add_custom_target(example_gemm_xdl)
+
+add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
+add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
+add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp)
+
+add_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
+add_dependencies(example_gemm_xdl example_gemm_xdl_bf16)
+add_dependencies(example_gemm_xdl example_gemm_xdl_int8)
+
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_gemm_xdl_int4 gemm_xdl_int4.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_int4)
+endif(USE_BITINT_EXTENSION_INT4)
+
+add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp)
+# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
+add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
+
+add_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
+add_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
diff --git a/3rdparty/composable_kernel/example/01_gemm/README.md b/3rdparty/composable_kernel/example/01_gemm/README.md
new file mode 100644
index 0000000..226783b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/README.md
@@ -0,0 +1,23 @@
+# Instructions for ```example_gemm_xdl```
+
+## Run ```example_gemm_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+./bin/example_gemm_xdl 0 1 5
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
+arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
+arg.c_grid_desc_m_n_{ 3840, 4096}
+launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+Warm up
+Start running 5 times...
+Perf: 1.19685 ms, 107.657 TFlops, 78.8501 GB/s
+```
diff --git a/3rdparty/composable_kernel/example/01_gemm/common.hpp b/3rdparty/composable_kernel/example/01_gemm/common.hpp
new file mode 100644
index 0000000..495a815
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/common.hpp
@@ -0,0 +1,89 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <numeric>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+struct ProblemSize final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+inline bool
+parse_cmd_args(int argc, char* argv[], ProblemSize& problem_size, ExecutionConfig& config)
+{
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        problem_size.M = std::stoi(argv[4]);
+        problem_size.N = std::stoi(argv[5]);
+        problem_size.K = std::stoi(argv[6]);
+
+        problem_size.StrideA = std::stoi(argv[7]);
+        problem_size.StrideB = std::stoi(argv[8]);
+        problem_size.StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
+                  << std::endl
+                  << "arg3: time kernel (0=no, 1=yes)" << std::endl
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl;
+        return false;
+    }
+
+    return true;
+}
diff --git a/3rdparty/composable_kernel/example/01_gemm/gemm_dl_fp16.cpp b/3rdparty/composable_kernel/example/01_gemm/gemm_dl_fp16.cpp
new file mode 100644
index 0000000..cf585a8
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/gemm_dl_fp16.cpp
@@ -0,0 +1,37 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+
+using ADataType   = ck::half_t;
+using BDataType   = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+
+using ALayout = Col;
+using BLayout = Row;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDl
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,               5,                  4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/01_gemm/gemm_dl_fp32.cpp b/3rdparty/composable_kernel/example/01_gemm/gemm_dl_fp32.cpp
new file mode 100644
index 0000000..93f085c
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/gemm_dl_fp32.cpp
@@ -0,0 +1,37 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+
+using ADataType   = float;
+using BDataType   = float;
+using CDataType   = float;
+using AccDataType = float;
+
+using ALayout = Col;
+using BLayout = Row;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDl
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 1>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<2, 1, 4, 1>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 1>, S<0, 1, 2, 3, 4, 5>,               5,                  4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/01_gemm/gemm_dl_int4.cpp b/3rdparty/composable_kernel/example/01_gemm/gemm_dl_int4.cpp
new file mode 100644
index 0000000..e392c49
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/gemm_dl_int4.cpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+
+using ADataType       = ck::int4_t;
+using BDataType       = ck::int4_t;
+using CDataType       = ck::int4_t;
+using KernelADataType = int8_t;
+using KernelBDataType = int8_t;
+using KernelCDataType = int8_t;
+using AccDataType     = int32_t;
+
+using ALayout = Col;
+using BLayout = Row;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDl
+// ######|           AData|           BData|           CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|            Type|            Type|            Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|                |                |                |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|                |                |                |            |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < KernelADataType, KernelBDataType, KernelCDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#define BUILD_INT4_EXAMPLE
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/01_gemm/gemm_dl_int8.cpp b/3rdparty/composable_kernel/example/01_gemm/gemm_dl_int8.cpp
new file mode 100644
index 0000000..be9e387
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/gemm_dl_int8.cpp
@@ -0,0 +1,37 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+
+using ADataType   = int8_t;
+using BDataType   = int8_t;
+using CDataType   = int8_t;
+using AccDataType = int32_t;
+
+using ALayout = Col;
+using BLayout = Row;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDl
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_bf16.cpp b/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_bf16.cpp
new file mode 100644
index 0000000..9aaae6a
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_bf16.cpp
@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = ck::bhalf_t;
+using CDataType        = ck::bhalf_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::bhalf_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_fp16.cpp b/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_fp16.cpp
new file mode 100644
index 0000000..488babb
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_fp16.cpp
@@ -0,0 +1,50 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+using ADataType        = ck::half_t;
+using BDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using CDataType        = ck::half_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance0 = ck::tensor_operation::device::DeviceGemmXdl
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+// ######|          |          |          |            |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>;
+// clang-format on
+
+// clang-format off
+using DeviceGemmInstance1 = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+using DeviceGemmInstance = DeviceGemmInstance0;
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_fp64.cpp b/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_fp64.cpp
new file mode 100644
index 0000000..99253b7
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_fp64.cpp
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+
+using ADataType   = double;
+using BDataType   = double;
+using CDataType   = double;
+using AccDataType = double;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+// ######|          |          |          |            |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+#if 0
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,   GemmDefault,    64,    32,    32,     4,  1,   16,   16,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,               7,               1>;
+#else
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,   GemmDefault,   256,   128,   128,     4,  2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>;
+#endif
+    // clang-format on
+
+    using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                            BDataType,
+                                                                            CDataType,
+                                                                            AccDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_int4.cpp b/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_int4.cpp
new file mode 100644
index 0000000..7f1283a
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_int4.cpp
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+using ADataType        = ck::int4_t;
+using BDataType        = ck::int4_t;
+using CDataType        = ck::int4_t;
+using KernelADataType  = int8_t;
+using KernelBDataType  = int8_t;
+using KernelCDataType  = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int8_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
+// ######| ALayout| BLayout| CLayout|           AData|           BData|           CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |            Type|            Type|            Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|        |        |        |                |                |                |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|        |        |        |                |                |                |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, KernelADataType, KernelBDataType, KernelCDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,          1,          1,           1,               S<1, 64, 1, 4>,              16>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#define BUILD_INT4_EXAMPLE
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_int8.cpp b/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_int8.cpp
new file mode 100644
index 0000000..e67594c
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_int8.cpp
@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+using ADataType        = int8_t;
+using BDataType        = int8_t;
+using CDataType        = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int8_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 64, 1, 4>,              16>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp b/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
new file mode 100644
index 0000000..12a6992
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
@@ -0,0 +1,249 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp"
+
+#include "ck/library/utility/literals.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+#define USING_SKIP_LDS 1
+
+// clang-format off
+#if USING_SKIP_LDS
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSkipBLds
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BThreadTransfer|  BBlock| CThreadTransfer| CThreadTransfer|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|       SrcScalar|  buffer| SrcDstVectorDim|       DstScalar|
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |       PerVector|  size  |                |       PerVector|
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |        |                |                |
+#if 0       
+                    <   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    16,   64,     4,  8,   16,   16,    1,    1,     S<16, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,            8,       8,        7,               1>;
+using ADataType   = ck::half_t;
+using BDataType   = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+#else  
+                    <   F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    16,   64,     4,  4,   16,   16,    1,    1,     S<16, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,             4,      4,        7,               1>;
+using ADataType   = float;
+using BDataType   = float;
+using CDataType   = float;
+using AccDataType = float;
+#endif
+
+#else
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+                   <   F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    16,   64,     4,  4,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,              1,      2>;
+using ADataType   = float;
+using BDataType   = float;
+using CDataType   = float;
+using AccDataType = float;
+
+#endif
+    // clang-format on
+
+    using ReferenceGemmInstance = ck::tensor_operation::host::
+        ReferenceGemm<ADataType, BDataType, CDataType, float, AElementOp, BElementOp, CElementOp>;
+
+template <typename DataType>
+std::ostream& show_2d_matrix(std::ostream& os, Tensor<DataType>& matrix)
+{
+    os << "[" << std::endl;
+    for(size_t x = 0; x < matrix.mDesc.GetLengths()[0]; x++)
+    {
+        os << "[";
+        for(size_t y = 0; y < matrix.mDesc.GetLengths()[1]; y++)
+        {
+            os << std::setw(5) << static_cast<float>(matrix(x, y));
+        }
+        os << "]" << std::endl;
+    }
+    os << "]";
+    return os;
+}
+int main(int argc, char* argv[])
+{
+    bool do_verification = 0;
+    int init_method      = 0;
+    bool time_kernel     = false;
+
+    // GEMM shape
+#if 1
+    ck::index_t M = 16;
+    ck::index_t N = 64 * 120;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideC = N;
+#else
+    ck::index_t M = 16;
+    ck::index_t N = 16;
+    ck::index_t K = 32;
+
+    ck::index_t StrideA = 8;
+    ck::index_t StrideB = 8;
+    ck::index_t StrideC = 16;
+#endif
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        // a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+    }
+
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                                      static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+                                      static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      StrideC,
+                                      a_element_op,
+                                      b_element_op,
+                                      c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+#if 0
+        {
+            show_2d_matrix(std::cout << "a : ", a_m_k) << std::endl;
+            show_2d_matrix(std::cout << "b: ", b_k_n) << std::endl;
+            show_2d_matrix(std::cout << "c_device: ", c_m_n_device_result) << std::endl;
+            show_2d_matrix(std::cout << "c_host  :", c_m_n_host_result) << std::endl;
+        }
+#endif
+        ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/01_gemm/run_gemm_example.inc b/3rdparty/composable_kernel/example/01_gemm/run_gemm_example.inc
new file mode 100644
index 0000000..4e2cedb
--- /dev/null
+++ b/3rdparty/composable_kernel/example/01_gemm/run_gemm_example.inc
@@ -0,0 +1,151 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+
+    using namespace ck::literals;
+
+    auto& [M, N, K, StrideA, StrideB, StrideC] = problem_size;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
+        break;
+    default:
+        ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
+    }
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+#ifdef BUILD_INT4_EXAMPLE
+    DeviceMem a_m_k_device_buf(sizeof(KernelADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(KernelBDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(KernelCDataType) *
+                               c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    const Tensor<KernelADataType> a_m_k_converted(a_m_k);
+    const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
+
+    a_m_k_device_buf.ToDevice(a_m_k_converted.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n_converted.mData.data());
+#else
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+#endif
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(
+#ifdef BUILD_INT4_EXAMPLE
+        static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+        static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+#else
+        static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+#endif
+        M,
+        N,
+        K,
+        StrideA,
+        StrideB,
+        StrideC,
+        a_element_op,
+        b_element_op,
+        c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return true;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop = 2_uz * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+#ifdef BUILD_INT4_EXAMPLE
+        Tensor<CDataType> c_m_n_device_result_converted(c_m_n_host_result.mDesc);
+
+        c_m_n_device_buf.FromDevice(c_m_n_device_result_converted.mData.data());
+
+        c_m_n_device_result = c_m_n_device_result_converted.CopyAsType<CDataType>();
+
+        return ck::utils::check_err(c_m_n_device_result_converted, c_m_n_host_result);
+#else
+        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+#endif
+    }
+
+    return true;
+}
+
+bool run_gemm_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
+}
diff --git a/3rdparty/composable_kernel/example/02_gemm_bilinear/CMakeLists.txt b/3rdparty/composable_kernel/example/02_gemm_bilinear/CMakeLists.txt
new file mode 100644
index 0000000..10ec0f1
--- /dev/null
+++ b/3rdparty/composable_kernel/example/02_gemm_bilinear/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
diff --git a/3rdparty/composable_kernel/example/02_gemm_bilinear/README.md b/3rdparty/composable_kernel/example/02_gemm_bilinear/README.md
new file mode 100644
index 0000000..9eb87e1
--- /dev/null
+++ b/3rdparty/composable_kernel/example/02_gemm_bilinear/README.md
@@ -0,0 +1,28 @@
+# Instructions for ```example_gemm_bilinear_xdl_fp16```
+
+## Run ```example_gemm_bilinear_xdl_fp16```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE
+#arg11 to 12: alpha, beta
+./bin/example_gemm_bilinear_xdl_fp16 1 1 1 3840 4096 4096 4096 4096 4096 4096 0.5 0.5
+```
+Result (MI100 @ 1502Mhz, 184.6TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
+arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
+arg.c0_grid_desc_m_n_{ 3840, 4096}
+arg.c_grid_desc_m_n_{ 3840, 4096}
+launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+Warm up
+Start running 1 times...
+Perf: 0.936965 ms, 137.517 TFlops, 102.959 GB/s
+error: 0
+max_diff: 0, 558.5, 558.5
+```
diff --git a/3rdparty/composable_kernel/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp b/3rdparty/composable_kernel/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
new file mode 100644
index 0000000..917b6b1
--- /dev/null
+++ b/3rdparty/composable_kernel/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
@@ -0,0 +1,306 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+struct AlphaBetaAdd
+{
+    AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
+
+    template <typename E, typename C, typename D>
+    __host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
+        ck::half_t& e, const float& c, const ck::half_t& d) const
+    {
+        e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
+    };
+
+    float alpha_;
+    float beta_;
+};
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F16;
+using EDataType        = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using DLayout = Row;
+using ELayout = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AlphaBetaAdd;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance =
+    ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<ALayout,
+                                                                   BLayout,
+                                                                   ck::Tuple<DLayout>,
+                                                                   ELayout,
+                                                                   ADataType,
+                                                                   BDataType,
+                                                                   AccDataType,
+                                                                   CShuffleDataType,
+                                                                   ck::Tuple<DDataType>,
+                                                                   EDataType,
+                                                                   AElementOp,
+                                                                   BElementOp,
+                                                                   CDEElementOp,
+                                                                   GemmSpec,
+                                                                   1,
+                                                                   256,
+                                                                   256,
+                                                                   128,
+                                                                   32,
+                                                                   8,
+                                                                   8,
+                                                                   32,
+                                                                   32,
+                                                                   4,
+                                                                   2,
+                                                                   S<4, 64, 1>,
+                                                                   S<1, 0, 2>,
+                                                                   S<1, 0, 2>,
+                                                                   2,
+                                                                   8,
+                                                                   8,
+                                                                   1,
+                                                                   S<4, 64, 1>,
+                                                                   S<1, 0, 2>,
+                                                                   S<1, 0, 2>,
+                                                                   2,
+                                                                   8,
+                                                                   8,
+                                                                   1,
+                                                                   1,
+                                                                   1,
+                                                                   S<1, 32, 1, 8>,
+                                                                   8>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideD = 4096;
+    ck::index_t StrideE = 4096;
+
+    float alpha = 1.0f;
+    float beta  = 1.0f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        alpha = std::stof(argv[4]);
+        beta  = std::stof(argv[5]);
+    }
+    else if(argc == 13)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+
+        alpha = std::stof(argv[11]);
+        beta  = std::stof(argv[12]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, alpha, "
+               "beta\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{alpha, beta};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, 1>{StrideD},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/03_gemm_bias_relu/CMakeLists.txt b/3rdparty/composable_kernel/example/03_gemm_bias_relu/CMakeLists.txt
new file mode 100644
index 0000000..35c54ab
--- /dev/null
+++ b/3rdparty/composable_kernel/example/03_gemm_bias_relu/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_gemm_bias_relu_xdl_fp16 gemm_bias_relu_xdl_fp16.cpp)
diff --git a/3rdparty/composable_kernel/example/03_gemm_bias_relu/README.md b/3rdparty/composable_kernel/example/03_gemm_bias_relu/README.md
new file mode 100644
index 0000000..f28a9a0
--- /dev/null
+++ b/3rdparty/composable_kernel/example/03_gemm_bias_relu/README.md
@@ -0,0 +1,10 @@
+# Instructions for ```example_gemm_bias_relu_xdl_fp16```
+
+## Run ```example_gemm_bias_relu_xdl_fp16```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE
+./bin/example_gemm_bias_relu_xdl_fp16 1 1 1 3840 4096 4096 4096 4096 4096
+```
diff --git a/3rdparty/composable_kernel/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp b/3rdparty/composable_kernel/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
new file mode 100644
index 0000000..aee51d0
--- /dev/null
+++ b/3rdparty/composable_kernel/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
@@ -0,0 +1,283 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// C = A * B
+// E = Relu(C + D);
+struct AddRelu
+{
+    __host__ __device__ void
+    operator()(ck::half_t& e, const ck::half_t& c, const ck::half_t& d) const
+    {
+        const ck::half_t x = c + d;
+
+        e = x > 0 ? x : 0;
+    }
+};
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using EDataType        = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using DLayout = Row;
+using ELayout = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddRelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance =
+    ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<ALayout,
+                                                                   BLayout,
+                                                                   ck::Tuple<DLayout>,
+                                                                   ELayout,
+                                                                   ADataType,
+                                                                   BDataType,
+                                                                   AccDataType,
+                                                                   CShuffleDataType,
+                                                                   ck::Tuple<DDataType>,
+                                                                   EDataType,
+                                                                   AElementOp,
+                                                                   BElementOp,
+                                                                   CDEElementOp,
+                                                                   GemmSpec,
+                                                                   1,
+                                                                   256,
+                                                                   256,
+                                                                   128,
+                                                                   32,
+                                                                   8,
+                                                                   8,
+                                                                   32,
+                                                                   32,
+                                                                   4,
+                                                                   2,
+                                                                   S<4, 64, 1>,
+                                                                   S<1, 0, 2>,
+                                                                   S<1, 0, 2>,
+                                                                   2,
+                                                                   8,
+                                                                   8,
+                                                                   1,
+                                                                   S<4, 64, 1>,
+                                                                   S<1, 0, 2>,
+                                                                   S<1, 0, 2>,
+                                                                   2,
+                                                                   8,
+                                                                   8,
+                                                                   1,
+                                                                   1,
+                                                                   1,
+                                                                   S<1, 32, 1, 8>,
+                                                                   8>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideE = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, 0, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+
+    auto invoker = device_op.MakeInvoker();
+
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, 1>{0},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(EDataType) * M * N + sizeof(EDataType) * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    if(do_verification)
+    {
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        Tensor<AccDataType> c_m_n(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
+            }
+        }
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/CMakeLists.txt b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/CMakeLists.txt
new file mode 100644
index 0000000..c75c5ba
--- /dev/null
+++ b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/CMakeLists.txt
@@ -0,0 +1,17 @@
+add_custom_target(example_gemm_add_add_fastgelu_xdl)
+
+add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp)
+add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)
+add_example_executable(example_gemm_add_add_fastgelu_xdl_fp32 gemm_add_add_fastgelu_xdl_fp32.cpp)
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_gemm_add_add_fastgelu_xdl_int4 gemm_add_add_fastgelu_xdl_int4.cpp)
+endif(USE_BITINT_EXTENSION_INT4)
+add_example_executable(example_gemm_add_add_fastgelu_xdl_int8 gemm_add_add_fastgelu_xdl_int8.cpp)
+
+add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_bf16)
+add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp16)
+add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp32)
+if(USE_BITINT_EXTENSION_INT4)
+  add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4)
+endif(USE_BITINT_EXTENSION_INT4)
+add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8)
diff --git a/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/README.md b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/README.md
new file mode 100644
index 0000000..08a55fb
--- /dev/null
+++ b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/README.md
@@ -0,0 +1,23 @@
+# Instructions for ```example_gemm_add_add_fastgelu_xdl_fp16```
+
+## Run ```example_gemm_add_add_fastgelu_xdl_fp16```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
+./bin/example_gemm_add_add_fastgelu_xdl_fp16 1 1 1
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
+d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 1.26914 ms, 101.525 TFlops, 100.804 GB/s, DeviceGemmMultipleD_Xdl_CShuffle<256, 256, 128, 32, 8, 8>
+```
diff --git a/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/common.hpp b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/common.hpp
new file mode 100644
index 0000000..3f9375e
--- /dev/null
+++ b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/common.hpp
@@ -0,0 +1,106 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cstddef>
+#include <iostream>
+#include <stdexcept>
+#include <string>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using I4 = ck::int4_t;
+#endif
+using I8  = int8_t;
+using I32 = int32_t;
+
+struct ProblemSize final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+inline bool
+parse_cmd_args(int argc, char* argv[], ProblemSize& problem_size, ExecutionConfig config)
+{
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 12)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        problem_size.M = std::stoi(argv[4]);
+        problem_size.N = std::stoi(argv[5]);
+        problem_size.K = std::stoi(argv[6]);
+
+        problem_size.StrideA  = std::stoi(argv[7]);
+        problem_size.StrideB  = std::stoi(argv[8]);
+        problem_size.StrideD0 = std::stoi(argv[9]);
+        problem_size.StrideD1 = std::stoi(argv[10]);
+        problem_size.StrideE  = std::stoi(argv[11]);
+    }
+    else
+    {
+        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
+                  << std::endl
+                  << "arg3: time kernel (0=no, 1=yes)" << std::endl
+                  << "arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, "
+                     "StrideE"
+                  << std::endl;
+        return false;
+    }
+
+    return true;
+}
diff --git a/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp
new file mode 100644
index 0000000..5e50c14
--- /dev/null
+++ b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType        = BF16;
+using BDataType        = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = BF16;
+using D1DataType       = BF16;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+using EDataType        = BF16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_add_fastgelu_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
new file mode 100644
index 0000000..6c7ca41
--- /dev/null
+++ b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F16;
+using D1DataType       = F16;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+using EDataType        = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_add_fastgelu_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp
new file mode 100644
index 0000000..1ef266f
--- /dev/null
+++ b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F32;
+using D1DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+using EDataType        = F32;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 32, 1, 8>,               4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_add_fastgelu_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp
new file mode 100644
index 0000000..8b5bc98
--- /dev/null
+++ b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#include "common.hpp"
+
+using ADataType        = I4;
+using BDataType        = I4;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using D0DataType       = I4;
+using D1DataType       = I4;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+using EDataType        = I4;
+
+using KernelADataType  = I8;
+using KernelBDataType  = I8;
+using KernelD0DataType = I8;
+using KernelD1DataType = I8;
+using KernelDsDataType = ck::Tuple<KernelD0DataType, KernelD1DataType>;
+using KernelEDataType  = I8;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
+//######| ALayout| BLayout| DsLayout| ELayout|           AData|           BData|     AccData|         CShuffle|           DsData|           EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |            Type|            Type|        Type|         DataType|             Type|            Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |                |                |            |                 |                 |                |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |                |                |            |                 |                 |                |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, KernelADataType, KernelBDataType, AccDataType, CShuffleDataType, KernelDsDataType, KernelEDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 64, 1, 4>,              16>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#define BUILD_INT4_EXAMPLE
+#include "run_gemm_add_add_fastgelu_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp
new file mode 100644
index 0000000..b236f5e
--- /dev/null
+++ b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType        = I8;
+using BDataType        = I8;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using D0DataType       = I8;
+using D1DataType       = I8;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+using EDataType        = I8;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 64, 1, 4>,              16>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_add_fastgelu_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
new file mode 100644
index 0000000..f3def33
--- /dev/null
+++ b/3rdparty/composable_kernel/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
@@ -0,0 +1,166 @@
+#pragma once
+
+bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+    using namespace ck::literals;
+
+    auto& [M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE] = problem_size;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<
+#ifdef BUILD_INT4_EXAMPLE
+        KernelEDataType
+#else
+        EDataType
+#endif
+        >
+        e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<KernelADataType> a_m_k_converted(a_m_k);
+    const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
+    const Tensor<KernelD0DataType> d0_m_n_converted(d0_m_n);
+    const Tensor<KernelD1DataType> d1_m_n_converted(d1_m_n);
+
+    a_device_buf.ToDevice(a_m_k_converted.mData.data());
+    b_device_buf.ToDevice(b_k_n_converted.mData.data());
+    d0_device_buf.ToDevice(d0_m_n_converted.mData.data());
+    d1_device_buf.ToDevice(d1_m_n_converted.mData.data());
+#else
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+#endif
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               {StrideD0, StrideD1},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = 2_uz * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(D0DataType) * N + sizeof(D1DataType) * M * N +
+                            sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << device_op.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<EDataType> e_m_n_device_result_converted(e_m_n_device_result);
+
+        return ck::utils::check_err(e_m_n_device_result_converted, e_m_n_host_result);
+#else
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+#endif
+    }
+
+    return true;
+}
+
+bool run_gemm_add_add_fastgelu_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) ||
+           run_gemm_add_add_fastgelu(problem_size, config);
+}
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/CMakeLists.txt b/3rdparty/composable_kernel/example/09_convnd_fwd/CMakeLists.txt
new file mode 100644
index 0000000..e0a5300
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/CMakeLists.txt
@@ -0,0 +1,11 @@
+add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
+add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
+add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
+add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
+# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
+add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
+
+add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp)
+add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp)
+add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp)
+
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/README.md b/3rdparty/composable_kernel/example/09_convnd_fwd/README.md
new file mode 100644
index 0000000..9ab5fee
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/README.md
@@ -0,0 +1,32 @@
+# Instructions for ```example_convnd_fwd_xdl```
+
+## Run ```example_convnd_fwd_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+#arg4: N spatial dimensions (default 2)
+#Following arguments (depending on number of spatial dims):
+# N, K, C, 
+# <filter spatial dimensions>, (ie Y, X for 2D)
+# <input image spatial dimensions>, (ie Hi, Wi for 2D)
+# <strides>, (ie Sy, Sx for 2D)
+# <dilations>, (ie Dy, Dx for 2D)
+# <left padding>, (ie LeftPy, LeftPx for 2D)
+# <right padding>, (ie RightPy, RightPx for 2D)
+./bin/example_convnd_fwd_xdl 0 1 100
+```
+
+Result (MI100 @ 1087Mhz, 33.4TFlops peak FP32)
+```
+input: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
+weights: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
+output: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
+arg.a_grid_desc_k0_m_k1_{432, 165888, 4}
+arg.b_grid_desc_k0_n_k1_{432, 256, 4}
+arg.c_grid_desc_m_n_{ 165888, 256}
+launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
+Warm up
+Start running 100 times...
+Perf: 4.43736 ms, 33.0753 TFlops, 150.357 GB/s
+```
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/compile.sh b/3rdparty/composable_kernel/example/09_convnd_fwd/compile.sh
new file mode 100644
index 0000000..3406886
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/compile.sh
@@ -0,0 +1,10 @@
+mkdir build -p
+cd build
+
+cmake                                                                                             \
+-D CMAKE_CXX_COMPILER=${ROCM_PATH}/bin/hipcc                                                    \
+-D CMAKE_CXX_FLAGS="-O3"                                                                          \
+-D CMAKE_BUILD_TYPE=Release                                                                       \
+..
+
+cd -
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_common.hpp b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_common.hpp
new file mode 100644
index 0000000..4c594cc
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_common.hpp
@@ -0,0 +1,172 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+void print_helper_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          int init_method,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      std::array<const void*, 0>{},
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  out_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(
+            out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return true;
+}
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_common.hpp b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_common.hpp
new file mode 100644
index 0000000..855710b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_common.hpp
@@ -0,0 +1,196 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+void print_helper_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd_dl(bool do_verification,
+                             int init_method,
+                             bool time_kernel,
+                             const ck::utils::conv::ConvParam& conv_param,
+                             const HostTensorDescriptor& in_g_n_c_wis_desc,
+                             const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                             const HostTensorDescriptor& out_g_n_k_wos_desc,
+                             const InElementOp& in_element_op,
+                             const WeiElementOp& wei_element_op,
+                             const OutElementOp& out_element_op)
+{
+    using DDataType = ck::remove_cvref_t<ck::tuple_element_t<0, DsDataType>>;
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<DDataType> bias(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 3});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 3});
+        bias.GenerateTensorValue(GeneratorTensor_2<DDataType>{-2, 3});
+        break;
+    case 2:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        bias.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
+        wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{-1});
+        bias.GenerateTensorValue(GeneratorTensor_1<DDataType>{1});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(DDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), d_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), d_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(
+        in_device_buf.GetDeviceBuffer(),
+        wei_device_buf.GetDeviceBuffer(),
+        std::array<const void*, 1>{bias_device_buf.GetDeviceBuffer()},
+        out_device_buf.GetDeviceBuffer(),
+        a_g_n_c_wis_lengths,
+        a_g_n_c_wis_strides,
+        b_g_k_c_xs_lengths,
+        b_g_k_c_xs_strides,
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_g_n_k_wos_lengths}},
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_g_n_k_wos_strides}},
+        e_g_n_k_wos_lengths,
+        e_g_n_k_wos_strides,
+        conv_filter_strides,
+        conv_filter_dilations,
+        input_left_pads,
+        input_right_pads,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        std::cout << "wrong! device_conv with the specified compilation parameters does not "
+                     "support this Conv problem"
+                  << std::endl;
+        return true;
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<
+            NDimSpatial,
+            InDataType,
+            WeiDataType,
+            OutDataType,
+            InElementOp,
+            WeiElementOp,
+            ck::tensor_operation::element_wise::PassThrough>();
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+        auto ref_argument =
+            ref_conv.MakeArgument(in,
+                                  wei,
+                                  out_host,
+                                  conv_param.conv_filter_strides_,
+                                  conv_param.conv_filter_dilations_,
+                                  conv_param.input_left_pads_,
+                                  conv_param.input_right_pads_,
+                                  in_element_op,
+                                  wei_element_op,
+                                  ck::tensor_operation::element_wise::PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // cde_elementwise
+        out_host.ForEach(
+            [&](auto&, auto idx) { out_element_op(out_host(idx), out_host(idx), bias(idx)); });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(
+            out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return true;
+}
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
new file mode 100644
index 0000000..db5a7f0
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_dl_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using AccDataType = float;
+using DsDataType  = ck::Tuple<ck::half_t>;
+using OutDataType = ck::half_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+// clang-format off
+using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
+// ######|        NDim|     InData|     WeiData|   MultpleD|     OutData|     AccData| InLayout| WeiLayout|            MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|     Spatial|       Type|        Type|       Type|        Type|        Type|         |          |               Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|            |           |            |           |            |            |         |          |                     |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|            |           |            |           |            |            |         |          |                     |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < NDimSpatial, InDataType, WeiDataType, DsDataType, OutDataType, AccDataType, InLayout, WeiLayout, ck::Tuple<OutLayout>, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
+// clang-format on
+
+#include "run_convnd_fwd_dl_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_dl_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
new file mode 100644
index 0000000..964d784
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_dl_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType  = float;
+using WeiDataType = float;
+using AccDataType = float;
+using DsDataType  = ck::Tuple<float>;
+using OutDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+// clang-format off
+using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
+// ######|        NDim|     InData|     WeiData|   MultpleD|     OutData|     AccData| InLayout| WeiLayout|            MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|     Spatial|       Type|        Type|       Type|        Type|        Type|         |          |               Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|            |           |            |           |            |            |         |          |                     |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|            |           |            |           |            |            |         |          |                     |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < NDimSpatial, InDataType, WeiDataType, DsDataType, OutDataType, AccDataType, InLayout, WeiLayout, ck::Tuple<OutLayout>, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
+// clang-format on
+
+#include "run_convnd_fwd_dl_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_dl_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
new file mode 100644
index 0000000..b0cd88f
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_dl_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType  = int8_t;
+using WeiDataType = int8_t;
+using AccDataType = int32_t;
+using DsDataType  = ck::Tuple<int8_t>;
+using OutDataType = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+// clang-format off
+using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
+// ######|        NDim|     InData|     WeiData|   MultpleD|     OutData|     AccData| InLayout| WeiLayout|            MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|     Spatial|       Type|        Type|       Type|        Type|        Type|         |          |               Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|            |           |            |           |            |            |         |          |                     |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|            |           |            |           |            |            |         |          |                     |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < NDimSpatial, InDataType, WeiDataType, DsDataType, OutDataType, AccDataType, InLayout, WeiLayout, ck::Tuple<OutLayout>, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
+// clang-format on
+
+#include "run_convnd_fwd_dl_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_dl_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
new file mode 100644
index 0000000..d55d315
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType       = ck::bhalf_t;
+using WeiDataType      = ck::bhalf_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using OutDataType      = ck::bhalf_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        32,          // KPerBlock
+        8,           // AK1
+        8,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 32, 1, 8>,
+        8>;
+
+#include "run_convnd_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
new file mode 100644
index 0000000..d84afba
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType       = ck::half_t;
+using WeiDataType      = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::half_t;
+using OutDataType      = ck::half_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::UnaryConvert;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        32,          // KPerBlock
+        8,           // AK1
+        8,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 32, 1, 8>,
+        8>;
+
+#include "run_convnd_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
new file mode 100644
index 0000000..f5acc54
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType       = float;
+using WeiDataType      = float;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using OutDataType      = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        16,          // KPerBlock
+        4,           // AK1
+        4,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        4,           // ABlockTransferSrcScalarPerVector
+        4,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        4,           // BBlockTransferSrcScalarPerVector
+        4,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 16, 1, 16>,
+        4>;
+
+#include "run_convnd_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
new file mode 100644
index 0000000..8d69797
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType       = double;
+using WeiDataType      = double;
+using AccDataType      = double;
+using CShuffleDataType = double;
+using OutDataType      = double;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        128,         // NPerBlock
+        8,           // KPerBlock
+        2,           // AK1
+        2,           // BK1
+        16,          // MPerXdl
+        16,          // NPerXdl
+        4,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        2,           // ABlockTransferSrcScalarPerVector
+        2,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        2,           // BBlockTransferSrcScalarPerVector
+        2,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 16, 1, 16>,
+        1>;
+
+#include "run_convnd_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
new file mode 100644
index 0000000..99f7f25
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType       = int8_t;
+using WeiDataType      = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int8_t;
+using OutDataType      = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        64,          // KPerBlock
+        16,          // AK1
+        16,          // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        16,          // ABlockTransferSrcScalarPerVector
+        16,          // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        16,          // BBlockTransferSrcScalarPerVector
+        16,          // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 64, 1, 4>,
+        16>;
+
+#include "run_convnd_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc b/3rdparty/composable_kernel/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
new file mode 100644
index 0000000..697ada1
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_convnd_fwd_dl_example(int argc, char* argv[])
+{
+    print_helper_msg();
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv_param{
+        2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        do_verification                   = std::stoi(argv[1]);
+        init_method                       = std::stoi(argv[2]);
+        time_kernel                       = std::stoi(argv[3]);
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+
+        conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto run = [&](auto ndim_spatial, auto in_layout, auto wei_layout, auto out_layout) {
+        constexpr ck::index_t ndim_spatial_value = ndim_spatial.value;
+        std::cout << "ndim_spatial_value: " << ndim_spatial_value << std::endl;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_grouped_conv_fwd_dl<
+            ndim_spatial_value,
+            InDataType,
+            WeiDataType,
+            DsDataType,
+            OutDataType,
+            InElementOp,
+            WeiElementOp,
+            OutElementOp,
+            DeviceGroupedConvNDFwdInstance<ndim_spatial_value, InLayout, WeiLayout, OutLayout>>(
+            do_verification,
+            init_method,
+            time_kernel,
+            conv_param,
+            in_g_n_c_wis_desc,
+            wei_g_k_c_xs_desc,
+            out_g_n_k_wos_desc,
+            in_element_op,
+            wei_element_op,
+            out_element_op);
+    };
+
+    namespace ctc = ck::tensor_layout::convolution;
+
+    if(conv_param.num_dim_spatial_ == 1)
+    {
+        return run(ck::Number<1>{}, ctc::GNWC{}, ctc::GKXC{}, ctc::GNWK{});
+    }
+    else if(conv_param.num_dim_spatial_ == 2)
+    {
+        return run(ck::Number<2>{}, ctc::GNHWC{}, ctc::GKYXC{}, ctc::GNHWK{});
+    }
+    else if(conv_param.num_dim_spatial_ == 3)
+    {
+        return run(ck::Number<3>{}, ctc::GNDHWC{}, ctc::GKZYXC{}, ctc::GNDHWK{});
+    }
+
+    return true;
+}
diff --git a/3rdparty/composable_kernel/example/09_convnd_fwd/run_convnd_fwd_example.inc b/3rdparty/composable_kernel/example/09_convnd_fwd/run_convnd_fwd_example.inc
new file mode 100644
index 0000000..36a6805
--- /dev/null
+++ b/3rdparty/composable_kernel/example/09_convnd_fwd/run_convnd_fwd_example.inc
@@ -0,0 +1,97 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_convnd_fwd_example(int argc, char* argv[])
+{
+    print_helper_msg();
+
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv_param{
+        2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        do_verification                   = std::stoi(argv[1]);
+        init_method                       = std::stoi(argv[2]);
+        time_kernel                       = std::stoi(argv[3]);
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+
+        conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto run = [&](auto ndim_spatial, auto in_layout, auto wei_layout, auto out_layout) {
+        constexpr ck::index_t ndim_spatial_value = ndim_spatial.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_grouped_conv_fwd<
+            ndim_spatial_value,
+            InDataType,
+            WeiDataType,
+            OutDataType,
+            InElementOp,
+            WeiElementOp,
+            OutElementOp,
+            DeviceGroupedConvNDFwdInstance<ndim_spatial_value, InLayout, WeiLayout, OutLayout>>(
+            do_verification,
+            init_method,
+            time_kernel,
+            conv_param,
+            in_g_n_c_wis_desc,
+            wei_g_k_c_xs_desc,
+            out_g_n_k_wos_desc,
+            in_element_op,
+            wei_element_op,
+            out_element_op);
+    };
+
+    namespace ctc = ck::tensor_layout::convolution;
+
+    if(conv_param.num_dim_spatial_ == 1)
+    {
+        return run(ck::Number<1>{}, ctc::GNWC{}, ctc::GKXC{}, ctc::GNWK{});
+    }
+    else if(conv_param.num_dim_spatial_ == 2)
+    {
+        return run(ck::Number<2>{}, ctc::GNHWC{}, ctc::GKYXC{}, ctc::GNHWK{});
+    }
+    else if(conv_param.num_dim_spatial_ == 3)
+    {
+        return run(ck::Number<3>{}, ctc::GNDHWC{}, ctc::GKZYXC{}, ctc::GNDHWK{});
+    }
+
+    return true;
+}
diff --git a/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
new file mode 100644
index 0000000..98941b4
--- /dev/null
+++ b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
@@ -0,0 +1,16 @@
+add_custom_target(example_convnd_fwd_reduce_xdl)
+
+add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
+add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
+add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp)
+add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
+
+add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
+add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
+add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
+add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
+
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp)
+  add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4)
+endif(USE_BITINT_EXTENSION_INT4)
diff --git a/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
new file mode 100644
index 0000000..00e370f
--- /dev/null
+++ b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
@@ -0,0 +1,159 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cassert>
+#include <cstdint>
+#include <cstdlib>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <type_traits>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using BF16 = ck::bhalf_t;
+using FP16 = ck::half_t;
+using FP32 = float;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using I4 = ck::int4_t;
+#endif
+using I8  = std::int8_t;
+using I32 = std::int32_t;
+
+template <typename ALay, typename BLay, typename DELay, typename RLay>
+struct LayoutSetting
+{
+    using ALayout  = ALay;
+    using BLayout  = BLay;
+    using DELayout = DELay;
+    using RLayout  = RLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct LayoutSettingSelector;
+
+namespace ctl = ck::tensor_layout::convolution;
+
+template <>
+struct LayoutSettingSelector<1> final : LayoutSetting<ctl::GNWC, ctl::GKXC, ctl::GNWK, ctl::GNW>
+{
+};
+
+template <>
+struct LayoutSettingSelector<2> final : LayoutSetting<ctl::GNHWC, ctl::GKYXC, ctl::GNHWK, ctl::GNHW>
+{
+};
+
+template <>
+struct LayoutSettingSelector<3> final
+    : LayoutSetting<ctl::GNDHWC, ctl::GKZYXC, ctl::GNDHWK, ctl::GNDHW>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using ALayout = typename LayoutSettingSelector<NDimSpatial>::ALayout;
+
+template <ck::index_t NDimSpatial>
+using BLayout = typename LayoutSettingSelector<NDimSpatial>::BLayout;
+
+template <ck::index_t NDimSpatial>
+using DELayout = typename LayoutSettingSelector<NDimSpatial>::DELayout;
+
+template <ck::index_t NDimSpatial>
+using RLayout = typename LayoutSettingSelector<NDimSpatial>::RLayout;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ck::utils::conv::ConvParam& problem_size,
+                           ExecutionConfig& config)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        problem_size                      = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
+
+inline HostTensorDescriptor
+make_r0_host_tensor_descriptor(const ck::utils::conv::ConvParam& problem_size)
+{
+    std::vector<ck::index_t> dimensions{problem_size.G_, problem_size.N_};
+
+    ck::ranges::copy(problem_size.output_spatial_lengths_, std::back_inserter(dimensions));
+
+    return HostTensorDescriptor(dimensions);
+}
+
+template <typename Lengths, typename Strides>
+void unpack_host_tensor_descriptor(const HostTensorDescriptor& descriptor,
+                                   Lengths& lengths,
+                                   Strides& strides)
+{
+    assert(size(descriptor.GetLengths()) == size(lengths));
+    std::copy_n(begin(descriptor.GetLengths()), size(descriptor.GetLengths()), begin(lengths));
+
+    assert(size(descriptor.GetStrides()) == size(strides));
+    std::copy_n(begin(descriptor.GetStrides()), size(descriptor.GetStrides()), begin(strides));
+}
diff --git a/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp
new file mode 100644
index 0000000..6ff29b4
--- /dev/null
+++ b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType         = BF16;
+using BDataType         = BF16;
+using AccDataType       = FP32;
+using CShuffleDataType  = FP32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = BF16;
+using ReduceAccDataType = FP32;
+using R0DataType        = FP32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp
new file mode 100644
index 0000000..02c19c2
--- /dev/null
+++ b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType         = FP16;
+using BDataType         = FP16;
+using AccDataType       = FP32;
+using CShuffleDataType  = FP32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = FP16;
+using ReduceAccDataType = FP32;
+using R0DataType        = FP32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp
new file mode 100644
index 0000000..679bb5c
--- /dev/null
+++ b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType         = FP32;
+using BDataType         = FP32;
+using AccDataType       = FP32;
+using CShuffleDataType  = FP32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = FP32;
+using ReduceAccDataType = FP32;
+using R0DataType        = FP32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp
new file mode 100644
index 0000000..abdbdaf
--- /dev/null
+++ b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#define BUILD_INT4_EXAMPLE
+
+#include "common.hpp"
+
+using ADataType         = I4;
+using BDataType         = I4;
+using KernelADataType   = I8;
+using KernelBDataType   = I8;
+using AccDataType       = I32;
+using CShuffleDataType  = I32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = I32;
+using ReduceAccDataType = I32;
+using R0DataType        = I32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp
new file mode 100644
index 0000000..cf86afa
--- /dev/null
+++ b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType         = I8;
+using BDataType         = I8;
+using AccDataType       = I32;
+using CShuffleDataType  = I32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = I32;
+using ReduceAccDataType = I32;
+using R0DataType        = I32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
new file mode 100644
index 0000000..b3a3891
--- /dev/null
+++ b/3rdparty/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
@@ -0,0 +1,307 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+template <ck::index_t NDimSpatial>
+using DeviceInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
+//######| NDimSpatial|              ALayout|              BLayout|              DELayout|              RLayout|           AData|           BData|     AccData|         CShuffle|     DsData|     EData|     ReduceAccData|     RsData|           A|           B|          CDE|          Qs|          Rs|           Thread|           Global|           Conv|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|    CDRThreadTransfer|                  CDE|    RThreadTransfer|
+//######|            |                     |                     |                      |                     |            Type|            Type|        Type|         DataType|       Type|      Type|              Type|       Type| Elementwise| Elementwise|  Elementwise| Elementwise| Elementwise|           Reduce|           Reduce|            Fwd|Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|       ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
+//######|            |                     |                     |                      |                     |                |                |            |                 |           |          |                  |           |   Operation|   Operation|    Operation|   Operation|   Operation|        Operation|        Operation| Specialization|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _MPerBlock_NPerBlock|      ScalarPerVector|         _MPerBlock|
+//######|            |                     |                     |                      |                     |                |                |            |                 |           |          |                  |           |            |            |             |            |            |                 |                 |               |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |           _NPerBlock|                   |
+#ifdef BUILD_INT4_EXAMPLE
+        < NDimSpatial, ALayout<NDimSpatial>, BLayout<NDimSpatial>, DELayout<NDimSpatial>, RLayout<NDimSpatial>, KernelADataType, KernelBDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType,  AElementOp,  BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp,       ConvSpec,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<64, 4>,                    4,                  1>;
+#else
+        < NDimSpatial, ALayout<NDimSpatial>, BLayout<NDimSpatial>, DELayout<NDimSpatial>, RLayout<NDimSpatial>,       ADataType,       BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType,  AElementOp,  BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp,       ConvSpec,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<64, 4>,                    4,                  1>;
+#endif
+
+template <ck::index_t NDimSpatial>
+using HostInstance = ck::tensor_operation::host::ReferenceConvFwd
+        <NDimSpatial, ADataType, BDataType, EDataType, AElementOp, BElementOp, PassThrough>;
+// clang-format on
+
+template <ck::index_t NDimSpatial>
+bool run_convnd_fwd_max(const ck::utils::conv::ConvParam& problem_size,
+                        const ExecutionConfig& config)
+{
+    static_assert(1 <= NDimSpatial && NDimSpatial <= 3, "Unsupported NDimSpatial");
+
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+
+    const auto conv_input_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ALayout<NDimSpatial>>(
+            problem_size);
+
+    const auto conv_weight_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<BLayout<NDimSpatial>>(
+            problem_size);
+
+    const auto conv_output_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<DELayout<NDimSpatial>>(
+            problem_size);
+
+    const auto r0_desc = make_r0_host_tensor_descriptor(problem_size);
+
+    Tensor<ADataType> conv_input(conv_input_g_n_c_wis_desc);
+    Tensor<BDataType> conv_weight(conv_weight_g_k_c_xs_desc);
+    Tensor<EDataType> conv_output_device(conv_output_g_n_k_wos_desc);
+    Tensor<R0DataType> r0_device(r0_desc);
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-8, 7}(conv_input);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-8, 7}(conv_weight);
+        break;
+    default:
+        ck::utils::FillUniformDistribution<ADataType>{-5, 5}(conv_input);
+        ck::utils::FillUniformDistribution<BDataType>{-5, 5}(conv_weight);
+    }
+
+    DeviceMem conv_input_device_buf(sizeof(ADataType) * conv_input.mDesc.GetElementSpaceSize());
+    DeviceMem conv_weight_device_buf(sizeof(BDataType) * conv_weight.mDesc.GetElementSpaceSize());
+    DeviceMem conv_output_device_buf(sizeof(EDataType) *
+                                     conv_output_device.mDesc.GetElementSpaceSize());
+    DeviceMem r0_device_buf(sizeof(R0DataType) * r0_device.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<KernelADataType> conv_input_converted(conv_input);
+    const Tensor<KernelBDataType> conv_weight_converted(conv_weight);
+
+    conv_input_device_buf.ToDevice(conv_input_converted.mData.data());
+    conv_weight_device_buf.ToDevice(conv_weight_converted.mData.data());
+#else
+    conv_input_device_buf.ToDevice(conv_input.mData.data());
+    conv_weight_device_buf.ToDevice(conv_weight.mData.data());
+#endif
+
+    std::array<ck::index_t, NDimSpatial + 3> conv_input_g_n_c_wis_lengths{},
+        conv_input_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> conv_weight_g_k_c_xs_lengths{},
+        conv_weight_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> conv_output_g_n_k_wos_lengths{},
+        conv_output_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 2> r0_lengths{}, r0_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{}, conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{}, input_right_pads{};
+
+    unpack_host_tensor_descriptor(
+        conv_input_g_n_c_wis_desc, conv_input_g_n_c_wis_lengths, conv_input_g_n_c_wis_strides);
+    unpack_host_tensor_descriptor(
+        conv_weight_g_k_c_xs_desc, conv_weight_g_k_c_xs_lengths, conv_weight_g_k_c_xs_strides);
+    unpack_host_tensor_descriptor(
+        conv_output_g_n_k_wos_desc, conv_output_g_n_k_wos_lengths, conv_output_g_n_k_wos_strides);
+    unpack_host_tensor_descriptor(r0_desc, r0_lengths, r0_strides);
+
+    ck::ranges::copy(problem_size.conv_filter_strides_, begin(conv_filter_strides));
+    ck::ranges::copy(problem_size.conv_filter_dilations_, begin(conv_filter_dilations));
+    ck::ranges::copy(problem_size.input_left_pads_, begin(input_left_pads));
+    ck::ranges::copy(problem_size.input_right_pads_, begin(input_right_pads));
+
+    // run Conv + Reduction on device
+    auto conv     = DeviceInstance<NDimSpatial>{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(conv_input_device_buf.GetDeviceBuffer(),
+                                      conv_weight_device_buf.GetDeviceBuffer(),
+                                      std::array<const void*, 0>{},
+                                      conv_output_device_buf.GetDeviceBuffer(),
+                                      {r0_device_buf.GetDeviceBuffer()},
+                                      conv_input_g_n_c_wis_lengths,
+                                      conv_input_g_n_c_wis_strides,
+                                      conv_weight_g_k_c_xs_lengths,
+                                      conv_weight_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      conv_output_g_n_k_wos_lengths,
+                                      conv_output_g_n_k_wos_strides,
+                                      r0_lengths,
+                                      r0_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      AElementOp{},
+                                      BElementOp{},
+                                      CDEElementOp{},
+                                      QsElementOp{},
+                                      RsElementOp{});
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_conv with the specified compilation parameters does "
+                     "not support this Conv problem"
+                  << std::endl;
+        return false;
+    }
+
+    const float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    const std::size_t flop      = problem_size.GetFlops();
+    const std::size_t num_btype = problem_size.GetByte<ADataType, BDataType, EDataType>();
+
+    const float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    const float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        Tensor<EDataType> conv_output_host(conv_output_g_n_k_wos_desc);
+
+        // run Conv + Reduction on host
+        auto ref_conv     = HostInstance<NDimSpatial>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(conv_input,
+                                                  conv_weight,
+                                                  conv_output_host,
+                                                  problem_size.conv_filter_strides_,
+                                                  problem_size.conv_filter_dilations_,
+                                                  problem_size.input_left_pads_,
+                                                  problem_size.input_right_pads_,
+                                                  AElementOp{},
+                                                  BElementOp{},
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        Tensor<R0DataType> r0_host(r0_device.mDesc);
+
+        auto reduce0_op = RsThreadReduceOp{}[ck::Number<0>{}];
+
+        auto& output_dims = conv_output_g_n_k_wos_desc.GetLengths();
+
+        if constexpr(NDimSpatial == 1)
+        {
+            for(std::size_t g = 0; g < output_dims[0]; ++g)
+            {
+                for(std::size_t n = 0; n < output_dims[1]; ++n)
+                {
+                    for(std::size_t w = 0; w < output_dims[3]; ++w)
+                    {
+                        auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                        for(std::size_t k = 0; k < output_dims[2]; ++k)
+                        {
+
+                            auto e_val =
+                                ck::type_convert<ReduceAccDataType>(conv_output_host(g, n, k, w));
+                            reduce0_op(reduce0_acc, e_val);
+                        }
+                        r0_host(g, n, w) = ck::type_convert<R0DataType>(reduce0_acc);
+                    }
+                }
+            }
+        }
+        else if constexpr(NDimSpatial == 2)
+        {
+            for(std::size_t g = 0; g < output_dims[0]; ++g)
+            {
+                for(std::size_t n = 0; n < output_dims[1]; ++n)
+                {
+                    for(std::size_t h = 0; h < output_dims[3]; ++h)
+                    {
+                        for(std::size_t w = 0; w < output_dims[4]; ++w)
+                        {
+                            auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                            for(std::size_t k = 0; k < output_dims[2]; ++k)
+                            {
+
+                                auto e_val = ck::type_convert<ReduceAccDataType>(
+                                    conv_output_host(g, n, k, h, w));
+                                reduce0_op(reduce0_acc, e_val);
+                            }
+                            r0_host(g, n, h, w) = ck::type_convert<R0DataType>(reduce0_acc);
+                        }
+                    }
+                }
+            }
+        }
+        else if constexpr(NDimSpatial == 3)
+        {
+            for(std::size_t g = 0; g < output_dims[0]; ++g)
+            {
+                for(std::size_t n = 0; n < output_dims[1]; ++n)
+                {
+                    for(std::size_t d = 0; d < output_dims[3]; ++d)
+                    {
+                        for(std::size_t h = 0; h < output_dims[4]; ++h)
+                        {
+                            for(std::size_t w = 0; w < output_dims[5]; ++w)
+                            {
+                                auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                                for(std::size_t k = 0; k < output_dims[2]; ++k)
+                                {
+
+                                    auto e_val = ck::type_convert<ReduceAccDataType>(
+                                        conv_output_host(g, n, k, d, h, w));
+                                    reduce0_op(reduce0_acc, e_val);
+                                }
+                                r0_host(g, n, d, h, w) = ck::type_convert<R0DataType>(reduce0_acc);
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        conv_output_device_buf.FromDevice(conv_output_device.mData.data());
+        r0_device_buf.FromDevice(r0_device.mData.data());
+
+        return ck::utils::check_err(conv_output_device,
+                                    conv_output_host,
+                                    "Error: incorrect results! (Matrix E)",
+                                    1e-5f,
+                                    1e-4f) &&
+               ck::utils::check_err(
+                   r0_device, r0_host, "Error: incorrect results! (Matrix R0)", 1e-5f, 1e-4f);
+    }
+
+    return true;
+}
+
+bool run_convnd_fwd_max_example(int argc, char* argv[])
+{
+    ck::utils::conv::ConvParam problem_size{
+        2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+    ExecutionConfig config;
+
+    if(!parse_cmd_args(argc, argv, problem_size, config))
+    {
+        return false;
+    }
+
+    switch(problem_size.num_dim_spatial_)
+    {
+    case 1: return run_convnd_fwd_max<1>(problem_size, config);
+    case 2: return run_convnd_fwd_max<2>(problem_size, config);
+    case 3: return run_convnd_fwd_max<3>(problem_size, config);
+    }
+
+    return false;
+}
diff --git a/3rdparty/composable_kernel/example/12_reduce/CMakeLists.txt b/3rdparty/composable_kernel/example/12_reduce/CMakeLists.txt
new file mode 100644
index 0000000..6e58ed9
--- /dev/null
+++ b/3rdparty/composable_kernel/example/12_reduce/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_example_executable(example_reduce_blockwise reduce_blockwise.cpp)
+add_example_executable(example_reduce_multiblock_atomic_add reduce_multiblock_atomic_add.cpp)
+add_example_executable(example_reduce_blockwise_two_call reduce_blockwise_two_call.cpp)
diff --git a/3rdparty/composable_kernel/example/12_reduce/README.md b/3rdparty/composable_kernel/example/12_reduce/README.md
new file mode 100644
index 0000000..76d2852
--- /dev/null
+++ b/3rdparty/composable_kernel/example/12_reduce/README.md
@@ -0,0 +1,62 @@
+# Instructions for ```example_reduce_blockwise```
+
+## Run ```example_reduce_blockwise```
+```bash
+# -D <xxx> : input 3d/4d/5d tensor lengths
+# -R <xxx> : reduce dimension ids
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64, 7: int4)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 0 2 1
+```
+
+Result
+```
+./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 0 2 1
+launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.238063 ms, 264.285 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
+```
+
+## Run ```example_reduce_multiblock_atomic_add```
+```bash
+# -D <xxx> : input 3d/4d/5d tensor lengths
+# -R <xxx> : reduce dimension ids
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: data type (0: fp32, 1: fp64)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+./bin/example_reduce_multiblock_atomic_add -D 16,64,32,960 -v 1 0 2 0
+```
+
+Result
+```
+./bin/example_reduce_multiblock_atomic_add -D 16,64,32,960 -v 1 0 2 0
+Perf: 0 ms, inf GB/s, DeviceReduceMultiBlock<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
+echo $?
+0
+```
+
+# Instructions for ```example_reduce_blockwise_two_call```
+
+## Run ```example_reduce_blockwise_two_call```
+```bash
+#arg1:  verification (0=no, 1=yes(
+#arg2:  initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3:  time kernel (0=no, 1=yes)
+./bin/example_reduce_blockwise_two_call 1 2 1
+```
+
+Result
+```
+./bin/example_reduce_blockwise_two_call 1 2 1
+launch_and_time_kernel: grid_dim {204800, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {6400, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 2.1791 ms, 771.42 GB/s, DeviceReduceBlockWise<256,M_C32_S1,K_C8_S1,InSrcVectorDim_1_InSrcVectorSize_1_OutDstVectorSize_1> => DeviceReduceBlockWise<256,M_C256_S1,K_C1_S1,InSrcVectorDim_1_InSrcVectorSize_1_OutDstVectorSize_1>
+```
diff --git a/3rdparty/composable_kernel/example/12_reduce/reduce_blockwise.cpp b/3rdparty/composable_kernel/example/12_reduce/reduce_blockwise.cpp
new file mode 100644
index 0000000..a7ee999
--- /dev/null
+++ b/3rdparty/composable_kernel/example/12_reduce/reduce_blockwise.cpp
@@ -0,0 +1,299 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/utility/reduction_enums.hpp"
+#include "reduce_blockwise_impl.hpp"
+#include "reduce_example_common.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths = {16, 64, 32, 960};
+    std::vector<int> reduceDims   = {0, 1, 2};
+    std::vector<float> scales     = {1.0f, 0.0f};
+
+    bool do_verification = true;
+    int data_type        = 1;
+    int init_method      = 2;
+    bool time_kernel     = true;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of to-reduce dimensions"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64, 7: int4)"
+                  << std::endl;
+        std::cout << "Arg2 -- init method (0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg3 -- time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:l:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 3 > argc)
+        {
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+        };
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        return (0);
+    };
+};
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ReduceTensorOp ReduceOpId,
+          index_t PropagateNan,
+          index_t OutputIndex>
+bool reduce_blockwise_test(bool do_verification,
+                           int init_method,
+                           bool time_kernel,
+                           const std::vector<size_t>& inLengths,
+                           const std::vector<int>& reduceDims,
+                           float alpha,
+                           float beta)
+{
+    bool matched = false;
+    int result   = 0;
+
+    const auto tuple_object = reduce_shape_instances{};
+
+    static_for<0, std::tuple_size<reduce_shape_instances>::value, 1>{}([&](auto i) {
+        if(matched)
+            return;
+
+        using ShapeType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
+
+        if(ShapeType::Rank_ != inLengths.size() || ShapeType::NumReduceDim_ != reduceDims.size())
+            return;
+
+        std::array<int, ShapeType::NumReduceDim_> arrReduceDims;
+
+        ck::ranges::copy(reduceDims, arrReduceDims.begin());
+
+        result = reduce_blockwise_impl<InOutDataType,
+                                       AccDataType,
+                                       ReduceOpId,
+                                       ShapeType::Rank_,
+                                       ShapeType::NumReduceDim_,
+                                       PropagateNan,
+                                       OutputIndex>(
+            do_verification, init_method, time_kernel, inLengths, arrReduceDims, alpha, beta);
+
+        matched = true;
+    });
+
+    return (result == 0) ? true : false;
+};
+
+constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::AVG;
+constexpr bool PropagateNan         = true;
+constexpr bool OutputIndex          = false;
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        SimpleAppArgs arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            pass = reduce_blockwise_test<ck::half_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 1)
+        {
+            pass = reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 3)
+        {
+            pass = reduce_blockwise_test<int8_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 5)
+        {
+            pass = reduce_blockwise_test<ck::bhalf_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 6)
+        {
+            pass = reduce_blockwise_test<double, double, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        else if(arg.data_type == 7)
+        {
+            pass = reduce_blockwise_test<int4_t, int32_t, ReduceTensorOp::AVG, false, false>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+
+            pass = pass && reduce_blockwise_test<int4_t, int8_t, ReduceTensorOp::MAX, false, false>(
+                               arg.do_verification,
+                               arg.init_method,
+                               arg.time_kernel,
+                               arg.inLengths,
+                               arg.reduceDims,
+                               arg.scales[0],
+                               arg.scales[1]);
+        }
+#endif
+    }
+    else
+    {
+        // for testing half_t
+        pass =
+            pass && reduce_blockwise_test<ck::half_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                        true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing float
+        pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                           true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing double
+        pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                           true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing bhalf_t
+        pass = pass &&
+               reduce_blockwise_test<ck::bhalf_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                   true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing int8_t
+        pass =
+            pass && reduce_blockwise_test<int8_t, int32_t, ReduceOpId, PropagateNan, OutputIndex>(
+                        true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        // for testing int4_t using AVG operation
+        pass = pass && reduce_blockwise_test<int4_t, int32_t, ReduceTensorOp::AVG, false, false>(
+                           true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing int4_t using MAX operation
+        pass = pass && reduce_blockwise_test<int4_t, int8_t, ReduceTensorOp::MAX, false, false>(
+                           true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+#endif
+        // for testing 3D input
+        pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                           true, 2, true, {16, 64, 960}, {0, 1}, 1.0f, 0.0f);
+
+        // for testing 5D input
+        pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                           true, 2, true, {16, 64, 32, 2, 960}, {0, 1, 2, 3}, 1.0f, 0.0f);
+    };
+
+    return (pass ? 0 : 1);
+};
diff --git a/3rdparty/composable_kernel/example/12_reduce/reduce_blockwise_impl.hpp b/3rdparty/composable_kernel/example/12_reduce/reduce_blockwise_impl.hpp
new file mode 100644
index 0000000..7bafd2d
--- /dev/null
+++ b/3rdparty/composable_kernel/example/12_reduce/reduce_blockwise_impl.hpp
@@ -0,0 +1,338 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_reduction.hpp"
+
+#include "reduce_example_common.hpp"
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          ck::index_t Rank,
+          ck::index_t NumReduceDim,
+          bool PropagateNan,
+          bool OutputIndex>
+int reduce_blockwise_impl(bool do_verification,
+                          int init_method,
+                          bool time_kernel,
+                          const std::vector<size_t>& inLengths,
+                          const std::array<int, NumReduceDim>& reduceDims,
+                          float alpha,
+                          float beta)
+
+{
+    using namespace ck;
+    using namespace ck::tensor_operation::device;
+
+    constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
+    constexpr bool op_support_indices =
+        (ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
+         ReduceOpId == ReduceTensorOp::AMAX);
+
+    constexpr bool invalid_reduce_1 = OutputIndex && !op_support_indices;
+
+    // 1) If InOutDataType is half_t, must use half_t as AccDataType for indexable reduction
+    // operations 2) If InOutDataType is half_t, must use float as AccDataType for non-indexable
+    // reduction operations
+    constexpr bool invalid_reduce_2 =
+        std::is_same<InOutDataType, half_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, float>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, half_t>::value));
+
+    // 1) If InOutDataType is float, must use float as AccDataType for indexable reduction
+    // operations
+    constexpr bool invalid_reduce_3 =
+        std::is_same<InOutDataType, float>::value &&
+        (op_support_indices && !std::is_same<AccDataType, float>::value);
+
+    // 1) If InOutDataType is int8_t or int4_t, must use int8_t as AccDataType for indexable
+    // reduction operations 2) If InOutDataType is int8_t or int4_t, must use int32_t as AccDataType
+    // for non-indexable reduction operations
+    constexpr bool invalid_reduce_4 =
+        std::is_same<InOutDataType, int8_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, int8_t>::value));
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    constexpr bool invalid_reduce_4_2 =
+        std::is_same<InOutDataType, int4_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, int8_t>::value));
+#endif
+
+    // 1) If InOutDataType is int8_t or int4_t, the supported operation must be either indexable
+    // operations or ADD/AVG
+    constexpr bool invalid_reduce_5 = std::is_same<InOutDataType, int8_t>::value &&
+                                      (!op_support_indices && ReduceOpId != ReduceTensorOp::ADD &&
+                                       ReduceOpId != ReduceTensorOp::AVG);
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    constexpr bool invalid_reduce_5_2 = std::is_same<InOutDataType, int4_t>::value &&
+                                        (!op_support_indices && ReduceOpId != ReduceTensorOp::ADD &&
+                                         ReduceOpId != ReduceTensorOp::AVG);
+#endif
+
+    // 1) If InOutDataType is bhalf_t, must use float as AccDataType for all reduction operations
+    constexpr bool invalid_reduce_6 =
+        std::is_same<InOutDataType, bhalf_t>::value && !std::is_same<AccDataType, float>::value;
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    constexpr bool invalid_reduce =
+        (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 || invalid_reduce_4 ||
+         invalid_reduce_5 || invalid_reduce_6 || invalid_reduce_4_2 || invalid_reduce_5_2);
+#else
+    constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
+                                     invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
+#endif
+
+    if constexpr(invalid_reduce)
+    {
+        std::cerr << "The reduction setting is invalid, exiting!" << std::endl;
+        return (-1);
+    };
+
+    using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+    using InElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+    using AccElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    using InOutDataTypeInDevice = typename std::
+        conditional<std::is_same<InOutDataType, int4_t>::value, int8_t, InOutDataType>::type;
+#else
+    using InOutDataTypeInDevice   = InOutDataType;
+#endif
+
+    using DeviceReduceInstance =
+        ck::tensor_operation::device::DeviceReduceMultiBlock<InOutDataTypeInDevice,
+                                                             AccDataType,
+                                                             InOutDataTypeInDevice,
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             ReduceOperation,
+                                                             InElementwiseOperation,
+                                                             AccElementwiseOperation,
+                                                             InMemoryDataOperationEnum::Set,
+                                                             PropagateNan,
+                                                             OutputIndex,
+                                                             false, // HaveIndexInputIfOutputIndex
+                                                             256,   // BlockSize
+                                                             4,     // MThreadClusterSize
+                                                             64,    // KThreadClusterSize
+                                                             1,     // MThreadSliceSize
+                                                             1,     // KThreadSliceSize
+                                                             0,     // InSrcVectorDim
+                                                             1,     // InSrceVectorSize
+                                                             1>;    // OutDstVectorSize
+
+    Tensor<InOutDataType> in(inLengths);
+
+    std::vector<size_t> outLengths;
+
+    auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
+
+    if(invariantDims.empty())
+        outLengths.push_back(1);
+    else
+        for(auto dim : invariantDims)
+            outLengths.push_back(inLengths[dim]);
+
+    Tensor<InOutDataType> out_ref(outLengths);
+    Tensor<InOutDataType> out(outLengths);
+    Tensor<int> out_indices_ref(outLengths);
+    Tensor<int> out_indices(outLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = out.mDesc.GetStrides();
+
+    size_t invariant_total_length = out.mDesc.GetElementSize();
+    size_t reduce_total_length    = in.mDesc.GetElementSize() / invariant_total_length;
+
+    std::size_t num_thread = 1;
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            in.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
+                                            num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InOutDataTypeInDevice) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(InOutDataTypeInDevice) * out.mDesc.GetElementSpaceSize());
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    if(std::is_same<InOutDataType, int4_t>::value)
+    {
+        std::vector<InOutDataTypeInDevice> tmp_buf(in.mData.size());
+
+        std::copy_n(in.mData.data(), in.mData.size(), tmp_buf.data());
+        in_dev.ToDevice(tmp_buf.data());
+    }
+    else
+#endif
+        in_dev.ToDevice(in.mData.data());
+
+    if(beta != 0.0f)
+    {
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        if(std::is_same<InOutDataType, int4_t>::value)
+        {
+            std::vector<InOutDataTypeInDevice> tmp_buf(in.mData.size());
+
+            std::copy_n(out.mData.data(), out.mData.size(), tmp_buf.data());
+            out_dev.ToDevice(tmp_buf.data());
+        }
+        else
+#endif
+            out_dev.ToDevice(out.mData.data());
+    };
+
+    size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0;
+
+    DeviceMem out_index_dev(indicesSizeInBytes);
+
+    InElementwiseOperation in_elementwise_op;
+    AccElementwiseOperation acc_elementwise_op;
+
+    std::tie(in_elementwise_op, acc_elementwise_op) =
+        reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+            static_cast<int32_t>(reduce_total_length));
+
+    if(do_verification)
+    {
+        ReductionHost<InOutDataType,
+                      AccDataType,
+                      InOutDataType,
+                      ReduceOperation,
+                      InElementwiseOperation,
+                      AccElementwiseOperation,
+                      Rank,
+                      NumReduceDim,
+                      PropagateNan,
+                      OutputIndex>
+            hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
+
+        hostReduce.Run(alpha,
+                       in.mData.data(),
+                       beta,
+                       out_ref.mData.data(),
+                       out_indices_ref.mData.data(),
+                       in_elementwise_op,
+                       acc_elementwise_op);
+    };
+
+    std::array<index_t, Rank> arrInLengths;
+    std::array<index_t, Rank> arrInStrides;
+    std::array<index_t, NumOutDim> arrOutLengths;
+    std::array<index_t, NumOutDim> arrOutStrides;
+
+    ck::ranges::copy(inLengths, arrInLengths.begin());
+    ck::ranges::copy(inStrides, arrInStrides.begin());
+    ck::ranges::copy(outLengths, arrOutLengths.begin());
+    ck::ranges::copy(outStrides, arrOutStrides.begin());
+
+    auto reduce = DeviceReduceInstance{};
+
+    auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
+                                                   arrInStrides,
+                                                   arrOutLengths,
+                                                   arrOutStrides,
+                                                   reduceDims,
+                                                   alpha,
+                                                   beta,
+                                                   in_dev.GetDeviceBuffer(),
+                                                   nullptr,
+                                                   out_dev.GetDeviceBuffer(),
+                                                   out_index_dev.GetDeviceBuffer(),
+                                                   in_elementwise_op,
+                                                   acc_elementwise_op);
+
+    if(!reduce.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cerr
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+
+        return (-2);
+    };
+
+    std::string reduce_name = reduce.GetTypeString();
+
+    auto invoker_ptr = reduce.MakeInvokerPointer();
+
+    float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
+                            invariant_total_length * sizeof(InOutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        if(std::is_same<InOutDataType, int4_t>::value)
+        {
+            std::vector<InOutDataTypeInDevice> tmp_buf(out.mData.size());
+
+            out_dev.FromDevice(tmp_buf.data());
+
+            std::copy_n(tmp_buf.data(), out.mData.size(), out.mData.data());
+        }
+        else
+#endif
+            out_dev.FromDevice(out.mData.data());
+
+        pass = pass && ck::utils::check_err(out, out_ref);
+
+        if(OutputIndex)
+        {
+            out_index_dev.FromDevice(out_indices.mData.data());
+            pass = pass && ck::utils::check_err(out_indices, out_indices_ref);
+        };
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/12_reduce/reduce_blockwise_two_call.cpp b/3rdparty/composable_kernel/example/12_reduce/reduce_blockwise_two_call.cpp
new file mode 100644
index 0000000..39821f2
--- /dev/null
+++ b/3rdparty/composable_kernel/example/12_reduce/reduce_blockwise_two_call.cpp
@@ -0,0 +1,301 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <sstream>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_reduction.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+using InOutDataType = ck::half_t;
+using InOutDataType = ck::half_t;
+using AccDataType   = float;
+
+constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2;
+constexpr bool PropagateNan         = true;
+constexpr bool OutputIndex          = false;
+
+using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+using InElementwiseOperation =
+    typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+using AccElementwiseOperation =
+    typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+using DeviceReduceInstance_1 = DeviceReduceMultiBlock<InOutDataType,
+                                                      AccDataType,
+                                                      InOutDataType,
+                                                      5, // Rank
+                                                      1, // NumReduceDim
+                                                      ReduceOperation,
+                                                      InElementwiseOperation,
+                                                      PassThroughOp,
+                                                      InMemoryDataOperationEnum::Set,
+                                                      PropagateNan,
+                                                      OutputIndex,
+                                                      false, // HaveIndexInputIfOutputIndex
+                                                      256,
+                                                      32,
+                                                      8,
+                                                      1,
+                                                      1,
+                                                      1, // vector dim
+                                                      1,
+                                                      1>;
+
+using DeviceReduceInstance_2 = DeviceReduceMultiBlock<InOutDataType,
+                                                      AccDataType,
+                                                      InOutDataType,
+                                                      4, // Rank
+                                                      1, // NumReduceDim
+                                                      ReduceOperation,
+                                                      PassThroughOp,
+                                                      AccElementwiseOperation,
+                                                      InMemoryDataOperationEnum::Set,
+                                                      PropagateNan,
+                                                      OutputIndex,
+                                                      false, // HaveIndexInputIfOutputIndex
+                                                      256,
+                                                      128,
+                                                      2,
+                                                      1,
+                                                      1,
+                                                      1, // vector dim
+                                                      1,
+                                                      1>;
+
+static bool do_verify;
+static int init_method;
+static float alpha;
+static float beta;
+static bool time_kernel;
+
+int main(int argc, char* argv[])
+{
+    // used by the device reduction
+    const std::array<int, 1> reduceDims_1 = {4};
+    // const std::array<int, 4> invariantDims_1 = {0, 1, 2, 3};
+
+    const std::array<int, 1> reduceDims_2 = {3};
+    // const std::array<int, 3> invariantDims_2 = {0, 1, 2};
+
+    // used by the host reduction
+    const std::array<int, 2> reduceDims    = {3, 4};
+    const std::array<int, 3> invariantDims = {0, 1, 2};
+
+    const std::vector<size_t> inLengths_1 = {64, 320, 80, 4, 128};
+
+    // input lengths of the second reduction, which is also the output lengths of the first
+    // reduction
+    const std::vector<size_t> inLengths_2 = {64, 320, 80, 4};
+
+    const std::vector<size_t> outLengths = {64, 320, 80};
+
+    if(argc == 1)
+    {
+        do_verify   = true;
+        init_method = 2;
+        time_kernel = true;
+    }
+    else if(argc == 4)
+    {
+        do_verify   = static_cast<bool>(argv[1]);
+        init_method = atoi(argv[2]);
+        time_kernel = static_cast<bool>(atoi(argv[3]));
+    }
+    else
+    {
+        std::ostringstream ostr;
+
+        ostr << "Wrong parameter! " << std::endl
+             << "Usage: " << argv[0] << "[verify 0/1] init_method time_kernel" << std::endl;
+
+        throw std::runtime_error(ostr.str());
+    };
+
+    alpha = 1.0f;
+    beta  = 0.0f;
+
+    Tensor<InOutDataType> in_1(inLengths_1);
+
+    Tensor<InOutDataType> out_ref(outLengths);
+    Tensor<InOutDataType> in_2(inLengths_2); // also the output tensor of the first reduction
+    Tensor<InOutDataType> out(outLengths);
+
+    auto inStrides_1 = in_1.mDesc.GetStrides();
+    auto inStrides_2 = in_2.mDesc.GetStrides();
+    auto outStrides  = out.mDesc.GetStrides();
+
+    size_t invariant_total_length = out.mDesc.GetElementSize();
+    size_t reduce_total_length    = in_1.mDesc.GetElementSize() / invariant_total_length;
+
+    std::size_t num_thread = 1;
+
+    if(do_verify)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            in_1.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in_1.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in_1.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
+                                            num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+
+    DeviceMem in_1_dev(sizeof(InOutDataType) * in_1.mDesc.GetElementSpaceSize());
+    DeviceMem in_2_dev(sizeof(InOutDataType) * in_2.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(InOutDataType) * out.mDesc.GetElementSpaceSize());
+
+    in_1_dev.ToDevice(in_1.mData.data());
+
+    if(beta != 0.0f)
+        out_dev.ToDevice(out.mData.data());
+
+    InElementwiseOperation in_elementwise_op;
+    AccElementwiseOperation acc_elementwise_op;
+
+    std::tie(in_elementwise_op, acc_elementwise_op) =
+        reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+            static_cast<int32_t>(reduce_total_length));
+
+    if(do_verify)
+    {
+        ReductionHost<InOutDataType,
+                      AccDataType,
+                      InOutDataType,
+                      ReduceOperation,
+                      InElementwiseOperation,
+                      AccElementwiseOperation,
+                      5, // Rank
+                      2, // NumReduceDim
+                      PropagateNan,
+                      OutputIndex>
+            hostReduce(in_1.mDesc, out_ref.mDesc, invariantDims, reduceDims);
+
+        hostReduce.Run(alpha,
+                       in_1.mData.data(),
+                       beta,
+                       out_ref.mData.data(),
+                       nullptr,
+                       in_elementwise_op,
+                       acc_elementwise_op);
+    };
+
+    std::array<index_t, 5> arrInLengths_1;
+    std::array<index_t, 5> arrInStrides_1;
+    std::array<index_t, 4> arrInLengths_2;
+    std::array<index_t, 4> arrInStrides_2;
+    std::array<index_t, 3> arrOutLengths;
+    std::array<index_t, 3> arrOutStrides;
+
+    ck::ranges::copy(inLengths_1, arrInLengths_1.begin());
+    ck::ranges::copy(inStrides_1, arrInStrides_1.begin());
+    ck::ranges::copy(inLengths_2, arrInLengths_2.begin());
+    ck::ranges::copy(inStrides_2, arrInStrides_2.begin());
+    ck::ranges::copy(outLengths, arrOutLengths.begin());
+    ck::ranges::copy(outStrides, arrOutStrides.begin());
+
+    auto reduce_1 = DeviceReduceInstance_1{};
+
+    auto argument_ptr_1 = reduce_1.MakeArgumentPointer(arrInLengths_1,
+                                                       arrInStrides_1,
+                                                       arrInLengths_2,
+                                                       arrInStrides_2,
+                                                       reduceDims_1,
+                                                       1.0f,
+                                                       0.0f,
+                                                       in_1_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       in_2_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       in_elementwise_op,
+                                                       PassThroughOp{});
+
+    if(!reduce_1.IsSupportedArgument(argument_ptr_1.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+    };
+
+    auto invoker_ptr_1 = reduce_1.MakeInvokerPointer();
+
+    auto reduce_2 = DeviceReduceInstance_2{};
+
+    auto argument_ptr_2 = reduce_2.MakeArgumentPointer(arrInLengths_2,
+                                                       arrInStrides_2,
+                                                       arrOutLengths,
+                                                       arrOutStrides,
+                                                       reduceDims_2,
+                                                       alpha,
+                                                       beta,
+                                                       in_2_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       out_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       PassThroughOp{},
+                                                       acc_elementwise_op);
+
+    if(!reduce_2.IsSupportedArgument(argument_ptr_2.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+    };
+
+    auto invoker_ptr_2 = reduce_2.MakeInvokerPointer();
+
+    float avg_time_1 = invoker_ptr_1->Run(argument_ptr_1.get(), StreamConfig{nullptr, time_kernel});
+    float avg_time_2 = invoker_ptr_2->Run(argument_ptr_2.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
+                            invariant_total_length * sizeof(InOutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / (avg_time_1 + avg_time_2);
+
+    std::cout << "Perf: " << avg_time_1 + avg_time_2 << " ms, " << gb_per_sec << " GB/s, "
+              << reduce_1.GetTypeString() << " => " << reduce_2.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verify)
+    {
+        out_dev.FromDevice(out.mData.data());
+        pass = pass && ck::utils::check_err(out, out_ref);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/12_reduce/reduce_example_common.hpp b/3rdparty/composable_kernel/example/12_reduce/reduce_example_common.hpp
new file mode 100644
index 0000000..05f0a0e
--- /dev/null
+++ b/3rdparty/composable_kernel/example/12_reduce/reduce_example_common.hpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+
+template <int Rank, int NumReduceDim>
+static inline std::array<int, Rank - NumReduceDim>
+get_invariant_dims(const std::array<int, NumReduceDim>& reduceDims)
+{
+    int reduceFlag = 0;
+
+    // flag the bits for the reduceDims
+    for(int i = 0; i < NumReduceDim; i++)
+    {
+        reduceFlag |= 1 << reduceDims[i];
+    };
+
+    std::array<int, Rank - NumReduceDim> invariantDims;
+
+    // collect invariant dimensions
+    int dim = 0;
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) == 0)
+        {
+            invariantDims[dim] = i;
+            dim++;
+        };
+
+    return invariantDims;
+};
+
+template <ck::index_t Rank, ck::index_t NumReduceDim>
+struct ReduceShape
+{
+    static constexpr ck::index_t Rank_         = Rank;
+    static constexpr ck::index_t NumReduceDim_ = NumReduceDim;
+};
+
+using reduce_shape_instances = std::tuple<ReduceShape<3, 1>,
+                                          ReduceShape<3, 2>,
+                                          ReduceShape<4, 1>,
+                                          ReduceShape<4, 2>,
+                                          ReduceShape<4, 3>,
+                                          ReduceShape<5, 1>,
+                                          ReduceShape<5, 2>,
+                                          ReduceShape<5, 3>,
+                                          ReduceShape<5, 4>>;
diff --git a/3rdparty/composable_kernel/example/12_reduce/reduce_multiblock_atomic_add.cpp b/3rdparty/composable_kernel/example/12_reduce/reduce_multiblock_atomic_add.cpp
new file mode 100644
index 0000000..c4d63a3
--- /dev/null
+++ b/3rdparty/composable_kernel/example/12_reduce/reduce_multiblock_atomic_add.cpp
@@ -0,0 +1,216 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/utility/reduction_enums.hpp"
+#include "reduce_multiblock_atomic_add_impl.hpp"
+#include "reduce_example_common.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths = {16, 64, 32, 960};
+    std::vector<int> reduceDims   = {0, 1, 2};
+    std::vector<float> scales     = {1.0f, 0.0f};
+
+    bool do_verification = true;
+    int data_type        = 1;
+    int init_method      = 2;
+    bool time_kernel     = true;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of to-reduce dimensions"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp32, 1: fp64)" << std::endl;
+        std::cout << "Arg2 -- init method (0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg3 -- time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:l:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 3 > argc)
+        {
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+        };
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        return (0);
+    };
+};
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ReduceTensorOp ReduceOpId,
+          index_t PropagateNan>
+bool reduce_multiblock_atomic_add_test(bool do_verification,
+                                       int init_method,
+                                       bool time_kernel,
+                                       const std::vector<size_t>& inLengths,
+                                       const std::vector<int>& reduceDims,
+                                       float alpha,
+                                       float beta)
+{
+    bool matched = false;
+    int result   = 0;
+
+    const auto tuple_object = reduce_shape_instances{};
+
+    static_for<0, std::tuple_size<reduce_shape_instances>::value, 1>{}([&](auto i) {
+        if(matched)
+            return;
+
+        using ShapeType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
+
+        if(ShapeType::Rank_ != inLengths.size() || ShapeType::NumReduceDim_ != reduceDims.size())
+            return;
+
+        std::array<int, ShapeType::NumReduceDim_> a_reduceDims;
+
+        ck::ranges::copy(reduceDims, a_reduceDims.begin());
+
+        result = reduce_multiblock_atomic_add_impl<InOutDataType,
+                                                   AccDataType,
+                                                   ReduceOpId,
+                                                   ShapeType::Rank_,
+                                                   ShapeType::NumReduceDim_,
+                                                   PropagateNan>(
+            do_verification, init_method, time_kernel, inLengths, a_reduceDims, alpha, beta);
+
+        matched = true;
+    });
+
+    return (result == 0) ? true : false;
+};
+
+constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::AVG;
+constexpr bool PropagateNan         = true;
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        SimpleAppArgs arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            pass = reduce_multiblock_atomic_add_test<float, float, ReduceOpId, PropagateNan>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 1)
+        {
+            pass = reduce_multiblock_atomic_add_test<double, double, ReduceOpId, PropagateNan>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+    }
+    else
+    {
+        // for testing float
+        pass = pass && reduce_multiblock_atomic_add_test<float, float, ReduceOpId, PropagateNan>(
+                           true, 2, false, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing double
+        pass = pass && reduce_multiblock_atomic_add_test<double, double, ReduceOpId, PropagateNan>(
+                           true, 2, false, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing 3D input
+        pass = pass && reduce_multiblock_atomic_add_test<float, float, ReduceOpId, PropagateNan>(
+                           true, 2, false, {16, 64, 960}, {0, 1}, 1.0f, 0.0f);
+
+        // for testing 5D input
+        pass = pass && reduce_multiblock_atomic_add_test<float, float, ReduceOpId, PropagateNan>(
+                           true, 2, false, {16, 64, 32, 2, 960}, {0, 1, 2, 3}, 1.0f, 0.0f);
+    };
+
+    return (pass ? 0 : 1);
+};
diff --git a/3rdparty/composable_kernel/example/12_reduce/reduce_multiblock_atomic_add_impl.hpp b/3rdparty/composable_kernel/example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
new file mode 100644
index 0000000..94867ae
--- /dev/null
+++ b/3rdparty/composable_kernel/example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
@@ -0,0 +1,233 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_reduction.hpp"
+
+#include "reduce_example_common.hpp"
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          ck::index_t Rank,
+          ck::index_t NumReduceDim,
+          bool PropagateNan>
+int reduce_multiblock_atomic_add_impl(bool do_verification,
+                                      int init_method,
+                                      bool time_kernel,
+                                      const std::vector<size_t>& inLengths,
+                                      const std::array<int, NumReduceDim>& reduceDims,
+                                      float alpha,
+                                      float beta)
+
+{
+    using namespace ck;
+    using namespace ck::tensor_operation::device;
+
+    constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
+    constexpr bool op_support_atomic_add =
+        (ReduceOpId == ReduceTensorOp::ADD || ReduceOpId == ReduceTensorOp::AVG);
+
+    constexpr bool invalid_reduce_1 = !op_support_atomic_add;
+    constexpr bool invalid_reduce_2 =
+        !(std::is_same<InOutDataType, float>::value || std::is_same<InOutDataType, double>::value);
+
+    constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2);
+
+    if(invalid_reduce)
+    {
+        std::cerr << "The reduction setting is invalid, exiting!" << std::endl;
+        return (-1);
+    };
+
+    using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+    using InElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+    using AccElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+    using DeviceReduceInstance =
+        ck::tensor_operation::device::DeviceReduceMultiBlock<InOutDataType,
+                                                             AccDataType,
+                                                             InOutDataType,
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             ReduceOperation,
+                                                             InElementwiseOperation,
+                                                             AccElementwiseOperation,
+                                                             InMemoryDataOperationEnum::AtomicAdd,
+                                                             PropagateNan,
+                                                             false,
+                                                             false, // HaveIndexInputIfOutputIndex
+                                                             256,
+                                                             4,
+                                                             64,
+                                                             1,
+                                                             1,
+                                                             0,
+                                                             1,
+                                                             1>;
+
+    Tensor<InOutDataType> in(inLengths);
+
+    std::vector<size_t> outLengths;
+
+    auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
+
+    if(invariantDims.empty())
+        outLengths.push_back(1);
+    else
+        for(auto dim : invariantDims)
+            outLengths.push_back(inLengths[dim]);
+
+    Tensor<InOutDataType> out_ref(outLengths);
+    Tensor<InOutDataType> out(outLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = out.mDesc.GetStrides();
+
+    size_t invariant_total_length = out.mDesc.GetElementSize();
+    size_t reduce_total_length    = in.mDesc.GetElementSize() / invariant_total_length;
+
+    std::size_t num_thread = 1;
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            in.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
+                                            num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InOutDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(InOutDataType) * out.mDesc.GetElementSpaceSize());
+
+    in_dev.ToDevice(in.mData.data());
+
+    if(beta != 0.0f)
+        out_dev.ToDevice(out.mData.data());
+
+    InElementwiseOperation in_elementwise_op;
+    AccElementwiseOperation acc_elementwise_op;
+
+    std::tie(in_elementwise_op, acc_elementwise_op) =
+        reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+            static_cast<int32_t>(reduce_total_length));
+
+    if(do_verification)
+    {
+        ReductionHost<InOutDataType,
+                      AccDataType,
+                      InOutDataType,
+                      ReduceOperation,
+                      InElementwiseOperation,
+                      AccElementwiseOperation,
+                      Rank,
+                      NumReduceDim,
+                      PropagateNan,
+                      false>
+            hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
+
+        hostReduce.Run(alpha,
+                       in.mData.data(),
+                       beta,
+                       out_ref.mData.data(),
+                       nullptr,
+                       in_elementwise_op,
+                       acc_elementwise_op);
+    };
+
+    std::array<index_t, Rank> arrInLengths;
+    std::array<index_t, Rank> arrInStrides;
+    std::array<index_t, NumOutDim> arrOutLengths;
+    std::array<index_t, NumOutDim> arrOutStrides;
+
+    ck::ranges::copy(inLengths, arrInLengths.begin());
+    ck::ranges::copy(inStrides, arrInStrides.begin());
+    ck::ranges::copy(outLengths, arrOutLengths.begin());
+    ck::ranges::copy(outStrides, arrOutStrides.begin());
+
+    auto reduce = DeviceReduceInstance{};
+
+    auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
+                                                   arrInStrides,
+                                                   arrOutLengths,
+                                                   arrOutStrides,
+                                                   reduceDims,
+                                                   alpha,
+                                                   beta,
+                                                   in_dev.GetDeviceBuffer(),
+                                                   nullptr,
+                                                   out_dev.GetDeviceBuffer(),
+                                                   nullptr,
+                                                   in_elementwise_op,
+                                                   acc_elementwise_op);
+
+    if(!reduce.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cerr
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+
+        return (-2);
+    };
+
+    std::string reduce_name = reduce.GetTypeString();
+
+    auto invoker_ptr = reduce.MakeInvokerPointer();
+
+    float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
+                            invariant_total_length * sizeof(InOutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        out_dev.FromDevice(out.mData.data());
+        pass = pass && ck::utils::check_err(out, out_ref);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/13_pool2d_fwd/CMakeLists.txt b/3rdparty/composable_kernel/example/13_pool2d_fwd/CMakeLists.txt
new file mode 100644
index 0000000..db09c03
--- /dev/null
+++ b/3rdparty/composable_kernel/example/13_pool2d_fwd/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
+add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)
+
diff --git a/3rdparty/composable_kernel/example/13_pool2d_fwd/README.md b/3rdparty/composable_kernel/example/13_pool2d_fwd/README.md
new file mode 100644
index 0000000..9b01773
--- /dev/null
+++ b/3rdparty/composable_kernel/example/13_pool2d_fwd/README.md
@@ -0,0 +1,41 @@
+# Instructions for ```example_pool2d_fwd``` Examples
+
+## Run ```example_pool2d_fwd_fp16```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, RightPx
+./bin/example_pool2d_fwd_fp16 1 1 1
+```
+
+Result 
+```
+in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
+out_n_c_ho_wo: dim 4, lengths {128, 192, 36, 36}, strides {248832, 1, 6912, 192}
+launch_and_time_kernel: grid_dim {124416, 1, 1}, block_dim {64, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.397436 ms, 1.44252 TFlops, 783.713 GB/s
+```
+
+## Run ```example_pool2d_fwd_fp32```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, RightPx
+./bin/example_pool2d_fwd_fp32 1 1 1
+```
+
+
+Result 
+```
+./bin/example_pool2d_fwd_fp32 1 1 1
+in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
+out_n_c_ho_wo: dim 4, lengths {128, 192, 36, 36}, strides {248832, 1, 6912, 192}
+launch_and_time_kernel: grid_dim {124416, 1, 1}, block_dim {64, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 1.01823 ms, 0.563045 TFlops, 611.8 GB/s
+```
diff --git a/3rdparty/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_common.hpp b/3rdparty/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_common.hpp
new file mode 100644
index 0000000..b83cb6a
--- /dev/null
+++ b/3rdparty/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_common.hpp
@@ -0,0 +1,283 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool OutputIndex>
+static void pool_host_verify(const Tensor<InDataType>& in,
+                             Tensor<OutDataType>& out,
+                             Tensor<IndexDataType>& out_indices,
+                             const std::array<ck::index_t, 2>& window_spatial_lengths,
+                             const std::array<ck::index_t, 2>& window_strides,
+                             const std::array<ck::index_t, 2>& in_left_pads,
+                             const std::array<ck::index_t, 2>& /*in_right_pads*/)
+{
+    const int32_t reduceLength = window_spatial_lengths[0] * window_spatial_lengths[1];
+
+    using ReduceOperation = typename ck::reduce_binary_operator<ReduceOpId>::opType;
+
+    auto elementwise_ops =
+        ck::reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(reduceLength);
+
+    auto in_elementwise_op  = std::get<0>(elementwise_ops);
+    auto acc_elementwise_op = std::get<1>(elementwise_ops);
+
+    if constexpr(!OutputIndex)
+    {
+        using Accumulation =
+            ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+        auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
+            auto accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+            for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
+            {
+                ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
+                for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
+                {
+                    ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
+                    if(hi >= 0 && hi < static_cast<ck::index_t>(in.mDesc.GetLengths()[2]) &&
+                       wi >= 0 && wi < static_cast<ck::index_t>(in.mDesc.GetLengths()[3]))
+                    {
+                        AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
+
+                        in_elementwise_op(currVal, currVal);
+
+                        Accumulation::Calculate(accuVal, currVal);
+                    }
+                }
+            }
+
+            acc_elementwise_op(accuVal, accuVal);
+
+            out(n, c, ho, wo) = accuVal;
+        };
+
+        make_ParallelTensorFunctor(f_nchw,
+                                   out.mDesc.GetLengths()[0],
+                                   out.mDesc.GetLengths()[1],
+                                   out.mDesc.GetLengths()[2],
+                                   out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
+    }
+    else
+    {
+        using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
+                                                                        ReduceOperation,
+                                                                        AccDataType,
+                                                                        IndexDataType>;
+        auto f_nchw        = [&](auto n, auto c, auto ho, auto wo) {
+            auto accuVal            = ReduceOperation::template GetIdentityValue<AccDataType>();
+            IndexDataType accuIndex = 0;
+
+            for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
+            {
+                ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
+                for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
+                {
+                    ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
+                    if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
+                       wi < in.mDesc.GetLengths()[3])
+                    {
+                        AccDataType currVal     = static_cast<AccDataType>(in(n, c, hi, wi));
+                        IndexDataType currIndex = y * window_spatial_lengths[1] + x;
+
+                        in_elementwise_op(currVal, currVal);
+
+                        Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+                    }
+                }
+            }
+
+            acc_elementwise_op(accuVal, accuVal);
+
+            out(n, c, ho, wo)         = accuVal;
+            out_indices(n, c, ho, wo) = accuIndex;
+        };
+
+        make_ParallelTensorFunctor(f_nchw,
+                                   out.mDesc.GetLengths()[0],
+                                   out.mDesc.GetLengths()[1],
+                                   out.mDesc.GetLengths()[2],
+                                   out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
+    };
+}
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          typename InLayout,
+          typename OutLayout,
+          ck::ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool OutputIndex>
+bool pool_test(bool do_verification,
+               int init_method,
+               bool time_kernel,
+               ck::index_t N,
+               ck::index_t C,
+               ck::index_t Y,
+               ck::index_t X,
+               ck::index_t Hi,
+               ck::index_t Wi,
+               ck::index_t window_stride_h,
+               ck::index_t window_stride_w,
+               ck::index_t in_left_pad_h,
+               ck::index_t in_left_pad_w,
+               ck::index_t in_right_pad_h,
+               ck::index_t in_right_pad_w)
+{
+    using DevicePoolFwdInstance =
+        ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<
+            InDataType,  // InDataType
+            OutDataType, // OutDataType
+            AccDataType, // AccDataType
+            ReduceOpId,
+            OutputIndex,
+            64, // BlockSize
+            64, // ReduceMThreadClusterSize
+            1,  // ReduceKThreadClusterSize
+            4,  // ReduceMThreadSliceSize
+            1,  // ReduceKThreadSliceSize
+            4>; // InSrcOutDstVectorSize
+
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
+
+    const std::array<ck::index_t, 2> window_spatial_lengths{{Y, X}};
+    const std::array<ck::index_t, 2> window_strides{{window_stride_h, window_stride_w}};
+    const std::array<ck::index_t, 2> input_left_pads{{in_left_pad_h, in_left_pad_w}};
+    const std::array<ck::index_t, 2> input_right_pads{{in_right_pad_h, in_right_pad_w}};
+
+    // tensor layout
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
+            using namespace ck::literals;
+
+            if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, H * W, W, 1_uz});
+            }
+            else if constexpr(ck::is_same<decltype(layout),
+                                          ck::tensor_layout::convolution::NHWC>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
+            }
+        };
+
+    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
+    Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+    Tensor<IndexDataType> out_indices_n_c_ho_wo_host(
+        f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+    Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+    Tensor<IndexDataType> out_indices_n_c_ho_wo_device(
+        f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "out_n_c_ho_wo: " << out_n_c_ho_wo_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}); break;
+    case 2: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
+    default: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
+    DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
+                                     out_indices_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+
+    auto pool         = DevicePoolFwdInstance{};
+    auto invoker_ptr  = pool.MakeInvokerPointer();
+    auto argument_ptr = pool.MakeArgumentPointer(
+        static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+        static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+        N,
+        C,
+        std::array<ck::index_t, 2>{{Hi, Wi}},
+        std::array<ck::index_t, 2>{{Y, X}},
+        std::array<ck::index_t, 2>{{Ho, Wo}},
+        window_strides,
+        input_left_pads,
+        input_right_pads);
+
+    if(!pool.IsSupportedArgument(argument_ptr.get()))
+    {
+        throw std::runtime_error("wrong! device_op with the specified compilation parameters does "
+                                 "not support this problem");
+    }
+
+    float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * N * C * Ho * Wo * Y * X;
+
+    std::size_t num_btype =
+        sizeof(InDataType) * (N * C * Hi * Wi) + sizeof(OutDataType) * (N * C * Ho * Wo);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        pool_host_verify<InDataType,
+                         OutDataType,
+                         AccDataType,
+                         IndexDataType,
+                         ReduceOpId,
+                         PropagateNan,
+                         OutputIndex>(in_n_c_hi_wi,
+                                      out_n_c_ho_wo_host,
+                                      out_indices_n_c_ho_wo_host,
+                                      window_spatial_lengths,
+                                      window_strides,
+                                      input_left_pads,
+                                      input_right_pads);
+
+        out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
+
+        pass = pass && ck::utils::check_err(out_n_c_ho_wo_device, out_n_c_ho_wo_host);
+
+        if constexpr(OutputIndex)
+        {
+            out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
+
+            pass = pass &&
+                   ck::utils::check_err(out_indices_n_c_ho_wo_device, out_indices_n_c_ho_wo_host);
+        };
+    }
+
+    return (pass);
+};
diff --git a/3rdparty/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_fp16.cpp b/3rdparty/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
new file mode 100644
index 0000000..659f325
--- /dev/null
+++ b/3rdparty/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
@@ -0,0 +1,117 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "pool2d_fwd_common.hpp"
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+using IndexDataType = int32_t;
+
+using InLayout  = ck::tensor_layout::convolution::NHWC;
+using OutLayout = ck::tensor_layout::convolution::NHWC;
+
+#if 1
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+#else
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+#endif
+
+static constexpr bool OutputIndex  = false;
+static constexpr bool PropagateNan = false;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification;
+    int init_method;
+    bool time_kernel;
+
+    // Pool shape
+    ck::index_t N               = 128;
+    ck::index_t C               = 192;
+    ck::index_t Y               = 3;
+    ck::index_t X               = 3;
+    ck::index_t Hi              = 71;
+    ck::index_t Wi              = 71;
+    ck::index_t window_stride_h = 2;
+    ck::index_t window_stride_w = 2;
+    ck::index_t in_left_pad_h   = 1;
+    ck::index_t in_left_pad_w   = 1;
+    ck::index_t in_right_pad_h  = 1;
+    ck::index_t in_right_pad_w  = 1;
+
+    if(argc == 1)
+    {
+        do_verification = true;
+        init_method     = 1;
+        time_kernel     = true;
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+    }
+    else if(argc == 16)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+
+        N               = std::stoi(argv[4]);
+        C               = std::stoi(argv[5]);
+        Y               = std::stoi(argv[6]);
+        X               = std::stoi(argv[7]);
+        Hi              = std::stoi(argv[8]);
+        Wi              = std::stoi(argv[9]);
+        window_stride_h = std::stoi(argv[10]);
+        window_stride_w = std::stoi(argv[11]);
+        in_left_pad_h   = std::stoi(argv[12]);
+        in_left_pad_w   = std::stoi(argv[13]);
+        in_right_pad_h  = std::stoi(argv[14]);
+        in_right_pad_w  = std::stoi(argv[15]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(0);
+    }
+
+    bool pass = pool_test<InDataType,
+                          OutDataType,
+                          AccDataType,
+                          IndexDataType,
+                          InLayout,
+                          OutLayout,
+                          ReduceOpId,
+                          PropagateNan,
+                          OutputIndex>(do_verification,
+                                       init_method,
+                                       time_kernel,
+                                       N,
+                                       C,
+                                       Y,
+                                       X,
+                                       Hi,
+                                       Wi,
+                                       window_stride_h,
+                                       window_stride_w,
+                                       in_left_pad_h,
+                                       in_left_pad_w,
+                                       in_right_pad_h,
+                                       in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_fp32.cpp b/3rdparty/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
new file mode 100644
index 0000000..f47c7ff
--- /dev/null
+++ b/3rdparty/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
@@ -0,0 +1,117 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "pool2d_fwd_common.hpp"
+
+using InDataType  = float;
+using OutDataType = float;
+using AccDataType = float;
+
+using IndexDataType = int32_t;
+
+using InLayout  = ck::tensor_layout::convolution::NHWC;
+using OutLayout = ck::tensor_layout::convolution::NHWC;
+
+#if 1
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+#else
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+#endif
+
+static constexpr bool OutputIndex  = false;
+static constexpr bool PropagateNan = false;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification;
+    int init_method;
+    bool time_kernel;
+
+    // Pool shape
+    ck::index_t N               = 128;
+    ck::index_t C               = 192;
+    ck::index_t Y               = 3;
+    ck::index_t X               = 3;
+    ck::index_t Hi              = 71;
+    ck::index_t Wi              = 71;
+    ck::index_t window_stride_h = 2;
+    ck::index_t window_stride_w = 2;
+    ck::index_t in_left_pad_h   = 1;
+    ck::index_t in_left_pad_w   = 1;
+    ck::index_t in_right_pad_h  = 1;
+    ck::index_t in_right_pad_w  = 1;
+
+    if(argc == 1)
+    {
+        do_verification = true;
+        init_method     = 1;
+        time_kernel     = true;
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+    }
+    else if(argc == 16)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+
+        N               = std::stoi(argv[4]);
+        C               = std::stoi(argv[5]);
+        Y               = std::stoi(argv[6]);
+        X               = std::stoi(argv[7]);
+        Hi              = std::stoi(argv[8]);
+        Wi              = std::stoi(argv[9]);
+        window_stride_h = std::stoi(argv[10]);
+        window_stride_w = std::stoi(argv[11]);
+        in_left_pad_h   = std::stoi(argv[12]);
+        in_left_pad_w   = std::stoi(argv[13]);
+        in_right_pad_h  = std::stoi(argv[14]);
+        in_right_pad_w  = std::stoi(argv[15]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(0);
+    }
+
+    bool pass = pool_test<InDataType,
+                          OutDataType,
+                          AccDataType,
+                          IndexDataType,
+                          InLayout,
+                          OutLayout,
+                          ReduceOpId,
+                          PropagateNan,
+                          OutputIndex>(do_verification,
+                                       init_method,
+                                       time_kernel,
+                                       N,
+                                       C,
+                                       Y,
+                                       X,
+                                       Hi,
+                                       Wi,
+                                       window_stride_h,
+                                       window_stride_w,
+                                       in_left_pad_h,
+                                       in_left_pad_w,
+                                       in_right_pad_h,
+                                       in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/14_gemm_quantization/CMakeLists.txt b/3rdparty/composable_kernel/example/14_gemm_quantization/CMakeLists.txt
new file mode 100644
index 0000000..ca09c48
--- /dev/null
+++ b/3rdparty/composable_kernel/example/14_gemm_quantization/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp)
+add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp)
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp b/3rdparty/composable_kernel/example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp
new file mode 100644
index 0000000..d5f4e6f
--- /dev/null
+++ b/3rdparty/composable_kernel/example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp
@@ -0,0 +1,235 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using I8  = int8_t;
+using I32 = int32_t;
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using CDEElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
+
+using ADataType        = I8;
+using BDataType        = I8;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using BiasDataType     = I32;
+using DsDataType       = ck::Tuple<BiasDataType>;
+using EDataType        = I8;
+
+using ALayout    = Row;
+using BLayout    = Col;
+using BiasLayout = Row;
+using DsLayout   = ck::Tuple<BiasLayout>;
+using ELayout    = Row;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<
+     ALayout,
+     BLayout,
+     DsLayout,
+     ELayout,
+     ADataType,
+     BDataType,
+     AccDataType,
+     CShuffleDataType,
+     DsDataType,
+     EDataType,
+     PassThrough,                // AElementwiseOperation,
+     PassThrough,                // BElementwiseOperation,
+     CDEElementOp,               // CDEElementwiseOperation,
+     GemmDefault,                // GemmSpecialization GemmSpec,
+     1,                          // NumGemmKPrefetchStage,
+     256,                        // BlockSize,
+     256,                        // MPerBlock,
+     128,                        // NPerBlock,
+     64,                         // KPerBlock,
+     16,                         // AK1,
+     16,                         // BK1,
+     32,                         // MPerXDL,
+     32,                         // NPerXDL,
+     4,                          // MXdlPerWave,
+     2,                          // NXdlPerWave,
+     S<4, 64, 1>,                // ABlockTransferThreadClusterLengths_AK0_M_AK1,
+     S<1, 0, 2>,                 // ABlockTransferThreadClusterArrangeOrder,
+     S<1, 0, 2>,                 // ABlockTransferSrcAccessOrder,
+     2,                          // index_t ABlockTransferSrcVectorDim,
+     16,                         // index_t ABlockTransferSrcScalarPerVector,
+     16,                         // index_t ABlockTransferDstScalarPerVector_AK1,
+     1,                          // bool ABlockLdsExtraM,
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder,
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder,
+     2,                          // index_t BBlockTransferSrcVectorDim,
+     8,                          // index_t BBlockTransferSrcScalarPerVector,
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1,
+     1,                          // bool BBlockLdsExtraN,
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle,
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle,
+     S<1, 64, 1, 4>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+     8>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        PassThrough>;
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA    = 1024;
+    ck::index_t StrideB    = 1024;
+    ck::index_t StrideBias = 0;
+    ck::index_t StrideE    = 1024;
+
+    float requant_scale = 0.03;
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1_uz}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1_uz, stride}));
+            }
+        };
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                    std::vector<std::size_t>({stride}));
+    };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<BiasDataType> bias_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "bias_n: " << bias_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-128, 127});
+    b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-128, 127});
+    bias_n.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-128, 127});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    bias_device_buf.ToDevice(bias_n.mData.data());
+
+    auto a_element_op   = PassThrough{};
+    auto b_element_op   = PassThrough{};
+    auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {bias_device_buf.GetDeviceBuffer()},
+                                      e_device_buf.GetDeviceBuffer(),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      {StrideBias},
+                                      StrideE,
+                                      a_element_op,
+                                      b_element_op,
+                                      cde_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), bias_n(n));
+            }
+        }
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp b/3rdparty/composable_kernel/example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
new file mode 100644
index 0000000..2371737
--- /dev/null
+++ b/3rdparty/composable_kernel/example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
@@ -0,0 +1,207 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using I8  = int8_t;
+using I32 = int32_t;
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+
+using ADataType        = I8;
+using BDataType        = I8;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = I8;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<
+     ALayout,
+     BLayout,
+     DsLayout,
+     ELayout,
+     ADataType,
+     BDataType,
+     AccDataType,
+     CShuffleDataType,
+     DsDataType,
+     EDataType,
+     PassThrough,                // AElementwiseOperation,
+     PassThrough,                // BElementwiseOperation,
+     CDEElementOp,               // CDEElementwiseOperation,
+     GemmDefault,                // GemmSpecialization GemmSpec,
+     1,                          // NumGemmKPrefetchStage,
+     256,                        // BlockSize,
+     256,                        // MPerBlock,
+     128,                        // NPerBlock,
+     64,                         // KPerBlock,
+     16,                         // AK1,
+     16,                         // BK1,
+     32,                         // MPerXDL,
+     32,                         // NPerXDL,
+     4,                          // MXdlPerWave,
+     2,                          // NXdlPerWave,
+     S<4, 64, 1>,                // ABlockTransferThreadClusterLengths_AK0_M_AK1,
+     S<1, 0, 2>,                 // ABlockTransferThreadClusterArrangeOrder,
+     S<1, 0, 2>,                 // ABlockTransferSrcAccessOrder,
+     2,                          // index_t ABlockTransferSrcVectorDim,
+     16,                         // index_t ABlockTransferSrcScalarPerVector,
+     16,                         // index_t ABlockTransferDstScalarPerVector_AK1,
+     1,                          // bool ABlockLdsExtraM,
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder,
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder,
+     2,                          // index_t BBlockTransferSrcVectorDim,
+     8,                          // index_t BBlockTransferSrcScalarPerVector,
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1,
+     1,                          // bool BBlockLdsExtraN,
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle,
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle,
+     S<1, 64, 1, 4>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+     16>;                        // index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, EDataType, float, PassThrough, PassThrough, CDEElementOp>;
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    float requant_scale = 0.03;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1_uz}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1_uz, stride}));
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-128, 127});
+    b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-128, 127});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op   = PassThrough{};
+    auto b_element_op   = PassThrough{};
+    auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {},
+                                      e_device_buf.GetDeviceBuffer(),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      {},
+                                      StrideE,
+                                      a_element_op,
+                                      b_element_op,
+                                      cde_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, e_m_n_host_result, a_element_op, b_element_op, cde_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/15_grouped_gemm/CMakeLists.txt b/3rdparty/composable_kernel/example/15_grouped_gemm/CMakeLists.txt
new file mode 100644
index 0000000..67f6160
--- /dev/null
+++ b/3rdparty/composable_kernel/example/15_grouped_gemm/CMakeLists.txt
@@ -0,0 +1,17 @@
+add_custom_target(example_grouped_gemm_xdl)
+
+add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
+add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
+add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
+add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
+
+add_dependencies(example_grouped_gemm_xdl
+                 example_grouped_gemm_xdl_fp32
+                 example_grouped_gemm_xdl_fp16
+                 example_grouped_gemm_xdl_bfp16
+                 example_grouped_gemm_xdl_int8)
+
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
+  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
+endif()
diff --git a/3rdparty/composable_kernel/example/15_grouped_gemm/README.md b/3rdparty/composable_kernel/example/15_grouped_gemm/README.md
new file mode 100644
index 0000000..c83b23e
--- /dev/null
+++ b/3rdparty/composable_kernel/example/15_grouped_gemm/README.md
@@ -0,0 +1,25 @@
+# Instructions for ```example_grouped_gemm_xdl```
+
+## Run ```example_grouped_gemm_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+./bin/example_grouped_gemm_xdl_fp16 0 1 5
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+gemm[0] a_m_k: dim 2, lengths {256, 64}, strides {64, 1} b_k_n: dim 2, lengths {64, 128}, strides {1, 64} c_m_n: dim 2, lengths {256, 128}, strides {128, 1}
+gemm[1] a_m_k: dim 2, lengths {512, 128}, strides {128, 1} b_k_n: dim 2, lengths {128, 256}, strides {1, 128} c_m_n: dim 2, lengths {512, 256}, strides {256, 1}
+gemm[2] a_m_k: dim 2, lengths {768, 192}, strides {192, 1} b_k_n: dim 2, lengths {192, 384}, strides {1, 192} c_m_n: dim 2, lengths {768, 384}, strides {384, 1}
+gemm[3] a_m_k: dim 2, lengths {1024, 256}, strides {256, 1} b_k_n: dim 2, lengths {256, 512}, strides {1, 256} c_m_n: dim 2, lengths {1024, 512}, strides {512, 1}
+group: 0 arg.a_grid_desc_k0_m_k1_{8, 256, 8}, arg.b_grid_desc_k0_n_k1_{8, 128, 8}, arg.c_grid_desc_m_n_{ 256, 128}
+group: 1 arg.a_grid_desc_k0_m_k1_{16, 512, 8}, arg.b_grid_desc_k0_n_k1_{16, 256, 8}, arg.c_grid_desc_m_n_{ 512, 256}
+group: 2 arg.a_grid_desc_k0_m_k1_{24, 768, 8}, arg.b_grid_desc_k0_n_k1_{24, 384, 8}, arg.c_grid_desc_m_n_{ 768, 384}
+group: 3 arg.a_grid_desc_k0_m_k1_{32, 1024, 8}, arg.b_grid_desc_k0_n_k1_{32, 512, 8}, arg.c_grid_desc_m_n_{ 1024, 512}
+launch_and_time_kernel: grid_dim {30, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 5 times...
+Perf: 0.037887 ms, 11.0706 TFlops, 90.8132 GB/s, DeviceGroupedGemmXdl<256, 256, 128, 4, 8, 32, 32, 4, 2>
+```
diff --git a/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_bfp16.cpp b/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_bfp16.cpp
new file mode 100644
index 0000000..05d572a
--- /dev/null
+++ b/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_bfp16.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = BF16;
+using BDataType        = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp b/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
new file mode 100644
index 0000000..3f78daf
--- /dev/null
+++ b/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp b/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp
new file mode 100644
index 0000000..fd93bb5
--- /dev/null
+++ b/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 32, 1, 8>,               4>;
+// clang-format on
+
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp b/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp
new file mode 100644
index 0000000..faf41bb
--- /dev/null
+++ b/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = ck::int4_t;
+using BDataType        = ck::int4_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = ck::int4_t;
+
+using KernelADataType = int8_t;
+using KernelBDataType = int8_t;
+using KernelEDataType = int8_t;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+        < ALayout,              //ALayout
+          BLayout,              //BLayout
+          DsLayout,             //DsLayout
+          ELayout,              //ELayout
+          KernelADataType,      //ADataType    
+          KernelBDataType,      //BDataType   
+          AccDataType,          //AccDataType
+          CShuffleDataType,     //CShuffleDataType
+          DsDataType,           //DsDataType
+          KernelEDataType,      //EDataType
+          AElementOp,           //AElementwiseOperation
+          BElementOp,           //BElementwiseOperation
+          CDEElementOp,         //CDEElementwiseOperation
+          GemmDefault,          //GEMMSpecialization
+          1,                    // NumGemmKPrefetchStage
+          256,                  // BlockSize
+          256,                  // MPerBlock
+          128,                  // NPerBlock
+          64,                   // KPerBlock
+          16,                   // AK1
+          16,                   // BK1
+          32,                   // MPerXdl
+          32,                   // NPerXdl
+          4,                    // MXdlPerWave
+          2,                    // NXdlPerWave
+          S<4, 64, 1>,          // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+          S<1, 0, 2>,           // ABlockTransfer ThreadCluster ArrangeOrder
+          S<1, 0, 2>,           // ABlockTransfer SrcAccessOrder
+          2,                    // ABlockTransfer SrcVectorDim
+          16,                   // ABlockTransfer SrcScalarPerVector
+          16,                   // ABlockTransfer DstScalarPerVector_K1
+          1,                    // ABlockLdsExtraM
+          S<4, 64, 1>,          // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+          S<1, 0, 2>,           // BBlockTransfer ThreadCluster ArrangeOrder
+          S<1, 0, 2>,           // BBlockTransfer SrcAccessOrder
+          2,                    // BBlockTransfer SrcVectorDim
+          16,                   // BBlockTransfer SrcScalarPerVector
+          16,                   // BBlockTransfer DstScalarPerVector_K1
+          1,                    // BBlockLdsExtraN
+          1,                    // CShuffleMXdlPerWavePerShuffle
+          1,                    // CShuffleNXdlPerWavePerShuffle
+          S<1, 64, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MWaveMPerXdl_NBlock_NWaveNPerXdl
+          16>;                  // CBlockTransferScalarPerVector_NWaveNPerXdl
+// clang-format on
+
+#define BUILD_INT4_EXAMPLE
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp b/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp
new file mode 100644
index 0000000..7cb0977
--- /dev/null
+++ b/3rdparty/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = int8_t;
+using BDataType        = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int8_t;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = int8_t;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>;
+// clang-format on
+
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/15_grouped_gemm/run_grouped_gemm_example.inc b/3rdparty/composable_kernel/example/15_grouped_gemm/run_grouped_gemm_example.inc
new file mode 100644
index 0000000..324e177
--- /dev/null
+++ b/3rdparty/composable_kernel/example/15_grouped_gemm/run_grouped_gemm_example.inc
@@ -0,0 +1,265 @@
+#pragma once
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+    static_assert(sizeof(ADataType) == sizeof(KernelADataType));
+    static_assert(sizeof(BDataType) == sizeof(KernelBDataType));
+    static_assert(sizeof(EDataType) == sizeof(KernelEDataType));
+#endif
+    int group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_c;
+
+    gemm_descs.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        int M = problem_size.Ms[i];
+        int N = problem_size.Ns[i];
+        int K = problem_size.Ks[i];
+
+        int stride_A = problem_size.stride_As[i];
+        int stride_B = problem_size.stride_Bs[i];
+        int stride_C = problem_size.stride_Cs[i];
+
+        gemm_descs.push_back({M, N, K, stride_A, stride_B, stride_C, {}});
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<BDataType>> b_tensors;
+    std::vector<Tensor<EDataType>> c_host_tensors;
+#ifdef BUILD_INT4_EXAMPLE
+    std::vector<Tensor<KernelEDataType>> c_device_tensors;
+#else
+    std::vector<Tensor<EDataType>> c_device_tensors;
+#endif
+
+    a_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    c_host_tensors.reserve(group_count);
+    c_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
+
+    a_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+    {
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].K_, gemm_descs[i].stride_A_, ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].K_, gemm_descs[i].N_, gemm_descs[i].stride_B_, BLayout{})));
+        c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+#ifdef BUILD_INT4_EXAMPLE
+        c_device_tensors.push_back(Tensor<KernelEDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+#else
+        c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+#endif
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
+                  << std::endl;
+
+        flop += std::size_t(2) * gemm_descs[i].M_ * gemm_descs[i].K_ * gemm_descs[i].N_;
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
+
+        switch(config.init_method)
+        {
+        case 0: break;
+        case 1:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            break;
+        case 2:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+    }
+
+    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+    {
+        a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpaceSize()));
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpaceSize()));
+        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSpaceSize()));
+
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<KernelADataType> a_converted(a_tensors[i]);
+        const Tensor<KernelBDataType> b_converted(b_tensors[i]);
+
+        a_tensors_device[i]->ToDevice(a_converted.mData.data());
+        b_tensors_device[i]->ToDevice(b_converted.mData.data());
+#else
+        a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
+#endif
+
+        p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
+        p_b.push_back(b_tensors_device[i]->GetDeviceBuffer());
+        p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
+    }
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<std::array<const void*, 0>> p_Ds = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_a, p_b, p_Ds, p_c, gemm_descs, a_element_op, b_element_op, c_element_op);
+
+    DeviceMem gemm_desc_workspace(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, gemm_desc_workspace.GetDeviceBuffer());
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                EDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CDEElementOp>;
+
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      c_host_tensors[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+
+#ifdef BUILD_INT4_EXAMPLE
+            const Tensor<EDataType> c_device_result_converted(c_device_tensors[i]);
+            pass &= ck::utils::check_err(c_device_result_converted, c_host_tensors[i]);
+
+#else
+            pass &= ck::utils::check_err(c_device_tensors[i], c_host_tensors[i]);
+#endif
+        }
+    }
+
+    if(config.time_kernel)
+    {
+        float ave_time   = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+        float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    }
+
+    return pass;
+}
+
+bool run_grouped_gemm_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ms.push_back(256 + 256 * i);
+        problem_size.Ns.push_back(128 + 128 * i);
+        problem_size.Ks.push_back(128 + 64 * i);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ks[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+    }
+
+    if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    return run_grouped_gemm(problem_size, config);
+}
diff --git a/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
new file mode 100644
index 0000000..226656a
--- /dev/null
+++ b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
@@ -0,0 +1,40 @@
+add_custom_target(example_gemm_reduce_xdl)
+add_custom_target(example_gemm_reduce_xdl_max)
+add_custom_target(example_gemm_reduce_xdl_mean_meansquare)
+add_custom_target(example_gemm_add_add_mean_meansquare_xdl)
+
+add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
+add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp)
+add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp)
+add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp)
+
+add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
+
+add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
+add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
+add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
+add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
+
+add_dependencies(example_gemm_reduce_xdl_max
+                 example_gemm_max_xdl_bf16
+                 example_gemm_max_xdl_fp16
+                 example_gemm_max_xdl_fp32
+                 example_gemm_max_xdl_int8)
+
+add_dependencies(example_gemm_reduce_xdl_mean_meansquare
+                 example_gemm_mean_meansquare_xdl_fp16
+                 example_gemm_mean_meansquare_xdl_fp32
+                 example_gemm_mean_meansquare_xdl_bf16
+                 example_gemm_add_addsquare_xdl_int8)
+
+add_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
+
+add_dependencies(example_gemm_reduce_xdl
+                 example_gemm_reduce_xdl_mean_meansquare
+                 example_gemm_reduce_xdl_max
+                 example_gemm_add_add_mean_meansquare_xdl)
+
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp)
+  add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
+endif()
diff --git a/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp
new file mode 100644
index 0000000..eb3832a
--- /dev/null
+++ b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp
@@ -0,0 +1,276 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+// DataType
+using ADataType         = F16;
+using BDataType         = F16;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using D0DataType        = F16;
+using D1DataType        = F16;
+using DsDataType        = ck::Tuple<D0DataType, D1DataType>;
+using EDataType         = F16;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using R1DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType, R1DataType>;
+
+// Layout
+using ALayout  = Row;
+using BLayout  = Col;
+using D1Layout = Row;
+using ELayout  = D1Layout;
+
+// Elementwise op
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using AddAdd       = ck::tensor_operation::element_wise::AddAdd;
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using Div          = ck::tensor_operation::element_wise::UnaryDivide;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAdd;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<Div, Div>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+//######| ALayout| BLayout| ELayout|     AData|     BData|     GemmAccData|         CShuffle|     DsData|     EData|     ReduceAccData|     RsData|           A|           B|          CDE|          Qs|          Rs|           Thread|           Global|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|    CDRThreadTransfer|                  CDE|    RThreadTransfer|
+//######|        |        |        |      Type|      Type|            Type|         DataType|       Type|      Type|              Type|       Type| Elementwise| Elementwise|  Elementwise| Elementwise| Elementwise|           Reduce|           Reduce| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|       ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
+//######|        |        |        |          |          |                |                 |           |          |                  |           |   Operation|   Operation|    Operation|   Operation|   Operation|        Operation|        Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _MPerBlock_NPerBlock|      ScalarPerVector|         _MPerBlock|
+//######|        |        |        |          |          |                |                 |           |          |                  |           |            |            |             |            |            |                 |                 |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |           _NPerBlock|                   |
+        < ALayout, BLayout, ELayout, ADataType, BDataType, GemmAccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType,  AElementOp,  BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<64, 4>,                    4,                  1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        EDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+template <typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType,
+          typename R0DataType,
+          typename R1DataType>
+void DumpPerf(float ave_time, int M, int N, int K)
+{
+    std::size_t flop          = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(D0DataType) * M * N + sizeof(D1DataType) * M * N +
+                                sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
+                                sizeof(R1DataType) * M;
+
+    float tflops          = static_cast<float>(flop) / 1.E9 / ave_time;
+    float gemm_gb_per_sec = gemm_num_byte / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gemm_gb_per_sec
+              << " GB/s, " << std::endl;
+}
+
+auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+    return HostTensorDescriptor({len}, {stride});
+};
+
+auto f_host_tensor_descriptor2d =
+    [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({row, col}, {stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({row, col}, {1_uz, stride});
+        }
+    };
+
+int main()
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = 1024;
+    ck::index_t StrideB  = 1024;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 1024;
+    ck::index_t StrideE  = 1024;
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor2d(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> r0_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> r1_m(f_host_tensor_descriptor1d(M, 1));
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
+    b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
+    d0_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-1, 1});
+    d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-1, 1});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem r0_device_buf(sizeof(R0DataType) * r0_m.mDesc.GetElementSpaceSize());
+    DeviceMem r1_device_buf(sizeof(R1DataType) * r1_m.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_device_buf.ToDevice(d0_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto qs_element_op  = QsElementOp{};
+    auto rs_element_op  = RsElementOp{N, N};
+
+    // Prepare GEMM, mean, mean_square
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               {r0_device_buf.GetDeviceBuffer(), r1_device_buf.GetDeviceBuffer()},
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               {StrideD0, StrideD1},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op,
+                               qs_element_op,
+                               rs_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    // init reducetion buffer to 0
+    r0_device_buf.SetZero();
+    r1_device_buf.SetZero();
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    bool do_verification = true;
+    bool pass            = true;
+
+    if(do_verification)
+    {
+        auto I0 = ck::Number<0>{};
+        auto I1 = ck::Number<1>{};
+
+        Tensor<EDataType> e_m_n_host(e_m_n.mDesc);
+        Tensor<R0DataType> r0_m_host(r0_m.mDesc);
+        Tensor<R1DataType> r1_m_host(r1_m.mDesc);
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, e_m_n_host, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        auto reduce0_op = R0ThreadReduceOp{};
+        auto reduce1_op = R1ThreadReduceOp{};
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+            auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType square_e_val;
+
+                auto e_val  = ck::type_convert<GemmAccDataType>(e_m_n_host(m, n));
+                auto d0_val = ck::type_convert<GemmAccDataType>(d0_n(n));
+                auto d1_val = ck::type_convert<GemmAccDataType>(d1_m_n(m, n));
+                cde_element_op(e_val, e_val, d0_val, d1_val);
+                e_m_n_host(m, n) = ck::type_convert<EDataType>(e_val);
+
+                auto e_val_reduce = ck::type_convert<ReduceAccDataType>(e_val);
+                qs_element_op[I1](square_e_val, e_val_reduce);
+
+                reduce0_op(reduce0_acc, e_val_reduce);
+                reduce1_op(reduce1_acc, square_e_val);
+            }
+
+            rs_element_op[I0](reduce0_acc, reduce0_acc);
+            rs_element_op[I1](reduce1_acc, reduce1_acc);
+            r0_m_host(m) = ck::type_convert<R0DataType>(reduce0_acc);
+            r1_m_host(m) = ck::type_convert<R1DataType>(reduce1_acc);
+        }
+
+        e_device_buf.FromDevice(e_m_n.mData.data());
+        r0_device_buf.FromDevice(r0_m.mData.data());
+        r1_device_buf.FromDevice(r1_m.mData.data());
+
+        pass = ck::utils::check_err(e_m_n, e_m_n_host, "Error: Incorrect results c", 1e-2, 1e-2);
+        pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
+        pass &= ck::utils::check_err(r1_m, r1_m_host, "Error: Incorrect results d1", 1e-2, 1e-2);
+    }
+
+    bool time_kernel = true;
+    if(time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+        DumpPerf<ADataType, BDataType, D0DataType, D1DataType, EDataType, R0DataType, R1DataType>(
+            ave_time, M, N, K);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp
new file mode 100644
index 0000000..e124800
--- /dev/null
+++ b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp
@@ -0,0 +1,364 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = INT8;
+using BDataType         = INT8;
+using GemmAccDataType   = INT32;
+using CShuffleDataType  = INT32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = INT8;
+using ReduceAccDataType = INT32;
+using R0DataType        = INT32;
+using R1DataType        = INT32;
+using RsDataType        = ck::Tuple<R0DataType, R1DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<PassThrough, PassThrough>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         64,                        // KPerBlock
+         16,                        // AK1
+         16,                        // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         16,                        // ABlockTransfer SrcScalarPerVector
+         16,                        // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         16,                        // BBlockTransfer SrcScalarPerVector
+         16,                        // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+using namespace ck::literals;
+
+template <typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename R0DataType,
+          typename R1DataType,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp,
+          typename QsElementOp,
+          typename RsElementOp,
+          typename RsThreadReduceOp,
+          typename ReduceAccDataType,
+          typename DeviceOpInstance,
+          typename ReferenceGemmInstance>
+bool run_gemm_reduce_add_addsquare_xdl(ck::index_t M,
+                                       ck::index_t N,
+                                       ck::index_t K,
+                                       ck::index_t StrideA,
+                                       ck::index_t StrideB,
+                                       ck::index_t StrideE,
+                                       bool do_verification,
+                                       int init_method,
+                                       bool time_kernel)
+{
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> r0_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> r1_m(f_host_tensor_descriptor1d(M, 1));
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
+        break;
+    default:
+        ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem r0_device_buf(sizeof(R0DataType) * r0_m.mDesc.GetElementSpaceSize());
+    DeviceMem r1_device_buf(sizeof(R1DataType) * r1_m.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto qs_element_op  = QsElementOp{};
+    auto rs_element_op  = RsElementOp{};
+
+    // Prepare GEMM, add, add_square
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               {},
+                               e_device_buf.GetDeviceBuffer(),
+                               {r0_device_buf.GetDeviceBuffer(), r1_device_buf.GetDeviceBuffer()},
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               {},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op,
+                               qs_element_op,
+                               rs_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    // init reducetion buffer to 0
+    r0_device_buf.SetZero();
+    r1_device_buf.SetZero();
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        auto I0 = ck::Number<0>{};
+        auto I1 = ck::Number<1>{};
+
+        Tensor<ReduceAccDataType> e_m_n_host(e_m_n.mDesc);
+        Tensor<R0DataType> r0_m_host(r0_m.mDesc);
+        Tensor<R1DataType> r1_m_host(r1_m.mDesc);
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, e_m_n_host, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        auto reduce0_op = RsThreadReduceOp{}[I0];
+        auto reduce1_op = RsThreadReduceOp{}[I1];
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto reduce0_acc = reduce0_op.template GetIdentityValue<ReduceAccDataType>();
+            auto reduce1_acc = reduce1_op.template GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType square_e_val;
+                auto e_val = ck::type_convert<ReduceAccDataType>(e_m_n_host(m, n));
+                qs_element_op[I1](square_e_val, e_val);
+
+                reduce0_op(reduce0_acc, e_val);
+                reduce1_op(reduce1_acc, square_e_val);
+            }
+
+            r0_m_host(m) = ck::type_convert<R0DataType>(reduce0_acc);
+            r1_m_host(m) = ck::type_convert<R1DataType>(reduce1_acc);
+        }
+        e_device_buf.FromDevice(e_m_n.mData.data());
+
+        Tensor<EDataType> e_m_n_host_converted(e_m_n_host);
+
+        pass = ck::utils::check_err(
+            e_m_n, e_m_n_host_converted, "Error: Incorrect results c", 1e-2, 1e-2);
+
+        r0_device_buf.FromDevice(r0_m.mData.data());
+        r1_device_buf.FromDevice(r1_m.mData.data());
+
+        pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
+        pass &= ck::utils::check_err(r1_m, r1_m_host, "Error: Incorrect results d1", 1e-2, 1e-2);
+
+        if(pass)
+        {
+            std::cout << "Success!" << std::endl;
+        }
+    }
+
+    if(time_kernel)
+    {
+        float ave_time            = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+        std::size_t flop          = 2_uz * M * N * K + 3_uz * M * N;
+        std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                    sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
+                                    sizeof(R1DataType) * M;
+
+        float tflops          = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gemm_gb_per_sec = gemm_num_byte / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gemm_gb_per_sec
+                  << " GB/s, " << std::endl;
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1152;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = 512;
+    ck::index_t StrideB = 512;
+    ck::index_t StrideE = 1152;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return !run_gemm_reduce_add_addsquare_xdl<ADataType,
+                                              BDataType,
+                                              EDataType,
+                                              R0DataType,
+                                              R1DataType,
+                                              ALayout,
+                                              BLayout,
+                                              ELayout,
+                                              AElementOp,
+                                              BElementOp,
+                                              CDEElementOp,
+                                              QsElementOp,
+                                              RsElementOp,
+                                              RsThreadReduceOp,
+                                              ReduceAccDataType,
+                                              DeviceOpInstance,
+                                              ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp
new file mode 100644
index 0000000..c2feffe
--- /dev/null
+++ b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = BF16;
+using BDataType         = BF16;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = BF16;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         32,                        // KPerBlock
+         8,                         // AK1
+         8,                         // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         8,                         // ABlockTransfer SrcScalarPerVector
+         8,                         // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         8,                         // BBlockTransfer SrcScalarPerVector
+         8,                         // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1152;
+    ck::index_t K = 256;
+
+    ck::index_t StrideA = 256;
+    ck::index_t StrideB = 256;
+    ck::index_t StrideE = 1152;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return run_gemm_reduce_max_xdl<ADataType,
+                                   BDataType,
+                                   EDataType,
+                                   R0DataType,
+                                   ALayout,
+                                   BLayout,
+                                   ELayout,
+                                   AElementOp,
+                                   BElementOp,
+                                   CDEElementOp,
+                                   QsElementOp,
+                                   RsElementOp,
+                                   RsThreadReduceOp,
+                                   ReduceAccDataType,
+                                   DeviceOpInstance,
+                                   ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp
new file mode 100644
index 0000000..363390a
--- /dev/null
+++ b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = F16;
+using BDataType         = F16;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = F16;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         32,                        // KPerBlock
+         8,                         // AK1
+         8,                         // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         8,                         // ABlockTransfer SrcScalarPerVector
+         8,                         // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         8,                         // BBlockTransfer SrcScalarPerVector
+         8,                         // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return run_gemm_reduce_max_xdl<ADataType,
+                                   BDataType,
+                                   EDataType,
+                                   R0DataType,
+                                   ALayout,
+                                   BLayout,
+                                   ELayout,
+                                   AElementOp,
+                                   BElementOp,
+                                   CDEElementOp,
+                                   QsElementOp,
+                                   RsElementOp,
+                                   RsThreadReduceOp,
+                                   ReduceAccDataType,
+                                   DeviceOpInstance,
+                                   ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp
new file mode 100644
index 0000000..de6b7eb
--- /dev/null
+++ b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp
@@ -0,0 +1,166 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = F32;
+using BDataType         = F32;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = F32;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         16,                        // KPerBlock
+         4,                         // AK1
+         4,                         // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         4,                         // ABlockTransfer SrcScalarPerVector
+         4,                         // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         4,                         // BBlockTransfer SrcScalarPerVector
+         4,                         // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: Measure kernel execution time (1=ON, 0=Off)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    return run_gemm_reduce_max_xdl<ADataType,
+                                   BDataType,
+                                   EDataType,
+                                   R0DataType,
+                                   ALayout,
+                                   BLayout,
+                                   ELayout,
+                                   AElementOp,
+                                   BElementOp,
+                                   CDEElementOp,
+                                   QsElementOp,
+                                   RsElementOp,
+                                   RsThreadReduceOp,
+                                   ReduceAccDataType,
+                                   DeviceOpInstance,
+                                   ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp
new file mode 100644
index 0000000..9666fc6
--- /dev/null
+++ b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp
@@ -0,0 +1,172 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType         = INT4;
+using ADataKernelType   = INT8;
+using BDataType         = INT4;
+using BDataKernelType   = INT8;
+using GemmAccDataType   = INT32;
+using CShuffleDataType  = INT32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = INT4;
+using EDataKernelType   = INT8;
+using ReduceAccDataType = INT32;
+using R0DataType        = INT32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataKernelType,           // ADataType
+         BDataKernelType,           // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataKernelType,           // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         64,                        // KPerBlock
+         16,                        // AK1
+         16,                        // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         16,                        // ABlockTransfer SrcScalarPerVector
+         16,                        // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         16,                        // BBlockTransfer SrcScalarPerVector
+         16,                        // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1152;
+    ck::index_t K = 256;
+
+    ck::index_t StrideA = 256;
+    ck::index_t StrideB = 256;
+    ck::index_t StrideE = 1152;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return run_gemm_reduce_max_xdl<ADataType,
+                                   BDataType,
+                                   EDataType,
+                                   R0DataType,
+                                   ALayout,
+                                   BLayout,
+                                   ELayout,
+                                   AElementOp,
+                                   BElementOp,
+                                   CDEElementOp,
+                                   QsElementOp,
+                                   RsElementOp,
+                                   RsThreadReduceOp,
+                                   ReduceAccDataType,
+                                   DeviceOpInstance,
+                                   ReferenceGemmInstance,
+                                   ADataKernelType,
+                                   BDataKernelType,
+                                   EDataKernelType>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp
new file mode 100644
index 0000000..00e0b76
--- /dev/null
+++ b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp
@@ -0,0 +1,166 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType         = INT8;
+using BDataType         = INT8;
+using GemmAccDataType   = INT32;
+using CShuffleDataType  = INT32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = INT8;
+using ReduceAccDataType = INT32;
+using R0DataType        = INT32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         64,                        // KPerBlock
+         16,                        // AK1
+         16,                        // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         16,                        // ABlockTransfer SrcScalarPerVector
+         16,                        // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         16,                        // BBlockTransfer SrcScalarPerVector
+         16,                        // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1152;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = 512;
+    ck::index_t StrideB = 512;
+    ck::index_t StrideE = 1152;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return run_gemm_reduce_max_xdl<ADataType,
+                                   BDataType,
+                                   EDataType,
+                                   R0DataType,
+                                   ALayout,
+                                   BLayout,
+                                   ELayout,
+                                   AElementOp,
+                                   BElementOp,
+                                   CDEElementOp,
+                                   QsElementOp,
+                                   RsElementOp,
+                                   RsThreadReduceOp,
+                                   ReduceAccDataType,
+                                   DeviceOpInstance,
+                                   ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp
new file mode 100644
index 0000000..652c0e6
--- /dev/null
+++ b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp
@@ -0,0 +1,174 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = BF16;
+using BDataType         = BF16;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = BF16;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using R1DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType, R1DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using Div          = ck::tensor_operation::element_wise::UnaryDivide;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<Div, Div>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         32,                        // KPerBlock
+         8,                         // AK1
+         8,                         // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         8,                         // ABlockTransfer SrcScalarPerVector
+         8,                         // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         8,                         // BBlockTransfer SrcScalarPerVector
+         8,                         // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1152;
+    ck::index_t K = 192;
+
+    ck::index_t StrideA = 192;
+    ck::index_t StrideB = 192;
+    ck::index_t StrideE = 1152;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return !run_gemm_reduce_mean_meansquare_xdl<ADataType,
+                                                BDataType,
+                                                EDataType,
+                                                R0DataType,
+                                                R1DataType,
+                                                ALayout,
+                                                BLayout,
+                                                ELayout,
+                                                AElementOp,
+                                                BElementOp,
+                                                CDEElementOp,
+                                                QsElementOp,
+                                                RsElementOp,
+                                                RsThreadReduceOp,
+                                                ReduceAccDataType,
+                                                DeviceOpInstance,
+                                                ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp
new file mode 100644
index 0000000..7eee24f
--- /dev/null
+++ b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp
@@ -0,0 +1,174 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = F16;
+using BDataType         = F16;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = F16;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using R1DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType, R1DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using Div          = ck::tensor_operation::element_wise::UnaryDivide;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<Div, Div>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         32,                        // KPerBlock
+         8,                         // AK1
+         8,                         // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         8,                         // ABlockTransfer SrcScalarPerVector
+         8,                         // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         8,                         // BBlockTransfer SrcScalarPerVector
+         8,                         // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return !run_gemm_reduce_mean_meansquare_xdl<ADataType,
+                                                BDataType,
+                                                EDataType,
+                                                R0DataType,
+                                                R1DataType,
+                                                ALayout,
+                                                BLayout,
+                                                ELayout,
+                                                AElementOp,
+                                                BElementOp,
+                                                CDEElementOp,
+                                                QsElementOp,
+                                                RsElementOp,
+                                                RsThreadReduceOp,
+                                                ReduceAccDataType,
+                                                DeviceOpInstance,
+                                                ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp
new file mode 100644
index 0000000..c250b99
--- /dev/null
+++ b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp
@@ -0,0 +1,174 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = F32;
+using BDataType         = F32;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = F32;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using R1DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType, R1DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using Div          = ck::tensor_operation::element_wise::UnaryDivide;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<Div, Div>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         16,                        // KPerBlock
+         4,                         // AK1
+         4,                         // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         4,                         // ABlockTransfer SrcScalarPerVector
+         4,                         // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         4,                         // BBlockTransfer SrcScalarPerVector
+         4,                         // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return !run_gemm_reduce_mean_meansquare_xdl<ADataType,
+                                                BDataType,
+                                                EDataType,
+                                                R0DataType,
+                                                R1DataType,
+                                                ALayout,
+                                                BLayout,
+                                                ELayout,
+                                                AElementOp,
+                                                BElementOp,
+                                                CDEElementOp,
+                                                QsElementOp,
+                                                RsElementOp,
+                                                RsThreadReduceOp,
+                                                ReduceAccDataType,
+                                                DeviceOpInstance,
+                                                ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
new file mode 100644
index 0000000..62992de
--- /dev/null
+++ b/3rdparty/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
@@ -0,0 +1,491 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/host_utility/io.hpp"
+#include "ck/stream_config.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using F16         = ck::half_t;
+using BF16        = ck::bhalf_t;
+using F32         = float;
+using F64         = double;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using INT4 = ck::int4_t;
+#endif
+using INT8  = std::int8_t;
+using INT32 = std::int32_t;
+
+template <typename ADataType, typename BDataType, typename EDataType, typename R0DataType>
+void DumpGemmReduceMaxPerf(float ave_time, int M, int N, int K)
+{
+    using namespace ck::literals;
+
+    std::size_t flop          = 2_uz * M * N * K;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(EDataType) * M * N + sizeof(R0DataType) * M;
+
+    float tflops          = static_cast<float>(flop) / 1.E9 / ave_time;
+    float gemm_gb_per_sec = gemm_num_byte / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gemm_gb_per_sec
+              << " GB/s, " << std::endl;
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename R0DataType,
+          typename R1DataType>
+void DumpGemmReduceMeanSquareMeanPerf(float ave_time, int M, int N, int K)
+{
+    using namespace ck::literals;
+
+    std::size_t flop          = 2_uz * M * N * K + M * (3_uz * N + 2_uz);
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
+                                sizeof(R1DataType) * M;
+
+    float tflops          = static_cast<float>(flop) / 1.E9 / ave_time;
+    float gemm_gb_per_sec = gemm_num_byte / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gemm_gb_per_sec
+              << " GB/s, " << std::endl;
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename R0DataType,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp,
+          typename QsElementOp,
+          typename RsElementOp,
+          typename RsThreadReduceOp,
+          typename ReduceAccDataType,
+          typename DeviceOpInstance,
+          typename ReferenceGemmInstance,
+          typename ADataKernelType = ADataType,
+          typename BDataKernelType = BDataType,
+          typename EDataKernelType = EDataType>
+auto run_gemm_reduce_max_xdl(ck::index_t M,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t StrideA,
+                             ck::index_t StrideB,
+                             ck::index_t StrideE,
+                             bool do_verification,
+                             int init_method,
+                             bool time_kernel)
+{
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+    static_assert(sizeof(ADataType) == sizeof(ADataKernelType));
+    static_assert(sizeof(BDataType) == sizeof(BDataKernelType));
+    static_assert(sizeof(EDataType) == sizeof(EDataKernelType));
+#endif
+    using namespace ck::literals;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<EDataKernelType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> r0_m(f_host_tensor_descriptor1d(M, 1));
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
+        break;
+    default:
+        ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataKernelType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataKernelType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataKernelType) * e_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem r0_device_buf(sizeof(R0DataType) * r0_m.mDesc.GetElementSpaceSize());
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    if constexpr(std::is_same_v<ADataType, ck::int4_t>)
+    {
+        Tensor<ADataKernelType> a_m_k_converted = a_m_k.template CopyAsType<ADataKernelType>();
+        Tensor<BDataKernelType> b_k_n_converted = b_k_n.template CopyAsType<BDataKernelType>();
+
+        a_device_buf.ToDevice(a_m_k_converted.mData.data());
+        b_device_buf.ToDevice(b_k_n_converted.mData.data());
+    }
+    else
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    {
+        a_device_buf.ToDevice(a_m_k.mData.data());
+        b_device_buf.ToDevice(b_k_n.mData.data());
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto qs_element_op  = QsElementOp{};
+    auto rs_element_op  = RsElementOp{};
+
+    // Prepare GEMM, max
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument  = device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                           b_device_buf.GetDeviceBuffer(),
+                                           {},
+                                           e_device_buf.GetDeviceBuffer(),
+                                           {r0_device_buf.GetDeviceBuffer()},
+                                           M,
+                                           N,
+                                           K,
+                                           StrideA,
+                                           StrideB,
+                                           {},
+                                           StrideE,
+                                           a_element_op,
+                                           b_element_op,
+                                           cde_element_op,
+                                           qs_element_op,
+                                           rs_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    // [CAUTION]: launch_and_time_kernel will not initialize D.
+    // If we evaluate kernel multiple time but without initialize D. Verification will fail
+    r0_device_buf.SetValue(ck::NumericLimits<R0DataType>::Lowest());
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        auto I0 = ck::Number<0>{};
+
+        Tensor<ReduceAccDataType> e_m_n_host(e_m_n.mDesc);
+        Tensor<R0DataType> r0_m_host(r0_m.mDesc);
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, e_m_n_host, a_element_op, b_element_op, cde_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        auto reduce0_op = RsThreadReduceOp{}[I0];
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto reduce0_acc = reduce0_op.template GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                auto e_val = e_m_n_host(m, n);
+                reduce0_op(reduce0_acc, e_val);
+            };
+
+            r0_m_host(m) = ck::type_convert<R0DataType>(reduce0_acc);
+        }
+
+        e_device_buf.FromDevice(e_m_n.mData.data());
+        Tensor<EDataType> e_m_n_host_converted(e_m_n_host);
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        if constexpr(std::is_same_v<ADataType, ck::int4_t>)
+        {
+            Tensor<EDataType> e_m_n_device_converted(e_m_n);
+            pass = ck::utils::check_err(e_m_n_device_converted,
+                                        e_m_n_host_converted,
+                                        "Error: Incorrect results c",
+                                        1e-2,
+                                        1e-2);
+        }
+        else
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        {
+            pass = ck::utils::check_err(
+                e_m_n, e_m_n_host_converted, "Error: Incorrect results c", 1e-2, 1e-2);
+        }
+
+        r0_device_buf.FromDevice(r0_m.mData.data());
+        pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
+
+        if(pass)
+        {
+            std::cout << "Success!" << std::endl;
+        }
+    }
+
+    if(time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+        DumpGemmReduceMaxPerf<ADataType, BDataType, EDataType, R0DataType>(ave_time, M, N, K);
+    }
+
+    return pass ? 0 : 1;
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename R0DataType,
+          typename R1DataType,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp,
+          typename QsElementOp,
+          typename RsElementOp,
+          typename RsThreadReduceOp,
+          typename ReduceAccDataType,
+          typename DeviceOpInstance,
+          typename ReferenceGemmInstance,
+          typename ADataKernelType = ADataType,
+          typename BDataKernelType = BDataType,
+          typename EDataKernelType = EDataType>
+bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M,
+                                         ck::index_t N,
+                                         ck::index_t K,
+                                         ck::index_t StrideA,
+                                         ck::index_t StrideB,
+                                         ck::index_t StrideE,
+                                         bool do_verification,
+                                         int init_method,
+                                         bool time_kernel)
+{
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+    static_assert(sizeof(ADataType) == sizeof(ADataKernelType));
+    static_assert(sizeof(BDataType) == sizeof(BDataKernelType));
+    static_assert(sizeof(EDataType) == sizeof(EDataKernelType));
+#endif
+    using namespace ck::literals;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<EDataKernelType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> r0_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> r1_m(f_host_tensor_descriptor1d(M, 1));
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
+        break;
+    default:
+        ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataKernelType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataKernelType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataKernelType) * e_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem r0_device_buf(sizeof(R0DataType) * r0_m.mDesc.GetElementSpaceSize());
+    DeviceMem r1_device_buf(sizeof(R1DataType) * r1_m.mDesc.GetElementSpaceSize());
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    if constexpr(std::is_same_v<ADataType, ck::int4_t>)
+    {
+        Tensor<ADataKernelType> a_m_k_converted = a_m_k.template CopyAsType<ADataKernelType>();
+        Tensor<BDataKernelType> b_k_n_converted = b_k_n.template CopyAsType<BDataKernelType>();
+
+        a_device_buf.ToDevice(a_m_k_converted.mData.data());
+        b_device_buf.ToDevice(b_k_n_converted.mData.data());
+    }
+    else
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    {
+        a_device_buf.ToDevice(a_m_k.mData.data());
+        b_device_buf.ToDevice(b_k_n.mData.data());
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto qs_element_op  = QsElementOp{};
+    auto rs_element_op  = RsElementOp{N, N};
+
+    // Prepare GEMM, mean, mean_square
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               {},
+                               e_device_buf.GetDeviceBuffer(),
+                               {r0_device_buf.GetDeviceBuffer(), r1_device_buf.GetDeviceBuffer()},
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               {},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op,
+                               qs_element_op,
+                               rs_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    // init reducetion buffer to 0
+    r0_device_buf.SetZero();
+    r1_device_buf.SetZero();
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        auto I0 = ck::Number<0>{};
+        auto I1 = ck::Number<1>{};
+
+        Tensor<ReduceAccDataType> e_m_n_host(e_m_n.mDesc);
+        Tensor<R0DataType> r0_m_host(r0_m.mDesc);
+        Tensor<R1DataType> r1_m_host(r1_m.mDesc);
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, e_m_n_host, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        auto reduce0_op = RsThreadReduceOp{}[I0];
+        auto reduce1_op = RsThreadReduceOp{}[I1];
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto reduce0_acc = reduce0_op.template GetIdentityValue<ReduceAccDataType>();
+            auto reduce1_acc = reduce1_op.template GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType square_e_val;
+                auto e_val = ck::type_convert<ReduceAccDataType>(e_m_n_host(m, n));
+                qs_element_op[I1](square_e_val, e_val);
+
+                reduce0_op(reduce0_acc, e_val);
+                reduce1_op(reduce1_acc, square_e_val);
+            }
+
+            rs_element_op[I0](reduce0_acc, reduce0_acc);
+            rs_element_op[I1](reduce1_acc, reduce1_acc);
+            r0_m_host(m) = ck::type_convert<R0DataType>(reduce0_acc);
+            r1_m_host(m) = ck::type_convert<R1DataType>(reduce1_acc);
+        }
+        e_device_buf.FromDevice(e_m_n.mData.data());
+        Tensor<EDataType> e_m_n_host_converted(e_m_n_host);
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        if constexpr(std::is_same_v<ADataType, ck::int4_t>)
+        {
+            Tensor<EDataType> e_m_n_device_converted(e_m_n);
+            pass = ck::utils::check_err(e_m_n_device_converted,
+                                        e_m_n_host_converted,
+                                        "Error: Incorrect results c",
+                                        1e-2,
+                                        1e-2);
+        }
+        else
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        {
+            pass = ck::utils::check_err(
+                e_m_n, e_m_n_host_converted, "Error: Incorrect results c", 1e-2, 1e-2);
+        }
+
+        r0_device_buf.FromDevice(r0_m.mData.data());
+        r1_device_buf.FromDevice(r1_m.mData.data());
+
+        pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
+        pass &= ck::utils::check_err(r1_m, r1_m_host, "Error: Incorrect results d1", 1e-2, 1e-2);
+
+        if(pass)
+        {
+            std::cout << "Success!" << std::endl;
+        }
+    }
+
+    if(time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+        DumpGemmReduceMeanSquareMeanPerf<ADataType, BDataType, EDataType, R0DataType, R1DataType>(
+            ave_time, M, N, K);
+    }
+
+    return pass;
+}
diff --git a/3rdparty/composable_kernel/example/17_convnd_bwd_data/CMakeLists.txt b/3rdparty/composable_kernel/example/17_convnd_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000..fa4e65d
--- /dev/null
+++ b/3rdparty/composable_kernel/example/17_convnd_bwd_data/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp)
+target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)
+
+add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
+target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
diff --git a/3rdparty/composable_kernel/example/17_convnd_bwd_data/README.md b/3rdparty/composable_kernel/example/17_convnd_bwd_data/README.md
new file mode 100644
index 0000000..b5c8281
--- /dev/null
+++ b/3rdparty/composable_kernel/example/17_convnd_bwd_data/README.md
@@ -0,0 +1,47 @@
+# Instructions for ```example_convnd_bwd_data_xdl```
+
+## Run ```example_example_convnd_bwd_data_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+#arg4: num_dim_spatial(1|2|3)
+#arg5 to ...: N, K, C, [Z,] [Y,] X, [Di,] [Hi,] Wi, S[z,] [Sy,] Sx, [Dz,] [Dy,] Dx, [LeftPz,] [LeftPy,] LeftPx, [RightPy,] [RightPy,] RightPx
+./bin/example_convnd_bwd_data_xdl 0 1 5 
+```
+
+Result
+```
+in_n_c_hi_wi: dim 4, lengths {128, 128, 71, 71}, strides {645248, 1, 9088, 128}
+wei_k_c_y_x: dim 4, lengths {256, 128, 3, 3}, strides {1152, 1, 384, 128}
+out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
+arg.a_grid_desc_k0_m_k1_container_{128, 175232, 8}
+arg.b_grid_desc_k0_n_k1_container_{128, 128, 8}
+arg.c_grid_desc_m_n_container_{ 175232, 128}
+arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 2, 2, 2, 4, 2 ) 
+launch_and_time_kernel: grid_dim {1369, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 1 times...
+arg.a_grid_desc_k0_m_k1_container_{64, 175232, 8}
+arg.b_grid_desc_k0_n_k1_container_{64, 128, 8}
+arg.c_grid_desc_m_n_container_{ 175232, 128}
+arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 2, 2, 2, 4, 2 ) 
+launch_and_time_kernel: grid_dim {1369, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 1 times...
+arg.a_grid_desc_k0_m_k1_container_{64, 175232, 8}
+arg.b_grid_desc_k0_n_k1_container_{64, 128, 8}
+arg.c_grid_desc_m_n_container_{ 175232, 128}
+arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 2, 2, 2, 4, 2 ) 
+launch_and_time_kernel: grid_dim {1369, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 1 times...
+arg.a_grid_desc_k0_m_k1_container_{32, 175232, 8}
+arg.b_grid_desc_k0_n_k1_container_{32, 128, 8}
+arg.c_grid_desc_m_n_container_{ 175232, 128}
+arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 2, 2, 2, 4, 2 ) 
+launch_and_time_kernel: grid_dim {1369, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 1 times...
+Perf: 1.40031 ms, 69.8734 TFlops, 179.037 GB/s
+```
diff --git a/3rdparty/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_common.hpp b/3rdparty/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
new file mode 100644
index 0000000..26fa9e9
--- /dev/null
+++ b/3rdparty/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
@@ -0,0 +1,152 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
+
+void print_helper_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNdBwdDataInstance>
+int run_conv_bwd_data(bool do_verification,
+                      int init_method,
+                      bool time_kernel,
+                      const ck::utils::conv::ConvParam& conv_param,
+                      const HostTensorDescriptor& in_g_n_c_wis_desc,
+                      const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                      const HostTensorDescriptor& out_g_n_k_wos_desc,
+                      const InElementOp& in_element_op,
+                      const WeiElementOp& wei_element_op,
+                      const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in_host(in_g_n_c_wis_desc);
+    Tensor<InDataType> in_device(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> out(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in_host.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    case 2:
+        out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        break;
+    default:
+        out.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
+        wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+
+    out_device_buf.ToDevice(out.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+
+    // reset input to zero
+    in_device_buf.SetZero();
+
+    // do GEMM
+    auto conv     = DeviceConvNdBwdDataInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                      static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                      static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                      conv_param.N_,
+                                      conv_param.K_,
+                                      conv_param.C_,
+                                      conv_param.input_spatial_lengths_,
+                                      conv_param.filter_spatial_lengths_,
+                                      conv_param.GetOutputSpatialLengths(),
+                                      conv_param.conv_filter_strides_,
+                                      conv_param.conv_filter_dilations_,
+                                      conv_param.input_left_pads_,
+                                      conv_param.input_right_pads_,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        std::cout << "Not support,please check parameters or device";
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
+                                                                         InDataType,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         OutElementOp>();
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(in_host,
+                                                  wei,
+                                                  out,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        in_device_buf.FromDevice(in_device.mData.data());
+
+        return ck::utils::check_err(in_device, in_host) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp b/3rdparty/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp
new file mode 100644
index 0000000..f0896e9
--- /dev/null
+++ b/3rdparty/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp
@@ -0,0 +1,180 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_bwd_data_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+template <ck::index_t NDimSpatial>
+// clang-format off
+using DeviceConvNdBwdDataInstance = ck::tensor_operation::device::DeviceConvNdBwdDataNwcKxcNwk_Dl<
+//        ######|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|    Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+//        ######|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|        Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+//        ######|           |           |            |            |            |    Operation|     Operation|     Operation| Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+//        ######|           |           |            |            |            |             |              |              |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+                 NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 1, 8, 2>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 8, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    namespace ctc = ck::tensor_layout::convolution;
+
+    print_helper_msg();
+
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv_param{
+        2, 1, 128, 256, 256, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        do_verification                   = std::stoi(argv[1]);
+        init_method                       = std::stoi(argv[2]);
+        time_kernel                       = std::stoi(argv[3]);
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+
+        conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(conv_param.num_dim_spatial_ == 1)
+    {
+        using InLayout  = ctc::GNWC;
+        using WeiLayout = ctc::GKXC;
+        using OutLayout = ctc::GNWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<1,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<1>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+    else if(conv_param.num_dim_spatial_ == 2)
+    {
+        using InLayout  = ctc::GNHWC;
+        using WeiLayout = ctc::GKYXC;
+        using OutLayout = ctc::GNHWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<2,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<2>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+    else if(conv_param.num_dim_spatial_ == 3)
+    {
+        using InLayout  = ctc::GNDHWC;
+        using WeiLayout = ctc::GKZYXC;
+        using OutLayout = ctc::GNDHWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<3,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<3>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp b/3rdparty/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp
new file mode 100644
index 0000000..c4f2c1f
--- /dev/null
+++ b/3rdparty/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp
@@ -0,0 +1,207 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_bwd_data_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvNdBwdDataInstance = ck::tensor_operation::device::DeviceConvNdBwdDataNwcKxcNwk_Xdl<
+    NDimSpatial,    // NDimSpatial
+    InDataType,     // InDataType
+    WeiDataType,    // WeiDataType
+    OutDataType,    // OutDataType
+    AccDataType,    // AccDataType
+    InElementOp,    // InElementwiseOperation
+    WeiElementOp,   // WeiElementwiseOperation
+    OutElementOp,   // OutElementwiseOperation
+    ConvBwdDefault, // ConvolutionBackwardDataSpecialization
+    256,            // BlockSize
+    128,            // MPerBlock
+    128,            // NPerBlock
+    4,              // K0PerBlock
+    8,              // K1
+    32,             // MPerXdl
+    32,             // NPerXdl
+    2,              // MXdlPerWave
+    2,              // NXdlPerWave
+    S<4, 64, 1>,    // ABlockTransferThreadClusterLengths_K0_M_K1
+    S<1, 0, 2>,     // ABlockTransferThreadClusterArrangeOrder
+    S<1, 0, 2>,     // ABlockTransferSrcAccessOrder
+    2,              // ABlockTransferSrcVectorDim
+    8,              // ABlockTransferSrcScalarPerVector
+    8,              // ABlockTransferDstScalarPerVector_K1
+    true,           // ABlockLdsAddExtraM
+    S<4, 64, 1>,    // BBlockTransferThreadClusterLengths_K0_N_K1
+    S<2, 0, 1>,     // BBlockTransferThreadClusterArrangeOrder
+    S<0, 2, 1>,     // BBlockTransferSrcAccessOrder
+    1,              // BBlockTransferSrcVectorDim
+    2,              // BBlockTransferSrcScalarPerVector
+    8,              // BBlockTransferDstScalarPerVector_K1
+    true,           // BBlockLdsAddExtraN
+    7,
+    1>; // GemmCThreadTransferDstScalarPerVector
+
+int main(int argc, char* argv[])
+{
+    namespace ctc = ck::tensor_layout::convolution;
+
+    print_helper_msg();
+
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv_param{
+        2, 1, 128, 256, 256, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        do_verification                   = std::stoi(argv[1]);
+        init_method                       = std::stoi(argv[2]);
+        time_kernel                       = std::stoi(argv[3]);
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+
+        conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(conv_param.num_dim_spatial_ == 1)
+    {
+        using InLayout  = ctc::GNWC;
+        using WeiLayout = ctc::GKXC;
+        using OutLayout = ctc::GNWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<1,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<1>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+    else if(conv_param.num_dim_spatial_ == 2)
+    {
+        using InLayout  = ctc::GNHWC;
+        using WeiLayout = ctc::GKYXC;
+        using OutLayout = ctc::GNHWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<2,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<2>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+    else if(conv_param.num_dim_spatial_ == 3)
+    {
+        using InLayout  = ctc::GNDHWC;
+        using WeiLayout = ctc::GKZYXC;
+        using OutLayout = ctc::GNDHWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<3,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<3>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/18_batched_gemm_reduce/CMakeLists.txt b/3rdparty/composable_kernel/example/18_batched_gemm_reduce/CMakeLists.txt
new file mode 100644
index 0000000..99fc004
--- /dev/null
+++ b/3rdparty/composable_kernel/example/18_batched_gemm_reduce/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_example_executable(example_batched_gemm_reduce_xdl_fp16 batched_gemm_reduce_xdl_fp16.cpp)
+
diff --git a/3rdparty/composable_kernel/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp b/3rdparty/composable_kernel/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
new file mode 100644
index 0000000..c2e3602
--- /dev/null
+++ b/3rdparty/composable_kernel/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
@@ -0,0 +1,311 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using ADataType         = F16;
+using BDataType         = F16;
+using CDataType         = F16;
+using ReduceAccDataType = F32;
+using ReduceDataType    = F32;
+using ReducePtrsGlobal  = ck::Tuple<ReduceDataType*, ReduceDataType*>;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+using ReduceOp0  = ck::reduce::Add;
+using ReduceOp1  = ck::reduce::Add;
+using ReduceOps  = ck::Tuple<ReduceOp0, ReduceOp1>;
+
+using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
+using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps    = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
+using ReduceOutElementOps   = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
+
+using ReduceGlobalMemOps =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmSpecialization =
+    ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatchedGemmReduce_Xdl_CShuffle
+//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         DData|           A|           B|           C|         Dxs|     DxsInEleOp|     DxsAccEleOp|            D|               GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+//######|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|    Type Tuple| Elementwise| Elementwise| Elementwise|      Reduce|               |                |   MemoryData|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+//######|        |        |        |     |      |      |         |         |          |              |   Operation|   Operation|   Operation|   Operation|               |                |    Operation|                   |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+//######|        |        |        |     |      |      |         |         |          |              |            |            |            |            |               |                |             |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal,  AElementOp,  BElementOp,  CElementOp, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceGlobalMemOps, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>;
+// clang-format on
+
+using ReferenceBatchedGemmInstance =
+    ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                     BDataType,
+                                                     CDataType,
+                                                     ReduceAccDataType,
+                                                     AElementOp,
+                                                     BElementOp,
+                                                     CElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 2048;
+    ck::index_t N = 1920;
+    ck::index_t K = 2048;
+
+    ck::index_t StrideA = 2048;
+    ck::index_t StrideB = 2048;
+    ck::index_t StrideC = 1920;
+
+    ck::index_t BatchCount = 4;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 11)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+
+        BatchCount = std::stoi(argv[10]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, BatchCount\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {row * stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {col * stride, 1_uz, stride});
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(f_host_tensor_descriptor(BatchCount, M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_g_k_n(f_host_tensor_descriptor(BatchCount, K, N, StrideB, BLayout{}));
+
+    Tensor<CDataType> c_g_m_n_host_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> d0_g_m_host_result({BatchCount, M});
+    Tensor<ReduceDataType> d1_g_m_host_result({BatchCount, M});
+
+    Tensor<CDataType> c_g_m_n_device_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> d0_g_m_device_result({BatchCount, M});
+    Tensor<ReduceDataType> d1_g_m_device_result({BatchCount, M});
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
+    std::cout << "d0_g_m: " << d0_g_m_host_result.mDesc << std::endl;
+    std::cout << "d1_g_m: " << d1_g_m_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
+                                 d0_g_m_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
+                                 d1_g_m_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_g_m_k.mData.data());
+    b_device_buf.ToDevice(b_g_k_n.mData.data());
+
+    auto a_element_op                     = AElementOp{};
+    auto b_element_op                     = BElementOp{};
+    auto c_element_op                     = CElementOp{};
+    std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
+
+    auto passthrough                            = UnaryIdenticElementOp{};
+    auto square                                 = UnarySquareElementOp{};
+    std::array<void*, 2> reduce_in_element_ops  = {&passthrough, &square};
+    std::array<void*, 2> reduce_out_element_ops = {&passthrough, &passthrough};
+
+    std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
+                                      reduce1_device_buf.GetDeviceBuffer()};
+
+    // do GEMM
+    auto batched_gemm = DeviceBatchedGemmReduceInstance{};
+    auto invoker      = batched_gemm.MakeInvoker();
+    auto argument     = batched_gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                              b_device_buf.GetDeviceBuffer(),
+                                              nullptr,
+                                              {},
+                                              c_device_buf.GetDeviceBuffer(),
+                                              p_reduces,
+                                              M,
+                                              N,
+                                              K,
+                                              StrideA,
+                                              StrideB,
+                                              StrideC,
+                                              {},
+                                              gemm_element_ops,
+                                              {},
+                                              reduce_in_element_ops,
+                                              reduce_out_element_ops,
+                                              BatchCount);
+
+    if(!batched_gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    // init DO, D1 to 0
+    reduce0_device_buf.SetZero();
+    reduce1_device_buf.SetZero();
+
+    // if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
+    // will not be correct. need to set time_kernel = false for correctness test
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = std::size_t(2) * BatchCount * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * BatchCount * M * K +
+                            sizeof(BDataType) * BatchCount * K * N +
+                            sizeof(CDataType) * BatchCount * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << batched_gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
+        reduce0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
+        reduce1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        auto reduce0_op = ReduceOp0{};
+        auto reduce1_op = ReduceOp1{};
+
+        for(int batch = 0; batch < BatchCount; ++batch)
+        {
+            for(int m = 0; m < M; ++m)
+            {
+                auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
+
+                for(int n = 0; n < N; ++n)
+                {
+                    auto c_val =
+                        ck::type_convert<ReduceAccDataType>(c_g_m_n_host_result(batch, m, n));
+                    ReduceAccDataType d0_val;
+                    ReduceAccDataType d1_val;
+
+                    UnaryIdenticElementOp{}(d0_val, c_val);
+                    UnarySquareElementOp{}(d1_val, c_val);
+                    reduce0_op(reduce0_acc, d0_val);
+                    reduce1_op(reduce1_acc, d1_val);
+                }
+
+                d0_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce0_acc);
+                d1_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce1_acc);
+            }
+        }
+
+        pass = ck::utils::check_err(
+                   c_g_m_n_host_result, c_g_m_n_device_result, "Error: Incorrect results c") &&
+               ck::utils::check_err(d0_g_m_device_result,
+                                    d0_g_m_host_result,
+                                    "Error: Incorrect results! D0",
+                                    1e-4,
+                                    1e-5) &&
+               ck::utils::check_err(d1_g_m_device_result,
+                                    d1_g_m_host_result,
+                                    "Error: Incorrect results! D1",
+                                    1e-3,
+                                    1e-5);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/19_binary_elementwise/CMakeLists.txt b/3rdparty/composable_kernel/example/19_binary_elementwise/CMakeLists.txt
new file mode 100644
index 0000000..39646e0
--- /dev/null
+++ b/3rdparty/composable_kernel/example/19_binary_elementwise/CMakeLists.txt
@@ -0,0 +1,4 @@
+add_example_executable(example_broadcast_add_2d_amn_bn broadcast_add_2d_amn_bn.cpp)
+add_example_executable(example_broadcast_add_3d_am_bmnk broadcast_add_3d_am_bmnk.cpp)
+add_example_executable(example_elementwise_add_1d elementwise_add_1d.cpp)
+add_example_executable(example_elementwise_add_4d elementwise_add_4d.cpp)
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp b/3rdparty/composable_kernel/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
new file mode 100644
index 0000000..9eae27c
--- /dev/null
+++ b/3rdparty/composable_kernel/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType = F16;
+using CDataType  = F16;
+
+using Add = ck::tensor_operation::element_wise::Add;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
+                                                    ck::Tuple<CDataType>,
+                                                    Add,
+                                                    2,
+                                                    8,
+                                                    ck::Sequence<8, 8>,
+                                                    ck::Sequence<8>>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename Functor,
+          int broadcastDim>
+void host_broadcast2D(
+    HostTensorC& C, const HostTensorA& A, const HostTensorB& B, int M, int N, Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(int m = 0; m < M; ++m)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            auto Amn  = A(m, n);
+            ctype Cmn = 0;
+            if constexpr(broadcastDim == 0)
+            {
+                auto Bn = B(n);
+                functor(Cmn, Amn, Bn);
+            }
+            else
+            {
+                auto Bm = B(m);
+                functor(Cmn, Amn, Bm);
+            }
+            C(m, n) = Cmn;
+        }
+    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    ck::index_t M      = 1024;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = 1024;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
+        using namespace ck::literals;
+
+        return HostTensorDescriptor({row, col}, {stride, 1_uz});
+    };
+
+    Tensor<ABDataType> a_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+    Tensor<ABDataType> b_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+
+    a_m_n.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_n.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_n_device_buf(sizeof(ABDataType) * a_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem b_n_device_buf(sizeof(ABDataType) * b_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n.mDesc.GetElementSpaceSize());
+
+    a_m_n_device_buf.ToDevice(a_m_n.mData.data());
+    b_n_device_buf.ToDevice(b_n.mData.data());
+
+    std::array<const void*, 2> input = {a_m_n_device_buf.GetDeviceBuffer(),
+                                        b_n_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {c_m_n_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 2> abc_lengths = {M, N};
+    std::array<ck::index_t, 2> a_strides   = {Stride, 1};
+    std::array<ck::index_t, 2> b_strides   = {0, 1};
+    std::array<ck::index_t, 2> c_strides   = {Stride, 1};
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(
+        abc_lengths, {a_strides, b_strides}, {c_strides}, input, output, Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_n_device_buf.FromDevice(c_m_n.mData.data());
+        Tensor<CDataType> host_c_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+
+        host_broadcast2D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add, 0>(
+            host_c_m_n, a_m_n, b_n, M, N, Add{});
+
+        pass &= ck::utils::check_err(c_m_n, host_c_m_n, "Error: Incorrect results c", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp b/3rdparty/composable_kernel/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
new file mode 100644
index 0000000..813d38b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
@@ -0,0 +1,120 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType = F16;
+using CDataType  = F16;
+
+using Add = ck::tensor_operation::element_wise::Add;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
+                                                    ck::Tuple<CDataType>,
+                                                    Add,
+                                                    3,
+                                                    8,
+                                                    ck::Sequence<1, 8>,
+                                                    ck::Sequence<8>>;
+
+template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
+void host_broadcast3D_am_bmnk(HostTensorC& C,
+                              const HostTensorA& A,
+                              const HostTensorB& B,
+                              const std::vector<std::size_t>& shape,
+                              Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(std::size_t m = 0; m < shape[0]; ++m)
+        for(std::size_t n = 0; n < shape[1]; ++n)
+            for(std::size_t k = 0; k < shape[2]; ++k)
+            {
+                auto a_val  = A(m);
+                auto b_val  = B(m, n, k);
+                ctype c_val = 0;
+                functor(c_val, a_val, b_val);
+                C(m, n, k) = c_val;
+            }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    std::vector<std::size_t> mnk = {4, 16, 32};
+    ck::index_t M                = mnk[0];
+
+    Tensor<ABDataType> a_m({M});
+    Tensor<ABDataType> b_m_n_k(mnk);
+    Tensor<CDataType> c_m_n_k(mnk);
+
+    a_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_m_n_k.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_device_buf(sizeof(ABDataType) * a_m.mDesc.GetElementSpaceSize());
+    DeviceMem b_m_n_k_device_buf(sizeof(ABDataType) * b_m_n_k.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_k_device_buf(sizeof(CDataType) * c_m_n_k.mDesc.GetElementSpaceSize());
+
+    a_m_device_buf.ToDevice(a_m.mData.data());
+    b_m_n_k_device_buf.ToDevice(b_m_n_k.mData.data());
+
+    std::array<const void*, 2> input = {a_m_device_buf.GetDeviceBuffer(),
+                                        b_m_n_k_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {c_m_n_k_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 3> abc_lengths;
+    std::array<ck::index_t, 3> a_strides = {1, 0, 0};
+    std::array<ck::index_t, 3> b_strides;
+    std::array<ck::index_t, 3> c_strides;
+
+    ck::ranges::copy(mnk, abc_lengths.begin());
+    ck::ranges::copy(b_m_n_k.mDesc.GetStrides(), b_strides.begin());
+    ck::ranges::copy(c_m_n_k.mDesc.GetStrides(), c_strides.begin());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(
+        abc_lengths, {a_strides, b_strides}, {c_strides}, input, output, Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_n_k_device_buf.FromDevice(c_m_n_k.mData.data());
+        Tensor<CDataType> host_c_m_n_k(mnk);
+
+        host_broadcast3D_am_bmnk<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
+            host_c_m_n_k, a_m, b_m_n_k, mnk, Add{});
+
+        pass &=
+            ck::utils::check_err(c_m_n_k, host_c_m_n_k, "Error: Incorrect results c", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/19_binary_elementwise/elementwise_add_1d.cpp b/3rdparty/composable_kernel/example/19_binary_elementwise/elementwise_add_1d.cpp
new file mode 100644
index 0000000..a1ca937
--- /dev/null
+++ b/3rdparty/composable_kernel/example/19_binary_elementwise/elementwise_add_1d.cpp
@@ -0,0 +1,111 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType = F16;
+using CDataType  = F16;
+
+using Add = ck::tensor_operation::element_wise::Add;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
+                                                    ck::Tuple<CDataType>,
+                                                    Add,
+                                                    1,
+                                                    8,
+                                                    ck::Sequence<8, 8>,
+                                                    ck::Sequence<8>>;
+
+template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
+void host_elementwise1D(
+    HostTensorC& C, const HostTensorA& A, const HostTensorB& B, int M, Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0))>;
+
+    for(int m = 0; m < M; ++m)
+    {
+        auto Am  = A(m);
+        auto Bm  = B(m);
+        ctype Cm = 0;
+        functor(Cm, Am, Bm);
+        C(m) = Cm;
+    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    ck::index_t M = 1024;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    Tensor<ABDataType> a_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<ABDataType> b_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<CDataType> c_m(f_host_tensor_descriptor1d(M, 1));
+
+    a_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_device_buf(sizeof(ABDataType) * a_m.mDesc.GetElementSpaceSize());
+    DeviceMem b_m_device_buf(sizeof(ABDataType) * b_m.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_device_buf(sizeof(CDataType) * c_m.mDesc.GetElementSpaceSize());
+
+    a_m_device_buf.ToDevice(a_m.mData.data());
+    b_m_device_buf.ToDevice(b_m.mData.data());
+
+    std::array<const void*, 2> input = {a_m_device_buf.GetDeviceBuffer(),
+                                        b_m_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {c_m_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 1> abc_lengths = {M};
+    std::array<ck::index_t, 1> a_strides   = {1};
+    std::array<ck::index_t, 1> b_strides   = {1};
+    std::array<ck::index_t, 1> c_strides   = {1};
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(
+        abc_lengths, {a_strides, b_strides}, {c_strides}, input, output, Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_device_buf.FromDevice(c_m.mData.data());
+        Tensor<CDataType> host_c_m(f_host_tensor_descriptor1d(M, 1));
+
+        host_elementwise1D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
+            host_c_m, a_m, b_m, M, Add{});
+
+        pass &= ck::utils::check_err(c_m, host_c_m, "Error: Incorrect results c", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/19_binary_elementwise/elementwise_add_4d.cpp b/3rdparty/composable_kernel/example/19_binary_elementwise/elementwise_add_4d.cpp
new file mode 100644
index 0000000..27e1001
--- /dev/null
+++ b/3rdparty/composable_kernel/example/19_binary_elementwise/elementwise_add_4d.cpp
@@ -0,0 +1,120 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType = F16;
+using CDataType  = F16;
+
+using Add = ck::tensor_operation::element_wise::Add;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
+                                                    ck::Tuple<CDataType>,
+                                                    Add,
+                                                    4,
+                                                    8,
+                                                    ck::Sequence<8, 8>,
+                                                    ck::Sequence<8>>;
+
+template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
+void host_elementwise4D(HostTensorC& C,
+                        const HostTensorA& A,
+                        const HostTensorB& B,
+                        const std::vector<std::size_t>& shape,
+                        Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0, 0, 0))>;
+
+    for(std::size_t n = 0; n < shape[0]; ++n)
+        for(std::size_t c = 0; c < shape[1]; ++c)
+            for(std::size_t h = 0; h < shape[2]; ++h)
+                for(std::size_t w = 0; w < shape[3]; ++w)
+                {
+                    auto a_val  = A(n, c, h, w);
+                    auto b_val  = B(n, c, h, w);
+                    ctype c_val = 0;
+                    functor(c_val, a_val, b_val);
+                    C(n, c, h, w) = c_val;
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    std::vector<std::size_t> nchw = {4, 16, 32, 32};
+
+    Tensor<ABDataType> a(nchw);
+    Tensor<ABDataType> b(nchw);
+    Tensor<CDataType> c(nchw);
+
+    a.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ABDataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(ABDataType) * b.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+    b_device_buf.ToDevice(b.mData.data());
+
+    std::array<const void*, 2> input = {a_device_buf.GetDeviceBuffer(),
+                                        b_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {c_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 4> abc_lengths;
+    std::array<ck::index_t, 4> a_strides;
+    std::array<ck::index_t, 4> b_strides;
+    std::array<ck::index_t, 4> c_strides;
+
+    ck::ranges::copy(nchw, abc_lengths.begin());
+    ck::ranges::copy(a.mDesc.GetStrides(), a_strides.begin());
+    ck::ranges::copy(b.mDesc.GetStrides(), b_strides.begin());
+    ck::ranges::copy(c.mDesc.GetStrides(), c_strides.begin());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(
+        abc_lengths, {a_strides, b_strides}, {c_strides}, input, output, Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c.mData.data());
+        Tensor<CDataType> host_c(nchw);
+
+        host_elementwise4D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
+            host_c, a, b, nchw, Add{});
+
+        pass &= ck::utils::check_err(c, host_c, "Error: Incorrect results c", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/CMakeLists.txt b/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/CMakeLists.txt
new file mode 100644
index 0000000..557f797
--- /dev/null
+++ b/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/CMakeLists.txt
@@ -0,0 +1,8 @@
+add_custom_target(example_grouped_conv_bwd_weight)
+
+add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
+add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
+
+
+add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16
+                                                 example_grouped_conv_bwd_weight_xdl_bf16)
diff --git a/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/common.hpp b/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/common.hpp
new file mode 100644
index 0000000..d2a8bed
--- /dev/null
+++ b/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/common.hpp
@@ -0,0 +1,138 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <iostream>
+#include <iterator>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp"
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+template <typename InputLay, typename WeightLay, typename OutputLay>
+struct CommonLayoutSetting
+{
+    using InputLayout  = InputLay;
+    using WeightLayout = WeightLay;
+    using OutputLayout = OutputLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct CommonLayoutSettingSelector;
+
+namespace ctl = ck::tensor_layout::convolution;
+
+template <>
+struct CommonLayoutSettingSelector<1> final : CommonLayoutSetting<ctl::GNWC, ctl::GKXC, ctl::GNWK>
+{
+};
+
+template <>
+struct CommonLayoutSettingSelector<2> final
+    : CommonLayoutSetting<ctl::GNHWC, ctl::GKYXC, ctl::GNHWK>
+{
+};
+
+template <>
+struct CommonLayoutSettingSelector<3> final
+    : CommonLayoutSetting<ctl::GNDHWC, ctl::GKZYXC, ctl::GNDHWK>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using InputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::InputLayout;
+
+template <ck::index_t NDimSpatial>
+using WeightLayout = typename CommonLayoutSettingSelector<NDimSpatial>::WeightLayout;
+
+template <ck::index_t NDimSpatial>
+using OutputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::OutputLayout;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+#define DefaultConvParam                                                      \
+    ck::utils::conv::ConvParam                                                \
+    {                                                                         \
+        2, 4, 1, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, { 1, 1 } \
+    }
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ExecutionConfig& config,
+                           ck::utils::conv::ConvParam& conv_param)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        conv_param                        = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
diff --git a/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp b/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
new file mode 100644
index 0000000..9035309
--- /dev/null
+++ b/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using InDataType = BF16;
+// bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory
+using WeiDataType = F32;
+using OutDataType = BF16;
+using AccDataType = F32;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = PassThrough;
+
+#include "run_grouped_conv_bwd_weight_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp b/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
new file mode 100644
index 0000000..6791b0b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
@@ -0,0 +1,17 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using InDataType  = F16;
+using WeiDataType = F16;
+using OutDataType = F16;
+using AccDataType = F32;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = PassThrough;
+
+#include "run_grouped_conv_bwd_weight_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc b/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
new file mode 100644
index 0000000..5264c85
--- /dev/null
+++ b/3rdparty/composable_kernel/example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
@@ -0,0 +1,206 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+template <ck::index_t NDimSpatial>
+using DeviceConvBwdWeightInstance =
+    ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
+        NDimSpatial,          // NDimSpatial
+        InDataType,           // InDataType
+        WeiDataType,          // WeiDataType
+        OutDataType,          // OutDataType
+        AccDataType,          // AccDataType
+        InElementOp,          // InElementwiseOperation
+        WeiElementOp,         // WeiElementwiseOperation
+        OutElementOp,         // OutElementwiseOperation
+        ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
+        256,                  // BlockSize
+        128,                  // MPerBlock
+        128,                  // NPerBlock
+        4,                    // K0PerBlock
+        8,                    // K1
+        32,                   // MPerXdl
+        32,                   // NPerXdl
+        2,                    // MXdlPerWave
+        2,                    // NXdlPerWave
+        S<1, 4, 16, 4>,       // ABlockTransferThreadClusterLengths_K0_M_K1
+        S<0, 3, 1, 2>,        // ABlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,        // ABlockTransferSrcAccessOrder
+        2,                    // ABlockTransferSrcVectorDim
+        8,                    // ABlockTransferSrcScalarPerVector
+        2,                    // ABlockTransferDstScalarPerVector_K1
+        true,                 // ABlockLdsAddExtraM
+        S<1, 4, 16, 4>,       // BBlockTransferThreadClusterLengths_K0_N_K1
+        S<0, 3, 1, 2>,        // BBlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,        // BBlockTransferSrcAccessOrder
+        2,                    // BBlockTransferSrcVectorDim
+        8,                    // BBlockTransferSrcScalarPerVector
+        2,                    // BBlockTransferDstScalarPerVector_K1
+        true,                 // BBlockLdsAddExtraN
+        1,                    // CShuffleMXdlPerWavePerShuffle
+        1,                    // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
+
+template <ck::index_t NDimSpatial>
+using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+                                                                                     InDataType,
+                                                                                     WeiDataType,
+                                                                                     OutDataType,
+                                                                                     InElementOp,
+                                                                                     WeiElementOp,
+                                                                                     OutElementOp>;
+
+template <ck::index_t NDimSpatial>
+bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
+                                 const ck::utils::conv::ConvParam& conv_param)
+{
+    constexpr ck::index_t split_k = 2;
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<
+            InputLayout<NDimSpatial>>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<
+            WeightLayout<NDimSpatial>>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<
+            OutputLayout<NDimSpatial>>(conv_param);
+
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei_host_result(wei_g_k_c_xs_desc);
+    Tensor<WeiDataType> wei_device_result(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> out(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei_host_result.mDesc << std::endl;
+    std::cout << "out: " << out.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    out_device_buf.ToDevice(out.mData.data());
+
+    // init to 0
+    wei_device_buf.SetZero();
+
+    std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto range_copy = [](const auto& from, auto to) { std::copy(begin(from), end(from), to); };
+
+    range_copy(conv_param.input_spatial_lengths_, begin(input_spatial_lengths));
+    range_copy(conv_param.filter_spatial_lengths_, begin(filter_spatial_lengths));
+    range_copy(conv_param.output_spatial_lengths_, begin(output_spatial_lengths));
+    range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
+    range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
+    range_copy(conv_param.input_left_pads_, begin(input_left_pads));
+    range_copy(conv_param.input_right_pads_, begin(input_right_pads));
+
+    // do GEMM
+    auto conv     = DeviceConvBwdWeightInstance<NDimSpatial>{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                      static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                      static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                      conv_param.G_,
+                                      conv_param.N_,
+                                      conv_param.K_,
+                                      conv_param.C_,
+                                      input_spatial_lengths,
+                                      filter_spatial_lengths,
+                                      output_spatial_lengths,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      InElementOp{},
+                                      WeiElementOp{},
+                                      OutElementOp{},
+                                      split_k);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_conv with the specified compilation parameters does "
+                     "not support this Conv problem"
+                  << std::endl;
+        return false;
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+    std::cerr << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl
+              << "DeviceOp: " << conv.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_conv     = HostConvBwdWeightInstance<NDimSpatial>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei_host_result,
+                                                  out,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  InElementOp{},
+                                                  WeiElementOp{},
+                                                  OutElementOp{});
+
+        ref_invoker.Run(ref_argument);
+
+        wei_device_buf.FromDevice(wei_device_result.mData.data());
+
+        return ck::utils::check_err(wei_device_result.mData, wei_host_result.mData);
+    }
+
+    return true;
+}
+
+bool run_grouped_conv_bwd_weight_example(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return false;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return run_grouped_conv_bwd_weight<1>(config, conv_param);
+    case 2: return run_grouped_conv_bwd_weight<2>(config, conv_param);
+    case 3: return run_grouped_conv_bwd_weight<3>(config, conv_param);
+    }
+
+    return false;
+}
diff --git a/3rdparty/composable_kernel/example/21_gemm_layernorm/CMakeLists.txt b/3rdparty/composable_kernel/example/21_gemm_layernorm/CMakeLists.txt
new file mode 100644
index 0000000..78d3a5d
--- /dev/null
+++ b/3rdparty/composable_kernel/example/21_gemm_layernorm/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_fp16 gemm_bias_relu_add_layernorm_xdl_fp16.cpp)
+add_example_executable(example_gemm_layernorm_xdl_fp16 gemm_layernorm_xdl_fp16.cpp)
+add_example_executable(example_gemm_xdl_layernorm_single_kernel_fp16 gemm_xdl_layernorm_single_kernel_fp16.cpp)
diff --git a/3rdparty/composable_kernel/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_fp16.cpp b/3rdparty/composable_kernel/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_fp16.cpp
new file mode 100644
index 0000000..e37555e
--- /dev/null
+++ b/3rdparty/composable_kernel/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_fp16.cpp
@@ -0,0 +1,406 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+// DataType
+using ADataType                = F16;
+using BDataType                = F16;
+using GemmAccDataType          = F32;
+using CShuffleDataType         = F32;
+using D0DataType               = F16;
+using D1DataType               = F16;
+using DsDataType               = ck::Tuple<D0DataType, D1DataType>;
+using EDataType                = F16;
+using ReduceAccDataType        = F32;
+using R0DataType               = F32;
+using R1DataType               = F32;
+using RsDataType               = ck::Tuple<R0DataType, R1DataType>;
+using GammaDataType            = F16;
+using BetaDataType             = F16;
+using LayerNormOutDataType     = F16;
+using NormalizeComputeDataType = F32;
+
+// Layout
+using ALayout  = Row;
+using BLayout  = Col;
+using D1Layout = Row;
+using ELayout  = D1Layout;
+
+// Elementwise op
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd   = ck::tensor_operation::element_wise::AddReluAdd;
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using Div          = ck::tensor_operation::element_wise::UnaryDivide;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddReluAdd;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<Div, Div>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+//######| ALayout| BLayout| ELayout|     AData|     BData|     GemmAccData|         CShuffle|     DsData|     EData|     ReduceAccData|     RsData|           A|           B|          CDE|          Qs|          Rs|           Thread|           Global|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|    CDRThreadTransfer|                  CDE|    RThreadTransfer|
+//######|        |        |        |      Type|      Type|            Type|         DataType|       Type|      Type|              Type|       Type| Elementwise| Elementwise|  Elementwise| Elementwise| Elementwise|           Reduce|           Reduce| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|       ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
+//######|        |        |        |          |          |                |                 |           |          |                  |           |   Operation|   Operation|    Operation|   Operation|   Operation|        Operation|        Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _MPerBlock_NPerBlock|      ScalarPerVector|         _MPerBlock|
+//######|        |        |        |          |          |                |                 |           |          |                  |           |            |            |             |            |            |                 |                 |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |           _NPerBlock|                   |
+        < ALayout, BLayout, ELayout, ADataType, BDataType, GemmAccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType,  AElementOp,  BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<64, 4>,                    4,                  1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        EDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
+
+// A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
+using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
+    ck::Tuple<EDataType,
+              R0DataType,
+              R1DataType,
+              GammaDataType,
+              BetaDataType>,         // x(gemm_out), mean, meansquare, gamma, beta
+    ck::Tuple<LayerNormOutDataType>, // y
+    NormalizeFunctor,
+    2,
+    8,                           // MPerthread
+    ck::Sequence<8, 1, 1, 8, 8>, // scalarPerVector: x(gemm_out), mean, meansquare, gamma, beta
+    ck::Sequence<8>>;            // scalarPerVector: y(layerNorm_out)
+
+auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+    return HostTensorDescriptor({len}, {stride});
+};
+
+auto f_host_tensor_descriptor2d =
+    [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({row, col}, {stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({row, col}, {1_uz, stride});
+        }
+    };
+
+void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
+                         const Tensor<ADataType>& a_m_k,
+                         const Tensor<BDataType>& b_k_n,
+                         const Tensor<D0DataType>& bias_n,
+                         const Tensor<D1DataType>& d1_m_n,
+                         const Tensor<GammaDataType>& gamma_n,
+                         const Tensor<BetaDataType>& beta_n,
+                         AElementOp a_element_op,
+                         BElementOp b_element_op,
+                         CDEElementOp cde_element_op,
+                         int M,
+                         int N)
+{
+
+    int StrideE = N;
+    Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> mean_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
+    auto averageOpInst = Div{N};
+
+    auto ref_gemm    = ReferenceGemmInstance{};
+    auto ref_invoker = ref_gemm.MakeInvoker();
+
+    auto ref_argument =
+        ref_gemm.MakeArgument(a_m_k, b_k_n, e_m_n, a_element_op, b_element_op, PassThrough{});
+
+    ref_invoker.Run(ref_argument);
+
+    // c = activation(c + bias) + c1_functor(c1)
+    for(int m = 0; m < M; ++m)
+        for(int n = 0; n < N; ++n)
+        {
+            auto acc = ck::type_convert<GemmAccDataType>(e_m_n(m, n));
+            cde_element_op(e_m_n(m, n), acc, bias_n(n), d1_m_n(m, n));
+        }
+
+    // reduce_mean and reduce_square_mean
+    auto r0Op = R0ThreadReduceOp{};
+    auto r1Op = R1ThreadReduceOp{};
+    for(int m = 0; m < M; ++m)
+    {
+        auto mean_acc        = r0Op.GetIdentityValue<ReduceAccDataType>();
+        auto mean_square_acc = r1Op.GetIdentityValue<ReduceAccDataType>();
+
+        for(int n = 0; n < N; ++n)
+        {
+            auto e_val                     = ck::type_convert<ReduceAccDataType>(e_m_n(m, n));
+            ReduceAccDataType square_e_val = 0;
+            Square{}(square_e_val, e_val);
+
+            r0Op(mean_acc, e_val);
+            r1Op(mean_square_acc, square_e_val);
+        }
+
+        averageOpInst(mean_acc, mean_acc);
+        averageOpInst(mean_square_acc, mean_square_acc);
+        mean_m(m)       = ck::type_convert<R0DataType>(mean_acc);
+        meanSquare_m(m) = ck::type_convert<R1DataType>(mean_square_acc);
+    }
+
+    // LayerNorm
+    auto layerNormInst = NormalizeFunctor{};
+    for(int m = 0; m < M; ++m)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            LayerNormOutDataType out_val = 0;
+            layerNormInst(out_val, e_m_n(m, n), mean_m(m), meanSquare_m(m), gamma_n(n), beta_n(n));
+            out_m_n(m, n) = out_val;
+        }
+    }
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename R0DataType,
+          typename R1DataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename NormalizeDataType>
+void DumpGemmLayerNormPerf(float gemm_reduce_time, float normalize_time, int M, int N, int K)
+{
+    std::size_t gemm_flop     = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(EDataType) * M * N + sizeof(D0DataType) * M * N +
+                                sizeof(D0DataType) * M * N + sizeof(R0DataType) * M +
+                                sizeof(R1DataType) * M;
+
+    std::size_t normalize_num_byte = sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
+                                     sizeof(R1DataType) * M + sizeof(GammaDataType) * N +
+                                     sizeof(BetaDataType) * N + sizeof(NormalizeDataType) * M * N;
+
+    float tflops               = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
+    float gemm_gb_per_sec      = gemm_num_byte / 1.E6 / gemm_reduce_time;
+    float normalize_gb_per_sec = normalize_num_byte / 1.E6 / normalize_time;
+
+    std::cout << "gemm + reduce_mean + reduce_square_mean Perf: " << gemm_reduce_time << " ms, "
+              << tflops << " TFlops, " << gemm_gb_per_sec << " GB/s, " << std::endl;
+
+    std::cout << "5-ary elementwise Perf: " << normalize_time << " ms, " << normalize_gb_per_sec
+              << " GB/s, " << std::endl;
+}
+
+int main()
+{
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = 1024;
+    ck::index_t StrideB  = 1024;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 1024;
+    ck::index_t StrideE  = 1024;
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> bias_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor2d(M, N, StrideD1, ELayout{}));
+    Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> r0_Mean_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> r1_MeanSquare_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<LayerNormOutDataType> layerNorm_m_n(
+        f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
+    b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
+    bias_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-1, 1});
+    d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-5, 5});
+    gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
+    beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(D0DataType) * bias_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem r0_Mean_device_buf(sizeof(R0DataType) * r0_Mean_m.mDesc.GetElementSpaceSize());
+    DeviceMem r1_MeanSquare_device_buf(sizeof(R1DataType) *
+                                       r1_MeanSquare_m.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpaceSize());
+    DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpaceSize());
+    DeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) *
+                                   layerNorm_m_n.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    bias_device_buf.ToDevice(bias_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+    gamma_device_buf.ToDevice(gamma_n.mData.data());
+    beta_device_buf.ToDevice(beta_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto qs_element_op  = QsElementOp{};
+    auto rs_element_op  = RsElementOp{N, N};
+
+    // Prepare GEMM, mean, mean_square
+    auto gemmReduce          = DeviceOpInstance{};
+    auto gemmReduce_invoker  = gemmReduce.MakeInvoker();
+    auto gemmReduce_argument = gemmReduce.MakeArgument(
+        a_device_buf.GetDeviceBuffer(),
+        b_device_buf.GetDeviceBuffer(),
+        {bias_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+        e_device_buf.GetDeviceBuffer(),
+        {r0_Mean_device_buf.GetDeviceBuffer(), r1_MeanSquare_device_buf.GetDeviceBuffer()},
+        M,
+        N,
+        K,
+        StrideA,
+        StrideB,
+        {StrideD0, StrideD1},
+        StrideE,
+        a_element_op,
+        b_element_op,
+        cde_element_op,
+        qs_element_op,
+        rs_element_op);
+
+    if(!gemmReduce.IsSupportedArgument(gemmReduce_argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    // init reducetion buffer to 0
+    r0_Mean_device_buf.SetZero();
+    r1_MeanSquare_device_buf.SetZero();
+
+    // Prepare LayerNorm
+    std::array<const void*, 5> input = {e_device_buf.GetDeviceBuffer(),
+                                        r0_Mean_device_buf.GetDeviceBuffer(),
+                                        r1_MeanSquare_device_buf.GetDeviceBuffer(),
+                                        gamma_device_buf.GetDeviceBuffer(),
+                                        beta_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {layerNorm_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 2> xyLengths = {M, N};
+    std::array<ck::index_t, 2> xyStrides = {StrideE, 1};
+
+    auto normalize         = DeviceNormalizeInstance{};
+    auto normalize_invoker = normalize.MakeInvoker();
+    auto normalize_argument_ptr =
+        normalize.MakeArgumentPointer(xyLengths,
+                                      {xyStrides, {1, 0}, {1, 0}, {0, 1}, {0, 1}},
+                                      {xyStrides},
+                                      input,
+                                      output,
+                                      NormalizeFunctor{});
+
+    if(!normalize.IsSupportedArgument(normalize_argument_ptr.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device, exiting!");
+    }
+
+    // run kernel
+    gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, false});
+    normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, false});
+
+    bool pass = true;
+    {
+        // verification
+        Tensor<LayerNormOutDataType> host_layerNorm_m_n(
+            f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+
+        host_gemm_layernorm(host_layerNorm_m_n,
+                            a_m_k,
+                            b_k_n,
+                            bias_n,
+                            d1_m_n,
+                            gamma_n,
+                            beta_n,
+                            a_element_op,
+                            b_element_op,
+                            cde_element_op,
+                            M,
+                            N);
+
+        layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
+        pass &= ck::utils::check_err(layerNorm_m_n,
+                                     host_layerNorm_m_n,
+                                     "Error: Incorrect results layerNorm_m_n",
+                                     1e-2,
+                                     1e-2);
+    }
+
+    {
+        // evaluate kernel perf
+        bool time_kernel = true;
+
+        float gemm_reduce_mean_reduce_square_mean_ave_time =
+            gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, time_kernel});
+        float normalize_ave_time =
+            normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+            DumpGemmLayerNormPerf<ADataType,
+                                  BDataType,
+                                  EDataType,
+                                  D0DataType,
+                                  D1DataType,
+                                  R0DataType,
+                                  R1DataType,
+                                  GammaDataType,
+                                  BetaDataType,
+                                  LayerNormOutDataType>(
+                gemm_reduce_mean_reduce_square_mean_ave_time, normalize_ave_time, M, N, K);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp b/3rdparty/composable_kernel/example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp
new file mode 100644
index 0000000..282c876
--- /dev/null
+++ b/3rdparty/composable_kernel/example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp
@@ -0,0 +1,375 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+// DataType
+using ADataType                = F16;
+using BDataType                = F16;
+using GemmAccDataType          = F32;
+using CShuffleDataType         = F32;
+using DsDataType               = ck::Tuple<>;
+using EDataType                = F16;
+using ReduceAccDataType        = F32;
+using R0DataType               = F32;
+using R1DataType               = F32;
+using RsDataType               = ck::Tuple<R0DataType, R1DataType>;
+using GammaDataType            = F16;
+using BetaDataType             = F16;
+using LayerNormOutDataType     = F16;
+using NormalizeComputeDataType = F32;
+
+// Layout
+using ALayout  = Row;
+using BLayout  = Col;
+using D1Layout = Row;
+using ELayout  = D1Layout;
+
+// Elementwise op
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using Div          = ck::tensor_operation::element_wise::UnaryDivide;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<Div, Div>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+//######| ALayout| BLayout| ELayout|     AData|     BData|     GemmAccData|         CShuffle|     DsData|     EData|     ReduceAccData|     RsData|           A|           B|          CDE|          Qs|          Rs|           Thread|           Global|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|    CDRThreadTransfer|                  CDE|    RThreadTransfer|
+//######|        |        |        |      Type|      Type|            Type|         DataType|       Type|      Type|              Type|       Type| Elementwise| Elementwise|  Elementwise| Elementwise| Elementwise|           Reduce|           Reduce| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|       ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
+//######|        |        |        |          |          |                |                 |           |          |                  |           |   Operation|   Operation|    Operation|   Operation|   Operation|        Operation|        Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _MPerBlock_NPerBlock|      ScalarPerVector|         _MPerBlock|
+//######|        |        |        |          |          |                |                 |           |          |                  |           |            |            |             |            |            |                 |                 |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |           _NPerBlock|                   |
+        < ALayout, BLayout, ELayout, ADataType, BDataType, GemmAccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType,  AElementOp,  BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<64, 4>,                    4,                  1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        EDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
+
+// A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
+using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
+    ck::Tuple<EDataType,
+              R0DataType,
+              R1DataType,
+              GammaDataType,
+              BetaDataType>,         // x(gemm_out), mean,
+                                     // meansquare,
+                                     // gamma, beta
+    ck::Tuple<LayerNormOutDataType>, // y
+    NormalizeFunctor,
+    2,
+    8,                           // MPerthread
+    ck::Sequence<8, 1, 1, 8, 8>, // scalarPerVector: x(gemm_out), mean, meansquare, gamma, beta
+    ck::Sequence<8>>;            // scalarPerVector: y(layerNorm_out)
+
+auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+    return HostTensorDescriptor({len}, {stride});
+};
+
+auto f_host_tensor_descriptor2d =
+    [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({row, col}, {stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({row, col}, {1_uz, stride});
+        }
+    };
+
+void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
+                         const Tensor<ADataType>& a_m_k,
+                         const Tensor<BDataType>& b_k_n,
+                         const Tensor<GammaDataType>& gamma_n,
+                         const Tensor<BetaDataType>& beta_n,
+                         AElementOp a_element_op,
+                         BElementOp b_element_op,
+                         CDEElementOp c_element_op,
+                         int M,
+                         int N)
+{
+    int StrideE = N;
+    Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> mean_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
+    auto averageOpInst = Div{N};
+
+    auto ref_gemm    = ReferenceGemmInstance{};
+    auto ref_invoker = ref_gemm.MakeInvoker();
+
+    auto ref_argument =
+        ref_gemm.MakeArgument(a_m_k, b_k_n, e_m_n, a_element_op, b_element_op, c_element_op);
+
+    ref_invoker.Run(ref_argument);
+
+    // reduce_mean and reduce_square_mean
+    auto r0Op = R0ThreadReduceOp{};
+    auto r1Op = R1ThreadReduceOp{};
+    for(int m = 0; m < M; ++m)
+    {
+        auto mean_acc        = r0Op.GetIdentityValue<ReduceAccDataType>();
+        auto mean_square_acc = r1Op.GetIdentityValue<ReduceAccDataType>();
+
+        for(int n = 0; n < N; ++n)
+        {
+            auto e_val                     = ck::type_convert<ReduceAccDataType>(e_m_n(m, n));
+            ReduceAccDataType square_e_val = 0;
+            Square{}(square_e_val, e_val);
+
+            r0Op(mean_acc, e_val);
+            r1Op(mean_square_acc, square_e_val);
+        }
+
+        averageOpInst(mean_acc, mean_acc);
+        averageOpInst(mean_square_acc, mean_square_acc);
+        mean_m(m)       = ck::type_convert<R0DataType>(mean_acc);
+        meanSquare_m(m) = ck::type_convert<R1DataType>(mean_square_acc);
+    }
+
+    // LayerNorm
+    auto layerNormInst = NormalizeFunctor{};
+    for(int m = 0; m < M; ++m)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            LayerNormOutDataType out_val = 0;
+            layerNormInst(out_val, e_m_n(m, n), mean_m(m), meanSquare_m(m), gamma_n(n), beta_n(n));
+            out_m_n(m, n) = out_val;
+        }
+    }
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename R0DataType,
+          typename R1DataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename NormalizeDataType>
+void DumpGemmLayerNormPerf(float gemm_reduce_time, float normalize_time, int M, int N, int K)
+{
+    std::size_t gemm_flop     = std::size_t(2) * M * N * K;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
+                                sizeof(R1DataType) * M;
+
+    std::size_t normalize_num_btye = sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
+                                     sizeof(R1DataType) * M + sizeof(GammaDataType) * N +
+                                     sizeof(BetaDataType) * N + sizeof(NormalizeDataType) * M * N;
+
+    float tflops               = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
+    float gemm_gb_per_sec      = gemm_num_byte / 1.E6 / gemm_reduce_time;
+    float normalize_gb_per_sec = normalize_num_btye / 1.E6 / normalize_time;
+
+    std::cout << "gemm + reduce_mean + reduce_square_mean Perf: " << gemm_reduce_time << " ms, "
+              << tflops << " TFlops, " << gemm_gb_per_sec << " GB/s, " << std::endl;
+
+    std::cout << "5-ary elementwise Perf: " << normalize_time << " ms, " << normalize_gb_per_sec
+              << " GB/s, " << std::endl;
+}
+
+int main()
+{
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> r0_Mean_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> r1_MeanSquare_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<LayerNormOutDataType> layerNorm_m_n(
+        f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
+    b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
+    gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
+    beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem r0_Mean_device_buf(sizeof(R0DataType) * r0_Mean_m.mDesc.GetElementSpaceSize());
+    DeviceMem r1_MeanSquare_device_buf(sizeof(R1DataType) *
+                                       r1_MeanSquare_m.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpaceSize());
+    DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpaceSize());
+    DeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) *
+                                   layerNorm_m_n.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    gamma_device_buf.ToDevice(gamma_n.mData.data());
+    beta_device_buf.ToDevice(beta_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto qs_element_op  = QsElementOp{};
+    auto rs_element_op  = RsElementOp{N, N};
+
+    // Prepare GEMM, mean, mean_square
+    auto gemmReduce          = DeviceOpInstance{};
+    auto gemmReduce_invoker  = gemmReduce.MakeInvoker();
+    auto gemmReduce_argument = gemmReduce.MakeArgument(
+        a_device_buf.GetDeviceBuffer(),
+        b_device_buf.GetDeviceBuffer(),
+        {},
+        e_device_buf.GetDeviceBuffer(),
+        {r0_Mean_device_buf.GetDeviceBuffer(), r1_MeanSquare_device_buf.GetDeviceBuffer()},
+        M,
+        N,
+        K,
+        StrideA,
+        StrideB,
+        {},
+        StrideE,
+        a_element_op,
+        b_element_op,
+        cde_element_op,
+        qs_element_op,
+        rs_element_op);
+
+    if(!gemmReduce.IsSupportedArgument(gemmReduce_argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    r0_Mean_device_buf.SetZero();
+    r1_MeanSquare_device_buf.SetZero();
+
+    // Prepare LayerNorm
+    std::array<const void*, 5> input = {e_device_buf.GetDeviceBuffer(),
+                                        r0_Mean_device_buf.GetDeviceBuffer(),
+                                        r1_MeanSquare_device_buf.GetDeviceBuffer(),
+                                        gamma_device_buf.GetDeviceBuffer(),
+                                        beta_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {layerNorm_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 2> xyLengths = {M, N};
+    std::array<ck::index_t, 2> xyStrides = {StrideE, 1};
+
+    auto normalize         = DeviceNormalizeInstance{};
+    auto normalize_invoker = normalize.MakeInvoker();
+    auto normalize_argument_ptr =
+        normalize.MakeArgumentPointer(xyLengths,
+                                      {xyStrides, {1, 0}, {1, 0}, {0, 1}, {0, 1}},
+                                      {xyStrides},
+                                      input,
+                                      output,
+                                      NormalizeFunctor{});
+
+    if(!normalize.IsSupportedArgument(normalize_argument_ptr.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device, exiting");
+    }
+
+    // run kernel
+    gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, false});
+    normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, false});
+
+    bool pass = true;
+    {
+        // verification
+        Tensor<LayerNormOutDataType> host_layerNorm_m_n(
+            f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+
+        host_gemm_layernorm(host_layerNorm_m_n,
+                            a_m_k,
+                            b_k_n,
+                            gamma_n,
+                            beta_n,
+                            a_element_op,
+                            b_element_op,
+                            cde_element_op,
+                            M,
+                            N);
+
+        layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
+        pass &= ck::utils::check_err(
+            layerNorm_m_n, host_layerNorm_m_n, "Error: Incorrect results d1", 1e-3, 1e-3);
+    }
+
+    {
+        // evaluate kernel perf
+        bool time_kernel = true;
+
+        float gemm_reduce_mean_reduce_square_mean_ave_time =
+            gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, time_kernel});
+        float normalize_ave_time =
+            normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+            DumpGemmLayerNormPerf<ADataType,
+                                  BDataType,
+                                  EDataType,
+                                  R0DataType,
+                                  R1DataType,
+                                  GammaDataType,
+                                  BetaDataType,
+                                  LayerNormOutDataType>(
+                gemm_reduce_mean_reduce_square_mean_ave_time, normalize_ave_time, M, N, K);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/21_gemm_layernorm/gemm_xdl_layernorm_single_kernel_fp16.cpp b/3rdparty/composable_kernel/example/21_gemm_layernorm/gemm_xdl_layernorm_single_kernel_fp16.cpp
new file mode 100644
index 0000000..3c3e36b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/21_gemm_layernorm/gemm_xdl_layernorm_single_kernel_fp16.cpp
@@ -0,0 +1,287 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// This example demonstrate a single kernel that runs GEMM layer and laynorm in one fused kernel
+//
+// The GEMM + Layernorm implementation is a specialized kernel which allows fusing both layers
+// together given the condition GEMM extents N of MNK is spanned by a single workgroup. For example,
+// a kernel configured with NPerBlock = 128 allows to operate on all GEMM sizes if N <= 128
+//
+// D = Layernorm(acc_element_op(A * B + broadcast(bias)) + add) * broadcast(gamma) + broadcast(beta)
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using CDataType        = F16;
+using C0DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+struct Relu
+{
+    template <typename OutT, typename InT>
+    __host__ __device__ void operator()(OutT& y, const InT& x) const
+    {
+        y = x > 0 ? x : 0;
+    }
+};
+
+using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+// Elementwise operation that operates on the output of matrix multiplication
+// i.e., AccElementOp(A * B + bias)
+using AccElementOp = Relu;
+// Elementwise operation that operates on the output of layer normalization
+using CElementOp = Relu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmLayerNorm_Xdl_CShuffle
+//######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     C0Data|     GemmAcc|         CShuffle|   ReduceAcc|           A|           B|          Acc|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce|     CReduceThreadCopy|
+//######|        |        |        |      Type|      Type|      Type|       Type|    DataType|         DataType|    DataType| Elementwise| Elementwise|  Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector|
+//######|        |        |        |          |          |          |           |            |                 |            |   Operation|   Operation|    Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|            _NPerBlock|
+//######|        |        |        |          |          |          |           |            |                 |            |            |            |             |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                      |
+        <     Row,     Col,     Row, ADataType, BDataType, CDataType, C0DataType, AccDataType, CShuffleDataType, AccDataType,  AElementOp,  BElementOp, AccElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           2,               S<1, 32, 1, 8>,               8,             S<64, 4>,                     4>;
+// clang-format on
+
+using ReferenceInstance = ck::tensor_operation::host::ReferenceGemmLayernorm<ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             C0DataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             AccElementOp,
+                                                                             CElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 128;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 128;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<AccDataType> acc_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<C0DataType> c0_n_bias({N});
+    Tensor<C0DataType> c0_m_n_add(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<C0DataType> c0_n_gamma({N});
+    Tensor<C0DataType> c0_n_beta({N});
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+    std::cout << "c0_n_bias: " << c0_n_bias.mDesc << std::endl;
+    std::cout << "c0_m_n_add: " << c0_m_n_add.mDesc << std::endl;
+    std::cout << "c0_n_gamma: " << c0_n_gamma.mDesc << std::endl;
+    std::cout << "c0_n_beta: " << c0_n_beta.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+        b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+    }
+
+    c0_n_bias.GenerateTensorValue(GeneratorTensor_2<C0DataType>{-5, 5});
+    c0_m_n_add.GenerateTensorValue(GeneratorTensor_2<C0DataType>{-5, 5});
+    c0_n_gamma.GenerateTensorValue(GeneratorTensor_2<C0DataType>{0, 2});
+    c0_n_beta.GenerateTensorValue(GeneratorTensor_2<C0DataType>{0, 5});
+    c_m_n_host_result.GenerateTensorValue(GeneratorTensor_1<CDataType>{0});
+    acc_m_n_host_result.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem c0_bias_buf(sizeof(C0DataType) * c0_n_bias.mDesc.GetElementSpaceSize());
+    DeviceMem c0_add_buf(sizeof(C0DataType) * c0_m_n_add.mDesc.GetElementSpaceSize());
+    DeviceMem c0_gamma_buf(sizeof(C0DataType) * c0_n_gamma.mDesc.GetElementSpaceSize());
+    DeviceMem c0_beta_buf(sizeof(C0DataType) * c0_n_beta.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    c0_bias_buf.ToDevice(c0_n_bias.mData.data());
+    c0_add_buf.ToDevice(c0_m_n_add.mData.data());
+    c0_gamma_buf.ToDevice(c0_n_gamma.mData.data());
+    c0_beta_buf.ToDevice(c0_n_beta.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto acc_element_op = AccElementOp{};
+    auto c_element_op   = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                      static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                      static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                      static_cast<C0DataType*>(c0_add_buf.GetDeviceBuffer()),
+                                      static_cast<C0DataType*>(c0_bias_buf.GetDeviceBuffer()),
+                                      static_cast<C0DataType*>(c0_gamma_buf.GetDeviceBuffer()),
+                                      static_cast<C0DataType*>(c0_beta_buf.GetDeviceBuffer()),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      StrideC,
+                                      a_element_op,
+                                      b_element_op,
+                                      acc_element_op,
+                                      c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    // extra 6MN flops due to: bias + add + gamma + beta + norm_sub + norm_div,
+    // excluding reduction steps
+    std::size_t flop = std::size_t(2) * M * N * K + std::size_t(6) * M * N;
+    // extra MN and 3N due to c0_add (MxN), bias (1xN), gamma (1xN), beta (1xN)
+    std::size_t bytes = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                        sizeof(CDataType) * 2 * M * N + sizeof(C0DataType) * 3 * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = bytes / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+        auto ref_gemm    = ReferenceInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(a_m_k,
+                                                  b_k_n,
+                                                  c_m_n_host_result,
+                                                  c0_n_bias,
+                                                  c0_m_n_add,
+                                                  c0_n_gamma,
+                                                  c0_n_beta,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  acc_element_op,
+                                                  c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        if constexpr(std::is_same<CShuffleDataType, F32>::value)
+        {
+            pass &= ck::utils::check_err(
+                c_m_n_device_result, c_m_n_host_result, "Error: Incorrect results c");
+        }
+        else if constexpr(std::is_same<CShuffleDataType, F16>::value)
+        {
+            pass &= ck::utils::check_err(
+                c_m_n_device_result, c_m_n_host_result, "Error: Incorrect results c", 1e-2, 1e-2);
+        }
+    }
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/22_cgemm/CMakeLists.txt b/3rdparty/composable_kernel/example/22_cgemm/CMakeLists.txt
new file mode 100644
index 0000000..1564561
--- /dev/null
+++ b/3rdparty/composable_kernel/example/22_cgemm/CMakeLists.txt
@@ -0,0 +1,17 @@
+add_custom_target(example_cgemm_xdl)
+
+add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
+add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
+add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp)
+add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
+
+add_dependencies(example_cgemm_xdl 
+                 example_cgemm_xdl_bf16
+                 example_cgemm_xdl_fp16
+                 example_cgemm_xdl_fp32
+                 example_cgemm_xdl_int8)
+
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp)
+  add_dependencies(example_cgemm_xdl example_cgemm_xdl_int4)
+endif()
diff --git a/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_bf16.cpp b/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_bf16.cpp
new file mode 100644
index 0000000..92ed90c
--- /dev/null
+++ b/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_bf16.cpp
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType   = BF16;
+using BDataType   = BF16;
+using CDataType   = BF16;
+using AccDataType = F32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CDataType,                  // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     32,                         // index_t KPerBlock
+     8,                          // index_t AK1
+     8,                          // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     8,                          // index_t ABlockTransferSrcScalarPerVector
+     8,                          // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     8,                          // index_t BBlockTransferSrcScalarPerVector
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 32, 1, 8>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     8>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 416;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: run kernel # of times (>1)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(0);
+    }
+
+    return !run_cgemm_xdl<ADataType,
+                          BDataType,
+                          CDataType,
+                          ALayout,
+                          BLayout,
+                          CLayout,
+                          PassThrough,
+                          PassThrough,
+                          PassThrough,
+                          DeviceCGemmInstance,
+                          ReferenceCGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_common.hpp b/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_common.hpp
new file mode 100644
index 0000000..6aa06b7
--- /dev/null
+++ b/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_common.hpp
@@ -0,0 +1,254 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/stream_config.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16   = ck::half_t;
+using F32   = float;
+using BF16  = ck::bhalf_t;
+using INT8  = std::int8_t;
+using INT32 = std::int32_t;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using INT4 = ck::int4_t;
+#endif
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename DeviceCGemmInstance,
+          typename ReferenceCGemmInstance,
+          typename KernelADataType = ADataType,
+          typename KernelBDataType = BDataType,
+          typename KernelCDataType = CDataType>
+bool run_cgemm_xdl(ck::index_t M,
+                   ck::index_t N,
+                   ck::index_t K,
+                   ck::index_t StrideA,
+                   ck::index_t StrideB,
+                   ck::index_t StrideC,
+                   bool do_verification,
+                   int init_method,
+                   bool time_kernel)
+{
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t),
+                  "sizeof ck::int4_t and int8_t is different!");
+    static_assert(sizeof(ADataType) == sizeof(KernelADataType),
+                  "sizeof ADataType and KernelADataType is different!");
+    static_assert(sizeof(BDataType) == sizeof(KernelBDataType),
+                  "sizeof BDataType and KernelBDataType is different!");
+    static_assert(sizeof(CDataType) == sizeof(KernelCDataType),
+                  "sizeof CDataType and KernelCDataType is different!");
+#endif
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k_real(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<ADataType> a_m_k_imag(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n_real(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<BDataType> b_k_n_imag(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<KernelCDataType> c_m_n_real_device_result(
+        f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<KernelCDataType> c_m_n_imag_device_result(
+        f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k_real: " << a_m_k_real.mDesc << std::endl;
+    std::cout << "a_m_k_imag: " << a_m_k_imag.mDesc << std::endl;
+    std::cout << "b_k_n_real: " << b_k_n_real.mDesc << std::endl;
+    std::cout << "b_k_n_imag: " << b_k_n_imag.mDesc << std::endl;
+    std::cout << "c_m_n_real: " << c_m_n_real_device_result.mDesc << std::endl;
+    std::cout << "c_m_n_imag: " << c_m_n_imag_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k_real.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        a_m_k_imag.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n_real.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        b_k_n_imag.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    default:
+        a_m_k_real.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        a_m_k_imag.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        b_k_n_real.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        b_k_n_imag.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    auto cgemm = DeviceCGemmInstance{};
+
+    DeviceMem a_m_k_real_device_buf(sizeof(KernelADataType) *
+                                    a_m_k_real.mDesc.GetElementSpaceSize());
+    DeviceMem a_m_k_imag_device_buf(sizeof(KernelADataType) *
+                                    a_m_k_imag.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_real_device_buf(sizeof(KernelBDataType) *
+                                    b_k_n_real.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_imag_device_buf(sizeof(KernelBDataType) *
+                                    b_k_n_imag.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_real_device_buf(sizeof(KernelCDataType) *
+                                    c_m_n_real_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_imag_device_buf(sizeof(KernelCDataType) *
+                                    c_m_n_imag_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem workspace_device_buf(cgemm.GetWorkspaceSize(M, N, K, StrideA, StrideB, StrideC));
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    if constexpr(std::is_same_v<ADataType, ck::int4_t>)
+    {
+        Tensor<KernelADataType> a_m_k_real_converted(a_m_k_real);
+        Tensor<KernelADataType> a_m_k_imag_converted(a_m_k_imag);
+        Tensor<KernelBDataType> b_k_n_real_converted(b_k_n_real);
+        Tensor<KernelBDataType> b_k_n_imag_converted(b_k_n_imag);
+
+        a_m_k_real_device_buf.ToDevice(a_m_k_real_converted.mData.data());
+        a_m_k_imag_device_buf.ToDevice(a_m_k_imag_converted.mData.data());
+        b_k_n_real_device_buf.ToDevice(b_k_n_real_converted.mData.data());
+        b_k_n_imag_device_buf.ToDevice(b_k_n_imag_converted.mData.data());
+    }
+    else
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    {
+        a_m_k_real_device_buf.ToDevice(a_m_k_real.mData.data());
+        a_m_k_imag_device_buf.ToDevice(a_m_k_imag.mData.data());
+        b_k_n_real_device_buf.ToDevice(b_k_n_real.mData.data());
+        b_k_n_imag_device_buf.ToDevice(b_k_n_imag.mData.data());
+    }
+
+    auto a_element_op = AElementwiseOperation{};
+    auto b_element_op = BElementwiseOperation{};
+    auto c_element_op = CElementwiseOperation{};
+
+    // do GEMM
+    auto invoker = cgemm.MakeInvoker();
+    auto argument =
+        cgemm.MakeArgument(static_cast<KernelADataType*>(a_m_k_real_device_buf.GetDeviceBuffer()),
+                           static_cast<KernelADataType*>(a_m_k_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<KernelBDataType*>(b_k_n_real_device_buf.GetDeviceBuffer()),
+                           static_cast<KernelBDataType*>(b_k_n_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<KernelCDataType*>(c_m_n_real_device_buf.GetDeviceBuffer()),
+                           static_cast<KernelCDataType*>(c_m_n_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<KernelCDataType*>(workspace_device_buf.GetDeviceBuffer()),
+                           M,
+                           N,
+                           K,
+                           StrideA,
+                           StrideB,
+                           StrideC,
+                           a_element_op,
+                           b_element_op,
+                           c_element_op);
+
+    if(!cgemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_cgemm with the specified compilation parameters does "
+            "not support this CGEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(8) * M * N * K;
+    std::size_t num_btype =
+        std::size_t(2) *
+        (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N);
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << cgemm.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        Tensor<CDataType> c_m_n_real_host_result(
+            f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+        Tensor<CDataType> c_m_n_imag_host_result(
+            f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+        auto ref_cgemm    = ReferenceCGemmInstance{};
+        auto ref_invoker  = ref_cgemm.MakeInvoker();
+        auto ref_argument = ref_cgemm.MakeArgument(a_m_k_real,
+                                                   a_m_k_imag,
+                                                   b_k_n_real,
+                                                   b_k_n_imag,
+                                                   c_m_n_real_host_result,
+                                                   c_m_n_imag_host_result,
+                                                   a_element_op,
+                                                   b_element_op,
+                                                   c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        c_m_n_real_device_buf.FromDevice(c_m_n_real_device_result.mData.data());
+        c_m_n_imag_device_buf.FromDevice(c_m_n_imag_device_result.mData.data());
+
+        bool result = true;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        if constexpr(std::is_same_v<ADataType, ck::int4_t>)
+        {
+            const Tensor<CDataType> c_m_n_real_device_result_converted(c_m_n_real_device_result);
+            const Tensor<CDataType> c_m_n_imag_device_result_converted(c_m_n_imag_device_result);
+
+            result = ck::utils::check_err(c_m_n_real_device_result_converted,
+                                          c_m_n_real_host_result,
+                                          "Verification error: incorrect results in real part!",
+                                          1e-2f,
+                                          1e-1f);
+            result = result && ck::utils::check_err(
+                                   c_m_n_imag_device_result_converted,
+                                   c_m_n_imag_host_result,
+                                   "Verification error: incorrect results in imaginary part!",
+                                   1e-2f,
+                                   1e-1f);
+        }
+        else
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        {
+            result = ck::utils::check_err(c_m_n_real_device_result,
+                                          c_m_n_real_host_result,
+                                          "Verification error: incorrect results in real part!",
+                                          1e-2f,
+                                          1e-1f);
+            result = result && ck::utils::check_err(
+                                   c_m_n_imag_device_result,
+                                   c_m_n_imag_host_result,
+                                   "Verification error: incorrect results in imaginary part!",
+                                   1e-2f,
+                                   1e-1f);
+        }
+
+        return result;
+    }
+    return true;
+}
diff --git a/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_fp16.cpp b/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_fp16.cpp
new file mode 100644
index 0000000..1137373
--- /dev/null
+++ b/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_fp16.cpp
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType        = F16;
+using BDataType        = F16;
+using CDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CShuffleDataType,           // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     32,                         // index_t KPerBlock
+     8,                          // index_t AK1
+     8,                          // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     8,                          // index_t ABlockTransferSrcScalarPerVector
+     8,                          // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     8,                          // index_t BBlockTransferSrcScalarPerVector
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 32, 1, 8>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     8>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: run kernel # of times (>1)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(0);
+    }
+
+    return !run_cgemm_xdl<ADataType,
+                          BDataType,
+                          CDataType,
+                          ALayout,
+                          BLayout,
+                          CLayout,
+                          PassThrough,
+                          PassThrough,
+                          PassThrough,
+                          DeviceCGemmInstance,
+                          ReferenceCGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_fp32.cpp b/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_fp32.cpp
new file mode 100644
index 0000000..0f45c18
--- /dev/null
+++ b/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_fp32.cpp
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType   = F32;
+using BDataType   = F32;
+using CDataType   = F32;
+using AccDataType = F32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CDataType,                  // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     16,                         // index_t KPerBlock
+     4,                          // index_t AK1
+     4,                          // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     4,                          // index_t ABlockTransferSrcScalarPerVector
+     4,                          // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     4,                          // index_t BBlockTransferSrcScalarPerVector
+     4,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 16, 1, 16>,            // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     4>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: run kernel # of times (>1)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(0);
+    }
+
+    return !run_cgemm_xdl<ADataType,
+                          BDataType,
+                          CDataType,
+                          ALayout,
+                          BLayout,
+                          CLayout,
+                          PassThrough,
+                          PassThrough,
+                          PassThrough,
+                          DeviceCGemmInstance,
+                          ReferenceCGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_int4.cpp b/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_int4.cpp
new file mode 100644
index 0000000..c26a83b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_int4.cpp
@@ -0,0 +1,140 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType        = INT4;
+using BDataType        = INT4;
+using CDataType        = INT4;
+using AccDataType      = INT32;
+using CShuffleDataType = INT32;
+
+using KernelADataType = INT8;
+using KernelBDataType = INT8;
+using KernelCDataType = INT8;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     KernelADataType,            // typename ADataType
+     KernelBDataType,            // typename BDataType
+     KernelCDataType,            // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CShuffleDataType,           // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     64,                         // index_t KPerBlock
+     16,                         // index_t AK1
+     16,                         // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     16,                         // index_t ABlockTransferSrcScalarPerVector
+     16,                         // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     8,                          // index_t BBlockTransferSrcScalarPerVector
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 64, 1, 4>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     16>;                        // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // CGEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1152;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideC = N;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: time kernel (0=no, 1=yes)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return !run_cgemm_xdl<ADataType,
+                          BDataType,
+                          CDataType,
+                          ALayout,
+                          BLayout,
+                          CLayout,
+                          PassThrough,
+                          PassThrough,
+                          PassThrough,
+                          DeviceCGemmInstance,
+                          ReferenceCGemmInstance,
+                          KernelADataType,
+                          KernelBDataType,
+                          KernelCDataType>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_int8.cpp b/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_int8.cpp
new file mode 100644
index 0000000..2f24189
--- /dev/null
+++ b/3rdparty/composable_kernel/example/22_cgemm/cgemm_xdl_int8.cpp
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType   = INT8;
+using BDataType   = INT8;
+using CDataType   = INT8;
+using AccDataType = INT32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CDataType,                  // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     64,                         // index_t KPerBlock
+     16,                         // index_t AK1
+     16,                         // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     16,                         // index_t ABlockTransferSrcScalarPerVector
+     16,                         // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     8,                          // index_t BBlockTransferSrcScalarPerVector
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 64, 1, 4>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     16>;                        // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: run kernel # of times (>1)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(0);
+    }
+
+    return !run_cgemm_xdl<ADataType,
+                          BDataType,
+                          CDataType,
+                          ALayout,
+                          BLayout,
+                          CLayout,
+                          PassThrough,
+                          PassThrough,
+                          PassThrough,
+                          DeviceCGemmInstance,
+                          ReferenceCGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}
diff --git a/3rdparty/composable_kernel/example/23_softmax/CMakeLists.txt b/3rdparty/composable_kernel/example/23_softmax/CMakeLists.txt
new file mode 100644
index 0000000..dafe655
--- /dev/null
+++ b/3rdparty/composable_kernel/example/23_softmax/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_softmax_blockwise softmax_blockwise.cpp)
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/example/23_softmax/README.md b/3rdparty/composable_kernel/example/23_softmax/README.md
new file mode 100644
index 0000000..37c43e9
--- /dev/null
+++ b/3rdparty/composable_kernel/example/23_softmax/README.md
@@ -0,0 +1,18 @@
+# Instructions for ```example_softmax_blockwise```
+
+## Run ```example_softmax_blockwise```
+```bash
+# -D <xxx> : input 3-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg2: time kernel (0=no, 1=yes)
+example_softmax_blockwise -D 4,128,2048 -v 1 1 1
+```
+
+Result
+```
+launch_and_time_kernel: grid_dim {64, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 0.0242877 ms, 259.039 GB/s, DeviceReduceSoftmax<256,M_C8_S1,K_C32_S8,InSrcVectorDim_1_InSrcVectorSize_8_OutDstVectorSize_8>
+```
diff --git a/3rdparty/composable_kernel/example/23_softmax/softmax_blockwise.cpp b/3rdparty/composable_kernel/example/23_softmax/softmax_blockwise.cpp
new file mode 100644
index 0000000..8854bf0
--- /dev/null
+++ b/3rdparty/composable_kernel/example/23_softmax/softmax_blockwise.cpp
@@ -0,0 +1,264 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+using namespace ck::tensor_operation::device;
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 3;
+constexpr int NumReduceDim = 1;
+
+using DeviceInstance = DeviceSoftmaxImpl<InDataType,
+                                         AccDataType,
+                                         OutDataType,
+                                         PassThrough, // InElementwiseOperation
+                                         PassThrough, // AccElementwiseOperation
+                                         Rank,
+                                         NumReduceDim,
+                                         256, // BlockSize
+                                         8,   // ClusterM
+                                         32,  // ClusterK
+                                         1,   // SliceM
+                                         8,   // SliceK
+                                         1,   // SrcVecDim (0=M, 1=K)
+                                         8,   // SrcScalarPerVector
+                                         8>;  // OutScalarPerVector
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths   = {8, 128, 2048};
+    std::vector<AccDataType> scales = {2.0f, 2.0f};
+
+    bool do_verification = true;
+    int init_method      = 2;
+    bool time_kernel     = true;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "Arg1 -- init method (0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg2 -- time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:l:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 2 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        return (0);
+    };
+};
+
+int main(int argc, char* argv[])
+{
+    // Example: batched gemm C[G, M, N] applies max/sum reduction along N internally
+    const std::vector<int> invariantDims{0, 1};
+    const std::vector<int> reduceDims{2};
+
+    SimpleAppArgs args;
+
+    if(argc > 1)
+    {
+        if(args.processArgs(argc, argv) < 0)
+            return (-1);
+    };
+
+    Tensor<InDataType> in(args.inLengths);
+    Tensor<OutDataType> out_ref(args.inLengths);
+    Tensor<OutDataType> out(args.inLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = out.mDesc.GetStrides();
+
+    AccDataType alpha = args.scales[0];
+    AccDataType beta  = args.scales[1];
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "out: " << out.mDesc << std::endl;
+
+    std::size_t num_thread = 1;
+
+    if(args.do_verification)
+    {
+        switch(args.init_method)
+        {
+        case 0: break;
+        case 1:
+            in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-5.0, 5.0}, num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+    // std::cout << "beta = " << beta << std::endl;
+    // LogRangeAsType<float>(std::cout << "tensor in: " , in.mData, ",") << std::endl;
+    // LogRangeAsType<float>(std::cout << "tensor prior out: " , out.mData, ",") << std::endl;
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+
+    in_dev.ToDevice(in.mData.data());
+
+    if(beta != 0.0f)
+        out_dev.ToDevice(out.mData.data());
+
+    if(args.do_verification)
+    {
+        using ReferenceInstance =
+            ck::tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>;
+        ReferenceInstance ref;
+        auto ref_arg = ref.MakeArgument(in, out_ref, alpha, beta, reduceDims);
+        auto invoker = ref.MakeInvoker();
+        invoker.Run(ref_arg);
+        // LogRangeAsType<float>(std::cout << "tensor out_ref: ", out_ref.mData, ",") << std::endl;
+    };
+
+    std::vector<ck::index_t> i_inLengths;
+    std::vector<ck::index_t> i_inStrides;
+
+    i_inLengths.assign(args.inLengths.begin(), args.inLengths.end());
+    i_inStrides.assign(inStrides.begin(), inStrides.end());
+
+    auto device_instance = DeviceInstance{};
+
+    std::cout << i_inLengths.size() << ", " << i_inStrides.size() << std::endl;
+
+    auto argument_ptr = device_instance.MakeArgumentPointer(i_inLengths,
+                                                            i_inStrides,
+                                                            reduceDims,
+                                                            &alpha,
+                                                            &beta,
+                                                            in_dev.GetDeviceBuffer(),
+                                                            out_dev.GetDeviceBuffer(),
+                                                            PassThrough{},
+                                                            PassThrough{});
+
+    if(!device_instance.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+        return 1;
+    };
+
+    std::string instance_name = device_instance.GetTypeString();
+
+    auto invoker_ptr = device_instance.MakeInvokerPointer();
+
+    bool pass = true;
+    if(args.do_verification)
+    {
+        invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        out_dev.FromDevice(out.mData.data());
+        // LogRangeAsType<float>(std::cout << "tensor out: " , out.mData, ",") << std::endl;
+        pass = pass && ck::utils::check_err(out, out_ref);
+    };
+
+    float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, args.time_kernel});
+
+    std::size_t num_bytes =
+        in.mDesc.GetElementSize() * sizeof(InDataType) +
+        (beta == 0.0f ? 1 : 2) * out.mDesc.GetElementSize() * sizeof(OutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << instance_name
+              << std::endl;
+
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/24_batched_gemm/CMakeLists.txt b/3rdparty/composable_kernel/example/24_batched_gemm/CMakeLists.txt
new file mode 100644
index 0000000..7962576
--- /dev/null
+++ b/3rdparty/composable_kernel/example/24_batched_gemm/CMakeLists.txt
@@ -0,0 +1,17 @@
+add_custom_target(example_batched_gemm_xdl)
+
+add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)
+add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
+add_example_executable(example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp)
+add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
+
+add_dependencies(example_batched_gemm_xdl
+                 example_batched_gemm_xdl_fp32
+                 example_batched_gemm_xdl_fp16
+                 example_batched_gemm_xdl_bfp16
+                 example_batched_gemm_xdl_int8)
+
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)
+endif()
diff --git a/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_bfp16.cpp b/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_bfp16.cpp
new file mode 100644
index 0000000..c684c13
--- /dev/null
+++ b/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_bfp16.cpp
@@ -0,0 +1,59 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = BF16;
+using BDataType        = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#include "run_batched_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp b/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
new file mode 100644
index 0000000..d1985f9
--- /dev/null
+++ b/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
@@ -0,0 +1,59 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#include "run_batched_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp b/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
new file mode 100644
index 0000000..a92a04d
--- /dev/null
+++ b/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
@@ -0,0 +1,58 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 32, 1, 8>,               4>;
+// clang-format on
+
+#include "run_batched_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_int4.cpp b/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_int4.cpp
new file mode 100644
index 0000000..5e82cfe
--- /dev/null
+++ b/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_int4.cpp
@@ -0,0 +1,99 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = ck::int4_t;
+using BDataType        = ck::int4_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = ck::int4_t;
+
+using KernelADataType = int8_t;
+using KernelBDataType = int8_t;
+using KernelEDataType = int8_t;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+    // clang-format off
+        < ALayout,              //ALayout
+          BLayout,              //BLayout
+          DsLayout,             //DsLayout
+          ELayout,              //ELayout
+          KernelADataType,      //ADataType    
+          KernelBDataType,      //BDataType   
+          AccDataType,          //AccDataType
+          CShuffleDataType,     //CShuffleDataType
+          DsDataType,           //DsDataType
+          KernelEDataType,      //EDataType
+          AElementOp,           //AElementwiseOperation
+          BElementOp,           //BElementwiseOperation
+          CDEElementOp,         //CDEElementwiseOperation
+          GemmDefault,          //GEMMSpecialization
+          1,                    // NumGemmKPrefetchStage
+          256,                  // BlockSize
+          256,                  // MPerBlock
+          128,                  // NPerBlock
+          64,                   // KPerBlock
+          16,                   // AK1
+          16,                   // BK1
+          32,                   // MPerXdl
+          32,                   // NPerXdl
+          4,                    // MXdlPerWave
+          2,                    // NXdlPerWave
+          S<4, 64, 1>,          // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+          S<1, 0, 2>,           // ABlockTransfer ThreadCluster ArrangeOrder
+          S<1, 0, 2>,           // ABlockTransfer SrcAccessOrder
+          2,                    // ABlockTransfer SrcVectorDim
+          16,                   // ABlockTransfer SrcScalarPerVector
+          16,                   // ABlockTransfer DstScalarPerVector_K1
+          1,                    // ABlockLdsExtraM
+          S<4, 64, 1>,          // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+          S<1, 0, 2>,           // BBlockTransfer ThreadCluster ArrangeOrder
+          S<1, 0, 2>,           // BBlockTransfer SrcAccessOrder
+          2,                    // BBlockTransfer SrcVectorDim
+          16,                   // BBlockTransfer SrcScalarPerVector
+          16,                   // BBlockTransfer DstScalarPerVector_K1
+          1,                    // BBlockLdsExtraN
+          1,                    // CShuffleMXdlPerWavePerShuffle
+          1,                    // CShuffleNXdlPerWavePerShuffle
+          S<1, 64, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MWaveMPerXdl_NBlock_NWaveNPerXdl
+          16>;                  // CBlockTransferScalarPerVector_NWaveNPerXdl
+// clang-format on
+
+#define BUILD_INT4_EXAMPLE
+#include "run_batched_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_int8.cpp b/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_int8.cpp
new file mode 100644
index 0000000..ad22227
--- /dev/null
+++ b/3rdparty/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_int8.cpp
@@ -0,0 +1,56 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = int8_t;
+using BDataType        = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int8_t;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = int8_t;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16>;
+// clang-format on
+
+#include "run_batched_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/24_batched_gemm/run_batched_gemm_example.inc b/3rdparty/composable_kernel/example/24_batched_gemm/run_batched_gemm_example.inc
new file mode 100644
index 0000000..21934ad
--- /dev/null
+++ b/3rdparty/composable_kernel/example/24_batched_gemm/run_batched_gemm_example.inc
@@ -0,0 +1,240 @@
+#include <random>
+
+#pragma once
+
+struct ProblemSize final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t stride_A = K;
+    ck::index_t stride_B = K;
+    ck::index_t stride_C = N;
+
+    ck::index_t batch_stride_A = M * K;
+    ck::index_t batch_stride_B = K * N;
+    ck::index_t batch_stride_C = M * N;
+
+    ck::index_t batch_count = 16;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    using namespace ck::literals;
+
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+    static_assert(sizeof(ADataType) == sizeof(KernelADataType));
+    static_assert(sizeof(BDataType) == sizeof(KernelBDataType));
+    static_assert(sizeof(EDataType) == sizeof(KernelEDataType));
+#endif
+
+    auto& [M,
+           N,
+           K,
+           stride_A,
+           stride_B,
+           stride_C,
+           batch_stride_A,
+           batch_stride_B,
+           batch_stride_C,
+           batch_count] = problem_size;
+
+    // GEMM shape
+    auto f_host_tensor_descriptor = [](std::size_t batch_count_,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({batch_count_, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count_, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
+    Tensor<BDataType> b_g_k_n(
+        f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
+#ifdef BUILD_INT4_EXAMPLE
+    Tensor<KernelEDataType> e_g_m_n_device_result(
+        f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
+#else
+    Tensor<EDataType> e_g_m_n_device_result(
+        f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
+#endif
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "e_g_m_n: " << e_g_m_n_device_result.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(EDataType) * e_g_m_n_device_result.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<KernelADataType> a_g_m_k_converted(a_g_m_k);
+    const Tensor<KernelBDataType> b_g_k_n_converted(b_g_k_n);
+
+    a_device_buf.ToDevice(a_g_m_k_converted.mData.data());
+    b_device_buf.ToDevice(b_g_k_n_converted.mData.data());
+#else
+    a_device_buf.ToDevice(a_g_m_k.mData.data());
+    b_device_buf.ToDevice(b_g_k_n.mData.data());
+#endif
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {},
+                                      c_device_buf.GetDeviceBuffer(),
+                                      M,
+                                      N,
+                                      K,
+                                      batch_count,
+                                      stride_A,
+                                      stride_B,
+                                      {},
+                                      stride_C,
+                                      batch_stride_A,
+                                      batch_stride_B,
+                                      {},
+                                      batch_stride_C,
+                                      a_element_op,
+                                      b_element_op,
+                                      cde_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+    bool pass = true;
+
+    if(config.do_verification)
+    {
+        c_device_buf.FromDevice(e_g_m_n_device_result.mData.data());
+
+        using ReferenceBatchedGemmInstance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                             BDataType,
+                                                             EDataType,
+                                                             AccDataType,
+                                                             AElementOp,
+                                                             BElementOp,
+                                                             CDEElementOp>;
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        Tensor<EDataType> e_g_m_n_host_result(
+            f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, e_g_m_n_host_result, a_element_op, b_element_op, cde_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<EDataType> e_device_result_converted(e_g_m_n_device_result);
+        pass &= ck::utils::check_err(e_device_result_converted, e_g_m_n_host_result);
+
+#else
+        pass = ck::utils::check_err(
+            e_g_m_n_device_result, e_g_m_n_host_result, "Error: Incorrect results c");
+#endif
+    }
+
+    if(config.time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+        std::size_t flop      = std::size_t(2) * batch_count * M * N * K;
+        std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
+                                sizeof(BDataType) * batch_count * K * N +
+                                sizeof(EDataType) * batch_count * M * N;
+
+        float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    }
+
+    return pass ? 0 : 1;
+}
+
+bool run_batched_gemm_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    std::mt19937 gen(11939);
+    std::uniform_int_distribution<int> dis(0, 15);
+
+    problem_size.M = 256 * (dis(gen) + 1);
+    problem_size.N = 128 * (dis(gen) + 1);
+    problem_size.K = 64 * (dis(gen) + 2);
+
+    problem_size.stride_A = problem_size.K;
+    problem_size.stride_B = problem_size.K;
+    problem_size.stride_C = problem_size.N;
+
+    problem_size.batch_stride_A = problem_size.M * problem_size.K;
+    problem_size.batch_stride_B = problem_size.K * problem_size.N;
+    problem_size.batch_stride_C = problem_size.M * problem_size.N;
+
+    problem_size.batch_count = 16;
+
+    if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    return run_batched_gemm(problem_size, config);
+}
diff --git a/3rdparty/composable_kernel/example/25_gemm_bias_e_permute/CMakeLists.txt b/3rdparty/composable_kernel/example/25_gemm_bias_e_permute/CMakeLists.txt
new file mode 100644
index 0000000..cbc3c00
--- /dev/null
+++ b/3rdparty/composable_kernel/example/25_gemm_bias_e_permute/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_example_executable(example_gemm_bias_e_permute_g1m3n2k1_xdl_fp16 gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp)
+add_example_executable(example_gemm_bias_e_permute_g1m2n3k1_xdl_fp16 gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp)
diff --git a/3rdparty/composable_kernel/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp b/3rdparty/composable_kernel/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp
new file mode 100644
index 0000000..c934d35
--- /dev/null
+++ b/3rdparty/composable_kernel/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp
@@ -0,0 +1,397 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 1;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 3;
+static constexpr ck::index_t NumDimK = 1;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //############################################| NumDimG| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           Gemm|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################################|        |        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################################|        |        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################################|        |        |        |        |      |      |        |         |           |      |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK,   F16,   F16,     F32,      F16, DsDataType,   F16,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1, 32, 1, 4>,               8>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimG == 1 && NumDimM == 2 && NumDimN == 3 && NumDimK == 1, bool> =
+              false>
+struct ReferenceContraction_G1_M2_N3_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_gs_ms_ks,
+                 const Tensor<BDataType>& b_gs_ns_ks,
+                 Tensor<EDataType>& e_gs_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_gs_ms_ks_{a_gs_ms_ks},
+              b_gs_ns_ks_{b_gs_ns_ks},
+              e_gs_ms_ns_{e_gs_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_gs_ms_ks_;
+        const Tensor<BDataType>& b_gs_ns_ks_;
+        Tensor<EDataType>& e_gs_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_G1_M2_N3_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_gs_ms_ns = [&](auto g0, auto m0, auto m1, auto n0, auto n1, auto n2) {
+                const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[3];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a, ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, m0, m1, k0)));
+                    arg.b_element_op_(
+                        v_b,
+                        ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, n0, n1, n2, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_gs_ms_ns_(g0, m0, m1, n0, n1, n2) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_gs_ms_ns,
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
+                             const Tensor<BDataType>& b_gs_ns_ks,
+                             Tensor<EDataType>& e_gs_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_M3_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::index_t G0 = 1;
+
+    ck::index_t M0 = 4;
+    ck::index_t M1 = 256;
+
+    ck::index_t N0 = 4;
+    ck::index_t N1 = 16;
+    ck::index_t N2 = 32;
+
+    ck::index_t K0 = 256;
+
+    // A[M0, M1, M2, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[N0, N1, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, N0, N1, N2, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{N0 * N1 * N2 * K0, N1 * N2 * K0, N2 * K0, K0, 1};
+
+    // D[N0, M0, N1, M1, N2]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, M0, M1, N0, N1, N2};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{N0 * N1 * N2, 0, 0, N1 * N2, N2, 1};
+    // E[N0, M0, N1, M1, N2]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, M0, M1, N0, N1, N2};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        M0 * M1 * N0 * N1 * N2, N1 * M1 * N2, N2, M0 * N1 * M1 * N2, M1 * N2, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
+    Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_device_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
+    std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
+    std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) *
+                           e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_gs_ms_ks_lengths,
+                                    a_gs_ms_ks_strides,
+                                    b_gs_ns_ks_lengths,
+                                    b_gs_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                    e_gs_ms_ns_lengths,
+                                    e_gs_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t M = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimG, NumDimM, 1, std::multiplies<>{});
+
+    std::size_t N = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    std::size_t K = ck::accumulate_n<ck::index_t>(
+        a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop      = std::size_t(2) * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+        using ReferenceOpInstance = ReferenceContraction_G1_M2_N3_K1<NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CShuffleDataType,
+                                                                     AccDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(a_gs_ms_ks,
+                                                  b_gs_ns_ks,
+                                                  c_gs_ms_ns_host_result,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t g0 = 0; g0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[0]; ++g0)
+        {
+            for(size_t m0 = 0; m0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[1]; ++m0)
+            {
+                for(size_t m1 = 0; m1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[2]; ++m1)
+                {
+                    for(size_t n0 = 0; n0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[3]; ++n0)
+                    {
+                        for(size_t n1 = 0; n1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[4]; ++n1)
+                        {
+                            for(size_t n2 = 0; n2 < e_gs_ms_ns_host_result.mDesc.GetLengths()[5];
+                                ++n2)
+                            {
+                                cde_element_op(e_gs_ms_ns_host_result(g0, m0, m1, n0, n1, n2),
+                                               c_gs_ms_ns_host_result(g0, m0, m1, n0, n1, n2),
+                                               d_gs_ms_ns(g0, m0, m1, n0, n1, n2));
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_gs_ms_ns_device_result, e_gs_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp b/3rdparty/composable_kernel/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp
new file mode 100644
index 0000000..98835f9
--- /dev/null
+++ b/3rdparty/composable_kernel/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp
@@ -0,0 +1,398 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 1;
+static constexpr ck::index_t NumDimM = 3;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 1;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //############################################| NumDimG| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           Gemm|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################################|        |        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################################|        |        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################################|        |        |        |        |      |      |        |         |           |      |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK,   F16,   F16,     F32,      F16, DsDataType,   F16,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1, 32, 1, 4>,               8>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+template <ck::index_t NumDimG,
+          ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimG == 1 && NumDimM == 3 && NumDimN == 2 && NumDimK == 1, bool> =
+              false>
+struct ReferenceContraction_G1_M3_N2_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_gs_ms_ks,
+                 const Tensor<BDataType>& b_gs_ns_ks,
+                 Tensor<EDataType>& e_gs_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_gs_ms_ks_{a_gs_ms_ks},
+              b_gs_ns_ks_{b_gs_ns_ks},
+              e_gs_ms_ns_{e_gs_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_gs_ms_ks_;
+        const Tensor<BDataType>& b_gs_ns_ks_;
+        Tensor<EDataType>& e_gs_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_G1_M3_N2_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_gs_ms_ns = [&](auto g0, auto m0, auto m1, auto m2, auto n0, auto n1) {
+                const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a,
+                        ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, m0, m1, m2, k0)));
+                    arg.b_element_op_(
+                        v_b, ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, n0, n1, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_gs_ms_ns_(g0, m0, m1, m2, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_gs_ms_ns,
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
+                             const Tensor<BDataType>& b_gs_ns_ks,
+                             Tensor<EDataType>& e_gs_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_G1_M3_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::index_t G0 = 1;
+
+    ck::index_t M0 = 4;
+    ck::index_t M1 = 8;
+    ck::index_t M2 = 256;
+
+    ck::index_t N0 = 32;
+    ck::index_t N1 = 128;
+
+    ck::index_t K0 = 1024;
+
+    // A[M0, M1, M2, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, M0, M1, M2, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M0 * M1 * M2 * K0, M1 * M2 * K0, M2 * K0, K0, 1};
+
+    // B[N0, N1, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, N0, N1, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{N0 * N1 * K0, N1 * K0, K0, 1};
+
+    // D[M0, N0, M1, N1, M2]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, M0, M1, M2, N0, N1};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{N0 * N1, 0, 0, 0, N1, 1};
+
+    // E[M1, M0, N0, M1, N1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, M0, M1, M2, N0, N1};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        M0 * M1 * M2 * N1 * N0, N0 * M1 * N1, N1, M0 * N0 * M1 * N1, M1 * N1, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
+    Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_device_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
+    std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
+    std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) *
+                           e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_gs_ms_ks_lengths,
+                                    a_gs_ms_ks_strides,
+                                    b_gs_ns_ks_lengths,
+                                    b_gs_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                    e_gs_ms_ns_lengths,
+                                    e_gs_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t M =
+        ck::accumulate_n<ck::index_t>(e_gs_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_gs_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop      = std::size_t(2) * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+        using ReferenceOpInstance = ReferenceContraction_G1_M3_N2_K1<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CShuffleDataType,
+                                                                     AccDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(a_gs_ms_ks,
+                                                  b_gs_ns_ks,
+                                                  c_gs_ms_ns_host_result,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t g0 = 0; g0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[0]; ++g0)
+        {
+            for(size_t m0 = 0; m0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[1]; ++m0)
+            {
+                for(size_t m1 = 0; m1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[2]; ++m1)
+                {
+                    for(size_t m2 = 0; m2 < e_gs_ms_ns_host_result.mDesc.GetLengths()[3]; ++m2)
+                    {
+                        for(size_t n0 = 0; n0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[4]; ++n0)
+                        {
+                            for(size_t n1 = 0; n1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[5];
+                                ++n1)
+                            {
+                                cde_element_op(e_gs_ms_ns_host_result(g0, m0, m1, m2, n0, n1),
+                                               c_gs_ms_ns_host_result(g0, m0, m1, m2, n0, n1),
+                                               d_gs_ms_ns(g0, m0, m1, m2, n0, n1));
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_gs_ms_ns_device_result, e_gs_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/26_contraction/CMakeLists.txt b/3rdparty/composable_kernel/example/26_contraction/CMakeLists.txt
new file mode 100644
index 0000000..87f4750
--- /dev/null
+++ b/3rdparty/composable_kernel/example/26_contraction/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_example_executable(example_contraction_bilinear_xdl_fp32 contraction_bilinear_xdl_fp32.cpp)
+add_example_executable(example_contraction_scale_xdl_fp32 contraction_scale_xdl_fp32.cpp)
diff --git a/3rdparty/composable_kernel/example/26_contraction/README.md b/3rdparty/composable_kernel/example/26_contraction/README.md
new file mode 100644
index 0000000..c88d93c
--- /dev/null
+++ b/3rdparty/composable_kernel/example/26_contraction/README.md
@@ -0,0 +1,20 @@
+# Instructions for ```example_contraction_bilinear_xdl_fp32```
+
+## Run
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+./bin/example_contraction_bilinear_xdl_fp32 1 1 1
+```
+
+Result (MI100 @ dynammic freq, 46TFlops peak FP32)
+```
+a_ms_ks: dim 4, lengths {30, 128, 32, 64}, strides {524288, 4096, 128, 1}
+b_ks_ns: dim 4, lengths {32, 64, 32, 64}, strides {128, 1, 524288, 4096}
+c_ms_ns: dim 4, lengths {30, 128, 32, 64}, strides {524288, 4096, 128, 1}
+launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 0.843286 ms, 38.1985 TFlops, 94.5014 GB/s, DeviceContractionMultipleD_Xdl_CShuffle<256, 256, 128, 16, 4, 4>
+```
diff --git a/3rdparty/composable_kernel/example/26_contraction/contraction_bilinear_xdl_fp32.cpp b/3rdparty/composable_kernel/example/26_contraction/contraction_bilinear_xdl_fp32.cpp
new file mode 100644
index 0000000..ea105e4
--- /dev/null
+++ b/3rdparty/composable_kernel/example/26_contraction/contraction_bilinear_xdl_fp32.cpp
@@ -0,0 +1,427 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F32;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F32;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>;
+
+using DeviceOpInstanceKNNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>;
+
+using DeviceOpInstanceMKNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>;
+
+using DeviceOpInstanceMNNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
+struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_ms_ks,
+                 const Tensor<BDataType>& b_ns_ks,
+                 Tensor<EDataType>& e_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_ms_ks_{a_ms_ks},
+              b_ns_ks_{b_ns_ks},
+              e_ms_ns_{e_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_ms_ks_;
+        const Tensor<BDataType>& b_ns_ks_;
+        Tensor<EDataType>& e_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_M2_N2_K2::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto m0, auto m1, auto n0, auto n1) {
+                const int K0 = arg.a_ms_ks_.mDesc.GetLengths()[2];
+                const int K1 = arg.a_ms_ks_.mDesc.GetLengths()[3];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    for(int k1 = 0; k1 < K1; ++k1)
+                    {
+                        AccDataType v_a;
+                        AccDataType v_b;
+
+                        arg.a_element_op_(
+                            v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
+                        arg.b_element_op_(
+                            v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
+
+                        v_acc += v_a * v_b;
+                    }
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_ms_ns_(m0, m1, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[3])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_ms_ks,
+                             const Tensor<BDataType>& b_ns_ks,
+                             Tensor<EDataType>& e_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{a_ms_ks, b_ns_ks, e_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_M2_N2_K2"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float alpha = 1.f;
+    float beta  = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 28)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t M0 = std::stoi(argv[4]);
+        const ck::index_t M1 = std::stoi(argv[5]);
+
+        const ck::index_t N0 = std::stoi(argv[6]);
+        const ck::index_t N1 = std::stoi(argv[7]);
+
+        const ck::index_t K0 = std::stoi(argv[8]);
+        const ck::index_t K1 = std::stoi(argv[9]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
+
+        d_ms_ns_lengths = {M0, M1, N0, N1};
+        d_ms_ns_strides = {
+            std::stoi(argv[18]), std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[22]), std::stoi(argv[23]), std::stoi(argv[24]), std::stoi(argv[25])};
+
+        alpha = std::stof(argv[26]);
+        beta  = std::stof(argv[27]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 7: M0, M1, N0, N1, K0, K1\n");
+        printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg18 to 21: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
+        printf("arg22 to 25: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg26 to 27: alpha, beta\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
+    Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
+    Tensor<EDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
+    std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
+    std::cout << "d_ms_ns: " << d_ms_ns.mDesc << std::endl;
+    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{alpha, beta};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_ms_ks_lengths,
+                                    a_ms_ks_strides,
+                                    b_ns_ks_lengths,
+                                    b_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                    e_ms_ns_lengths,
+                                    e_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t M =
+        ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop      = std::size_t(2) * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                  NumDimN,
+                                                                  NumDimK,
+                                                                  ADataType,
+                                                                  BDataType,
+                                                                  CShuffleDataType,
+                                                                  AccDataType,
+                                                                  AElementOp,
+                                                                  BElementOp,
+                                                                  PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
+                                       c_ms_ns_host_result(m0, m1, n0, n1),
+                                       d_ms_ns(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/26_contraction/contraction_scale_xdl_fp32.cpp b/3rdparty/composable_kernel/example/26_contraction/contraction_scale_xdl_fp32.cpp
new file mode 100644
index 0000000..26f176b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/26_contraction/contraction_scale_xdl_fp32.cpp
@@ -0,0 +1,409 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// clang-format off
+using DeviceOpInstanceKKN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>;
+
+using DeviceOpInstanceKNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>;
+
+using DeviceOpInstanceMKN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>;
+
+using DeviceOpInstanceMNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
+struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_ms_ks,
+                 const Tensor<BDataType>& b_ns_ks,
+                 Tensor<EDataType>& e_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_ms_ks_{a_ms_ks},
+              b_ns_ks_{b_ns_ks},
+              e_ms_ns_{e_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_ms_ks_;
+        const Tensor<BDataType>& b_ns_ks_;
+        Tensor<EDataType>& e_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_M2_N2_K2::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto m0, auto m1, auto n0, auto n1) {
+                const int K0 = arg.a_ms_ks_.mDesc.GetLengths()[2];
+                const int K1 = arg.a_ms_ks_.mDesc.GetLengths()[3];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    for(int k1 = 0; k1 < K1; ++k1)
+                    {
+                        AccDataType v_a;
+                        AccDataType v_b;
+
+                        arg.a_element_op_(
+                            v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
+                        arg.b_element_op_(
+                            v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
+
+                        v_acc += v_a * v_b;
+                    }
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_ms_ns_(m0, m1, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[3])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_ms_ks,
+                             const Tensor<BDataType>& b_ns_ks,
+                             Tensor<EDataType>& e_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{a_ms_ks, b_ns_ks, e_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_M2_N2_K2"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float scale = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 23)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t M0 = std::stoi(argv[4]);
+        const ck::index_t M1 = std::stoi(argv[5]);
+
+        const ck::index_t N0 = std::stoi(argv[6]);
+        const ck::index_t N1 = std::stoi(argv[7]);
+
+        const ck::index_t K0 = std::stoi(argv[8]);
+        const ck::index_t K1 = std::stoi(argv[9]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[18]), std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21])};
+
+        scale = std::stof(argv[22]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M0, M1, N0, N1, K0, K1\n");
+        printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg18 to 21: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg22: scale\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
+    Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
+    Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
+    std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
+    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_ns_ks.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{scale};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 0>{},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_ms_ks_lengths,
+                                    a_ms_ks_strides,
+                                    b_ns_ks_lengths,
+                                    b_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 0>{},
+                                    std::array<std::vector<ck::index_t>, 0>{},
+                                    e_ms_ns_lengths,
+                                    e_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t M =
+        ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + +sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                  NumDimN,
+                                                                  NumDimK,
+                                                                  ADataType,
+                                                                  BDataType,
+                                                                  CShuffleDataType,
+                                                                  AccDataType,
+                                                                  AElementOp,
+                                                                  BElementOp,
+                                                                  PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
+                                       c_ms_ns_host_result(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/27_layernorm/CMakeLists.txt b/3rdparty/composable_kernel/example/27_layernorm/CMakeLists.txt
new file mode 100644
index 0000000..d96deae
--- /dev/null
+++ b/3rdparty/composable_kernel/example/27_layernorm/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_layernorm_blockwise layernorm_blockwise.cpp)
diff --git a/3rdparty/composable_kernel/example/27_layernorm/layernorm_blockwise.cpp b/3rdparty/composable_kernel/example/27_layernorm/layernorm_blockwise.cpp
new file mode 100644
index 0000000..147307d
--- /dev/null
+++ b/3rdparty/composable_kernel/example/27_layernorm/layernorm_blockwise.cpp
@@ -0,0 +1,139 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+
+using XDataType     = ck::half_t;
+using GammaDataType = ck::half_t;
+using BetaDataType  = ck::half_t;
+using YDataType     = ck::half_t;
+using AccDataType   = float;
+using PassThrough   = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+using DeviceInstance =
+    ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
+                                                          GammaDataType,
+                                                          BetaDataType,
+                                                          AccDataType,
+                                                          YDataType,
+                                                          PassThrough,
+                                                          Rank,
+                                                          NumReduceDim,
+                                                          256, // BlockSize
+                                                          8,   // ClusterM
+                                                          32,  // ClusterK
+                                                          1,   // SliceM
+                                                          8,   // SliceK
+                                                          1,   // SrcVecDim (0=M, 1=K)
+                                                          8,   // SrcScalarPerVector
+                                                          1,   // GammaVecDim (0=M, 1=K)
+                                                          8,   // GammaScalarPerVector
+                                                          1,   // BetaVecDim (0=M, 1=K)
+                                                          8,   // BetaScalarPerVector
+                                                          8>;  // OutScalarPerVector
+
+int main()
+{
+    bool time_kernel = false;
+
+    ck::index_t M      = 1024;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = N;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
+        using namespace ck::literals;
+
+        return HostTensorDescriptor({row, col}, {stride, 1_uz});
+    };
+
+    Tensor<XDataType> x(f_host_tensor_descriptor2d(M, N, Stride));
+    Tensor<GammaDataType> gamma(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta(f_host_tensor_descriptor1d(N, 1));
+    Tensor<YDataType> y(f_host_tensor_descriptor2d(M, N, Stride));
+
+    x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
+    gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
+    beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{0.0, 1.0});
+
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    auto device_instance = DeviceInstance{};
+    auto argument_ptr    = device_instance.MakeArgumentPointer(
+        {M, N},
+        std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
+        {0, 1},
+        {0, 1},
+        std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+        {1},
+        1e-4,
+        x_dev.GetDeviceBuffer(),
+        gamma_dev.GetDeviceBuffer(),
+        beta_dev.GetDeviceBuffer(),
+        y_dev.GetDeviceBuffer(),
+        nullptr,
+        nullptr,
+        PassThrough{});
+
+    if(!device_instance.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported" << std::endl;
+        return 1;
+    };
+
+    auto invoker_ptr = device_instance.MakeInvokerPointer();
+    invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+    {
+        Tensor<YDataType> host_y(f_host_tensor_descriptor2d(M, N, Stride));
+        using ReferenceInstance = ck::tensor_operation::host::ReferenceLayernorm<XDataType,
+                                                                                 GammaDataType,
+                                                                                 BetaDataType,
+                                                                                 YDataType,
+                                                                                 AccDataType,
+                                                                                 PassThrough,
+                                                                                 Rank,
+                                                                                 NumReduceDim>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, {M, N}, {1}, 1e-4);
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+
+        y_dev.FromDevice(y.mData.data());
+        pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results d1", 1e-3, 1e-3);
+    }
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/28_grouped_gemm_bias_e_permute/CMakeLists.txt b/3rdparty/composable_kernel/example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
new file mode 100644
index 0000000..44ab168
--- /dev/null
+++ b/3rdparty/composable_kernel/example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_grouped_gemm_bias_e_permute_xdl_fp16 grouped_gemm_bias_e_permute_xdl_fp16.cpp)
diff --git a/3rdparty/composable_kernel/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp b/3rdparty/composable_kernel/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
new file mode 100644
index 0000000..f8e6501
--- /dev/null
+++ b/3rdparty/composable_kernel/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
@@ -0,0 +1,466 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimM = 3;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 1;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //############################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           Gemm|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################################|        |        |        |      |      |        |         |           |      |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F16,   F16,     F32,      F16, DsDataType,   F16,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1, 32, 1, 4>,               8>;
+// clang-format on
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimM == 3 && NumDimN == 2 && NumDimK == 1, bool> = false>
+struct ReferenceContraction_M3_N2_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_ms_ks,
+                 const Tensor<BDataType>& b_ns_ks,
+                 Tensor<EDataType>& e_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_ms_ks_{a_ms_ks},
+              b_ns_ks_{b_ns_ks},
+              e_ms_ns_{e_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_ms_ks_;
+        const Tensor<BDataType>& b_ns_ks_;
+        Tensor<EDataType>& e_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_M3_N2_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto m0, auto m1, auto m2, auto n0, auto n1) {
+                const int K0 = arg.a_ms_ks_.mDesc.GetLengths()[3];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, m2, k0)));
+                    arg.b_element_op_(
+                        v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_ms_ns_(m0, m1, m2, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[4])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_ms_ks,
+                             const Tensor<BDataType>& b_ns_ks,
+                             Tensor<EDataType>& e_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{a_ms_ks, b_ns_ks, e_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_M3_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    std::size_t group_count = rand() % 16 + 1;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::ContractionDesc<1>> contraction_descs;
+    std::vector<const void*> p_a, p_b;
+    std::vector<std::array<const void*, 1>> p_ds;
+    std::vector<void*> p_c;
+
+    contraction_descs.reserve(group_count);
+
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        int M0 = 4 * (rand() % 4 + 1);
+        int M1 = 4 * (rand() % 4 + 1);
+        int M2 = 256;
+
+        int N0 = 4 * (rand() % 4 + 1);
+        int N1 = 128;
+
+        int K0 = 64 * (rand() % 4 + 1);
+
+        // A[M0, M1, M2, K0]
+        std::vector<ck::index_t> a_ms_ks_lengths{M0, M1, M2, K0};
+        std::vector<ck::index_t> a_ms_ks_strides{M1 * M2 * K0, M2 * K0, K0, 1};
+        // B[N0, N1, K0]
+        std::vector<ck::index_t> b_ns_ks_lengths{N0, N1, K0};
+        std::vector<ck::index_t> b_ns_ks_strides{N1 * K0, K0, 1};
+#if 0
+        // D[M0, N0, M1, N1, M2]
+        std::vector<ck::index_t> d_ms_ns_lengths{M0, M1, M2, N0, N1};
+        std::vector<ck::index_t> d_ms_ns_strides{0, 0, 0, N1, 1};
+        // E[M0, N0, M1, N1, M2]
+        std::vector<ck::index_t> e_ms_ns_lengths{M0, M1, M2, N0, N1};
+        std::vector<ck::index_t> e_ms_ns_strides{N0 * M1 * N1 * M2, N1 * M2, 1, M1 * N1 * M2, M2};
+#else
+        // D[M0, N0, M1, N1, M2]
+        std::vector<ck::index_t> d_ms_ns_lengths{M0, M1, M2, N0, N1};
+        std::vector<ck::index_t> d_ms_ns_strides{0, 0, 0, N1, 1};
+        // E[M0, N0, M1, N1, M2]
+        std::vector<ck::index_t> e_ms_ns_lengths{M0, M1, M2, N0, N1};
+        std::vector<ck::index_t> e_ms_ns_strides{M1 * M2 * N0 * N1, M2 * N0 * N1, N0 * N1, N1, 1};
+#endif
+
+        contraction_descs.push_back(
+            ck::tensor_operation::device::ContractionDesc<1>{a_ms_ks_lengths,
+                                                             a_ms_ks_strides,
+                                                             b_ns_ks_lengths,
+                                                             b_ns_ks_strides,
+                                                             {d_ms_ns_lengths},
+                                                             {d_ms_ns_strides},
+                                                             e_ms_ns_lengths,
+                                                             e_ms_ns_strides});
+    }
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<BDataType>> b_tensors;
+    std::vector<Tensor<DDataType>> d_tensors;
+    std::vector<Tensor<EDataType>> e_device_tensors;
+
+    a_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    d_tensors.reserve(group_count);
+    e_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, d_tensors_device,
+        e_tensors_device;
+
+    a_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    d_tensors_device.reserve(group_count);
+    e_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(std::size_t i = 0; i < contraction_descs.size(); i++)
+    {
+        const auto a_ms_ks_lengths = contraction_descs[i].a_ms_ks_lengths;
+        const auto a_ms_ks_strides = contraction_descs[i].a_ms_ks_strides;
+
+        const auto b_ns_ks_lengths = contraction_descs[i].b_ns_ks_lengths;
+        const auto b_ns_ks_strides = contraction_descs[i].b_ns_ks_strides;
+
+        const auto d_ms_ns_lengths = contraction_descs[i].ds_ms_ns_lengths[0];
+        const auto d_ms_ns_strides = contraction_descs[i].ds_ms_ns_strides[0];
+
+        const auto e_ms_ns_lengths = contraction_descs[i].e_ms_ns_lengths;
+        const auto e_ms_ns_strides = contraction_descs[i].e_ms_ns_strides;
+
+        Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
+        Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
+        Tensor<DDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
+        Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        ck::index_t M_ =
+            ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+        ck::index_t N_ = ck::accumulate_n<ck::index_t>(
+            e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+        ck::index_t K_ = ck::accumulate_n<ck::index_t>(
+            a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+        a_tensors.push_back(a_ms_ks);
+        b_tensors.push_back(b_ns_ks);
+        d_tensors.push_back(d_ms_ns);
+
+        // e_host_tensors.push_back(e_ms_ns_host_result);
+        e_device_tensors.push_back(e_ms_ns_device_result);
+
+        flop += std::size_t(2) * M_ * K_ * N_;
+
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * e_device_tensors[i].mDesc.GetElementSize();
+
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_n_k: " << b_tensors[i].mDesc << " c_m_n: " << e_device_tensors[i].mDesc
+                  << std::endl;
+
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            d_tensors[i].GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+            break;
+        case 2:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            d_tensors[i].GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_1<ADataType>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_1<BDataType>{});
+            d_tensors[i].GenerateTensorValue(GeneratorTensor_1<DDataType>{});
+        }
+    }
+
+    for(std::size_t i = 0; i < contraction_descs.size(); i++)
+    {
+        a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpaceSize()));
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpaceSize()));
+        d_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(DDataType) * d_tensors[i].mDesc.GetElementSpaceSize()));
+        e_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(EDataType) * e_device_tensors[i].mDesc.GetElementSpaceSize()));
+
+        a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
+        d_tensors_device[i]->ToDevice(d_tensors[i].mData.data());
+
+        p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
+        p_b.push_back(b_tensors_device[i]->GetDeviceBuffer());
+        p_ds.push_back({d_tensors_device[i]->GetDeviceBuffer()});
+        p_c.push_back(e_tensors_device[i]->GetDeviceBuffer());
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceOpInstanceKKNN{};
+    auto invoker = gemm.MakeInvoker();
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_a, p_b, p_ds, p_c, contraction_descs, a_element_op, b_element_op, cde_element_op);
+
+    DeviceMem contraction_desc_workspace(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, contraction_desc_workspace.GetDeviceBuffer());
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < group_count; i++)
+        {
+            const auto e_ms_ns_lengths = contraction_descs[i].e_ms_ns_lengths;
+            const auto e_ms_ns_strides = contraction_descs[i].e_ms_ns_strides;
+
+            Tensor<EDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+            Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+            e_tensors_device[i]->FromDevice(e_device_tensors[i].mData.data());
+
+            using ReferenceOpInstance = ReferenceContraction_M3_N2_K1<NumDimM,
+                                                                      NumDimN,
+                                                                      NumDimK,
+                                                                      ADataType,
+                                                                      BDataType,
+                                                                      CShuffleDataType,
+                                                                      AccDataType,
+                                                                      AElementOp,
+                                                                      BElementOp,
+                                                                      PassThrough>;
+
+            auto ref_gemm    = ReferenceOpInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      c_ms_ns_host_result,
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      PassThrough{});
+
+            ref_invoker.Run(ref_argument);
+
+            for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
+            {
+                for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
+                {
+                    for(size_t m2 = 0; m2 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++m2)
+                    {
+                        for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n0)
+                        {
+                            for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[4]; ++n1)
+                            {
+                                cde_element_op(e_ms_ns_host_result(m0, m1, m2, n0, n1),
+                                               c_ms_ns_host_result(m0, m1, m2, n0, n1),
+                                               d_tensors[i](m0, m1, m2, n0, n1));
+                            }
+                        }
+                    }
+                }
+            }
+
+            pass &= ck::utils::check_err(e_device_tensors[i], e_ms_ns_host_result);
+        }
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/29_batched_gemm_bias_e_permute/CMakeLists.txt b/3rdparty/composable_kernel/example/29_batched_gemm_bias_e_permute/CMakeLists.txt
new file mode 100644
index 0000000..40470f2
--- /dev/null
+++ b/3rdparty/composable_kernel/example/29_batched_gemm_bias_e_permute/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
diff --git a/3rdparty/composable_kernel/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp b/3rdparty/composable_kernel/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp
new file mode 100644
index 0000000..25d815b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp
@@ -0,0 +1,397 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 1;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //############################################| NumDimG| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           Gemm|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################################|        |        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################################|        |        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################################|        |        |        |        |      |      |        |         |           |      |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK,   F16,   F16,     F32,      F16, DsDataType,   F16,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1, 32, 1, 4>,               8>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimG,
+          ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimG == 2 && NumDimM == 2 && NumDimN == 2 && NumDimK == 1, bool> =
+              false>
+struct ReferenceContraction_G2_M2_N2_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_gs_ms_ks,
+                 const Tensor<BDataType>& b_gs_ns_ks,
+                 Tensor<EDataType>& e_gs_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_gs_ms_ks_{a_gs_ms_ks},
+              b_gs_ns_ks_{b_gs_ns_ks},
+              e_gs_ms_ns_{e_gs_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_gs_ms_ks_;
+        const Tensor<BDataType>& b_gs_ns_ks_;
+        Tensor<EDataType>& e_gs_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_G2_M2_N2_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto g0, auto g1, auto m0, auto m1, auto n0, auto n1) {
+                const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a,
+                        ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, g1, m0, m1, k0)));
+                    arg.b_element_op_(
+                        v_b,
+                        ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, g1, n0, n1, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_gs_ms_ns_(g0, g1, m0, m1, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
+                             const Tensor<BDataType>& b_gs_ns_ks,
+                             Tensor<EDataType>& e_gs_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_G2_M2_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::index_t G0 = 1;
+    ck::index_t G1 = 2;
+
+    ck::index_t M0 = 4;
+    ck::index_t M1 = 256;
+
+    ck::index_t N0 = 16;
+    ck::index_t N1 = 128;
+
+    ck::index_t K0 = 64;
+
+    // A[G0, G1, M0, M1, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[G0, G1, N0, N1, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
+
+    // D[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
+    // E[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
+    Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_device_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
+    std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
+    std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) *
+                           e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_gs_ms_ks_lengths,
+                                    a_gs_ms_ks_strides,
+                                    b_gs_ns_ks_lengths,
+                                    b_gs_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                    e_gs_ms_ns_lengths,
+                                    e_gs_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t G =
+        ck::accumulate_n<ck::index_t>(e_gs_ms_ns_lengths.begin(), NumDimG, 1, std::multiplies<>{});
+
+    ck::index_t M = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimG, NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop      = std::size_t(2) * G * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * G * M * K + sizeof(BDataType) * G * K * N +
+                            sizeof(DDataType) * G * M * N + sizeof(EDataType) * G * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+        using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CShuffleDataType,
+                                                                     AccDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_gs_ms_ks, b_gs_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t g0 = 0; g0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[0]; ++g0)
+        {
+            for(size_t g1 = 0; g1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[1]; ++g1)
+            {
+                for(size_t m0 = 0; m0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[2]; ++m0)
+                {
+                    for(size_t m1 = 0; m1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[3]; ++m1)
+                    {
+                        for(size_t n0 = 0; n0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[4]; ++n0)
+                        {
+                            for(size_t n1 = 0; n1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[5];
+                                ++n1)
+                            {
+                                cde_element_op(e_gs_ms_ns_host_result(g0, g1, m0, m1, n0, n1),
+                                               c_ms_ns_host_result(g0, g1, m0, m1, n0, n1),
+                                               d_gs_ms_ns(g0, g1, m0, m1, n0, n1));
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_gs_ms_ns_device_result, e_gs_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
new file mode 100644
index 0000000..61b2b2f
--- /dev/null
+++ b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
@@ -0,0 +1,22 @@
+add_custom_target(example_grouped_conv_fwd_multiple_d)
+
+add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
+add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
+add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
+add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
+
+add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)
+add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
+add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
+add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
+
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int4 grouped_conv_fwd_bias_relu_add_xdl_int4.cpp)
+
+  add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4)
+endif() # USE_BITINT_EXTENSION_INT4
+
+
+add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
+
+add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
diff --git a/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/README.md b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/README.md
new file mode 100644
index 0000000..739a042
--- /dev/null
+++ b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/README.md
@@ -0,0 +1,30 @@
+Command
+```bash
+arg1: verification (0=no, 1=yes)
+arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+arg3: time kernel (0=no, 1=yes)
+Following arguments (depending on number of spatial dims):
+ Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
+ G, N, K, C,
+ <filter spatial dimensions>, (ie Y, X for 2D)
+ <input image spatial dimensions>, (ie Hi, Wi for 2D)
+ <strides>, (ie Sy, Sx for 2D)
+ <dilations>, (ie Dy, Dx for 2D)
+ <left padding>, (ie LeftPy, LeftPx for 2D)
+ <right padding>, (ie RightPy, RightPx for 2D)
+
+./bin/example_grouped_conv_fwd_bias_relu_add_xdl_fp16 1 1 1
+```
+
+Result (MI100)
+```
+in: dim 5, lengths {1, 128, 192, 71, 71}, strides {192, 967872, 1, 13632, 192}
+wei: dim 5, lengths {1, 256, 192, 3, 3}, strides {442368, 1728, 1, 576, 192}
+bias: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 0, 1, 0, 0}
+residual: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 0, 1, 0, 0}
+out: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 331776, 1, 9216, 256}
+launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 1.55981 ms, 94.0927 TFlops, 213.868 GB/s, DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<256, 128, 256, 16, Default>
+```
diff --git a/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/common.hpp b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/common.hpp
new file mode 100644
index 0000000..d6d6dd6
--- /dev/null
+++ b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/common.hpp
@@ -0,0 +1,355 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+#include <iostream>
+#include <string>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using BF16 = ck::bhalf_t;
+using FP16 = ck::half_t;
+using FP32 = float;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using I4 = ck::int4_t;
+#endif
+using I8  = std::int8_t;
+using I32 = std::int32_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <typename InputLay, typename WeightLay, typename OutputLay>
+struct CommonLayoutSetting
+{
+    using InputLayout  = InputLay;
+    using WeightLayout = WeightLay;
+    using OutputLayout = OutputLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct CommonLayoutSettingSelector;
+
+namespace ctl = ck::tensor_layout::convolution;
+
+template <>
+struct CommonLayoutSettingSelector<1> final
+    : CommonLayoutSetting<ctl::G_NW_C, ctl::G_K_X_C, ctl::G_NW_K>
+{
+};
+
+template <>
+struct CommonLayoutSettingSelector<2> final
+    : CommonLayoutSetting<ctl::G_NHW_C, ctl::G_K_YX_C, ctl::G_NHW_K>
+{
+};
+
+template <>
+struct CommonLayoutSettingSelector<3> final
+    : CommonLayoutSetting<ctl::G_NDHW_C, ctl::G_K_ZYX_C, ctl::G_NDHW_K>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using InputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::InputLayout;
+
+template <ck::index_t NDimSpatial>
+using WeightLayout = typename CommonLayoutSettingSelector<NDimSpatial>::WeightLayout;
+
+template <ck::index_t NDimSpatial>
+using OutputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::OutputLayout;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+};
+
+#define DefaultConvParam                                                       \
+    ck::utils::conv::ConvParam                                                 \
+    {                                                                          \
+        2, 32, 2, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, { 1, 1 } \
+    }
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ExecutionConfig& config,
+                           ck::utils::conv::ConvParam& conv_param)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        conv_param                        = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
+
+inline HostTensorDescriptor make_input_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.N_, conv_param.C_, conv_param.input_spatial_lengths_[0]},
+            {
+                conv_param.C_,                                                        // g
+                conv_param.input_spatial_lengths_[0] * conv_param.G_ * conv_param.C_, // n
+                1,                                                                    // c
+                conv_param.G_ * conv_param.C_                                         // wi
+            });
+
+    case 2:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.C_,
+             conv_param.input_spatial_lengths_[0],
+             conv_param.input_spatial_lengths_[1]},
+            {
+                conv_param.C_, // g
+                conv_param.input_spatial_lengths_[0] * conv_param.input_spatial_lengths_[1] *
+                    conv_param.G_ * conv_param.C_,                                    // n
+                1,                                                                    // c
+                conv_param.input_spatial_lengths_[1] * conv_param.G_ * conv_param.C_, // hi
+                conv_param.G_ * conv_param.C_                                         // wi
+            });
+
+    case 3:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.C_,
+             conv_param.input_spatial_lengths_[0],
+             conv_param.input_spatial_lengths_[1],
+             conv_param.input_spatial_lengths_[2]},
+            {
+                conv_param.C_, // g
+                conv_param.input_spatial_lengths_[0] * conv_param.input_spatial_lengths_[1] *
+                    conv_param.input_spatial_lengths_[2] * conv_param.G_ * conv_param.C_, // n
+                1,                                                                        // c
+                conv_param.input_spatial_lengths_[1] * conv_param.input_spatial_lengths_[2] *
+                    conv_param.G_ * conv_param.C_,                                    // di
+                conv_param.input_spatial_lengths_[2] * conv_param.G_ * conv_param.C_, // hi
+                conv_param.G_ * conv_param.C_                                         // wi
+            });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
+
+inline HostTensorDescriptor make_weight_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.K_, conv_param.C_, conv_param.filter_spatial_lengths_[0]},
+            {
+                conv_param.K_ * conv_param.filter_spatial_lengths_[0] * conv_param.C_, // g
+                conv_param.filter_spatial_lengths_[0] * conv_param.C_,                 // k
+                1,                                                                     // c
+                conv_param.C_                                                          // x
+            });
+    case 2:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.K_,
+             conv_param.C_,
+             conv_param.filter_spatial_lengths_[0],
+             conv_param.filter_spatial_lengths_[1]},
+            {
+                conv_param.K_ * conv_param.filter_spatial_lengths_[0] *
+                    conv_param.filter_spatial_lengths_[1] * conv_param.C_, // g
+                conv_param.filter_spatial_lengths_[0] * conv_param.filter_spatial_lengths_[1] *
+                    conv_param.C_,                                     // k
+                1,                                                     // c
+                conv_param.filter_spatial_lengths_[1] * conv_param.C_, // y
+                conv_param.C_                                          // x
+            });
+    case 3:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.K_,
+             conv_param.C_,
+             conv_param.filter_spatial_lengths_[0],
+             conv_param.filter_spatial_lengths_[1],
+             conv_param.filter_spatial_lengths_[2]},
+            {
+                conv_param.K_ * conv_param.filter_spatial_lengths_[0] *
+                    conv_param.filter_spatial_lengths_[1] * conv_param.filter_spatial_lengths_[2] *
+                    conv_param.C_, // g
+                conv_param.filter_spatial_lengths_[0] * conv_param.filter_spatial_lengths_[1] *
+                    conv_param.filter_spatial_lengths_[2] * conv_param.C_, // k
+                1,                                                         // c
+                conv_param.filter_spatial_lengths_[1] * conv_param.filter_spatial_lengths_[2] *
+                    conv_param.C_,                                     // z
+                conv_param.filter_spatial_lengths_[2] * conv_param.C_, // y
+                conv_param.C_                                          // x
+            });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
+
+inline HostTensorDescriptor make_bias_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
+            {
+                conv_param.K_, // g
+                0,             // k
+                1,             // c
+                0              // x
+            });
+    case 2:
+        return HostTensorDescriptor({conv_param.G_,
+                                     conv_param.N_,
+                                     conv_param.K_,
+                                     conv_param.output_spatial_lengths_[0],
+                                     conv_param.output_spatial_lengths_[1]},
+                                    {
+                                        conv_param.K_, // g
+                                        0,             // n
+                                        1,             // k
+                                        0,             // ho
+                                        0              // wo
+                                    });
+    case 3:
+        return HostTensorDescriptor({conv_param.G_,
+                                     conv_param.N_,
+                                     conv_param.K_,
+                                     conv_param.output_spatial_lengths_[0],
+                                     conv_param.output_spatial_lengths_[1],
+                                     conv_param.output_spatial_lengths_[2]},
+                                    {
+                                        conv_param.K_, // g
+                                        0,             // n
+                                        1,             // k
+                                        0,             // z
+                                        0,             // y
+                                        0              // x
+                                    });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
+
+inline HostTensorDescriptor make_output_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
+            {
+                conv_param.K_,                                                         // g
+                conv_param.output_spatial_lengths_[0] * conv_param.G_ * conv_param.K_, // n
+                1,                                                                     // k
+                conv_param.G_ * conv_param.K_                                          // wo
+            });
+    case 2:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.K_,
+             conv_param.output_spatial_lengths_[0],
+             conv_param.output_spatial_lengths_[1]},
+            {
+                conv_param.K_, // g
+                conv_param.output_spatial_lengths_[0] * conv_param.output_spatial_lengths_[1] *
+                    conv_param.G_ * conv_param.K_,                                     // n
+                1,                                                                     // k
+                conv_param.output_spatial_lengths_[1] * conv_param.G_ * conv_param.K_, // ho
+                conv_param.G_ * conv_param.K_                                          // wo
+            });
+
+    case 3:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.K_,
+             conv_param.output_spatial_lengths_[0],
+             conv_param.output_spatial_lengths_[1],
+             conv_param.output_spatial_lengths_[2]},
+            {
+                conv_param.K_, // g
+                conv_param.output_spatial_lengths_[0] * conv_param.output_spatial_lengths_[1] *
+                    conv_param.output_spatial_lengths_[2] * conv_param.G_ * conv_param.K_, // n
+                1,                                                                         // k
+                conv_param.output_spatial_lengths_[1] * conv_param.output_spatial_lengths_[2] *
+                    conv_param.G_ * conv_param.K_,                                     // do
+                conv_param.output_spatial_lengths_[2] * conv_param.G_ * conv_param.K_, // ho
+                conv_param.G_ * conv_param.K_                                          // wo
+            });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
diff --git a/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp
new file mode 100644
index 0000000..ee300d0
--- /dev/null
+++ b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+// kernel data types
+using InKernelDataType       = BF16;
+using WeiKernelDataType      = BF16;
+using AccDataType            = FP32;
+using CShuffleDataType       = FP32;
+using BiasKernelDataType     = BF16;
+using ResidualKernelDataType = BF16;
+using OutKernelDataType      = BF16;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp
new file mode 100644
index 0000000..5a9df0b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+// kernel data types
+using InKernelDataType       = FP16;
+using WeiKernelDataType      = FP16;
+using AccDataType            = FP32;
+using CShuffleDataType       = FP16;
+using BiasKernelDataType     = FP16;
+using ResidualKernelDataType = FP16;
+using OutKernelDataType      = FP16;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp
new file mode 100644
index 0000000..c2906cc
--- /dev/null
+++ b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+// kernel data types
+using InKernelDataType       = FP32;
+using WeiKernelDataType      = FP32;
+using AccDataType            = FP32;
+using CShuffleDataType       = FP32;
+using BiasKernelDataType     = FP32;
+using ResidualKernelDataType = FP32;
+using OutKernelDataType      = FP32;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp
new file mode 100644
index 0000000..3d5a243
--- /dev/null
+++ b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#include "common.hpp"
+
+// kernel data types
+using InKernelDataType       = I8;
+using WeiKernelDataType      = I8;
+using AccDataType            = I32;
+using CShuffleDataType       = I8;
+using BiasKernelDataType     = I8;
+using ResidualKernelDataType = I8;
+using OutKernelDataType      = I8;
+
+// tensor data types
+using InUserDataType  = I4;
+using WeiUserDataType = I4;
+using OutUserDataType = I4;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#define BUILD_INT4_EXAMPLE
+#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp
new file mode 100644
index 0000000..eaf680f
--- /dev/null
+++ b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+// kernel data types
+using InKernelDataType       = I8;
+using WeiKernelDataType      = I8;
+using AccDataType            = I32;
+using CShuffleDataType       = I8;
+using BiasKernelDataType     = I8;
+using ResidualKernelDataType = I8;
+using OutKernelDataType      = I8;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp
new file mode 100644
index 0000000..6de1daa
--- /dev/null
+++ b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+// kernel data types
+using InKernelDataType  = FP16;
+using WeiKernelDataType = FP16;
+using AccDataType       = FP32;
+using CShuffleDataType  = FP16;
+using OutKernelDataType = FP16;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = PassThrough;
+
+#include "run_grouped_conv_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc
new file mode 100644
index 0000000..4561156
--- /dev/null
+++ b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc
@@ -0,0 +1,288 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+template <typename BiasLay, typename ResidualLay>
+struct LayoutSetting
+{
+    using BiasLayout     = BiasLay;
+    using ResidualLayout = ResidualLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct LayoutSettingSelector;
+
+template <>
+struct LayoutSettingSelector<1> final : LayoutSetting<ctl::G_K, ctl::G_NW_K>
+{
+};
+
+template <>
+struct LayoutSettingSelector<2> final : LayoutSetting<ctl::G_K, ctl::G_NHW_K>
+{
+};
+
+template <>
+struct LayoutSettingSelector<3> final : LayoutSetting<ctl::G_K, ctl::G_NDHW_K>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using BiasLayout = typename LayoutSettingSelector<NDimSpatial>::BiasLayout;
+
+template <ck::index_t NDimSpatial>
+using ResidualLayout = typename LayoutSettingSelector<NDimSpatial>::ResidualLayout;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InputLayout<NDimSpatial>,
+        WeightLayout<NDimSpatial>,
+        ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>,
+        OutputLayout<NDimSpatial>,
+        InKernelDataType,
+        WeiKernelDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<BiasKernelDataType, ResidualKernelDataType>,
+        OutKernelDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        16,          // KPerBlock
+        4,           // AK1
+        4,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        4,           // ABlockTransferSrcScalarPerVector
+        4,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        4,           // BBlockTransferSrcScalarPerVector
+        4,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 16, 1, 16>,
+        4>;
+
+template <ck::index_t NDimSpatial>
+using HostConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                         InUserDataType,
+                                                                         WeiUserDataType,
+                                                                         CShuffleDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         PassThrough>;
+
+template <ck::index_t NDimSpatial>
+bool run_grouped_conv_fwd_bias_relu_add(const ExecutionConfig& config,
+                                        const ck::utils::conv::ConvParam& conv_param)
+{
+    static_assert(1 <= NDimSpatial && NDimSpatial <= 3, "Unsupported NDimSpatial");
+
+    const auto in_g_n_c_wis_desc   = make_input_descriptor(conv_param);
+    const auto wei_g_k_c_xs_desc   = make_weight_descriptor(conv_param);
+    const auto bias_g_n_k_wos_desc = make_bias_descriptor(conv_param);
+    const auto out_g_n_k_wos_desc  = make_output_descriptor(conv_param);
+
+    Tensor<InUserDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiUserDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutUserDataType> bias(bias_g_n_k_wos_desc);
+    Tensor<OutUserDataType> residual(bias_g_n_k_wos_desc);
+    Tensor<OutUserDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutKernelDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "bias: " << bias.mDesc << std::endl;
+    std::cout << "residual: " << residual.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InUserDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiUserDataType>{-5, 5});
+        bias.GenerateTensorValue(GeneratorTensor_2<OutUserDataType>{-5, 5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InUserDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiUserDataType>{-0.5, 0.5});
+        bias.GenerateTensorValue(GeneratorTensor_3<OutUserDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InKernelDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiKernelDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(OutKernelDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem residual_device_buf(sizeof(OutKernelDataType) * residual.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutKernelDataType) * out_device.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<InKernelDataType> in_converted(in);
+    const Tensor<WeiKernelDataType> wei_converted(wei);
+    const Tensor<OutKernelDataType> bias_converted(bias);
+    const Tensor<OutKernelDataType> residual_converted(residual);
+
+    in_device_buf.ToDevice(in_converted.mData.data());
+    wei_device_buf.ToDevice(wei_converted.mData.data());
+    bias_device_buf.ToDevice(bias_converted.mData.data());
+    residual_device_buf.ToDevice(residual_converted.mData.data());
+#else
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+    residual_device_buf.ToDevice(residual.mData.data());
+#endif
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(bias_g_n_k_wos_desc.GetLengths(), d0_g_n_k_wos_lengths);
+    copy(bias_g_n_k_wos_desc.GetStrides(), d0_g_n_k_wos_strides);
+    copy(bias_g_n_k_wos_desc.GetLengths(), d1_g_n_k_wos_lengths);
+    copy(bias_g_n_k_wos_desc.GetStrides(), d1_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv    = DeviceConvFwdInstance<NDimSpatial>{};
+    auto invoker = conv.MakeInvoker();
+    auto argument =
+        conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                          wei_device_buf.GetDeviceBuffer(),
+                          std::array<const void*, 2>{bias_device_buf.GetDeviceBuffer(),
+                                                     residual_device_buf.GetDeviceBuffer()},
+                          out_device_buf.GetDeviceBuffer(),
+                          a_g_n_c_wis_lengths,
+                          a_g_n_c_wis_strides,
+                          b_g_k_c_xs_lengths,
+                          b_g_k_c_xs_strides,
+                          std::array<std::array<ck::index_t, NDimSpatial + 3>, 2>{
+                              {d0_g_n_k_wos_lengths, d1_g_n_k_wos_lengths}},
+                          std::array<std::array<ck::index_t, NDimSpatial + 3>, 2>{
+                              {d0_g_n_k_wos_strides, d1_g_n_k_wos_strides}},
+                          e_g_n_k_wos_lengths,
+                          e_g_n_k_wos_strides,
+                          conv_filter_strides,
+                          conv_filter_dilations,
+                          input_left_pads,
+                          input_right_pads,
+                          InElementOp{},
+                          WeiElementOp{},
+                          OutElementOp{});
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InUserDataType, WeiUserDataType, OutUserDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        Tensor<CShuffleDataType> c_host(out_g_n_k_wos_desc);
+
+        auto ref_conv     = HostConvFwdInstance<NDimSpatial>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  c_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  InElementOp{},
+                                                  WeiElementOp{},
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        out_host.ForEach([&](auto&, auto idx) {
+            OutElementOp{}(out_host(idx), c_host(idx), bias(idx), residual(idx));
+        });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<OutUserDataType> out_device_converted(out_device);
+
+        return ck::utils::check_err(
+            out_device_converted, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+#else
+        return ck::utils::check_err(
+            out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+#endif
+    }
+
+    return true;
+}
+
+bool run_grouped_conv_fwd_bias_relu_add_example(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return false;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return run_grouped_conv_fwd_bias_relu_add<1>(config, conv_param);
+    case 2: return run_grouped_conv_fwd_bias_relu_add<2>(config, conv_param);
+    case 3: return run_grouped_conv_fwd_bias_relu_add<3>(config, conv_param);
+    }
+
+    return false;
+}
diff --git a/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc
new file mode 100644
index 0000000..d087c31
--- /dev/null
+++ b/3rdparty/composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc
@@ -0,0 +1,223 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+template <ck::index_t NDimSpatial>
+using DeviceConvFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InputLayout<NDimSpatial>,
+        WeightLayout<NDimSpatial>,
+        ck::Tuple<>,
+        OutputLayout<NDimSpatial>,
+        InKernelDataType,
+        WeiKernelDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutKernelDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        16,          // KPerBlock
+        4,           // AK1
+        4,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        4,           // ABlockTransferSrcScalarPerVector
+        4,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        4,           // BBlockTransferSrcScalarPerVector
+        4,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 16, 1, 16>,
+        4>;
+
+template <ck::index_t NDimSpatial>
+using HostConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                         InUserDataType,
+                                                                         WeiUserDataType,
+                                                                         CShuffleDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         PassThrough>;
+
+template <ck::index_t NDimSpatial>
+bool run_grouped_conv_fwd(const ExecutionConfig& config,
+                          const ck::utils::conv::ConvParam& conv_param)
+{
+    static_assert(1 <= NDimSpatial && NDimSpatial <= 3, "Unsupported NDimSpatial");
+
+    const auto in_g_n_c_wis_desc  = make_input_descriptor(conv_param);
+    const auto wei_g_k_c_xs_desc  = make_weight_descriptor(conv_param);
+    const auto out_g_n_k_wos_desc = make_output_descriptor(conv_param);
+
+    Tensor<InUserDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiUserDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutUserDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutKernelDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InUserDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiUserDataType>{-5, 5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InUserDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiUserDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InKernelDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiKernelDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutKernelDataType) * out_device.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<InKernelDataType> in_converted(in);
+    const Tensor<WeiKernelDataType> wei_converted(wei);
+
+    in_device_buf.ToDevice(in_converted.mData.data());
+    wei_device_buf.ToDevice(wei_converted.mData.data());
+#else
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+#endif
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvFwdInstance<NDimSpatial>{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      std::array<const void*, 0>{},
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      InElementOp{},
+                                      WeiElementOp{},
+                                      OutElementOp{});
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InUserDataType, WeiUserDataType, OutUserDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_conv     = HostConvFwdInstance<NDimSpatial>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  out_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  InElementOp{},
+                                                  WeiElementOp{},
+                                                  OutElementOp{});
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<OutUserDataType> out_device_converted(out_device);
+
+        return ck::utils::check_err(
+            out_device_converted.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+#else
+        return ck::utils::check_err(
+            out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+#endif
+    }
+
+    return true;
+}
+
+bool run_grouped_conv_fwd_example(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return false;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return run_grouped_conv_fwd<1>(config, conv_param);
+    case 2: return run_grouped_conv_fwd<2>(config, conv_param);
+    case 3: return run_grouped_conv_fwd<3>(config, conv_param);
+    }
+
+    return false;
+}
diff --git a/3rdparty/composable_kernel/example/31_batched_gemm_gemm/CMakeLists.txt b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/CMakeLists.txt
new file mode 100644
index 0000000..d792482
--- /dev/null
+++ b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/CMakeLists.txt
@@ -0,0 +1,8 @@
+add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp)
+add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
+add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp)
+add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
+
+if(USE_BITINT_EXTENSION_INT4)
+add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp)
+endif(USE_BITINT_EXTENSION_INT4)
diff --git a/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp
new file mode 100644
index 0000000..74e0e07
--- /dev/null
+++ b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
+                                              |------------|
+                                                   Gemm0
+                                              |---------------------|
+                                                       Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = BF16;
+using B0DataType       = BF16;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = BF16;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = PassThrough;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmDefault,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    32,          // KPerBlock
+    128,         // Gemm1NPerBlock
+    32,          // Gemm1KPerBlock
+    8,           // AK1
+    8,           // BK1
+    2,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    4,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<8, 32, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    2,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return run_batched_gemm_gemm_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp
new file mode 100644
index 0000000..d5fadb8
--- /dev/null
+++ b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
+                                              |------------|
+                                                   Gemm0
+                                              |---------------------|
+                                                       Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = PassThrough;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmDefault,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    32,          // KPerBlock
+    128,         // Gemm1NPerBlock
+    32,          // Gemm1KPerBlock
+    8,           // AK1
+    8,           // BK1
+    2,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    4,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<8, 32, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    2,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return run_batched_gemm_gemm_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp
new file mode 100644
index 0000000..0dd4e09
--- /dev/null
+++ b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp
@@ -0,0 +1,135 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
+                                              |------------|
+                                                   Gemm0
+                                              |---------------------|
+                                                       Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F32;
+using B0DataType       = F32;
+using B1DataType       = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F32;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = PassThrough;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmDefault,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    16,          // KPerBlock
+    128,         // Gemm1NPerBlock
+    16,          // Gemm1KPerBlock
+    4,           // AK1
+    4,           // BK1
+    1,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    4,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    4,
+    4,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    4,
+    4,
+    true,
+    S<8, 32, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    1,
+    false,
+    1,               // CShuffleMXdlPerWavePerShuffle
+    2,               // CShuffleNXdlPerWavePerShuffle
+    S<1, 16, 1, 16>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    4>;              // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return run_batched_gemm_gemm_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp
new file mode 100644
index 0000000..1fd9362
--- /dev/null
+++ b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp
@@ -0,0 +1,146 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
+                                              |------------|
+                                                   Gemm0
+                                              |---------------------|
+                                                       Gemm1
+*/
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = ck::int4_t;
+using B0DataType       = ck::int4_t;
+using B1DataType       = ck::int4_t;
+using KernelADataType  = int8_t;
+using KernelB0DataType = int8_t;
+using KernelB1DataType = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using CDataType        = ck::int4_t;
+using KernelCDataType  = int8_t;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = PassThrough;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    KernelADataType,
+    KernelB0DataType,
+    KernelB1DataType,
+    KernelCDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmDefault,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    64,          // KPerBlock
+    128,         // Gemm1NPerBlock
+    64,          // Gemm1KPerBlock
+    16,          // AK1
+    16,          // BK1
+    4,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    4,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    16,
+    16,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    16,
+    16,
+    true,
+    S<8, 32, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    4,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#define BUILD_INT4_EXAMPLE
+#include "run_batched_gemm_gemm_example.inc"
+
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+
+int main(int argc, char* argv[]) { return run_batched_gemm_gemm_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp
new file mode 100644
index 0000000..15d98ab
--- /dev/null
+++ b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp
@@ -0,0 +1,133 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
+                                              |------------|
+                                                   Gemm0
+                                              |---------------------|
+                                                       Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = int8_t;
+using B0DataType       = int8_t;
+using B1DataType       = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using CDataType        = int8_t;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = PassThrough;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmDefault,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    64,          // KPerBlock
+    128,         // Gemm1NPerBlock
+    64,          // Gemm1KPerBlock
+    16,          // AK1
+    16,          // BK1
+    4,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    4,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    16,
+    16,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    16,
+    16,
+    true,
+    S<8, 32, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    4,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return run_batched_gemm_gemm_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
new file mode 100644
index 0000000..7e5f161
--- /dev/null
+++ b/3rdparty/composable_kernel/example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
@@ -0,0 +1,277 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_batched_gemm_gemm_example(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M             = 1024;
+    ck::index_t N             = 1024;
+    ck::index_t K             = 64;
+    ck::index_t O             = 128;
+    ck::index_t BatchCount    = 4;
+    ck::index_t StrideA       = -1;
+    ck::index_t StrideB0      = -1;
+    ck::index_t StrideB1      = -1;
+    ck::index_t StrideC       = -1;
+    ck::index_t BatchStrideA  = -1;
+    ck::index_t BatchStrideB0 = -1;
+    ck::index_t BatchStrideB1 = -1;
+    ck::index_t BatchStrideC  = -1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 9)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+    }
+    else if(argc == 17)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+
+        StrideA  = std::stoi(argv[9]);
+        StrideB0 = std::stoi(argv[10]);
+        StrideB1 = std::stoi(argv[11]);
+        StrideC  = std::stoi(argv[12]);
+
+        BatchStrideA  = std::stoi(argv[13]);
+        BatchStrideB0 = std::stoi(argv[14]);
+        BatchStrideB1 = std::stoi(argv[15]);
+        BatchStrideC  = std::stoi(argv[16]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 17: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
+               "BatchStrideB0, BatchStrideB1, BatchStrideC\n");
+        exit(0);
+    }
+
+    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
+
+    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
+
+    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    // C_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<CDataType> c_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        break;
+    case 2:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+#ifdef BUILD_INT4_EXAMPLE
+    DeviceMem a_g_m_k_device_buf(sizeof(KernelADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(KernelB0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(KernelB1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(KernelCDataType) *
+                                 c_g_m_o_device_result.mDesc.GetElementSpaceSize());
+
+    const Tensor<KernelADataType> a_g_m_k_converted(a_g_m_k);
+    const Tensor<KernelB0DataType> b0_g_k_n_converted(b0_g_k_n);
+    const Tensor<KernelB1DataType> b1_g_n_o_converted(b1_g_n_o);
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k_converted.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n_converted.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o_converted.mData.data());
+#else
+    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) *
+                                 c_g_m_o_device_result.mDesc.GetElementSpaceSize());
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+#endif
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(
+#ifdef BUILD_INT4_EXAMPLE
+        static_cast<KernelADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<KernelB0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+        static_cast<KernelB1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+        static_cast<KernelCDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+#else
+        static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+        static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+#endif
+        M,
+        N,
+        K,
+        O,
+        BatchCount,
+        StrideA,
+        StrideB0,
+        StrideB1,
+        StrideC,
+        BatchStrideA,
+        BatchStrideB0,
+        BatchStrideB1,
+        BatchStrideC,
+        a_element_op,
+        b0_element_op,
+        acc0_element_op,
+        b1_element_op,
+        c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                            BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        // Output of Gemm0 is input A of Gemm1
+        Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, a1_g_m_n, a_element_op, b0_element_op, PassThrough{});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+#ifdef BUILD_INT4_EXAMPLE
+        Tensor<KernelCDataType> c_g_m_o_device_result_converted(c_g_m_o_host_result.mDesc);
+
+        c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result_converted.mData.data());
+
+        c_g_m_o_device_result = c_g_m_o_device_result_converted.CopyAsType<CDataType>();
+#else
+        c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
+#endif
+
+        return ck::utils::check_err(c_g_m_o_device_result, c_g_m_o_host_result);
+    }
+
+    return true;
+}
diff --git a/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
new file mode 100644
index 0000000..8d9aaec
--- /dev/null
+++ b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
@@ -0,0 +1,16 @@
+add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
+add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_bf16 batched_gemm_scale_softmax_gemm_xdl_bf16.cpp)
+add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
+add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
+add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
+add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
+add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
+
+add_custom_target(example_gemm_scale_softmax_gemm)
+add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
+add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
+add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
+add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
+add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
+add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
+add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
diff --git a/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
new file mode 100644
index 0000000..0eb1565
--- /dev/null
+++ b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
@@ -0,0 +1,160 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskOutUpperTriangle;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp
new file mode 100644
index 0000000..8f1db57
--- /dev/null
+++ b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp
@@ -0,0 +1,159 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = BF16;
+using B0DataType       = BF16;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = BF16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: bf16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, bf16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: bf16 in, bf16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
new file mode 100644
index 0000000..2ce91a8
--- /dev/null
+++ b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
@@ -0,0 +1,160 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp
new file mode 100644
index 0000000..1fd2bf6
--- /dev/null
+++ b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp
@@ -0,0 +1,143 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = BF16;
+using B0DataType       = BF16;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = BF16;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmSpec,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    32,          // KPerBlock
+    64,          // Gemm1NPerBlock
+    32,          // Gemm1KPerBlock
+    8,           // AK1
+    8,           // BK1
+    2,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    2,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<16, 16, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    2,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+    false>;
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp
new file mode 100644
index 0000000..f4a8589
--- /dev/null
+++ b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp
@@ -0,0 +1,144 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmSpec,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    32,          // KPerBlock
+    64,          // Gemm1NPerBlock
+    32,          // Gemm1KPerBlock
+    8,           // AK1
+    8,           // BK1
+    2,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    2,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<16, 16, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    2,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+    false>;
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
new file mode 100644
index 0000000..e4a71b0
--- /dev/null
+++ b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
@@ -0,0 +1,159 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskOutUpperTriangle;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_grouped_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
new file mode 100644
index 0000000..38b5bad
--- /dev/null
+++ b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
@@ -0,0 +1,160 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_grouped_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc
new file mode 100644
index 0000000..4e43dbd
--- /dev/null
+++ b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc
@@ -0,0 +1,261 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 2;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M             = 1020;
+    ck::index_t N             = 1020;
+    ck::index_t K             = 64;
+    ck::index_t O             = 128;
+    ck::index_t BatchCount    = 4;
+    ck::index_t StrideA       = -1;
+    ck::index_t StrideB0      = -1;
+    ck::index_t StrideB1      = -1;
+    ck::index_t StrideC       = -1;
+    ck::index_t BatchStrideA  = -1;
+    ck::index_t BatchStrideB0 = -1;
+    ck::index_t BatchStrideB1 = -1;
+    ck::index_t BatchStrideC  = -1;
+    float alpha               = 1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 9)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+    }
+    else if(argc == 18)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+
+        StrideA  = std::stoi(argv[9]);
+        StrideB0 = std::stoi(argv[10]);
+        StrideB1 = std::stoi(argv[11]);
+        StrideC  = std::stoi(argv[12]);
+
+        BatchStrideA  = std::stoi(argv[13]);
+        BatchStrideB0 = std::stoi(argv[14]);
+        BatchStrideB1 = std::stoi(argv[15]);
+        BatchStrideC  = std::stoi(argv[16]);
+
+        alpha = std::stof(argv[17]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 16: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
+               "BatchStrideB0, BatchStrideB1, BatchStrideC\n");
+        printf("arg17: scale (alpha)\n");
+        exit(0);
+    }
+
+    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
+
+    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
+
+    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+        }
+        else
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+        }
+    };
+
+    // C_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<CDataType> c_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        break;
+    case 2:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) *
+                                 c_g_m_o_device_result.mDesc.GetElementSpaceSize());
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+    auto argument =
+        gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+                          static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+                          static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+                          static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+                          M,
+                          N,
+                          K,
+                          O,
+                          BatchCount,
+                          StrideA,
+                          StrideB0,
+                          StrideB1,
+                          StrideC,
+                          BatchStrideA,
+                          BatchStrideB0,
+                          BatchStrideB1,
+                          BatchStrideC,
+                          a_element_op,
+                          b0_element_op,
+                          acc0_element_op,
+                          b1_element_op,
+                          c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                            BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
+
+    if(do_verification)
+    {
+        // Output of Gemm0 is input A of Gemm1
+        Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+        Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        return ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
new file mode 100644
index 0000000..0b876af
--- /dev/null
+++ b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
@@ -0,0 +1,278 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape for A/B0/B1/C
+    // C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
+    ck::index_t M = 120;
+    ck::index_t N = 1000;
+    ck::index_t K = 64;
+    ck::index_t O = 128;
+
+    // Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
+    // C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
+    // C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
+    ck::index_t G0 = 7;
+    ck::index_t G1 = 13;
+
+    float alpha = 1;
+
+    bool input_permute  = false;
+    bool output_permute = true;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 13)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M  = std::stoi(argv[4]);
+        N  = std::stoi(argv[5]);
+        K  = std::stoi(argv[6]);
+        O  = std::stoi(argv[7]);
+        G0 = std::stoi(argv[8]);
+        G1 = std::stoi(argv[9]);
+
+        alpha = std::stof(argv[10]);
+
+        input_permute  = std::stoi(argv[11]);
+        output_permute = std::stoi(argv[12]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 11: M, N, K, O, G0, G1\n");
+        printf("arg10: scale (alpha)\n");
+        printf("arg11 to 12: input / output permute\n");
+        exit(0);
+    }
+
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides =
+        input_permute
+            ? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
+            : std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
+
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides =
+        input_permute
+            ? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
+            : std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
+
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides =
+        input_permute
+            ? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
+            : std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
+
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides =
+        output_permute
+            ? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
+            : std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
+    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) *
+                           c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    // TODO ANT: replace array with vector?
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(
+        static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+        static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+        static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+        {}, // std::array<void*, 1> p_acc0_biases;
+        {}, // std::array<void*, 1> p_acc1_biases;
+        a_gs_ms_ks_lengths,
+        a_gs_ms_ks_strides,
+        b0_gs_ns_ks_lengths,
+        b0_gs_ns_ks_strides,
+        b1_gs_os_ns_lengths,
+        b1_gs_os_ns_strides,
+        c_gs_ms_os_lengths,
+        c_gs_ms_os_strides,
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
+        a_element_op,
+        b0_element_op,
+        acc0_element_op,
+        b1_element_op,
+        c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    ck::index_t BatchCount = G0 * G1;
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                            BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+
+        // gemm 0
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        // masking
+        const auto mask = DeviceGemmInstance::C0MatrixMask(N);
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(mask.IsMaskedElement(idx[1], idx[2]))
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        // softmax
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        // gemm1
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // permute
+        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+        });
+
+        // default absolute error and relative error is 0.001
+        double rtol = 1e-3;
+        double atol = 1e-3;
+
+        // when BF16 is taken, set absolute error and relative error to 0.01
+        if(std::is_same_v<ADataType, ck::bhalf_t> && std::is_same_v<B0DataType, ck::bhalf_t> &&
+           std::is_same_v<B1DataType, ck::bhalf_t> && std::is_same_v<CDataType, ck::bhalf_t>)
+        {
+            rtol = 1e-2;
+            atol = 1e-2;
+        }
+
+        return ck::utils::check_err(c_gs_ms_os_device_result.mData,
+                                    c_gs_ms_os_host_result.mData,
+                                    "Error: Incorrect results!",
+                                    rtol,
+                                    atol)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
new file mode 100644
index 0000000..ef2acf6
--- /dev/null
+++ b/3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
@@ -0,0 +1,319 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    bool input_permute  = false;
+    bool output_permute = true;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        input_permute  = std::stoi(argv[4]);
+        output_permute = std::stoi(argv[5]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 5: input / output permute\n");
+        exit(0);
+    }
+
+    float alpha = 1; // scaling after 1st gemm
+
+    std::size_t group_count = 7;
+
+    // Problem descs
+    std::vector<DeviceGemmInstance::ProblemDesc> problem_descs;
+    std::vector<const void*> p_a;
+    std::vector<const void*> p_b0;
+    std::vector<const void*> p_b1;
+    std::vector<void*> p_c;
+    std::vector<std::vector<int>> g0_g1_m_n_k_o;
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<B0DataType>> b0_tensors;
+    std::vector<Tensor<B1DataType>> b1_tensors;
+    std::vector<Tensor<CDataType>> c_tensors;
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+    std::vector<DeviceMemPtr> a_tensors_device;
+    std::vector<DeviceMemPtr> b0_tensors_device;
+    std::vector<DeviceMemPtr> b1_tensors_device;
+    std::vector<DeviceMemPtr> c_tensors_device;
+
+    std::size_t flop = 0, num_byte = 0;
+
+    std::cout << "group count " << group_count << ". printing first 4 groups\n";
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        int M  = 128 * (rand() % 8 + 1);
+        int N  = 128 * (rand() % 8 + 1);
+        int K  = 40;
+        int O  = 40 * (rand() % 2 + 1);
+        int G0 = rand() % 3 + 1;
+        int G1 = rand() % 5 + 1;
+
+        g0_g1_m_n_k_o.push_back({G0, G1, M, N, K, O});
+
+        std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+        std::vector<ck::index_t> a_gs_ms_ks_strides =
+            input_permute
+                ? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
+                : std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
+
+        std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+        std::vector<ck::index_t> b0_gs_ns_ks_strides =
+            input_permute
+                ? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
+                : std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
+
+        std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+        std::vector<ck::index_t> b1_gs_os_ns_strides =
+            input_permute
+                ? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
+                : std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
+
+        std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+        std::vector<ck::index_t> c_gs_ms_os_strides =
+            output_permute
+                ? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
+                : std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
+
+        problem_descs.push_back({a_gs_ms_ks_lengths,
+                                 a_gs_ms_ks_strides,
+                                 b0_gs_ns_ks_lengths,
+                                 b0_gs_ns_ks_strides,
+                                 b1_gs_os_ns_lengths,
+                                 b1_gs_os_ns_strides,
+                                 c_gs_ms_os_lengths,
+                                 c_gs_ms_os_strides,
+                                 {},   // acc0_biases_gs_ms_ns_lengths
+                                 {},   // acc0_biases_gs_ms_ns_strides
+                                 {},   // acc1_biases_gs_ms_os_lengths
+                                 {}}); // acc1_biases_gs_ms_os_strides
+
+        // C_m_o = A_m_k * B0_k_n * B1_n_o
+        Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+        Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+        Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+        Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+        int Batch = G0 * G1;
+        flop += (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * Batch;
+        num_byte += (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                    Batch;
+
+        if(i < 4)
+        {
+            std::cout << "a_gs_ms_ks[" << i << "]: " << a_gs_ms_ks.mDesc << ", "
+                      << "b0_gs_ns_ks[" << i << "]: " << b0_gs_ns_ks.mDesc << ", "
+                      << "b1_gs_os_ns[" << i << "]: " << b1_gs_os_ns.mDesc << ", "
+                      << "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << std::endl;
+        }
+
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+            break;
+        case 2:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+            break;
+        case 3:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+            break;
+        default:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        }
+
+        a_tensors.push_back(a_gs_ms_ks);
+        b0_tensors.push_back(b0_gs_ns_ks);
+        b1_tensors.push_back(b1_gs_os_ns);
+        c_tensors.push_back(c_gs_ms_os_device_result);
+
+        a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize()));
+        b0_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize()));
+        b1_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize()));
+        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize()));
+
+        a_tensors_device[i]->ToDevice(a_gs_ms_ks.mData.data());
+        b0_tensors_device[i]->ToDevice(b0_gs_ns_ks.mData.data());
+        b1_tensors_device[i]->ToDevice(b1_gs_os_ns.mData.data());
+
+        p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
+        p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer());
+        p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer());
+        p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
+    }
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(p_a,
+                                      p_b0,
+                                      p_b1,
+                                      p_c,
+                                      {}, // p_acc0_biases
+                                      {}, // p_acc1_biases
+                                      problem_descs,
+                                      a_element_op,
+                                      b0_element_op,
+                                      acc0_element_op,
+                                      b1_element_op,
+                                      c_element_op);
+
+    // specify workspace for problem_desc
+    DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, problem_desc_workspace.GetDeviceBuffer());
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < group_count; i++)
+        {
+            const int& G0 = g0_g1_m_n_k_o[i][0];
+            const int& G1 = g0_g1_m_n_k_o[i][1];
+            const int& M  = g0_g1_m_n_k_o[i][2];
+            const int& N  = g0_g1_m_n_k_o[i][3];
+            const int& K  = g0_g1_m_n_k_o[i][4];
+            const int& O  = g0_g1_m_n_k_o[i][5];
+
+            const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
+            const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
+
+            const auto& a_gs_ms_ks         = a_tensors[i];
+            const auto& b0_gs_ns_ks        = b0_tensors[i];
+            const auto& b1_gs_os_ns        = b1_tensors[i];
+            auto& c_gs_ms_os_device_result = c_tensors[i];
+            auto& c_gs_ms_os_device_buf    = *c_tensors_device[i];
+
+            c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+            Tensor<ADataType> a_g_m_k({G0 * G1, M, K});
+            Tensor<B0DataType> b0_g_k_n({G0 * G1, K, N});
+            Tensor<B1DataType> b1_g_n_o({G0 * G1, N, O});
+            Tensor<AccDataType> acc0_g_m_n({G0 * G1, M, N});        // scratch object after gemm0
+            Tensor<ADataType> a1_g_m_n({G0 * G1, M, N});            // scratch object after softmax
+            Tensor<CDataType> c_g_m_o_host_result({G0 * G1, M, O}); // scratch object after gemm1
+            Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+            // permute
+            a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+                a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+            });
+            b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+                b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+            });
+            b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+                b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+            });
+
+            // gemm 0
+            auto ref_gemm0          = ReferenceGemm0Instance{};
+            auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+            auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+                a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+            ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+            // masking
+            const auto mask = DeviceGemmInstance::C0MatrixMask(N);
+            acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+                if(mask.IsMaskedElement(idx[1], idx[2]))
+                    self(idx) = -ck::NumericLimits<float>::Infinity();
+            });
+
+            // softmax
+            auto ref_softmax          = ReferenceSoftmaxInstance{};
+            auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+            auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+            ref_softmax_invoker.Run(ref_softmax_argument);
+
+            // gemm 1
+            auto ref_gemm1          = ReferenceGemm1Instance{};
+            auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+            auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n,
+                                                             b1_g_n_o,
+                                                             c_g_m_o_host_result,
+                                                             PassThrough{},
+                                                             b1_element_op,
+                                                             c_element_op);
+
+            ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+            // permute
+            c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+                const size_t& g0 = idx[0];
+                const size_t& g1 = idx[1];
+
+                const size_t g = g0 * G1 + g1;
+
+                self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+            });
+
+            bool pass_ =
+                ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData);
+            pass &= pass_;
+        }
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/33_multiple_reduce/CMakeLists.txt b/3rdparty/composable_kernel/example/33_multiple_reduce/CMakeLists.txt
new file mode 100644
index 0000000..bc8c3eb
--- /dev/null
+++ b/3rdparty/composable_kernel/example/33_multiple_reduce/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_example_executable(example_dual_reduce_multiblock dual_reduce_multiblock.cpp)
+add_example_executable(example_dual_reduce_threadwise dual_reduce_threadwise.cpp)
diff --git a/3rdparty/composable_kernel/example/33_multiple_reduce/README.md b/3rdparty/composable_kernel/example/33_multiple_reduce/README.md
new file mode 100644
index 0000000..90762a6
--- /dev/null
+++ b/3rdparty/composable_kernel/example/33_multiple_reduce/README.md
@@ -0,0 +1,37 @@
+# Instructions for ```example_dual_reduce```
+
+## Run ```example_dual_reduce_multiblock```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg2: time kernel (0=no, 1=yes) 
+./bin/example_dual_reduce_multiblock -D 600,28,28,256 -v 1 2 1
+```
+
+Result
+```
+./bin/example_dual_reduce_multiblock -D 600,28,28,256 -v 1 2 1                        
+launch_and_time_kernel: grid_dim {150, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 1.19529 ms, 201.499 GB/s, DeviceMultipleReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_1_InSrcVectorSize_1,OutDstVectorSize_1_1>
+```
+
+## Run ```example_dual_reduce_threadwise```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg2: time kernel (0=no, 1=yes)
+./bin/example_dual_reduce_multiblock -D 8000,4,4,4 -v 1 2 1
+```
+
+Result
+```
+./bin/example_dual_reduce_threadwise -D 8000,4,4,4 -v 1 2 1
+launch_and_time_kernel: grid_dim {32, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.01512 ms, 71.9577 GB/s, DeviceMultipleReduceThreadwise<256,M_C256_S1,K_C1_S4,InSrcVectorDim_1_InSrcVectorSize_2,OutDstVectorSize_1_1>
+```
diff --git a/3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_common.hpp b/3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_common.hpp
new file mode 100644
index 0000000..376b95e
--- /dev/null
+++ b/3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_common.hpp
@@ -0,0 +1,314 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths = {600, 28, 28, 256};
+    size_t n, h, w, c;
+
+    bool do_verification = true;
+    int init_method      = 2;
+    bool time_kernel     = true;
+
+    public:
+    SimpleAppArgs()
+    {
+        n = inLengths[0];
+        h = inLengths[1];
+        w = inLengths[2];
+        c = inLengths[3];
+    };
+
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "Arg1 -- init method (0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg2 -- time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:l:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                if(inLengths.size() != 4)
+                    throw std::runtime_error(
+                        "Invalid option format! The number of integers is incorrect!");
+
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 2 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        n = inLengths[0];
+        h = inLengths[1];
+        w = inLengths[2];
+        c = inLengths[3];
+
+        return (0);
+    };
+};
+
+template <typename InDataType, typename OutDataType1, typename OutDataType2, typename AccDataType>
+static void mean_meansquare_host(const Tensor<InDataType>& in,
+                                 Tensor<OutDataType1>& mean_ref,
+                                 Tensor<OutDataType2>& meansquare_ref,
+                                 size_t n,
+                                 size_t h,
+                                 size_t w,
+                                 size_t c)
+
+{
+    auto thread_reduce_func = [&](auto iN) {
+        AccDataType mean       = ck::type_convert<AccDataType>(0.0f);
+        AccDataType meansquare = ck::type_convert<AccDataType>(0.0f);
+
+        // compute mean, meanquare, variance, invVariance
+        for(std::size_t iH = 0; iH < h; iH++)
+        {
+            for(std::size_t iW = 0; iW < w; iW++)
+            {
+                for(std::size_t iC = 0; iC < c; iC++)
+                {
+                    AccDataType curr_value = ck::type_convert<AccDataType>(in(iN, iH, iW, iC));
+
+                    mean += curr_value;
+                    meansquare += curr_value * curr_value;
+                };
+            }
+        };
+
+        mean       = mean / (h * w * c);
+        meansquare = meansquare / (h * w * c);
+
+        mean_ref(iN)       = ck::type_convert<OutDataType1>(mean);
+        meansquare_ref(iN) = ck::type_convert<OutDataType2>(meansquare);
+    };
+
+    std::size_t num_thread      = std::thread::hardware_concurrency();
+    std::size_t work_per_thread = (n + num_thread - 1) / num_thread;
+
+    std::vector<joinable_thread> threads(num_thread);
+
+    for(std::size_t it = 0; it < num_thread; it++)
+    {
+        std::size_t iN_begin = it * work_per_thread;
+        std::size_t iN_end   = std::min(static_cast<size_t>((it + 1) * work_per_thread), n);
+
+        auto f = [=] {
+            for(std::size_t iN = iN_begin; iN < iN_end; iN++)
+            {
+                thread_reduce_func(iN);
+            }
+        };
+
+        threads[it] = joinable_thread(f);
+    }
+};
+
+using ReduceOperation = ck::reduce::Add;
+
+using InElementwiseOperation_Mean  = ck::tensor_operation::element_wise::PassThrough;
+using AccElementwiseOperation_Mean = ck::tensor_operation::element_wise::UnaryDivide;
+
+using InElementwiseOperation_Meansquare  = ck::tensor_operation::element_wise::UnarySquare;
+using AccElementwiseOperation_Meansquare = ck::tensor_operation::element_wise::UnaryDivide;
+
+using InElementwiseOperationTuple =
+    ck::Tuple<InElementwiseOperation_Mean, InElementwiseOperation_Meansquare>;
+using AccElementwiseOperationTuple =
+    ck::Tuple<AccElementwiseOperation_Mean, AccElementwiseOperation_Meansquare>;
+
+template <typename DeviceDualReduce,
+          typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          int Rank,
+          int NumReduceDim>
+int mean_meansquare_dual_reduce_test(size_t n,
+                                     size_t h,
+                                     size_t w,
+                                     size_t c,
+                                     bool do_verification,
+                                     int init_method,
+                                     bool time_kernel,
+                                     const std::array<int, NumReduceDim> reduceDims)
+{
+    const std::vector<size_t> inLengths = {n, h, w, c};
+
+    Tensor<InDataType> in(inLengths);
+
+    std::vector<size_t> outLengths{n};
+
+    Tensor<OutDataType> mean_ref(outLengths);
+    Tensor<OutDataType> mean(outLengths);
+    Tensor<OutDataType> meansquare_ref(outLengths);
+    Tensor<OutDataType> meansquare(outLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = mean.mDesc.GetStrides();
+
+    size_t invariant_total_length = n;
+    size_t reduce_total_length    = h * w * c;
+
+    const AccDataType alpha = ck::type_convert<AccDataType>(1.0f);
+    const AccDataType beta  = ck::type_convert<AccDataType>(0.0f);
+
+    std::size_t num_thread = 1;
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1: in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread); break;
+        case 2: in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread); break;
+        default: in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem mean_dev(sizeof(OutDataType) * mean.mDesc.GetElementSpaceSize());
+    DeviceMem meansquare_dev(sizeof(OutDataType) * meansquare.mDesc.GetElementSpaceSize());
+
+    in_dev.ToDevice(in.mData.data());
+
+    if(do_verification)
+    {
+        mean_meansquare_host<InDataType, OutDataType, OutDataType, AccDataType>(
+            in, mean_ref, meansquare_ref, n, h, w, c);
+    };
+
+    constexpr ck::index_t NumInputDim  = Rank;
+    constexpr ck::index_t NumOutputDim = (Rank - NumReduceDim > 1) ? Rank - NumReduceDim : 1;
+
+    std::array<ck::index_t, NumInputDim> i_inLengths;
+    std::array<ck::index_t, NumInputDim> i_inStrides;
+    std::array<ck::index_t, NumOutputDim> i_outLengths;
+    std::array<ck::index_t, NumOutputDim> i_outStrides;
+
+    ck::ranges::copy(inLengths, i_inLengths.begin());
+    ck::ranges::copy(inStrides, i_inStrides.begin());
+    ck::ranges::copy(outLengths, i_outLengths.begin());
+    ck::ranges::copy(outStrides, i_outStrides.begin());
+
+    auto dual_reduce_op = DeviceDualReduce{};
+
+    auto argument_ptr = dual_reduce_op.MakeArgumentPointer(
+        i_inLengths,
+        i_inStrides,
+        i_outLengths,
+        {i_outStrides, i_outStrides},
+        reduceDims,
+        {&alpha, &alpha},
+        {&beta, &beta},
+        in_dev.GetDeviceBuffer(),
+        {mean_dev.GetDeviceBuffer(), meansquare_dev.GetDeviceBuffer()},
+        ck::make_tuple(InElementwiseOperation_Mean{}, InElementwiseOperation_Meansquare{}),
+        ck::make_tuple(
+            AccElementwiseOperation_Mean{static_cast<int32_t>(reduce_total_length)},
+            AccElementwiseOperation_Meansquare{static_cast<int32_t>(reduce_total_length)}));
+
+    if(!dual_reduce_op.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+        return (-1);
+    };
+
+    std::string reduce_name = dual_reduce_op.GetTypeString();
+
+    auto invoker_ptr = dual_reduce_op.MakeInvokerPointer();
+
+    float avg_time = 0.0f;
+
+    avg_time += invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InDataType) +
+                            2 * invariant_total_length * sizeof(OutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        mean_dev.FromDevice(mean.mData.data());
+        meansquare_dev.FromDevice(meansquare.mData.data());
+        pass = pass && ck::utils::check_err(mean, mean_ref);
+        pass = pass && ck::utils::check_err(meansquare, meansquare_ref);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_multiblock.cpp b/3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_multiblock.cpp
new file mode 100644
index 0000000..9360599
--- /dev/null
+++ b/3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_multiblock.cpp
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "dual_reduce_common.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+using InDataType       = ck::half_t;
+using OutDataType      = float;
+using OutDataTypeTuple = Tuple<OutDataType, OutDataType>;
+using AccDataType      = float;
+
+// for NHWC layer-norm calculation of mean and meansquare
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 3;
+
+constexpr bool PropagateNan = false;
+
+constexpr InMemoryDataOperationEnum OutMemoryDataOperation = InMemoryDataOperationEnum::Set;
+
+using DeviceDualReduce = DeviceMultipleReduceMultiBlock<2,
+                                                        InDataType,
+                                                        AccDataType,
+                                                        OutDataTypeTuple,
+                                                        Rank,
+                                                        NumReduceDim,
+                                                        ReduceOperation,
+                                                        InElementwiseOperationTuple,
+                                                        AccElementwiseOperationTuple,
+                                                        OutMemoryDataOperation,
+                                                        PropagateNan,
+                                                        256,
+                                                        4,
+                                                        64,
+                                                        1,
+                                                        1,
+                                                        1, // InSrcVectorDim
+                                                        1,
+                                                        ck::Sequence<1, 1>>;
+
+int main(int argc, char* argv[])
+{
+    int retval = 0;
+
+    if(argc > 1)
+    {
+        SimpleAppArgs arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        std::array<int, NumReduceDim> reduceDims = {1, 2, 3};
+
+        retval = mean_meansquare_dual_reduce_test<DeviceDualReduce,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  AccDataType,
+                                                  Rank,
+                                                  NumReduceDim>(arg.n,
+                                                                arg.h,
+                                                                arg.w,
+                                                                arg.c,
+                                                                arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                reduceDims);
+    }
+    else
+    {
+        std::array<int, NumReduceDim> reduceDims = {1, 2, 3};
+
+        retval = mean_meansquare_dual_reduce_test<DeviceDualReduce,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  AccDataType,
+                                                  Rank,
+                                                  NumReduceDim>(
+            600, 28, 28, 256, true, 2, true, reduceDims);
+    };
+
+    return (retval);
+}
diff --git a/3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_threadwise.cpp b/3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_threadwise.cpp
new file mode 100644
index 0000000..5625583
--- /dev/null
+++ b/3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_threadwise.cpp
@@ -0,0 +1,93 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "dual_reduce_common.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+using InDataType       = ck::half_t;
+using OutDataType      = float;
+using OutDataTypeTuple = Tuple<OutDataType, OutDataType>;
+using AccDataType      = float;
+
+// for NHWC layer-norm calculation of mean and meansquare
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 3;
+
+constexpr bool PropagateNan = false;
+
+using DeviceDualReduce = DeviceMultipleReduceThreadWise<2,
+                                                        InDataType,
+                                                        AccDataType,
+                                                        OutDataTypeTuple,
+                                                        Rank,
+                                                        NumReduceDim,
+                                                        ReduceOperation,
+                                                        InElementwiseOperationTuple,
+                                                        AccElementwiseOperationTuple,
+                                                        PropagateNan,
+                                                        256,
+                                                        1,
+                                                        4,
+                                                        1, // InSrcVectorDim
+                                                        2,
+                                                        ck::Sequence<1, 1>>;
+
+int main(int argc, char* argv[])
+{
+    int retval = 0;
+
+    if(argc > 1)
+    {
+        SimpleAppArgs arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        std::array<int, NumReduceDim> reduceDims = {1, 2, 3};
+
+        retval = mean_meansquare_dual_reduce_test<DeviceDualReduce,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  AccDataType,
+                                                  Rank,
+                                                  NumReduceDim>(arg.n,
+                                                                arg.h,
+                                                                arg.w,
+                                                                arg.c,
+                                                                arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                reduceDims);
+    }
+    else
+    {
+        std::array<int, NumReduceDim> reduceDims = {1, 2, 3};
+
+        retval = mean_meansquare_dual_reduce_test<DeviceDualReduce,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  AccDataType,
+                                                  Rank,
+                                                  NumReduceDim>(
+            8000, 4, 4, 4, true, 2, true, reduceDims);
+    };
+
+    return (retval);
+}
diff --git a/3rdparty/composable_kernel/example/34_batchnorm/CMakeLists.txt b/3rdparty/composable_kernel/example/34_batchnorm/CMakeLists.txt
new file mode 100644
index 0000000..d964f40
--- /dev/null
+++ b/3rdparty/composable_kernel/example/34_batchnorm/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_example_executable(example_batchnorm_forward_training batchnorm_forward_training_nhwc.cpp)
+add_example_executable(example_batchnorm_forward_inferring batchnorm_forward_inferring_nhwc.cpp)
+add_example_executable(example_batchnorm_backward batchnorm_backward_nhwc.cpp)
diff --git a/3rdparty/composable_kernel/example/34_batchnorm/README.md b/3rdparty/composable_kernel/example/34_batchnorm/README.md
new file mode 100644
index 0000000..294e32b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/34_batchnorm/README.md
@@ -0,0 +1,81 @@
+# Instructions for ```batchnorm nhwc``` Example
+
+## Run ```batchnorm forward nhwc```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1:  data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
+#arg2: 1/0 to indicate whether to update the moving average and variance (0=no, 1=yes)
+#arg3: 1/0 to indicate whether to save result mean/invVariance (0=no, 1=yes)
+#arg4: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg5: time kernel (0=no, 1=yes) 
+./bin/example_batchnorm_forward -D 128,16,16,1024 -v 1 0 0 1 2 1
+```
+
+Result 
+```
+./bin/example_batchnorm_forward -D 128,16,16,1024 -v 1 0 0 1 2 1
+launch_and_time_kernel: grid_dim {64, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {120, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {120, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 2.08231 ms, 354.519 GB/s
+```
+
+Result
+```
+./bin/example_batchnorm_forward -D 128,16,16,1024 -v 1 0 1 0 2 0
+echo $?
+0
+```
+
+## Run ```batchnorm infer nhwc```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1:  data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+./bin/example_batchnorm_infer -D 128,16,16,1024 -v 1 0 2 1
+```
+
+Result
+```
+./bin/example_batchnorm_infer -D 128,16,16,1024 -v 1 0 2 1
+launch_and_time_kernel: grid_dim {120, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 1.28235 ms, 523.329 GB/s
+```
+
+## Run ```batchnorm backward nhwc```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
+Arg2 -- 1/0 to indicate whether to use saved mean and invVariance
+Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+Arg4 -- time kernel (0=no, 1=yes)
+Arg5: use multi-block welford (0=n0, 1=yes)
+./bin/example_batchnorm_backward -D 128,16,3,1024 -v 1 0 0 3 1 1
+```
+
+Result 
+```
+./bin/example_batchnorm_backward -D 128,16,3,1024 -v 1 0 0 3 1 1
+launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.411026 ms, 91.8702 GB/s
+```
diff --git a/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_backward_nhwc.cpp b/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_backward_nhwc.cpp
new file mode 100644
index 0000000..a6ca9d1
--- /dev/null
+++ b/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_backward_nhwc.cpp
@@ -0,0 +1,506 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <limits>
+#include <iostream>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
+
+static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class BatchNormBwdArg
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inOutLengths;
+
+    bool do_verification = false;
+
+    bool haveSavedMeanInvVar;
+
+    int data_type               = 0;
+    int init_method             = 3;
+    bool time_kernel            = false;
+    bool use_multiblock_welford = false;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        // clang-format off
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2 -- 1/0 to indicate whether to use saved mean and invVariance" << std::endl;
+        std::cout << "Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
+        std::cout << "Arg4 -- time kernel (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg5: use multi-block welford (0=n0, 1=yes)" << std::endl;
+        // clang-format on
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inOutLengths = getTypeValuesFromString<size_t>(optarg);
+
+                if(inOutLengths.size() != 4)
+                    throw std::runtime_error(
+                        "NHWC tensor layout should have 4 length values specified!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 5 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type              = std::atoi(argv[optind++]);
+        haveSavedMeanInvVar    = std::atoi(argv[optind++]);
+        init_method            = std::atoi(argv[optind++]);
+        time_kernel            = static_cast<bool>(std::atoi(argv[optind++]));
+        use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
+
+        return (0);
+    };
+};
+
+using namespace ck;
+
+template <typename XDataType, typename AccDataType, bool UseMultiblockInK>
+bool bnorm_bwd_nhwc_test(bool do_verification,
+                         int init_method,
+                         bool time_kernel,
+                         const std::vector<size_t> inOutLengths,
+                         bool haveSavedMeanInvVar,
+                         double epsilon)
+{
+    // for NHWC BatchNorm calculation of mean and meansquare
+    constexpr index_t Rank         = 4;
+    constexpr index_t NumReduceDim = 3;
+
+    using ScaleDataType = XDataType;
+
+    const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
+
+    // input data of the batchnorm backward algorithm
+    Tensor<XDataType> x(inOutLengths);
+    Tensor<AccDataType> dy(inOutLengths);
+
+    Tensor<ScaleDataType> bnScale(scaleBiasMeanVarLengths);
+
+    Tensor<AccDataType> savedMean(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> savedInvVar(scaleBiasMeanVarLengths);
+    // savedVariance is only used for initializing savedInvVar
+    Tensor<AccDataType> savedVariance(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm backward algorithm
+    Tensor<AccDataType> dx_ref(inOutLengths);
+    Tensor<AccDataType> dx(inOutLengths);
+
+    Tensor<AccDataType> dscale(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> dbias(scaleBiasMeanVarLengths);
+
+    Tensor<AccDataType> dscale_ref(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> dbias_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = dy.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = dscale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(haveSavedMeanInvVar)
+    {
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 1.0f;
+        const float noise_stddev = 0.0001f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+
+        // initialize the savedMean to be values with tiny variation to the mean of the x values
+        savedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
+                                      num_thread);
+
+        // initialize the variance to be values with tiny variation to the variance of the x values
+        savedVariance.GenerateTensorValue(
+            GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+
+        auto it_src       = savedVariance.mData.begin();
+        auto it_dst       = savedInvVar.mData.begin();
+        float tmp_epsilon = std::numeric_limits<float>::epsilon();
+
+        while(it_src != savedVariance.mData.end())
+        {
+            *it_dst = type_convert<AccDataType>(
+                1.0f / std::sqrtf(type_convert<float>(*it_src) + tmp_epsilon));
+
+            it_src++;
+            it_dst++;
+        };
+    }
+    else
+    {
+        const float x_mean   = 0.0f;
+        const float x_stddev = 1.0f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            dy.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_0<ScaleDataType>{}, num_thread);
+            break;
+        case 1:
+            dy.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_1<ScaleDataType>{1}, num_thread);
+            break;
+        case 2:
+            dy.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-2, 2}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_2<ScaleDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            dy.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-0.2f, 0.2f}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_3<ScaleDataType>{-0.5f, 0.5f}, num_thread);
+        }
+    };
+
+    // input data of the batchnorm backward algorithm
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem dy_dev(sizeof(AccDataType) * dy.mDesc.GetElementSpaceSize());
+
+    DeviceMem bnScale_dev(sizeof(ScaleDataType) * bnScale.mDesc.GetElementSpaceSize());
+
+    DeviceMem savedMean_dev(sizeof(AccDataType) * savedMean.mDesc.GetElementSpaceSize());
+    DeviceMem savedInvVar_dev(sizeof(AccDataType) * savedInvVar.mDesc.GetElementSpaceSize());
+
+    // output data of the batchnorm backward algorithm
+    DeviceMem dx_dev(sizeof(AccDataType) * dx.mDesc.GetElementSpaceSize());
+
+    DeviceMem dscale_dev(sizeof(AccDataType) * dscale.mDesc.GetElementSpaceSize());
+    DeviceMem dbias_dev(sizeof(AccDataType) * dbias.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    dy_dev.ToDevice(dy.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+
+    if(haveSavedMeanInvVar)
+    {
+        savedMean_dev.ToDevice(savedMean.mData.data());
+        savedInvVar_dev.ToDevice(savedInvVar.mData.data());
+    };
+
+    std::array<index_t, Rank> i_inOutLengths;
+    std::array<index_t, Rank> i_inOutStrides;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
+
+    std::copy(inOutLengths.begin(), inOutLengths.end(), i_inOutLengths.begin());
+    std::copy(inOutStrides.begin(), inOutStrides.end(), i_inOutStrides.begin());
+    std::copy(scaleBiasMeanVarLengths.begin(),
+              scaleBiasMeanVarLengths.end(),
+              i_scaleBiasMeanVarLengths.begin());
+    std::copy(scaleBiasMeanVarStrides.begin(),
+              scaleBiasMeanVarStrides.end(),
+              i_scaleBiasMeanVarStrides.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    using DeviceBatchNormBwdInstance =
+        ck::tensor_operation::device::DeviceBatchNormBwdImpl<XDataType,
+                                                             AccDataType,
+                                                             AccDataType,
+                                                             AccDataType,
+                                                             ScaleDataType, // ScaleDataType
+                                                             AccDataType,   // DscaleDbiasDataType
+                                                             AccDataType,   // MeanVarDataType
+                                                             PassThroughOp,
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             UseMultiblockInK,
+                                                             256,
+                                                             16,
+                                                             16,
+                                                             1,
+                                                             2,
+                                                             0,
+                                                             1,  // XSrcVectorSize
+                                                             1,  // DySrcVectorSize
+                                                             1,  // DxDstVectorSize
+                                                             1,  // ScaleSrcVectorSize
+                                                             1,  // DscaleDbiasDstVectorSize
+                                                             1>; // MeanVarSrcVectorSize
+
+    auto batchnorm_bwd = DeviceBatchNormBwdInstance{};
+
+    auto argument_ptr = batchnorm_bwd.MakeArgumentPointer(
+        i_inOutLengths,
+        i_inOutStrides,
+        i_inOutStrides,
+        i_inOutStrides,
+        {0, 1, 2},
+        i_scaleBiasMeanVarLengths,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        x_dev.GetDeviceBuffer(),
+        dy_dev.GetDeviceBuffer(),
+        bnScale_dev.GetDeviceBuffer(),
+        haveSavedMeanInvVar ? savedMean_dev.GetDeviceBuffer() : nullptr,
+        haveSavedMeanInvVar ? savedInvVar_dev.GetDeviceBuffer() : nullptr,
+        epsilon,
+        PassThroughOp{},
+        dx_dev.GetDeviceBuffer(),
+        dscale_dev.GetDeviceBuffer(),
+        dbias_dev.GetDeviceBuffer());
+
+    if(!batchnorm_bwd.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
+                     "exiting!"
+                  << std::endl;
+        return (false);
+    };
+
+    size_t workspace_sz = batchnorm_bwd.GetWorkSpaceSize(argument_ptr.get());
+
+    DeviceMem workspace_dev(workspace_sz);
+
+    batchnorm_bwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+    auto invoker_ptr = batchnorm_bwd.MakeInvokerPointer();
+
+    if(time_kernel)
+    {
+        float avg_time   = 0.0f;
+        size_t num_bytes = 0;
+
+        size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
+        size_t invariant_length = inOutLengths[3];
+
+        avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        // inputing of x, dy, scale, outputing of dx, dscale, dbias
+        num_bytes +=
+            total_length * sizeof(XDataType) * 3 + invariant_length * sizeof(AccDataType) * 3;
+
+        // outputing of mean, inv-variance
+        num_bytes += haveSavedMeanInvVar ? invariant_length * sizeof(AccDataType) * 2 : 0;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+    }
+    else
+        (void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        using ReferenceBatchNormBwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormBwd<XDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              ScaleDataType, // ScaleDataType
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumReduceDim>;
+
+        auto batchNormBwd_ref = ReferenceBatchNormBwdInstance{};
+
+        auto argument_ptr_ref = batchNormBwd_ref.MakeArgumentPointer(
+            i_inOutLengths,
+            i_inOutStrides,
+            i_inOutStrides,
+            i_inOutStrides,
+            {0, 1, 2},
+            i_scaleBiasMeanVarLengths,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            x.mData.data(),
+            dy.mData.data(),
+            bnScale.mData.data(),
+            haveSavedMeanInvVar ? savedMean.mData.data() : nullptr,
+            haveSavedMeanInvVar ? savedInvVar.mData.data() : nullptr,
+            epsilon,
+            PassThroughOp{},
+            dx_ref.mData.data(),
+            dscale_ref.mData.data(),
+            dbias_ref.mData.data());
+
+        if(!batchNormBwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout
+                << "The runtime parameters seems not supported by the device instance, exiting!"
+                << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormBwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+
+        dx_dev.FromDevice(dx.mData.data());
+        dscale_dev.FromDevice(dscale.data());
+        dbias_dev.FromDevice(dbias.data());
+
+        // clang-format off
+        pass = pass && ck::utils::check_err(dbias.mData, dbias_ref.mData, "dBias result:", 2e-4, 2e-4);
+        pass = pass && ck::utils::check_err(dscale.mData, dscale_ref.mData, "dScale result:", 2e-4, 2e-4);
+        pass = pass && ck::utils::check_err(dx.mData, dx_ref.mData, "dx result:");
+        // clang-format on
+    };
+
+    return (pass);
+};
+
+static const double epsilon = std::numeric_limits<float>::epsilon();
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        BatchNormBwdArg arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
+                                                                    arg.init_method,
+                                                                    arg.time_kernel,
+                                                                    arg.inOutLengths,
+                                                                    arg.haveSavedMeanInvVar,
+                                                                    epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.haveSavedMeanInvVar,
+                                                                     epsilon);
+        }
+        else if(arg.data_type == 1)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<float, float, true>(arg.do_verification,
+                                                               arg.init_method,
+                                                               arg.time_kernel,
+                                                               arg.inOutLengths,
+                                                               arg.haveSavedMeanInvVar,
+                                                               epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<float, float, false>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.haveSavedMeanInvVar,
+                                                                epsilon);
+        }
+        else if(arg.data_type == 5)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.haveSavedMeanInvVar,
+                                                                     epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
+                                                                      arg.init_method,
+                                                                      arg.time_kernel,
+                                                                      arg.inOutLengths,
+                                                                      arg.haveSavedMeanInvVar,
+                                                                      epsilon);
+        }
+        else if(arg.data_type == 6)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<double, double, true>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.haveSavedMeanInvVar,
+                                                                 epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<double, double, false>(arg.do_verification,
+                                                                  arg.init_method,
+                                                                  arg.time_kernel,
+                                                                  arg.inOutLengths,
+                                                                  arg.haveSavedMeanInvVar,
+                                                                  epsilon);
+        }
+    }
+    else
+    {
+        pass = bnorm_bwd_nhwc_test<ck::half_t, float, true>(true,
+                                                            3,
+                                                            false, // don't time kernel
+                                                            {128, 16, 6, 512},
+                                                            false,
+                                                            epsilon);
+
+        pass = pass && bnorm_bwd_nhwc_test<ck::half_t, float, false>(true,
+                                                                     3,
+                                                                     false, // don't time kernel
+                                                                     {128, 16, 3, 1024},
+                                                                     false,
+                                                                     epsilon);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_common.hpp b/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_common.hpp
new file mode 100644
index 0000000..bdbc8ea
--- /dev/null
+++ b/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_common.hpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cassert>
+#include <vector>
+#include <array>
+#include <type_traits>
+
+#include "ck/utility/data_type.hpp"
+
+struct NormalizeInInfer
+{
+    NormalizeInInfer(double epsilon = 1e-4) : epsilon_(epsilon) {}
+
+    template <typename T1, typename T2, typename T3, typename T4>
+    __host__ __device__ constexpr void operator()(T1& y,
+                                                  const T1& x,
+                                                  const T2& mean,
+                                                  const T2& variance,
+                                                  const T3& gamma,
+                                                  const T4& beta) const
+    {
+        static_assert(std::is_same<T2, float>::value || std::is_same<T2, double>::value,
+                      "Data type is not supported by this operation!");
+
+        using ck::type_convert;
+        using ck::math::sqrt;
+
+        T2 tmp_x, tmp_y;
+
+        tmp_x = type_convert<T2>(x);
+
+        tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) *
+                    type_convert<T2>(gamma) +
+                type_convert<T2>(beta);
+        y = type_convert<T1>(tmp_y);
+    };
+
+    double epsilon_;
+};
+
+template <int Rank, int NumReduceDim>
+static inline std::array<int, Rank - NumReduceDim>
+get_invariant_dims(const std::array<int, NumReduceDim>& reduceDims)
+{
+    int reduceFlag = 0;
+
+    // flag the bits for the reduceDims
+    for(int i = 0; i < NumReduceDim; i++)
+    {
+        reduceFlag |= 1 << reduceDims[i];
+    };
+
+    std::array<int, Rank - NumReduceDim> invariantDims;
+
+    // collect invariant dimensions
+    int dim = 0;
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) == 0)
+        {
+            invariantDims[dim] = i;
+            dim++;
+        };
+
+    return invariantDims;
+};
diff --git a/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp b/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp
new file mode 100644
index 0000000..dc29848
--- /dev/null
+++ b/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp
@@ -0,0 +1,366 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <limits>
+#include <iostream>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp"
+
+#include "batchnorm_infer_impl.hpp"
+
+static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class BatchNormInferArg
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inOutLengths;
+
+    bool do_verification = false;
+
+    int data_type    = 0;
+    int init_method  = 2;
+    bool time_kernel = false;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension "
+                     "lengths, must have 4 integers for nhwc"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the batch-normalization "
+                     "result by "
+                     "comparing with the host-based batch-normalization"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2: init method used for bnScale and bnBias (0=no init, 1=single integer "
+                     "value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg3: time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inOutLengths = getTypeValuesFromString<size_t>(optarg);
+
+                if(inOutLengths.size() != 4)
+                    throw std::runtime_error(
+                        "NHWC tensor layout should have 4 length values specified!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 3 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return (-1);
+
+        return (0);
+    };
+};
+
+using namespace ck;
+
+template <typename InOutDataType, typename AccDataType>
+bool bnorm_infer_nhwc_test(bool do_verification,
+                           int init_method,
+                           bool time_kernel,
+                           const std::vector<size_t> inOutLengths,
+                           double epsilon)
+{
+    // for NHWC BatchNorm calculation of mean and meansquare
+    constexpr int Rank         = 4;
+    constexpr int NumReduceDim = 3;
+
+    // when using lengths[] to create a tensor, lengths[0] is the length of highest dimension
+    // eg. N of NHWC, so lengths[3] is the dimension C length of NHWC
+    const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
+
+    // input data of the batchnorm forward algorithm
+    Tensor<InOutDataType> x(inOutLengths);
+    Tensor<AccDataType> bnScale(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> bnBias(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm forward algorithm
+    Tensor<InOutDataType> y_ref(inOutLengths);
+    Tensor<InOutDataType> y(inOutLengths);
+
+    Tensor<AccDataType> estimatedMean(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> estimatedVariance(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if constexpr(std::is_same<InOutDataType, int8_t>::value)
+    {
+        x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 2.5f;
+        const float noise_stddev = 0.0001f;
+
+        estimatedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
+                                          num_thread);
+
+        estimatedVariance.GenerateTensorValue(
+            GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+    }
+    else
+    {
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 1.0f;
+        const float noise_stddev = 0.0001f;
+
+        x.GenerateTensorValue(GeneratorTensor_4<InOutDataType>{x_mean, x_stddev}, num_thread);
+
+        // initialize the savedMean to be values with tiny variation to the mean of the x values
+        estimatedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
+                                          num_thread);
+
+        // initialize the variance to be values with tiny variation to the variance of the x values
+        estimatedVariance.GenerateTensorValue(
+            GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            bnScale.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            break;
+        case 1:
+            bnScale.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0}, num_thread);
+            break;
+        case 2:
+            bnScale.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            bnScale.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem x_dev(sizeof(InOutDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(InOutDataType) * y.mDesc.GetElementSpaceSize());
+    DeviceMem bnScale_dev(sizeof(AccDataType) * bnScale.mDesc.GetElementSpaceSize());
+    DeviceMem bnBias_dev(sizeof(AccDataType) * bnBias.mDesc.GetElementSpaceSize());
+
+    // mean_dev or resultSaveMean_dev
+    DeviceMem estimatedMean_dev(sizeof(AccDataType) * estimatedMean.mDesc.GetElementSpaceSize());
+    // meansquare_dev or resultSaveInvVariance_dev
+    DeviceMem estimatedVariance_dev(sizeof(AccDataType) *
+                                    estimatedVariance.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+    bnBias_dev.ToDevice(bnBias.mData.data());
+    estimatedMean_dev.ToDevice(estimatedMean.mData.data());
+    estimatedVariance_dev.ToDevice(estimatedVariance.mData.data());
+
+    using ck::index_t;
+
+    std::array<index_t, Rank> i_inOutLengths;
+    std::array<index_t, Rank> i_inOutStrides;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
+
+    ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
+    ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
+    ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
+    ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
+
+    int result = 0;
+
+    result = bnorm_infer<InOutDataType,
+                         InOutDataType,
+                         AccDataType,
+                         AccDataType,
+                         AccDataType,
+                         AccDataType,
+                         Rank,
+                         NumReduceDim,
+                         false>(time_kernel,
+                                {0, 1, 2},
+                                i_inOutLengths,
+                                i_inOutStrides,
+                                i_inOutStrides,
+                                i_scaleBiasMeanVarLengths,
+                                i_scaleBiasMeanVarStrides,
+                                i_scaleBiasMeanVarStrides,
+                                i_scaleBiasMeanVarStrides,
+                                x_dev.GetDeviceBuffer(),
+                                bnScale_dev.GetDeviceBuffer(),
+                                bnBias_dev.GetDeviceBuffer(),
+                                epsilon,
+                                estimatedMean_dev.GetDeviceBuffer(),
+                                estimatedVariance_dev.GetDeviceBuffer(),
+                                y_dev.GetDeviceBuffer());
+
+    if(result < 0)
+        return (false);
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+        using ReferenceBatchNormInferInstance =
+            ck::tensor_operation::host::ReferenceBatchNormInfer<InOutDataType,
+                                                                InOutDataType,
+                                                                AccDataType,
+                                                                AccDataType,
+                                                                AccDataType,
+                                                                AccDataType,
+                                                                PassThroughOp,
+                                                                Rank,
+                                                                NumReduceDim>;
+        auto batchNormInfer_ref = ReferenceBatchNormInferInstance{};
+
+        auto argument_ptr_ref =
+            batchNormInfer_ref.MakeArgumentPointer(i_inOutLengths,
+                                                   i_inOutStrides,
+                                                   i_inOutStrides,
+                                                   {0, 1, 2},
+                                                   i_scaleBiasMeanVarLengths,
+                                                   i_scaleBiasMeanVarStrides,
+                                                   i_scaleBiasMeanVarStrides,
+                                                   i_scaleBiasMeanVarStrides,
+                                                   x.mData.data(),
+                                                   bnScale.mData.data(),
+                                                   bnBias.mData.data(),
+                                                   epsilon,
+                                                   PassThroughOp{},
+                                                   estimatedMean.mData.data(),
+                                                   estimatedVariance.mData.data(),
+                                                   y_ref.mData.data());
+
+        if(!batchNormInfer_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout
+                << "The runtime parameters seems not supported by the BatchNorm instance, exiting!"
+                << std::endl;
+            return (-2);
+        };
+
+        auto invoker_ptr_ref = batchNormInfer_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+
+        y_dev.FromDevice(y.mData.data());
+        pass = pass && ck::utils::check_err(y, y_ref);
+    };
+
+    return (pass);
+};
+
+static const double epsilon = std::numeric_limits<float>::epsilon();
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        BatchNormInferArg arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            pass = bnorm_infer_nhwc_test<ck::half_t, float>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        }
+        else if(arg.data_type == 1)
+        {
+            pass = bnorm_infer_nhwc_test<float, float>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        }
+        else if(arg.data_type == 3)
+        {
+            pass = bnorm_infer_nhwc_test<int8_t, float>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        }
+        else if(arg.data_type == 5)
+        {
+            pass = bnorm_infer_nhwc_test<ck::bhalf_t, float>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        }
+        else if(arg.data_type == 6)
+        {
+            pass = bnorm_infer_nhwc_test<double, double>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        };
+    }
+    else
+    {
+        pass = bnorm_infer_nhwc_test<ck::half_t, float>(true,
+                                                        2,
+                                                        false, // don't time kernel
+                                                        {128, 16, 16, 1024},
+                                                        epsilon);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp b/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp
new file mode 100644
index 0000000..da36d65
--- /dev/null
+++ b/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp
@@ -0,0 +1,591 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <limits>
+#include <iostream>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class BatchNormFwdArg
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inOutLengths;
+
+    bool do_verification = false;
+
+    bool updateMovingAverage;
+    bool saveMeanAndInvVariance;
+
+    int data_type               = 0;
+    int init_method             = 2;
+    bool time_kernel            = false;
+    bool use_multiblock_welford = false;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension "
+                     "lengths, must have 4 integers for nhwc"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the batch-normalization "
+                     "result by "
+                     "comparing with the host-based batch-normalization"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2: 1/0 to indicate whether to update the moving average and variance "
+                     "(0=no, 1=yes)"
+                  << std::endl;
+        std::cout << "Arg3: 1/0 to indicate whether to save the calculated mean and invVariance "
+                     "(0=no, 1=yes)"
+                  << std::endl;
+        std::cout << "Arg4: init method used for bnScale and bnBias (0=no init, 1=single integer "
+                     "value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg5: time kernel (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg6: use multi-block welford (0=n0, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inOutLengths = getTypeValuesFromString<size_t>(optarg);
+
+                if(inOutLengths.size() != 4)
+                    throw std::runtime_error(
+                        "NHWC tensor layout should have 4 length values specified!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 6 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type              = std::atoi(argv[optind++]);
+        updateMovingAverage    = std::atoi(argv[optind++]);
+        saveMeanAndInvVariance = std::atoi(argv[optind++]);
+        init_method            = std::atoi(argv[optind++]);
+        time_kernel            = static_cast<bool>(std::atoi(argv[optind++]));
+        use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return (-1);
+
+        return (0);
+    };
+};
+
+using namespace ck;
+
+template <typename InOutDataType, typename AccDataType, bool UseMultiblockInK>
+bool bnorm_fwd_nhwc_test(bool do_verification,
+                         int init_method,
+                         bool time_kernel,
+                         const std::vector<size_t> inOutLengths,
+                         bool updateMovingAverage,
+                         bool saveMeanAndInvVariance,
+                         double averageFactor,
+                         double epsilon)
+{
+    // for NHWC BatchNorm calculation of mean and meansquare
+    constexpr int Rank         = 4;
+    constexpr int NumReduceDim = 3;
+
+    // when using lengths[] to create a tensor, lengths[0] is the length of highest dimension
+    // eg. N of NHWC, so lengths[3] is the dimension C length of NHWC
+    const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
+
+    // input data of the batchnorm forward algorithm
+    Tensor<InOutDataType> x(inOutLengths);
+    Tensor<AccDataType> bnScale(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> bnBias(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm forward algorithm
+    Tensor<InOutDataType> y_ref(inOutLengths);
+    Tensor<InOutDataType> y(inOutLengths);
+
+    Tensor<AccDataType> resultSaveMean_ref(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> resultSaveInvVariance_ref(scaleBiasMeanVarLengths);
+
+    Tensor<AccDataType> resultRunningMean_ref(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> resultRunningVariance_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(updateMovingAverage)
+    {
+        if constexpr(std::is_same<InOutDataType, int8_t>::value)
+        {
+            x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+
+            const float x_mean       = 0.0f;
+            const float x_stddev     = 2.5f;
+            const float noise_stddev = 0.04f;
+
+            resultRunningMean_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
+
+            resultRunningVariance_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+        }
+        else
+        {
+            const float x_mean       = 0.0f;
+            const float x_stddev     = 1.0f;
+            const float noise_stddev = 0.04f;
+
+            // input data in normal distribution
+            x.GenerateTensorValue(GeneratorTensor_4<InOutDataType>{x_mean, x_stddev}, num_thread);
+
+            // initialize the runningMean to be values with tiny variation to the mean of the x
+            // values
+            resultRunningMean_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
+
+            // initialize the runningVariance to be values with tiny variation to the variance of
+            // the x values
+            resultRunningVariance_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+        };
+    }
+    else
+    {
+        if constexpr(std::is_same<InOutDataType, int8_t>::value)
+            x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+        else
+            x.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0f, 5.0f}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            bnScale.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            break;
+        case 1:
+            bnScale.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0}, num_thread);
+            break;
+        case 2:
+            bnScale.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            bnScale.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem x_dev(sizeof(InOutDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(InOutDataType) * y.mDesc.GetElementSpaceSize());
+    DeviceMem bnScale_dev(sizeof(AccDataType) * bnScale.mDesc.GetElementSpaceSize());
+    DeviceMem bnBias_dev(sizeof(AccDataType) * bnBias.mDesc.GetElementSpaceSize());
+
+    // mean_dev or resultSaveMean_dev
+    DeviceMem resultSaveMean_dev(sizeof(AccDataType) *
+                                 resultSaveMean_ref.mDesc.GetElementSpaceSize());
+    // meansquare_dev or resultSaveInvVariance_dev
+    DeviceMem resultSaveInvVariance_dev(sizeof(AccDataType) *
+                                        resultSaveInvVariance_ref.mDesc.GetElementSpaceSize());
+    // resultRunningMean_dev
+    DeviceMem resultRunningMean_dev(sizeof(AccDataType) *
+                                    resultRunningMean_ref.mDesc.GetElementSpaceSize());
+    // resultRunningVariance_dev
+    DeviceMem resultRunningVariance_dev(sizeof(AccDataType) *
+                                        resultRunningVariance_ref.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+    bnBias_dev.ToDevice(bnBias.mData.data());
+
+    if(updateMovingAverage)
+    {
+        resultRunningMean_dev.ToDevice(resultRunningMean_ref.mData.data());
+        resultRunningVariance_dev.ToDevice(resultRunningVariance_ref.mData.data());
+    };
+
+    std::array<index_t, Rank> i_inOutLengths;
+    std::array<index_t, Rank> i_inOutStrides;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
+
+    ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
+    ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
+    ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
+    ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    using DeviceBatchNormFwdInstance =
+        ck::tensor_operation::device::DeviceBatchNormFwdImpl<InOutDataType,
+                                                             InOutDataType,
+                                                             AccDataType,
+                                                             AccDataType,   // ScaleDataType
+                                                             AccDataType,   // BiasDataType
+                                                             AccDataType,   // MeanVarDataType
+                                                             PassThroughOp, // YElementwiseOp
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             UseMultiblockInK,
+                                                             256,
+                                                             16,
+                                                             16,
+                                                             1,
+                                                             2,
+                                                             0,
+                                                             1,
+                                                             1,
+                                                             1,
+                                                             1,
+                                                             1>;
+
+    auto batchnorm_fwd = DeviceBatchNormFwdInstance{};
+
+    auto argument_ptr = batchnorm_fwd.MakeArgumentPointer(
+        i_inOutLengths,
+        i_inOutStrides,
+        i_inOutStrides,
+        {0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
+        i_scaleBiasMeanVarLengths,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        x_dev.GetDeviceBuffer(),
+        bnScale_dev.GetDeviceBuffer(),
+        bnBias_dev.GetDeviceBuffer(),
+        epsilon,
+        PassThroughOp{},
+        y_dev.GetDeviceBuffer(),
+        saveMeanAndInvVariance ? resultSaveMean_dev.GetDeviceBuffer() : nullptr,
+        saveMeanAndInvVariance ? resultSaveInvVariance_dev.GetDeviceBuffer() : nullptr,
+        averageFactor,
+        updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
+        updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr);
+
+    if(!batchnorm_fwd.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
+                     "exiting!"
+                  << std::endl;
+        return (false);
+    };
+
+    size_t workspace_sz = batchnorm_fwd.GetWorkSpaceSize(argument_ptr.get());
+
+    DeviceMem workspace_dev(workspace_sz);
+
+    batchnorm_fwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+    auto invoker_ptr = batchnorm_fwd.MakeInvokerPointer();
+
+    if(time_kernel)
+    {
+        float avg_time   = 0.0f;
+        size_t num_bytes = 0;
+
+        size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
+        size_t invariant_length = inOutLengths[3];
+
+        avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        // inputing of x, scale, bias, outputing of y
+        num_bytes +=
+            total_length * sizeof(InOutDataType) * 2 + invariant_length * sizeof(AccDataType) * 2;
+
+        // outputing of mean, inv-variance
+        num_bytes += saveMeanAndInvVariance ? invariant_length * sizeof(AccDataType) * 2 : 0;
+
+        // updating of moving mean, variance
+        num_bytes += updateMovingAverage ? invariant_length * sizeof(AccDataType) * 4 : 0;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+    }
+    else
+        (void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+
+        using ReferenceBatchNormFwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormFwd<InOutDataType,
+                                                              InOutDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumReduceDim>;
+
+        auto batchNormFwd_ref = ReferenceBatchNormFwdInstance{};
+
+        auto argument_ptr_ref = batchNormFwd_ref.MakeArgumentPointer(
+            i_inOutLengths,
+            i_inOutStrides,
+            i_inOutStrides,
+            {0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
+            i_scaleBiasMeanVarLengths,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            x.mData.data(),
+            bnScale.mData.data(),
+            bnBias.mData.data(),
+            epsilon,
+            PassThroughOp{},
+            y_ref.mData.data(),
+            saveMeanAndInvVariance ? resultSaveMean_ref.mData.data() : nullptr,
+            saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr,
+            averageFactor,
+            updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr,
+            updateMovingAverage ? resultRunningVariance_ref.mData.data() : nullptr);
+
+        if(!batchNormFwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters seems not supported by the BatchNorm reference "
+                         "instance, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormFwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+
+        y_dev.FromDevice(y.mData.data());
+        pass = pass && ck::utils::check_err(y, y_ref);
+
+        if(updateMovingAverage)
+        {
+            Tensor<AccDataType> resultRunningMean(scaleBiasMeanVarLengths);
+            Tensor<AccDataType> resultRunningVariance(scaleBiasMeanVarLengths);
+
+            resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
+            resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
+
+            pass = pass && ck::utils::check_err(resultRunningMean, resultRunningMean_ref);
+            pass = pass && ck::utils::check_err(resultRunningVariance, resultRunningVariance_ref);
+        };
+
+        if(saveMeanAndInvVariance)
+        {
+            using ck::host_common::dumpBufferToFile;
+
+            Tensor<AccDataType> resultSaveMean(scaleBiasMeanVarLengths);
+            Tensor<AccDataType> resultSaveInvVariance(scaleBiasMeanVarLengths);
+
+            resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
+            resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
+
+            pass = pass && ck::utils::check_err(resultSaveMean, resultSaveMean_ref);
+            pass = pass && ck::utils::check_err(resultSaveInvVariance, resultSaveInvVariance_ref);
+        };
+    };
+
+    return (pass);
+};
+
+const double epsilon              = std::numeric_limits<float>::epsilon();
+static const double averageFactor = 0.1;
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        BatchNormFwdArg arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
+                                                                    arg.init_method,
+                                                                    arg.time_kernel,
+                                                                    arg.inOutLengths,
+                                                                    arg.updateMovingAverage,
+                                                                    arg.saveMeanAndInvVariance,
+                                                                    averageFactor,
+                                                                    epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.updateMovingAverage,
+                                                                     arg.saveMeanAndInvVariance,
+                                                                     averageFactor,
+                                                                     epsilon);
+        }
+        else if(arg.data_type == 1)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<float, float, true>(arg.do_verification,
+                                                               arg.init_method,
+                                                               arg.time_kernel,
+                                                               arg.inOutLengths,
+                                                               arg.updateMovingAverage,
+                                                               arg.saveMeanAndInvVariance,
+                                                               averageFactor,
+                                                               epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<float, float, false>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.updateMovingAverage,
+                                                                arg.saveMeanAndInvVariance,
+                                                                averageFactor,
+                                                                epsilon);
+        }
+        else if(arg.data_type == 3)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<int8_t, float, true>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.updateMovingAverage,
+                                                                arg.saveMeanAndInvVariance,
+                                                                averageFactor,
+                                                                epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<int8_t, float, false>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.updateMovingAverage,
+                                                                 arg.saveMeanAndInvVariance,
+                                                                 averageFactor,
+                                                                 epsilon);
+        }
+        else if(arg.data_type == 5)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.updateMovingAverage,
+                                                                     arg.saveMeanAndInvVariance,
+                                                                     averageFactor,
+                                                                     epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
+                                                                      arg.init_method,
+                                                                      arg.time_kernel,
+                                                                      arg.inOutLengths,
+                                                                      arg.updateMovingAverage,
+                                                                      arg.saveMeanAndInvVariance,
+                                                                      averageFactor,
+                                                                      epsilon);
+        }
+        else if(arg.data_type == 6)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<double, double, true>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.updateMovingAverage,
+                                                                 arg.saveMeanAndInvVariance,
+                                                                 averageFactor,
+                                                                 epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<double, double, false>(arg.do_verification,
+                                                                  arg.init_method,
+                                                                  arg.time_kernel,
+                                                                  arg.inOutLengths,
+                                                                  arg.updateMovingAverage,
+                                                                  arg.saveMeanAndInvVariance,
+                                                                  averageFactor,
+                                                                  epsilon);
+        }
+    }
+    else
+    {
+        pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(true,
+                                                            2,
+                                                            false, // don't time kernel
+                                                            {128, 16, 6, 512},
+                                                            true,
+                                                            true,
+                                                            averageFactor,
+                                                            epsilon);
+
+        pass = pass && bnorm_fwd_nhwc_test<ck::half_t, float, false>(true,
+                                                                     2,
+                                                                     false, // don't time kernel
+                                                                     {128, 16, 3, 1024},
+                                                                     true,
+                                                                     true,
+                                                                     averageFactor,
+                                                                     epsilon);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_infer_impl.hpp b/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_infer_impl.hpp
new file mode 100644
index 0000000..e457df8
--- /dev/null
+++ b/3rdparty/composable_kernel/example/34_batchnorm/batchnorm_infer_impl.hpp
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cassert>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+
+#include "batchnorm_common.hpp"
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          ck::index_t Rank,
+          ck::index_t NumBatchNormReduceDim,
+          bool fastest_dim_is_reduced = false>
+int bnorm_infer(
+    bool time_kernel,
+    const std::array<int, NumBatchNormReduceDim> reduceDims,
+    const std::array<ck::index_t, Rank> xyLengths,
+    const std::array<ck::index_t, Rank> xStrides,
+    const std::array<ck::index_t, Rank> yStrides,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+    const void* p_x,
+    const void* p_scale,
+    const void* p_bias,
+    double epsilon,
+    const void* p_estimatedMean,
+    const void* p_estimatedVariance,
+    void* p_y)
+{
+    (void)bnScaleBiasMeanVarLengths;
+
+    static_assert(NumBatchNormReduceDim < Rank,
+                  "Invalid number of reduced dimensions for batchnorm!");
+
+    using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<XDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean,
+                                                                                  // variance,
+                                                                                  // scale,
+                                                                                  // bias,
+        ck::Tuple<YDataType>,                                                     // y
+        NormalizeInInfer,
+        Rank,
+        2,                           // MPerthread
+        ck::Sequence<1, 1, 1, 1, 1>, // x, mean, variance, scale, bias
+        ck::Sequence<1>>;            // scalarPerVector: y
+
+    auto invariantDims = get_invariant_dims<Rank, NumBatchNormReduceDim>(reduceDims);
+    std::array<ck::index_t, Rank> aligned_bnScaleStrides{0};
+    std::array<ck::index_t, Rank> aligned_bnBiasStrides{0};
+    std::array<ck::index_t, Rank> aligned_bnMeanVarStrides{0};
+
+    int i = 0;
+    for(auto dim : invariantDims)
+    {
+        assert(xyLengths[dim] == bnScaleBiasMeanVarLengths[i]);
+
+        aligned_bnScaleStrides[dim]   = bnScaleStrides[i];
+        aligned_bnBiasStrides[dim]    = bnBiasStrides[i];
+        aligned_bnMeanVarStrides[dim] = bnMeanVarStrides[i];
+        i++;
+    };
+
+    int32_t reduceLength = 1;
+
+    for(auto dim : reduceDims)
+        reduceLength *= xyLengths[dim];
+
+    int32_t invariantLength = 1;
+
+    for(auto dim : invariantDims)
+        invariantLength *= xyLengths[dim];
+
+    size_t total_length = static_cast<size_t>(invariantLength) * reduceLength;
+
+    float avg_time        = 0.0f;
+    std::size_t num_bytes = 0;
+
+    auto dev_normalize = DeviceNormalizeInstance{};
+
+    auto argument_ptr1 = dev_normalize.MakeArgumentPointer(
+        xyLengths,
+        {xStrides,
+         aligned_bnMeanVarStrides,
+         aligned_bnMeanVarStrides,
+         aligned_bnScaleStrides,
+         aligned_bnBiasStrides},
+        {yStrides},
+        {p_x, p_estimatedMean, p_estimatedVariance, p_scale, p_bias},
+        {p_y},
+        NormalizeInInfer{epsilon});
+
+    if(!dev_normalize.IsSupportedArgument(argument_ptr1.get()))
+    {
+        std::cout << "The runtime parameters seems not supported by the Devic, exiting!"
+                  << std::endl;
+
+        return (-1);
+    };
+
+    auto invoker_ptr1 = dev_normalize.MakeInvokerPointer();
+
+    avg_time += invoker_ptr1->Run(argument_ptr1.get(), StreamConfig{nullptr, time_kernel});
+
+    num_bytes += total_length * sizeof(XDataType) +
+                 invariantLength *
+                     (sizeof(ScaleDataType) + sizeof(BiasDataType) + 2 * sizeof(MeanVarDataType)) +
+                 total_length * sizeof(YDataType);
+
+    if(time_kernel)
+    {
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+    };
+
+    return (0);
+};
diff --git a/3rdparty/composable_kernel/example/35_splitK_gemm/CMakeLists.txt b/3rdparty/composable_kernel/example/35_splitK_gemm/CMakeLists.txt
new file mode 100644
index 0000000..7945839
--- /dev/null
+++ b/3rdparty/composable_kernel/example/35_splitK_gemm/CMakeLists.txt
@@ -0,0 +1,17 @@
+add_custom_target(example_splitK_gemm_xdl)
+
+add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
+add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
+add_example_executable(example_splitK_gemm_xdl_bfp16 splitK_gemm_xdl_bfp16.cpp)
+add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
+
+add_dependencies(example_splitK_gemm_xdl
+                 example_splitK_gemm_xdl_fp32
+                 example_splitK_gemm_xdl_fp16
+                 example_splitK_gemm_xdl_bfp16
+                 example_splitK_gemm_xdl_int8)
+
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp)
+  add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)
+endif()
diff --git a/3rdparty/composable_kernel/example/35_splitK_gemm/run_splitK_gemm_example.inc b/3rdparty/composable_kernel/example/35_splitK_gemm/run_splitK_gemm_example.inc
new file mode 100644
index 0000000..e9bd5c5
--- /dev/null
+++ b/3rdparty/composable_kernel/example/35_splitK_gemm/run_splitK_gemm_example.inc
@@ -0,0 +1,217 @@
+#pragma once
+
+struct ProblemSize final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t stride_A = K;
+    ck::index_t stride_B = K;
+    ck::index_t stride_C = N;
+
+    ck::index_t k_batch = 4;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    using namespace ck::literals;
+
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+    static_assert(sizeof(ADataType) == sizeof(KernelADataType));
+    static_assert(sizeof(BDataType) == sizeof(KernelBDataType));
+#endif
+
+    auto& [M, N, K, StrideA, StrideB, StrideC, KBatch] = problem_size;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+        b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+    }
+
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<KernelADataType> a_m_k_converted(a_m_k);
+    const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
+
+    a_m_k_device_buf.ToDevice(a_m_k_converted.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n_converted.mData.data());
+#else
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+#endif
+    c_m_n_device_buf.SetZero();
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(
+#ifdef BUILD_INT4_EXAMPLE
+        static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+#else
+        static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+#endif
+        static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+        M,
+        N,
+        K,
+        StrideA,
+        StrideB,
+        StrideC,
+        a_element_op,
+        b_element_op,
+        c_element_op,
+        KBatch);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+    bool pass = true;
+
+    if(config.do_verification)
+    {
+        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        if(std::is_same<CDataType, ck::half_t>::value)
+        {
+            pass &= ck::utils::check_err(
+                c_m_n_device_result, c_m_n_host_result, "fp16 incorrect result", 3e-3, 1e-3);
+        }
+        else
+        {
+            pass &= ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+        }
+    }
+
+    if(config.time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+        std::size_t flop = std::size_t(2) * M * N * K;
+        std::size_t num_btype =
+            sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+        float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    }
+
+    return pass;
+}
+
+bool run_splitK_gemm_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 5)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+        problem_size.k_batch   = std::stoi(argv[4]);
+    }
+    else if(argc == 11)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+        problem_size.k_batch   = std::stoi(argv[4]);
+
+        problem_size.M = std::stoi(argv[5]);
+        problem_size.N = std::stoi(argv[6]);
+        problem_size.K = std::stoi(argv[7]);
+
+        problem_size.stride_A = std::stoi(argv[8]);
+        problem_size.stride_B = std::stoi(argv[9]);
+        problem_size.stride_C = std::stoi(argv[10]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4: KBatch\n");
+        printf("arg5 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
+        exit(0);
+    }
+
+    return run_splitK_gemm(problem_size, config);
+}
diff --git a/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_bfp16.cpp b/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_bfp16.cpp
new file mode 100644
index 0000000..7191ecf
--- /dev/null
+++ b/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_bfp16.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = BF16;
+using BDataType   = BF16;
+using AccDataType = F32;
+using CDataType   = F32;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
+    // clang-format off
+//######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+//######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+//######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   256,   128,     4,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               4>;
+// clang-format on
+
+#include "run_splitK_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp b/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp
new file mode 100644
index 0000000..efdb315
--- /dev/null
+++ b/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = F32;
+using CDataType   = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
+    // clang-format off
+//######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+//######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+//######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   256,   128,     4,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#include "run_splitK_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp b/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp
new file mode 100644
index 0000000..bc2e3d1
--- /dev/null
+++ b/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = F32;
+using BDataType   = F32;
+using AccDataType = F32;
+using CDataType   = F32;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
+    // clang-format off
+//######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+//######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+//######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   256,   128,     4,   4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>;
+// clang-format on
+
+#include "run_splitK_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp b/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp
new file mode 100644
index 0000000..4eb2782
--- /dev/null
+++ b/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp
@@ -0,0 +1,92 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = ck::int4_t;
+using BDataType   = ck::int4_t;
+using AccDataType = int32_t;
+using CDataType   = int32_t;
+
+using KernelADataType = int8_t;
+using KernelBDataType = int8_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
+    // clang-format off
+        <KernelADataType,      //ADataType    
+         KernelBDataType,      //BDataType   
+         CDataType,            //EDataType
+         AccDataType,          //AccDataType
+         ALayout,              //ALayout
+         BLayout,              //BLayout
+         CLayout,              //ELayout
+         AElementOp,           //AElementwiseOperation
+         BElementOp,           //BElementwiseOperation
+         CElementOp,           //CElementwiseOperation
+         GemmDefault,          //GEMMSpecialization
+         256,                  // BlockSize
+         256,                  // MPerBlock
+         128,                  // NPerBlock
+         4,                    // KPerBlock
+         16,                   // K1
+         32,                   // MPerXdl
+         32,                   // NPerXdl
+         4,                    // MXdlPerWave
+         2,                    // NXdlPerWave
+         S<1, 4, 64, 1>,       // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<0, 2, 1, 3>,        // ABlockTransfer ThreadCluster ArrangeOrder
+         S<0, 2, 1, 3>,        // ABlockTransfer SrcAccessOrder
+         3,                    // ABlockTransfer SrcVectorDim
+         16,                   // ABlockTransfer SrcScalarPerVector
+         16,                   // ABlockTransfer DstScalarPerVector_K1
+         true,                 // ABlockLdsExtraM
+         S<1, 4, 64, 1>,       // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<0, 1, 3, 2>,        // BBlockTransfer ThreadCluster ArrangeOrder
+         S<0, 1, 3, 2>,        // BBlockTransfer SrcAccessOrder
+         3,                    // BBlockTransfer SrcVectorDim
+         16,                   // BBlockTransfer SrcScalarPerVector
+         16,                   // BBlockTransfer DstScalarPerVector_K1
+         true,                 // BBlockLdsExtraN
+         1,                    // CShuffleMXdlPerWavePerShuffle
+         1,                    // CShuffleNXdlPerWavePerShuffle
+         S<1, 32, 1, 8>,       // CBlockTransferClusterLengths _MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+         4>;                   // CBlockTransferScalarPerVector_NWaveNPerXdl
+// clang-format on
+
+#define BUILD_INT4_EXAMPLE
+#include "run_splitK_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp b/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp
new file mode 100644
index 0000000..eefdbca
--- /dev/null
+++ b/3rdparty/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp
@@ -0,0 +1,55 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = int8_t;
+using BDataType   = int8_t;
+using AccDataType = int32_t;
+using CDataType   = int32_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
+    // clang-format off
+//######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+//######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+//######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   256,   128,     4,  16,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,             16,             16,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,             16,             16,      true,           1,           1,                   S<1, 32, 1, 8>,               4>;
+// clang-format on
+
+#include "run_splitK_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/36_sparse_embedding/CMakeLists.txt b/3rdparty/composable_kernel/example/36_sparse_embedding/CMakeLists.txt
new file mode 100644
index 0000000..16ca6ea
--- /dev/null
+++ b/3rdparty/composable_kernel/example/36_sparse_embedding/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_sparse_embedding3_forward_layernorm sparse_embedding3_forward_layernorm.cpp)
diff --git a/3rdparty/composable_kernel/example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp b/3rdparty/composable_kernel/example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
new file mode 100644
index 0000000..f5eb4c3
--- /dev/null
+++ b/3rdparty/composable_kernel/example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
@@ -0,0 +1,219 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+#include <ctime>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_sparse_embedding3_forward_layernorm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp"
+
+// using EmbType       = float;
+// using IndexType     = int64_t;
+// using GammaDataType = float;
+// using BetaDataType  = float;
+// using AccDataType   = float;
+// using OutType       = float;
+
+using EmbType       = ck::half_t;
+using IndexType     = int64_t;
+using GammaDataType = ck::half_t;
+using BetaDataType  = ck::half_t;
+using AccDataType   = float;
+using OutType       = ck::half_t;
+
+// clang-format off
+//                                                                                                         BlockSize, DimClusterSize, RowClusterSize, DimPerBlock, RowPerBlock, DimThreadSize, RowVectorSize
+using DeviceInstance_fp32_e256   = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  256,   1, 1>;
+using DeviceInstance_fp32_e512   = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  512,   1, 1>;
+using DeviceInstance_fp32_e768   = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  768,   1, 1>;
+using DeviceInstance_fp32_e1024  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  1024,  1, 1>;
+using DeviceInstance_fp32_e1536  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  1536,  1, 1>;
+using DeviceInstance_fp32_e2048  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  2048,  1, 4>;
+using DeviceInstance_fp32_e4096  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  4096,  1, 4>;
+using DeviceInstance_fp32_e8192  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  8192,  1, 4>;
+using DeviceInstance_fp32_e16384 = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  16384, 1, 4>;
+
+using DeviceInstance_fp16_e256   = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  256,   1, 1>;
+using DeviceInstance_fp16_e512   = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  512,   1, 2>;
+using DeviceInstance_fp16_e768   = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  768,   1, 1>;
+using DeviceInstance_fp16_e1024  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  1024,  1, 2>;
+using DeviceInstance_fp16_e1536  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  1536,  1, 2>;
+using DeviceInstance_fp16_e2048  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  2048,  1, 2>;
+using DeviceInstance_fp16_e4096  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  4096,  1, 8>;
+using DeviceInstance_fp16_e8192  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  8192,  1, 8>;
+
+template<typename emb_type, ck::index_t dim> struct emb_kernel{};
+
+template<> struct emb_kernel<float, 256>  { using kernel_type = DeviceInstance_fp32_e256; };
+template<> struct emb_kernel<float, 512>  { using kernel_type = DeviceInstance_fp32_e512; };
+template<> struct emb_kernel<float, 768>  { using kernel_type = DeviceInstance_fp32_e768; };
+template<> struct emb_kernel<float, 1024> { using kernel_type = DeviceInstance_fp32_e1024;};
+template<> struct emb_kernel<float, 1536> { using kernel_type = DeviceInstance_fp32_e1536;};
+template<> struct emb_kernel<float, 2048> { using kernel_type = DeviceInstance_fp32_e2048;};
+template<> struct emb_kernel<float, 4096> { using kernel_type = DeviceInstance_fp32_e4096;};
+template<> struct emb_kernel<float, 8192> { using kernel_type = DeviceInstance_fp32_e8192;};
+template<> struct emb_kernel<float, 16384>{ using kernel_type = DeviceInstance_fp32_e16384;};
+
+template<> struct emb_kernel<ck::half_t, 256>  { using kernel_type = DeviceInstance_fp16_e256; };
+template<> struct emb_kernel<ck::half_t, 512>  { using kernel_type = DeviceInstance_fp16_e512; };
+template<> struct emb_kernel<ck::half_t, 768>  { using kernel_type = DeviceInstance_fp16_e768; };
+template<> struct emb_kernel<ck::half_t, 1024> { using kernel_type = DeviceInstance_fp16_e1024; };
+template<> struct emb_kernel<ck::half_t, 1536> { using kernel_type = DeviceInstance_fp16_e1536; };
+template<> struct emb_kernel<ck::half_t, 2048> { using kernel_type = DeviceInstance_fp16_e2048; };
+template<> struct emb_kernel<ck::half_t, 4096> { using kernel_type = DeviceInstance_fp16_e4096; };
+template<> struct emb_kernel<ck::half_t, 8192> { using kernel_type = DeviceInstance_fp16_e8192; };
+
+// clang-format on
+
+int main()
+{
+    bool time_kernel = true;
+
+    constexpr auto num_rows = 65536;
+    constexpr auto dims     = ck::Sequence<256, 512, 768, 1024, 1536, 2048, 4096, 8192>{};
+    // constexpr auto dims = ck::Sequence<256, 512>{};
+    constexpr auto index_length   = 2048;
+    constexpr AccDataType epsilon = 1e-4;
+
+    auto f_host_tensor_desc_1d = [](std::size_t len_) { return HostTensorDescriptor({len_}); };
+
+    auto f_host_tensor_desc_2d = [](std::size_t rows_, std::size_t cols_) {
+        return HostTensorDescriptor({rows_, cols_});
+    };
+
+    using ReferenceInstance =
+        ck::tensor_operation::host::ReferenceSparseEmbedding3ForwardLayernorm<EmbType,
+                                                                              IndexType,
+                                                                              GammaDataType,
+                                                                              BetaDataType,
+                                                                              AccDataType,
+                                                                              OutType>;
+
+    ck::static_for<0, dims.Size(), 1>{}([&](auto I) {
+        std::srand(std::time(nullptr));
+        constexpr auto current_dim = dims.At(I);
+        Tensor<EmbType> emb_a(f_host_tensor_desc_2d(num_rows, current_dim));
+        Tensor<EmbType> emb_b(f_host_tensor_desc_2d(num_rows, current_dim));
+        Tensor<EmbType> emb_c(f_host_tensor_desc_2d(num_rows, current_dim));
+
+        Tensor<IndexType> index_a(f_host_tensor_desc_1d(index_length));
+        Tensor<IndexType> index_b(f_host_tensor_desc_1d(index_length));
+        Tensor<IndexType> index_c(f_host_tensor_desc_1d(index_length));
+
+        Tensor<GammaDataType> gamma(f_host_tensor_desc_1d(current_dim));
+        Tensor<BetaDataType> beta(f_host_tensor_desc_1d(current_dim));
+
+        Tensor<OutType> out(f_host_tensor_desc_2d(index_length, current_dim));
+
+        emb_a.GenerateTensorValue(GeneratorTensor_3<EmbType>{0.0, 1.0});
+        emb_b.GenerateTensorValue(GeneratorTensor_3<EmbType>{0.0, 1.0});
+        emb_c.GenerateTensorValue(GeneratorTensor_3<EmbType>{0.0, 1.0});
+
+        index_a.GenerateTensorValue(GeneratorTensor_2<IndexType>{0, num_rows});
+        index_b.GenerateTensorValue(GeneratorTensor_2<IndexType>{0, num_rows});
+        index_c.GenerateTensorValue(GeneratorTensor_2<IndexType>{0, num_rows});
+
+        gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
+        beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{0.0, 1.0});
+
+        DeviceMem emb_a_dev(sizeof(EmbType) * emb_a.mDesc.GetElementSpaceSize());
+        DeviceMem emb_b_dev(sizeof(EmbType) * emb_b.mDesc.GetElementSpaceSize());
+        DeviceMem emb_c_dev(sizeof(EmbType) * emb_c.mDesc.GetElementSpaceSize());
+
+        DeviceMem index_a_dev(sizeof(IndexType) * index_a.mDesc.GetElementSpaceSize());
+        DeviceMem index_b_dev(sizeof(IndexType) * index_b.mDesc.GetElementSpaceSize());
+        DeviceMem index_c_dev(sizeof(IndexType) * index_c.mDesc.GetElementSpaceSize());
+
+        DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+        DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+
+        DeviceMem out_dev(sizeof(OutType) * out.mDesc.GetElementSpaceSize());
+
+        emb_a_dev.ToDevice(emb_a.mData.data());
+        emb_b_dev.ToDevice(emb_b.mData.data());
+        emb_c_dev.ToDevice(emb_c.mData.data());
+
+        index_a_dev.ToDevice(index_a.mData.data());
+        index_b_dev.ToDevice(index_b.mData.data());
+        index_c_dev.ToDevice(index_c.mData.data());
+
+        gamma_dev.ToDevice(gamma.mData.data());
+        beta_dev.ToDevice(beta.mData.data());
+
+        auto device_instance = typename emb_kernel<EmbType, current_dim>::kernel_type{};
+        auto argument_ptr    = device_instance.MakeArgumentPointer(out_dev.GetDeviceBuffer(),
+                                                                emb_a_dev.GetDeviceBuffer(),
+                                                                emb_b_dev.GetDeviceBuffer(),
+                                                                emb_c_dev.GetDeviceBuffer(),
+                                                                index_a_dev.GetDeviceBuffer(),
+                                                                index_b_dev.GetDeviceBuffer(),
+                                                                index_c_dev.GetDeviceBuffer(),
+                                                                gamma_dev.GetDeviceBuffer(),
+                                                                beta_dev.GetDeviceBuffer(),
+                                                                num_rows,
+                                                                current_dim,
+                                                                index_length,
+                                                                epsilon);
+        std::cout << "Dim:" << current_dim << ", kernel:" << device_instance.GetTypeString()
+                  << std::endl
+                  << std::flush;
+
+        bool is_supported = device_instance.IsSupportedArgument(argument_ptr.get());
+
+        if(!is_supported)
+        {
+            std::cout << "Runtime parameters are not supported" << std::endl;
+            return;
+        }
+
+        auto invoker_ptr = device_instance.MakeInvokerPointer();
+        float time_ms    = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        bool pass = true;
+        {
+            Tensor<OutType> out_from_dev(f_host_tensor_desc_2d(index_length, current_dim));
+            ReferenceInstance ref;
+            auto ref_argument = ref.MakeArgument(out,
+                                                 emb_a,
+                                                 emb_b,
+                                                 emb_c,
+                                                 index_a,
+                                                 index_b,
+                                                 index_c,
+                                                 gamma,
+                                                 beta,
+                                                 num_rows,
+                                                 current_dim,
+                                                 index_length,
+                                                 epsilon);
+            auto ref_invoker  = ref.MakeInvoker();
+            ref_invoker.Run(ref_argument);
+
+            out_dev.FromDevice(out_from_dev.mData.data());
+            pass &= ck::utils::check_err(out_from_dev, out, "Error: Incorrect results", 1e-3, 1e-3);
+        }
+
+        double total_read = current_dim * index_length * 3 * sizeof(EmbType) +
+                            current_dim * sizeof(GammaDataType) +
+                            current_dim * sizeof(BetaDataType);
+        double total_write = current_dim * index_length * sizeof(OutType);
+        double gbps        = (total_read + total_write) / time_ms / 1e6;
+
+        std::cout << ", total bytes:" << (total_read + total_write) << ", time:" << time_ms
+                  << ", gbps:" << gbps << ", valid:" << (pass ? "y" : "n") << std::endl
+                  << std::flush;
+    });
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt b/3rdparty/composable_kernel/example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
new file mode 100644
index 0000000..a9be3a7
--- /dev/null
+++ b/3rdparty/composable_kernel/example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp)
diff --git a/3rdparty/composable_kernel/example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp b/3rdparty/composable_kernel/example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
new file mode 100644
index 0000000..071e8a7
--- /dev/null
+++ b/3rdparty/composable_kernel/example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
@@ -0,0 +1,519 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Computes C_m_o = Relu(A0[m, k] * B0[n, k] + D00[m, n] + D01[mn]) * B1[n, o] + D1[m, o]
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using A0DataType        = F16;
+using B0DataType        = F16;
+using Acc0DataType      = F32;
+using D00DataType       = F16;
+using D01DataType       = F16;
+using B1DataType        = F16;
+using Acc1DataType      = F32;
+using C1ShuffleDataType = F32;
+using D1DataType        = F16;
+using E1DataType        = F16;
+
+using A0Layout  = Row;
+using B0Layout  = Col;
+using D00Layout = Row;
+using D01Layout = Row;
+using B1Layout  = Row;
+using D1Layout  = Row;
+using E1Layout  = Row;
+
+// E = Relu(C + D0 + D1)
+struct AddAddRelu
+{
+    __host__ __device__ void
+    operator()(ck::half_t& e, const ck::half_t& c, const ck::half_t& d0, const ck::half_t& d1) const
+    {
+        const ck::half_t x = c + d0 + d1;
+
+        ck::tensor_operation::element_wise::Relu{}.template operator()<ck::half_t>(e, x);
+    }
+    __host__ __device__ void
+    operator()(float& e, const float& c, const ck::half_t& d0, const ck::half_t& d1) const
+    {
+        const float x = c + (d0 + d1);
+
+        ck::tensor_operation::element_wise::Relu{}.template operator()<float>(e, x);
+    }
+};
+
+// E = Gelu(C + D0 + D1)
+struct AddAddGelu
+{
+    __host__ __device__ void
+    operator()(ck::half_t& e, const ck::half_t& c, const ck::half_t& d0, const ck::half_t& d1) const
+    {
+        const ck::half_t x = c + d0 + d1;
+
+        ck::tensor_operation::element_wise::Gelu{}.template operator()<ck::half_t, ck::half_t>(e,
+                                                                                               x);
+    }
+
+    __host__ __device__ void
+    operator()(float& e, const float& c, const ck::half_t& d0, const ck::half_t& d1) const
+    {
+        const float x = c + (d0 + d1);
+
+        ck::tensor_operation::element_wise::Gelu{}.template operator()<float, float>(e, x);
+    }
+};
+
+// E = FastGelu(C + D0 + D1)
+struct AddAddFastGelu
+{
+    __host__ __device__ void
+    operator()(float& e, const float& c, const ck::half_t& d0, const ck::half_t& d1) const
+    {
+        const float x = c + (d0 + d1);
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(e, x);
+    }
+};
+
+using A0ElementOp   = PassThrough;
+using B0ElementOp   = PassThrough;
+using CDE0ElementOp = AddAddRelu;
+using A1ElementOp   = PassThrough;
+using B1ElementOp   = PassThrough;
+using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr bool PadGemm0M = false;
+static constexpr bool PadGemm0N = false;
+static constexpr bool PadGemm0K = false;
+static constexpr bool PadGemm1N = false;
+static constexpr bool PadGemm1K = false;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<
+        A0Layout,
+        B0Layout,
+        ck::Tuple<D00Layout, D01Layout>,
+        B1Layout,
+        ck::Tuple<D1Layout>,
+        E1Layout,
+        A0DataType,
+        B0DataType,
+        Acc0DataType,
+        ck::Tuple<D00DataType, D01DataType>,
+        B1DataType,
+        Acc1DataType,
+        C1ShuffleDataType,
+        ck::Tuple<D1DataType>,
+        E1DataType,
+        A0ElementOp,
+        B0ElementOp,
+        CDE0ElementOp,
+        B1ElementOp,
+        CDE1ElementOp,
+        PadGemm0M,
+        PadGemm0N,
+        PadGemm0K,
+        PadGemm1N,
+        PadGemm1K,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<8, 32, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M              = 1024;
+    ck::index_t N              = 1024;
+    ck::index_t K              = 64;
+    ck::index_t O              = 128;
+    ck::index_t BatchCount     = 4;
+    ck::index_t StrideA0       = -1;
+    ck::index_t StrideB0       = -1;
+    ck::index_t StrideD00      = -1;
+    ck::index_t StrideD01      = -1;
+    ck::index_t StrideB1       = -1;
+    ck::index_t StrideD1       = -1;
+    ck::index_t StrideE1       = -1;
+    ck::index_t BatchStrideA0  = -1;
+    ck::index_t BatchStrideB0  = -1;
+    ck::index_t BatchStrideD00 = -1;
+    ck::index_t BatchStrideD01 = -1;
+    ck::index_t BatchStrideB1  = -1;
+    ck::index_t BatchStrideD1  = -1;
+    ck::index_t BatchStrideE1  = -1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 9)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+    }
+    else if(argc == 23)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+
+        StrideA0  = std::stoi(argv[9]);
+        StrideB0  = std::stoi(argv[10]);
+        StrideD00 = std::stoi(argv[11]);
+        StrideD01 = std::stoi(argv[12]);
+        StrideB1  = std::stoi(argv[13]);
+        StrideD1  = std::stoi(argv[14]);
+        StrideE1  = std::stoi(argv[15]);
+
+        BatchStrideA0  = std::stoi(argv[16]);
+        BatchStrideB0  = std::stoi(argv[17]);
+        BatchStrideD00 = std::stoi(argv[18]);
+        BatchStrideD01 = std::stoi(argv[19]);
+        BatchStrideB1  = std::stoi(argv[20]);
+        BatchStrideD1  = std::stoi(argv[21]);
+        BatchStrideE1  = std::stoi(argv[22]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 8: M, N, K, O, Batch\n");
+        printf(
+            "arg9 to 15: StrideA0, StrideB0, StrideD00, StrideD01, StrideB1, StrideD1, StrideE1\n");
+        printf("arg16 to 22: BatchStrideA0, BatchStrideB0, BatchStrideD00, BatchStrideD01, "
+               "BatchStrideB1, BatchStrideD1, BatchStrideE1 \n");
+        exit(0);
+    }
+
+    const int DefaultStrideA0  = ck::is_same_v<A0Layout, Row> ? K : M;
+    const int DefaultStrideB0  = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideD00 = ck::is_same_v<D00Layout, Row> ? N : M;
+    const int DefaultStrideD01 = ck::is_same_v<D01Layout, Row> ? N : M;
+    const int DefaultStrideB1  = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideD1  = ck::is_same_v<D1Layout, Row> ? O : M;
+    const int DefaultStrideE1  = ck::is_same_v<E1Layout, Row> ? O : M;
+
+    StrideA0  = (StrideA0 < 0) ? DefaultStrideA0 : StrideA0;
+    StrideB0  = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideD00 = (StrideD00 < 0) ? DefaultStrideD00 : StrideD00;
+    StrideD01 = (StrideD01 < 0) ? DefaultStrideD01 : StrideD01;
+    StrideB1  = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideD1  = (StrideD1 < 0) ? DefaultStrideD1 : StrideD1;
+    StrideE1  = (StrideE1 < 0) ? DefaultStrideE1 : StrideE1;
+
+    const int DefaultBatchStrideA0  = (ck::is_same_v<A0Layout, Col> ? K : M) * StrideA0;
+    const int DefaultBatchStrideB0  = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideD00 = (ck::is_same_v<D00Layout, Col> ? N : M) * StrideD00;
+    const int DefaultBatchStrideD01 = (ck::is_same_v<D01Layout, Col> ? N : M) * StrideD01;
+    const int DefaultBatchStrideB1  = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideD1  = (ck::is_same_v<D1Layout, Col> ? O : M) * StrideD1;
+    const int DefaultBatchStrideE1  = (ck::is_same_v<E1Layout, Col> ? O : M) * StrideE1;
+
+    BatchStrideA0  = BatchStrideA0 < 0 ? DefaultBatchStrideA0 : BatchStrideA0;
+    BatchStrideB0  = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideD00 = BatchStrideD00 < 0 ? DefaultBatchStrideD00 : BatchStrideD00;
+    BatchStrideD01 = BatchStrideD01 < 0 ? DefaultBatchStrideD01 : BatchStrideD01;
+    BatchStrideB1  = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideD1  = BatchStrideD1 < 0 ? DefaultBatchStrideD1 : BatchStrideD1;
+    BatchStrideE1  = BatchStrideE1 < 0 ? DefaultBatchStrideE1 : BatchStrideE1;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    // E_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<A0DataType> a0_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA0, BatchStrideA0, A0Layout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<D00DataType> d00_g_m_n(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideD00, BatchStrideD00, D00Layout{}));
+    Tensor<D01DataType> d01_g_m_n(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideD01, BatchStrideD01, D01Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<D1DataType> d1_g_m_o(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideD1, BatchStrideD1, D1Layout{}));
+    Tensor<E1DataType> e1_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideE1, BatchStrideE1, E1Layout{}));
+    Tensor<E1DataType> e1_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideE1, BatchStrideE1, E1Layout{}));
+
+    std::cout << "a0_g_m_k: " << a0_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "d00_g_m_n: " << d00_g_m_n.mDesc
+              << " size: " << d00_g_m_n.mDesc.GetElementSpaceSize() << std::endl;
+    std::cout << "d01_g_m_n: " << d01_g_m_n.mDesc
+              << " size: " << d01_g_m_n.mDesc.GetElementSpaceSize() << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "e1_g_m_o: " << e1_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_g_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 3});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 3});
+        d00_g_m_n.GenerateTensorValue(GeneratorTensor_2<D00DataType>{-2, 3});
+        d01_g_m_n.GenerateTensorValue(GeneratorTensor_2<D01DataType>{-2, 3});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 3});
+        d1_g_m_o.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-2, 3});
+        break;
+    case 2:
+        a0_g_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        d00_g_m_n.GenerateTensorValue(GeneratorTensor_3<D00DataType>{0.0, 1.0});
+        d01_g_m_n.GenerateTensorValue(GeneratorTensor_3<D01DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        d1_g_m_o.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+        break;
+    default:
+        a0_g_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        d00_g_m_n.GenerateTensorValue(GeneratorTensor_1<D00DataType>{1});
+        d01_g_m_n.GenerateTensorValue(GeneratorTensor_1<D01DataType>{1});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d1_g_m_o.GenerateTensorValue(GeneratorTensor_1<D1DataType>{1});
+    }
+
+    DeviceMem a0_g_m_k_device_buf(sizeof(A0DataType) * a0_g_m_k.mDesc.GetElementSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
+    DeviceMem d00_g_m_n_device_buf(sizeof(D00DataType) * d00_g_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d01_g_m_n_device_buf(sizeof(D01DataType) * d01_g_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
+    DeviceMem e1_g_m_o_device_buf(sizeof(E1DataType) *
+                                  e1_g_m_o_device_result.mDesc.GetElementSize());
+    DeviceMem d1_g_m_o_device_buf(sizeof(D1DataType) * d1_g_m_o.mDesc.GetElementSpaceSize());
+
+    a0_g_m_k_device_buf.ToDevice(a0_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    d00_g_m_n_device_buf.ToDevice(d00_g_m_n.mData.data());
+    d01_g_m_n_device_buf.ToDevice(d01_g_m_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+    d1_g_m_o_device_buf.ToDevice(d1_g_m_o.mData.data());
+
+    auto a0_element_op   = A0ElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto cde0_element_op = CDE0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto cde1_element_op = CDE1ElementOp{};
+
+    // do GEMM
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+    auto argument =
+        gemm.MakeArgument(static_cast<A0DataType*>(a0_g_m_k_device_buf.GetDeviceBuffer()),
+                          static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+                          std::array<const void*, 2>{d00_g_m_n_device_buf.GetDeviceBuffer(),
+                                                     d01_g_m_n_device_buf.GetDeviceBuffer()},
+                          static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+                          std::array<const void*, 1>{d1_g_m_o_device_buf.GetDeviceBuffer()},
+                          static_cast<E1DataType*>(e1_g_m_o_device_buf.GetDeviceBuffer()),
+                          M,
+                          N,
+                          K,
+                          O,
+                          BatchCount,
+                          StrideA0,
+                          StrideB0,
+                          std::array<ck::index_t, 2>{StrideD00, StrideD01},
+                          StrideB1,
+                          std::array<ck::index_t, 1>{StrideD1},
+                          StrideE1,
+                          BatchStrideA0,
+                          BatchStrideB0,
+                          std::array<ck::index_t, 2>{BatchStrideD00, BatchStrideD01},
+                          BatchStrideB1,
+                          std::array<ck::index_t, 1>{BatchStrideD1},
+                          BatchStrideE1,
+                          a0_element_op,
+                          b0_element_op,
+                          cde0_element_op,
+                          b1_element_op,
+                          cde1_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype =
+        (sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(D00DataType) * N +
+         sizeof(D01DataType) * N + sizeof(B1DataType) * N * O + sizeof(E1DataType) * M * O +
+         sizeof(D1DataType) * O) *
+        BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    e1_g_m_o_device_buf.FromDevice(e1_g_m_o_device_result.mData.data());
+
+    if(do_verification)
+    {
+        using ReferenceGemm0Instance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<A0DataType,
+                                                             B0DataType,
+                                                             Acc0DataType,
+                                                             Acc0DataType,
+                                                             A0ElementOp,
+                                                             B0ElementOp,
+                                                             PassThrough>;
+
+        using ReferenceGemm1Instance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<Acc0DataType,
+                                                             B1DataType,
+                                                             Acc1DataType,
+                                                             Acc1DataType,
+                                                             PassThrough,
+                                                             B1ElementOp,
+                                                             PassThrough>;
+
+        // Output of Gemm0 is input A of Gemm1
+        Tensor<Acc0DataType> c0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+        Tensor<Acc0DataType> e0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+        Tensor<Acc1DataType> c1_g_m_o(f_host_tensor_descriptor(BatchCount, M, O, O, M * O, Row{}));
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a0_g_m_k, b0_g_k_n, c0_g_m_n, a0_element_op, b0_element_op, PassThrough{});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        // bias+bias+relu
+        e0_g_m_n.ForEach([&](auto&, auto idx) {
+            cde0_element_op(e0_g_m_n(idx), c0_g_m_n(idx), d00_g_m_n(idx), d01_g_m_n(idx));
+        });
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            e0_g_m_n, b1_g_n_o, c1_g_m_o, PassThrough{}, b1_element_op, PassThrough{});
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // bias
+        e1_g_m_o_host_result.ForEach([&](auto&, auto idx) {
+            cde1_element_op(e1_g_m_o_host_result(idx), c1_g_m_o(idx), d1_g_m_o(idx));
+        });
+
+        return ck::utils::check_err(e1_g_m_o_device_result, e1_g_m_o_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt b/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
new file mode 100644
index 0000000..9cf960c
--- /dev/null
+++ b/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
@@ -0,0 +1,7 @@
+add_custom_target(example_grouped_conv_bwd_data)
+
+add_example_executable(example_grouped_conv_bwd_data_fp16 grouped_conv_bwd_data_fp16.cpp)
+add_example_executable(example_grouped_conv_bwd_data_bias_relu_fp16 grouped_conv_bwd_data_bias_relu_fp16.cpp)
+
+add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_fp16)
+add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_fp16)
diff --git a/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/common.hpp b/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/common.hpp
new file mode 100644
index 0000000..d07ee7b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/common.hpp
@@ -0,0 +1,103 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <initializer_list>
+#include <iostream>
+#include <numeric>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static inline constexpr ck::index_t NDimSpatial = 2;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+using FP16 = ck::half_t;
+using FP32 = float;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+};
+
+#define DefaultConvParams                                                                \
+    ck::utils::conv::ConvParam                                                           \
+    {                                                                                    \
+        NDimSpatial, 32, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, { 1, 1 } \
+    }
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ExecutionConfig& config,
+                           ck::utils::conv::ConvParam& conv_params)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+        config = ExecutionConfig{};
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        conv_params                       = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
diff --git a/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_fp16.cpp b/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_fp16.cpp
new file mode 100644
index 0000000..55ea8c3
--- /dev/null
+++ b/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_fp16.cpp
@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using OutDataType      = FP16;
+using WeiDataType      = FP16;
+using AccDataType      = FP32;
+using CShuffleDataType = FP16;
+using BiasDataType     = FP16; // bias
+using InDataType       = FP16;
+
+using OutLayout  = ck::tensor_layout::convolution::GNHWK;
+using WeiLayout  = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout = ck::Tuple<ck::tensor_layout::convolution::G_C>;
+using InLayout   = ck::tensor_layout::convolution::GNHWC;
+
+using OutElementOp = PassThrough;
+using WeiElementOp = PassThrough;
+using InElementOp  = ck::tensor_operation::element_wise::AddRelu;
+
+// clang-format off
+using DeviceConvInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
+// ######| NDimSpatial|   ALayout|   BLayout|   DsLayout|  ELayout|       AData|       BData|     AccData|         CShuffle|                  DsData|      EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+// ######|            |          |          |           |         |        Type|        Type|        Type|         DataType|                    Type|       Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+// ######|            |          |          |           |         |            |            |            |                 |                        |           |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+// ######|            |          |          |           |         |            |            |            |                 |                        |           |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+         < NDimSpatial, OutLayout, WeiLayout, BiasLayout, InLayout, OutDataType, WeiDataType, AccDataType, CShuffleDataType, ck::Tuple<BiasDataType>, InDataType, OutElementOp, WeiElementOp,    InElementOp,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>;
+// clang-format on
+
+#include "run_grouped_conv_bwd_data_bias_relu_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_bwd_data_bias_relu_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_fp16.cpp b/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_fp16.cpp
new file mode 100644
index 0000000..ddf82ec
--- /dev/null
+++ b/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_fp16.cpp
@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using OutDataType      = FP16;
+using WeiDataType      = FP16;
+using AccDataType      = FP32;
+using CShuffleDataType = FP16;
+using DsDataType       = ck::Tuple<>;
+using InDataType       = FP16;
+
+using OutLayout = ck::tensor_layout::convolution::GNHWK;
+using WeiLayout = ck::tensor_layout::convolution::GKYXC;
+using DsLayout  = ck::Tuple<>;
+using InLayout  = ck::tensor_layout::convolution::GNHWC;
+
+using OutElementOp = PassThrough;
+using WeiElementOp = PassThrough;
+using InElementOp  = PassThrough;
+
+// clang-format off
+using DeviceConvInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
+// ######| NDimSpatial|   ALayout|   BLayout|   DsLayout|  ELayout|       AData|       BData|     AccData|         CShuffle|       DsData|      EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+// ######|            |          |          |           |         |        Type|        Type|        Type|         DataType|         Type|       Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+// ######|            |          |          |           |         |            |            |            |                 |             |           |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+// ######|            |          |          |           |         |            |            |            |                 |             |           |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+         < NDimSpatial, OutLayout, WeiLayout,   DsLayout, InLayout, OutDataType, WeiDataType, AccDataType, CShuffleDataType,   DsDataType, InDataType, OutElementOp, WeiElementOp,    InElementOp,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>;
+// clang-format on
+
+#include "run_grouped_conv_bwd_data_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_bwd_data_example(argc, argv); }
diff --git a/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc b/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc
new file mode 100644
index 0000000..0afd8bd
--- /dev/null
+++ b/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc
@@ -0,0 +1,224 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+bool run_conv_bwd_data_bias_relu(const ExecutionConfig& config,
+                                 const ck::utils::conv::ConvParam& conv_params,
+                                 const HostTensorDescriptor& out_g_n_k_wos_desc,
+                                 const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                                 const HostTensorDescriptor& bias_g_n_c_wis_desc,
+                                 const HostTensorDescriptor& in_g_n_c_wis_desc,
+                                 const OutElementOp& out_element_op,
+                                 const WeiElementOp& wei_element_op,
+                                 const InElementOp& in_element_op)
+{
+    Tensor<OutDataType> out(out_g_n_k_wos_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<BiasDataType> bias(bias_g_n_c_wis_desc);
+    Tensor<InDataType> in_host(in_g_n_c_wis_desc);
+    Tensor<InDataType> in_device(in_g_n_c_wis_desc);
+
+    std::cout << "out: " << out.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "bias: " << bias.mDesc << std::endl;
+    std::cout << "in: " << in_host.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        bias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-5, 5});
+        break;
+    default:
+        out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        bias.GenerateTensorValue(GeneratorTensor_3<BiasDataType>{0.0, 1.0});
+    }
+
+    DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
+
+    out_device_buf.ToDevice(out.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+
+    // reset input to zero
+    in_device_buf.SetZero();
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(out_g_n_k_wos_desc.GetLengths(), a_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), a_g_n_k_wos_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(bias_g_n_c_wis_desc.GetLengths(), d0_g_n_c_wis_lengths);
+    copy(bias_g_n_c_wis_desc.GetStrides(), d0_g_n_c_wis_strides);
+    copy(in_g_n_c_wis_desc.GetLengths(), e_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), e_g_n_c_wis_strides);
+    copy(conv_params.conv_filter_strides_, conv_filter_strides);
+    copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_params.input_left_pads_, input_left_pads);
+    copy(conv_params.input_right_pads_, input_right_pads);
+
+    static_assert(std::is_default_constructible_v<DeviceConvInstance>);
+
+    // do conv
+    auto conv     = DeviceConvInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(
+        out_device_buf.GetDeviceBuffer(),
+        wei_device_buf.GetDeviceBuffer(),
+        std::array<const void*, 1>{bias_device_buf.GetDeviceBuffer()},
+        in_device_buf.GetDeviceBuffer(),
+        a_g_n_k_wos_lengths,
+        a_g_n_k_wos_strides,
+        b_g_k_c_xs_lengths,
+        b_g_k_c_xs_strides,
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{d0_g_n_c_wis_lengths},
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{d0_g_n_c_wis_strides},
+        e_g_n_c_wis_lengths,
+        e_g_n_c_wis_strides,
+        conv_filter_strides,
+        conv_filter_dilations,
+        input_left_pads,
+        input_right_pads,
+        out_element_op,
+        wei_element_op,
+        in_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_conv with the specified compilation parameters does "
+                     "not support this Conv problem"
+                  << std::endl;
+
+        return false;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = conv_params.GetFlops();
+    std::size_t num_btype = conv_params.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    if(config.do_verification)
+    {
+        // c doesn't physically exist, any layout is fine
+        Tensor<float> c_host(in_g_n_c_wis_desc);
+
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
+                                                                         float,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         PassThrough,
+                                                                         WeiElementOp,
+                                                                         OutElementOp>();
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(c_host,
+                                                  wei,
+                                                  out,
+                                                  conv_params.conv_filter_strides_,
+                                                  conv_params.conv_filter_dilations_,
+                                                  conv_params.input_left_pads_,
+                                                  conv_params.input_right_pads_,
+                                                  PassThrough{},
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        in_host.ForEach(
+            [&](auto&, auto idx) { in_element_op(in_host(idx), c_host(idx), bias(idx)); });
+
+        in_device_buf.FromDevice(in_device.mData.data());
+
+        return ck::utils::check_err(in_device, in_host);
+    }
+
+    return true;
+}
+
+int run_grouped_conv_bwd_data_bias_relu_example(int argc, char* argv[])
+{
+    namespace ctc = ck::tensor_layout::convolution;
+
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_params = DefaultConvParams;
+
+    if(!parse_cmd_args(argc, argv, config, conv_params))
+    {
+        return EXIT_FAILURE;
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(conv_params.num_dim_spatial_ != NDimSpatial)
+    {
+        std::cerr << "unsupported # of spatials dimensions" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    // output image: GNHWK
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+            conv_params);
+
+    // weight: GKYXC
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_params);
+
+    // input image bias: G_C
+    const auto bias_g_n_c_wis_desc = HostTensorDescriptor({conv_params.G_,
+                                                           conv_params.N_,
+                                                           conv_params.C_,
+                                                           conv_params.input_spatial_lengths_[0],
+                                                           conv_params.input_spatial_lengths_[1]},
+                                                          {
+                                                              conv_params.C_, // g
+                                                              0,              // n
+                                                              1,              // c
+                                                              0,              // hi
+                                                              0               // wi
+                                                          });
+
+    // input image: GNHWC
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_params);
+
+    return !run_conv_bwd_data_bias_relu(config,
+                                        conv_params,
+                                        out_g_n_k_wos_desc,
+                                        wei_g_k_c_xs_desc,
+                                        bias_g_n_c_wis_desc,
+                                        in_g_n_c_wis_desc,
+                                        wei_element_op,
+                                        out_element_op,
+                                        in_element_op);
+}
diff --git a/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc b/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc
new file mode 100644
index 0000000..e50c98b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc
@@ -0,0 +1,190 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+bool run_conv_bwd_data(const ExecutionConfig& config,
+                       const ck::utils::conv::ConvParam& conv_params,
+                       const HostTensorDescriptor& out_g_n_k_wos_desc,
+                       const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                       const HostTensorDescriptor& in_g_n_c_wis_desc,
+                       const OutElementOp& out_element_op,
+                       const WeiElementOp& wei_element_op,
+                       const InElementOp& in_element_op)
+{
+    Tensor<OutDataType> out(out_g_n_k_wos_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<InDataType> in_host(in_g_n_c_wis_desc);
+    Tensor<InDataType> in_device(in_g_n_c_wis_desc);
+
+    std::cout << "out: " << out.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "in: " << in_host.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
+
+    out_device_buf.ToDevice(out.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+
+    // reset input to zero
+    in_device_buf.SetZero();
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(out_g_n_k_wos_desc.GetLengths(), a_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), a_g_n_k_wos_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(in_g_n_c_wis_desc.GetLengths(), e_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), e_g_n_c_wis_strides);
+    copy(conv_params.conv_filter_strides_, conv_filter_strides);
+    copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_params.input_left_pads_, input_left_pads);
+    copy(conv_params.input_right_pads_, input_right_pads);
+
+    static_assert(std::is_default_constructible_v<DeviceConvInstance>);
+
+    // do conv
+    auto conv     = DeviceConvInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(out_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      std::array<const void*, 0>{},
+                                      in_device_buf.GetDeviceBuffer(),
+                                      a_g_n_k_wos_lengths,
+                                      a_g_n_k_wos_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
+                                      e_g_n_c_wis_lengths,
+                                      e_g_n_c_wis_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      out_element_op,
+                                      wei_element_op,
+                                      in_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_conv with the specified compilation parameters does "
+                     "not support this Conv problem"
+                  << std::endl;
+
+        return false;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = conv_params.GetFlops();
+    std::size_t num_btype = conv_params.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
+                                                                         InDataType,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         PassThrough,
+                                                                         WeiElementOp,
+                                                                         OutElementOp>();
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(in_host,
+                                                  wei,
+                                                  out,
+                                                  conv_params.conv_filter_strides_,
+                                                  conv_params.conv_filter_dilations_,
+                                                  conv_params.input_left_pads_,
+                                                  conv_params.input_right_pads_,
+                                                  PassThrough{},
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        in_device_buf.FromDevice(in_device.mData.data());
+
+        return ck::utils::check_err(in_device.mData, in_host.mData);
+    }
+
+    return true;
+}
+
+int run_grouped_conv_bwd_data_example(int argc, char* argv[])
+{
+    namespace ctc = ck::tensor_layout::convolution;
+
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_params = DefaultConvParams;
+
+    if(!parse_cmd_args(argc, argv, config, conv_params))
+    {
+        return EXIT_FAILURE;
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(conv_params.num_dim_spatial_ != NDimSpatial)
+    {
+        std::cerr << "unsupported # of spatials dimensions" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    // output image: GNHWK
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+            conv_params);
+
+    // weight: GKYXC
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_params);
+
+    // input image: GNHWC
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_params);
+
+    return !run_conv_bwd_data(config,
+                              conv_params,
+                              out_g_n_k_wos_desc,
+                              wei_g_k_c_xs_desc,
+                              in_g_n_c_wis_desc,
+                              wei_element_op,
+                              out_element_op,
+                              in_element_op);
+}
diff --git a/3rdparty/composable_kernel/example/39_permute/CMakeLists.txt b/3rdparty/composable_kernel/example/39_permute/CMakeLists.txt
new file mode 100644
index 0000000..573ad72
--- /dev/null
+++ b/3rdparty/composable_kernel/example/39_permute/CMakeLists.txt
@@ -0,0 +1,9 @@
+add_custom_target(example_permute)
+
+add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)
+add_example_executable(example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp)
+add_example_executable(example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp)
+
+add_dependencies(example_permute example_permute_1xHxW_fp16)
+add_dependencies(example_permute example_permute_NxHxW_fp16)
+add_dependencies(example_permute example_permute_HxWx4_fp16)
diff --git a/3rdparty/composable_kernel/example/39_permute/common.hpp b/3rdparty/composable_kernel/example/39_permute/common.hpp
new file mode 100644
index 0000000..ab612ce
--- /dev/null
+++ b/3rdparty/composable_kernel/example/39_permute/common.hpp
@@ -0,0 +1,456 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cassert>
+#include <cstddef>
+#include <cstdlib>
+#include <cstring>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <type_traits>
+#include <utility>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/utility/type.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using F64 = double;
+
+struct Problem final
+{
+    static constexpr std::size_t NumDim = 3;
+
+    using Shape = std::array<std::size_t, NumDim>;
+    using Axes  = Shape;
+
+    Problem() = delete;
+
+    explicit Problem(const Shape& default_shape, const Axes& default_axes)
+        : shape(default_shape), axes(default_axes)
+    {
+    }
+
+    Shape shape;
+    Axes axes;
+};
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+namespace detail {
+
+template <typename Array, std::size_t Difference>
+struct enlarge_array_size;
+
+template <typename T, std::size_t Size, std::size_t Difference>
+struct enlarge_array_size<std::array<T, Size>, Difference>
+{
+    using type = std::array<T, Size + Difference>;
+};
+
+template <typename Array, std::size_t Difference>
+using enlarge_array_size_t = typename enlarge_array_size<Array, Difference>::type;
+
+template <typename Array>
+struct get_array_size;
+
+template <typename T, std::size_t Size>
+struct get_array_size<std::array<T, Size>> : std::integral_constant<std::size_t, Size>
+{
+};
+
+template <typename Array>
+inline constexpr std::size_t get_array_size_v = get_array_size<Array>::value;
+
+template <typename T, typename = void>
+struct is_iterator : std::false_type
+{
+};
+
+template <typename T>
+struct is_iterator<T,
+                   std::void_t<decltype(*std::declval<T>()),
+                               decltype(++std::declval<std::add_lvalue_reference_t<T>>()),
+                               decltype(std::declval<std::add_lvalue_reference_t<T>>()++)>>
+    : std::true_type
+{
+};
+
+template <typename T>
+inline constexpr bool is_iterator_v = is_iterator<T>::value;
+
+struct Placeholder final
+{
+    template <typename T>
+    constexpr inline operator T() const noexcept;
+};
+
+template <typename Iterator, typename = void>
+struct is_output_iterator : std::false_type
+{
+};
+
+template <typename Iterator>
+struct is_output_iterator<
+    Iterator,
+    std::void_t<decltype(*std::declval<Iterator>() = std::declval<Placeholder>())>>
+    : std::bool_constant<is_iterator_v<Iterator>>
+{
+};
+
+template <typename T>
+inline constexpr bool is_output_iterator_v = is_output_iterator<T>::value;
+
+template <typename Iterator, typename = void>
+struct is_bidirectional_iterator : std::false_type
+{
+};
+
+template <typename Iterator>
+struct is_bidirectional_iterator<
+    Iterator,
+    std::void_t<decltype(--std::declval<std::add_lvalue_reference_t<Iterator>>()),
+                decltype(std::declval<std::add_lvalue_reference_t<Iterator>>()--)>>
+    : std::bool_constant<is_iterator_v<Iterator>>
+{
+};
+
+template <typename Iterator>
+inline constexpr bool is_bidirectional_iterator_v = is_bidirectional_iterator<Iterator>::value;
+
+template <typename Iterator, typename = void>
+struct is_random_access_iterator : std::false_type
+{
+};
+
+template <typename Iterator>
+struct is_random_access_iterator<Iterator,
+                                 std::void_t<decltype(std::declval<Iterator>() + 1),
+                                             decltype(std::declval<Iterator>() - 1),
+                                             decltype(std::declval<Iterator>()[1])>>
+    : std::bool_constant<is_iterator_v<Iterator>>
+{
+};
+
+template <typename Iterator>
+inline constexpr bool is_random_access_iterator_v = is_random_access_iterator<Iterator>::value;
+
+template <typename T, typename = void>
+struct is_range : std::false_type
+{
+};
+
+template <typename T>
+struct is_range<T,
+                std::void_t<decltype(begin(std::declval<T>())),
+                            decltype(end(std::declval<T>())),
+                            decltype(begin(std::declval<T>()) != end(std::declval<T>()))>>
+    : std::bool_constant<is_iterator_v<ck::remove_cvref_t<decltype(begin(std::declval<T>()))>>>
+{
+};
+
+template <typename T>
+inline constexpr bool is_range_v = is_range<T>::value;
+
+template <typename Range, typename = void>
+struct is_sized_range : std::false_type
+{
+};
+
+template <typename Range>
+struct is_sized_range<Range, std::void_t<decltype(size(std::declval<Range>()))>>
+    : std::bool_constant<is_range_v<Range>>
+{
+};
+
+template <typename Range>
+inline constexpr bool is_sized_range_v = is_sized_range<Range>::value;
+
+template <typename Range, typename = void>
+struct is_bidirectional_range : std::false_type
+{
+};
+
+template <typename Range>
+struct is_bidirectional_range<Range, std::void_t<>>
+    : std::bool_constant<
+          is_range_v<Range> &&
+          is_bidirectional_iterator_v<ck::remove_cvref_t<decltype(begin(std::declval<Range>()))>>>
+{
+};
+
+template <typename Range>
+inline constexpr bool is_bidirectional_range_v = is_bidirectional_range<Range>::value;
+
+template <typename Range, typename = void>
+struct is_random_access_range : std::false_type
+{
+};
+
+template <typename Range>
+struct is_random_access_range<Range, std::void_t<>>
+    : std::bool_constant<
+          is_range_v<Range> &&
+          is_random_access_iterator_v<ck::remove_cvref_t<decltype(begin(std::declval<Range>()))>>>
+{
+};
+
+template <typename Range>
+inline constexpr bool is_random_access_range_v = is_random_access_range<Range>::value;
+
+template <typename Range>
+class to_array_proxy
+{
+    static_assert(is_range_v<Range>);
+
+    public:
+    explicit to_array_proxy(const Range& source) noexcept : source_(source) {}
+
+    template <typename T, std::size_t Size>
+    operator std::array<T, Size>() const
+    {
+        std::array<T, Size> destination;
+
+        std::copy_n(std::begin(source_),
+                    std::min<std::size_t>(Size, std::size(source_)),
+                    std::begin(destination));
+
+        return destination;
+    }
+
+    private:
+    const Range& source_;
+};
+
+} // namespace detail
+
+template <typename Range>
+inline auto to_array(Range& range) noexcept
+    -> std::enable_if_t<detail::is_range_v<Range>,
+                        detail::to_array_proxy<ck::remove_cvref_t<Range>>>
+{
+    return detail::to_array_proxy<ck::remove_cvref_t<Range>>{range};
+}
+
+template <typename Axes>
+inline auto is_valid_axes(const Axes& axes)
+    -> std::enable_if_t<detail::is_random_access_range_v<Axes>, bool>
+{
+    using std::empty;
+    if(empty(axes))
+    {
+        return false;
+    }
+
+    using std::begin, std::end;
+    std::vector<std::size_t> sorted_axes(begin(axes), end(axes));
+
+    std::sort(begin(sorted_axes), end(sorted_axes));
+    const auto last = std::unique(begin(sorted_axes), end(sorted_axes));
+
+    return (last == end(sorted_axes)) && (*begin(sorted_axes) == 0) &&
+           (*std::prev(last) == size(axes) - 1);
+}
+
+template <typename Shape>
+inline auto is_valid_shape(const Shape& shape) -> std::enable_if_t<detail::is_range_v<Shape>, bool>
+{
+    static_assert(std::is_unsigned_v<ck::remove_cvref_t<decltype(*std::begin(shape))>>);
+
+    using std::begin, std::end;
+    using std::empty;
+    return !empty(shape) && std::all_of(begin(shape), end(shape), [](auto dim) { return 0 < dim; });
+}
+
+template <typename Shape, typename Indices>
+inline auto is_valid_indices(const Shape& shape, const Indices& indices)
+    -> std::enable_if_t<detail::is_sized_range_v<Shape> && detail::is_sized_range_v<Indices>, bool>
+{
+    static_assert(std::is_unsigned_v<ck::remove_cvref_t<decltype(*std::begin(indices))>>);
+
+    if(!is_valid_shape(shape))
+    {
+        return false;
+    }
+
+    using std::empty;
+    if(empty(indices))
+    {
+        return false;
+    }
+
+    using std::size;
+    if(size(shape) != size(indices))
+    {
+        return false;
+    }
+
+    using std::begin, std::end;
+
+    auto dim = begin(shape);
+    auto idx = begin(indices);
+    for(; dim != end(shape) && idx != end(indices); ++dim, ++idx)
+    {
+        if(*dim <= *idx)
+        {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+template <std::size_t Size>
+std::array<std::size_t, Size> transpose(const std::array<std::size_t, Size>& shape,
+                                        const std::array<std::size_t, Size>& axes)
+{
+    assert(is_valid_shape(shape) && is_valid_axes(axes));
+
+    std::array<std::size_t, Size> transposed;
+    auto iter = std::begin(transposed);
+    for(const auto axis : axes)
+    {
+        *iter++ = shape[axis];
+    }
+
+    return transposed;
+}
+
+auto extend_shape(const Problem::Shape& shape, std::size_t new_dim)
+{
+    detail::enlarge_array_size_t<Problem::Shape, 1> extended_shape;
+
+    using std::begin, std::end;
+
+    ck::ranges::copy(shape, begin(extended_shape));
+    extended_shape.back() = new_dim;
+
+    return extended_shape;
+}
+
+auto extend_axes(const Problem::Axes& axes)
+{
+    detail::enlarge_array_size_t<Problem::Axes, 1> extended_axes;
+
+    using std::begin, std::end;
+
+    ck::ranges::copy(axes, begin(extended_axes));
+    extended_axes.back() = detail::get_array_size_v<Problem::Axes>;
+
+    return extended_axes;
+}
+
+template <typename Shape, typename Indices>
+auto advance_indices(const Shape& shape, Indices& indices) -> std::enable_if_t<
+    detail::is_bidirectional_range_v<Shape> && detail::is_sized_range_v<Shape> &&
+        detail::is_bidirectional_range_v<Indices> && detail::is_sized_range_v<Indices>,
+    bool>
+{
+    using std::size;
+    if(!(is_valid_shape(shape) && is_valid_indices(shape, indices) && size(shape) == size(indices)))
+    {
+        return false;
+    }
+
+    bool carry = true;
+
+    using std::rbegin, std::rend;
+    auto dim = rbegin(shape);
+    auto idx = rbegin(indices);
+    for(; carry && dim != rend(shape) && idx != rend(indices); ++dim, ++idx)
+    {
+        *idx  = (*idx + carry);
+        carry = ((*idx == *dim) ? (*idx = 0, true) : false);
+    }
+
+    return !carry;
+}
+
+template <typename Src, typename Axes, typename Functor, typename Dest>
+auto host_permute(const Tensor<Src>& src, const Axes& axes, Functor functor, Tensor<Dest>& dest)
+    -> std::enable_if_t<detail::is_random_access_range_v<Axes> && detail::is_sized_range_v<Axes> &&
+                            std::is_invocable_v<Functor,
+                                                std::add_lvalue_reference_t<Dest>,
+                                                std::add_lvalue_reference_t<Src>>,
+                        bool>
+{
+    const auto& shape            = src.mDesc.GetLengths();
+    const auto& transposed_shape = dest.mDesc.GetLengths();
+    if(!(is_valid_shape(shape) && is_valid_shape(transposed_shape)))
+    {
+        return false;
+    }
+
+    using std::size;
+    if(!is_valid_axes(axes))
+    {
+        return false;
+    }
+
+    static_assert(detail::is_sized_range_v<ck::remove_cvref_t<decltype(shape)>> &&
+                  detail::is_sized_range_v<ck::remove_cvref_t<decltype(transposed_shape)>>);
+
+    if(size(shape) != size(transposed_shape))
+    {
+        return false;
+    }
+
+    static_assert(detail::is_random_access_range_v<ck::remove_cvref_t<decltype(shape)>> &&
+                  detail::is_random_access_range_v<ck::remove_cvref_t<decltype(transposed_shape)>>);
+    {
+        for(std::size_t idx = 0; idx < size(shape); ++idx)
+        {
+            if(transposed_shape[idx] != shape[axes[idx]])
+            {
+                return false;
+            }
+        }
+    }
+
+    std::vector<std::size_t> indices(size(shape), 0);
+    if(!is_valid_indices(shape, indices))
+    {
+        return false;
+    }
+
+    switch(size(shape))
+    {
+    case 3: {
+        do
+        {
+            Dest output = 0;
+            functor(output, src(indices[0], indices[1], indices[2]));
+            dest(indices[axes[0]], indices[axes[1]], indices[axes[2]]) = output;
+        } while(advance_indices(shape, indices));
+    }
+    break;
+    case 4: {
+        do
+        {
+            Dest output = 0;
+            functor(output, src(indices[0], indices[1], indices[2], indices[3]));
+            dest(indices[axes[0]], indices[axes[1]], indices[axes[2]], indices[axes[3]]) = output;
+        } while(advance_indices(shape, indices));
+    }
+    break;
+    default: return false;
+    }
+
+    return true;
+}
diff --git a/3rdparty/composable_kernel/example/39_permute/permute_1xHxW_fp16.cpp b/3rdparty/composable_kernel/example/39_permute/permute_1xHxW_fp16.cpp
new file mode 100644
index 0000000..d7f9b80
--- /dev/null
+++ b/3rdparty/composable_kernel/example/39_permute/permute_1xHxW_fp16.cpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using InDataType  = F16;
+using OutDataType = F16;
+
+// clang-format off
+using DevicePermuteInstance = ck::tensor_operation::device::DevicePermuteImpl
+// ######| NumDim|     InData|     OutData| Elementwise| Block|  NPer|  HPer|  WPer|   InBlock|      InBlockTransfer|           InBlockTransfer|       Src|       Dst|             Src|             Dst|
+// ######|       |       Type|        Type|   Operation|  Size| Block| Block| Block| LdsExtraW| ThreadClusterLengths| ThreadClusterArrangeOrder| VectorDim| VectorDim| ScalarPerVector| ScalarPerVector|
+// ######|       |           |            |            |      |      |      |      |          |                     |                          |          |          |                |                |
+// ######|       |           |            |            |      |      |      |      |          |                     |                          |          |          |                |                |
+         <      3, InDataType, OutDataType, PassThrough,   256,     1,    32,    32,         3,         S<1, 32,  8>,                S<0, 1, 2>,         2,         1,               2,               1>;
+// clang-format on
+
+#include "run_permute_element_example.inc"
+
+int main() { return !run_permute_element_example({1, 32000, 80}, {0, 2, 1}); }
diff --git a/3rdparty/composable_kernel/example/39_permute/permute_HxWx4_fp16.cpp b/3rdparty/composable_kernel/example/39_permute/permute_HxWx4_fp16.cpp
new file mode 100644
index 0000000..342aa13
--- /dev/null
+++ b/3rdparty/composable_kernel/example/39_permute/permute_HxWx4_fp16.cpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using DataType   = F16;
+using BundleType = F64;
+
+static_assert(sizeof(BundleType) % sizeof(DataType) == 0);
+
+// clang-format off
+using DevicePermuteInstance = ck::tensor_operation::device::DevicePermuteImpl
+// ######| NumDim|     InData|     OutData| Elementwise| Block|  NPer|  HPer|  WPer|   InBlock|      InBlockTransfer|           InBlockTransfer|       Src|       Dst|             Src|             Dst|
+// ######|       |       Type|        Type|   Operation|  Size| Block| Block| Block| LdsExtraW| ThreadClusterLengths| ThreadClusterArrangeOrder| VectorDim| VectorDim| ScalarPerVector| ScalarPerVector|
+// ######|       |           |            |            |      |      |      |      |          |                     |                          |          |          |                |                |
+// ######|       |           |            |            |      |      |      |      |          |                     |                          |          |          |                |                |
+         <       3, BundleType, BundleType, PassThrough,   256,     1,    32,    32,         5,         S<1, 32,  8>,                S<0, 1, 2>,         2,         1,               4,               1>;
+// clang-format on
+
+#include "run_permute_bundle_example.inc"
+
+int main() { return !run_permute_bundle_example({1, 80, 32000}, {0, 2, 1}); }
diff --git a/3rdparty/composable_kernel/example/39_permute/permute_NxHxW_fp16.cpp b/3rdparty/composable_kernel/example/39_permute/permute_NxHxW_fp16.cpp
new file mode 100644
index 0000000..b53975e
--- /dev/null
+++ b/3rdparty/composable_kernel/example/39_permute/permute_NxHxW_fp16.cpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using InDataType  = F16;
+using OutDataType = F16;
+
+// clang-format off
+using DevicePermuteInstance = ck::tensor_operation::device::DevicePermuteImpl
+// ######| NumDim|     InData|     OutData| Elementwise| Block|  NPer|  HPer|  WPer|   InBlock|      InBlockTransfer|           InBlockTransfer|       Src|       Dst|             Src|             Dst|
+// ######|       |       Type|        Type|   Operation|  Size| Block| Block| Block| LdsExtraW| ThreadClusterLengths| ThreadClusterArrangeOrder| VectorDim| VectorDim| ScalarPerVector| ScalarPerVector|
+// ######|       |           |            |            |      |      |      |      |          |                     |                          |          |          |                |                |
+// ######|       |           |            |            |      |      |      |      |          |                     |                          |          |          |                |                |
+         <      3, InDataType, OutDataType, PassThrough,   128,     4,    16,     8,         6,          S<2, 16, 4>,                S<0, 1, 2>,         2,         1,               2,               1>;
+// clang-format on
+
+#include "run_permute_element_example.inc"
+
+int main() { return !run_permute_element_example({121, 768, 80}, {0, 2, 1}); }
diff --git a/3rdparty/composable_kernel/example/39_permute/run_permute_bundle_example.inc b/3rdparty/composable_kernel/example/39_permute/run_permute_bundle_example.inc
new file mode 100644
index 0000000..70406d6
--- /dev/null
+++ b/3rdparty/composable_kernel/example/39_permute/run_permute_bundle_example.inc
@@ -0,0 +1,78 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_permute_bundle(const Problem& problem)
+{
+    const auto& input_bundle_shape = problem.shape;
+    const auto& input_bundle_axes  = problem.axes;
+
+    const auto output_bundle_shape = transpose(input_bundle_shape, input_bundle_axes);
+
+    Tensor<BundleType> input_bundle_tensor(input_bundle_shape);
+    Tensor<BundleType> output_bundle_tensor(output_bundle_shape);
+
+    // initialize tensor by assigning DataType values
+    ck::utils::FillUniformDistribution<DataType>{-1.f, 1.f}(input_bundle_tensor.AsSpan<DataType>());
+
+    DeviceMem input_device_buf(input_bundle_tensor.GetElementSpaceSizeInBytes());
+    DeviceMem output_device_buf(output_bundle_tensor.GetElementSpaceSizeInBytes());
+
+    using std::data;
+    input_device_buf.ToDevice(data(input_bundle_tensor));
+
+    static_assert(std::is_default_constructible_v<DevicePermuteInstance>);
+
+    auto permute  = DevicePermuteInstance{};
+    auto argument = permute.MakeArgument(to_array(input_bundle_shape),
+                                         to_array(input_bundle_tensor.GetStrides()),
+                                         to_array(output_bundle_shape),
+                                         to_array(output_bundle_tensor.GetStrides()),
+                                         input_device_buf.GetDeviceBuffer(),
+                                         output_device_buf.GetDeviceBuffer(),
+                                         PassThrough{});
+
+    if(!permute.IsSupportedArgument(argument))
+    {
+        std::cerr << "The runtime parameters seems not supported by the device instance, exiting!"
+                  << std::endl;
+        return false;
+    };
+
+    auto invoker   = permute.MakeInvoker();
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    output_device_buf.FromDevice(data(output_bundle_tensor));
+
+    constexpr std::size_t NumElemsInBundle = sizeof(BundleType) / sizeof(DataType);
+
+    // extend tensor shape from [N, H, W] to [N, H, W, NumElemsInBundle]
+    //               axes  from [0, 2, 1] to [0, 2, 1, 3]
+    const auto input_shape = extend_shape(input_bundle_shape, NumElemsInBundle);
+    const auto input_axes  = extend_axes(input_bundle_axes);
+
+    using std::begin;
+
+    Tensor<DataType> input_tensor(input_shape);
+    ck::ranges::copy(input_bundle_tensor.AsSpan<const DataType>(), begin(input_tensor));
+
+    Tensor<DataType> output_tensor(transpose(input_shape, input_axes));
+    if(!host_permute(input_tensor, input_axes, PassThrough{}, output_tensor))
+    {
+        return false;
+    }
+
+    return ck::utils::check_err(output_bundle_tensor.AsSpan<const DataType>(),
+                                output_tensor.AsSpan<const DataType>(),
+                                "Error: incorrect results in output tensor",
+                                1e-6,
+                                1e-6);
+}
+
+bool run_permute_bundle_example(const Problem::Shape& shape, const Problem::Axes& axes)
+{
+    return run_permute_bundle(Problem{shape, axes});
+}
diff --git a/3rdparty/composable_kernel/example/39_permute/run_permute_element_example.inc b/3rdparty/composable_kernel/example/39_permute/run_permute_element_example.inc
new file mode 100644
index 0000000..bc62353
--- /dev/null
+++ b/3rdparty/composable_kernel/example/39_permute/run_permute_element_example.inc
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_permute_element(const Problem& problem)
+{
+    const auto& input_shape = problem.shape;
+    const auto& input_axes  = problem.axes;
+
+    const auto output_shape = transpose(input_shape, input_axes);
+
+    Tensor<InDataType> input_tensor(input_shape);
+    Tensor<OutDataType> output_tensor(output_shape);
+
+    ck::utils::FillUniformDistribution<InDataType>{-1.f, 1.f}(input_tensor);
+
+    DeviceMem input_device_buf(input_tensor.GetElementSpaceSizeInBytes());
+    DeviceMem output_device_buf(output_tensor.GetElementSpaceSizeInBytes());
+
+    using std::data;
+    input_device_buf.ToDevice(data(input_tensor));
+
+    static_assert(std::is_default_constructible_v<DevicePermuteInstance>);
+
+    auto permute  = DevicePermuteInstance{};
+    auto argument = permute.MakeArgument(to_array(input_shape),
+                                         to_array(input_tensor.GetStrides()),
+                                         to_array(output_shape),
+                                         to_array(output_tensor.GetStrides()),
+                                         input_device_buf.GetDeviceBuffer(),
+                                         output_device_buf.GetDeviceBuffer(),
+                                         PassThrough{});
+
+    if(!permute.IsSupportedArgument(argument))
+    {
+        std::cerr << "The runtime parameters seems not supported by the device instance, exiting!"
+                  << std::endl;
+        return false;
+    };
+
+    auto invoker   = permute.MakeInvoker();
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    output_device_buf.FromDevice(data(output_tensor));
+
+    Tensor<OutDataType> output_tensor_host(output_shape);
+    if(!host_permute(input_tensor, input_axes, PassThrough{}, output_tensor_host))
+    {
+        return false;
+    }
+
+    return ck::utils::check_err(output_tensor.AsSpan<const OutDataType>(),
+                                output_tensor_host.AsSpan<const OutDataType>(),
+                                "Error: incorrect results in output tensor",
+                                1e-6,
+                                1e-6);
+}
+
+bool run_permute_element_example(const Problem::Shape& shape, const Problem::Axes& axes)
+{
+    return run_permute_element(Problem{shape, axes});
+}
diff --git a/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/CMakeLists.txt b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/CMakeLists.txt
new file mode 100644
index 0000000..9cb30f6
--- /dev/null
+++ b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/CMakeLists.txt
@@ -0,0 +1,8 @@
+add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
+add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
+add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
+add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
+
+if(USE_BITINT_EXTENSION_INT4)
+add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp)
+endif(USE_BITINT_EXTENSION_INT4)
diff --git a/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp
new file mode 100644
index 0000000..2aea08c
--- /dev/null
+++ b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using In0DataType       = ck::bhalf_t;
+using Wei0DataType      = ck::bhalf_t;
+using Acc0DataType      = float;
+using Wei1DataType      = ck::bhalf_t;
+using Acc1DataType      = float;
+using C1ShuffleDataType = float;
+using Out1DataType      = ck::bhalf_t;
+
+// This is used for reference code
+using Out0DataType = ck::bhalf_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using In0ElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using Wei0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Wei1ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out1ElementOp = ck::tensor_operation::element_wise::UnaryConvert;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceBatchedGemmGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+        Row,               // ALayout
+        Col,               // B0Layout
+        Col,               // B1Layout
+        Row,               // CLayout
+        In0DataType,       // ADataType,
+        Wei0DataType,      // B0DataType,
+        Wei1DataType,      // B1DataType,
+        Out1DataType,      // CDataType,
+        Acc0DataType,      // AccDataType,
+        C1ShuffleDataType, // CShuffleDataType,
+        In0ElementOp,      // AElementOp,
+        Wei0ElementOp,     // B0ElementOp,
+        Out0ElementOp,     // Acc0ElementOp,
+        Wei1ElementOp,     // B1ElementOp,
+        Out1ElementOp,     // CElementOp,
+        GemmDefault,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        4,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // B1BlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        4,
+        4,
+        true,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+#include "run_grouped_conv_conv_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_conv_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp
new file mode 100644
index 0000000..b7f80e7
--- /dev/null
+++ b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using In0DataType       = ck::half_t;
+using Wei0DataType      = ck::half_t;
+using Acc0DataType      = float;
+using Wei1DataType      = ck::half_t;
+using Acc1DataType      = float;
+using C1ShuffleDataType = float;
+using Out1DataType      = ck::half_t;
+
+// This is used for reference code
+using Out0DataType = ck::half_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using In0ElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using Wei0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Wei1ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out1ElementOp = ck::tensor_operation::element_wise::UnaryConvert;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceBatchedGemmGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+        Row,               // ALayout
+        Col,               // B0Layout
+        Col,               // B1Layout
+        Row,               // CLayout
+        In0DataType,       // ADataType,
+        Wei0DataType,      // B0DataType,
+        Wei1DataType,      // B1DataType,
+        Out1DataType,      // CDataType,
+        Acc0DataType,      // AccDataType,
+        C1ShuffleDataType, // CShuffleDataType,
+        In0ElementOp,      // AElementOp,
+        Wei0ElementOp,     // B0ElementOp,
+        Out0ElementOp,     // Acc0ElementOp,
+        Wei1ElementOp,     // B1ElementOp,
+        Out1ElementOp,     // CElementOp,
+        GemmDefault,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        4,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // B1BlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        4,
+        4,
+        true,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+#include "run_grouped_conv_conv_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_conv_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp
new file mode 100644
index 0000000..15e4609
--- /dev/null
+++ b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using In0DataType       = float;
+using Wei0DataType      = float;
+using Acc0DataType      = float;
+using Wei1DataType      = float;
+using Acc1DataType      = float;
+using C1ShuffleDataType = float;
+using Out1DataType      = float;
+
+// This is used for reference code
+using Out0DataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using In0ElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using Wei0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Wei1ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out1ElementOp = ck::tensor_operation::element_wise::UnaryConvert;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceBatchedGemmGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+        Row,               // ALayout
+        Col,               // B0Layout
+        Col,               // B1Layout
+        Row,               // CLayout
+        In0DataType,       // ADataType,
+        Wei0DataType,      // B0DataType,
+        Wei1DataType,      // B1DataType,
+        Out1DataType,      // CDataType,
+        Acc0DataType,      // AccDataType,
+        C1ShuffleDataType, // CShuffleDataType,
+        In0ElementOp,      // AElementOp,
+        Wei0ElementOp,     // B0ElementOp,
+        Out0ElementOp,     // Acc0ElementOp,
+        Wei1ElementOp,     // B1ElementOp,
+        Out1ElementOp,     // CElementOp,
+        GemmDefault,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        16,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        16,          // Gemm1KPerBlock
+        4,           // AK1
+        4,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        4,
+        4,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        4,
+        4,
+        true,
+        S<4, 64, 1>, // B1BlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        2,
+        2,
+        true,
+        1,               // CShuffleMXdlPerWavePerShuffle
+        2,               // CShuffleNXdlPerWavePerShuffle
+        S<1, 16, 1, 16>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        4>;              // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+#include "run_grouped_conv_conv_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_conv_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp
new file mode 100644
index 0000000..2cc4c07
--- /dev/null
+++ b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using In0DataType        = ck::int4_t;
+using Wei0DataType       = ck::int4_t;
+using KernelIn0DataType  = int8_t;
+using KernelWei0DataType = int8_t;
+using Acc0DataType       = int32_t;
+using Wei1DataType       = ck::int4_t;
+using KernelWei1DataType = int8_t;
+using Acc1DataType       = int32_t;
+using C1ShuffleDataType  = int32_t;
+using Out1DataType       = ck::int4_t;
+using KernelOut1DataType = int8_t;
+
+// This is used for reference code
+using Out0DataType = ck::int4_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using In0ElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using Wei0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Wei1ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out1ElementOp = ck::tensor_operation::element_wise::UnaryConvert;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceBatchedGemmGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+        Row,                // ALayout
+        Col,                // B0Layout
+        Col,                // B1Layout
+        Row,                // CLayout
+        KernelIn0DataType,  // ADataType,
+        KernelWei0DataType, // B0DataType,
+        KernelWei1DataType, // B1DataType,
+        KernelOut1DataType, // CDataType,
+        Acc0DataType,       // AccDataType,
+        C1ShuffleDataType,  // CShuffleDataType,
+        In0ElementOp,       // AElementOp,
+        Wei0ElementOp,      // B0ElementOp,
+        Out0ElementOp,      // Acc0ElementOp,
+        Wei1ElementOp,      // B1ElementOp,
+        Out1ElementOp,      // CElementOp,
+        GemmDefault,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        64,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        64,          // Gemm1KPerBlock
+        16,          // AK1
+        16,          // BK1
+        4,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        16,
+        16,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        16,
+        16,
+        true,
+        S<4, 64, 1>, // B1BlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        4,
+        4,
+        true,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+#define BUILD_INT4_EXAMPLE
+#include "run_grouped_conv_conv_fwd_example.inc"
+
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+
+int main(int argc, char* argv[]) { return run_grouped_conv_conv_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp
new file mode 100644
index 0000000..40ff0f6
--- /dev/null
+++ b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using In0DataType       = int8_t;
+using Wei0DataType      = int8_t;
+using Acc0DataType      = int32_t;
+using Wei1DataType      = int8_t;
+using Acc1DataType      = int32_t;
+using C1ShuffleDataType = int32_t;
+using Out1DataType      = int8_t;
+
+// This is used for reference code
+using Out0DataType = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using In0ElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using Wei0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Wei1ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out1ElementOp = ck::tensor_operation::element_wise::UnaryConvert;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceBatchedGemmGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+        Row,               // ALayout
+        Col,               // B0Layout
+        Col,               // B1Layout
+        Row,               // CLayout
+        In0DataType,       // ADataType,
+        Wei0DataType,      // B0DataType,
+        Wei1DataType,      // B1DataType,
+        Out1DataType,      // CDataType,
+        Acc0DataType,      // AccDataType,
+        C1ShuffleDataType, // CShuffleDataType,
+        In0ElementOp,      // AElementOp,
+        Wei0ElementOp,     // B0ElementOp,
+        Out0ElementOp,     // Acc0ElementOp,
+        Wei1ElementOp,     // B1ElementOp,
+        Out1ElementOp,     // CElementOp,
+        GemmDefault,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        64,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        64,          // Gemm1KPerBlock
+        16,          // AK1
+        16,          // BK1
+        4,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        16,
+        16,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        16,
+        16,
+        true,
+        S<4, 64, 1>, // B1BlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        4,
+        4,
+        true,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+#include "run_grouped_conv_conv_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_conv_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc
new file mode 100644
index 0000000..a2c97f4
--- /dev/null
+++ b/3rdparty/composable_kernel/example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc
@@ -0,0 +1,379 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <ck::index_t NDimSpatial,
+          typename In0DataType,
+          typename Wei0DataType,
+          typename Out0DataType,
+          typename Wei1DataType,
+          typename Out1DataType,
+          typename In0ElementOp,
+          typename Wei0ElementOp,
+          typename Out0ElementOp,
+          typename Wei1ElementOp,
+          typename Out1ElementOp,
+          typename DeviceOpInstance>
+bool run_grouped_conv_conv_fwd(bool do_verification,
+                               int init_method,
+                               bool time_kernel,
+                               const ck::utils::conv::ConvParam& conv0_param,
+                               const ck::utils::conv::ConvParam& conv1_param,
+                               const HostTensorDescriptor& in0_g_n_c_wis_desc,
+                               const HostTensorDescriptor& wei0_g_k_c_xs_desc,
+                               const HostTensorDescriptor& out0_g_n_k_wos_desc,
+                               const HostTensorDescriptor& wei1_g_k_c_xs_desc,
+                               const HostTensorDescriptor& out1_g_n_k_wos_desc,
+                               const In0ElementOp& in0_element_op,
+                               const Wei0ElementOp& wei0_element_op,
+                               const Wei1ElementOp& wei1_element_op,
+                               const Out0ElementOp& out0_element_op,
+                               const Out1ElementOp& out1_element_op)
+{
+    Tensor<In0DataType> in0(in0_g_n_c_wis_desc);
+    Tensor<Wei0DataType> wei0(wei0_g_k_c_xs_desc);
+    Tensor<Wei1DataType> wei1(wei1_g_k_c_xs_desc);
+    Tensor<Out1DataType> out1_host(out1_g_n_k_wos_desc);
+    Tensor<Out1DataType> out1_device(out1_g_n_k_wos_desc);
+
+    std::cout << "in0: " << in0.mDesc << std::endl;
+    std::cout << "wei0: " << wei0.mDesc << std::endl;
+    std::cout << "wei1: " << wei1.mDesc << std::endl;
+    std::cout << "out1: " << out1_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in0.GenerateTensorValue(GeneratorTensor_2<In0DataType>{-5, 5});
+        wei0.GenerateTensorValue(GeneratorTensor_2<Wei0DataType>{-5, 5});
+        wei1.GenerateTensorValue(GeneratorTensor_2<Wei1DataType>{-5, 5});
+        break;
+    default:
+        in0.GenerateTensorValue(GeneratorTensor_3<In0DataType>{0.0, 1.0});
+        wei0.GenerateTensorValue(GeneratorTensor_3<Wei0DataType>{-0.5, 0.5});
+        wei1.GenerateTensorValue(GeneratorTensor_3<Wei1DataType>{-0.5, 0.5});
+    }
+
+#ifdef BUILD_INT4_EXAMPLE
+    DeviceMem in0_device_buf(sizeof(KernelIn0DataType) * in0.mDesc.GetElementSpaceSize());
+    DeviceMem wei0_device_buf(sizeof(KernelWei0DataType) * wei0.mDesc.GetElementSpaceSize());
+    DeviceMem wei1_device_buf(sizeof(KernelWei1DataType) * wei1.mDesc.GetElementSpaceSize());
+    DeviceMem out1_device_buf(sizeof(KernelOut1DataType) * out1_device.mDesc.GetElementSpaceSize());
+
+    const Tensor<KernelIn0DataType> in0_converted(in0);
+    const Tensor<KernelWei0DataType> wei0_converted(wei0);
+    const Tensor<KernelWei1DataType> wei1_converted(wei1);
+
+    in0_device_buf.ToDevice(in0_converted.mData.data());
+    wei0_device_buf.ToDevice(wei0_converted.mData.data());
+    wei1_device_buf.ToDevice(wei1_converted.mData.data());
+#else
+    DeviceMem in0_device_buf(sizeof(In0DataType) * in0.mDesc.GetElementSpaceSize());
+    DeviceMem wei0_device_buf(sizeof(Wei0DataType) * wei0.mDesc.GetElementSpaceSize());
+    DeviceMem wei1_device_buf(sizeof(Wei1DataType) * wei1.mDesc.GetElementSpaceSize());
+    DeviceMem out1_device_buf(sizeof(Out1DataType) * out1_device.mDesc.GetElementSpaceSize());
+
+    in0_device_buf.ToDevice(in0.mData.data());
+    wei0_device_buf.ToDevice(wei0.mData.data());
+    wei1_device_buf.ToDevice(wei1.mData.data());
+#endif
+
+    std::array<ck::index_t, NDimSpatial + 3> a0_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a0_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b0_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b0_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b1_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b1_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e1_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e1_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv0_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv0_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input0_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input0_right_pads{};
+    std::array<ck::index_t, NDimSpatial> conv1_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv1_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input1_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input1_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in0_g_n_c_wis_desc.GetLengths(), a0_g_n_c_wis_lengths);
+    copy(in0_g_n_c_wis_desc.GetStrides(), a0_g_n_c_wis_strides);
+    copy(wei0_g_k_c_xs_desc.GetLengths(), b0_g_k_c_xs_lengths);
+    copy(wei0_g_k_c_xs_desc.GetStrides(), b0_g_k_c_xs_strides);
+    copy(wei1_g_k_c_xs_desc.GetLengths(), b1_g_k_c_xs_lengths);
+    copy(wei1_g_k_c_xs_desc.GetStrides(), b1_g_k_c_xs_strides);
+    copy(out1_g_n_k_wos_desc.GetLengths(), e1_g_n_k_wos_lengths);
+    copy(out1_g_n_k_wos_desc.GetStrides(), e1_g_n_k_wos_strides);
+    copy(conv0_param.conv_filter_strides_, conv0_filter_strides);
+    copy(conv0_param.conv_filter_dilations_, conv0_filter_dilations);
+    copy(conv0_param.input_left_pads_, input0_left_pads);
+    copy(conv0_param.input_right_pads_, input0_right_pads);
+    copy(conv1_param.conv_filter_strides_, conv1_filter_strides);
+    copy(conv1_param.conv_filter_dilations_, conv1_filter_dilations);
+    copy(conv1_param.input_left_pads_, input1_left_pads);
+    copy(conv1_param.input_right_pads_, input1_right_pads);
+
+    // do Conv using GEMM, only works for 1x1 conv for now
+    const ck::index_t gemm_batch = a0_g_n_c_wis_lengths[0];
+
+    const ck::index_t gemm0_m_length =
+        e1_g_n_k_wos_lengths[1] *
+        ck::accumulate_n<ck::index_t>(
+            e1_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>{});
+
+    const ck::index_t gemm0_n_length = b0_g_k_c_xs_lengths[1];
+
+    const ck::index_t gemm0_k_length = ck::accumulate_n<ck::index_t>(
+        b0_g_k_c_xs_lengths.begin() + 2, NDimSpatial + 1, 1, std::multiplies<>{});
+
+    const ck::index_t gemm1_n_length = b1_g_k_c_xs_lengths[1];
+
+    //
+    const ck::index_t a0_stride = a0_g_n_c_wis_strides[2 + NDimSpatial];
+    const ck::index_t b0_stride = b0_g_k_c_xs_strides[2 + NDimSpatial];
+    const ck::index_t b1_stride = b1_g_k_c_xs_strides[2 + NDimSpatial];
+    const ck::index_t e1_stride = e1_g_n_k_wos_strides[2 + NDimSpatial];
+
+    //
+    const ck::index_t a0_batch_stride = a0_g_n_c_wis_strides[0];
+    const ck::index_t b0_batch_stride = b0_g_k_c_xs_strides[0];
+    const ck::index_t b1_batch_stride = b1_g_k_c_xs_strides[0];
+    const ck::index_t e1_batch_stride = e1_g_n_k_wos_strides[0];
+
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument  = device_op.MakeArgument(
+#ifdef BUILD_INT4_EXAMPLE
+        static_cast<KernelIn0DataType*>(in0_device_buf.GetDeviceBuffer()),
+        static_cast<KernelWei0DataType*>(wei0_device_buf.GetDeviceBuffer()),
+        static_cast<KernelWei1DataType*>(wei1_device_buf.GetDeviceBuffer()),
+        static_cast<KernelOut1DataType*>(out1_device_buf.GetDeviceBuffer()),
+#else
+        static_cast<In0DataType*>(in0_device_buf.GetDeviceBuffer()),
+        static_cast<Wei0DataType*>(wei0_device_buf.GetDeviceBuffer()),
+        static_cast<Wei1DataType*>(wei1_device_buf.GetDeviceBuffer()),
+        static_cast<Out1DataType*>(out1_device_buf.GetDeviceBuffer()),
+#endif
+        gemm0_m_length,
+        gemm0_n_length,
+        gemm0_k_length,
+        gemm1_n_length,
+        gemm_batch,
+        a0_stride,
+        b0_stride,
+        b1_stride,
+        e1_stride,
+        a0_batch_stride,
+        b0_batch_stride,
+        b1_batch_stride,
+        e1_batch_stride,
+        in0_element_op,
+        wei0_element_op,
+        out0_element_op,
+        wei1_element_op,
+        out1_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv0_param.GetFlops() + conv1_param.GetFlops();
+    std::size_t num_btype = conv0_param.template GetInputByte<In0DataType>() +
+                            conv0_param.template GetWeightByte<Wei0DataType>() +
+                            conv1_param.template GetWeightByte<Wei1DataType>() +
+                            conv1_param.template GetOutputByte<Out1DataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << device_op.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+        Tensor<Out0DataType> out0_host(out0_g_n_k_wos_desc);
+
+        auto ref_conv0 = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                      In0DataType,
+                                                                      Wei0DataType,
+                                                                      Out0DataType,
+                                                                      In0ElementOp,
+                                                                      Wei0ElementOp,
+                                                                      Out0ElementOp>();
+
+        auto ref_conv1 = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                      Out0DataType,
+                                                                      Wei1DataType,
+                                                                      Out1DataType,
+                                                                      PassThrough,
+                                                                      Wei1ElementOp,
+                                                                      Out1ElementOp>();
+
+        auto ref_conv0_invoker = ref_conv0.MakeInvoker();
+        auto ref_conv1_invoker = ref_conv1.MakeInvoker();
+
+        auto ref_conv0_argument = ref_conv0.MakeArgument(in0,
+                                                         wei0,
+                                                         out0_host,
+                                                         conv0_param.conv_filter_strides_,
+                                                         conv0_param.conv_filter_dilations_,
+                                                         conv0_param.input_left_pads_,
+                                                         conv0_param.input_right_pads_,
+                                                         in0_element_op,
+                                                         wei0_element_op,
+                                                         out0_element_op);
+
+        auto ref_conv1_argument = ref_conv1.MakeArgument(out0_host,
+                                                         wei1,
+                                                         out1_host,
+                                                         conv1_param.conv_filter_strides_,
+                                                         conv1_param.conv_filter_dilations_,
+                                                         conv1_param.input_left_pads_,
+                                                         conv1_param.input_right_pads_,
+                                                         out0_element_op,
+                                                         wei1_element_op,
+                                                         out1_element_op);
+
+        ref_conv0_invoker.Run(ref_conv0_argument);
+        ref_conv1_invoker.Run(ref_conv1_argument);
+
+#ifdef BUILD_INT4_EXAMPLE
+        Tensor<KernelOut1DataType> out1_device_converted(out1_host.mDesc);
+
+        out1_device_buf.FromDevice(out1_device_converted.mData.data());
+
+        out1_device = out1_device_converted.CopyAsType<Out1DataType>();
+#else
+        out1_device_buf.FromDevice(out1_device.mData.data());
+#endif
+
+        return ck::utils::check_err(
+            out1_device, out1_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return true;
+}
+
+bool run_grouped_conv_conv_fwd_example(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv0_param{
+        2, 1, 128, 512, 128, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
+
+    ck::utils::conv::ConvParam conv1_param{
+        2, 1, 128, 128, 512, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    const auto in0_element_op  = In0ElementOp{};
+    const auto wei0_element_op = Wei0ElementOp{};
+    const auto wei1_element_op = Wei1ElementOp{};
+    const auto out0_element_op = Out0ElementOp{};
+    const auto out1_element_op = Out1ElementOp{};
+
+    const auto run = [&](auto ndim_spatial,
+                         auto in0_layout,
+                         auto wei0_layout,
+                         auto wei1_layout,
+                         auto out1_layout) {
+        constexpr ck::index_t ndim_spatial_value = ndim_spatial.value;
+
+        using In0Layout  = decltype(in0_layout);
+        using Wei0Layout = decltype(wei0_layout);
+        using Wei1Layout = decltype(wei1_layout);
+        using Out1Layout = decltype(out1_layout);
+
+        const auto in0_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<In0Layout>(
+                conv0_param);
+
+        const auto wei0_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<Wei0Layout>(
+                conv0_param);
+
+        // out0 doesn't physical exist, any layout for host verification is OK
+        const auto out0_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<Out1Layout>(
+                conv0_param);
+
+        const auto wei1_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<Wei1Layout>(
+                conv1_param);
+
+        const auto out1_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<Out1Layout>(
+                conv1_param);
+
+        return run_grouped_conv_conv_fwd<ndim_spatial_value,
+                                         In0DataType,
+                                         Wei0DataType,
+                                         Out0DataType,
+                                         Wei1DataType,
+                                         Out1DataType,
+                                         In0ElementOp,
+                                         Wei0ElementOp,
+                                         Out0ElementOp,
+                                         Wei1ElementOp,
+                                         Out1ElementOp,
+                                         DeviceBatchedGemmGemmInstance>(do_verification,
+                                                                        init_method,
+                                                                        time_kernel,
+                                                                        conv0_param,
+                                                                        conv1_param,
+                                                                        in0_g_n_c_wis_desc,
+                                                                        wei0_g_k_c_xs_desc,
+                                                                        out0_g_n_k_wos_desc,
+                                                                        wei1_g_k_c_xs_desc,
+                                                                        out1_g_n_k_wos_desc,
+                                                                        in0_element_op,
+                                                                        wei0_element_op,
+                                                                        wei1_element_op,
+                                                                        out0_element_op,
+                                                                        out1_element_op);
+    };
+
+    namespace ctc = ck::tensor_layout::convolution;
+
+    if(conv0_param.num_dim_spatial_ == 1)
+    {
+        return run(ck::Number<1>{}, ctc::GNWC{}, ctc::GKXC{}, ctc::GKXC{}, ctc::GNWK{});
+    }
+    else if(conv0_param.num_dim_spatial_ == 2)
+    {
+        return run(ck::Number<2>{}, ctc::GNHWC{}, ctc::GKYXC{}, ctc::GKYXC{}, ctc::GNHWK{});
+    }
+    else if(conv0_param.num_dim_spatial_ == 3)
+    {
+        return run(ck::Number<3>{}, ctc::GNDHWC{}, ctc::GKZYXC{}, ctc::GKZYXC{}, ctc::GNDHWK{});
+    }
+
+    return true;
+}
diff --git a/3rdparty/composable_kernel/example/42_groupnorm/CMakeLists.txt b/3rdparty/composable_kernel/example/42_groupnorm/CMakeLists.txt
new file mode 100644
index 0000000..c3b7b82
--- /dev/null
+++ b/3rdparty/composable_kernel/example/42_groupnorm/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_groupnorm_sigmoid_fp16 groupnorm_sigmoid_fp16.cpp)
diff --git a/3rdparty/composable_kernel/example/42_groupnorm/groupnorm_sigmoid_fp16.cpp b/3rdparty/composable_kernel/example/42_groupnorm/groupnorm_sigmoid_fp16.cpp
new file mode 100644
index 0000000..e62001d
--- /dev/null
+++ b/3rdparty/composable_kernel/example/42_groupnorm/groupnorm_sigmoid_fp16.cpp
@@ -0,0 +1,174 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp"
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+using XDataType     = ck::half_t;
+using GammaDataType = ck::half_t;
+using BetaDataType  = ck::half_t;
+using YDataType     = ck::half_t;
+using AccDataType   = float;
+
+struct YElementOp
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(ck::is_same<T, float>::value || ck::is_same<T, double>::value ||
+                          ck::is_same<T, ck::half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        T a;
+
+        ck::tensor_operation::element_wise::Sigmoid{}(a, x);
+
+        y = x * a;
+    };
+};
+
+using DeviceInstance =
+    ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
+                                                          GammaDataType,
+                                                          BetaDataType,
+                                                          AccDataType,
+                                                          YDataType,
+                                                          YElementOp,
+                                                          Rank,
+                                                          NumReduceDim,
+                                                          1024, // BlockSize
+                                                          1,    // ClusterM
+                                                          1024, // ClusterK
+                                                          1,    // SliceM
+                                                          32,   // SliceK
+                                                          1,    // SrcVecDim (0=M, 1=K)
+                                                          2,    // SrcScalarPerVector
+                                                          1,    // GammaVecDim (0=M, 1=K)
+                                                          2,    // GammaScalarPerVector
+                                                          1,    // BetaVecDim (0=M, 1=K)
+                                                          2,    // BetaScalarPerVector
+                                                          2>;   // OutScalarPerVector
+
+int main(int argc, char* argv[])
+{
+    ck::index_t N = 2;
+    ck::index_t H = 32;
+    ck::index_t W = 32;
+    ck::index_t G = 32;
+    ck::index_t C = 30;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 6)
+    {
+        N = std::stoi(argv[1]);
+        H = std::stoi(argv[2]);
+        W = std::stoi(argv[3]);
+        G = std::stoi(argv[4]);
+        C = std::stoi(argv[5]);
+    }
+    else
+    {
+        std::cerr << "arg1 to 5: N, H, W, G, C" << std::endl;
+
+        return 1;
+    }
+
+    Tensor<XDataType> x({N, H, W, G, C});
+    Tensor<YDataType> y({N, H, W, G, C});
+    Tensor<GammaDataType> gamma({G, C});
+    Tensor<BetaDataType> beta({G, C});
+
+    ck::utils::FillUniformDistribution<XDataType>{0.f, 1.f}(x);
+    ck::utils::FillUniformDistribution<GammaDataType>{0.f, 1.f}(gamma);
+    ck::utils::FillUniformDistribution<BetaDataType>{0.f, 1.f}(beta);
+
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    const auto y_element_op = YElementOp{};
+
+    auto device_instance = DeviceInstance{};
+    auto argument_ptr    = device_instance.MakeArgumentPointer(
+        {N, H, W, G, C},
+        std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
+        {0, 0, 0, C, 1},
+        {0, 0, 0, C, 1},
+        std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+        {1, 2, 4}, // reduction dimension: [H, W, C]
+        1e-6,
+        x_dev.GetDeviceBuffer(),
+        gamma_dev.GetDeviceBuffer(),
+        beta_dev.GetDeviceBuffer(),
+        y_dev.GetDeviceBuffer(),
+        nullptr,
+        nullptr,
+        y_element_op);
+
+    if(!device_instance.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported" << std::endl;
+        return 1;
+    };
+
+    auto invoker_ptr = device_instance.MakeInvokerPointer();
+    float ave_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, true});
+
+    std::size_t num_btype = sizeof(XDataType) * N * H * W * G * C +
+                            sizeof(YDataType) * N * H * W * G * C + sizeof(GammaDataType) * G * C +
+                            sizeof(BetaDataType) * G * C;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s, "
+              << device_instance.GetTypeString() << std::endl;
+
+    bool pass = true;
+    {
+        Tensor<YDataType> host_y({N, H, W, G, C});
+        using ReferenceInstance = ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
+                                                                                 GammaDataType,
+                                                                                 BetaDataType,
+                                                                                 YDataType,
+                                                                                 AccDataType,
+                                                                                 YElementOp>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(x, gamma, beta, host_y, y_element_op, {N, H, W, G, C}, 1e-6);
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+
+        y_dev.FromDevice(y.mData.data());
+        pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
+    }
+
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/43_splitk_gemm_bias_e_permute/CMakeLists.txt b/3rdparty/composable_kernel/example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
new file mode 100644
index 0000000..c29f18f
--- /dev/null
+++ b/3rdparty/composable_kernel/example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
+add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
diff --git a/3rdparty/composable_kernel/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp b/3rdparty/composable_kernel/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp
new file mode 100644
index 0000000..7ac4b68
--- /dev/null
+++ b/3rdparty/composable_kernel/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp
@@ -0,0 +1,407 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 1;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //############################################| NumDimG| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           Gemm|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|CBlockTransferClusterLengths|   CBlockTransfer|
+        //############################################|        |        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|        _MBlock_MWaveMPerXdl|  ScalarPerVector|
+        //############################################|        |        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|        _NBlock_NWaveNPerXdl|    _NWaveNPerXdl|
+        //############################################|        |        |        |        |      |      |        |         |           |      |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                            |                 |
+        DeviceSplitKContractionMultipleD_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK,   F16,   F16,     F32,      F16, DsDataType,   F16,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,              S<1, 32, 1, 4>,               8>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimG,
+          ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimG == 2 && NumDimM == 2 && NumDimN == 2 && NumDimK == 1, bool> =
+              false>
+struct ReferenceContraction_G2_M2_N2_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_gs_ms_ks,
+                 const Tensor<BDataType>& b_gs_ns_ks,
+                 Tensor<EDataType>& e_gs_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_gs_ms_ks_{a_gs_ms_ks},
+              b_gs_ns_ks_{b_gs_ns_ks},
+              e_gs_ms_ns_{e_gs_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_gs_ms_ks_;
+        const Tensor<BDataType>& b_gs_ns_ks_;
+        Tensor<EDataType>& e_gs_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_G2_M2_N2_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto g0, auto g1, auto m0, auto m1, auto n0, auto n1) {
+                const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a,
+                        ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, g1, m0, m1, k0)));
+                    arg.b_element_op_(
+                        v_b,
+                        ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, g1, n0, n1, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_gs_ms_ns_(g0, g1, m0, m1, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
+                             const Tensor<BDataType>& b_gs_ns_ks,
+                             Tensor<EDataType>& e_gs_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_G2_M2_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int split_k          = 1;
+
+    ck::index_t G0 = 1;
+    ck::index_t G1 = 2;
+
+    ck::index_t M0 = 4;
+    ck::index_t M1 = 256;
+
+    ck::index_t N0 = 16;
+    ck::index_t N1 = 128;
+
+    ck::index_t K0 = 64 * 2;
+
+    // A[G0, G1, M0, M1, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[G0, G1, N0, N1, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
+
+    // D[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
+    // E[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 5)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        split_k         = std::stoi(argv[4]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_gs_ms_ks(
+        std::vector<std::size_t>(a_gs_ms_ks_lengths.begin(), a_gs_ms_ks_lengths.end()),
+        std::vector<std::size_t>(a_gs_ms_ks_strides.begin(), a_gs_ms_ks_strides.end()));
+    Tensor<BDataType> b_gs_ns_ks(
+        std::vector<std::size_t>(b_gs_ns_ks_lengths.begin(), b_gs_ns_ks_lengths.end()),
+        std::vector<std::size_t>(b_gs_ns_ks_strides.begin(), b_gs_ns_ks_strides.end()));
+    Tensor<DDataType> d_gs_ms_ns(
+        std::vector<std::size_t>(d_gs_ms_ns_lengths.begin(), d_gs_ms_ns_lengths.end()),
+        std::vector<std::size_t>(d_gs_ms_ns_strides.begin(), d_gs_ms_ns_strides.end()));
+    Tensor<EDataType> e_gs_ms_ns_host_result(
+        std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
+        std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
+    Tensor<EDataType> e_gs_ms_ns_device_result(
+        std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
+        std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
+    std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
+    std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) *
+                           e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_gs_ms_ks_lengths,
+                                    a_gs_ms_ks_strides,
+                                    b_gs_ns_ks_lengths,
+                                    b_gs_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                    e_gs_ms_ns_lengths,
+                                    e_gs_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op,
+                                    split_k);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t G = std::accumulate(e_gs_ms_ns_lengths.begin(),
+                                    e_gs_ms_ns_lengths.begin() + NumDimG,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    ck::index_t M = std::accumulate(e_gs_ms_ns_lengths.begin() + NumDimG,
+                                    e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    ck::index_t N = std::accumulate(e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM,
+                                    e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM + NumDimN,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    ck::index_t K = std::accumulate(a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM,
+                                    a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM + NumDimK,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    std::size_t flop      = std::size_t(2) * G * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * G * M * K + sizeof(BDataType) * G * K * N +
+                            sizeof(DDataType) * G * M * N + sizeof(EDataType) * G * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(
+            std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
+            std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
+
+        using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CShuffleDataType,
+                                                                     AccDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_gs_ms_ks, b_gs_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        e_gs_ms_ns_host_result.ForEach([&](auto&, auto idx) {
+            cde_element_op(e_gs_ms_ns_host_result(idx), c_ms_ns_host_result(idx), d_gs_ms_ns(idx));
+        });
+
+        return ck::utils::check_err(e_gs_ms_ns_device_result.mData, e_gs_ms_ns_host_result.mData)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp b/3rdparty/composable_kernel/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp
new file mode 100644
index 0000000..764e55e
--- /dev/null
+++ b/3rdparty/composable_kernel/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp
@@ -0,0 +1,407 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F32;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F32;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 1;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //############################################| NumDimG| NumDimM| NumDimN| NumDimK|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|            A|           B|          CDE|           Gemm|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|CBlockTransferClusterLengths|   CBlockTransfer|
+        //############################################|        |        |        |        |      Type|      Type|        Type|         DataType|       Type|      Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|        _MBlock_MWaveMPerXdl|  ScalarPerVector|
+        //############################################|        |        |        |        |          |          |            |                 |           |          |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|        _NBlock_NWaveNPerXdl|    _NWaveNPerXdl|
+        //############################################|        |        |        |        |          |          |            |                 |           |          |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                            |                 |
+        DeviceSplitKContractionMultipleD_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   4,   4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,         1,           1,           1,              S<1, 32, 1, 4>,               4>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimG,
+          ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimG == 2 && NumDimM == 2 && NumDimN == 2 && NumDimK == 1, bool> =
+              false>
+struct ReferenceContraction_G2_M2_N2_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_gs_ms_ks,
+                 const Tensor<BDataType>& b_gs_ns_ks,
+                 Tensor<EDataType>& e_gs_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_gs_ms_ks_{a_gs_ms_ks},
+              b_gs_ns_ks_{b_gs_ns_ks},
+              e_gs_ms_ns_{e_gs_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_gs_ms_ks_;
+        const Tensor<BDataType>& b_gs_ns_ks_;
+        Tensor<EDataType>& e_gs_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_G2_M2_N2_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto g0, auto g1, auto m0, auto m1, auto n0, auto n1) {
+                const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a,
+                        ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, g1, m0, m1, k0)));
+                    arg.b_element_op_(
+                        v_b,
+                        ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, g1, n0, n1, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_gs_ms_ns_(g0, g1, m0, m1, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
+                             const Tensor<BDataType>& b_gs_ns_ks,
+                             Tensor<EDataType>& e_gs_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_G2_M2_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int split_k          = 1;
+
+    ck::index_t G0 = 1;
+    ck::index_t G1 = 2;
+
+    ck::index_t M0 = 4;
+    ck::index_t M1 = 256;
+
+    ck::index_t N0 = 16;
+    ck::index_t N1 = 128;
+
+    ck::index_t K0 = 64 * 2;
+
+    // A[G0, G1, M0, M1, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[G0, G1, N0, N1, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
+
+    // D[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
+    // E[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 5)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        split_k         = std::stoi(argv[4]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_gs_ms_ks(
+        std::vector<std::size_t>(a_gs_ms_ks_lengths.begin(), a_gs_ms_ks_lengths.end()),
+        std::vector<std::size_t>(a_gs_ms_ks_strides.begin(), a_gs_ms_ks_strides.end()));
+    Tensor<BDataType> b_gs_ns_ks(
+        std::vector<std::size_t>(b_gs_ns_ks_lengths.begin(), b_gs_ns_ks_lengths.end()),
+        std::vector<std::size_t>(b_gs_ns_ks_strides.begin(), b_gs_ns_ks_strides.end()));
+    Tensor<DDataType> d_gs_ms_ns(
+        std::vector<std::size_t>(d_gs_ms_ns_lengths.begin(), d_gs_ms_ns_lengths.end()),
+        std::vector<std::size_t>(d_gs_ms_ns_strides.begin(), d_gs_ms_ns_strides.end()));
+    Tensor<EDataType> e_gs_ms_ns_host_result(
+        std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
+        std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
+    Tensor<EDataType> e_gs_ms_ns_device_result(
+        std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
+        std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
+    std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
+    std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) *
+                           e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_gs_ms_ks_lengths,
+                                    a_gs_ms_ks_strides,
+                                    b_gs_ns_ks_lengths,
+                                    b_gs_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                    e_gs_ms_ns_lengths,
+                                    e_gs_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op,
+                                    split_k);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t G = std::accumulate(e_gs_ms_ns_lengths.begin(),
+                                    e_gs_ms_ns_lengths.begin() + NumDimG,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    ck::index_t M = std::accumulate(e_gs_ms_ns_lengths.begin() + NumDimG,
+                                    e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    ck::index_t N = std::accumulate(e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM,
+                                    e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM + NumDimN,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    ck::index_t K = std::accumulate(a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM,
+                                    a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM + NumDimK,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    std::size_t flop      = std::size_t(2) * G * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * G * M * K + sizeof(BDataType) * G * K * N +
+                            sizeof(DDataType) * G * M * N + sizeof(EDataType) * G * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(
+            std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
+            std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
+
+        using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CShuffleDataType,
+                                                                     AccDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_gs_ms_ks, b_gs_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        e_gs_ms_ns_host_result.ForEach([&](auto&, auto idx) {
+            cde_element_op(e_gs_ms_ns_host_result(idx), c_ms_ns_host_result(idx), d_gs_ms_ns(idx));
+        });
+
+        return ck::utils::check_err(e_gs_ms_ns_device_result.mData, e_gs_ms_ns_host_result.mData)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}
diff --git a/3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/CMakeLists.txt b/3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/CMakeLists.txt
new file mode 100644
index 0000000..f02e511
--- /dev/null
+++ b/3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_example_executable(example_conv2d_fwd_xdl_perchannel_quantization_int8 conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp)
+add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp)
+add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp)
diff --git a/3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp b/3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp
new file mode 100644
index 0000000..832665e
--- /dev/null
+++ b/3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp
@@ -0,0 +1,342 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using BiasDataType         = int32_t;
+using RequantScaleDataType = float;
+using AccDataType          = int32_t;
+using CShuffleDataType     = int32_t;
+using OutDataType          = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename BiasLayout,
+          typename RequantScaleLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout, RequantScaleLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<BiasDataType, RequantScaleDataType>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        64,          // KPerBlock
+        16,          // AK1
+        16,          // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        16,          // ABlockTransferSrcScalarPerVector
+        16,          // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        16,          // BBlockTransferSrcScalarPerVector
+        16,          // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 64, 1, 4>,
+        8>;
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& bias_g_k_desc,
+                          const HostTensorDescriptor& requant_scale_g_k_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<BiasDataType> bias(bias_g_k_desc);
+    Tensor<RequantScaleDataType> requant_scale(requant_scale_g_k_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "bias: " << bias.mDesc << std::endl;
+    std::cout << "requant_scale: " << requant_scale.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-128, 127});
+    wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-128, 127});
+    bias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-128, 127});
+    requant_scale.GenerateTensorValue(GeneratorTensor_2<RequantScaleDataType>{0, 1});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem requant_scale_device_buf(sizeof(RequantScaleDataType) *
+                                       requant_scale.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+    requant_scale_device_buf.ToDevice(requant_scale.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(bias_g_k_desc.GetLengths(), d0_g_n_k_wos_lengths);
+    copy(bias_g_k_desc.GetStrides(), d0_g_n_k_wos_strides);
+    copy(requant_scale_g_k_desc.GetLengths(), d1_g_n_k_wos_lengths);
+    copy(requant_scale_g_k_desc.GetStrides(), d1_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(
+        in_device_buf.GetDeviceBuffer(),
+        wei_device_buf.GetDeviceBuffer(),
+        {bias_device_buf.GetDeviceBuffer(), requant_scale_device_buf.GetDeviceBuffer()},
+        out_device_buf.GetDeviceBuffer(),
+        a_g_n_c_wis_lengths,
+        a_g_n_c_wis_strides,
+        b_g_k_c_xs_lengths,
+        b_g_k_c_xs_strides,
+        {d0_g_n_k_wos_lengths, d1_g_n_k_wos_lengths},
+        {d0_g_n_k_wos_strides, d1_g_n_k_wos_strides},
+        e_g_n_k_wos_lengths,
+        e_g_n_k_wos_strides,
+        conv_filter_strides,
+        conv_filter_dilations,
+        input_left_pads,
+        input_right_pads,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_host(out_g_n_k_wos_desc);
+
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     CShuffleDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     PassThrough>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  c_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        out_host.ForEach([&](auto&, auto idx) {
+            out_element_op(out_host(idx), c_host(idx), bias(idx), requant_scale(idx));
+        });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        pass &=
+            ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return (pass ? 0 : 1);
+}
+
+int main()
+{
+    bool do_verification           = true;
+    bool time_kernel               = true;
+    const ck::index_t ndim_spatial = 2;
+
+    ck::utils::conv::ConvParam conv_param{
+        ndim_spatial, // n_dim
+        1,            // group
+        4,            // batch
+        64,           // output channels
+        32,           // input chanels
+        {3, 3},       // weight HW
+        {71, 71},     // x HW
+        {2, 2},       // strides
+        {1, 1},       // dilations
+        {1, 1},       // left_pads
+        {1, 1}        // right_pads
+    };
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{ActivationOp{}};
+
+    using InLayout           = ck::tensor_layout::convolution::GNHWC;
+    using WeiLayout          = ck::tensor_layout::convolution::GKYXC;
+    using BiasLayout         = ck::tensor_layout::convolution::G_K;
+    using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+    using OutLayout          = ck::tensor_layout::convolution::GNHWK;
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    // TODO - make_bias_host_tensor_descriptor_g_n_k_wos_packed()
+    const auto bias_g_k_desc = HostTensorDescriptor({conv_param.G_,
+                                                     conv_param.N_,
+                                                     conv_param.K_,
+                                                     conv_param.output_spatial_lengths_[0],
+                                                     conv_param.output_spatial_lengths_[1]},
+                                                    {
+                                                        conv_param.K_, // g
+                                                        0,             // n
+                                                        1,             // k
+                                                        0,             // ho
+                                                        0              // wo
+                                                    });
+
+    const auto requant_scale_g_k_desc = bias_g_k_desc;
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    std::cout << out_g_n_k_wos_desc << std::endl;
+
+    using deviceOp = DeviceGroupedConvNDFwdInstance<ndim_spatial,
+                                                    InLayout,
+                                                    WeiLayout,
+                                                    BiasLayout,
+                                                    RequantScaleLayout,
+                                                    OutLayout>;
+
+    return run_grouped_conv_fwd<ndim_spatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InElementOp,
+                                WeiElementOp,
+                                OutElementOp,
+                                deviceOp>(do_verification,
+                                          time_kernel,
+                                          conv_param,
+                                          in_g_n_c_wis_desc,
+                                          wei_g_k_c_xs_desc,
+                                          bias_g_k_desc,
+                                          requant_scale_g_k_desc,
+                                          out_g_n_k_wos_desc,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+}
diff --git a/3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp b/3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp
new file mode 100644
index 0000000..f540135
--- /dev/null
+++ b/3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp
@@ -0,0 +1,318 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using InDataType       = int8_t;
+using WeiDataType      = int8_t;
+using BiasDataType     = int32_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using OutDataType      = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename BiasLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<BiasDataType>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        64,          // KPerBlock
+        16,          // AK1
+        16,          // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        16,          // ABlockTransferSrcScalarPerVector
+        16,          // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        16,          // BBlockTransferSrcScalarPerVector
+        16,          // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 64, 1, 4>,
+        8>;
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& bias_g_k_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<BiasDataType> bias(bias_g_k_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "bias: " << bias.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+    wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+    bias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-5, 5});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(bias_g_k_desc.GetLengths(), d0_g_n_k_wos_lengths);
+    copy(bias_g_k_desc.GetStrides(), d0_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      {bias_device_buf.GetDeviceBuffer()},
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      {d0_g_n_k_wos_lengths},
+                                      {d0_g_n_k_wos_strides},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_host(out_g_n_k_wos_desc);
+
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     CShuffleDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     PassThrough>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  c_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        out_host.ForEach(
+            [&](auto&, auto idx) { out_element_op(out_host(idx), c_host(idx), bias(idx)); });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        pass &=
+            ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return (pass ? 0 : 1);
+}
+
+int main()
+{
+    bool do_verification           = true;
+    bool time_kernel               = true;
+    const ck::index_t ndim_spatial = 2;
+
+    ck::utils::conv::ConvParam conv_param{
+        ndim_spatial, // n_dim
+        1,            // group
+        4,            // batch
+        64,           // output channels
+        32,           // input chanels
+        {3, 3},       // weight HW
+        {71, 71},     // x HW
+        {2, 2},       // strides
+        {1, 1},       // dilations
+        {1, 1},       // left_pads
+        {1, 1}        // right_pads
+    };
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{0.5f, ActivationOp{}};
+
+    using InLayout   = ck::tensor_layout::convolution::GNHWC;
+    using WeiLayout  = ck::tensor_layout::convolution::GKYXC;
+    using BiasLayout = ck::tensor_layout::convolution::G_K;
+    using OutLayout  = ck::tensor_layout::convolution::GNHWK;
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    // TODO - make_bias_host_tensor_descriptor_g_n_k_wos_packed()
+    const auto bias_g_k_desc = HostTensorDescriptor({conv_param.G_,
+                                                     conv_param.N_,
+                                                     conv_param.K_,
+                                                     conv_param.output_spatial_lengths_[0],
+                                                     conv_param.output_spatial_lengths_[1]},
+                                                    {
+                                                        conv_param.K_, // g
+                                                        0,             // n
+                                                        1,             // k
+                                                        0,             // ho
+                                                        0              // wo
+                                                    });
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    std::cout << out_g_n_k_wos_desc << std::endl;
+
+    return run_grouped_conv_fwd<
+        ndim_spatial,
+        InDataType,
+        WeiDataType,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        DeviceGroupedConvNDFwdInstance<ndim_spatial, InLayout, WeiLayout, BiasLayout, OutLayout>>(
+        do_verification,
+        time_kernel,
+        conv_param,
+        in_g_n_c_wis_desc,
+        wei_g_k_c_xs_desc,
+        bias_g_k_desc,
+        out_g_n_k_wos_desc,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
+}
diff --git a/3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp b/3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp
new file mode 100644
index 0000000..2d46d86
--- /dev/null
+++ b/3rdparty/composable_kernel/example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp
@@ -0,0 +1,279 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using InDataType       = int8_t;
+using WeiDataType      = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using OutDataType      = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        64,          // KPerBlock
+        16,          // AK1
+        16,          // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        16,          // ABlockTransferSrcScalarPerVector
+        16,          // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        16,          // BBlockTransferSrcScalarPerVector
+        16,          // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 64, 1, 4>,
+        16>;
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+    wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      {},
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      {},
+                                      {},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  out_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        pass &=
+            ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return (pass ? 0 : 1);
+}
+
+int main()
+{
+    bool do_verification           = true;
+    bool time_kernel               = true;
+    const ck::index_t ndim_spatial = 2;
+
+    ck::utils::conv::ConvParam conv_param{
+        ndim_spatial, // n_dim
+        1,            // group
+        4,            // batch
+        64,           // output channels
+        32,           // input chanels
+        {3, 3},       // weight HW
+        {71, 71},     // x HW
+        {2, 2},       // strides
+        {1, 1},       // dilations
+        {1, 1},       // left_pads
+        {1, 1}        // right_pads
+    };
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{0.5f, ActivationOp{}};
+
+    using InLayout  = ck::tensor_layout::convolution::GNHWC;
+    using WeiLayout = ck::tensor_layout::convolution::GKYXC;
+    using OutLayout = ck::tensor_layout::convolution::GNHWK;
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    return run_grouped_conv_fwd<
+        ndim_spatial,
+        InDataType,
+        WeiDataType,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        DeviceGroupedConvNDFwdInstance<ndim_spatial, InLayout, WeiLayout, OutLayout>>(
+        do_verification,
+        time_kernel,
+        conv_param,
+        in_g_n_c_wis_desc,
+        wei_g_k_c_xs_desc,
+        out_g_n_k_wos_desc,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
+}
diff --git a/3rdparty/composable_kernel/example/44_elementwise_permute/CMakeLists.txt b/3rdparty/composable_kernel/example/44_elementwise_permute/CMakeLists.txt
new file mode 100644
index 0000000..0e0091a
--- /dev/null
+++ b/3rdparty/composable_kernel/example/44_elementwise_permute/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
+add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
diff --git a/3rdparty/composable_kernel/example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp b/3rdparty/composable_kernel/example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp
new file mode 100644
index 0000000..0bbdbe5
--- /dev/null
+++ b/3rdparty/composable_kernel/example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp
@@ -0,0 +1,116 @@
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType = F16;
+using BDataType = F16;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using DeviceElementwisePermuteInstance =
+    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ADataType>,
+                                                    ck::Tuple<BDataType>,
+                                                    PassThrough,
+                                                    4,
+                                                    8,
+                                                    ck::Sequence<8>,
+                                                    ck::Sequence<1>>;
+
+template <typename HostTensorA, typename HostTensorB, typename Functor>
+void host_elementwise4D(HostTensorB& B_nhwc, const HostTensorA& A_nchw, Functor functor)
+{
+    for(std::size_t n = 0; n < A_nchw.mDesc.GetLengths()[0]; ++n)
+        for(std::size_t c = 0; c < A_nchw.mDesc.GetLengths()[1]; ++c)
+            for(std::size_t h = 0; h < A_nchw.mDesc.GetLengths()[2]; ++h)
+                for(std::size_t w = 0; w < A_nchw.mDesc.GetLengths()[3]; ++w)
+                {
+                    auto a_val = A_nchw(n, c, h, w);
+                    functor(B_nhwc(n, h, w, c), a_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    std::vector<std::size_t> nchw = {16, 128, 32, 64};
+    std::vector<std::size_t> nhwc = {16, 32, 64, 128};
+    Tensor<ADataType> a(nchw);
+    Tensor<BDataType> b(nhwc);
+
+    a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 4> ab_lengths;
+    std::array<ck::index_t, 4> a_strides = {static_cast<int>(nchw[1] * nchw[2] * nchw[3]),
+                                            static_cast<int>(nchw[2] * nchw[3]),
+                                            static_cast<int>(nchw[3]),
+                                            1};
+    std::array<ck::index_t, 4> b_strides = {static_cast<int>(nhwc[1] * nhwc[2] * nhwc[3]),
+                                            1,
+                                            static_cast<int>(nhwc[2] * nhwc[3]),
+                                            static_cast<int>(nhwc[3])};
+
+    ck::ranges::copy(nchw, ab_lengths.begin());
+
+    auto broadcastPermute = DeviceElementwisePermuteInstance{};
+    auto argument         = broadcastPermute.MakeArgumentPointer(
+        ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+    if(!broadcastPermute.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    std::cout << "A (nchw): " << a.mDesc << std::endl;
+    std::cout << "B (nhwc): " << b.mDesc << std::endl;
+
+    auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
+    float ave_time =
+        broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+    std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
+
+    std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
+                            sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        b_device_buf.FromDevice(b.mData.data());
+        Tensor<BDataType> host_b(nhwc);
+        host_elementwise4D(host_b, a, PassThrough{});
+
+        pass &=
+            ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp b/3rdparty/composable_kernel/example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
new file mode 100644
index 0000000..f16ad3b
--- /dev/null
+++ b/3rdparty/composable_kernel/example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
@@ -0,0 +1,130 @@
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise_2d.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+
+using ADataType = F16;
+using BDataType = F16;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using DeviceElementwisePermuteInstance =
+    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ADataType>,
+                                                    ck::Tuple<BDataType>,
+                                                    PassThrough,
+                                                    3, // NumDim_M
+                                                    1, // NumDim_N
+                                                    8,
+                                                    8,
+                                                    ck::Sequence<8>,
+                                                    ck::Sequence<8>>;
+
+template <typename HostTensorA, typename HostTensorB, typename Functor>
+void host_elementwise4D(HostTensorB& B_nhwc,
+                        const HostTensorA& A_nchw,
+                        const std::vector<std::size_t>& shape_nchw,
+                        Functor functor)
+{
+    for(std::size_t n = 0; n < shape_nchw[0]; ++n)
+        for(std::size_t c = 0; c < shape_nchw[1]; ++c)
+            for(std::size_t h = 0; h < shape_nchw[2]; ++h)
+                for(std::size_t w = 0; w < shape_nchw[3]; ++w)
+                {
+                    auto a_val = A_nchw(n, c, h, w);
+                    functor(B_nhwc(n, h, w, c), a_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    const int N = 120;
+    const int C = 128;
+    const int H = 32;
+    const int W = 1024;
+
+    /**const int N = 120;
+    const int H = 32;
+    const int W = 64;
+
+    const int C = 128;**/
+
+    std::vector<std::size_t> nchw = {N, C, H, W};
+    std::vector<std::size_t> nhwc = {N, H, W, C};
+
+    Tensor<ADataType> a(nchw);
+    Tensor<BDataType> b(nhwc);
+
+    a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+    // LogRangeAsType<float>(std::cout << "Tensor a  : ", a.mData, ",") << std::endl;
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 4> ab_lengths{N, H, W, C};
+
+    std::array<ck::index_t, 4> a_strides = {C * H * W, W, 1, H * W};
+    std::array<ck::index_t, 4> b_strides = {H * W * C, W * C, C, 1};
+
+    auto broadcastPermute = DeviceElementwisePermuteInstance{};
+    auto argument         = broadcastPermute.MakeArgumentPointer(
+        ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+    if(!broadcastPermute.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    std::cout << "A (nchw): " << a.mDesc << std::endl;
+    std::cout << "B (nhwc): " << b.mDesc << std::endl;
+
+    auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
+    float ave_time =
+        broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
+
+    std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
+                            sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        b_device_buf.FromDevice(b.mData.data());
+        // LogRangeAsType<float>(std::cout << "Tensor b  : ", b.mData, ",") << std::endl;
+
+        Tensor<BDataType> host_b(nhwc);
+        host_elementwise4D<Tensor<ADataType>, Tensor<BDataType>, PassThrough>(
+            host_b, a, nchw, PassThrough{});
+
+        // LogRangeAsType<float>(std::cout << "Host b  : ", host_b.mData, ",") << std::endl;
+        pass &=
+            ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/example/45_elementwise_normalization/CMakeLists.txt b/3rdparty/composable_kernel/example/45_elementwise_normalization/CMakeLists.txt
new file mode 100644
index 0000000..8f5b9d4
--- /dev/null
+++ b/3rdparty/composable_kernel/example/45_elementwise_normalization/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_elementwise_layernorm_blockwise elementwise_layernorm_blockwise.cpp)
diff --git a/3rdparty/composable_kernel/example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp b/3rdparty/composable_kernel/example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp
new file mode 100644
index 0000000..12f6bc6
--- /dev/null
+++ b/3rdparty/composable_kernel/example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp
@@ -0,0 +1,195 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+
+using ADataType             = ck::half_t; // Input 1
+using BDataType             = ck::half_t; // Input 2
+using XDataType             = ck::half_t;
+using GammaDataType         = ck::half_t;
+using BetaDataType          = ck::half_t;
+using YDataType             = ck::half_t;
+using AccDataType           = float;
+using XElementwiseOperation = ck::tensor_operation::element_wise::Add;
+using YElementwiseOperation = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+// X = Elementwise(input1, input2, input3, ...)
+// Y = Layernorm(X, beta, gamma)
+using DeviceInstance = ck::tensor_operation::device::DeviceElementwiseNormalizationImpl<
+    ck::Tuple<ADataType, BDataType>,
+    GammaDataType,
+    BetaDataType,
+    AccDataType,
+    YDataType,
+    XElementwiseOperation,
+    YElementwiseOperation,
+    Rank,
+    NumReduceDim,
+    256, // BlockSize
+    8,   // ClusterM
+    32,  // ClusterK
+    1,   // SliceM
+    32,  // SliceK
+    1,   // SrcVecDim (0=M, 1=K)
+    8,   // SrcScalarPerVector
+    1,   // GammaVecDim (0=M, 1=K)
+    8,   // GammaScalarPerVector
+    1,   // BetaVecDim (0=M, 1=K)
+    8,   // BetaScalarPerVector
+    8>;  // OutScalarPerVector
+
+template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
+void host_elementwise2D(HostTensorC& C,
+                        const HostTensorA& A,
+                        const HostTensorB& B,
+                        const std::vector<std::size_t>& shape,
+                        Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(std::size_t m = 0; m < shape[0]; ++m)
+        for(std::size_t n = 0; n < shape[1]; ++n)
+        {
+            auto a_val  = A(m, n);
+            auto b_val  = B(m, n);
+            ctype c_val = 0;
+            functor(c_val, a_val, b_val);
+            C(m, n) = c_val;
+        }
+}
+
+int main()
+{
+    bool time_kernel = true;
+
+    ck::index_t M      = 48 * 256;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = N;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                    std::vector<std::size_t>({stride}));
+    };
+
+    auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                    std::vector<std::size_t>({stride, 1}));
+    };
+
+    Tensor<ADataType> a(f_host_tensor_descriptor2d(M, N, Stride));
+    Tensor<BDataType> b(f_host_tensor_descriptor2d(M, N, Stride));
+    Tensor<GammaDataType> gamma(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta(f_host_tensor_descriptor1d(N, 1));
+    Tensor<YDataType> y(f_host_tensor_descriptor2d(M, N, Stride));
+
+    a.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+    b.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+    gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
+    beta.GenerateTensorValue(GeneratorTensor_2<BetaDataType>{-5, 5});
+
+    DeviceMem a_dev(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_dev(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    a_dev.ToDevice(a.mData.data());
+    b_dev.ToDevice(b.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    std::array<const void*, 2> input = {a_dev.GetDeviceBuffer(), b_dev.GetDeviceBuffer()};
+
+    auto device_instance = DeviceInstance{};
+    auto argument_ptr    = device_instance.MakeArgumentPointer(
+        {M, N},
+        {
+            std::vector<ck::index_t>{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()},
+            std::vector<ck::index_t>{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()},
+        },
+        {0, 1},
+        {0, 1},
+        std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+        {1},
+        1e-4,
+        input,
+        gamma_dev.GetDeviceBuffer(),
+        beta_dev.GetDeviceBuffer(),
+        y_dev.GetDeviceBuffer(),
+        XElementwiseOperation{},
+        YElementwiseOperation{});
+
+    if(!device_instance.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported" << std::endl;
+        return 1;
+    };
+
+    auto invoker_ptr = device_instance.MakeInvokerPointer();
+    float ela_time   = 0;
+    ela_time         = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    float data_mem_size = M * N * sizeof(ADataType) + M * N * sizeof(BDataType) +
+                          M * N * sizeof(YDataType) + N * sizeof(GammaDataType) +
+                          N * sizeof(BetaDataType);
+    float bandwidth = data_mem_size * 1000 / ela_time / 1024 / 1024 / 1024;
+
+    std::cout << "Bandwidth is : " << bandwidth << "GB/s . " << std::endl;
+    std::cout << "Time elapase is : " << ela_time << " ms . " << std::endl;
+
+    bool pass = true;
+    {
+        std::vector<std::size_t> mn = {static_cast<unsigned long>(M),
+                                       static_cast<unsigned long>(N)};
+        Tensor<XDataType> x(f_host_tensor_descriptor2d(M, N, Stride));
+        host_elementwise2D<Tensor<ADataType>,
+                           Tensor<BDataType>,
+                           Tensor<XDataType>,
+                           XElementwiseOperation>(x, a, b, mn, XElementwiseOperation{});
+
+        Tensor<YDataType> host_y(f_host_tensor_descriptor2d(M, N, Stride));
+        using ReferenceInstance =
+            ck::tensor_operation::host::ReferenceLayernorm<XDataType,
+                                                           GammaDataType,
+                                                           BetaDataType,
+                                                           YDataType,
+                                                           AccDataType,
+                                                           YElementwiseOperation,
+                                                           Rank,
+                                                           NumReduceDim>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(x, gamma, beta, host_y, YElementwiseOperation{}, {M, N}, {1}, 1e-4);
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+
+        y_dev.FromDevice(y.mData.data());
+        pass &=
+            ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
+        if(!(pass))
+        {
+            std::cout << "layernorm wrong" << std::endl;
+        }
+    }
+    return (pass ? 0 : 1);
+}
diff --git a/3rdparty/composable_kernel/example/CMakeLists.txt b/3rdparty/composable_kernel/example/CMakeLists.txt
new file mode 100644
index 0000000..02a77f9
--- /dev/null
+++ b/3rdparty/composable_kernel/example/CMakeLists.txt
@@ -0,0 +1,34 @@
+include_directories(BEFORE
+    ${PROJECT_SOURCE_DIR}/include
+    ${PROJECT_SOURCE_DIR}/library/include
+)
+
+add_custom_target(examples)
+
+function(add_example_executable EXAMPLE_NAME FILE_NAME)
+    message("adding example ${EXAMPLE_NAME}")
+    add_executable(${EXAMPLE_NAME} ${FILE_NAME})
+    target_link_libraries(${EXAMPLE_NAME} PRIVATE utility)
+    # HC
+    target_compile_options(${EXAMPLE_NAME} PRIVATE --gpu-max-threads-per-block=1024)
+    add_test(NAME ${EXAMPLE_NAME} COMMAND $<TARGET_FILE:${EXAMPLE_NAME}> ${ARGN})
+    add_dependencies(examples ${EXAMPLE_NAME})
+    add_dependencies(check ${EXAMPLE_NAME})
+    rocm_install(TARGETS ${EXAMPLE_NAME} COMPONENT examples)
+endfunction(add_example_executable EXAMPLE_NAME)
+
+function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
+    message("adding example ${EXAMPLE_NAME}")
+    add_executable(${EXAMPLE_NAME} ${FILE_NAME})
+    target_link_libraries(${EXAMPLE_NAME} PRIVATE utility)
+    add_dependencies(examples ${EXAMPLE_NAME})
+    rocm_install(TARGETS ${EXAMPLE_NAME} COMPONENT examples)
+endfunction(add_example_executable_no_testing EXAMPLE_NAME)
+
+# add all example subdir
+file(GLOB dir_list LIST_DIRECTORIES true *)
+FOREACH(subdir ${dir_list})
+    IF(IS_DIRECTORY "${subdir}")
+        add_subdirectory(${subdir})
+    ENDIF()
+ENDFOREACH()
diff --git a/3rdparty/composable_kernel/include/ck/ck.hpp b/3rdparty/composable_kernel/include/ck/ck.hpp
new file mode 100644
index 0000000..0e71df0
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/ck.hpp
@@ -0,0 +1,201 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#ifndef CK_DONT_USE_HIP_RUNTIME_HEADERS
+#include "hip/hip_runtime.h"
+#include "hip/hip_fp16.h"
+#endif
+
+#define CK_TIME_KERNEL 1
+
+// constant address space for kernel parameter
+// https://llvm.org/docs/AMDGPUUsage.html#address-spaces
+#define CK_CONSTANT_ADDRESS_SPACE __attribute__((address_space(4)))
+
+// launch bounds
+#define CK_USE_LAUNCH_BOUNDS 1
+
+#ifdef CK_USE_LAUNCH_BOUNDS
+#define CK_MAX_THREAD_PER_BLOCK 256
+#define CK_MIN_BLOCK_PER_CU 2
+#endif
+
+// check GPU target
+#ifdef __HIP_DEVICE_COMPILE__
+#if !(defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx926__) || defined(__gfx908__) || \
+      defined(__gfx90a__) || defined(__gfx1030__) || defined(__gfx1100__))
+#error Not supported target
+#endif
+#endif
+
+// buffer resource
+#ifndef __HIP_DEVICE_COMPILE__ // for host code
+#define CK_BUFFER_RESOURCE_3RD_DWORD -1
+#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__)||defined(__gfx926__)  || defined(__gfx908__) || \
+    defined(__gfx90a__) // for GPU code
+#define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
+#elif defined(__gfx1030__) // for GPU code
+#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
+#elif defined(__gfx1100__) // for GPU code
+#define CK_BUFFER_RESOURCE_3RD_DWORD 0x10020000
+#endif
+
+// FMA instruction
+#ifndef __HIP_DEVICE_COMPILE__                   // for host code, define nothing
+#elif defined(__gfx803__) || defined(__gfx900__) // for GPU code
+#define CK_USE_AMD_V_MAC_F32
+#elif defined(__gfx906__)|| defined(__gfx926__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx1030__) // for GPU code
+#define CK_USE_AMD_V_FMAC_F32
+#define CK_USE_AMD_V_DOT2_F32_F16
+#define CK_USE_AMD_V_DOT4_I32_I8
+#endif
+
+// MFMA instruction
+#ifndef __HIP_DEVICE_COMPILE__ // for host code
+#define CK_USE_AMD_MFMA
+#elif defined(__gfx908__) || defined(__gfx90a__) // for GPU code
+#define CK_USE_AMD_MFMA
+#endif
+
+#if defined(__gfx90a__)
+#define CK_USE_AMD_MFMA_BF16_1K_OP
+#endif
+
+// WMMA instruction
+#ifndef __HIP_DEVICE_COMPILE__ // for host code
+#define CK_USE_AMD_WMMA
+#elif defined(__gfx1100__) // for GPU code
+#define CK_USE_AMD_WMMA
+#endif
+
+// buffer load
+#define CK_USE_AMD_BUFFER_LOAD 1
+
+// buffer store
+#define CK_USE_AMD_BUFFER_STORE 1
+
+// buffer atomic add: integer
+#define CK_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER 1
+
+// buffer atomic add: floating point
+#ifndef __HIP_DEVICE_COMPILE__ // for host code
+#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
+#elif defined(__gfx908__) || defined(__gfx90a__) // for GPU code
+#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
+#else // for GPU code
+#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0
+#endif
+
+#if defined(__gfx90a__) // for GPU code
+#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1
+#else
+#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0
+#endif
+
+// inline asm
+#define CK_USE_AMD_INLINE_ASM 1
+
+// inner product (DLOP)
+#define CK_USE_AMD_INNER_PRODUCT_INLINE_ASM 1
+
+// block synchronization only s_wait lgkmcnt(0), not vmcnt(0)
+#define CK_EXPERIMENTAL_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM 1
+
+// experimental feature: multi index implemented as array
+#define CK_EXPERIMENTAL_USE_DYNAMICALLY_INDEXED_MULTI_INDEX 0
+
+// experimental feature: static tensor descriptor
+#define CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR 0
+
+// experimental feature: buffer load/store/atomic-add/ OOB trick
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting. Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter for each usage
+#ifndef CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 0
+#endif
+#define CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK 1
+#define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_ADD_OOB_CHECK_OFFSET_TRICK 1
+#define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_MAX_OOB_CHECK_OFFSET_TRICK 1
+
+// experimental feature: in-regsiter sub-dword transpose
+#define CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE 1
+
+// experimental feature: merge transformation use magic number division
+#define CK_EXPERIMENTAL_MERGE_USE_MAGIC_DIVISION 1
+
+// experimental feature: use __builtin_memcpy instead of pointer cast to access a vector from
+// pointer of scalar
+#define CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS 0
+
+// experimental feature: use __builtin_memcpy instead of union to do bit_cast
+#define CK_EXPERIMENTAL_USE_MEMCPY_FOR_BIT_CAST 1
+
+// experimental feature: optimize for inter-wave scheduling policy
+#define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING 1
+#define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS 1
+// this will let make_default_loop_scheduler() return interwave scheduling flag by default
+#define CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING 0
+// experimental feature: add instances using interwave scheduling
+#define CK_EXPERIMENTAL_INTER_WAVE_INSTANCES 1
+// experimental feature: add instances using pipeline v2
+#define CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES 1
+
+// hack: have underlying assumption that need to be satsified, otherwise it's a bug
+// hack for forcing register to keep idx_diff_low_const in SGPR. idx_diff_low_const must be
+// thread-invariant, otherwise it's a bug
+// TODO: separate index calculation into "compile-time", "global", "block", "wave", "thread"
+#define CK_HACK_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE 0
+
+// workaround: compiler crash when compiling recursive lambda
+#define CK_WORKAROUND_SWDEV_275126 1
+
+// workaround: compiler crash when using buffer load/store for i8
+#define CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE 1
+
+// workaround: compiler gnerating inefficient ds_write instructions
+#define CK_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE 1
+
+// workaround: verifaction failure, due to compiler regression, for conv bwd-data fp16 using some
+// tuning parameter
+#define CK_WORKAROUND_SWDEV_325164 0
+
+// workaround: a BF16 attention kernel for gfx908 is likely affected by a compiler issue
+#ifdef __gfx908__
+#define CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE 1
+#else // __gfx90a__, ...
+#define CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE 0
+#endif // __gfx908__
+
+namespace ck {
+
+enum struct InMemoryDataOperationEnum
+{
+    Set,
+    AtomicAdd,
+    AtomicMax,
+    Add
+};
+
+// FIXME: use regular Sequence and remove this
+template <InMemoryDataOperationEnum... Is>
+struct InMemoryDataOperationEnumSequence
+{
+    static constexpr int mSize = sizeof...(Is);
+
+    __host__ __device__ static constexpr InMemoryDataOperationEnum At(int I)
+    {
+        // the last dummy element is to prevent compiler complain about empty array, when mSize = 0
+        const InMemoryDataOperationEnum mData[mSize + 1] = {Is..., InMemoryDataOperationEnum::Set};
+        return mData[I];
+    }
+};
+
+// index type
+using index_t      = int32_t;
+using long_index_t = int64_t;
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/host_utility/device_prop.hpp b/3rdparty/composable_kernel/include/ck/host_utility/device_prop.hpp
new file mode 100644
index 0000000..e2cbdb7
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/host_utility/device_prop.hpp
@@ -0,0 +1,54 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <string>
+#include <map>
+#include <hip/hip_runtime.h>
+
+namespace ck {
+
+inline std::string get_device_name()
+{
+    hipDeviceProp_t props{};
+    int device;
+    auto status = hipGetDevice(&device);
+    if(status != hipSuccess)
+    {
+        return std::string();
+    }
+
+    status = hipGetDeviceProperties(&props, device);
+    if(status != hipSuccess)
+    {
+        return std::string();
+    }
+    const std::string raw_name(props.gcnArchName);
+
+    // https://github.com/ROCmSoftwarePlatform/MIOpen/blob/8498875aef84878e04c1eabefdf6571514891086/src/target_properties.cpp#L40
+    static std::map<std::string, std::string> device_name_map = {
+        {"Ellesmere", "gfx803"},
+        {"Baffin", "gfx803"},
+        {"RacerX", "gfx803"},
+        {"Polaris10", "gfx803"},
+        {"Polaris11", "gfx803"},
+        {"Tonga", "gfx803"},
+        {"Fiji", "gfx803"},
+        {"gfx800", "gfx803"},
+        {"gfx802", "gfx803"},
+        {"gfx804", "gfx803"},
+        {"Vega10", "gfx900"},
+        {"gfx901", "gfx900"},
+        {"10.3.0 Sienna_Cichlid 18", "gfx1030"},
+    };
+
+    const auto name = raw_name.substr(0, raw_name.find(':')); // str.substr(0, npos) returns str.
+
+    auto match = device_name_map.find(name);
+    if(match != device_name_map.end())
+        return match->second;
+    return name;
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/host_utility/hip_check_error.hpp b/3rdparty/composable_kernel/include/ck/host_utility/hip_check_error.hpp
new file mode 100644
index 0000000..d3dc8ea
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/host_utility/hip_check_error.hpp
@@ -0,0 +1,17 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+
+inline void hip_check_error(hipError_t x)
+{
+    if(x != hipSuccess)
+    {
+        std::ostringstream ss;
+        ss << "HIP runtime error: " << hipGetErrorString(x) << ". " << __FILE__ << ": " << __LINE__
+           << "in function: " << __func__;
+        throw std::runtime_error(ss.str());
+    }
+}
diff --git a/3rdparty/composable_kernel/include/ck/host_utility/io.hpp b/3rdparty/composable_kernel/include/ck/host_utility/io.hpp
new file mode 100644
index 0000000..ac87195
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/host_utility/io.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <iostream>
+#include <vector>
+#include <iterator>
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+
+template <typename T>
+std::ostream& operator<<(std::ostream& os, const std::vector<T>& v)
+{
+    std::copy(std::begin(v), std::end(v), std::ostream_iterator<T>(os, " "));
+    return os;
+}
+
+template <typename T, std::size_t N>
+std::ostream& operator<<(std::ostream& os, const std::array<T, N>& v)
+{
+    std::copy(std::begin(v), std::end(v), std::ostream_iterator<T>(os, " "));
+    return os;
+}
+
+template <typename... Ts>
+std::ostream& operator<<(std::ostream& os, const ck::TensorDescriptor<Ts...>& desc)
+{
+    constexpr ck::index_t nDim = ck::remove_cvref_t<decltype(desc)>::GetNumOfDimension();
+
+    os << "{";
+
+    ck::static_for<0, nDim - 1, 1>{}([&](auto i) { os << desc.GetLength(i) << ", "; });
+
+    os << desc.GetLength(ck::Number<nDim - 1>{});
+
+    os << "}";
+
+    return os;
+}
diff --git a/3rdparty/composable_kernel/include/ck/host_utility/kernel_launch.hpp b/3rdparty/composable_kernel/include/ck/host_utility/kernel_launch.hpp
new file mode 100644
index 0000000..ed6e2f0
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/host_utility/kernel_launch.hpp
@@ -0,0 +1,74 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+
+#include "ck/ck.hpp"
+#include "ck/stream_config.hpp"
+#include "ck/host_utility/hip_check_error.hpp"
+
+template <typename... Args, typename F>
+float launch_and_time_kernel(const StreamConfig& stream_config,
+                             F kernel,
+                             dim3 grid_dim,
+                             dim3 block_dim,
+                             std::size_t lds_byte,
+                             Args... args)
+{
+#if CK_TIME_KERNEL
+    if(stream_config.time_kernel_)
+    {
+        printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
+               __func__,
+               grid_dim.x,
+               grid_dim.y,
+               grid_dim.z,
+               block_dim.x,
+               block_dim.y,
+               block_dim.z);
+
+        const int nrepeat = 10;
+
+        printf("Warm up 1 time\n");
+
+        // warm up
+        kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+
+        printf("Start running %d times...\n", nrepeat);
+
+        hipEvent_t start, stop;
+
+        hip_check_error(hipEventCreate(&start));
+        hip_check_error(hipEventCreate(&stop));
+
+        hip_check_error(hipDeviceSynchronize());
+        hip_check_error(hipEventRecord(start, stream_config.stream_id_));
+
+        for(int i = 0; i < nrepeat; ++i)
+        {
+            kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+        }
+
+        hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
+        hip_check_error(hipEventSynchronize(stop));
+
+        float total_time = 0;
+
+        hip_check_error(hipEventElapsedTime(&total_time, start, stop));
+
+        return total_time / nrepeat;
+    }
+    else
+    {
+        kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+
+        return 0;
+    }
+#else
+    kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+
+    return 0;
+#endif
+}
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_data_convolution_into_gemm_v4r1_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_data_convolution_into_gemm_v4r1_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..db8e48d
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_data_convolution_into_gemm_v4r1_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,275 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// Number of GEMMs = YTilde * XTilde
+// GemmM = C
+// GemmN = N * HTildeSlice * WTildeSlice
+// GemmK = K * YDotSlice * XDotSlice
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t IYTildeValue,
+          index_t IXTildeValue,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto
+transform_backward_data_convolution_into_gemm_v4r1_nhwc_kyxc_nhwk(
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<IYTildeValue>,
+    Number<IXTildeValue>,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1  = Number<GemmK1Value>{};
+    constexpr auto IYTilde = Number<IYTildeValue>{};
+    constexpr auto IXTilde = Number<IXTildeValue>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+    const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+    const auto YTilde = ConvStrideH / GcdStrideDilationH;
+    const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+    const auto YDot = math::integer_divide_ceil(Y, YTilde);
+    const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+    const auto HTilde = Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+    const auto WTilde = Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+    // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+    const auto IHTildeSliceBegin = math::integer_divide_floor(
+        math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+    const auto IWTildeSliceBegin = math::integer_divide_floor(
+        math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+    const auto IHTildeSliceEnd =
+        math::min(HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+    const auto IWTildeSliceEnd =
+        math::min(WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+    const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+    const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+    // GemmK is different for each GEMM
+    const auto YDotSlice = math::integer_divide_ceil(Y - IYTilde, YTilde);
+    const auto XDotSlice = math::integer_divide_ceil(X - IXTilde, XTilde);
+
+    const auto K1 = GemmK1;
+    const auto K0 = K / K1;
+
+    // weight tensor
+    const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+        wei_k_y_x_c_grid_desc,
+        make_tuple(make_pass_through_transform(K),
+                   make_embed_transform(make_tuple(YDot, YTilde),
+                                        make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                   make_embed_transform(make_tuple(XDot, XTilde),
+                                        make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto wei_k0_k1_ydotslice_xdotslice_c_grid_desc =
+        transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                               make_slice_transform(YDot, I0, YDotSlice),
+                                               make_slice_transform(XDot, I0, XDotSlice),
+                                               make_freeze_transform(IYTilde),
+                                               make_freeze_transform(IXTilde),
+                                               make_pass_through_transform(C)),
+                                    make_tuple(Sequence<0>{},
+                                               Sequence<1>{},
+                                               Sequence<3>{},
+                                               Sequence<2>{},
+                                               Sequence<4>{},
+                                               Sequence<5>{}),
+                                    make_tuple(Sequence<0, 1>{},
+                                               Sequence<2>{},
+                                               Sequence<3>{},
+                                               Sequence<>{},
+                                               Sequence<>{},
+                                               Sequence<4>{}));
+
+#if 1
+    const auto wei_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
+        make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                   make_pass_through_transform(C),
+                   make_pass_through_transform(K1)),
+        make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+#else
+    const auto wei_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
+        make_tuple(make_merge_transform(make_tuple(K0, YDotSlice, XDotSlice)),
+                   make_pass_through_transform(C),
+                   make_pass_through_transform(K1)),
+        make_tuple(Sequence<0, 2, 3>{}, Sequence<4>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+#endif
+
+    // output tensor
+    // this add padding check
+    const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+        out_n_ho_wo_k_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Ho, I0, I0),
+                   make_pad_transform(Wo, I0, I0),
+                   make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+        out_n_hop_wop_k_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(YDot, HTilde),
+                                        make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                   make_embed_transform(make_tuple(XDot, WTilde),
+                                        make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                   make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
+        transform_tensor_descriptor(
+            out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
+            make_tuple(make_pass_through_transform(N),
+                       make_slice_transform(YDot, I0, YDotSlice),
+                       make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                       make_slice_transform(XDot, I0, XDotSlice),
+                       make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                       make_unmerge_transform(make_tuple(K0, K1))),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5>{}),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5, 6>{}));
+
+#if 1
+    const auto out_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+        make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                   make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                   make_pass_through_transform(K1)),
+        make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+#else
+    const auto out_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+        make_tuple(make_merge_transform(make_tuple(K0, YDotSlice, XDotSlice)),
+                   make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                   make_pass_through_transform(K1)),
+        make_tuple(Sequence<5, 1, 3>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+#endif
+
+    // input tensor
+    const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+        in_n_hi_wi_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(YTilde, HTilde),
+                                        make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(XTilde, WTilde),
+                                        make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+        in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_freeze_transform(IYTilde),
+                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                   make_freeze_transform(IXTilde),
+                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{},
+                   Sequence<1>{},
+                   Sequence<2>{},
+                   Sequence<3>{},
+                   Sequence<4>{},
+                   Sequence<5>{}),
+        make_tuple(Sequence<0>{},
+                   Sequence<>{},
+                   Sequence<1>{},
+                   Sequence<>{},
+                   Sequence<2>{},
+                   Sequence<3>{}));
+
+    const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        in_n_htildeslice_wtildeslice_c_grid_desc,
+        make_tuple(make_pass_through_transform(C),
+                   make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice))),
+        make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(wei_gemmk0_gemmm_gemmk1_grid_desc,
+                      out_gemmk0_gemmn_gemmk1_grid_desc,
+                      in_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..5391b59
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,355 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1R2_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1R2_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: out
+// B: wei
+// C: in
+// Number of GEMMs = YTilde * XTilde
+// GemmM = N * HTildeSlice * WTildeSlice
+// GemmN = C
+// GemmK = K * YDotSlice * XDotSlice
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          typename IYTilde,
+          typename IXTilde,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto
+transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk(
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    IYTilde i_ytilde,
+    IXTilde i_xtilde,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+    const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+    const auto YTilde = ConvStrideH / GcdStrideDilationH;
+    const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+    const auto YDot = math::integer_divide_ceil(Y, YTilde);
+    const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+    const auto HTilde = Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+    const auto WTilde = Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+    // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+    const auto IHTildeSliceBegin = math::integer_divide_floor(
+        math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+    const auto IWTildeSliceBegin = math::integer_divide_floor(
+        math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+    const auto IHTildeSliceEnd =
+        math::min(HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+    const auto IWTildeSliceEnd =
+        math::min(WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+    const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+    const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+    // GemmK is different for each GEMM
+    const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+    const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+    const auto K1 = GemmK1;
+    const auto K0 = K / K1;
+
+    // A: output tensor
+    // this add padding check
+    const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+        out_n_ho_wo_k_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Ho, I0, I0),
+                   make_pad_transform(Wo, I0, I0),
+                   make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+        out_n_hop_wop_k_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(YDot, HTilde),
+                                        make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                   make_embed_transform(make_tuple(XDot, WTilde),
+                                        make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                   make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
+        transform_tensor_descriptor(
+            out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
+            make_tuple(make_pass_through_transform(N),
+                       make_slice_transform(YDot, I0, YDotSlice),
+                       make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                       make_slice_transform(XDot, I0, XDotSlice),
+                       make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                       make_unmerge_transform(make_tuple(K0, K1))),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5>{}),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5, 6>{}));
+
+#if 1
+    const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+        make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                   make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                   make_pass_through_transform(K1)),
+        make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+#else
+    const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+        make_tuple(make_merge_transform(make_tuple(K0, YDotSlice, XDotSlice)),
+                   make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                   make_pass_through_transform(K1)),
+        make_tuple(Sequence<5, 1, 3>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+#endif
+
+    // B: weight tensor
+    const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+        wei_k_y_x_c_grid_desc,
+        make_tuple(make_pass_through_transform(K),
+                   make_embed_transform(make_tuple(YDot, YTilde),
+                                        make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                   make_embed_transform(make_tuple(XDot, XTilde),
+                                        make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto wei_k0_k1_ydotslice_xdotslice_c_grid_desc =
+        transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                               make_slice_transform(YDot, I0, YDotSlice),
+                                               make_slice_transform(XDot, I0, XDotSlice),
+                                               make_freeze_transform(i_ytilde),
+                                               make_freeze_transform(i_xtilde),
+                                               make_pass_through_transform(C)),
+                                    make_tuple(Sequence<0>{},
+                                               Sequence<1>{},
+                                               Sequence<3>{},
+                                               Sequence<2>{},
+                                               Sequence<4>{},
+                                               Sequence<5>{}),
+                                    make_tuple(Sequence<0, 1>{},
+                                               Sequence<2>{},
+                                               Sequence<3>{},
+                                               Sequence<>{},
+                                               Sequence<>{},
+                                               Sequence<4>{}));
+
+#if 1
+    const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
+        make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                   make_pass_through_transform(C),
+                   make_pass_through_transform(K1)),
+        make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+#else
+    const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
+        make_tuple(make_merge_transform(make_tuple(K0, YDotSlice, XDotSlice)),
+                   make_pass_through_transform(C),
+                   make_pass_through_transform(K1)),
+        make_tuple(Sequence<0, 2, 3>{}, Sequence<4>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+#endif
+
+    // C: input tensor
+    const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+        in_n_hi_wi_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(YTilde, HTilde),
+                                        make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(XTilde, WTilde),
+                                        make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+        in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_freeze_transform(i_ytilde),
+                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                   make_freeze_transform(i_xtilde),
+                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{},
+                   Sequence<1>{},
+                   Sequence<2>{},
+                   Sequence<3>{},
+                   Sequence<4>{},
+                   Sequence<5>{}),
+        make_tuple(Sequence<0>{},
+                   Sequence<>{},
+                   Sequence<1>{},
+                   Sequence<>{},
+                   Sequence<2>{},
+                   Sequence<3>{}));
+
+    const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        in_n_htildeslice_wtildeslice_c_grid_desc,
+        make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                      wei_gemmk0_gemmn_gemmk1_grid_desc,
+                      in_gemmm_gemmn_grid_desc);
+}
+
+// A: out
+// B: wei
+// C: in
+// Number of GEMMs = 1
+// GemmM = N * Ho * Wo
+// GemmN = C
+// GemmK = K
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto
+transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk_1x1(
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const TensorDescriptor<Wei...>& /* wei_k_y_x_c_grid_desc */,
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const ConvStrides& conv_strides,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto K1 = GemmK1;
+    const auto K0 = K / K1;
+
+    // A: output tensor
+    const auto out_gemmk0_gemmm_gemmk1_grid_desc =
+        transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+                                    make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                               make_unmerge_transform(make_tuple(K0, K1))),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+    // B: weight tensor
+    const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+        make_tuple(make_unmerge_transform(make_tuple(K0, K1)), make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // C: input tensor
+    const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hi_wi_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                   make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        in_n_y_ho_x_wo_c_grid_desc,
+        make_tuple(make_freeze_transform(I0),
+                   make_freeze_transform(I0),
+                   make_merge_transform(make_tuple(N, Ho, Wo)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
+        make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                      wei_gemmk0_gemmn_gemmk1_grid_desc,
+                      in_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw.hpp
new file mode 100644
index 0000000..bb1dc23
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw.hpp
@@ -0,0 +1,150 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_ATOMIC_NCHW_KCYX_NKHW_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_ATOMIC_NCHW_KCYX_NKHW_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// GemmM = K
+// GemmK = N * Ho * Wo
+// GemmN = C * Y * X
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value,
+          typename GemmKBatchType,
+          typename GemmKPadType>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw_pad(
+    const TensorDescriptor<Wei...>& wei_k_c_y_x_grid_desc,
+    const TensorDescriptor<In...>& in_n_c_hi_wi_grid_desc,
+    const TensorDescriptor<Out...>& out_n_k_ho_wo_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>,
+    GemmKBatchType GemmKBatch,
+    GemmKPadType GemmKPad)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_c_hi_wi_grid_desc.GetLength(I0);
+    const auto C = in_n_c_hi_wi_grid_desc.GetLength(I1);
+    const auto K = out_n_k_ho_wo_grid_desc.GetLength(I1);
+
+    const auto Hi = in_n_c_hi_wi_grid_desc.GetLength(I2);
+    const auto Wi = in_n_c_hi_wi_grid_desc.GetLength(I3);
+
+    const auto Ho = out_n_k_ho_wo_grid_desc.GetLength(I2);
+    const auto Wo = out_n_k_ho_wo_grid_desc.GetLength(I3);
+
+    const auto Y = wei_k_c_y_x_grid_desc.GetLength(I2);
+    const auto X = wei_k_c_y_x_grid_desc.GetLength(I3);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM      = K;
+    const auto GemmN      = C * Y * X;
+    const auto GemmKTotal = N * Ho * Wo;
+    const index_t GemmK0  = GemmKPad / (GemmKBatch * GemmK1);
+
+    // A: output tensor
+    const auto out_gemmktotal_gemmm_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho * Wo)),
+        make_tuple(make_pass_through_transform(K), make_merge_transform(make_tuple(N, Ho * Wo))),
+        make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+        out_gemmktotal_gemmm_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_gemmkpad_gemmm_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // B: input tensor
+    const auto in_n_c_hip_wip_grid_desc = transform_tensor_descriptor(
+        in_n_c_hi_wi_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_c_y_ho_x_wo_grid_desc = transform_tensor_descriptor(
+        in_n_c_hip_wip_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW))),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}));
+
+    const auto in_gemmktotal_gemmn_grid_desc =
+        transform_tensor_descriptor(in_n_c_y_ho_x_wo_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(C, Y, X)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 2, 4>{}, Sequence<0, 3, 5>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+        in_gemmktotal_gemmn_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        in_gemmkpad_gemmn_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // C: weight tensor
+    const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, C * Y * X)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(C * Y * X)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                      in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp
new file mode 100644
index 0000000..ca53093
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_NCHW_KCYX_NKHW_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_NCHW_KCYX_NKHW_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// GemmM = K
+// GemmK = N * Ho * Wo
+// GemmN = C * Y * X
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw_pad(
+    const TensorDescriptor<Wei...>& wei_k_c_y_x_grid_desc,
+    const TensorDescriptor<In...>& in_n_c_hi_wi_grid_desc,
+    const TensorDescriptor<Out...>& out_n_k_ho_wo_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_c_hi_wi_grid_desc.GetLength(I0);
+    const auto C = in_n_c_hi_wi_grid_desc.GetLength(I1);
+    const auto K = out_n_k_ho_wo_grid_desc.GetLength(I1);
+
+    const auto Hi = in_n_c_hi_wi_grid_desc.GetLength(I2);
+    const auto Wi = in_n_c_hi_wi_grid_desc.GetLength(I3);
+
+    const auto Ho = out_n_k_ho_wo_grid_desc.GetLength(I2);
+    const auto Wo = out_n_k_ho_wo_grid_desc.GetLength(I3);
+
+    const auto Y = wei_k_c_y_x_grid_desc.GetLength(I2);
+    const auto X = wei_k_c_y_x_grid_desc.GetLength(I3);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM  = K;
+    const auto GemmN  = C * Y * X;
+    const auto GemmK  = N * Ho * Wo;
+    const auto GemmK0 = GemmK / GemmK1;
+
+    // weight tensor
+    const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, C * Y * X)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(C * Y * X)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    // input tensor
+    const auto in_n_c_hip_wip_grid_desc = transform_tensor_descriptor(
+        in_n_c_hi_wi_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_c_y_ho_x_wo_grid_desc = transform_tensor_descriptor(
+        in_n_c_hip_wip_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW))),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}));
+
+    const auto in_gemmk_gemmn_grid_desc =
+        transform_tensor_descriptor(in_n_c_y_ho_x_wo_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(C, Y, X)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 2, 4>{}, Sequence<0, 3, 5>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmk0_gemmn_gemmk1_grid_desc =
+        transform_tensor_descriptor(in_gemmk_gemmn_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmN)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // output tensor
+    const auto out_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho * Wo)),
+        make_tuple(make_pass_through_transform(K), make_merge_transform(make_tuple(N, Ho * Wo))),
+        make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto out_gemmk0_gemmm_gemmk1_grid_desc =
+        transform_tensor_descriptor(out_gemmk_gemmm_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmM)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                      in_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..e960f90
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,150 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_ATOMIC_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_ATOMIC_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: in
+// B: wei
+// C: out
+// GemmM = N * Ho * Wo
+// GemmN = K
+// GemmK = Y * X * C
+template <typename... In,
+          typename... Wei,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value,
+          typename GemmKBatchType,
+          typename GemmKPadType>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk_pad(
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>,
+    GemmKBatchType GemmKBatch,
+    GemmKPadType GemmKPad)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM      = Y * X * C;
+    const auto GemmN      = K;
+    const auto GemmKTotal = N * Ho * Wo;
+    const index_t GemmK0  = GemmKPad / (GemmKBatch * GemmK1);
+
+    // A: input tensor
+    const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+        in_n_hi_wi_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmktotal_gemmm_grid_desc =
+        transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+        in_gemmktotal_gemmm_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        in_gemmkpad_gemmm_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // B: output tensor
+    const auto out_gemmktotal_gemmn_grid_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+    const auto out_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+        out_gemmktotal_gemmn_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_gemmkpad_gemmn_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // C: weight tensor
+    const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    return make_tuple(in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                      out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..052bab4
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,135 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: in
+// B: wei
+// C: out
+// GemmM = N * Ho * Wo
+// GemmN = K
+// GemmK = Y * X * C
+template <typename... In,
+          typename... Wei,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk_pad(
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM  = Y * X * C;
+    const auto GemmN  = K;
+    const auto GemmK  = N * Ho * Wo;
+    const auto GemmK0 = GemmK / GemmK1;
+
+    // A: input tensor
+    const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+        in_n_hi_wi_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmk_gemmm_grid_desc =
+        transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmk0_gemmm_gemmk1_grid_desc =
+        transform_tensor_descriptor(in_gemmk_gemmm_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmM)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // B: output tensor
+    const auto out_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+        make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto out_gemmk0_gemmn_gemmk1_grid_desc =
+        transform_tensor_descriptor(out_gemmk_gemmn_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmN)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // C: weight tensor
+    const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                      out_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..c301a9e
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R5_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R5_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: out
+// B: in
+// C: wei
+// GemmM = K
+// GemmN = Y * X * C
+// GemmKTotal = N * Ho * Wo
+template <typename... In,
+          typename... Wei,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value,
+          typename GemmKBatchType,
+          typename GemmKPadType>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk_pad(
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>,
+    GemmKBatchType GemmKBatch,
+    GemmKPadType GemmKPad)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM      = K;
+    const auto GemmN      = Y * X * C;
+    const auto GemmKTotal = N * Ho * Wo;
+    const index_t GemmK0  = GemmKPad / (GemmKBatch * GemmK1);
+
+    // A: output tensor
+    const auto out_gemmktotal_gemmm_grid_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+    const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+        out_gemmktotal_gemmm_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_gemmkpad_gemmm_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // B: input tensor
+    const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+        in_n_hi_wi_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmktotal_gemmn_grid_desc =
+        transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+        in_gemmktotal_gemmn_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        in_gemmkpad_gemmn_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // C: weight tensor
+    const auto wei_gemmm_gemmn_grid_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+    return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                      in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp
new file mode 100644
index 0000000..4126753
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp
@@ -0,0 +1,153 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION3D_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_FORWARD_CONVOLUTION3D_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: in
+// B: wei
+// C: out
+// GemmM = N * Do * Ho * Wo
+// GemmN = K
+// GemmK = Z * Y * X * C
+template <typename... In,
+          typename... Wei,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto
+transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk_pad(
+    const TensorDescriptor<In...>& in_grid_desc_n_di_hi_wi_c,
+    const TensorDescriptor<Wei...>& wei_k_z_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_do_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+    constexpr auto I4 = Number<4>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_grid_desc_n_di_hi_wi_c.GetLength(I0);
+    const auto K = out_n_do_ho_wo_k_grid_desc.GetLength(I4);
+    const auto C = in_grid_desc_n_di_hi_wi_c.GetLength(I4);
+
+    const auto Di = in_grid_desc_n_di_hi_wi_c.GetLength(I1);
+    const auto Hi = in_grid_desc_n_di_hi_wi_c.GetLength(I2);
+    const auto Wi = in_grid_desc_n_di_hi_wi_c.GetLength(I3);
+
+    const auto Do = out_n_do_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Ho = out_n_do_ho_wo_k_grid_desc.GetLength(I2);
+    const auto Wo = out_n_do_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Z = wei_k_z_y_x_c_grid_desc.GetLength(I1);
+    const auto Y = wei_k_z_y_x_c_grid_desc.GetLength(I2);
+    const auto X = wei_k_z_y_x_c_grid_desc.GetLength(I3);
+
+    const auto ConvStrideD = conv_strides[I0];
+    const auto ConvStrideH = conv_strides[I1];
+    const auto ConvStrideW = conv_strides[I2];
+
+    const auto ConvDilationD = conv_dilations[I0];
+    const auto ConvDilationH = conv_dilations[I1];
+    const auto ConvDilationW = conv_dilations[I2];
+
+    const auto InLeftPadD = in_left_pads[I0];
+    const auto InLeftPadH = in_left_pads[I1];
+    const auto InLeftPadW = in_left_pads[I2];
+
+    const auto InRightPadD = in_right_pads[I0];
+    const auto InRightPadH = in_right_pads[I1];
+    const auto InRightPadW = in_right_pads[I2];
+
+    const auto GemmM  = N * Do * Ho * Wo;
+    const auto GemmN  = K;
+    const auto GemmK  = Z * Y * X * C;
+    const auto GemmK0 = GemmK / GemmK1;
+
+    // A: input tensor
+    const auto in_grid_desc_n_dip_hip_wip_c = transform_tensor_descriptor(
+        in_grid_desc_n_di_hi_wi_c,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Di, InLeftPadD, InRightPadD),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+    const auto in_grid_desc_n_z_do_y_ho_x_wo_c = transform_tensor_descriptor(
+        in_grid_desc_n_dip_hip_wip_c,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+        make_tuple(
+            Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5, 6>{}, Sequence<7>{}));
+
+    const auto in_grid_desc_gemmk_gemmm =
+        transform_tensor_descriptor(in_grid_desc_n_z_do_y_ho_x_wo_c,
+                                    make_tuple(make_merge_transform(make_tuple(Z, Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Do, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_grid_desc_gemmk0_gemmm_gemmk1 =
+        transform_tensor_descriptor(in_grid_desc_gemmk_gemmm,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmM)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // B: weight tensor
+    const auto wei_grid_desc_gemmk_gemmn = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(Z * Y * X * C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto wei_grid_desc_gemmk0_gemmn_gemmk1 =
+        transform_tensor_descriptor(wei_grid_desc_gemmk_gemmn,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmN)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // C: output tensor
+    const auto out_grid_desc_gemmm_gemmn = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K)),
+        make_tuple(make_pass_through_transform(N * Do * Ho * Wo), make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    // const auto out_grid_desc_gemmm_gemmn = transform_tensor_descriptor(
+    //     out_n_do_ho_wo_k_grid_desc,
+    //     make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+    //                make_pass_through_transform(K)),
+    //     make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<3>{}),
+    //     make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(in_grid_desc_gemmk0_gemmm_gemmk1,
+                      wei_grid_desc_gemmk0_gemmn_gemmk1,
+                      out_grid_desc_gemmm_gemmn);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw.hpp
new file mode 100644
index 0000000..381f9ac
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw.hpp
@@ -0,0 +1,260 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4_NCHW_KCYX_NKHW_HPP
+#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4_NCHW_KCYX_NKHW_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// GemmM = K
+// GemmN = N * Ho * Wo
+// GemmK = C * Y * X
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads>
+__host__ __device__ constexpr auto transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw_pad(
+    const TensorDescriptor<Wei...>& wei_k_c_y_x_global_desc,
+    const TensorDescriptor<In...>& in_n_c_hi_wi_global_desc,
+    const TensorDescriptor<Out...>& out_n_k_ho_wo_global_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    const auto N = in_n_c_hi_wi_global_desc.GetLength(I0);
+    const auto C = in_n_c_hi_wi_global_desc.GetLength(I1);
+    const auto K = out_n_k_ho_wo_global_desc.GetLength(I1);
+
+    const auto Hi = in_n_c_hi_wi_global_desc.GetLength(I2);
+    const auto Wi = in_n_c_hi_wi_global_desc.GetLength(I3);
+
+    const auto Ho = out_n_k_ho_wo_global_desc.GetLength(I2);
+    const auto Wo = out_n_k_ho_wo_global_desc.GetLength(I3);
+
+    const auto Y = wei_k_c_y_x_global_desc.GetLength(I2);
+    const auto X = wei_k_c_y_x_global_desc.GetLength(I3);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    // weight tensor
+    const auto wei_gemmk_gemmm_global_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, C * Y * X)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(C * Y * X)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    // input tensor
+    const auto in_n_c_hip_wip_global_desc = transform_tensor_descriptor(
+        in_n_c_hi_wi_global_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_c_y_ho_x_wo_global_desc = transform_tensor_descriptor(
+        in_n_c_hip_wip_global_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW))),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}));
+
+    const auto in_gemmk_gemmn_global_desc =
+        transform_tensor_descriptor(in_n_c_y_ho_x_wo_global_desc,
+                                    make_tuple(make_merge_transform(make_tuple(C, Y, X)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 2, 4>{}, Sequence<0, 3, 5>{}),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    // output tensor
+    const auto out_gemmm_gemmn_global_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho * Wo)),
+        make_tuple(make_pass_through_transform(K), make_merge_transform(make_tuple(N, Ho * Wo))),
+        make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(
+        wei_gemmk_gemmm_global_desc, in_gemmk_gemmn_global_desc, out_gemmm_gemmn_global_desc);
+}
+
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads>
+__host__ __device__ constexpr auto
+transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw_no_pad(
+    const TensorDescriptor<Wei...>& wei_k_c_y_x_global_desc,
+    const TensorDescriptor<In...>& in_n_c_hi_wi_global_desc,
+    const TensorDescriptor<Out...>& out_n_k_ho_wo_global_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    const auto N = in_n_c_hi_wi_global_desc.GetLength(I0);
+    const auto C = in_n_c_hi_wi_global_desc.GetLength(I1);
+    const auto K = out_n_k_ho_wo_global_desc.GetLength(I1);
+
+    const auto Ho = out_n_k_ho_wo_global_desc.GetLength(I2);
+    const auto Wo = out_n_k_ho_wo_global_desc.GetLength(I3);
+
+    const auto Y = wei_k_c_y_x_global_desc.GetLength(I2);
+    const auto X = wei_k_c_y_x_global_desc.GetLength(I3);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    assert(InLeftPadH == 0 && InLeftPadW == 0 && InRightPadH == 0 && InRightPadW == 0);
+
+    // weight tensor
+    const auto wei_gemmk_gemmm_global_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, C * Y * X)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(C * Y * X)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    // input tensor
+    const auto in_n_c_y_ho_x_wo_global_desc = transform_tensor_descriptor(
+        in_n_c_hi_wi_global_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW))),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}));
+
+    const auto in_gemmk_gemmn_global_desc =
+        transform_tensor_descriptor(in_n_c_y_ho_x_wo_global_desc,
+                                    make_tuple(make_merge_transform(make_tuple(C, Y, X)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 2, 4>{}, Sequence<0, 3, 5>{}),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    // output tensor
+    const auto out_gemmm_gemmn_global_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho * Wo)),
+        make_tuple(make_pass_through_transform(K), make_merge_transform(make_tuple(N, Ho * Wo))),
+        make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(
+        wei_gemmk_gemmm_global_desc, in_gemmk_gemmn_global_desc, out_gemmm_gemmn_global_desc);
+}
+
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads>
+__host__ __device__ constexpr auto transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw_1x1(
+    const TensorDescriptor<Wei...>& wei_k_c_y_x_global_desc,
+    const TensorDescriptor<In...>& in_n_c_hi_wi_global_desc,
+    const TensorDescriptor<Out...>& out_n_k_ho_wo_global_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    const auto N = in_n_c_hi_wi_global_desc.GetLength(I0);
+    const auto C = in_n_c_hi_wi_global_desc.GetLength(I1);
+    const auto K = out_n_k_ho_wo_global_desc.GetLength(I1);
+
+    const auto Ho = out_n_k_ho_wo_global_desc.GetLength(I2);
+    const auto Wo = out_n_k_ho_wo_global_desc.GetLength(I3);
+
+    const auto Y = wei_k_c_y_x_global_desc.GetLength(I2);
+    const auto X = wei_k_c_y_x_global_desc.GetLength(I3);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    assert(Y == 1 && X == 1 && ConvStrideH == 1 && ConvStrideW == 1 && ConvDilationH == 1 &&
+           ConvDilationW == 1 && InLeftPadH == 0 && InLeftPadW == 0 && InRightPadH == 0 &&
+           InRightPadW == 0);
+
+    // weight tensor
+    const auto wei_gemmk_gemmm_global_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    // input tensor
+    const auto in_gemmk_gemmn_global_desc = transform_tensor_descriptor(
+        in_n_c_hi_wi_global_desc,
+        make_tuple(make_pass_through_transform(C), make_merge_transform(make_tuple(N, Ho, Wo))),
+        make_tuple(Sequence<1>{}, Sequence<0, 2, 3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    // output tensor
+    const auto out_gemmm_gemmn_global_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho * Wo)),
+        make_tuple(make_pass_through_transform(K), make_merge_transform(make_tuple(N, Ho * Wo))),
+        make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(
+        wei_gemmk_gemmm_global_desc, in_gemmk_gemmn_global_desc, out_gemmm_gemmn_global_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..ebfaabb
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,179 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// GemmM = K
+// GemmN = N * Ho * Wo
+// GemmK = C * Y * X
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads>
+__host__ __device__ constexpr auto transform_forward_convolution_into_gemm_v4r4_nhwc_kyxc_nhwk_pad(
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    // weight tensor
+    const auto wei_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    // input tensor
+    const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+        in_n_hi_wi_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmk_gemmn_grid_desc =
+        transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    // output tensor
+    const auto out_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+        make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    return make_tuple(
+        wei_gemmk_gemmm_grid_desc, in_gemmk_gemmn_grid_desc, out_gemmm_gemmn_grid_desc);
+}
+
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads>
+__host__ __device__ constexpr auto transform_forward_convolution_into_gemm_v4r4_nhwc_kyxc_nhwk_1x1(
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    assert(Y == 1 && X == 1 && ConvStrideH == 1 && ConvStrideW == 1 && ConvDilationH == 1 &&
+           ConvDilationW == 1 && InLeftPadH == 0 && InLeftPadW == 0 && InRightPadH == 0 &&
+           InRightPadW == 0);
+
+    // weight tensor
+    const auto wei_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    // input tensor
+    const auto in_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, C)),
+        make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    // output tensor
+    const auto out_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+        make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    return make_tuple(
+        wei_gemmk_gemmm_grid_desc, in_gemmk_gemmn_grid_desc, out_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp
new file mode 100644
index 0000000..6e576d6
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R2_NCHW_KCYX_NKHW_HPP
+#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R2_NCHW_KCYX_NKHW_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// GemmM = K
+// GemmN = N * Ho * Wo
+// GemmK = C * Y * X
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto
+transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw_pad(
+    const TensorDescriptor<Wei...>& wei_k_c_y_x_grid_desc,
+    const TensorDescriptor<In...>& in_n_c_hi_wi_grid_desc,
+    const TensorDescriptor<Out...>& out_n_k_ho_wo_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_c_hi_wi_grid_desc.GetLength(I0);
+    const auto C = in_n_c_hi_wi_grid_desc.GetLength(I1);
+    const auto K = out_n_k_ho_wo_grid_desc.GetLength(I1);
+
+    const auto Hi = in_n_c_hi_wi_grid_desc.GetLength(I2);
+    const auto Wi = in_n_c_hi_wi_grid_desc.GetLength(I3);
+
+    const auto Ho = out_n_k_ho_wo_grid_desc.GetLength(I2);
+    const auto Wo = out_n_k_ho_wo_grid_desc.GetLength(I3);
+
+    const auto Y = wei_k_c_y_x_grid_desc.GetLength(I2);
+    const auto X = wei_k_c_y_x_grid_desc.GetLength(I3);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM  = K;
+    const auto GemmN  = N * Ho * Wo;
+    const auto GemmK  = C * Y * X;
+    const auto GemmK0 = GemmK / GemmK1;
+
+    // weight tensor
+    const auto wei_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, C * Y * X)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(C * Y * X)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto wei_gemmk0_gemmm_gemmk1_grid_desc =
+        transform_tensor_descriptor(wei_gemmk_gemmm_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmM)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // input tensor
+    const auto in_n_c_hip_wip_grid_desc = transform_tensor_descriptor(
+        in_n_c_hi_wi_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_c_y_ho_x_wo_grid_desc = transform_tensor_descriptor(
+        in_n_c_hip_wip_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW))),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}));
+
+    const auto in_gemmk_gemmn_grid_desc =
+        transform_tensor_descriptor(in_n_c_y_ho_x_wo_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(C, Y, X)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 2, 4>{}, Sequence<0, 3, 5>{}),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_gemmk0_gemmn_gemmk1_grid_desc =
+        transform_tensor_descriptor(in_gemmk_gemmn_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmN)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // output tensor
+    const auto out_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho * Wo)),
+        make_tuple(make_pass_through_transform(K), make_merge_transform(make_tuple(N, Ho * Wo))),
+        make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(wei_gemmk0_gemmm_gemmk1_grid_desc,
+                      in_gemmk0_gemmn_gemmk1_grid_desc,
+                      out_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..13e1bf2
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R2_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R2_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// GemmM = K
+// GemmN = N * Ho * Wo
+// GemmK = C * Y * X
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto
+transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk_pad(
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM  = K;
+    const auto GemmN  = N * Ho * Wo;
+    const auto GemmK  = C * Y * X;
+    const auto GemmK0 = GemmK / GemmK1;
+
+    // weight tensor
+    const auto wei_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto wei_gemmk0_gemmm_gemmk1_grid_desc =
+        transform_tensor_descriptor(wei_gemmk_gemmm_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmM)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // input tensor
+    const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+        in_n_hi_wi_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmk_gemmn_grid_desc =
+        transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_gemmk0_gemmn_gemmk1_grid_desc =
+        transform_tensor_descriptor(in_gemmk_gemmn_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmN)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // output tensor
+    const auto out_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+        make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    return make_tuple(wei_gemmk0_gemmm_gemmk1_grid_desc,
+                      in_gemmk0_gemmn_gemmk1_grid_desc,
+                      out_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..088d14b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,134 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: in
+// B: wei
+// C: out
+// GemmM = N * Ho * Wo
+// GemmN = K
+// GemmK = Y * X * C
+template <typename... In,
+          typename... Wei,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk(
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM  = N * Ho * Wo;
+    const auto GemmN  = K;
+    const auto GemmK  = Y * X * C;
+    const auto GemmK0 = GemmK / GemmK1;
+
+    // A: input tensor
+    const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+        in_n_hi_wi_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmk_gemmm_grid_desc =
+        transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_gemmk0_gemmm_gemmk1_grid_desc =
+        transform_tensor_descriptor(in_gemmk_gemmm_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmM)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // B: weight tensor
+    const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+        transform_tensor_descriptor(wei_gemmk_gemmn_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmN)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // C: output tensor
+    const auto out_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+        make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                      wei_gemmk0_gemmn_gemmk1_grid_desc,
+                      out_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v6r1_nchw_kcyx_nkhw.hpp b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v6r1_nchw_kcyx_nkhw.hpp
new file mode 100644
index 0000000..a6785d5
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/problem_transform/transform_forward_convolution_into_gemm_v6r1_nchw_kcyx_nkhw.hpp
@@ -0,0 +1,135 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_CONTRACTION_V6R1_NCHW_KCYX_NKHW_HPP
+#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_CONTRACTION_V6R1_NCHW_KCYX_NKHW_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// GemmM0 = 1
+// GemmM1 = K
+// GemmN0 = N0
+// GemmN1 = (N / N0) * Ho * Wo
+// GemmK0 = (C / C0) * Y * X
+// GemmK1 = C0
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          typename N0Type,
+          typename C0Type>
+__host__ __device__ constexpr auto
+transform_forward_convolution_into_contraction_v6r1_nchw_kcyx_nkhw_pad(
+    const TensorDescriptor<Wei...>& wei_k_c_y_x_grid_desc,
+    const TensorDescriptor<In...>& in_n_c_hi_wi_grid_desc,
+    const TensorDescriptor<Out...>& out_n_k_ho_wo_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    const N0Type& N0,
+    const C0Type& C0)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    const auto N = in_n_c_hi_wi_grid_desc.GetLength(I0);
+    const auto C = in_n_c_hi_wi_grid_desc.GetLength(I1);
+    const auto K = out_n_k_ho_wo_grid_desc.GetLength(I1);
+
+    const auto Hi = in_n_c_hi_wi_grid_desc.GetLength(I2);
+    const auto Wi = in_n_c_hi_wi_grid_desc.GetLength(I3);
+
+    const auto Ho = out_n_k_ho_wo_grid_desc.GetLength(I2);
+    const auto Wo = out_n_k_ho_wo_grid_desc.GetLength(I3);
+
+    const auto Y = wei_k_c_y_x_grid_desc.GetLength(I2);
+    const auto X = wei_k_c_y_x_grid_desc.GetLength(I3);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto N1 = N / N0;
+    const auto C1 = C / C0;
+
+    // weight tensor
+    const auto wei_gk0_gm0_gm1_gk1_grid_desc =
+        transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C * Y * X)),
+                                    make_tuple(make_unmerge_transform(make_tuple(I1, K)),
+                                               make_unmerge_transform(make_tuple(C0, C1 * Y * X))),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<1, 2>{}, Sequence<3, 0>{}));
+
+    // input tensor
+    const auto in_n_c_hip_wip_grid_desc = transform_tensor_descriptor(
+        in_n_c_hi_wi_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n0_n1_c0_c1_y_ho_x_wo_grid_desc = transform_tensor_descriptor(
+        in_n_c_hip_wip_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(N0, N1)),
+                   make_unmerge_transform(make_tuple(C0, C1)),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW))),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}, Sequence<6, 7>{}));
+
+    const auto in_gk0_gn0_gn1_gk1_grid_desc = transform_tensor_descriptor(
+        in_n0_n1_c0_c1_y_ho_x_wo_grid_desc,
+        make_tuple(make_merge_transform(make_tuple(C1, Y, X)),
+                   make_pass_through_transform(N0),
+                   make_merge_transform(make_tuple(N1, Ho, Wo)),
+                   make_pass_through_transform(C0)),
+        make_tuple(Sequence<3, 4, 6>{}, Sequence<0>{}, Sequence<1, 5, 7>{}, Sequence<2>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    // output tensor
+    const auto out_n_k_howo_grid_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho * Wo));
+
+    const auto out_n0_n1_1_k_howo_grid_desc =
+        transform_tensor_descriptor(out_n_k_howo_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(N0, N1)),
+                                               make_unmerge_transform(make_tuple(I1, K)),
+                                               make_pass_through_transform(Ho * Wo)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                    make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}, Sequence<4>{}));
+
+    const auto out_gm0_gm1_gn0_gn1_grid_desc = transform_tensor_descriptor(
+        out_n0_n1_1_k_howo_grid_desc,
+        make_tuple(make_pass_through_transform(I1),
+                   make_pass_through_transform(K),
+                   make_pass_through_transform(N0),
+                   make_merge_transform_v2_magic_division(make_tuple(N1, Ho * Wo))),
+        make_tuple(Sequence<2>{}, Sequence<3>{}, Sequence<0>{}, Sequence<1, 4>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    return make_tuple(
+        wei_gk0_gm0_gm1_gk1_grid_desc, in_gk0_gn0_gn1_gk1_grid_desc, out_gm0_gm1_gn0_gn1_grid_desc);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/stream_config.hpp b/3rdparty/composable_kernel/include/ck/stream_config.hpp
new file mode 100644
index 0000000..70ca345
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/stream_config.hpp
@@ -0,0 +1,14 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+#include <hip/hip_fp16.h>
+
+struct StreamConfig
+{
+    hipStream_t stream_id_ = nullptr;
+    bool time_kernel_      = false;
+    int log_level_         = 0;
+};
diff --git a/3rdparty/composable_kernel/include/ck/tensor/static_tensor.hpp b/3rdparty/composable_kernel/include/ck/tensor/static_tensor.hpp
new file mode 100644
index 0000000..fee679f
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor/static_tensor.hpp
@@ -0,0 +1,273 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_STATIC_TENSOR_HPP
+#define CK_STATIC_TENSOR_HPP
+
+namespace ck {
+
+// StaticTensor for Scalar
+template <AddressSpaceEnum AddressSpace,
+          typename T,
+          typename TensorDesc,
+          bool InvalidElementUseNumericalZeroValue,
+          typename enable_if<TensorDesc::IsKnownAtCompileTime(), bool>::type = false>
+struct StaticTensor
+{
+    static constexpr auto desc_                  = TensorDesc{};
+    static constexpr index_t ndim_               = TensorDesc::GetNumOfDimension();
+    static constexpr index_t element_space_size_ = desc_.GetElementSpaceSize();
+
+    __host__ __device__ constexpr StaticTensor() : invalid_element_scalar_value_{0} {}
+
+    __host__ __device__ constexpr StaticTensor(T invalid_element_value)
+        : invalid_element_scalar_value_{invalid_element_value}
+    {
+    }
+
+    // read access
+    template <typename Idx,
+              typename enable_if<is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
+                                 bool>::type = false>
+    __host__ __device__ constexpr const T& operator[](Idx) const
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        if constexpr(is_valid)
+        {
+            return data_[Number<offset>{}];
+        }
+        else
+        {
+            if constexpr(InvalidElementUseNumericalZeroValue)
+            {
+                return zero_scalar_value_;
+            }
+            else
+            {
+                return invalid_element_scalar_value_;
+            }
+        }
+    }
+
+    // write access
+    template <typename Idx,
+              typename enable_if<is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
+                                 bool>::type = false>
+    __host__ __device__ constexpr T& operator()(Idx)
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        if constexpr(is_valid)
+        {
+            return data_(Number<offset>{});
+        }
+        else
+        {
+            return ignored_element_scalar_;
+        }
+    }
+
+    StaticBuffer<AddressSpace, T, element_space_size_, true> data_;
+    static constexpr T zero_scalar_value_ = T{0};
+    const T invalid_element_scalar_value_;
+    T ignored_element_scalar_;
+};
+
+// StaticTensor for vector
+template <AddressSpaceEnum AddressSpace,
+          typename S,
+          index_t ScalarPerVector,
+          typename TensorDesc,
+          bool InvalidElementUseNumericalZeroValue,
+          typename enable_if<TensorDesc::IsKnownAtCompileTime(), bool>::type = false>
+struct StaticTensorTupleOfVectorBuffer
+{
+    static constexpr auto desc_                  = TensorDesc{};
+    static constexpr index_t ndim_               = TensorDesc::GetNumOfDimension();
+    static constexpr index_t element_space_size_ = desc_.GetElementSpaceSize();
+
+    static constexpr index_t num_of_vector_ =
+        math::integer_divide_ceil(element_space_size_, ScalarPerVector);
+
+    using V = vector_type<S, ScalarPerVector>;
+
+    __host__ __device__ constexpr StaticTensorTupleOfVectorBuffer()
+        : invalid_element_scalar_value_{0}
+    {
+    }
+
+    __host__ __device__ constexpr StaticTensorTupleOfVectorBuffer(S invalid_element_value)
+        : invalid_element_scalar_value_{invalid_element_value}
+    {
+    }
+
+    // Get S
+    // Idx is for S, not V
+    template <typename Idx,
+              typename enable_if<is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
+                                 bool>::type = false>
+    __host__ __device__ constexpr const S& operator[](Idx) const
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        if constexpr(is_valid)
+        {
+            return data_[Number<offset>{}];
+        }
+        else
+        {
+            if constexpr(InvalidElementUseNumericalZeroValue)
+            {
+                return zero_scalar_value_;
+            }
+            else
+            {
+                return invalid_element_scalar_value_;
+            }
+        }
+    }
+
+    // Set S
+    // Idx is for S, not V
+    template <typename Idx,
+              typename enable_if<is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
+                                 bool>::type = false>
+    __host__ __device__ constexpr S& operator()(Idx)
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        if constexpr(is_valid)
+        {
+            return data_(Number<offset>{});
+        }
+        else
+        {
+            return ignored_element_scalar_;
+        }
+    }
+
+    // Get X
+    // Idx is for S, not X. Idx should be aligned with X
+    template <typename X,
+              typename Idx,
+              typename enable_if<has_same_scalar_type<S, X>::value &&
+                                     is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
+                                 bool>::type = false>
+    __host__ __device__ constexpr X GetAsType(Idx) const
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        if constexpr(is_valid)
+        {
+            return data_.template GetAsType<X>(Number<offset>{});
+        }
+        else
+        {
+            if constexpr(InvalidElementUseNumericalZeroValue)
+            {
+                // TODO: is this right way to initialize a vector?
+                return X{0};
+            }
+            else
+            {
+                // TODO: is this right way to initialize a vector?
+                return X{invalid_element_scalar_value_};
+            }
+        }
+    }
+
+    // Set X
+    // Idx is for S, not X. Idx should be aligned with X
+    template <typename X,
+              typename Idx,
+              typename enable_if<has_same_scalar_type<S, X>::value &&
+                                     is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
+                                 bool>::type = false>
+    __host__ __device__ constexpr void SetAsType(Idx, X x)
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        if constexpr(is_valid)
+        {
+            data_.template SetAsType<X>(Number<offset>{}, x);
+        }
+    }
+
+    // Get read access to V. No is_valid check
+    // Idx is for S, not V. Idx should be aligned with V
+    template <typename Idx>
+    __host__ __device__ constexpr const V& GetVectorTypeReference(Idx) const
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        return data_.GetVectorTypeReference(Number<offset>{});
+    }
+
+    // Get read access to V. No is_valid check
+    // Idx is for S, not V. Idx should be aligned with V
+    template <typename Idx>
+    __host__ __device__ constexpr V& GetVectorTypeReference(Idx)
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        return data_.GetVectorTypeReference(Number<offset>{});
+    }
+
+    StaticBufferTupleOfVector<AddressSpace, S, num_of_vector_, ScalarPerVector, true> data_;
+    static constexpr S zero_scalar_value_ = S{0};
+    const S invalid_element_scalar_value_ = S{0};
+    S ignored_element_scalar_;
+};
+
+template <AddressSpaceEnum AddressSpace,
+          typename T,
+          typename TensorDesc,
+          typename enable_if<TensorDesc::IsKnownAtCompileTime(), bool>::type = false>
+__host__ __device__ constexpr auto make_static_tensor(TensorDesc)
+{
+    return StaticTensor<AddressSpace, T, TensorDesc, true>{};
+}
+
+template <
+    AddressSpaceEnum AddressSpace,
+    typename T,
+    typename TensorDesc,
+    typename X,
+    typename enable_if<TensorDesc::IsKnownAtCompileTime(), bool>::type                   = false,
+    typename enable_if<is_same<remove_cvref_t<T>, remove_cvref_t<X>>::value, bool>::type = false>
+__host__ __device__ constexpr auto make_static_tensor(TensorDesc, X invalid_element_value)
+{
+    return StaticTensor<AddressSpace, T, TensorDesc, true>{invalid_element_value};
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/tensor_description/cluster_descriptor.hpp b/3rdparty/composable_kernel/include/ck/tensor_description/cluster_descriptor.hpp
new file mode 100644
index 0000000..0c9ea2f
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_description/cluster_descriptor.hpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+template <typename Lengths,
+          typename ArrangeOrder = typename arithmetic_sequence_gen<0, Lengths::Size(), 1>::type>
+__host__ __device__ constexpr auto make_cluster_descriptor(
+    const Lengths& lengths,
+    ArrangeOrder order = typename arithmetic_sequence_gen<0, Lengths::Size(), 1>::type{})
+{
+    constexpr index_t ndim_low = Lengths::Size();
+
+    const auto reordered_lengths = container_reorder_given_new2old(lengths, order);
+
+    const auto low_lengths = generate_tuple(
+        [&](auto idim_low) { return reordered_lengths[idim_low]; }, Number<ndim_low>{});
+
+    const auto transform = make_merge_transform(low_lengths);
+
+    constexpr auto low_dim_old_top_ids = ArrangeOrder{};
+
+    constexpr auto up_dim_new_top_ids = Sequence<0>{};
+
+    return make_single_stage_tensor_adaptor(
+        make_tuple(transform), make_tuple(low_dim_old_top_ids), make_tuple(up_dim_new_top_ids));
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_description/multi_index_transform.hpp b/3rdparty/composable_kernel/include/ck/tensor_description/multi_index_transform.hpp
new file mode 100644
index 0000000..4e4d759
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_description/multi_index_transform.hpp
@@ -0,0 +1,1954 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/multi_index.hpp"
+
+namespace ck {
+
+template <typename LowLength>
+struct PassThrough
+{
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths = decltype(make_tuple(LowLength{}));
+
+    UpLengths up_lengths_;
+
+    __host__ __device__ constexpr PassThrough() = default;
+
+    __host__ __device__ constexpr PassThrough(const LowLength& low_length)
+        : up_lengths_{make_tuple(low_length)}
+    {
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return 1; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ static constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                                  const UpIdx& idx_up)
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}];
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ static void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                                     const UpIdxDiff& idx_diff_up,
+                                                     LowIdx& idx_low,
+                                                     const UpIdx&,
+                                                     Number<Hack>)
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        idx_diff_low(I0) = idx_diff_up[I0];
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return true; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("PassThrough, ");
+        printf("up_lengths_");
+        print_multi_index(up_lengths_);
+        printf("}");
+    }
+};
+
+template <typename LowLength,
+          typename LeftPadLength,
+          typename RightPadLength,
+          bool SkipIsValidCheck = false>
+struct Pad
+{
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths = decltype(make_tuple(LowLength{} + LeftPadLength{} + RightPadLength{}));
+
+    UpLengths up_lengths_;
+    LeftPadLength left_pad_length_;
+    RightPadLength right_pad_length_;
+
+    __host__ __device__ constexpr Pad() = default;
+
+    __host__ __device__ constexpr Pad(const LowLength& low_length,
+                                      const LeftPadLength& left_pad_length,
+                                      const RightPadLength& right_pad_length)
+        : up_lengths_{make_tuple(low_length + left_pad_length + right_pad_length)},
+          left_pad_length_{left_pad_length},
+          right_pad_length_{right_pad_length}
+    {
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return 1; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}] - left_pad_length_;
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ static void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                                     const UpIdxDiff& idx_diff_up,
+                                                     LowIdx& idx_low,
+                                                     const UpIdx&,
+                                                     Number<Hack>)
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        idx_diff_low(I0) = idx_diff_up[I0];
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return true; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return SkipIsValidCheck;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& idx_up) const
+    {
+        return SkipIsValidCheck ||
+               ((idx_up[Number<0>{}] >= left_pad_length_) &&
+                (idx_up[Number<0>{}] < up_lengths_[Number<0>{}] - right_pad_length_));
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<LeftPadLength>::value &&
+               is_known_at_compile_time<RightPadLength>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("Pad, ");
+        printf("up_lengths_");
+        print_multi_index(up_lengths_);
+        printf("left_pad_length %d", index_t{left_pad_length_});
+        printf("right_pad_length %d", index_t{right_pad_length_});
+        printf("}");
+    }
+};
+
+template <typename LowLength, typename LeftPadLength, bool SkipIsValidCheck = false>
+struct LeftPad
+{
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths = decltype(make_tuple(LowLength{} + LeftPadLength{}));
+
+    UpLengths up_lengths_;
+    LeftPadLength left_pad_length_;
+
+    __host__ __device__ constexpr LeftPad() = default;
+
+    __host__ __device__ constexpr LeftPad(const LowLength& low_length,
+                                          const LeftPadLength& left_pad_length)
+        : up_lengths_{make_tuple(low_length + left_pad_length)}, left_pad_length_{left_pad_length}
+    {
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return 1; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}] - left_pad_length_;
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ static void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                                     const UpIdxDiff& idx_diff_up,
+                                                     LowIdx& idx_low,
+                                                     const UpIdx&,
+                                                     Number<Hack>)
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        idx_diff_low(I0) = idx_diff_up[I0];
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return true; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return SkipIsValidCheck;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& idx_up) const
+    {
+        return SkipIsValidCheck || (idx_up[Number<0>{}] >= left_pad_length_);
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<LeftPadLength>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("LeftPad, ");
+        printf("up_lengths_");
+        print_multi_index(up_lengths_);
+        printf("left_pad_length_ %d", index_t{left_pad_length_});
+        printf("}");
+    }
+};
+
+template <typename LowLength, typename RightPadLength, bool SkipIsValidCheck = false>
+struct RightPad
+{
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths = decltype(make_tuple(LowLength{} + RightPadLength{}));
+
+    UpLengths up_lengths_;
+    LowLength low_length_;
+    RightPadLength right_pad_length_;
+
+    __host__ __device__ constexpr RightPad() = default;
+
+    __host__ __device__ constexpr RightPad(const LowLength& low_length,
+                                           const RightPadLength& right_pad_length)
+        : up_lengths_{make_tuple(low_length + right_pad_length)},
+          low_length_{low_length},
+          right_pad_length_{right_pad_length}
+    {
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return 1; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ static constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                                  const UpIdx& idx_up)
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}];
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ static void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                                     const UpIdxDiff& idx_diff_up,
+                                                     LowIdx& idx_low,
+                                                     const UpIdx&,
+                                                     Number<Hack>)
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        idx_diff_low(I0) = idx_diff_up[I0];
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return true; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return SkipIsValidCheck;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& idx_up) const
+    {
+        return SkipIsValidCheck || (idx_up[Number<0>{}] < low_length_);
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<LowLength>::value &&
+               is_known_at_compile_time<RightPadLength>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("RightPad, ");
+        printf("up_lengths_");
+        print_multi_index(up_lengths_);
+        printf("low_length_ %d", index_t{low_length_});
+        printf("left_pad_length_ %d", index_t{right_pad_length_});
+        printf("}");
+    }
+};
+
+// idx_low = coefficients[0, ...nDimUp-1] * idx_up[0, ...nDimUp-1]
+// UpLengths and Coefficients can be either of the followings:
+//   1) Tuple of index_t, which is known at run-time, or
+//   2) Tuple of Number, which is known at compile-time, or
+//   3) Tuple of mixture of index_t and Number, which is known partially at run-time and partially
+//   at compile-time
+template <typename UpLengths,
+          typename Coefficients,
+          typename enable_if<UpLengths::Size() == Coefficients::Size(), bool>::type = false>
+struct Embed
+{
+    static constexpr index_t NDimUp = UpLengths::Size();
+
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<NDimUp>;
+
+    UpLengths up_lengths_;
+    Coefficients coefficients_;
+
+    __host__ __device__ constexpr Embed() = default;
+
+    __host__ __device__ constexpr Embed(const UpLengths& up_lengths,
+                                        const Coefficients& coefficients)
+        : up_lengths_{up_lengths}, coefficients_{coefficients}
+    {
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return 1; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return NDimUp; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == NDimUp,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = 0;
+
+        static_for<0, NDimUp, 1>{}([&idx_low, &idx_up, this](auto i) {
+            idx_low(Number<0>{}) += idx_up[i] * this->coefficients_[i];
+        });
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff& idx_diff_up,
+                                              LowIdx& idx_low,
+                                              const UpIdx&,
+                                              Number<Hack>) const
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == NDimUp &&
+                          LowIdx::Size() == 1 && UpIdx::Size() == NDimUp,
+                      "wrong! inconsistent # of dimension");
+
+        idx_diff_low(Number<0>{}) = 0;
+
+        static_for<0, NDimUp, 1>{}(
+            [&](auto i) { idx_diff_low(Number<0>{}) += idx_diff_up[i] * coefficients_[i]; });
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return true; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<Coefficients>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("Embed, ");
+        printf("up_lengths_ ");
+        print_multi_index(up_lengths_);
+        printf("coefficients_ ");
+        print_multi_index(coefficients_);
+        printf("}");
+    }
+};
+
+// Implementation of "Merge" transformation primitive that uses regular to do lowering of
+// multi-index and use carry-and-borrow check to do lowering of multi-index delta
+template <typename LowLengths>
+struct Merge_v1_carry_check
+{
+    static constexpr index_t NDimLow = LowLengths::Size();
+
+    using LowerIndex = MultiIndex<NDimLow>;
+    using UpperIndex = MultiIndex<1>;
+
+    using LowLengthsScan =
+        decltype(container_reverse_exclusive_scan(LowLengths{}, math::multiplies{}, Number<1>{}));
+
+    using UpLengths =
+        decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies{}, Number<1>{})));
+
+    LowLengths low_lengths_;
+    LowLengthsScan low_lengths_scan_;
+    UpLengths up_lengths_;
+
+    __host__ __device__ constexpr Merge_v1_carry_check() = default;
+
+    __host__ __device__ constexpr Merge_v1_carry_check(const LowLengths& low_lengths)
+        : low_lengths_{low_lengths},
+          low_lengths_scan_{
+              container_reverse_exclusive_scan(low_lengths, math::multiplies{}, Number<1>{})},
+          up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies{}, Number<1>{}))}
+    {
+        static_assert(LowerIndex::Size() == NDimLow, "wrong!");
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return NDimLow; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        index_t tmp = idx_up[Number<0>{}];
+
+        // normal division
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            idx_low(i) = tmp / this->low_lengths_scan_[i];
+            tmp -= idx_low[i] * this->low_lengths_scan_[i];
+        });
+
+        idx_low(Number<NDimLow - 1>{}) = tmp;
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex_1a(LowIdxDiff& idx_diff_low,
+                                                 const UpIdxDiff& idx_diff_up,
+                                                 LowIdx& idx_low,
+                                                 const UpIdx& /* idx_up_new */,
+                                                 Number<Hack>) const
+    {
+        static_assert(LowIdxDiff::Size() == NDimLow && UpIdxDiff::Size() == 1 &&
+                          LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        // CalculateLowerIndex(idx_diff_low_const) has multiple integer divisions.
+        // However,
+        //   1) If idx_diff_up is known at compile-time, then idx_diff_low_const
+        //   can be calculated at compile-time.
+        //   2) If idx_diff_up is not known at compile-time, but its value
+        //   doesn't change during the whole kernel execution, then
+        //   idx_diff_low_const also
+        //   doesn't change during the whole kernel execution. Compiler generated
+        //   ISA should
+        //   only caclculate idx_diff_low_const once and save it durinng the whole
+        //   kernel execution
+        // If neither 1) nor 2) is satisfied, then the calculation will also be
+        // computed at
+        //   run-time each time this function is called, and can be very expensive.
+        LowerIndex idx_diff_low_const;
+        LowerIndex idx_low_length_minus_idx_diff_low_const;
+        LowerIndex idx_low_length_plus_idx_diff_low_const;
+
+#if !CK_HACK_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE
+        index_t tmp = idx_diff_up[Number<0>{}];
+
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            idx_diff_low_const(i) = tmp / low_lengths_scan_[i];
+            tmp -= idx_diff_low_const[i] * low_lengths_scan_[i];
+        });
+
+        idx_diff_low_const(Number<NDimLow - 1>{}) = tmp;
+
+        static_for<0, NDimLow, 1>{}([&](auto i) {
+            idx_low_length_minus_idx_diff_low_const(i) = low_lengths_[i] - idx_diff_low_const[i];
+
+            idx_low_length_plus_idx_diff_low_const(i) = low_lengths_[i] + idx_diff_low_const[i];
+        });
+#else
+        // Hack: this force result into SGPR. Need to make sure the result is thread invariant
+        index_t tmp = idx_diff_up[Number<0>{}];
+
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            idx_diff_low_const(i) = __builtin_amdgcn_readfirstlane(tmp / low_lengths_scan_[i]);
+            tmp -= idx_diff_low_const[i] * low_lengths_scan_[i];
+        });
+
+        idx_diff_low_const(Number<NDimLow - 1>{}) = __builtin_amdgcn_readfirstlane(tmp);
+
+        static_for<0, NDimLow, 1>{}([&](auto i) {
+            idx_low_length_minus_idx_diff_low_const(i) =
+                __builtin_amdgcn_readfirstlane(low_lengths_[i] - idx_diff_low_const[i]);
+
+            idx_low_length_plus_idx_diff_low_const(i) =
+                __builtin_amdgcn_readfirstlane(low_lengths_[i] + idx_diff_low_const[i]);
+        });
+#endif
+
+        if constexpr(Hack == 1)
+        {
+            // do carry check on each low dimension in reversed order
+            // do not need to check the first dimension
+            index_t carry = 0;
+
+            static_for<NDimLow - 1, 0, -1>{}([&](auto i) {
+                index_t idx_low_tmp = idx_low[i] + carry;
+
+                bool do_carry = idx_low_tmp >= idx_low_length_minus_idx_diff_low_const[i];
+
+                idx_diff_low(i) =
+                    do_carry ? -idx_low_length_minus_idx_diff_low_const[i] : idx_diff_low_const[i];
+
+                idx_diff_low(i) += carry;
+
+                carry = do_carry ? 1 : 0;
+            });
+
+            idx_diff_low(Number<0>{}) = idx_diff_low_const[Number<0>{}] + carry;
+
+            idx_low += idx_diff_low;
+        }
+        else if constexpr(Hack == 2)
+        {
+            // do carry check on each low dimension in reversed order
+            // do not need to check the first dimension
+            index_t borrow = 0;
+
+            static_for<NDimLow - 1, 0, -1>{}([&](auto i) {
+                index_t idx_low_tmp = idx_low[i] - borrow;
+
+                bool do_borrow = idx_low_tmp < -idx_diff_low_const[i];
+
+                idx_diff_low(i) =
+                    do_borrow ? idx_low_length_plus_idx_diff_low_const[i] : idx_diff_low_const[i];
+
+                idx_diff_low(i) -= borrow;
+
+                borrow = do_borrow ? 1 : 0;
+            });
+
+            idx_diff_low(Number<0>{}) = idx_diff_low_const[Number<0>{}] - borrow;
+
+            idx_low += idx_diff_low;
+        }
+        else
+        {
+            // do carry check on each low dimension in reversed order
+            // do not need to check the first dimension
+            index_t carry = 0;
+
+            static_for<NDimLow - 1, 0, -1>{}([&](auto i) {
+                index_t idx_low_tmp = idx_low[i] + carry;
+
+                bool do_carry  = idx_low_tmp >= idx_low_length_minus_idx_diff_low_const[i];
+                bool do_borrow = idx_low_tmp < -idx_diff_low_const[i];
+
+                idx_diff_low(i) =
+                    do_carry ? -idx_low_length_minus_idx_diff_low_const[i] : idx_diff_low_const[i];
+                idx_diff_low(i) =
+                    do_borrow ? idx_low_length_plus_idx_diff_low_const[i] : idx_diff_low[i];
+
+                idx_diff_low(i) += carry;
+
+                carry = do_carry ? 1 : 0;
+                carry = do_borrow ? -1 : carry;
+            });
+
+            idx_diff_low(Number<0>{}) = idx_diff_low_const[Number<0>{}] + carry;
+
+            idx_low += idx_diff_low;
+        }
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex_1b(LowIdxDiff& idx_diff_low,
+                                                 const UpIdxDiff& idx_diff_up,
+                                                 LowIdx& idx_low,
+                                                 const UpIdx& /* idx_up_new */,
+                                                 Number<Hack>) const
+    {
+        static_assert(LowIdxDiff::Size() == NDimLow && UpIdxDiff::Size() == 1 &&
+                          LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        // CalculateLowerIndex(idx_diff_low_const) has multiple integer divisions.
+        // However,
+        //   1) If idx_diff_up is known at compile-time, then idx_diff_low_const
+        //   can be calculated at compile-time.
+        //   2) If idx_diff_up is not known at compile-time, but its value
+        //   doesn't change during the whole kernel execution, then
+        //   idx_diff_low_const also
+        //   doesn't change during the whole kernel execution. Compiler generated
+        //   ISA should
+        //   only caclculate idx_diff_low_const once and save it durinng the whole
+        //   kernel execution
+        // If neither 1) nor 2) is satisfied, then the calculation will also be
+        // computed at
+        //   run-time each time this function is called, and can be very expensive.
+        LowerIndex idx_diff_low_const;
+        LowerIndex idx_low_length_minus_idx_diff_low_const;
+        LowerIndex idx_low_length_plus_idx_diff_low_const;
+
+#if !CK_HACK_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE
+        index_t tmp = idx_diff_up[Number<0>{}];
+
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            idx_diff_low_const(i) = tmp / low_lengths_scan_[i];
+            tmp -= idx_diff_low_const[i] * low_lengths_scan_[i];
+        });
+
+        idx_diff_low_const(Number<NDimLow - 1>{}) = tmp;
+
+        static_for<0, NDimLow, 1>{}([&](auto i) {
+            idx_low_length_minus_idx_diff_low_const(i) = low_lengths_[i] - idx_diff_low_const[i];
+
+            idx_low_length_plus_idx_diff_low_const(i) = low_lengths_[i] + idx_diff_low_const[i];
+        });
+#else
+        // Hack: this force result into SGPR. Need to make sure the result is thread invariant
+        index_t tmp = idx_diff_up[Number<0>{}];
+
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            idx_diff_low_const(i) = __builtin_amdgcn_readfirstlane(tmp / low_lengths_scan_[i]);
+            tmp -= idx_diff_low_const[i] * low_lengths_scan_[i];
+        });
+
+        idx_diff_low_const(Number<NDimLow - 1>{}) = __builtin_amdgcn_readfirstlane(tmp);
+
+        static_for<0, NDimLow, 1>{}([&](auto i) {
+            idx_low_length_minus_idx_diff_low_const(i) =
+                __builtin_amdgcn_readfirstlane(low_lengths_[i] - idx_diff_low_const[i]);
+
+            idx_low_length_plus_idx_diff_low_const(i) = low_lengths_[i] + idx_diff_low_const[i];
+        });
+#endif
+
+        if constexpr(Hack == 1)
+        {
+            // do carry check on each low dimension in reversed order
+            // do not need to check the first dimension
+            index_t carry = 0;
+
+            static_for<NDimLow - 1, 0, -1>{}([&](auto i) {
+                index_t idx_low_tmp = idx_low[i] + carry;
+
+                bool do_carry = idx_low_tmp >= idx_low_length_minus_idx_diff_low_const[i];
+
+                idx_diff_low(i) =
+                    do_carry ? -idx_low_length_minus_idx_diff_low_const[i] : idx_diff_low_const[i];
+
+                idx_diff_low(i) += carry;
+
+                carry = do_carry ? 1 : 0;
+            });
+
+            idx_diff_low(Number<0>{}) = idx_diff_low_const[Number<0>{}] + carry;
+
+            idx_low += idx_diff_low;
+        }
+        else if constexpr(Hack == 2)
+        {
+            // do carry check on each low dimension in reversed order
+            // do not need to check the first dimension
+            index_t borrow = 0;
+
+            static_for<NDimLow - 1, 0, -1>{}([&](auto i) {
+                index_t negative_idx_low_tmp = borrow - idx_low[i];
+
+                bool do_borrow = negative_idx_low_tmp > idx_diff_low_const[i];
+
+                idx_diff_low(i) =
+                    do_borrow ? idx_low_length_plus_idx_diff_low_const[i] : idx_diff_low_const[i];
+
+                idx_diff_low(i) -= borrow;
+
+                borrow = do_borrow ? 1 : 0;
+            });
+
+            idx_diff_low(Number<0>{}) = idx_diff_low_const[Number<0>{}] - borrow;
+
+            idx_low += idx_diff_low;
+        }
+        else
+        {
+            // do carry check on each low dimension in reversed order
+            // do not need to check the first dimension
+            index_t carry = 0;
+
+            static_for<NDimLow - 1, 0, -1>{}([&](auto i) {
+                index_t idx_low_tmp = idx_low[i] + carry;
+
+                bool do_carry  = idx_low_tmp >= idx_low_length_minus_idx_diff_low_const[i];
+                bool do_borrow = idx_low_tmp < -idx_diff_low_const[i];
+
+                idx_diff_low(i) =
+                    do_carry ? -idx_low_length_minus_idx_diff_low_const[i] : idx_diff_low_const[i];
+                idx_diff_low(i) =
+                    do_borrow ? idx_low_length_plus_idx_diff_low_const[i] : idx_diff_low[i];
+
+                idx_diff_low(i) += carry;
+
+                carry = do_carry ? 1 : 0;
+                carry = do_borrow ? -1 : carry;
+            });
+
+            idx_diff_low(Number<0>{}) = idx_diff_low_const[Number<0>{}] + carry;
+
+            idx_low += idx_diff_low;
+        }
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex_2(LowIdxDiff& idx_diff_low,
+                                                const UpIdxDiff& idx_diff_up,
+                                                LowIdx& idx_low,
+                                                const UpIdx& /* idx_up_new */,
+                                                Number<Hack>) const
+    {
+        static_assert(LowIdxDiff::Size() == NDimLow && UpIdxDiff::Size() == 1 &&
+                          LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        // CalculateLowerIndex(idx_diff_low_const) has multiple integer divisions.
+        // However,
+        //   1) If idx_diff_up is known at compile-time, then idx_diff_low_const
+        //   can be calculated at compile-time.
+        //   2) If idx_diff_up is not known at compile-time, but its value
+        //   doesn't change during the whole kernel execution, then
+        //   idx_diff_low_const also
+        //   doesn't change during the whole kernel execution. Compiler generated
+        //   ISA should
+        //   only caclculate idx_diff_low_const once and save it durinng the whole
+        //   kernel execution
+        // If neither 1) nor 2) is satisfied, then the calculation will also be
+        //   computed at run-time each time this function is called, and can be
+        //   very expensive.
+        LowerIndex idx_diff_low_const;
+
+#if !CK_HACK_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE
+        index_t tmp = idx_diff_up[Number<0>{}];
+
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            idx_diff_low_const(i) = tmp / low_lengths_scan_[i];
+            tmp -= idx_diff_low_const[i] * low_lengths_scan_[i];
+        });
+
+        idx_diff_low_const(Number<NDimLow - 1>{}) = tmp;
+#else
+        // Hack: this force result into SGPR. Need to make sure the result is thread invariant
+        index_t tmp = idx_diff_up[Number<0>{}];
+
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            idx_diff_low_const(i) = __builtin_amdgcn_readfirstlane(tmp / low_lengths_scan_[i]);
+            tmp -= idx_diff_low_const[i] * low_lengths_scan_[i];
+        });
+
+        idx_diff_low_const(Number<NDimLow - 1>{}) = __builtin_amdgcn_readfirstlane(tmp);
+#endif
+
+        if constexpr(Hack == 1)
+        {
+            // do carry check on each low dimension in reversed order
+            // do not need to check the first dimension
+            bool do_carry = 0;
+
+            static_for<NDimLow - 1, 0, -1>{}([&](auto i) {
+                idx_diff_low(i) = idx_diff_low_const[i] + do_carry;
+
+                index_t idx_low_tmp = idx_low[i] + idx_diff_low[i];
+
+                do_carry = idx_low_tmp >= low_lengths_[i];
+
+#if 0
+                // TODO: use exec-mask inline asm, which use 1 VALU
+                if(do_carry)
+                {
+                    idx_diff_low(i) -= low_lengths_[i];
+                }
+#elif 1
+                // this use 2 VALU
+                idx_diff_low(i) = do_carry ? idx_diff_low[i] - low_lengths_[i] : idx_diff_low[i];
+#elif 1
+                // this use 2 VALU
+                index_t idx_diff_low_tmp = idx_diff_low[i] - low_lengths_[i];
+                idx_diff_low(i)          = do_carry ? idx_diff_low_tmp : idx_diff_low[i];
+#endif
+
+                idx_low(i) += idx_diff_low[i];
+            });
+
+            constexpr auto I0 = Number<0>{};
+
+            idx_diff_low(I0) = idx_diff_low_const[I0] + do_carry;
+
+            idx_low(I0) += idx_diff_low[I0];
+        }
+        else if constexpr(Hack == 2)
+        {
+            // do borrow check on each low dimension in reversed order
+            // do not need to check the first dimension
+            bool do_borrow = 0;
+
+            static_for<NDimLow - 1, 0, -1>{}([&](auto i) {
+                idx_diff_low(i) = idx_diff_low_const[i] - do_borrow;
+
+                index_t idx_low_tmp = idx_low[i] + idx_diff_low[i];
+
+                do_borrow = idx_low_tmp < 0;
+
+#if 0
+                // TODO: use exec-mask inline asm
+                if(do_borrow)
+                {
+                    idx_diff_low(i) += low_lengths_[i];
+                }
+#elif 1
+                idx_diff_low(i) = do_borrow ? idx_diff_low[i] + low_lengths_[i] : idx_diff_low[i];
+#elif 1
+                index_t idx_diff_low_tmp = idx_diff_low[i] + low_lengths_[i];
+                idx_diff_low(i)          = do_borrow ? idx_diff_low_tmp : idx_diff_low[i];
+#endif
+
+                idx_low(i) += idx_diff_low[i];
+            });
+
+            constexpr auto I0 = Number<0>{};
+
+            idx_diff_low(I0) = idx_diff_low_const[I0] - do_borrow;
+
+            idx_low(I0) += idx_diff_low[I0];
+        }
+        else
+        {
+            // not implemented
+        }
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff& idx_diff_up,
+                                              LowIdx& idx_low,
+                                              const UpIdx& idx_up_new,
+                                              Number<Hack>) const
+    {
+#if 1
+        UpdateLowerIndex_1a(idx_diff_low, idx_diff_up, idx_low, idx_up_new, Number<Hack>{});
+#elif 0
+        UpdateLowerIndex_1b(idx_diff_low, idx_diff_up, idx_low, idx_up_new, Number<Hack>{});
+#else
+        UpdateLowerIndex_2(idx_diff_low, idx_diff_up, idx_low, idx_up_new, Number<Hack>{});
+#endif
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return false; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<LowLengths>::value &&
+               is_known_at_compile_time<LowLengthsScan>::value &&
+               is_known_at_compile_time<UpLengths>::value;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("Merge_v1_carry_check, ");
+        printf("low_lengths_ ");
+        print_multi_index(low_lengths_);
+        printf("low_lengths_scan_ ");
+        print_multi_index(low_lengths_scan_);
+        printf("up_lengths_ ");
+        print_multi_index(up_lengths_);
+        printf("}");
+    }
+};
+
+template <typename LowLengths>
+struct lambda_merge_generate_MagicDivision_calculate_magic_multiplier
+{
+    template <index_t I>
+    __host__ __device__ constexpr auto operator()(Number<I> i) const
+    {
+        return MagicDivision::CalculateMagicMultiplier(LowLengths{}[i]);
+    }
+};
+
+template <typename LowLengths>
+struct lambda_merge_generate_MagicDivision_calculate_magic_shift
+{
+    template <index_t I>
+    __host__ __device__ constexpr auto operator()(Number<I> i) const
+    {
+        return MagicDivision::CalculateMagicShift(LowLengths{}[i]);
+    }
+};
+
+// Implementation of "Merge" transformation primitive that uses magic-number-division to do lowering
+// of both multi-index and delta of multi-index
+// Caution:
+//   1. The magic number division implementation being used would produce correct result if the
+//   dividended is uint32_t and its value is with in 31-bit value range of uint32_t.
+//   2. The magic number division for int32_t dividened has not been implemented, the int32_t
+//   dividend would be bit-wise interpreted as uint32_t and magic number division implementation for
+//   uint32_t is then used.
+//   3. For Merge primitive, upper-index is the dividend.
+//   4. When upper-index is uint32_t, its value need to be within 31-bit range.
+//   5. When upper-index is int32_t type (when index_t is int32_t), its value need to be
+//   non-negative.
+template <typename LowLengths>
+struct Merge_v2_magic_division
+{
+    static constexpr index_t NDimLow = LowLengths::Size();
+
+    using LowerIndex = MultiIndex<NDimLow>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths =
+        decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies{}, Number<1>{})));
+
+    using LowLengthsMagicDivisorMultipiler = decltype(
+        generate_tuple(lambda_merge_generate_MagicDivision_calculate_magic_multiplier<LowLengths>{},
+                       Number<NDimLow>{}));
+
+    using LowLengthsMagicDivisorShift = decltype(
+        generate_tuple(lambda_merge_generate_MagicDivision_calculate_magic_shift<LowLengths>{},
+                       Number<NDimLow>{}));
+
+    LowLengths low_lengths_;
+    LowLengthsMagicDivisorMultipiler low_lengths_magic_divisor_multiplier_;
+    LowLengthsMagicDivisorShift low_lengths_magic_divisor_shift_;
+    UpLengths up_lengths_;
+
+    __host__ __device__ constexpr Merge_v2_magic_division() = default;
+
+    __host__ __device__ constexpr Merge_v2_magic_division(const LowLengths& low_lengths)
+        : low_lengths_{low_lengths},
+          low_lengths_magic_divisor_multiplier_{generate_tuple(
+              [&](auto i) { return MagicDivision::CalculateMagicMultiplier(low_lengths[i]); },
+              Number<NDimLow>{})},
+          low_lengths_magic_divisor_shift_{generate_tuple(
+              [&](auto i) { return MagicDivision::CalculateMagicShift(low_lengths[i]); },
+              Number<NDimLow>{})},
+          up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies{}, Number<1>{}))}
+    {
+        static_assert(LowerIndex::Size() == NDimLow, "wrong!");
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return NDimLow; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        index_t tmp = idx_up[Number<0>{}];
+
+        static_for<NDimLow - 1, 0, -1>{}([&, this](auto i) {
+            index_t tmp2 =
+                MagicDivision::DoMagicDivision(tmp,
+                                               this->low_lengths_magic_divisor_multiplier_[i],
+                                               this->low_lengths_magic_divisor_shift_[i]);
+            idx_low(i) = tmp - tmp2 * this->low_lengths_[i];
+            tmp        = tmp2;
+        });
+
+        idx_low(Number<0>{}) = tmp;
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff&,
+                                              LowIdx& idx_low,
+                                              const UpIdx& idx_up_new,
+                                              Number<Hack>) const
+    {
+        static_assert(LowIdxDiff::Size() == NDimLow && UpIdxDiff::Size() == 1 &&
+                          LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        index_t tmp = idx_up_new[Number<0>{}];
+
+        static_for<NDimLow - 1, 0, -1>{}([&, this](auto i) {
+            index_t tmp2 =
+                MagicDivision::DoMagicDivision(tmp,
+                                               this->low_lengths_magic_divisor_multiplier_[i],
+                                               this->low_lengths_magic_divisor_shift_[i]);
+
+            index_t idx_low_old = idx_low[i];
+
+            idx_low(i) = tmp - tmp2 * this->low_lengths_[i];
+            tmp        = tmp2;
+
+            idx_diff_low(i) = idx_low[i] - idx_low_old;
+        });
+
+        idx_diff_low(Number<0>{}) = tmp - idx_low(Number<0>{});
+
+        idx_low(Number<0>{}) = tmp;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return false; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<LowLengths>::value &&
+               is_known_at_compile_time<LowLengthsMagicDivisorMultipiler>::value &&
+               is_known_at_compile_time<LowLengthsMagicDivisorShift>::value &&
+               is_known_at_compile_time<UpLengths>::value;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("Merge_v2_magic_division, ");
+        printf("low_lengths_ ");
+        print_multi_index(low_lengths_);
+        printf("low_lengths_magic_divisor_multiplier_ ");
+        print_multi_index(low_lengths_magic_divisor_multiplier_);
+        printf("low_lengths_magic_divisor_shift_ ");
+        print_multi_index(low_lengths_magic_divisor_shift_);
+        printf("up_lengths_ ");
+        print_multi_index(up_lengths_);
+        printf("}");
+    }
+};
+
+// Implementation of "Merge" transformation primitive that uses magic-number-division to do lowering
+// of both multi-index and delta of multi-index
+// Caution:
+//   1. The magic number division implementation being used would produce correct result if the
+//   dividended is uint32_t and its value is with in 31-bit value range of uint32_t.
+//   2. The magic number division for int32_t dividened has not been implemented, the int32_t
+//   dividend would be bit-wise interpreted as uint32_t and magic number division implementation for
+//   uint32_t is then used.
+//   3. For Merge primitive, upper-index is the dividend.
+//   4. When upper-index is uint32_t, its value need to be within 31-bit range.
+//   5. When upper-index is int32_t type (when index_t is int32_t), its value need to be
+//   non-negative.
+template <typename LowLengths>
+struct Merge_v2r2_magic_division
+{
+    static constexpr index_t NDimLow = LowLengths::Size();
+
+    using LowerIndex = MultiIndex<NDimLow>;
+    using UpperIndex = MultiIndex<1>;
+
+    using LowLengthsScan =
+        decltype(container_reverse_exclusive_scan(LowLengths{}, math::multiplies{}, Number<1>{}));
+
+    using UpLengths =
+        decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies{}, Number<1>{})));
+
+    using LowLengthsScanMagicDivisorMultipiler = decltype(generate_tuple(
+        lambda_merge_generate_MagicDivision_calculate_magic_multiplier<LowLengthsScan>{},
+        Number<NDimLow>{}));
+
+    using LowLengthsScanMagicDivisorShift = decltype(
+        generate_tuple(lambda_merge_generate_MagicDivision_calculate_magic_shift<LowLengthsScan>{},
+                       Number<NDimLow>{}));
+
+    LowLengths low_lengths_;
+    LowLengthsScan low_lengths_scan_;
+    LowLengthsScanMagicDivisorMultipiler low_lengths_scan_magic_divisor_multiplier_;
+    LowLengthsScanMagicDivisorShift low_lengths_scan_magic_divisor_shift_;
+    UpLengths up_lengths_;
+
+    __host__ __device__ constexpr Merge_v2r2_magic_division() = default;
+
+    __host__ __device__ constexpr Merge_v2r2_magic_division(const LowLengths& low_lengths)
+        : low_lengths_{low_lengths},
+          low_lengths_scan_{
+              container_reverse_exclusive_scan(low_lengths, math::multiplies{}, Number<1>{})},
+          low_lengths_scan_magic_divisor_multiplier_{generate_tuple(
+              [&](auto i) { return MagicDivision::CalculateMagicMultiplier(low_lengths_scan_[i]); },
+              Number<NDimLow>{})},
+          low_lengths_scan_magic_divisor_shift_{generate_tuple(
+              [&](auto i) { return MagicDivision::CalculateMagicShift(low_lengths_scan_[i]); },
+              Number<NDimLow>{})},
+          up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies{}, Number<1>{}))}
+    {
+        static_assert(LowerIndex::Size() == NDimLow, "wrong!");
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return NDimLow; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        index_t tmp = idx_up[Number<0>{}];
+
+        static_for<0, NDimLow - 1, 1>{}([&, this](auto i) {
+            idx_low(i) =
+                MagicDivision::DoMagicDivision(tmp,
+                                               this->low_lengths_scan_magic_divisor_multiplier_[i],
+                                               this->low_lengths_scan_magic_divisor_shift_[i]);
+
+            tmp -= idx_low[i] * this->low_lengths_scan_[i];
+        });
+
+        idx_low(Number<NDimLow - 1>{}) = tmp;
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff&,
+                                              LowIdx& idx_low,
+                                              const UpIdx& idx_up_new,
+                                              Number<Hack>) const
+    {
+        static_assert(LowIdxDiff::Size() == NDimLow && UpIdxDiff::Size() == 1 &&
+                          LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        index_t tmp = idx_up_new[Number<0>{}];
+
+        static_for<0, NDimLow - 1, 1>{}([&, this](auto i) {
+            index_t idx_low_old = idx_low[i];
+
+            idx_low(i) =
+                MagicDivision::DoMagicDivision(tmp,
+                                               this->low_lengths_scan_magic_divisor_multiplier_[i],
+                                               this->low_lengths_scan_magic_divisor_shift_[i]);
+
+            idx_diff_low(i) = idx_low[i] - idx_low_old;
+
+            tmp -= idx_low[i] * this->low_lengths_scan_[i];
+        });
+
+        idx_diff_low(Number<NDimLow - 1>{}) = tmp - idx_low[Number<NDimLow - 1>{}];
+
+        idx_low(Number<NDimLow - 1>{}) = tmp;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return false; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<LowLengths>::value &&
+               is_known_at_compile_time<LowLengthsScanMagicDivisorMultipiler>::value &&
+               is_known_at_compile_time<LowLengthsScanMagicDivisorShift>::value &&
+               is_known_at_compile_time<UpLengths>::value;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("Merge_v2r2_magic_division, ");
+        printf("low_lengths_ ");
+        print_multi_index(low_lengths_);
+        printf("low_lengths_scan ");
+        print_multi_index(low_lengths_scan_);
+        printf("low_lengths_scan_magic_divisor_multiplier_ ");
+        print_multi_index(low_lengths_scan_magic_divisor_multiplier_);
+        printf("low_lengths_scan_magic_divisor_shift_ ");
+        print_multi_index(low_lengths_scan_magic_divisor_shift_);
+        printf("up_lengths_ ");
+        print_multi_index(up_lengths_);
+        printf("}");
+    }
+};
+
+// Implementation of "Merge" transformation primitive that uses division and mod. It is supposed to
+// be used for low_lengths that are known at compile time and are power of 2, otherwise performance
+// will be very bad
+template <typename LowLengths>
+struct Merge_v3_division_mod
+{
+    static constexpr index_t NDimLow = LowLengths::Size();
+
+    using LowerIndex = MultiIndex<NDimLow>;
+    using UpperIndex = MultiIndex<1>;
+
+    using LowLengthsScan =
+        decltype(container_reverse_exclusive_scan(LowLengths{}, math::multiplies{}, Number<1>{}));
+
+    using UpLengths =
+        decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies{}, Number<1>{})));
+
+    LowLengths low_lengths_;
+    LowLengthsScan low_lengths_scan_;
+    UpLengths up_lengths_;
+
+    __host__ __device__ constexpr Merge_v3_division_mod() = default;
+
+    __host__ __device__ constexpr Merge_v3_division_mod(const LowLengths& low_lengths)
+        : low_lengths_{low_lengths},
+          low_lengths_scan_{
+              container_reverse_exclusive_scan(low_lengths, math::multiplies{}, Number<1>{})},
+          up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies{}, Number<1>{}))}
+    {
+        static_assert(LowerIndex::Size() == NDimLow, "wrong!");
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return NDimLow; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        index_t tmp = idx_up[Number<0>{}];
+
+        // division and mod
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            idx_low(i) = tmp / this->low_lengths_scan_[i];
+            tmp %= this->low_lengths_scan_[i];
+        });
+
+        idx_low(Number<NDimLow - 1>{}) = tmp;
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff&,
+                                              LowIdx& idx_low,
+                                              const UpIdx& idx_up_new,
+                                              Number<Hack>) const
+    {
+        static_assert(LowIdxDiff::Size() == NDimLow && UpIdxDiff::Size() == 1 &&
+                          LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0   = Number<0>{};
+        constexpr auto INm1 = Number<NDimLow - 1>{};
+
+        index_t tmp = idx_up_new[I0];
+
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            const index_t tmp2 = idx_low[i];
+            idx_low(i)         = tmp / this->low_lengths_scan_[i];
+            idx_diff_low(i)    = idx_low[i] - tmp2;
+            tmp %= this->low_lengths_scan_[i];
+        });
+
+        const index_t tmp2 = idx_low[INm1];
+        idx_low(INm1)      = tmp;
+        idx_diff_low(INm1) = idx_low[INm1] - tmp2;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return false; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<LowLengths>::value &&
+               is_known_at_compile_time<LowLengthsScan>::value &&
+               is_known_at_compile_time<UpLengths>::value;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("Merge_v3_direct_division_mod, ");
+        printf("low_lengths_ ");
+        print_multi_index(low_lengths_);
+        printf("low_lengths_scan_ ");
+        print_multi_index(low_lengths_scan_);
+        printf("up_lengths_ ");
+        print_multi_index(up_lengths_);
+        printf("}");
+    }
+};
+
+template <typename UpLengths, bool Use24BitIntegerCalculation>
+struct UnMerge
+{
+    static constexpr index_t NDimUp = UpLengths::Size();
+
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<NDimUp>;
+
+    using UpLengthsScan =
+        decltype(container_reverse_exclusive_scan(UpLengths{}, math::multiplies{}, Number<1>{}));
+
+    UpLengths up_lengths_;
+    UpLengthsScan up_lengths_scan_;
+
+    __host__ __device__ constexpr UnMerge() = default;
+
+    __host__ __device__ constexpr UnMerge(const UpLengths& up_lengths)
+        : up_lengths_{up_lengths},
+          up_lengths_scan_{
+              container_reverse_exclusive_scan(up_lengths, math::multiplies{}, Number<1>{})}
+    {
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return 1; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return NDimUp; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        if constexpr(!Use24BitIntegerCalculation)
+        {
+            idx_low(Number<0>{}) = idx_up[Number<NDimUp - 1>{}];
+
+            static_for<0, NDimUp - 1, 1>{}(
+                [&](auto i) { idx_low(Number<0>{}) += idx_up[i] * up_lengths_scan_[i]; });
+        }
+        else
+        {
+            idx_low(Number<0>{}) = idx_up[Number<NDimUp - 1>{}];
+
+            static_for<0, NDimUp - 1, 1>{}([&](auto i) {
+                idx_low(Number<0>{}) =
+                    (0x00ffffff & idx_low[Number<0>{}]) +
+                    (0x00ffffff & idx_up[i]) * (0x00ffffff & up_lengths_scan_[i]);
+            });
+        }
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff& idx_diff_up,
+                                              LowIdx& idx_low,
+                                              const UpIdx&,
+                                              Number<Hack>) const
+    {
+        CalculateLowerIndex(idx_diff_low, idx_diff_up);
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return true; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<UpLengthsScan>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("UnMerge, ");
+        printf("up_lengths_");
+        print_multi_index(up_lengths_);
+        printf("up_lengths_scan_");
+        print_multi_index(up_lengths_scan_);
+        printf("}");
+    }
+};
+
+template <typename LowerIndex>
+struct Freeze
+{
+    LowerIndex low_idx_;
+
+    __host__ __device__ constexpr Freeze() = default;
+
+    __host__ __device__ constexpr Freeze(const LowerIndex& low_idx) : low_idx_{low_idx} {}
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return 1; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 0; }
+
+    __host__ __device__ static constexpr auto GetUpperLengths() { return Tuple<>{}; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& /* idx_up */) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 0,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = low_idx_;
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ static void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                                     const UpIdxDiff& /* idx_diff_up */,
+                                                     LowIdx& /* idx_low */,
+                                                     const UpIdx& /* idx_up_new */,
+                                                     Number<Hack>)
+    {
+        idx_diff_low(Number<0>{}) = 0;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return true; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<LowerIndex>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Freeze");
+        printf("low_idx_ %d", index_t{low_idx_});
+    }
+};
+
+// Insert a dangling upper dimension without lower dimension
+template <typename UpperLength>
+struct Insert
+{
+    using UpLengths = decltype(make_tuple(UpperLength{}));
+
+    UpLengths up_lengths_;
+
+    __host__ __device__ constexpr Insert() = default;
+
+    __host__ __device__ constexpr Insert(const UpperLength& up_length)
+        : up_lengths_{make_tuple(up_length)}
+    {
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return 0; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr auto GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx&, const UpIdx&) const
+    {
+        static_assert(LowIdx::Size() == 0 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ static void
+    UpdateLowerIndex(LowIdxDiff&, const UpIdxDiff&, LowIdx&, const UpIdx&, Number<Hack>)
+    {
+        static_assert(LowIdxDiff::Size() == 0 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 0 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return true; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpperLength>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Insert");
+        print_multi_index(up_lengths_);
+    }
+};
+
+template <typename VectorSize, typename UpLength>
+struct Vectorize
+{
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths = decltype(make_tuple(UpLength{}));
+
+    UpLengths up_lengths_;
+    VectorSize vector_size_;
+
+    __host__ __device__ constexpr Vectorize() = default;
+
+    __host__ __device__ constexpr Vectorize(const VectorSize& vector_size,
+                                            const UpLength& up_length)
+        : vector_size_{vector_size}, up_lengths_{make_tuple(up_length)}
+    {
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return 1; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = vector_size_ * idx_up[Number<0>{}];
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff& idx_diff_up,
+                                              LowIdx& idx_low,
+                                              const UpIdx&,
+                                              Number<Hack>) const
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        idx_diff_low(I0) = vector_size_ * idx_diff_up[I0];
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return true; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("Vectorize, ");
+        printf("up_lengths_");
+        print_multi_index(up_lengths_);
+        printf("}");
+    }
+};
+
+template <typename LowLength, typename SliceBegin, typename SliceEnd>
+struct Slice
+{
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths = decltype(make_tuple(SliceEnd{} - SliceBegin{}));
+
+    UpLengths up_lengths_;
+    SliceBegin slice_begin_;
+    SliceEnd slice_end_;
+
+    __host__ __device__ constexpr Slice() = default;
+
+    __host__ __device__ constexpr Slice(const LowLength&,
+                                        const SliceBegin& slice_begin,
+                                        const SliceEnd& slice_end)
+        : up_lengths_{make_tuple(slice_end - slice_begin)},
+          slice_begin_{slice_begin},
+          slice_end_{slice_end}
+    {
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return 1; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}] + slice_begin_;
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ static void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                                     const UpIdxDiff& idx_diff_up,
+                                                     LowIdx& idx_low,
+                                                     const UpIdx&,
+                                                     Number<Hack>)
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        idx_diff_low(I0) = idx_diff_up[I0];
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return true; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ constexpr bool IsValidUpperIndexMappedToValidLowerIndex(const UpIdx&) const
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<SliceBegin>::value &&
+               is_known_at_compile_time<SliceEnd>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("Slice, ");
+        printf("up_lengths_");
+        print_multi_index(up_lengths_);
+        printf("slice_begin_ %d", index_t{slice_begin_});
+        printf("slice_end %d", index_t{slice_end_});
+        printf("}");
+    }
+};
+
+/*
+ * \brief lower_idx = upper_idx % modulus.
+ * TODO: Need an improved implementation since the modulo operation is expensive.
+ */
+template <typename Modulus, typename UpLength>
+struct Modulo
+{
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+    using UpLengths  = decltype(make_tuple(UpLength{}));
+
+    Modulus modulus_;
+    UpLengths up_lengths_;
+
+    __host__ __device__ constexpr Modulo() = default;
+
+    __host__ __device__ constexpr Modulo(const Modulus& modulus, const UpLength& up_length)
+        : modulus_{modulus}, up_lengths_{make_tuple(up_length)}
+    {
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return 1; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}] % modulus_;
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff& idx_diff_up,
+                                              LowIdx& idx_low,
+                                              const UpIdx& up_idx,
+                                              Number<Hack>) const
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        const auto idx_low_old = idx_low;
+        idx_low(I0)            = (up_idx(I0) + idx_diff_up(I0)) % modulus_;
+        idx_diff_low(I0)       = idx_low - idx_low_old;
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return false; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("Modulus, ");
+        printf("up_lengths_");
+        print_multi_index(up_lengths_);
+        printf("}");
+    }
+};
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_description/multi_index_transform_helper.hpp b/3rdparty/composable_kernel/include/ck/tensor_description/multi_index_transform_helper.hpp
new file mode 100644
index 0000000..044a903
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_description/multi_index_transform_helper.hpp
@@ -0,0 +1,130 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform.hpp"
+
+namespace ck {
+
+template <typename LowLength>
+__host__ __device__ constexpr auto make_pass_through_transform(const LowLength& low_length)
+{
+    return PassThrough<LowLength>{low_length};
+}
+
+template <typename LowLength, typename LeftPad, typename RightPad, bool SkipIsValidCheck = false>
+__host__ __device__ constexpr auto
+make_pad_transform(const LowLength& low_length,
+                   const LeftPad& left_pad,
+                   const RightPad& right_pad,
+                   integral_constant<bool, SkipIsValidCheck> = integral_constant<bool, false>{})
+{
+    return Pad<LowLength, LeftPad, RightPad, SkipIsValidCheck>{low_length, left_pad, right_pad};
+}
+
+template <typename LowLength, typename LeftPadLength, bool SkipIsValidCheck = false>
+__host__ __device__ constexpr auto make_left_pad_transform(
+    const LowLength& low_length,
+    const LeftPadLength& left_pad,
+    integral_constant<bool, SkipIsValidCheck> = integral_constant<bool, false>{})
+{
+    return LeftPad<LowLength, LeftPadLength, SkipIsValidCheck>{low_length, left_pad};
+}
+
+template <typename LowLength, typename RightPadLength, bool SkipIsValidCheck = false>
+__host__ __device__ constexpr auto make_right_pad_transform(
+    const LowLength& low_length,
+    const RightPadLength& right_pad,
+    integral_constant<bool, SkipIsValidCheck> = integral_constant<bool, false>{})
+{
+    return RightPad<LowLength, RightPadLength, SkipIsValidCheck>{low_length, right_pad};
+}
+
+template <typename UpLengths,
+          typename Coefficients,
+          typename enable_if<UpLengths::Size() == Coefficients::Size(), bool>::type = false>
+__host__ __device__ constexpr auto make_embed_transform(const UpLengths& up_lengths,
+                                                        const Coefficients& coefficients)
+{
+    return Embed<UpLengths, Coefficients>{up_lengths, coefficients};
+}
+
+template <typename LowLengths>
+__host__ __device__ constexpr auto make_merge_transform(const LowLengths& low_lengths)
+{
+#if CK_EXPERIMENTAL_MERGE_USE_MAGIC_DIVISION
+    return make_merge_transform_v2_magic_division(low_lengths);
+#else
+    return make_merge_transform_v1_carry_check(low_lengths);
+#endif
+}
+
+template <typename LowLengths>
+__host__ __device__ constexpr auto
+make_merge_transform_v1_carry_check(const LowLengths& low_lengths)
+{
+    return Merge_v1_carry_check<LowLengths>{low_lengths};
+}
+
+template <typename LowLengths>
+__host__ __device__ constexpr auto
+make_merge_transform_v2_magic_division(const LowLengths& low_lengths)
+{
+#if 1
+    return Merge_v2_magic_division<LowLengths>{low_lengths};
+#else
+    return Merge_v2r2_magic_division<LowLengths>{low_lengths};
+#endif
+}
+
+template <typename LowLengths>
+__host__ __device__ constexpr auto
+make_merge_transform_v3_division_mod(const LowLengths& low_lengths)
+{
+    return Merge_v3_division_mod<LowLengths>{low_lengths};
+}
+
+template <typename UpLengths, bool Use24BitIntegerCalculation = false>
+__host__ __device__ constexpr auto make_unmerge_transform(
+    const UpLengths& up_lengths,
+    integral_constant<bool, Use24BitIntegerCalculation> = integral_constant<bool, false>{})
+{
+    return UnMerge<UpLengths, Use24BitIntegerCalculation>{up_lengths};
+}
+
+template <typename LowerIndex>
+__host__ __device__ constexpr auto make_freeze_transform(const LowerIndex& low_idx)
+{
+    return Freeze<LowerIndex>{low_idx};
+}
+
+template <typename UpperIndex>
+__host__ __device__ constexpr auto make_insert_transform(const UpperIndex& up_idx)
+{
+    return Insert<UpperIndex>{up_idx};
+}
+
+template <typename LowLength, typename SliceBegin, typename SliceEnd>
+__host__ __device__ constexpr auto make_slice_transform(const LowLength& low_length,
+                                                        const SliceBegin& slice_begin,
+                                                        const SliceEnd& slice_end)
+{
+    return Slice<LowLength, SliceBegin, SliceEnd>{low_length, slice_begin, slice_end};
+}
+
+template <typename VectorSize, typename UpLength>
+__host__ __device__ constexpr auto make_vectorize_transform(const VectorSize& vector_size,
+                                                            const UpLength& up_length)
+{
+    return Vectorize<VectorSize, UpLength>{vector_size, up_length};
+}
+
+template <typename Modulus, typename UpLength>
+__host__ __device__ constexpr auto make_modulo_transform(const Modulus& modulus,
+                                                         const UpLength& up_length)
+{
+    return Modulo<Modulus, UpLength>{modulus, up_length};
+}
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_description/tensor_adaptor.hpp b/3rdparty/composable_kernel/include/ck/tensor_description/tensor_adaptor.hpp
new file mode 100644
index 0000000..d42e0a6
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_description/tensor_adaptor.hpp
@@ -0,0 +1,482 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// Transforms: Tuple<transforms...>
+// LowerDimensionHiddenIdss : Tuple<Sequence<...>, ...>
+// UpperDimensionHiddenIdss : Tuple<Sequence<...>, ...>
+// BottomDimensionHiddenIds : Sequence<...>
+// TopDimensionHiddenIds : Sequence<...>
+template <typename Transforms,
+          typename LowerDimensionHiddenIdss,
+          typename UpperDimensionHiddenIdss,
+          typename BottomDimensionHiddenIds,
+          typename TopDimensionHiddenIds>
+struct TensorAdaptor
+{
+    __host__ __device__ static constexpr index_t GetNumOfTransform() { return Transforms::Size(); }
+
+    __host__ __device__ constexpr const auto& GetTransforms() const { return transforms_; }
+
+    __host__ __device__ static constexpr auto GetLowerDimensionHiddenIdss()
+    {
+        return LowerDimensionHiddenIdss{};
+    }
+
+    __host__ __device__ static constexpr auto GetUpperDimensionHiddenIdss()
+    {
+        return UpperDimensionHiddenIdss{};
+    }
+
+    __host__ __device__ static constexpr auto GetTopDimensionHiddenIds()
+    {
+        return TopDimensionHiddenIds{};
+    }
+
+    __host__ __device__ static constexpr auto GetBottomDimensionHiddenIds()
+    {
+        return BottomDimensionHiddenIds{};
+    }
+
+    __host__ __device__ static constexpr auto InitializeElementSize(const Transforms& transforms)
+    {
+        const auto lengths = generate_tuple(
+            [&](auto idim_top) {
+                constexpr auto tmp = GetTransformAndItsUpperDimension(idim_top);
+
+                constexpr index_t itran   = tmp[Number<0>{}];
+                constexpr index_t idim_up = tmp[Number<1>{}];
+                constexpr bool found      = tmp[Number<2>{}];
+
+                static_assert(found == true,
+                              "wrong! not found matching transformation and upper-dimension");
+
+                const auto length =
+                    transforms[Number<itran>{}].GetUpperLengths()[Number<idim_up>{}];
+
+                return length;
+            },
+            Number<ndim_top_>{});
+
+        // TODO: make container_reduce support tuple of Number and index_t
+        return container_reduce(lengths, math::multiplies{}, Number<1>{});
+    }
+
+    template <index_t IDim>
+    __host__ __device__ static constexpr auto GetTransformAndItsUpperDimension(Number<IDim>)
+    {
+        constexpr auto idim_top = Number<IDim>{};
+
+        constexpr index_t idim_hidden = TopDimensionHiddenIds::At(idim_top);
+
+        index_t itran_found   = 0;
+        index_t idim_up_found = 0;
+        bool found            = false;
+
+        static_for<0, ntransform_, 1>{}([&](auto itran) {
+            constexpr auto up_dim_ids = UpperDimensionHiddenIdss{}[itran];
+
+            static_for<0, up_dim_ids.Size(), 1>{}([&](auto idim_up) {
+                if constexpr(up_dim_ids[idim_up] == idim_hidden)
+                {
+                    itran_found   = itran;
+                    idim_up_found = idim_up;
+                    found         = true;
+                }
+            });
+        });
+
+        return make_tuple(itran_found, idim_up_found, found);
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfBottomDimension()
+    {
+        return BottomDimensionHiddenIds::Size();
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfTopDimension()
+    {
+        return TopDimensionHiddenIds::Size();
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfHiddenDimension()
+    {
+        constexpr auto all_low_dim_ids = unpack(
+            [](auto&&... xs) constexpr { return merge_sequences(xs...); },
+            LowerDimensionHiddenIdss{});
+
+        constexpr auto all_up_dim_ids = unpack(
+            [](auto&&... xs) constexpr { return merge_sequences(xs...); },
+            UpperDimensionHiddenIdss{});
+
+        constexpr auto all_dim_ids = merge_sequences(all_low_dim_ids, all_up_dim_ids);
+
+        using unique_sort_all_dim_ids = typename sequence_unique_sort<decltype(all_dim_ids),
+                                                                      math::less<index_t>,
+                                                                      math::equal<index_t>>::type;
+
+        return unique_sort_all_dim_ids::Size();
+    }
+
+    constexpr static index_t ntransform_  = GetNumOfTransform();
+    constexpr static index_t ndim_hidden_ = GetNumOfHiddenDimension();
+    constexpr static index_t ndim_bottom_ = GetNumOfBottomDimension();
+    constexpr static index_t ndim_top_    = GetNumOfTopDimension();
+
+    using HiddenIndex = MultiIndex<ndim_hidden_>;
+    using BottomIndex = MultiIndex<ndim_bottom_>;
+    using TopIndex    = MultiIndex<ndim_top_>;
+
+    // may be index_t or Number<>
+    using ElementSize = remove_cv_t<decltype(InitializeElementSize(Transforms{}))>;
+
+    public:
+#if 0 // workaround compiler complaint about constexpr
+    __host__ __device__ constexpr TensorAdaptor() = default;
+#else
+    __host__ __device__ constexpr TensorAdaptor() : transforms_{}, element_size_{} {}
+#endif
+
+    __host__ __device__ constexpr TensorAdaptor(const Transforms& transforms)
+        : transforms_{transforms}, element_size_{InitializeElementSize(transforms)}
+    {
+        static_assert(Transforms::Size() == ntransform_ &&
+                          LowerDimensionHiddenIdss::Size() == ntransform_ &&
+                          UpperDimensionHiddenIdss::Size() == ntransform_,
+                      "wrong! inconsistent # of transformations");
+
+        // TODO check dependency of dimensions is valid
+    }
+
+    __host__ __device__ constexpr auto GetElementSize() const { return element_size_; }
+
+#if 0 // debug
+    template <index_t I>
+    __host__ __device__ constexpr index_t GetTopDimensionLength(Number<I> idim) const
+    {
+        // TODO: not implemented
+    }
+
+    template <index_t I>
+    __host__ __device__ constexpr index_t GetBottomDimensionLength(Number<I> idim) const
+    {
+        // TODO: not implemented
+    }
+#endif
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        static_assert(TopIdx::Size() == TopDimensionHiddenIds::Size(),
+                      "wrong! # of dimension inconsistent");
+
+        constexpr index_t ntransform  = GetNumOfTransform();
+        constexpr index_t ndim_hidden = GetNumOfHiddenDimension();
+
+        MultiIndex<ndim_hidden> idx_hidden;
+
+        // initialize uppest index
+        set_container_subset(idx_hidden, GetTopDimensionHiddenIds(), idx_top);
+
+        // calculate hidden index
+        static_for<ntransform, 0, -1>{}([&](auto itran_p1) {
+            auto itran              = itran_p1 - Number<1>{};
+            const auto& tran        = GetTransforms().At(itran);
+            constexpr auto dims_low = GetLowerDimensionHiddenIdss().At(itran);
+            constexpr auto dims_up  = GetUpperDimensionHiddenIdss().At(itran);
+
+            const auto idx_up = get_container_subset(idx_hidden, dims_up);
+
+            MultiIndex<dims_low.Size()> idx_low;
+
+            tran.CalculateLowerIndex(idx_low, idx_up);
+
+            set_container_subset(idx_hidden, dims_low, idx_low);
+        });
+
+        return get_container_subset(idx_hidden, BottomDimensionHiddenIds{});
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        bool is_known = true;
+
+        static_for<0, Transforms::Size(), 1>{}([&](auto i) {
+            is_known &= remove_cvref_t<decltype(Transforms{}[i])>::IsKnownAtCompileTime();
+        });
+
+        return is_known && is_known_at_compile_time<ElementSize>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("TensorAdaptor, ");
+        static_for<0, ntransform_, 1>{}([&](auto i) {
+            printf("transforms: ");
+            transforms_[i].Print();
+            printf("LowerDimensionHiddenIds:");
+            LowerDimensionHiddenIdss{}.At(i).Print();
+            printf("UpperDimensionHiddenIds:");
+            UpperDimensionHiddenIdss{}.At(i).Print();
+        });
+
+        printf("BottomDimensionHiddenIds:");
+        BottomDimensionHiddenIds::Print();
+        printf("TopDimensionHiddenIds:");
+        TopDimensionHiddenIds::Print();
+
+        printf("}");
+    }
+
+    private:
+    Transforms transforms_;
+    ElementSize element_size_;
+};
+
+template <typename TensorAdaptor0, typename TensorAdaptor1>
+__host__ __device__ constexpr auto chain_tensor_adaptors(const TensorAdaptor0& adaptor0,
+                                                         const TensorAdaptor1& adaptor1)
+{
+    static_assert(TensorAdaptor0::GetNumOfTopDimension() ==
+                      TensorAdaptor1::GetNumOfBottomDimension(),
+                  "wrong!");
+
+    // all_transforms = transform0 + transform1
+    const auto all_transforms =
+        container_concat(adaptor0.GetTransforms(), adaptor1.GetTransforms());
+
+    // shift
+    constexpr index_t adaptor0_max_hidden_id = [&]() {
+        index_t adaptor0_max_hidden_id_ = NumericLimits<index_t>::Min();
+
+        static_for<0, TensorAdaptor0::GetNumOfTransform(), 1>{}([&](auto itran) {
+            constexpr index_t ndim_low =
+                TensorAdaptor0{}.GetTransforms()[itran].GetNumOfLowerDimension();
+
+            static_for<0, ndim_low, 1>{}([&](auto idim_low) {
+                adaptor0_max_hidden_id_ =
+                    math::max(adaptor0_max_hidden_id_,
+                              TensorAdaptor0::GetLowerDimensionHiddenIdss()[itran][idim_low].value);
+            });
+
+            constexpr index_t ndim_up =
+                TensorAdaptor0{}.GetTransforms()[itran].GetNumOfUpperDimension();
+
+            static_for<0, ndim_up, 1>{}([&](auto idim_up) {
+                adaptor0_max_hidden_id_ =
+                    math::max(adaptor0_max_hidden_id_,
+                              TensorAdaptor0::GetUpperDimensionHiddenIdss()[itran][idim_up].value);
+            });
+        });
+
+        return adaptor0_max_hidden_id_;
+    }();
+
+    constexpr index_t adaptor1_min_hidden_id = [&]() {
+        index_t adaptor1_min_hidden_id_ = NumericLimits<index_t>::Max();
+
+        static_for<0, TensorAdaptor1::GetNumOfTransform(), 1>{}([&](auto itran) {
+            constexpr index_t ndim_low =
+                TensorAdaptor1{}.GetTransforms()[itran].GetNumOfLowerDimension();
+
+            // get the min of all lower dimenions, but not bottom dimension (because their id will
+            // be matched with top id from adaptor0)
+            static_for<0, ndim_low, 1>{}([&](auto idim_low) {
+                constexpr index_t low_dim_hidden_id =
+                    TensorAdaptor1::GetLowerDimensionHiddenIdss()[itran][idim_low].value;
+
+                bool is_bottom_dim = false;
+                static_for<0, TensorAdaptor1::GetNumOfBottomDimension(), 1>{}([&](auto i) {
+                    if constexpr(low_dim_hidden_id ==
+                                 TensorAdaptor1::GetBottomDimensionHiddenIds()[i])
+                    {
+                        is_bottom_dim = true;
+                    }
+                });
+
+                if(!is_bottom_dim)
+                {
+                    adaptor1_min_hidden_id_ = math::min(adaptor1_min_hidden_id_, low_dim_hidden_id);
+                }
+            });
+
+            constexpr index_t ndim_up =
+                TensorAdaptor1{}.GetTransforms()[itran].GetNumOfUpperDimension();
+
+            // get the min of all upper dimensions
+            static_for<0, ndim_up, 1>{}([&](auto idim_up) {
+                adaptor1_min_hidden_id_ =
+                    math::min(adaptor1_min_hidden_id_,
+                              TensorAdaptor1::GetUpperDimensionHiddenIdss()[itran][idim_up].value);
+            });
+        });
+
+        return adaptor1_min_hidden_id_;
+    }();
+
+    constexpr index_t adaptor1_hidden_id_shift =
+        adaptor0_max_hidden_id + 1 - adaptor1_min_hidden_id;
+
+    constexpr index_t ndim_bottom_1 = TensorAdaptor1::GetNumOfBottomDimension();
+
+    // all_low_dim_hidden_idss =
+    // low_dim_hidden_idss_0 + match_hidden_id_for_1(shift_hidden_id_for_1(low_dim_hiden_idss_1))
+    constexpr auto low_dim_hidden_idss_1 = generate_tuple(
+        // generate sequence of ids for a transform
+        [&](auto itran) {
+            constexpr auto ndim_low_1 = TensorAdaptor1::GetLowerDimensionHiddenIdss()[itran].Size();
+
+            constexpr auto low_dim_hidden_ids_1 =
+                TensorAdaptor1::GetLowerDimensionHiddenIdss()[itran];
+
+            // sequence in, sequence out
+            constexpr auto low_dim_hidden_ids_1_mod = [&]() constexpr
+            {
+                auto low_dim_hidden_ids_1_mod_ = to_multi_index(low_dim_hidden_ids_1);
+
+                // shift hidden id so every dim id is unique
+                static_for<0, ndim_low_1, 1>{}([&](auto idim_low_1) {
+                    low_dim_hidden_ids_1_mod_(idim_low_1) += adaptor1_hidden_id_shift;
+                });
+
+                // match hidden id
+                static_for<0, ndim_low_1, 1>{}([&](auto idim_low_1) {
+                    static_for<0, ndim_bottom_1, 1>{}([&](auto idim_bottom_1) {
+                        // if this low dim is bottom dim, then do id matching
+                        if constexpr(low_dim_hidden_ids_1[idim_low_1] ==
+                                     TensorAdaptor1::GetBottomDimensionHiddenIds()[idim_bottom_1])
+                        {
+                            low_dim_hidden_ids_1_mod_(idim_low_1) =
+                                TensorAdaptor0::GetTopDimensionHiddenIds()[idim_bottom_1];
+                        }
+                    });
+                });
+
+                return low_dim_hidden_ids_1_mod_;
+            }
+            ();
+
+            return generate_sequence_v2(
+                [&](auto i) constexpr { return Number<low_dim_hidden_ids_1_mod[i]>{}; },
+                Number<ndim_low_1>{});
+        },
+        Number<TensorAdaptor1::GetNumOfTransform()>{});
+
+    constexpr auto all_low_dim_hidden_idss =
+        container_concat(TensorAdaptor0::GetLowerDimensionHiddenIdss(), low_dim_hidden_idss_1);
+
+    // all_up_dim_hidden_idss =
+    // up_dim_hidden_idss_0 + shift_hidden_id_for_1(up_dim_hiden_idss_1)
+    constexpr auto up_dim_hidden_idss_1 = generate_tuple(
+        // generate sequence of ids for a transform
+        [&](auto itran) {
+            constexpr auto ndim_up_1 = TensorAdaptor1::GetUpperDimensionHiddenIdss()[itran].Size();
+
+            constexpr auto up_dim_hidden_ids_1 =
+                TensorAdaptor1::GetUpperDimensionHiddenIdss()[itran];
+
+            // sequence in, constexpr tuple out
+            constexpr auto up_dim_hidden_ids_1_mod = [&]() constexpr
+            {
+                auto up_dim_hidden_ids_1_mod_ = to_multi_index(up_dim_hidden_ids_1);
+
+                // shift hidden id
+                static_for<0, ndim_up_1, 1>{}([&](auto idim_up_1) {
+                    up_dim_hidden_ids_1_mod_(idim_up_1) += adaptor1_hidden_id_shift;
+                });
+
+                return up_dim_hidden_ids_1_mod_;
+            }
+            ();
+
+            // constexpr tuple to sequence
+            return generate_sequence_v2(
+                [&](auto i) constexpr { return Number<up_dim_hidden_ids_1_mod[i]>{}; },
+                Number<ndim_up_1>{});
+        },
+        Number<TensorAdaptor1::GetNumOfTransform()>{});
+
+    constexpr auto all_up_dim_hidden_idss =
+        container_concat(TensorAdaptor0::GetUpperDimensionHiddenIdss(), up_dim_hidden_idss_1);
+
+    // bottom_dim_hidden_ids = bottom_dim_hidden_ids_0
+    constexpr auto bottom_dim_hidden_ids = TensorAdaptor0::GetBottomDimensionHiddenIds();
+
+    // top_dim_hidden_ids = shift_hidden_id(top_dim_hidden_ids_1)
+    constexpr auto top_dim_hidden_ids =
+        TensorAdaptor1::GetTopDimensionHiddenIds() + Number<adaptor1_hidden_id_shift>{};
+
+    // put everything together
+    return TensorAdaptor<remove_cv_t<decltype(all_transforms)>,
+                         remove_cv_t<decltype(all_low_dim_hidden_idss)>,
+                         remove_cv_t<decltype(all_up_dim_hidden_idss)>,
+                         remove_cv_t<decltype(bottom_dim_hidden_ids)>,
+                         remove_cv_t<decltype(top_dim_hidden_ids)>>{all_transforms};
+}
+
+// Transforms: Tuple<transforms...>
+// LowerDimensionOldTopIdss: Tuple<Sequence<...>, ...>
+// UpperDimensionNewTopIdss: Tuple<Sequence<...>, ...>
+template <typename Transforms, typename LowerDimensionOldTopIdss, typename UpperDimensionNewTopIdss>
+__host__ __device__ constexpr auto make_single_stage_tensor_adaptor(const Transforms& transforms,
+                                                                    LowerDimensionOldTopIdss,
+                                                                    UpperDimensionNewTopIdss)
+{
+    constexpr index_t ntransform = Transforms::Size();
+
+    static_assert(LowerDimensionOldTopIdss::Size() == ntransform &&
+                      UpperDimensionNewTopIdss::Size() == ntransform,
+                  "wrong!");
+
+    // sanity check on LowerDimensionOldTopIdss and UpperDimensionNewTopIdss
+    constexpr auto all_low_dim_old_top_ids = unpack(
+        [](auto&&... xs) constexpr { return merge_sequences(xs...); }, LowerDimensionOldTopIdss{});
+
+    constexpr auto all_up_dim_new_top_ids = unpack(
+        [](auto&&... xs) constexpr { return merge_sequences(xs...); }, UpperDimensionNewTopIdss{});
+
+    static_assert(is_valid_sequence_map<decltype(all_low_dim_old_top_ids)>::value &&
+                      is_valid_sequence_map<decltype(all_up_dim_new_top_ids)>::value,
+                  "wrong!");
+
+    constexpr index_t ndim_old_top = all_low_dim_old_top_ids.Size();
+    constexpr index_t ndim_new_top = all_up_dim_new_top_ids.Size();
+
+    // low_dim_hidden_idss
+    constexpr auto low_dim_hidden_idss = LowerDimensionOldTopIdss{};
+
+    // up_dim_hidden_idss: shift UpperDimensionNewTopIdss by ndim_bottom
+    constexpr auto up_dim_hidden_idss = generate_tuple(
+        [](auto itran) { return UpperDimensionNewTopIdss{}[itran] + Number<ndim_old_top>{}; },
+        Number<ntransform>{});
+
+    // bottom_dim_hidden_ids
+    constexpr auto bottom_dim_hidden_ids =
+        typename arithmetic_sequence_gen<0, ndim_old_top, 1>::type{};
+
+    // top_dim_hidden_ids
+    constexpr auto top_dim_hidden_ids =
+        typename arithmetic_sequence_gen<0, ndim_new_top, 1>::type{} + Number<ndim_old_top>{};
+
+    return TensorAdaptor<remove_cv_t<Transforms>,
+                         remove_cv_t<decltype(low_dim_hidden_idss)>,
+                         remove_cv_t<decltype(up_dim_hidden_idss)>,
+                         remove_cv_t<decltype(bottom_dim_hidden_ids)>,
+                         remove_cv_t<decltype(top_dim_hidden_ids)>>{transforms};
+}
+
+template <typename X, typename... Xs, typename enable_if<sizeof...(Xs) >= 2, bool>::type = false>
+__host__ __device__ constexpr auto chain_tensor_adaptors(const X& x, const Xs&... xs)
+{
+    return chain_tensor_adaptors(x, chain_tensor_adaptors(xs...));
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_description/tensor_descriptor.hpp b/3rdparty/composable_kernel/include/ck/tensor_description/tensor_descriptor.hpp
new file mode 100644
index 0000000..f07d5b1
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_description/tensor_descriptor.hpp
@@ -0,0 +1,615 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/sequence_helper.hpp"
+#include "ck/tensor_description/multi_index_transform.hpp"
+
+namespace ck {
+
+template <index_t NDimHidden, typename VisibleDimensionIds>
+struct TensorCoordinate;
+
+template <index_t NTransform, index_t NDimVisible, typename UpdateLowerIndexHack>
+struct TensorCoordinateStep;
+
+// Transforms: Tuple<transforms...>
+// LowerDimensionIdss : Tuple<Sequence<...>, ...>
+// UpperDimensionIdss : Tuple<Sequence<...>, ...>
+// VisibleDimensionIds> : Sequence<...>
+template <typename Transforms,
+          typename LowerDimensionIdss,
+          typename UpperDimensionIdss,
+          typename VisibleDimensionIds,
+          typename ElementSpaceSize>
+struct TensorDescriptor
+{
+    // TODO make these private
+    __host__ __device__ static constexpr index_t GetNumOfTransform() { return Transforms::Size(); }
+
+    __host__ __device__ static constexpr index_t GetNumOfVisibleDimension()
+    {
+        return VisibleDimensionIds::Size();
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfHiddenDimension()
+    {
+        constexpr auto all_low_dim_ids = unpack(
+            [](auto&&... xs) constexpr { return merge_sequences(xs...); }, LowerDimensionIdss{});
+
+        constexpr auto all_up_dim_ids = unpack(
+            [](auto&&... xs) constexpr { return merge_sequences(xs...); }, UpperDimensionIdss{});
+
+        constexpr auto all_dim_ids = merge_sequences(all_low_dim_ids, all_up_dim_ids);
+
+        using unique_sort_all_dim_ids = typename sequence_unique_sort<decltype(all_dim_ids),
+                                                                      math::less<index_t>,
+                                                                      math::equal<index_t>>::type;
+
+        return unique_sort_all_dim_ids::Size();
+    }
+
+    __host__ __device__ static constexpr auto InitializeElementSize(const Transforms& transforms)
+    {
+        const auto lengths = generate_tuple(
+            [&](auto idim_visible) {
+                constexpr auto tmp = GetTransformAndItsUpperDimension(idim_visible);
+
+                constexpr index_t itran   = tmp[Number<0>{}];
+                constexpr index_t idim_up = tmp[Number<1>{}];
+                constexpr bool found      = tmp[Number<2>{}];
+
+                static_assert(found == true,
+                              "wrong! not found matching transformation and upper-dimension");
+
+                const auto length =
+                    transforms[Number<itran>{}].GetUpperLengths()[Number<idim_up>{}];
+
+                return length;
+            },
+            Number<ndim_visible_>{});
+
+        // TODO: make container_reduce support tuple of Number and index_t
+        return container_reduce(lengths, math::multiplies{}, Number<1>{});
+    }
+
+    template <index_t IDim>
+    __host__ __device__ static constexpr auto GetTransformAndItsUpperDimension(Number<IDim>)
+    {
+        constexpr auto idim_visible = Number<IDim>{};
+
+        constexpr index_t idim_hidden = VisibleDimensionIds::At(idim_visible);
+
+        index_t itran_found   = 0;
+        index_t idim_up_found = 0;
+        bool found            = false;
+
+        static_for<0, ntransform_, 1>{}([&](auto itran) {
+            constexpr auto up_dim_ids = UpperDimensionIdss{}[itran];
+
+            static_for<0, up_dim_ids.Size(), 1>{}([&](auto idim_up) {
+                if constexpr(up_dim_ids[idim_up] == idim_hidden)
+                {
+                    itran_found   = itran;
+                    idim_up_found = idim_up;
+                    found         = true;
+                }
+            });
+        });
+
+        return make_tuple(itran_found, idim_up_found, found);
+    }
+
+    constexpr static index_t ntransform_   = GetNumOfTransform();
+    constexpr static index_t ndim_visible_ = GetNumOfVisibleDimension();
+    constexpr static index_t ndim_hidden_  = GetNumOfHiddenDimension();
+
+    using VisibleIndex = MultiIndex<ndim_visible_>;
+    using HiddenIndex  = MultiIndex<ndim_hidden_>;
+    using Coordinate   = TensorCoordinate<ndim_hidden_, VisibleDimensionIds>;
+
+    // may be index_t or Number<>
+    using ElementSize = remove_cv_t<decltype(InitializeElementSize(Transforms{}))>;
+
+    public:
+#if 0 // workaround compiler complaint about constexpr
+    __host__ __device__ constexpr TensorDescriptor() = default;
+#else
+    __host__ __device__ constexpr TensorDescriptor()
+        : transforms_{}, element_size_{}, element_space_size_{}
+    {
+    }
+#endif
+
+    __host__ __device__ constexpr TensorDescriptor(const Transforms& transforms,
+                                                   ElementSpaceSize element_space_size)
+        : transforms_{transforms},
+          element_size_{InitializeElementSize(transforms)},
+          element_space_size_{element_space_size}
+
+    {
+        static_assert(Transforms::Size() == ntransform_ &&
+                          LowerDimensionIdss::Size() == ntransform_ &&
+                          UpperDimensionIdss::Size() == ntransform_,
+                      "wrong! inconsistent # of transformations");
+
+        // TODO check dependency of dimensions is valid
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfDimension()
+    {
+        return GetNumOfVisibleDimension();
+    }
+
+    template <index_t IDim>
+    __host__ __device__ constexpr auto GetLength(Number<IDim>) const
+    {
+        static_assert(IDim >= 0 && IDim < ndim_visible_, "wrong! out of range");
+
+        constexpr auto tmp = GetTransformAndItsUpperDimension(Number<IDim>{});
+
+        constexpr index_t itran   = tmp[Number<0>{}];
+        constexpr index_t idim_up = tmp[Number<1>{}];
+        constexpr bool found      = tmp[Number<2>{}];
+
+        static_assert(found == true,
+                      "wrong! not found matching transformation and upper-dimension");
+
+        return transforms_[Number<itran>{}].GetUpperLengths()[Number<idim_up>{}];
+    }
+
+    __host__ __device__ constexpr auto GetLengths() const
+    {
+        // FIXME: use Tuple of reference instead
+        return generate_sequence_v2([&](auto I) { return GetLength(I); }, Number<ndim_visible_>{});
+    }
+
+    __host__ __device__ constexpr auto GetElementSize() const { return element_size_; }
+
+    __host__ __device__ constexpr auto GetElementSpaceSize() const { return element_space_size_; }
+
+    template <typename Idx>
+    __host__ __device__ constexpr index_t CalculateOffset(const Idx& idx) const
+    {
+        static_assert(Idx::Size() == GetNumOfDimension(), "wrong! inconsistent # of dimension");
+
+        return make_tensor_coordinate(*this, idx).GetOffset();
+    }
+
+    // TODO make these private
+    __host__ __device__ constexpr const auto& GetTransforms() const { return transforms_; }
+
+    __host__ __device__ static constexpr auto GetLowerDimensionIdss()
+    {
+        return LowerDimensionIdss{};
+    }
+
+    __host__ __device__ static constexpr auto GetUpperDimensionIdss()
+    {
+        return UpperDimensionIdss{};
+    }
+
+    __host__ __device__ static constexpr auto GetVisibleDimensionIds()
+    {
+        return VisibleDimensionIds{};
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        bool is_known = true;
+
+        static_for<0, Transforms::Size(), 1>{}([&](auto i) {
+            is_known &= remove_cvref_t<decltype(Transforms{}[i])>::IsKnownAtCompileTime();
+        });
+
+        return is_known && is_known_at_compile_time<ElementSize>::value &&
+               is_known_at_compile_time<ElementSpaceSize>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("TensorDescriptor, ");
+        static_for<0, ntransform_, 1>{}([&](auto i) {
+            printf("transforms: ");
+            transforms_[i].Print();
+            printf("LowerDimensionIds:");
+            LowerDimensionIdss{}.At(i).Print();
+            printf("UpperDimensionIds:");
+            UpperDimensionIdss{}.At(i).Print();
+        });
+        printf("}");
+
+        VisibleDimensionIds::Print();
+    }
+
+    // TODO make these private
+    Transforms transforms_;
+    ElementSize element_size_;
+    ElementSpaceSize element_space_size_;
+};
+
+template <index_t NDimHidden, typename VisibleDimensionIds>
+struct TensorCoordinate
+{
+    // TODO make these private
+    static constexpr index_t ndim_visible_ = VisibleDimensionIds::Size();
+
+    using HiddenIndex  = MultiIndex<NDimHidden>;
+    using VisibleIndex = MultiIndex<ndim_visible_>;
+
+    public:
+    __host__ __device__ constexpr TensorCoordinate() = default;
+
+    __host__ __device__ constexpr TensorCoordinate(const HiddenIndex& idx_hidden)
+        : idx_hidden_{idx_hidden}
+    {
+    }
+
+    __host__ __device__ constexpr auto GetIndex() const { return GetVisibleIndex(); }
+
+    __host__ __device__ constexpr index_t GetOffset() const { return idx_hidden_[Number<0>{}]; }
+
+    // TODO make these private
+    __host__ __device__ constexpr const auto& GetHiddenIndex() const { return idx_hidden_; }
+
+    __host__ __device__ auto& GetHiddenIndex() { return idx_hidden_; }
+
+    __host__ __device__ constexpr auto GetVisibleIndex() const
+    {
+        return get_container_subset(idx_hidden_, VisibleDimensionIds{});
+    }
+
+    // TODO make these private
+    HiddenIndex idx_hidden_;
+};
+
+template <index_t NTransform, index_t NDimVisible, typename UpdateLowerIndexHack>
+struct TensorCoordinateStep
+{
+    // TODO make these private
+    using VisibleIndex = MultiIndex<NDimVisible>;
+
+    public:
+    __host__ __device__ constexpr TensorCoordinateStep() = default;
+
+    __host__ __device__ constexpr TensorCoordinateStep(const VisibleIndex& idx_diff_visible,
+                                                       const MultiIndex<NTransform>& do_transforms)
+        : idx_diff_visible_{idx_diff_visible}, do_transforms_{do_transforms}
+    {
+    }
+
+    __host__ __device__ constexpr const auto& GetIndexDiff() const { return GetVisibleIndexDiff(); }
+
+    // TODO make these private
+    __host__ __device__ constexpr const auto& GetVisibleIndexDiff() const
+    {
+        return idx_diff_visible_;
+    }
+
+    VisibleIndex idx_diff_visible_;
+    MultiIndex<NTransform> do_transforms_;
+
+    // HACK: control UpdateLowerIndex()
+    static constexpr UpdateLowerIndexHack update_lower_index_hack_;
+};
+
+// TODO: How to fix this? It uses an struct instead of lambda because lambda
+// doesn't have constructor, and to put it outside the scope where it is used
+// (transform_tensor_descriptor) because template cannot be defined inside a function
+// template
+template <typename NewTransforms>
+struct lambda_get_up_dim_num
+{
+    template <typename I>
+    __host__ __device__ constexpr auto operator()(I) const
+    {
+        using Tran = remove_reference_t<decltype(NewTransforms{}.At(I{}))>;
+        return Number<Tran::GetNumOfUpperDimension()>{};
+    }
+};
+
+template <typename OldTensorDescriptor,
+          typename NewTransforms,
+          typename NewLowerDimensionOldVisibleIdss,
+          typename NewUpperDimensionNewVisibleIdss>
+__host__ __device__ constexpr auto
+transform_tensor_descriptor(const OldTensorDescriptor& old_tensor_desc,
+                            const NewTransforms& new_transforms,
+                            NewLowerDimensionOldVisibleIdss,
+                            NewUpperDimensionNewVisibleIdss)
+{
+    // sanity check
+    {
+        static_assert(NewTransforms::Size() == NewLowerDimensionOldVisibleIdss::Size() &&
+                          NewTransforms::Size() == NewUpperDimensionNewVisibleIdss::Size(),
+                      "wrong! inconsitent number of transform");
+
+        constexpr auto all_old_top_ids = unpack([](auto... xs) { return merge_sequences(xs...); },
+                                                NewLowerDimensionOldVisibleIdss{});
+
+        constexpr auto all_new_top_ids = unpack([](auto... xs) { return merge_sequences(xs...); },
+                                                NewUpperDimensionNewVisibleIdss{});
+
+        static_assert(is_valid_sequence_map<decltype(all_old_top_ids)>::value &&
+                          is_valid_sequence_map<decltype(all_new_top_ids)>::value,
+                      "wrong!");
+    }
+
+    // lower dimension's hidden idss
+    // convert lower dimension visible idss (tuple of sequences) to hidden idss (tuple of
+    // sequences)
+    constexpr auto low_dim_hidden_idss = transform_tuples(
+        // convert lower dimension visible ids (a sequence) to hidden ids (a sequence)
+        [](auto low_dim_visible_ids) constexpr {
+            return transform_sequences(
+                // convert lower dimension visible id to hidden id
+                [](auto low_dim_visible_id) constexpr {
+                    return OldTensorDescriptor::GetVisibleDimensionIds()[low_dim_visible_id];
+                },
+                low_dim_visible_ids);
+        },
+        NewLowerDimensionOldVisibleIdss{});
+
+    constexpr index_t num_new_transform = NewTransforms::Size();
+
+    // upper dimension's hidden idss
+    constexpr index_t old_hidden_dim_number = OldTensorDescriptor::GetNumOfHiddenDimension();
+
+    constexpr auto up_dim_numbers =
+        generate_sequence(lambda_get_up_dim_num<NewTransforms>{}, Number<num_new_transform>{});
+
+    constexpr auto up_dim_numbers_scan = merge_sequences(
+        Sequence<0>{}, inclusive_scan_sequence(up_dim_numbers, math::plus<index_t>{}, Number<0>{}));
+
+    constexpr auto up_dim_hidden_idss = generate_tuple(
+        [ old_hidden_dim_number, up_dim_numbers_scan ](auto i) constexpr {
+            return
+                typename arithmetic_sequence_gen<old_hidden_dim_number + up_dim_numbers_scan[i],
+                                                 old_hidden_dim_number + up_dim_numbers_scan[i + 1],
+                                                 1>::type{};
+        },
+        Number<num_new_transform>{});
+
+    // new visible dimension's hidden ids
+    constexpr auto unordered_new_visible_dim_hidden_ids = unpack(
+        [](auto... xs) constexpr { return merge_sequences(xs...); }, up_dim_hidden_idss);
+
+    constexpr auto new_visible_dim_unordered2ordered = unpack(
+        [](auto... xs) constexpr { return merge_sequences(xs...); },
+        NewUpperDimensionNewVisibleIdss{});
+
+    constexpr auto new_visible_dim_hidden_ids =
+        unordered_new_visible_dim_hidden_ids.ReorderGivenOld2New(new_visible_dim_unordered2ordered);
+
+    // put everything together
+    const auto all_transforms = container_concat(old_tensor_desc.GetTransforms(), new_transforms);
+
+    constexpr auto all_low_dim_hidden_idss =
+        container_concat(OldTensorDescriptor::GetLowerDimensionIdss(), low_dim_hidden_idss);
+
+    constexpr auto all_up_dim_hidden_idss =
+        container_concat(OldTensorDescriptor::GetUpperDimensionIdss(), up_dim_hidden_idss);
+
+    const auto element_space_size = old_tensor_desc.GetElementSpaceSize();
+
+    return TensorDescriptor<remove_cv_t<decltype(all_transforms)>,
+                            remove_cv_t<decltype(all_low_dim_hidden_idss)>,
+                            remove_cv_t<decltype(all_up_dim_hidden_idss)>,
+                            remove_cv_t<decltype(new_visible_dim_hidden_ids)>,
+                            remove_cv_t<decltype(element_space_size)>>{all_transforms,
+                                                                       element_space_size};
+}
+
+template <typename TensorDesc, typename VisibleIndex>
+__host__ __device__ constexpr auto make_tensor_coordinate(const TensorDesc& tensor_desc,
+                                                          const VisibleIndex& idx_visible)
+{
+    static_assert(TensorDesc::GetNumOfDimension() == VisibleIndex::Size(),
+                  "wrong! # of dimension inconsistent");
+
+    constexpr index_t ntransform   = TensorDesc::GetNumOfTransform();
+    constexpr index_t ndim_hidden  = TensorDesc::GetNumOfHiddenDimension();
+    constexpr auto visible_dim_ids = TensorDesc::GetVisibleDimensionIds();
+
+    MultiIndex<ndim_hidden> idx_hidden;
+
+    // initialize visible index
+    set_container_subset(idx_hidden, visible_dim_ids, idx_visible);
+
+    // calculate hidden index
+    static_for<ntransform, 0, -1>{}([&tensor_desc, &idx_hidden](auto itran_p1) {
+        auto itran              = itran_p1 - Number<1>{};
+        const auto& tran        = tensor_desc.GetTransforms().At(itran);
+        constexpr auto dims_low = TensorDesc::GetLowerDimensionIdss().At(itran);
+        constexpr auto dims_up  = TensorDesc::GetUpperDimensionIdss().At(itran);
+
+        const auto idx_up = get_container_subset(idx_hidden, dims_up);
+
+        MultiIndex<dims_low.Size()> idx_low;
+
+        tran.CalculateLowerIndex(idx_low, idx_up);
+
+        set_container_subset(idx_hidden, dims_low, idx_low);
+    });
+
+    return TensorCoordinate<ndim_hidden, decltype(visible_dim_ids)>{idx_hidden};
+}
+
+// UpdateLowerIndexHack: Sequence<...>
+// HACK: control UpdateLowerIndex
+template <typename TensorDesc, typename VisibleIndex, typename UpdateLowerIndexHack>
+__host__ __device__ constexpr auto make_tensor_coordinate_step(const TensorDesc&,
+                                                               const VisibleIndex& idx_diff_visible,
+                                                               UpdateLowerIndexHack)
+{
+    static_assert(TensorDesc::GetNumOfDimension() == VisibleIndex::Size(),
+                  "wrong! # of dimension inconsistent");
+
+    constexpr index_t ntransform   = TensorDesc::GetNumOfTransform();
+    constexpr index_t ndim_hidden  = TensorDesc::GetNumOfHiddenDimension();
+    constexpr index_t ndim_visible = TensorDesc::GetNumOfVisibleDimension();
+    constexpr auto visible_dim_ids = TensorDesc::GetVisibleDimensionIds();
+
+    static_assert(UpdateLowerIndexHack::Size() == ntransform, "wrong!");
+
+    // use index_t for boolean type
+    auto do_transforms    = make_zero_multi_index<ntransform>();
+    auto is_non_zero_diff = make_zero_multi_index<ndim_hidden>();
+
+    // decide do_transform by checkout non-zero index diff components
+    MultiIndex<VisibleIndex::Size()> non_zero_diff_pick_visible;
+
+    static_for<0, ndim_visible, 1>{}(
+        [&](auto i) { non_zero_diff_pick_visible(i) = (idx_diff_visible[i] != 0); });
+
+    set_container_subset(is_non_zero_diff, visible_dim_ids, non_zero_diff_pick_visible);
+
+    static_for<ntransform - 1, -1, -1>{}([&](auto itran) {
+        constexpr auto dims_low = TensorDesc::GetLowerDimensionIdss().At(itran);
+        constexpr auto dims_up  = TensorDesc::GetUpperDimensionIdss().At(itran);
+
+        const auto non_zero_diff_pick_up = get_container_subset(is_non_zero_diff, dims_up);
+
+        MultiIndex<dims_low.Size()> non_zero_diff_pick_low;
+
+        // if any of upper index diff components is non-zero, then
+        //   1) Need to do this transform
+        //   2) all components of lower index diff will assume to be non-zero and need to be
+        //   computed
+        const bool idx_diff_up_has_non_zero = container_reduce(
+            non_zero_diff_pick_up, [](auto a, auto b) constexpr { return a or b; }, false);
+
+        do_transforms(itran) = idx_diff_up_has_non_zero;
+
+        static_for<0, dims_low.Size(), 1>{}(
+            [&](auto i) { non_zero_diff_pick_low(i) = idx_diff_up_has_non_zero; });
+
+        set_container_subset(is_non_zero_diff, dims_low, non_zero_diff_pick_low);
+    });
+
+    return TensorCoordinateStep<ntransform, ndim_visible, UpdateLowerIndexHack>{idx_diff_visible,
+                                                                                do_transforms};
+}
+
+template <typename TensorDesc, typename VisibleIndex>
+__host__ __device__ constexpr auto make_tensor_coordinate_step(const TensorDesc&,
+                                                               const VisibleIndex& idx_diff_visible)
+{
+    constexpr index_t ntransform = TensorDesc::GetNumOfTransform();
+
+    return make_tensor_coordinate_step(
+        TensorDesc{}, idx_diff_visible, typename uniform_sequence_gen<ntransform, 0>::type{});
+}
+
+template <typename TensorDesc, typename TensorCoord, typename TensorCoordStep>
+__host__ __device__ constexpr void move_tensor_coordinate(const TensorDesc& tensor_desc,
+                                                          TensorCoord& coord,
+                                                          const TensorCoordStep& coord_step)
+{
+    constexpr index_t ndim_hidden = TensorDesc::GetNumOfHiddenDimension();
+    constexpr index_t ntransform  = TensorDesc::GetNumOfTransform();
+
+    // this is what needs to be calculated
+    auto idx_diff_hidden = make_zero_multi_index<ndim_hidden>();
+
+    // initialize visible index diff
+    set_container_subset(
+        idx_diff_hidden, TensorDesc::GetVisibleDimensionIds(), coord_step.GetVisibleIndexDiff());
+
+    // this is what needs to be updated
+    auto& idx_hidden = coord.GetHiddenIndex();
+
+    // update visible index
+    auto idx_hidden_pick_visible =
+        get_container_subset(idx_hidden, TensorDesc::GetVisibleDimensionIds());
+
+    idx_hidden_pick_visible += coord_step.GetIndexDiff();
+
+    set_container_subset(idx_hidden, TensorDesc::GetVisibleDimensionIds(), idx_hidden_pick_visible);
+
+    // update rest of hidden index
+    static_for<ntransform - 1, -1, -1>{}([&](auto itran) {
+        if(coord_step.do_transforms_[itran])
+        {
+            const auto& tran        = tensor_desc.GetTransforms().At(itran);
+            constexpr auto dims_low = TensorDesc::GetLowerDimensionIdss().At(itran);
+            constexpr auto dims_up  = TensorDesc::GetUpperDimensionIdss().At(itran);
+
+            const auto idx_up_new  = get_container_subset(idx_hidden, dims_up);
+            auto idx_low           = get_container_subset(idx_hidden, dims_low);
+            const auto idx_diff_up = get_container_subset(idx_diff_hidden, dims_up);
+
+            MultiIndex<dims_low.Size()> idx_diff_low;
+
+            // HACK: control UpdateLowerIndex for Merge using hack
+            constexpr index_t Hack = decltype(coord_step.update_lower_index_hack_)::At(itran);
+
+            tran.UpdateLowerIndex(idx_diff_low, idx_diff_up, idx_low, idx_up_new, Number<Hack>{});
+
+            set_container_subset(idx_diff_hidden, dims_low, idx_diff_low);
+            set_container_subset(idx_hidden, dims_low, idx_low);
+        }
+    });
+}
+
+template <typename TensorDesc, typename TensorCoord>
+__host__ __device__ constexpr bool
+coordinate_has_valid_offset_assuming_visible_index_is_valid(const TensorDesc& tensor_desc,
+                                                            const TensorCoord& coord)
+{
+    bool valid = true;
+
+    constexpr index_t ntransform = TensorDesc::GetNumOfTransform();
+
+    const auto& idx_hidden = coord.GetHiddenIndex();
+
+    static_for<ntransform - 1, -1, -1>{}([&tensor_desc, &idx_hidden, &valid](auto itran) {
+        const auto tran = tensor_desc.GetTransforms().At(itran);
+
+        // check validity, only if current transformation does not always has a valid mapping
+        if constexpr(!decltype(tran)::IsValidUpperIndexAlwaysMappedToValidLowerIndex())
+        {
+            const auto idx_up =
+                get_container_subset(idx_hidden, TensorDesc::GetUpperDimensionIdss().At(itran));
+
+            // Comment: using valid = valid && .. will result in weird control flow in ISA
+            valid &= tran.IsValidUpperIndexMappedToValidLowerIndex(idx_up);
+        }
+    });
+
+    return valid;
+}
+
+template <typename TensorDesc, typename TensorCoord>
+__host__ __device__ constexpr bool coordinate_has_valid_offset(const TensorDesc& tensor_desc,
+                                                               const TensorCoord& coord)
+{
+    // check visible index
+    const auto& idx_visible = coord.GetVisibleIndex();
+
+    bool is_visible_index_valid = true;
+
+    static_for<0, TensorDesc::GetNumOfDimension(), 1>{}(
+        [&is_visible_index_valid, &idx_visible, &tensor_desc](auto i) {
+            is_visible_index_valid =
+                is_visible_index_valid &&
+                (idx_visible[i] >= 0 && idx_visible[i] < tensor_desc.GetLength(i));
+        });
+
+    // check other hidden index
+    return is_visible_index_valid &&
+           coordinate_has_valid_offset_assuming_visible_index_is_valid(tensor_desc, coord);
+}
+
+template <typename TensorDesc>
+using TensorCoordinate_t = decltype(make_tensor_coordinate(
+    TensorDesc{}, MultiIndex<remove_cvref_t<TensorDesc>::GetNumOfDimension()>{}));
+
+template <typename TensorDesc>
+using TensorCoordinateStep_t = decltype(make_tensor_coordinate_step(
+    TensorDesc{}, MultiIndex<remove_cvref_t<TensorDesc>::GetNumOfDimension()>{}));
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_description/tensor_descriptor_helper.hpp b/3rdparty/composable_kernel/include/ck/tensor_description/tensor_descriptor_helper.hpp
new file mode 100644
index 0000000..461aae7
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_description/tensor_descriptor_helper.hpp
@@ -0,0 +1,165 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+
+namespace ck {
+
+/*
+ * These functions create tensor descriptor at runtime. If they are not constexpr, you will
+ * likely see usage of scratch memory during construction of these tensor descriptors. So
+ * it's better to call these functions on host and then pass the constructed tensor descritpors
+ * to GPU. If the tensor descritpors being constructed are constexpr, then you can call these
+ * functions on GPU without worrying about scratch memory usage.
+ */
+
+#if CK_WORKAROUND_SWDEV_275126
+template <typename Lengths, typename Strides, index_t I, typename AccOld>
+__host__ __device__ constexpr auto calculate_element_space_size_impl(const Lengths& lengths,
+                                                                     const Strides& strides,
+                                                                     Number<I> i,
+                                                                     AccOld acc_old)
+{
+    auto acc_new = acc_old + (lengths[i] - Number<1>{}) * strides[i];
+
+    if constexpr(i.value < Lengths::Size() - 1)
+    {
+        return calculate_element_space_size_impl(lengths, strides, i + Number<1>{}, acc_new);
+    }
+    else
+    {
+        return acc_new;
+    }
+}
+#endif
+
+// Lengths..., Strides... could be:
+//   1) index_t, which is known at run-time, or
+//   2) Number<>, which is known at compile-time
+// element_space_size could be:
+//   1) long_index_t, or
+//   2) LongNumber<>
+template <typename... Lengths,
+          typename... Strides,
+          typename enable_if<sizeof...(Lengths) == sizeof...(Strides), bool>::type = false>
+__host__ __device__ constexpr auto make_naive_tensor_descriptor(const Tuple<Lengths...>& lengths,
+                                                                const Tuple<Strides...>& strides)
+{
+    constexpr index_t N = sizeof...(Lengths);
+
+    const auto transforms = make_tuple(make_embed_transform(lengths, strides));
+
+    constexpr auto low_dim_hidden_idss = make_tuple(Sequence<0>{});
+
+    constexpr auto up_dim_hidden_idss =
+        make_tuple(typename arithmetic_sequence_gen<1, N + 1, 1>::type{});
+
+    constexpr auto visible_dim_hidden_ids = typename arithmetic_sequence_gen<1, N + 1, 1>::type{};
+
+#if !CK_WORKAROUND_SWDEV_275126
+    // rocm-4.1 compiler would crash for recursive labmda
+    // recursive function for reduction
+    auto f = [&](auto fs, auto i, auto acc_old) {
+        auto acc_new = acc_old + (lengths[i] - Number<1>{}) * strides[i];
+
+        if constexpr(i.value < N - 1)
+        {
+            return fs(fs, i + Number<1>{}, acc_new);
+        }
+        else
+        {
+            return acc_new;
+        }
+    };
+
+    const auto element_space_size = f(f, Number<0>{}, LongNumber<1>{});
+#else
+    const auto element_space_size =
+        calculate_element_space_size_impl(lengths, strides, Number<0>{}, LongNumber<1>{});
+#endif
+
+    return TensorDescriptor<remove_cv_t<decltype(transforms)>,
+                            remove_cv_t<decltype(low_dim_hidden_idss)>,
+                            remove_cv_t<decltype(up_dim_hidden_idss)>,
+                            remove_cv_t<decltype(visible_dim_hidden_ids)>,
+                            remove_cv_t<decltype(element_space_size)>>{transforms,
+                                                                       element_space_size};
+}
+
+// Lengths... could be:
+//   1) index_t, which is known at run-time, or
+//   2) Number<>, which is known at compile-time
+// element_space_size could be:
+//   1) long_index_t, or
+//   2) LongNumber<>
+template <typename... Lengths>
+__host__ __device__ constexpr auto
+make_naive_tensor_descriptor_packed(const Tuple<Lengths...>& lengths)
+{
+    constexpr index_t N = sizeof...(Lengths);
+
+    const auto transforms = make_tuple(make_unmerge_transform(lengths));
+
+    constexpr auto low_dim_hidden_idss = make_tuple(Sequence<0>{});
+
+    constexpr auto up_dim_hidden_idss =
+        make_tuple(typename arithmetic_sequence_gen<1, N + 1, 1>::type{});
+
+    constexpr auto visible_dim_hidden_ids = typename arithmetic_sequence_gen<1, N + 1, 1>::type{};
+
+    const auto element_space_size = container_reduce(lengths, math::multiplies{}, LongNumber<1>{});
+
+    return TensorDescriptor<remove_cv_t<decltype(transforms)>,
+                            remove_cv_t<decltype(low_dim_hidden_idss)>,
+                            remove_cv_t<decltype(up_dim_hidden_idss)>,
+                            remove_cv_t<decltype(visible_dim_hidden_ids)>,
+                            remove_cv_t<decltype(element_space_size)>>{transforms,
+                                                                       element_space_size};
+}
+
+// Lengths... could be:
+//   1) index_t, which is known at run-time, or
+//   2) Number<>, which is known at compile-time
+// align could be:
+//   1) index_t, or
+//   2) Number<>
+template <typename... Lengths, typename Align>
+__host__ __device__ constexpr auto
+make_naive_tensor_descriptor_aligned(const Tuple<Lengths...>& lengths, Align align)
+{
+    constexpr auto I1 = Number<1>{};
+
+    constexpr index_t N = sizeof...(Lengths);
+
+    const auto stride_n_minus_2 = math::integer_least_multiple(lengths[Number<N - 1>{}], align);
+
+    auto strides = generate_tuple(
+        [&](auto i) {
+            if constexpr(i.value == N - 1)
+            {
+                return I1;
+            }
+            else if constexpr(i.value == N - 2)
+            {
+                return Number<stride_n_minus_2>{};
+            }
+            else
+            {
+                return container_reduce(lengths,
+                                        math::multiplies{},
+                                        Number<stride_n_minus_2>{},
+                                        i + I1,
+                                        Number<N - 1>{},
+                                        I1);
+            }
+        },
+        Number<N>{});
+
+    return make_naive_tensor_descriptor(lengths, strides);
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_description/tensor_space_filling_curve.hpp b/3rdparty/composable_kernel/include/ck/tensor_description/tensor_space_filling_curve.hpp
new file mode 100644
index 0000000..17c9100
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_description/tensor_space_filling_curve.hpp
@@ -0,0 +1,162 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/sequence_helper.hpp"
+#include "ck/utility/statically_indexed_array_multi_index.hpp"
+#include "ck/utility/tuple_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+template <typename TensorLengths,
+          typename DimAccessOrder,
+          typename ScalarsPerAccess,
+          bool SnakeCurved = true> // # of scalars per access in each dimension
+struct SpaceFillingCurve
+{
+    static constexpr index_t nDim = TensorLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    static constexpr index_t ScalarPerVector =
+        reduce_on_sequence(ScalarsPerAccess{}, math::multiplies{}, Number<1>{});
+
+    static constexpr auto access_lengths   = TensorLengths{} / ScalarsPerAccess{};
+    static constexpr auto dim_access_order = DimAccessOrder{};
+    static constexpr auto ordered_access_lengths =
+        container_reorder_given_new2old(access_lengths, dim_access_order);
+
+    static constexpr auto to_index_adaptor = make_single_stage_tensor_adaptor(
+        make_tuple(make_merge_transform(ordered_access_lengths)),
+        make_tuple(typename arithmetic_sequence_gen<0, nDim, 1>::type{}),
+        make_tuple(Sequence<0>{}));
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    __host__ __device__ static constexpr index_t GetNumOfAccess()
+    {
+        static_assert(TensorLengths::Size() == ScalarsPerAccess::Size());
+        static_assert(TensorLengths{} % ScalarsPerAccess{} ==
+                      typename uniform_sequence_gen<TensorLengths::Size(), 0>::type{});
+
+        return reduce_on_sequence(TensorLengths{}, math::multiplies{}, Number<1>{}) /
+               ScalarPerVector;
+    }
+
+    template <index_t AccessIdx1dBegin, index_t AccessIdx1dEnd>
+    static __device__ __host__ constexpr auto GetStepBetween(Number<AccessIdx1dBegin>,
+                                                             Number<AccessIdx1dEnd>)
+    {
+        static_assert(AccessIdx1dBegin >= 0, "1D index should be non-negative");
+        static_assert(AccessIdx1dBegin < GetNumOfAccess(), "1D index should be larger than 0");
+        static_assert(AccessIdx1dEnd >= 0, "1D index should be non-negative");
+        static_assert(AccessIdx1dEnd < GetNumOfAccess(), "1D index should be larger than 0");
+
+        constexpr auto idx_begin = GetIndex(Number<AccessIdx1dBegin>{});
+        constexpr auto idx_end   = GetIndex(Number<AccessIdx1dEnd>{});
+        return idx_end - idx_begin;
+    }
+
+    template <index_t AccessIdx1d>
+    static __device__ __host__ constexpr auto GetForwardStep(Number<AccessIdx1d>)
+    {
+        static_assert(AccessIdx1d < GetNumOfAccess(), "1D index should be larger than 0");
+        return GetStepBetween(Number<AccessIdx1d>{}, Number<AccessIdx1d + 1>{});
+    }
+
+    template <index_t AccessIdx1d>
+    static __device__ __host__ constexpr auto GetBackwardStep(Number<AccessIdx1d>)
+    {
+        static_assert(AccessIdx1d > 0, "1D index should be larger than 0");
+
+        return GetStepBetween(Number<AccessIdx1d>{}, Number<AccessIdx1d - 1>{});
+    }
+
+    template <index_t AccessIdx1d>
+    static __device__ __host__ constexpr Index GetIndex(Number<AccessIdx1d>)
+    {
+#if 0
+        /*
+         * \todo: TensorAdaptor::CalculateBottomIndex does NOT return constexpr as expected.
+         */
+        constexpr auto ordered_access_idx = to_index_adaptor.CalculateBottomIndex(make_multi_index(Number<AccessIdx1d>{}));
+#else
+
+        constexpr auto access_strides = container_reverse_exclusive_scan(
+            ordered_access_lengths, math::multiplies{}, Number<1>{});
+
+        constexpr auto idx_1d = Number<AccessIdx1d>{};
+        // Given tensor strides \p access_lengths, and 1D index of space-filling-curve, compute the
+        // idim-th element of multidimensional index.
+        // All constexpr variables have to be captured by VALUE.
+        constexpr auto compute_index = [ idx_1d, access_strides ](auto idim) constexpr
+        {
+            constexpr auto compute_index_impl = [ idx_1d, access_strides ](auto jdim) constexpr
+            {
+                auto res = idx_1d.value;
+                auto id  = 0;
+
+                static_for<0, jdim.value + 1, 1>{}([&](auto kdim) {
+                    id = res / access_strides[kdim].value;
+                    res -= id * access_strides[kdim].value;
+                });
+
+                return id;
+            };
+
+            constexpr auto id = compute_index_impl(idim);
+            return Number<id>{};
+        };
+
+        constexpr auto ordered_access_idx = generate_tuple(compute_index, Number<nDim>{});
+#endif
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto idim) {
+                index_t tmp = ordered_access_idx[I0];
+
+                static_for<1, idim, 1>{}(
+                    [&](auto j) { tmp = tmp * ordered_access_lengths[j] + ordered_access_idx[j]; });
+
+                forward_sweep_(idim) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate multi-dim tensor index
+        auto idx_md = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto idim) {
+                ordered_idx(idim) =
+                    !SnakeCurved || forward_sweep[idim]
+                        ? ordered_access_idx[idim]
+                        : ordered_access_lengths[idim] - 1 - ordered_access_idx[idim];
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dim_access_order) *
+                   ScalarsPerAccess{};
+        }();
+        return idx_md;
+    }
+
+    // FIXME: rename this function
+    template <index_t AccessIdx1d>
+    static __device__ __host__ constexpr auto GetIndexTupleOfNumber(Number<AccessIdx1d>)
+    {
+        constexpr auto idx = GetIndex(Number<AccessIdx1d>{});
+
+        return generate_tuple([&](auto i) { return Number<idx[i]>{}; }, Number<nDim>{});
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp
new file mode 100644
index 0000000..8b1b7be
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp
@@ -0,0 +1,412 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp"
+
+namespace ck {
+
+// C[BM0, BM1, BN0, BN1] += transpose(A[K, BM0, BM1]) * B[K, BN0, BN1]
+// A and B are visable to the whole block, C is distributed among each thread
+// Assume:
+//   1. A:
+//     1. ABlockDesc_BK0_BM_BK1 is known at compile-time
+//     2. ABlockBuffer is DynamicBuffer
+//   2. B:
+//     1. BBlockDesc_BK0_BN_BK1 is known at compile-time
+//     2. BBlockBuffer is DynamicBuffer
+//   3. C:
+//     1. CThreadDesc_BM0_BM11_BN0_BN11 is known at compile-time
+//     2. CThreadBuffer is StaticBuffer
+// Also assume:
+//   BM10BN10ThreadClusterBM10Xs::Size() = BM10BN10ThreadClusterBN10Xs::Size() == 2
+//   BM0 = BN0 = 2. It will do 2x2 pipelined read and fma (ABBA optimization)
+template <index_t BlockSize,
+          typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename ABlockDesc_BK0_BM_BK1,
+          typename BBlockDesc_BK0_BN_BK1,
+          index_t BM1PerThreadBM11,
+          index_t BN1PerThreadBN11,
+          index_t BK0PerThread,
+          typename BM10BN10ThreadClusterBM10Xs, // Sequence<BM10BN10ThreadClusterBM100,
+                                                //          BM10BN10ThreadClusterBM101, ...>
+          typename BM10BN10ThreadClusterBN10Xs, // Sequence<BM10BN10ThreadClusterBN100,
+                                                //          BM10BN10ThreadClusterBN101, ...>
+          index_t AThreadCopyScalarPerVector_BM11,
+          index_t BThreadCopyScalarPerVector_BN11,
+          typename enable_if<ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
+                                 BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2
+{
+    using AIndex = MultiIndex<3>;
+    using BIndex = MultiIndex<3>;
+    using CIndex = MultiIndex<4>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr index_t BK0 = ABlockDesc_BK0_BM_BK1{}.GetLength(I0);
+    static constexpr index_t BK1 = ABlockDesc_BK0_BM_BK1{}.GetLength(I2);
+    static constexpr index_t BM  = ABlockDesc_BK0_BM_BK1{}.GetLength(I1);
+    static constexpr index_t BN  = BBlockDesc_BK0_BN_BK1{}.GetLength(I1);
+
+    static constexpr index_t BM100 = BM10BN10ThreadClusterBM10Xs{}[I0];
+    static constexpr index_t BN100 = BM10BN10ThreadClusterBN10Xs{}[I0];
+
+    static constexpr index_t BM101 = BM10BN10ThreadClusterBM10Xs{}[I1];
+    static constexpr index_t BN101 = BM10BN10ThreadClusterBN10Xs{}[I1];
+
+    static constexpr index_t BM11 = BM1PerThreadBM11;
+    static constexpr index_t BN11 = BN1PerThreadBN11;
+
+    static constexpr index_t BM1 = BM100 * BM101 * BM11;
+    static constexpr index_t BN1 = BN100 * BN101 * BN11;
+
+    static constexpr index_t BM0 = BM / BM1;
+    static constexpr index_t BN0 = BN / BN1;
+
+    __host__ __device__ static constexpr auto
+    MakeABlockDescriptor_BK0_BM0_BM1_BK1(const ABlockDesc_BK0_BM_BK1& a_block_desc_bk0_bm_bk1)
+    {
+        const auto a_block_bk0_bm0_bm1_bk1 = transform_tensor_descriptor(
+            a_block_desc_bk0_bm_bk1,
+            make_tuple(make_pass_through_transform(Number<BK0>{}),
+                       make_unmerge_transform(make_tuple(Number<BM0>{}, Number<BM1>{})),
+                       make_pass_through_transform(Number<BK1>{})),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return a_block_bk0_bm0_bm1_bk1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBBlockDescriptor_BK0_BN0_BN1_BK1(const BBlockDesc_BK0_BN_BK1& b_block_desc_bk0_bn_bk1)
+    {
+        const auto b_block_desc_bk0_bn0_bn1_bk1 = transform_tensor_descriptor(
+            b_block_desc_bk0_bn_bk1,
+            make_tuple(make_pass_through_transform(Number<BK0>{}),
+                       make_unmerge_transform(make_tuple(Number<BN0>{}, Number<BN1>{})),
+                       make_pass_through_transform(Number<BK1>{})),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return b_block_desc_bk0_bn0_bn1_bk1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM_BN()
+    {
+        // upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
+        // lower: [BM, BN]
+        constexpr auto c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_unmerge_transform(make_tuple(
+                               Number<BM0>{}, Number<BM100>{}, Number<BM101>{}, Number<BM11>{})),
+                           make_unmerge_transform(make_tuple(
+                               Number<BN0>{}, Number<BN100>{}, Number<BN101>{}, Number<BN11>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4, 5, 6, 7>{}));
+
+        return c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM0_BM1_BN0_BN1()
+    {
+        // upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
+        // lower: [BM0, BM1, BN0, BN1]
+        constexpr auto c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1 =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_pass_through_transform(Number<BM0>{}),
+                           make_unmerge_transform(
+                               make_tuple(Number<BM100>{}, Number<BM101>{}, Number<BM11>{})),
+                           make_pass_through_transform(Number<BN0>{}),
+                           make_unmerge_transform(
+                               make_tuple(Number<BN100>{}, Number<BN101>{}, Number<BN11>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}, Sequence<5, 6, 7>{}));
+
+        return c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1;
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadTensorLengths_BM0_BM1_BN0_BN1()
+    {
+        return Sequence<BM0, BM11, BN0, BN11>{};
+    }
+
+    static constexpr auto a_block_desc_bk0_bm0_bm1_bk1_ =
+        MakeABlockDescriptor_BK0_BM0_BM1_BK1(ABlockDesc_BK0_BM_BK1{});
+
+    static constexpr auto b_block_desc_bk0_bn0_bn1_bk1_ =
+        MakeBBlockDescriptor_BK0_BN0_BN1_BK1(BBlockDesc_BK0_BN_BK1{});
+
+    public:
+    __device__ BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2()
+        : c_thread_origin_data_idx_{CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
+              get_thread_local_1d_id())},
+          a_thread_copy_{
+              make_tuple(0, c_thread_origin_data_idx_[I0], c_thread_origin_data_idx_[I1], 0)},
+          b_thread_copy_{
+              make_tuple(0, c_thread_origin_data_idx_[I2], c_thread_origin_data_idx_[I3], 0)}
+    {
+        static_assert(ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
+                          BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(BlockSize == BM101 * BM100 * BN101 * BN100,
+                      "wrong! blocksize and cluster size not consistent");
+
+        static_assert(BM % BM1 == 0 && BN % BN1 == 0, "wrong!");
+
+        static_assert(ABlockDesc_BK0_BM_BK1{}.GetLength(I0) ==
+                          BBlockDesc_BK0_BN_BK1{}.GetLength(I0),
+                      "wrong! K dimension not consistent");
+
+        // TODO remove this restriction
+        static_assert(BM10BN10ThreadClusterBM10Xs::Size() == 2 &&
+                          BM10BN10ThreadClusterBN10Xs::Size() == 2,
+                      "wrong!");
+
+        // TODO: remove this restriction
+        static_assert(BM0 == 2, "wrong");
+        static_assert(BM0 == 2 && BN0 == 2, "wrong");
+    }
+
+    __device__ static CIndex CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(index_t thread_id)
+    {
+        // lower: [BM0, BM1, BN0, BN1]
+        // upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
+        constexpr auto adaptor0 =
+            MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM0_BM1_BN0_BN1();
+
+        // lower: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
+        // upper: [Tid, BM0, BM11, BN0, BN11]
+        constexpr auto adaptor1 = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(BM100, BN100, BM101, BN101)),
+                       make_pass_through_transform(BM0),
+                       make_pass_through_transform(BM11),
+                       make_pass_through_transform(BN0),
+                       make_pass_through_transform(BN11)),
+            make_tuple(
+                Sequence<1, 5, 2, 6>{}, Sequence<0>{}, Sequence<3>{}, Sequence<4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+        constexpr auto adaptor = chain_tensor_adaptors(adaptor0, adaptor1);
+
+        return adaptor.CalculateBottomIndex(make_multi_index(thread_id, 0, 0, 0, 0));
+    }
+
+    template <typename CThreadDesc_BM0_BM11_BN0_BN11,
+              typename ABlockBuffer,
+              typename BBlockBuffer,
+              typename CThreadBuffer>
+    __device__ void Run(const CThreadDesc_BM0_BM11_BN0_BN11&,
+                        const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        static_assert(CThreadDesc_BM0_BM11_BN0_BN11::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        // TODO: remove this restriction
+        static_assert(BM0 == 2 && BN0 == 2 &&
+                          CThreadDesc_BM0_BM11_BN0_BN11{}.GetLength(I0) == BM0 &&
+                          CThreadDesc_BM0_BM11_BN0_BN11{}.GetLength(I2) == BN0,
+                      "wrong");
+
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
+            a_thread_desc_bk0_bm0_bm1_bk1_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
+            b_thread_desc_bk0_bn0_bn1_bk1_.GetElementSpaceSize());
+
+        constexpr auto threadwise_contraction =
+            ThreadwiseContractionDl_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1<
+                FloatA,
+                FloatB,
+                FloatC,
+                decltype(a_thread_desc_bk0_bm0_bm1_bk1_),
+                decltype(b_thread_desc_bk0_bn0_bn1_bk1_),
+                CThreadDesc_BM0_BM11_BN0_BN11,
+                Sequence<BK0PerThread, BK1>,
+                Sequence<1, BM1PerThreadBM11>,
+                Sequence<1, BN1PerThreadBN11>>{};
+
+        // read A_sub_0
+        a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
+                           make_tuple(I0, I0, I0, I0),
+                           a_block_buf,
+                           a_thread_desc_bk0_bm0_bm1_bk1_,
+                           make_tuple(I0, I0, I0, I0),
+                           a_thread_buf);
+
+        // read B_sub_0
+        b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
+                           make_tuple(I0, I0, I0, I0),
+                           b_block_buf,
+                           b_thread_desc_bk0_bn0_bn1_bk1_,
+                           make_tuple(I0, I0, I0, I0),
+                           b_thread_buf);
+
+        // read B_sub_1
+        b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
+                           make_tuple(I0, I1, I0, I0),
+                           b_block_buf,
+                           b_thread_desc_bk0_bn0_bn1_bk1_,
+                           make_tuple(I0, I1, I0, I0),
+                           b_thread_buf);
+
+        // read A_sub_1
+        a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
+                           make_tuple(I0, I1, I0, I0),
+                           a_block_buf,
+                           a_thread_desc_bk0_bm0_bm1_bk1_,
+                           make_tuple(I0, I1, I0, I0),
+                           a_thread_buf);
+
+        // C_sub_00 += transpose(A_sub_0) * B_sub_0
+        threadwise_contraction.Run(a_thread_buf,
+                                   make_tuple(I0, I0, I0, I0),
+                                   b_thread_buf,
+                                   make_tuple(I0, I0, I0, I0),
+                                   c_thread_buf,
+                                   make_tuple(I0, I0, I0, I0));
+
+        // C_sub_01 += transpose(A_sub_0) * B_sub_1
+        threadwise_contraction.Run(a_thread_buf,
+                                   make_tuple(I0, I0, I0, I0),
+                                   b_thread_buf,
+                                   make_tuple(I0, I1, I0, I0),
+                                   c_thread_buf,
+                                   make_tuple(I0, I0, I1, I0));
+
+        // loop over rest of bk0
+        static_for<BK0PerThread, BK0, BK0PerThread>{}([&](auto bk0) {
+            // read A_sub_0
+            a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
+                               make_tuple(bk0, I0, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_bk0_bm0_bm1_bk1_,
+                               make_tuple(I0, I0, I0, I0),
+                               a_thread_buf);
+
+            // C_sub_10 += transpose(A_sub_1) * B_sub_0
+            threadwise_contraction.Run(a_thread_buf,
+                                       make_tuple(I0, I1, I0, I0),
+                                       b_thread_buf,
+                                       make_tuple(I0, I0, I0, I0),
+                                       c_thread_buf,
+                                       make_tuple(I1, I0, I0, I0));
+
+            // read B_sub_0
+            b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
+                               make_tuple(bk0, I0, I0, I0),
+                               b_block_buf,
+                               b_thread_desc_bk0_bn0_bn1_bk1_,
+                               make_tuple(I0, I0, I0, I0),
+                               b_thread_buf);
+
+            // C_sub_11 += transpose(A_sub_1) * B_sub_1
+            threadwise_contraction.Run(a_thread_buf,
+                                       make_tuple(I0, I1, I0, I0),
+                                       b_thread_buf,
+                                       make_tuple(I0, I1, I0, I0),
+                                       c_thread_buf,
+                                       make_tuple(I1, I0, I1, I0));
+
+            // read B_sub_1
+            b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
+                               make_tuple(bk0, I1, I0, I0),
+                               b_block_buf,
+                               b_thread_desc_bk0_bn0_bn1_bk1_,
+                               make_tuple(I0, I1, I0, I0),
+                               b_thread_buf);
+
+            // read A_sub_1
+            a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
+                               make_tuple(bk0, I1, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_bk0_bm0_bm1_bk1_,
+                               make_tuple(I0, I1, I0, I0),
+                               a_thread_buf);
+
+            // C_sub_00 += transpose(A_sub_0) * B_sub_0
+            threadwise_contraction.Run(a_thread_buf,
+                                       make_tuple(I0, I0, I0, I0),
+                                       b_thread_buf,
+                                       make_tuple(I0, I0, I0, I0),
+                                       c_thread_buf,
+                                       make_tuple(I0, I0, I0, I0));
+
+            // C_sub_01 += transpose(A_sub_0) * B_sub_1
+            threadwise_contraction.Run(a_thread_buf,
+                                       make_tuple(I0, I0, I0, I0),
+                                       b_thread_buf,
+                                       make_tuple(I0, I1, I0, I0),
+                                       c_thread_buf,
+                                       make_tuple(I0, I0, I1, I0));
+        });
+
+        // C_sub_10 += transpose(A_sub_1) * B_sub_0
+        threadwise_contraction.Run(a_thread_buf,
+                                   make_tuple(I0, I1, I0, I0),
+                                   b_thread_buf,
+                                   make_tuple(I0, I0, I0, I0),
+                                   c_thread_buf,
+                                   make_tuple(I1, I0, I0, I0));
+
+        // C_sub_11 += transpose(A_sub_1) * B_sub_1
+        threadwise_contraction.Run(a_thread_buf,
+                                   make_tuple(I0, I1, I0, I0),
+                                   b_thread_buf,
+                                   make_tuple(I0, I1, I0, I0),
+                                   c_thread_buf,
+                                   make_tuple(I1, I0, I1, I0));
+    }
+
+    private:
+    // A[BK0, BM0, BM1, BK1]
+    static constexpr auto a_thread_desc_bk0_bm0_bm1_bk1_ =
+        make_naive_tensor_descriptor_packed(make_tuple(
+            Number<BK0PerThread>{}, Number<BM0>{}, Number<BM1PerThreadBM11>{}, Number<BK1>{}));
+
+    // B[BK0, BN0, BN1, BK1]
+    static constexpr auto b_thread_desc_bk0_bn0_bn1_bk1_ =
+        make_naive_tensor_descriptor_packed(make_tuple(
+            Number<BK0PerThread>{}, Number<BN0>{}, Number<BN1PerThreadBN11>{}, Number<BK1>{}));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
+        FloatA,
+        FloatA,
+        decltype(a_block_desc_bk0_bm0_bm1_bk1_),
+        decltype(a_thread_desc_bk0_bm0_bm1_bk1_),
+        Sequence<BK0PerThread, 1, BM1PerThreadBM11, BK1>, // SliceLengths
+        Sequence<0, 1, 2, 3>,                             // DimAccessOrder
+        Sequence<1, 1, BM1PerThreadBM11, BK1>,            // SrcVectorTensorLengths
+        Sequence<0, 1, 2, 3>>;                            // SrcVectorTensorContiguousDimOrder
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
+        FloatB,
+        FloatB,
+        decltype(b_block_desc_bk0_bn0_bn1_bk1_),
+        decltype(b_thread_desc_bk0_bn0_bn1_bk1_),
+        Sequence<BK0PerThread, 1, BN1PerThreadBN11, BK1>, // SliceLengths
+        Sequence<0, 1, 2, 3>,                             // DimAccessOrder
+        Sequence<1, 1, BN1PerThreadBN11, BK1>,            // SrcVectorTensorLengths
+        Sequence<0, 1, 2, 3>>;                            // SrcVectorTensorContiguousDimOrder
+
+    CIndex c_thread_origin_data_idx_;
+
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp
new file mode 100644
index 0000000..33120bd
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp
@@ -0,0 +1,397 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_BLOCKWISE_GEMM_DLOPS_V2R2_HPP
+#define CK_BLOCKWISE_GEMM_DLOPS_V2R2_HPP
+
+#include "common_header.hpp"
+#include "tensor_adaptor.hpp"
+#include "threadwise_tensor_slice_transfer.hpp"
+#include "threadwise_contraction_dlops.hpp"
+
+namespace ck {
+
+// C[M0, M1, N0, N1] += transpose(A[K, M0, M1]) * B[K, N0, N1]
+// A and B are visable to the whole block, C is distributed among each thread
+// Assume:
+//   1. A:
+//     1. AKMBlockDesc is known at compile-time
+//     2. ABlockBuffer is DynamicBuffer
+//   2. B:
+//     1. BKNBlockDesc is known at compile-time
+//     2. BBlockBuffer is DynamicBuffer
+//   3. C:
+//     1. CM0M1N0N1ThreadDesc is known at compile-time
+//     2. CThreadBuffer is StaticBuffer
+// Also assume:
+//   M0 = N0 = 2. It will do 2x2 pipelined read and fma (ABBA optimization)
+template <
+    index_t BlockSize,
+    typename FloatA,
+    typename FloatB,
+    typename FloatC,
+    typename AKMBlockDesc,
+    typename BKNBlockDesc,
+    index_t M1PerThreadM11,
+    index_t N1PerThreadN11,
+    index_t KPerThread,
+    index_t M1N1ThreadClusterM100,
+    index_t M1N1ThreadClusterN100,
+    index_t M1N1ThreadClusterM101,
+    index_t M1N1ThreadClusterN101,
+    index_t AThreadCopyScalarPerVector_M11,
+    index_t BThreadCopyScalarPerVector_N11,
+    typename enable_if<AKMBlockDesc::IsKnownAtCompileTime() && BKNBlockDesc::IsKnownAtCompileTime(),
+                       bool>::type = false>
+struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v2r2_pipeline_2x2
+{
+    using AIndex = MultiIndex<3>;
+    using BIndex = MultiIndex<3>;
+    using CIndex = MultiIndex<4>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr index_t K = AKMBlockDesc{}.GetLength(I0);
+    static constexpr index_t M = AKMBlockDesc{}.GetLength(I1);
+    static constexpr index_t N = BKNBlockDesc{}.GetLength(I1);
+
+    static constexpr index_t M100 = M1N1ThreadClusterM100;
+    static constexpr index_t N100 = M1N1ThreadClusterN100;
+
+    static constexpr index_t M101 = M1N1ThreadClusterM101;
+    static constexpr index_t N101 = M1N1ThreadClusterN101;
+
+    static constexpr index_t M11 = M1PerThreadM11;
+    static constexpr index_t N11 = N1PerThreadN11;
+
+    static constexpr index_t M1 = M1N1ThreadClusterM100 * M1N1ThreadClusterM101 * M1PerThreadM11;
+    static constexpr index_t N1 = M1N1ThreadClusterN100 * M1N1ThreadClusterN101 * N1PerThreadN11;
+
+    static constexpr index_t M0 = M / M1;
+    static constexpr index_t N0 = N / N1;
+
+    __host__ __device__ static constexpr auto
+    MakeAKM0M1BlockDescriptor(const AKMBlockDesc& /* a_k_m_block_desc */)
+    {
+        const auto a_k_m0_m1_block_desc = transform_tensor_descriptor(
+            AKMBlockDesc{},
+            make_tuple(make_pass_through_transform(Number<K>{}),
+                       make_unmerge_transform(make_tuple(Number<M0>{}, Number<M1>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}));
+
+        return a_k_m0_m1_block_desc;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBKN0N1BlockDescriptor(const BKNBlockDesc& /* b_k_n_block_desc */)
+    {
+        const auto b_k_n0_n1_block_desc = transform_tensor_descriptor(
+            BKNBlockDesc{},
+            make_tuple(make_pass_through_transform(Number<K>{}),
+                       make_unmerge_transform(make_tuple(Number<N0>{}, Number<N1>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}));
+
+        return b_k_n0_n1_block_desc;
+    }
+
+    __host__ __device__ static constexpr auto MakeCM0M100M101M11N0N100N101N11ToMNBlockAdaptor()
+    {
+        // upper: [M0, M100, M101, M11, N0, N100, N101, N11]
+        // lower: [M, N]
+        constexpr auto c_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n_block_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_unmerge_transform(make_tuple(
+                               Number<M0>{}, Number<M100>{}, Number<M101>{}, Number<M11>{})),
+                           make_unmerge_transform(make_tuple(
+                               Number<N0>{}, Number<N100>{}, Number<N101>{}, Number<N11>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4, 5, 6, 7>{}));
+
+        return c_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n_block_adaptor;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCM0M100M101M11N0N100N101N11ToM0M1N0N1BlockAdaptor()
+    {
+        // upper: [M0, M100, M101, M11, N0, N100, N101, N11]
+        // lower: [M0, M1, N0, N1]
+        constexpr auto c_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1_block_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_pass_through_transform(Number<M0>{}),
+                           make_unmerge_transform(
+                               make_tuple(Number<M100>{}, Number<M101>{}, Number<M11>{})),
+                           make_pass_through_transform(Number<N0>{}),
+                           make_unmerge_transform(
+                               make_tuple(Number<N100>{}, Number<N101>{}, Number<N11>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}, Sequence<5, 6, 7>{}));
+
+        return c_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1_block_adaptor;
+    }
+
+    __host__ __device__ static constexpr auto GetCM0M1N0N1ThreadTensorLengths()
+    {
+        return Sequence<M0, M11, N0, N11>{};
+    }
+
+    static constexpr auto a_k_m0_m1_block_desc_ = MakeAKM0M1BlockDescriptor(AKMBlockDesc{});
+    static constexpr auto b_k_n0_n1_block_desc_ = MakeBKN0N1BlockDescriptor(BKNBlockDesc{});
+
+    public:
+    __device__ BlockwiseGemmDlops_km_kn_m0m1n0n1_v2r2_pipeline_2x2()
+        : c_thread_origin_data_idx_{CalculateCM0M1N0N1ThreadOriginOnBlock(
+              get_thread_local_1d_id())},
+          a_thread_copy_{
+              make_tuple(0, c_thread_origin_data_idx_[I0], c_thread_origin_data_idx_[I1])},
+          b_thread_copy_{
+              make_tuple(0, c_thread_origin_data_idx_[I2], c_thread_origin_data_idx_[I3])}
+    {
+        static_assert(AKMBlockDesc::IsKnownAtCompileTime() && BKNBlockDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(BlockSize == M101 * M100 * N101 * N100,
+                      "wrong! blocksize and cluster size not consistent");
+
+        static_assert(M % M1 == 0 && N % N1 == 0, "wrong!");
+
+        static_assert(AKMBlockDesc{}.GetLength(I0) == BKNBlockDesc{}.GetLength(I0),
+                      "wrong! K dimension not consistent");
+
+        // TODO: remove this restriction
+        static_assert(M0 == 2 && N0 == 2, "wrong");
+    }
+
+    __device__ static CIndex CalculateCM0M1N0N1ThreadOriginOnBlock(index_t thread_id)
+    {
+        // lower: [M0, M1, N0, N1]
+        // upper: [M0, M100, M101, M11, N0, N100, N101, N11]
+        constexpr auto adaptor0 = MakeCM0M100M101M11N0N100N101N11ToM0M1N0N1BlockAdaptor();
+
+        // lower: [M0, M100, M101, M11, N0, N100, N101, N11]
+        // upper: [Tid, M0, M11, N0, N11]
+        constexpr auto adaptor1 = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(M100, N100, M101, N101)),
+                       make_pass_through_transform(M0),
+                       make_pass_through_transform(M11),
+                       make_pass_through_transform(N0),
+                       make_pass_through_transform(N11)),
+            make_tuple(
+                Sequence<1, 5, 2, 6>{}, Sequence<0>{}, Sequence<3>{}, Sequence<4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+        constexpr auto adaptor = chain_tensor_adaptors(adaptor0, adaptor1);
+
+        return adaptor.CalculateBottomIndex(make_multi_index(thread_id, 0, 0, 0, 0));
+    }
+
+    __host__ __device__ static constexpr index_t GetABlockAlignment() { return M1PerThreadM11; }
+
+    __host__ __device__ static constexpr auto GetBBlockAlignment() { return N1PerThreadN11; }
+
+    template <typename CM0M1N0N1ThreadDesc,
+              typename ABlockBuffer,
+              typename BBlockBuffer,
+              typename CThreadBuffer>
+    __device__ void Run(const CM0M1N0N1ThreadDesc& /* c_m0_m1_n0_n1_thread_desc */,
+                        const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        static_assert(CM0M1N0N1ThreadDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        // TODO: remove this restriction
+        static_assert(M0 == 2 && N0 == 2 && CM0M1N0N1ThreadDesc{}.GetLength(I0) == M0 &&
+                          CM0M1N0N1ThreadDesc{}.GetLength(I2) == N0,
+                      "wrong");
+
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
+            a_k_m0_m1_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
+            b_k_n0_n1_thread_desc_.GetElementSpaceSize());
+
+        constexpr auto threadwise_gemm =
+            ThreadwiseGemmDlops_km0m1_kn0n1_m0m1n0n1<FloatA,
+                                                     FloatB,
+                                                     FloatC,
+                                                     decltype(a_k_m0_m1_thread_desc_),
+                                                     decltype(b_k_n0_n1_thread_desc_),
+                                                     CM0M1N0N1ThreadDesc,
+                                                     Sequence<KPerThread>,
+                                                     Sequence<1, M1PerThreadM11>,
+                                                     Sequence<1, N1PerThreadN11>>{};
+
+        // read A_sub_0
+        a_thread_copy_.Run(a_k_m0_m1_block_desc_,
+                           make_tuple(I0, I0, I0),
+                           a_block_buf,
+                           a_k_m0_m1_thread_desc_,
+                           make_tuple(I0, I0, I0),
+                           a_thread_buf);
+
+        // read B_sub_0
+        b_thread_copy_.Run(b_k_n0_n1_block_desc_,
+                           make_tuple(I0, I0, I0),
+                           b_block_buf,
+                           b_k_n0_n1_thread_desc_,
+                           make_tuple(I0, I0, I0),
+                           b_thread_buf);
+
+        // read B_sub_1
+        b_thread_copy_.Run(b_k_n0_n1_block_desc_,
+                           make_tuple(I0, I1, I0),
+                           b_block_buf,
+                           b_k_n0_n1_thread_desc_,
+                           make_tuple(I0, I1, I0),
+                           b_thread_buf);
+
+        // read A_sub_1
+        a_thread_copy_.Run(a_k_m0_m1_block_desc_,
+                           make_tuple(I0, I1, I0),
+                           a_block_buf,
+                           a_k_m0_m1_thread_desc_,
+                           make_tuple(I0, I1, I0),
+                           a_thread_buf);
+
+        // C_sub_00 += transpose(A_sub_0) * B_sub_0
+        threadwise_gemm.Run(a_thread_buf,
+                            make_tuple(I0, I0, I0),
+                            b_thread_buf,
+                            make_tuple(I0, I0, I0),
+                            c_thread_buf,
+                            make_tuple(I0, I0, I0, I0));
+
+        // C_sub_01 += transpose(A_sub_0) * B_sub_1
+        threadwise_gemm.Run(a_thread_buf,
+                            make_tuple(I0, I0, I0),
+                            b_thread_buf,
+                            make_tuple(I0, I1, I0),
+                            c_thread_buf,
+                            make_tuple(I0, I0, I1, I0));
+
+        // loop over rest of k
+        static_for<KPerThread, K, KPerThread>{}([&](auto k) {
+            // read A_sub_0
+            a_thread_copy_.Run(a_k_m0_m1_block_desc_,
+                               make_tuple(k, I0, I0),
+                               a_block_buf,
+                               a_k_m0_m1_thread_desc_,
+                               make_tuple(I0, I0, I0),
+                               a_thread_buf);
+
+            // C_sub_10 += transpose(A_sub_1) * B_sub_0
+            threadwise_gemm.Run(a_thread_buf,
+                                make_tuple(I0, I1, I0),
+                                b_thread_buf,
+                                make_tuple(I0, I0, I0),
+                                c_thread_buf,
+                                make_tuple(I1, I0, I0, I0));
+
+            // read B_sub_0
+            b_thread_copy_.Run(b_k_n0_n1_block_desc_,
+                               make_tuple(k, I0, I0),
+                               b_block_buf,
+                               b_k_n0_n1_thread_desc_,
+                               make_tuple(I0, I0, I0),
+                               b_thread_buf);
+
+            // C_sub_11 += transpose(A_sub_1) * B_sub_1
+            threadwise_gemm.Run(a_thread_buf,
+                                make_tuple(I0, I1, I0),
+                                b_thread_buf,
+                                make_tuple(I0, I1, I0),
+                                c_thread_buf,
+                                make_tuple(I1, I0, I1, I0));
+
+            // read B_sub_1
+            b_thread_copy_.Run(b_k_n0_n1_block_desc_,
+                               make_tuple(k, I1, I0),
+                               b_block_buf,
+                               b_k_n0_n1_thread_desc_,
+                               make_tuple(I0, I1, I0),
+                               b_thread_buf);
+
+            // read A_sub_1
+            a_thread_copy_.Run(a_k_m0_m1_block_desc_,
+                               make_tuple(k, I1, I0),
+                               a_block_buf,
+                               a_k_m0_m1_thread_desc_,
+                               make_tuple(I0, I1, I0),
+                               a_thread_buf);
+
+            // C_sub_00 += transpose(A_sub_0) * B_sub_0
+            threadwise_gemm.Run(a_thread_buf,
+                                make_tuple(I0, I0, I0),
+                                b_thread_buf,
+                                make_tuple(I0, I0, I0),
+                                c_thread_buf,
+                                make_tuple(I0, I0, I0, I0));
+
+            // C_sub_01 += transpose(A_sub_0) * B_sub_1
+            threadwise_gemm.Run(a_thread_buf,
+                                make_tuple(I0, I0, I0),
+                                b_thread_buf,
+                                make_tuple(I0, I1, I0),
+                                c_thread_buf,
+                                make_tuple(I0, I0, I1, I0));
+        });
+
+        // C_sub_10 += transpose(A_sub_1) * B_sub_0
+        threadwise_gemm.Run(a_thread_buf,
+                            make_tuple(I0, I1, I0),
+                            b_thread_buf,
+                            make_tuple(I0, I0, I0),
+                            c_thread_buf,
+                            make_tuple(I1, I0, I0, I0));
+
+        // C_sub_11 += transpose(A_sub_1) * B_sub_1
+        threadwise_gemm.Run(a_thread_buf,
+                            make_tuple(I0, I1, I0),
+                            b_thread_buf,
+                            make_tuple(I0, I1, I0),
+                            c_thread_buf,
+                            make_tuple(I1, I0, I1, I0));
+    }
+
+    private:
+    // A[K, M0, M1]
+    static constexpr auto a_k_m0_m1_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<KPerThread>{}, Number<M0>{}, Number<M1PerThreadM11>{}));
+
+    // B[K, N0, N1]
+    static constexpr auto b_k_n0_n1_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<KPerThread>{}, Number<N0>{}, Number<N1PerThreadN11>{}));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
+                                                         FloatA,
+                                                         decltype(a_k_m0_m1_block_desc_),
+                                                         decltype(a_k_m0_m1_thread_desc_),
+                                                         Sequence<KPerThread, 1, M1PerThreadM11>,
+                                                         Sequence<0, 1, 2>,
+                                                         2,
+                                                         AThreadCopyScalarPerVector_M11,
+                                                         1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatB,
+                                                         FloatB,
+                                                         decltype(b_k_n0_n1_block_desc_),
+                                                         decltype(b_k_n0_n1_thread_desc_),
+                                                         Sequence<KPerThread, 1, N1PerThreadN11>,
+                                                         Sequence<0, 1, 2>,
+                                                         2,
+                                                         BThreadCopyScalarPerVector_N11,
+                                                         1>;
+
+    CIndex c_thread_origin_data_idx_;
+
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
+};
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp
new file mode 100644
index 0000000..f456557
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp
@@ -0,0 +1,178 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_BLOCKWISE_GEMM_DLOPS_V3_HPP
+#define CK_BLOCKWISE_GEMM_DLOPS_V3_HPP
+
+#include "common_header.hpp"
+#include "threadwise_gemm_dlops_v3.hpp"
+
+namespace ck {
+
+template <index_t BlockSize,
+          typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename ABlockDesc_E1_K1_E2,
+          typename BBlockDesc_E1_N_Ho_Wo_E2,
+          typename CThreadDesc_K_N_Ho_Wo,
+          index_t EPerThreadLoop,
+          index_t KPerThreadLoop>
+struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+
+    using AIndex = MultiIndex<3>;
+    using BIndex = MultiIndex<3>;
+    using CIndex = MultiIndex<4>;
+
+    static constexpr auto E1        = ABlockDesc_E1_K1_E2{}.GetLength(I0);
+    static constexpr auto KPerBlock = ABlockDesc_E1_K1_E2{}.GetLength(I1);
+    static constexpr auto E2        = ABlockDesc_E1_K1_E2{}.GetLength(I2);
+
+    static constexpr auto HoPerBlock = BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I2);
+    static constexpr auto WoPerBlock = BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I3);
+
+    static constexpr auto KPerThread  = CThreadDesc_K_N_Ho_Wo{}.GetLength(I0);
+    static constexpr auto HoPerThread = CThreadDesc_K_N_Ho_Wo{}.GetLength(I2);
+    static constexpr auto WoPerThread = CThreadDesc_K_N_Ho_Wo{}.GetLength(I3);
+
+    static constexpr auto a_thread_mtx_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<EPerThreadLoop>{}, Number<KPerThreadLoop>{}, Number<E2>{}));
+
+    static constexpr auto b_thread_mtx_ =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<EPerThreadLoop>{},
+                                                       Number<1>{},
+                                                       Number<HoPerThread>{},
+                                                       Number<WoPerThread>{},
+                                                       Number<E2>{}));
+
+    static constexpr auto c_thread_mtx_ = make_naive_tensor_descriptor_packed(make_tuple(
+        Number<KPerThreadLoop>{}, Number<1>{}, Number<HoPerThread>{}, Number<WoPerThread>{}));
+
+    __device__ BlockwiseGemmDlops_km_kn_m0m1n0n1_v3()
+        : c_thread_origin_data_idx_{GetBeginOfCThreadDesc_K_N_Ho_Wo(get_thread_local_1d_id())},
+          a_thread_copy_{make_tuple(0, c_thread_origin_data_idx_[I0] * KPerThread, 0)}
+    {
+        static_assert(ABlockDesc_E1_K1_E2::IsKnownAtCompileTime() &&
+                          BBlockDesc_E1_N_Ho_Wo_E2::IsKnownAtCompileTime() &&
+                          CThreadDesc_K_N_Ho_Wo::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(
+            ABlockDesc_E1_K1_E2{}.GetLength(I0) == BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I0) &&
+                ABlockDesc_E1_K1_E2{}.GetLength(I2) == BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I4),
+            "wrong! E dimension not consistent\n");
+
+        static_assert(E1 % EPerThreadLoop == 0, "");
+        static_assert(KPerThread % KPerThreadLoop == 0, "");
+
+        static_assert(KPerBlock % KPerThread == 0 && HoPerBlock % HoPerThread == 0 &&
+                          WoPerBlock % WoPerThread == 0,
+                      "wrong! Cannot evenly divide work among\n");
+
+        constexpr auto KThreadCluster = KPerBlock / KPerThread;
+        constexpr auto HThreadCluster = HoPerBlock / HoPerThread;
+        constexpr auto WThreadCluster = WoPerBlock / WoPerThread;
+
+        static_assert(BlockSize == KThreadCluster * HThreadCluster * WThreadCluster,
+                      "wrong! wrong blocksize\n");
+    }
+
+    __device__ static constexpr auto GetCThreadDesc_K_N_Ho_WoLengths()
+    {
+        return Sequence<KPerThread, I1, HoPerThread, WoPerThread>{};
+    }
+
+    __device__ static CIndex GetBeginOfCThreadDesc_K_N_Ho_Wo(index_t thread_id)
+    {
+        constexpr auto K0 = KPerBlock / KPerThread;
+        constexpr auto N0 = I1;
+        constexpr auto H0 = HoPerBlock / HoPerThread;
+        constexpr auto W0 = WoPerBlock / WoPerThread;
+
+        constexpr auto c_threadid_to_k_n_h_w_thread_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(K0, N0, H0, W0))),
+                make_tuple(Sequence<0, 1, 2, 3>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto c_k_n_h_w_thread_cluster_idx =
+            c_threadid_to_k_n_h_w_thread_cluster_adaptor.CalculateBottomIndex(
+                make_multi_index(thread_id));
+
+        return c_k_n_h_w_thread_cluster_idx;
+    }
+
+    template <typename ABlockBuffer, typename BThreadBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BThreadBuffer& b_thread_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        static_assert(
+            is_same<remove_cvref_t<typename ABlockBuffer::type>, remove_cvref_t<FloatA>>::value &&
+            is_same<remove_cvref_t<typename BThreadBuffer::type>, remove_cvref_t<FloatB>>::value &&
+            is_same<remove_cvref_t<typename CThreadBuffer::type>, remove_cvref_t<FloatC>>::value &&
+            "wrong! inconsistent type");
+
+        constexpr auto a_block_mtx = ABlockDesc_E1_K1_E2{};
+
+        // thread A buffer for GEMM
+        StaticBuffer<AddressSpaceEnum::Vgpr, FloatA, a_thread_mtx_.GetElementSpaceSize(), true>
+            a_thread_buf;
+
+        constexpr auto threadwise_gemm = ThreadwiseGemmDlops_km_kn_mn_v3<FloatA,
+                                                                         FloatB,
+                                                                         FloatC,
+                                                                         decltype(a_thread_mtx_),
+                                                                         decltype(b_thread_mtx_),
+                                                                         decltype(c_thread_mtx_)>{};
+
+        static_for<0, E1, EPerThreadLoop>{}([&](auto e_begin) {
+            static_for<0, KPerThread, KPerThreadLoop>{}([&](auto k_begin) {
+                a_thread_copy_.Run(a_block_mtx,
+                                   make_tuple(e_begin, k_begin, I0),
+                                   a_block_buf,
+                                   a_thread_mtx_,
+                                   make_tuple(I0, I0, I0),
+                                   a_thread_buf);
+
+                threadwise_gemm.Run(a_thread_buf,
+                                    make_tuple(I0, I0, I0),
+                                    b_thread_buf,
+                                    make_tuple(e_begin, I0, I0, I0, I0),
+                                    c_thread_buf,
+                                    make_tuple(k_begin, I0, I0, I0));
+            });
+        });
+    }
+
+    template <typename ABlockSliceMoveStepIdx>
+    __device__ void MoveABlockSliceWindow(const ABlockSliceMoveStepIdx& a_block_slice_move_step_idx)
+    {
+        a_thread_copy_.MoveSrcSliceWindow(ABlockDesc_E1_K1_E2{}, a_block_slice_move_step_idx);
+    }
+
+    private:
+    using AThreadCopy =
+        ThreadwiseTensorSliceTransfer_v4<FloatA,
+                                         FloatA,
+                                         ABlockDesc_E1_K1_E2,
+                                         decltype(a_thread_mtx_),
+                                         Sequence<EPerThreadLoop, KPerThreadLoop, E2>,
+                                         Sequence<0, 1, 2>,
+                                         2,
+                                         E2,
+                                         E2>;
+
+    CIndex c_thread_origin_data_idx_;
+
+    AThreadCopy a_thread_copy_;
+};
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
new file mode 100644
index 0000000..d0e3be4
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
@@ -0,0 +1,1000 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+enum struct LoopScheduler
+{
+    Default,
+    Interwave,
+};
+
+constexpr LoopScheduler make_default_loop_scheduler()
+{
+#if CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING
+    return LoopScheduler::Interwave;
+#else
+    return LoopScheduler::Default;
+#endif // if CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING
+}
+
+template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+__host__ __device__ static constexpr auto
+MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(const TileDesc_K0_MN_K1&)
+{
+    constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{});
+    constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{});
+
+    return transform_tensor_descriptor(
+        TileDesc_K0_MN_K1{},
+        make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number<K0>{}, Number<K1>{})),
+                   make_unmerge_transform(
+                       make_tuple(Number<MNXdlPerWave>{}, Number<MNWaves>{}, Number<MNPerXdl>{}))),
+        make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+        make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename AK0MK1BlockDesc,
+          typename BK0NK1BlockDesc,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack>
+struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    static constexpr index_t WaveSize = get_warp_size();
+
+    static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1);
+    static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1);
+    static constexpr index_t KPerBlock =
+        BK0NK1BlockDesc{}.GetLength(I0) * BK0NK1BlockDesc{}.GetLength(I2);
+
+    static constexpr index_t A_K0 = AK0MK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t B_K0 = BK0NK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
+
+    static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack>{};
+
+    static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              xdlops_gemm.GetRegSizePerXdlops(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = ThisThreadBlock::GetThreadId();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPerThread * xdlops_a_idx[I0]);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPerThread * xdlops_b_idx[I0]);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
+
+        constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
+
+        return make_tuple(Number<m0>{},
+                          Number<n0>{},
+                          waveId_m,
+                          waveId_n,
+                          blk_idx[I0],
+                          blk_idx[I1],
+                          blk_idx[I2],
+                          blk_idx[I3]);
+    }
+
+    __host__ __device__ BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1()
+    {
+        static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
+                          BK0NK1BlockDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
+                      "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_block_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_grid_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto MakeABlockDescriptor_M0_M1_M2_K()
+    {
+        return transform_tensor_descriptor(
+            AK0MK1BlockDesc{},
+            make_tuple(
+                make_merge_transform_v3_division_mod(make_tuple(Number<A_K0>{}, Number<A_K1>{})),
+                make_unmerge_transform(
+                    make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    __host__ __device__ static constexpr auto MakeBBlockDescriptor_N0_N1_N2_K()
+    {
+        return transform_tensor_descriptor(
+            BK0NK1BlockDesc{},
+            make_tuple(
+                make_merge_transform_v3_division_mod(make_tuple(Number<B_K0>{}, Number<B_K1>{})),
+                make_unmerge_transform(
+                    make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerXDL>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k = MakeABlockDescriptor_M0_M1_M2_K();
+    static constexpr auto b_block_desc_n0_n1_n2_k = MakeBBlockDescriptor_N0_N1_N2_K();
+
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            // read A
+            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                               make_tuple(m0, I0, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_,
+                               make_tuple(I0, I0, I0, I0),
+                               a_thread_buf);
+
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                // read B
+                b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                   make_tuple(n0, I0, I0, I0),
+                                   b_block_buf,
+                                   b_thread_desc_,
+                                   make_tuple(I0, I0, I0, I0),
+                                   b_thread_buf);
+
+                static_for<0, KPerThread, KPack>{}([&](auto k) {
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+
+                    static_for<0, KPack, 1>{}([&](auto i) {
+                        a_thread_vec.template AsType<FloatAB>()(i) = a_thread_buf
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(i) = b_thread_buf
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+        });
+    }
+
+    protected:
+    // A[M0, M1, M2, KPerThread]
+    static constexpr auto a_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
+
+    // B[N0, N1, N2, KPerThread]
+    static constexpr auto b_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
+
+    // C[M, N, NumRegXdlops]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerThread>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(b_block_desc_n0_n1_n2_k),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerThread>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
+    BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
+};
+
+// Note: To facilitate the inter-wave loop scheduler, we need to explicitly set the macro
+// CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING=1 as a few intrinsics are not yet available in
+// the latest ROCm release. For unsupported compilers, inter-wave loop scheduler falls back to the
+// default loop scheduler which is given by the macro CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING=0
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename AK0MK1BlockDesc,
+          typename BK0NK1BlockDesc,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack,
+          index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS>
+struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
+    : public BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                 FloatAB,
+                                                                 FloatAcc,
+                                                                 AK0MK1BlockDesc,
+                                                                 BK0NK1BlockDesc,
+                                                                 MPerXDL,
+                                                                 NPerXDL,
+                                                                 MRepeat,
+                                                                 NRepeat,
+                                                                 KPack>
+{
+    using Base = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                     FloatAB,
+                                                                     FloatAcc,
+                                                                     AK0MK1BlockDesc,
+                                                                     BK0NK1BlockDesc,
+                                                                     MPerXDL,
+                                                                     NPerXDL,
+                                                                     MRepeat,
+                                                                     NRepeat,
+                                                                     KPack>;
+
+#if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
+    using Base::a_block_desc_m0_m1_m2_k;
+    using Base::A_K1;
+    using Base::b_block_desc_n0_n1_n2_k;
+    using Base::B_K1;
+    using Base::c_thread_buf_;
+    using Base::c_thread_desc_;
+    using Base::CalculateAThreadOriginDataIndex;
+    using Base::CalculateBThreadOriginDataIndex;
+    using Base::I0;
+    using Base::I1;
+    using Base::KPerThread;
+    using Base::xdlops_gemm;
+
+    static constexpr index_t KPerInnerLoop = math::max(KPerThread / NumMacClusters, KPack);
+
+    // 2-wave optimized blockwise gemm
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        static_for<0, KPerThread, KPerInnerLoop>{}([&](auto k) {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                // read A
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                   make_tuple(m0, I0, I0, k),
+                                   a_block_buf,
+                                   a_thread_desc_,
+                                   make_tuple(m0, I0, I0, I0),
+                                   a_thread_buf);
+            });
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                // read B
+                b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                   make_tuple(n0, I0, I0, k),
+                                   b_block_buf,
+                                   b_thread_desc_,
+                                   make_tuple(n0, I0, I0, I0),
+                                   b_thread_buf);
+            });
+            __builtin_amdgcn_s_barrier();
+            // NOTE: Synchronize threads in a workgroup at the start of each MAC cluster, but except
+            // the first, as we can shorten non-MAC cluster a bit and there's no observable negative
+            // impact. The desired effect is waves in a workgroup executing MAC in sync. This avoids
+            // some out-of-sync waves hijacking MAC resource from other workgroups and reducing the
+            // chance of latency hiding by waiting for the rest of the workgroup at the eventual
+            // sync point.
+            if constexpr(k.value != 0 || KPerInnerLoop == KPerThread)
+            {
+                asm volatile("s_barrier" ::);
+                __builtin_amdgcn_s_barrier();
+            }
+            static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto i) {
+                            a_thread_vec.template AsType<FloatAB>()(i) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, 0, 0, k_ + i))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(i) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, 0, 0, k_ + i))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // The block_sync_lds() here performs double duty:
+                        // A) safeguard against data hazard because barrier from blockwise_gemm is
+                        // moved here B) reduce VMEM FIFO congestion by applying small delays to
+                        // different wavefronts It is performed near the end of MAC cluster to
+                        // minimize lgkmcnt penalty
+                        if constexpr(k.value == KPerThread - KPerInnerLoop &&
+                                     k_.value == KPerInnerLoop - KPack && m0.value == MRepeat - 1 &&
+                                     n0.value == NRepeat - 1)
+                        {
+                            __builtin_amdgcn_s_barrier();
+                            block_sync_lds();
+                            __builtin_amdgcn_s_barrier();
+                        }
+
+                        // TODO: insert setprio in more precise manner since we
+                        // could have more than >1 MFMA instructions in single call
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
+                        {
+                            __builtin_amdgcn_s_barrier();
+                            //__builtin_amdgcn_s_setprio(1);
+                            asm("S_SETPRIO 1");
+                            __builtin_amdgcn_s_barrier();
+                        }
+                    });
+                });
+            });
+            __builtin_amdgcn_s_barrier();
+            //__builtin_amdgcn_s_setprio(0);
+            asm("S_SETPRIO 0");
+            __builtin_amdgcn_s_barrier();
+        });
+    }
+
+    protected:
+    // A[M0, M1, M2, KPerInnerLoop]
+    static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, I1, I1, Number<KPerInnerLoop>{}));
+
+    // B[N0, N1, N2, KPerInnerLoop]
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, I1, I1, Number<KPerInnerLoop>{}));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerInnerLoop>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(b_block_desc_n0_n1_n2_k),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerInnerLoop>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
+    BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
+
+#endif // #if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
+};
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename AK0MK1BlockDesc,
+          typename BK0NK1BlockDesc,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack,
+          LoopScheduler LoopSched>
+constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
+{
+    if constexpr(LoopSched == LoopScheduler::Default)
+    {
+        return BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                   FloatAB,
+                                                                   FloatAcc,
+                                                                   AK0MK1BlockDesc,
+                                                                   BK0NK1BlockDesc,
+                                                                   MPerXDL,
+                                                                   NPerXDL,
+                                                                   MRepeat,
+                                                                   NRepeat,
+                                                                   KPack>{};
+    }
+    else if constexpr(LoopSched == LoopScheduler::Interwave)
+    {
+        return BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                            FloatAB,
+                                                                            FloatAcc,
+                                                                            AK0MK1BlockDesc,
+                                                                            BK0NK1BlockDesc,
+                                                                            MPerXDL,
+                                                                            NPerXDL,
+                                                                            MRepeat,
+                                                                            NRepeat,
+                                                                            KPack>{};
+    }
+};
+
+// Blockwise gemm supporting
+// 1. regular XDL output M2_M3_M4_M2 and transposed XDL output M2_N2_N3_N4
+// 2. decoupled input tile descriptor and mma tile descriptor in order to support both vgpr and LDS
+// source buffer
+// 3. configurable k index starting position and step size after each FMA/XDL instruction
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack,
+          bool TransposeC = false,
+          index_t AMmaKStride =
+              KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, TransposeC>{}.K0PerXdlops,
+          index_t BMmaKStride =
+              KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, TransposeC>{}.K0PerXdlops>
+struct BlockwiseGemmXdlops_v2
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    static constexpr index_t WaveSize = get_warp_size();
+
+    static constexpr index_t A_K0 = ATileDesc{}.GetLength(I0);
+    static constexpr index_t B_K0 = BTileDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = ATileDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2);
+
+    static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, TransposeC>{};
+
+    static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
+
+    static_assert(KPerThread % KPack == 0,
+                  "Wrong KPack setting; try increasing KPerThread or decreasing KPack");
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              xdlops_gemm.GetRegSizePerXdlops(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = ThisThreadBlock::GetThreadId();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPack * xdlops_a_idx[I0]);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPack * xdlops_b_idx[I0]);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
+
+        constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
+
+        return make_tuple(
+            m0, n0, waveId_m, waveId_n, blk_idx[I0], blk_idx[I1], blk_idx[I2], blk_idx[I3]);
+    }
+
+    using Tuple4 = decltype(CalculateAThreadOriginDataIndex());
+
+    __host__ __device__ BlockwiseGemmXdlops_v2(Tuple4 a_origin = CalculateAThreadOriginDataIndex(),
+                                               Tuple4 b_origin = CalculateBThreadOriginDataIndex())
+        : a_thread_copy_(a_origin), b_thread_copy_(b_origin)
+    {
+        static_assert(AMmaTileDesc::IsKnownAtCompileTime() && BMmaTileDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
+                      "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    __host__ __device__ BlockwiseGemmXdlops_v2(const BlockwiseGemmXdlops_v2& other)
+        : a_thread_copy_(other.a_origin), b_thread_copy_(other.b_origin)
+    {
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, N, M0, M1, M2));
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_block_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_grid_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    static constexpr AMmaTileDesc a_block_desc_m0_m1_m2_k;
+    static constexpr BMmaTileDesc b_block_desc_n0_n1_n2_k;
+
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        static_for<0, KPerThread / KPack, 1>{}([&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                // read A
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                   make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                   a_block_buf,
+                                   a_thread_desc_,
+                                   make_tuple(I0, I0, I0, I0),
+                                   a_thread_buf);
+
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    // read B
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                       b_block_buf,
+                                       b_thread_desc_,
+                                       make_tuple(I0, I0, I0, I0),
+                                       b_thread_buf);
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+
+                    static_for<0, KPack, 1>{}([&](auto i) {
+                        a_thread_vec.template AsType<FloatAB>()(i) = a_thread_buf
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, i))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(i) = b_thread_buf
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, i))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+        });
+    }
+
+    protected:
+    // A[M0, M1, M2, KPack]
+    static constexpr auto a_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPack>{}));
+
+    // B[N0, N1, N2, KPack]
+    static constexpr auto b_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPack>{}));
+
+    // C[M, N, NumRegXdlops]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(b_block_desc_n0_n1_n2_k),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp
new file mode 100644
index 0000000..aa814ab
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp
@@ -0,0 +1,321 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename AK0MK1BlockDesc,
+          typename BK0K0BN0N1N2N3K1BlockDesc,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack>
+struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr index_t WaveSize = 64;
+
+    static constexpr index_t KPerBlock = K0PerBlock * KPack;
+
+    static constexpr index_t A_K0 = AK0MK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
+
+    static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack>{};
+
+    static constexpr index_t KPerThread  = KPerBlock / xdlops_gemm.K0PerXdlops;
+    static constexpr index_t K0PerThread = K0PerBlock / xdlops_gemm.K0PerXdlops;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              xdlops_gemm.GetRegSizePerXdlops(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPerThread * xdlops_a_idx[I0]);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPerThread * xdlops_b_idx[I0]);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
+
+        constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    __host__ __device__ BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1r1()
+    {
+        static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
+                          BK0K0BN0N1N2N3K1BlockDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(BlockSize == MWaves * NWaves * WaveSize,
+                      "BlockSize != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_block_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_grid_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto MakeABlockDescriptor_M0_M1_M2_K()
+    {
+        return transform_tensor_descriptor(
+            AK0MK1BlockDesc{},
+            make_tuple(
+                make_merge_transform_v3_division_mod(make_tuple(Number<A_K0>{}, Number<A_K1>{})),
+                make_unmerge_transform(
+                    make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    __device__ void MoveABlockSliceWindow()
+    {
+        a_thread_copy_.MoveSrcSliceWindow(a_block_desc_m0_m1_m2_k,
+                                          make_multi_index(0, 0, 0, K0PerBlock * KPack));
+    }
+    __device__ void ResetABlockStartWindow()
+    {
+        a_thread_copy_.SetSrcCoord(CalculateAThreadOriginDataIndex());
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k = MakeABlockDescriptor_M0_M1_M2_K();
+
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_thread_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            // read A
+            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                               make_tuple(m0, I0, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_,
+                               make_tuple(I0, I0, I0, I0),
+                               a_thread_buf);
+
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                // read B
+                static_for<0, KPerThread, KPack>{}([&](auto k) {
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+                    constexpr index_t k0 = k / KPack;
+                    static_for<0, KPack, 1>{}([&](auto i) {
+                        a_thread_vec.template AsType<FloatAB>()(i) = a_thread_buf
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(i) = b_thread_buf
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(k0, n0, i))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+        });
+    }
+
+    private:
+    // A[M0, M1, M2, KPerThread]
+    static constexpr auto a_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
+
+    // B[N0, N1, N2, KPerThread]
+    static constexpr auto b_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<K0PerThread>{}, // KPerThread
+                                                       Number<NRepeat>{},     // repeat
+                                                       Number<KPack>{}));
+
+    // C[M, N, NumRegXdlops]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerThread>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
new file mode 100644
index 0000000..d7ec177
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
@@ -0,0 +1,115 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+
+namespace ck {
+
+template <index_t BlockSize,
+          typename AccDataType,
+          typename ThreadMap_M_K, // thread_id to m_k
+          typename ThreadClusterDesc_M_K,
+          typename ThreadSliceDesc_M_K,
+          bool IgnoreNaN = false>
+struct BlockwiseSoftmax
+{
+    static constexpr auto I0         = Number<0>{};
+    static constexpr auto I1         = Number<1>{};
+    static constexpr index_t MRepeat = ThreadSliceDesc_M_K{}.GetLength(I0);
+    static constexpr index_t KRepeat = ThreadSliceDesc_M_K{}.GetLength(I1);
+
+    using ThreadSliceDesc_M = decltype(
+        make_naive_tensor_descriptor_packed(make_tuple(ThreadSliceDesc_M_K{}.GetLength(I0))));
+
+    using ThreadwiseMaxReduce = typename conditional<
+        IgnoreNaN,
+        ThreadwiseReduction<AccDataType,
+                            ThreadSliceDesc_M_K,
+                            ThreadSliceDesc_M,
+                            reduce::Max,
+                            false,
+                            detail::AccumulateWithNanIgnore<reduce::Max, AccDataType>>,
+        ThreadwiseReduction<AccDataType,
+                            ThreadSliceDesc_M_K,
+                            ThreadSliceDesc_M,
+                            reduce::Max,
+                            false>>::type;
+
+    using ThreadwiseSumReduce = typename conditional<
+        IgnoreNaN,
+        ThreadwiseReduction<AccDataType,
+                            ThreadSliceDesc_M_K,
+                            ThreadSliceDesc_M,
+                            reduce::Add,
+                            false,
+                            detail::AccumulateWithNanIgnore<reduce::Add, AccDataType>>,
+        ThreadwiseReduction<AccDataType,
+                            ThreadSliceDesc_M_K,
+                            ThreadSliceDesc_M,
+                            reduce::Add,
+                            false>>::type;
+
+    using ThreadClusterLengths_M_K = decltype(ThreadClusterDesc_M_K{}.GetLengths());
+
+    using BlockwiseMaxReduce = PartitionedBlockwiseReduction_v2<AccDataType,
+                                                                BlockSize,
+                                                                ThreadClusterLengths_M_K,
+                                                                ThreadMap_M_K,
+                                                                reduce::Max,
+                                                                false>;
+
+    using BlockwiseSumReduce = PartitionedBlockwiseReduction_v2<AccDataType,
+                                                                BlockSize,
+                                                                ThreadClusterLengths_M_K,
+                                                                ThreadMap_M_K,
+                                                                reduce::Add,
+                                                                false>;
+
+    using BufferType = StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MRepeat, true>;
+
+    template <typename CThreadBuffer, typename WorkspaceBuffer>
+    __host__ __device__ void Run(CThreadBuffer& in_thread_buf, WorkspaceBuffer& reduce_work_buf)
+    {
+        // find max value
+        static_for<0, MRepeat, 1>{}([&](auto I) {
+            max_value_buf(I) = reduce::Max::template GetIdentityValue<AccDataType>();
+        });
+        ThreadwiseMaxReduce::Reduce(in_thread_buf, max_value_buf);
+        static_for<0, MRepeat, 1>{}([&](auto I) {
+            BlockwiseMaxReduce::Reduce(reduce_work_buf, max_value_buf(I));
+            block_sync_lds();
+        });
+
+        // calculate exp for elements, P=exp(s-max)
+        static_for<0, MRepeat, 1>{}([&](auto iM) {
+            static_for<0, KRepeat, 1>{}([&](auto iK) {
+                auto offset = Number<ThreadSliceDesc_M_K{}.CalculateOffset(make_tuple(iM, iK))>{};
+                in_thread_buf(offset) = IgnoreNaN && ck::math::isnan(in_thread_buf[offset])
+                                            ? 0
+                                            : math::exp(in_thread_buf[offset] - max_value_buf(iM));
+            });
+        });
+
+        // sum data
+        static_for<0, MRepeat, 1>{}([&](auto I) {
+            sum_value_buf(I) = reduce::Add::template GetIdentityValue<AccDataType>();
+        });
+        ThreadwiseSumReduce::Reduce(in_thread_buf, sum_value_buf);
+        static_for<0, MRepeat, 1>{}([&](auto I) {
+            BlockwiseSumReduce::Reduce(reduce_work_buf, sum_value_buf(I));
+            block_sync_lds();
+        });
+    }
+
+    BufferType max_value_buf;
+    BufferType sum_value_buf;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp
new file mode 100644
index 0000000..03e4d42
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp
@@ -0,0 +1,156 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp"
+
+namespace ck {
+
+// this version does following things to avoid scratch memory issue
+// 1. Use StaticallyIndexedArray instead of C array for thread buffer
+// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
+// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
+template <index_t BlockSize,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename BlockSliceLengths,
+          typename ThreadSliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          typename SrcVectorTensorLengths,
+          typename DstVectorTensorLengths,
+          typename SrcVectorTensorContiguousDimOrder,
+          typename DstVectorTensorContiguousDimOrder,
+          bool ThreadTransferSrcResetCoordinateAfterRun,
+          bool ThreadTransferDstResetCoordinateAfterRun>
+struct BlockwiseTensorSliceTransfer_v5r1
+{
+    static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr BlockwiseTensorSliceTransfer_v5r1(const SrcDesc& src_desc,
+                                                           const Index& src_block_slice_origin,
+                                                           const DstDesc& dst_desc,
+                                                           const Index& dst_block_slice_origin)
+        : threadwise_transfer_(
+              src_desc, make_zero_multi_index<nDim>(), dst_desc, make_zero_multi_index<nDim>())
+
+    {
+        static_assert(nDim == remove_cvref_t<SrcDesc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
+                          nDim == BlockSliceLengths::Size() && nDim == ThreadSliceLengths::Size() &&
+                          nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<BlockSliceLengths, decltype(ThreadSliceLengths{} * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(BlockSize >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! BlockSize too small");
+
+        if(BlockSize == thread_cluster_desc_.GetElementSize() or
+           get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(get_thread_local_1d_id()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * ThreadSliceLengths{};
+
+            threadwise_transfer_.SetSrcSliceOrigin(src_desc,
+                                                   src_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    template <typename SrcBuffer>
+    __device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf)
+    {
+        if(BlockSize == thread_cluster_desc_.GetElementSize() or
+           get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.RunRead(src_desc, src_buf);
+        }
+    }
+
+    template <typename DstBuffer>
+    __device__ void RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf)
+    {
+        if(BlockSize == thread_cluster_desc_.GetElementSize() or
+           get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.RunWrite(dst_desc, dst_buf);
+        }
+    }
+
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& step)
+    {
+        if(BlockSize == thread_cluster_desc_.GetElementSize() or
+           get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(src_desc, step);
+        }
+    }
+
+    // SrcMoveSliceWindowStepHack to control index calculation move slice window
+    template <typename SrcMoveSliceWindowStepHack>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDesc& src_desc,
+                       const Index& step,
+                       const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
+    {
+        if(BlockSize == thread_cluster_desc_.GetElementSize() or
+           get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(
+                src_desc, step, src_move_slice_window_step_hack);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
+    {
+        if(BlockSize == thread_cluster_desc_.GetElementSize() or
+           get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v5r1<ThreadSliceLengths,
+                                           DstInMemOp,
+                                           SrcData,
+                                           DstData,
+                                           SrcDesc,
+                                           DstDesc,
+                                           SrcDimAccessOrder,
+                                           DstDimAccessOrder,
+                                           SrcVectorTensorLengths,
+                                           DstVectorTensorLengths,
+                                           SrcVectorTensorContiguousDimOrder,
+                                           DstVectorTensorContiguousDimOrder,
+                                           ThreadTransferSrcResetCoordinateAfterRun,
+                                           ThreadTransferDstResetCoordinateAfterRun>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_welford.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_welford.hpp
new file mode 100644
index 0000000..3165086
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_welford.hpp
@@ -0,0 +1,108 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/utility/reduction_common.hpp"
+
+namespace ck {
+
+// clang-format off
+// Assume:
+//  1) work_buffer is buffer (typically LDS) allocated outside as workspace
+//  2) work_buffer has T elements, and space size is no less than 3*BlockSize
+//  3) mean_value, var_value and count is the input data in vgpr from each thread
+//  4) mean_value, var_value and count is the over-written reduced output in vgpr for each thread
+//  5) Merge mean and M from ThreadwiseWelford
+// clang-format on
+template <typename T,
+          index_t BlockSize,
+          typename ThreadClusterLengths_M_K,
+          typename ThreadClusterArrangeOrder,
+          bool GetActualVariance = true>
+struct BlockwiseWelford
+{
+    static_assert(BlockSize == ThreadClusterLengths_M_K::At(0) * ThreadClusterLengths_M_K::At(1),
+                  "The product of cluster lengths should be same as BlockSize!");
+
+    static constexpr auto BufferLength_M = ThreadClusterLengths_M_K::At(0);
+    static constexpr auto BufferLength_K = ThreadClusterLengths_M_K::At(1);
+
+    static constexpr auto block_buf_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<BufferLength_M>{}, Number<BufferLength_K>{}));
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    __device__ static inline void
+    Merge(T& mean_a, T& var_a, int& count_a, T mean_b, T var_b, int count_b)
+    {
+        int count            = count_a + count_b;
+        T count_b_over_count = count == 0 ? type_convert<T>(0) : type_convert<T>(count_b) / count;
+        T delta              = mean_b - mean_a;
+        mean_a += delta * count_b_over_count;
+        var_a += var_b + delta * delta * count_a * count_b_over_count;
+        count_a = count;
+    }
+
+    __device__ static void Run(T& mean_value, T& var_value, int& count)
+    {
+        __shared__ T mean_block_buf[BlockSize];
+        __shared__ T var_block_buf[BlockSize];
+        __shared__ int count_block_buf[BlockSize];
+
+        constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[Number<0>{}];
+        const auto thread_k_cluster_id = thread_cluster_idx[Number<1>{}];
+
+        index_t offset1 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx);
+
+        mean_block_buf[offset1]  = mean_value;
+        var_block_buf[offset1]   = var_value;
+        count_block_buf[offset1] = count;
+
+        block_sync_lds();
+
+        static_for<0, cluster_len_shift, 1>{}([&](auto I) {
+            constexpr index_t indOffset = 1 << (cluster_len_shift - 1 - I());
+
+            if(thread_k_cluster_id < indOffset)
+            {
+                index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
+                                                                     make_tuple(0, indOffset));
+
+                T mean1    = mean_block_buf[offset1];
+                T var1     = var_block_buf[offset1];
+                int count1 = count_block_buf[offset1];
+
+                T mean2    = mean_block_buf[offset2];
+                T var2     = var_block_buf[offset2];
+                int count2 = count_block_buf[offset2];
+
+                Merge(mean1, var1, count1, mean2, var2, count2);
+
+                mean_block_buf[offset1]  = mean1;
+                var_block_buf[offset1]   = var1;
+                count_block_buf[offset1] = count1;
+            }
+
+            block_sync_lds();
+        });
+
+        index_t offset = block_buf_desc_m_k.CalculateOffset(make_tuple(thread_m_cluster_id, 0));
+
+        count      = count_block_buf[offset];
+        mean_value = mean_block_buf[offset];
+
+        if constexpr(GetActualVariance)
+            var_value = var_block_buf[offset] / count;
+        else
+            var_value = var_block_buf[offset];
+    };
+};
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp
new file mode 100644
index 0000000..2163ad3
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp
@@ -0,0 +1,244 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+
+namespace ck {
+
+// clang-format off
+// Assume:
+//  1) work_buffer is buffer (typically LDS) allocated outside as workspace, does not include any in/out data
+//  2) work_buffer has AccDataType elements, and space size is no less than BlockSize
+//  3) in_out_value is the input data in vgpr from each thread
+//  4) in_out_value is the over-written reduced output in vgpr for each thread
+// clang-format on
+template <typename AccDataType,
+          index_t BlockSize,
+          typename ThreadClusterLengths_M_K,
+          typename ThreadClusterArrangeOrder,
+          typename OpReduce,
+          bool PropagateNan,
+          typename Accumulation =
+              detail::AccumulateWithNanCheck<PropagateNan, OpReduce, AccDataType>>
+struct PartitionedBlockwiseReduction
+{
+    static_assert(BlockSize == ThreadClusterLengths_M_K::At(0) * ThreadClusterLengths_M_K::At(1),
+                  "The product of cluster lengths should be same as BlockSize!");
+
+    static constexpr auto BufferLength_M = ThreadClusterLengths_M_K::At(0);
+    static constexpr auto BufferLength_K = ThreadClusterLengths_M_K::At(1);
+
+    static_assert(BufferLength_K > 1, "Parallel reduction need work on at least two elements");
+
+    static constexpr auto block_buf_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<BufferLength_M>{}, Number<BufferLength_K>{}));
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    template <typename BufferType>
+    __device__ static void Reduce(BufferType& work_buffer, AccDataType& in_out_value)
+    {
+        static_assert(is_same<typename BufferType::type, AccDataType>{},
+                      "Buffer data type should be consistent as AccDataType!");
+
+        constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[Number<0>{}];
+        const auto thread_k_cluster_id = thread_cluster_idx[Number<1>{}];
+
+        work_buffer(block_buf_desc_m_k.CalculateOffset(thread_cluster_idx)) = in_out_value;
+
+        __syncthreads();
+
+        static_for<0, cluster_len_shift, 1>{}([&](auto I) {
+            constexpr index_t indOffset = 1 << (cluster_len_shift - 1 - I());
+
+            if(thread_k_cluster_id < indOffset)
+            {
+                index_t offset1 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx);
+                index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
+                                                                     make_tuple(0, indOffset));
+
+                AccDataType opData1 = work_buffer[offset1];
+                AccDataType opData2 = work_buffer[offset2];
+                Accumulation::Calculate(opData1, opData2);
+                work_buffer(offset1) = opData1;
+            }
+
+            __syncthreads();
+        });
+
+        index_t offset = block_buf_desc_m_k.CalculateOffset(make_tuple(thread_m_cluster_id, 0));
+
+        in_out_value = work_buffer[offset];
+    };
+};
+
+// clang-format off
+// Assume:
+//  1) work_buffer is buffer (typically LDS) allocated outside as workspace, does not include any in/out data
+//  2) work_buffer has AccDataType elements, and space size is no less than BlockSize
+//  3) in_out_value is the input data in vgpr from each thread
+//  4) in_out_value is the over-written reduced output in vgpr for each thread
+// clang-format on
+template <typename AccDataType,
+          index_t BlockSize,
+          typename ThreadClusterLengths_M_K,
+          typename ThreadClusterDesc,
+          typename OpReduce,
+          bool PropagateNan,
+          typename Accumulation =
+              detail::AccumulateWithNanCheck<PropagateNan, OpReduce, AccDataType>>
+struct PartitionedBlockwiseReduction_v2
+{
+    static_assert(BlockSize == ThreadClusterLengths_M_K::At(0) * ThreadClusterLengths_M_K::At(1),
+                  "The product of cluster lengths should be same as BlockSize!");
+
+    static constexpr auto BufferLength_M = ThreadClusterLengths_M_K::At(0);
+    static constexpr auto BufferLength_K = ThreadClusterLengths_M_K::At(1);
+
+    static_assert(BufferLength_K > 1, "Parallel reduction need work on at least two elements");
+
+    static constexpr auto block_buf_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<BufferLength_M>{}, Number<BufferLength_K>{}));
+
+    static constexpr auto thread_cluster_desc = ThreadClusterDesc{};
+
+    template <typename BufferType>
+    __device__ static void Reduce(BufferType& work_buffer, AccDataType& in_out_value)
+    {
+        static_assert(is_same<typename BufferType::type, AccDataType>{},
+                      "Buffer data type should be consistent as AccDataType!");
+
+        constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[Number<0>{}];
+        const auto thread_k_cluster_id = thread_cluster_idx[Number<1>{}];
+
+        work_buffer(block_buf_desc_m_k.CalculateOffset(thread_cluster_idx)) = in_out_value;
+
+        __syncthreads();
+
+        static_for<0, cluster_len_shift, 1>{}([&](auto I) {
+            constexpr index_t indOffset = 1 << (cluster_len_shift - 1 - I());
+
+            if(thread_k_cluster_id < indOffset)
+            {
+                index_t offset1 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx);
+                index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
+                                                                     make_tuple(0, indOffset));
+
+                AccDataType opData1 = work_buffer[offset1];
+                AccDataType opData2 = work_buffer[offset2];
+                Accumulation::Calculate(opData1, opData2);
+                work_buffer(offset1) = opData1;
+            }
+
+            __syncthreads();
+        });
+
+        index_t offset = block_buf_desc_m_k.CalculateOffset(make_tuple(thread_m_cluster_id, 0));
+
+        in_out_value = work_buffer[offset];
+    };
+};
+
+// clang-format off
+// Assume:
+//  1) work_val_buffer/work_idx_buffer is buffer (typically LDS) allocated outside as workspace, does not include any in/out data
+//  2) work_val_buffer/work_idx_buffer has AccDataType/IndexDataType elements, and space size is no less than BlockSize
+//  3) in_out_value/in_out_index is the input data in vgpr from each thread
+//  4) in_out_value/in_out_index is the over-written reduced output in vgpr for each thread
+// clang-format on
+template <
+    typename AccDataType,
+    typename IndexDataType,
+    index_t BlockSize,
+    typename ThreadClusterLengths_M_K,
+    typename ThreadClusterArrangeOrder,
+    typename OpReduce,
+    bool PropagateNan,
+    typename Accumulation =
+        detail::AccumulateWithIndexAndNanCheck<PropagateNan, OpReduce, AccDataType, IndexDataType>>
+struct PartitionedBlockwiseReductionWithIndex
+{
+    static_assert(BlockSize == ThreadClusterLengths_M_K::At(0) * ThreadClusterLengths_M_K::At(1),
+                  "The product of cluster lengths should be same as BlockSize!");
+
+    static constexpr auto BufferLength_M = ThreadClusterLengths_M_K::At(0);
+    static constexpr auto BufferLength_K = ThreadClusterLengths_M_K::At(1);
+
+    static_assert(BufferLength_K > 1, "Parallel reduction need work on at least two elements");
+
+    static constexpr auto block_buf_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<BufferLength_M>{}, Number<BufferLength_K>{}));
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    // This interface accumulates on both data values and indices
+    template <typename BufferType, typename IdxBufferType>
+    __device__ static void Reduce(BufferType& work_val_buffer,
+                                  IdxBufferType& work_idx_buffer,
+                                  AccDataType& in_out_value,
+                                  IndexDataType& in_out_index)
+    {
+        static_assert(is_same<typename BufferType::type, AccDataType>{},
+                      "Buffer data type should be consistent as AccDataType!");
+        static_assert(is_same<typename IdxBufferType::type, IndexDataType>{},
+                      "Buffer data type should be consistent as IndexDataType!");
+
+        constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[Number<0>{}];
+        const auto thread_k_cluster_id = thread_cluster_idx[Number<1>{}];
+
+        work_val_buffer(block_buf_desc_m_k.CalculateOffset(thread_cluster_idx)) = in_out_value;
+        work_idx_buffer(block_buf_desc_m_k.CalculateOffset(thread_cluster_idx)) = in_out_index;
+
+        __syncthreads();
+
+        static_for<0, cluster_len_shift, 1>{}([&](auto I) {
+            constexpr index_t indOffset = 1 << I();
+
+            if(thread_k_cluster_id % (indOffset * 2) == 0)
+            {
+                index_t offset1 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx);
+                index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
+                                                                     make_tuple(0, indOffset));
+
+                AccDataType opData1      = work_val_buffer[offset1];
+                AccDataType opData2      = work_val_buffer[offset2];
+                IndexDataType currIndex1 = work_idx_buffer[offset1];
+                IndexDataType currIndex2 = work_idx_buffer[offset2];
+
+                Accumulation::Calculate(opData1, opData2, currIndex1, currIndex2);
+                work_val_buffer(offset1) = opData1;
+                work_idx_buffer(offset1) = currIndex1;
+            }
+
+            __syncthreads();
+        });
+
+        index_t offset = block_buf_desc_m_k.CalculateOffset(make_tuple(thread_m_cluster_id, 0));
+
+        in_out_value = work_val_buffer[offset];
+        in_out_index = work_idx_buffer[offset];
+    };
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
new file mode 100644
index 0000000..0e5dfb3
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
@@ -0,0 +1,173 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp"
+
+namespace ck {
+
+// this version does following things to avoid scratch memory issue
+// 1. Use StaticallyIndexedArray instead of C array for thread buffer
+// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
+// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
+template <typename ThreadGroup,
+          typename SrcElementwiseOperation,
+          typename DstElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename BlockSliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          index_t DstScalarStrideInVector,
+          bool ThreadTransferSrcResetCoordinateAfterRun,
+          bool ThreadTransferDstResetCoordinateAfterRun,
+          index_t NumThreadScratch = 1>
+struct ThreadGroupTensorSliceTransfer_v4r1
+{
+    static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
+
+    static constexpr auto thread_slice_lengths = BlockSliceLengths{} / ThreadClusterLengths{};
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v4r1(
+        const SrcDesc& src_desc,
+        const Index& src_block_slice_origin,
+        const SrcElementwiseOperation& src_element_op,
+        const DstDesc& dst_desc,
+        const Index& dst_block_slice_origin,
+        const DstElementwiseOperation& dst_element_op)
+        : threadwise_transfer_(src_desc,
+                               make_zero_multi_index<nDim>(),
+                               src_element_op,
+                               dst_desc,
+                               make_zero_multi_index<nDim>(),
+                               dst_element_op)
+
+    {
+        static_assert(nDim == remove_cvref_t<SrcDesc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
+                          nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<BlockSliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrcSliceOrigin(src_desc,
+                                                   src_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    template <typename SrcBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunRead(const SrcDesc& src_desc,
+                            const SrcBuffer& src_buf,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.RunRead(src_desc, src_buf, thread_scratch_id);
+        }
+    }
+
+    template <typename DstBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunWrite(const DstDesc& dst_desc,
+                             DstBuffer& dst_buf,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.RunWrite(dst_desc, dst_buf, thread_scratch_id);
+        }
+    }
+
+    template <typename SrcBuffer, typename DstBuffer, index_t ThreadScratchId>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf,
+                        Number<ThreadScratchId> thread_scratch_id)
+    {
+        RunRead(src_desc, src_buf, thread_scratch_id);
+        RunWrite(dst_desc, dst_buf, thread_scratch_id);
+    }
+
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(src_desc, step);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v3r1<decltype(thread_slice_lengths),
+                                           SrcElementwiseOperation,
+                                           DstElementwiseOperation,
+                                           DstInMemOp,
+                                           SrcData,
+                                           DstData,
+                                           SrcDesc,
+                                           DstDesc,
+                                           SrcDimAccessOrder,
+                                           DstDimAccessOrder,
+                                           SrcVectorDim,
+                                           DstVectorDim,
+                                           SrcScalarPerVector,
+                                           DstScalarPerVector,
+                                           SrcScalarStrideInVector,
+                                           DstScalarStrideInVector,
+                                           ThreadTransferSrcResetCoordinateAfterRun,
+                                           ThreadTransferDstResetCoordinateAfterRun,
+                                           NumThreadScratch>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp
new file mode 100644
index 0000000..5c47a49
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp
@@ -0,0 +1,134 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp"
+
+namespace ck {
+
+// this version does following things to avoid scratch memory issue
+// 1. Use StaticallyIndexedArray instead of C array for thread buffer
+// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
+// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
+template <typename ThreadGroup,
+          typename ElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          bool ThreadTransferSrcResetCoordinateAfterRun,
+          bool ThreadTransferDstResetCoordinateAfterRun>
+struct ThreadGroupTensorSliceTransfer_v6r1
+{
+    static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
+
+    static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v6r1(const SrcDesc& src_desc,
+                                                             const Index& src_block_slice_origin,
+                                                             const DstDesc& dst_desc,
+                                                             const Index& dst_block_slice_origin,
+                                                             const ElementwiseOperation& element_op)
+        : threadwise_transfer_(src_desc,
+                               make_zero_multi_index<nDim>(),
+                               dst_desc,
+                               make_zero_multi_index<nDim>(),
+                               element_op)
+
+    {
+        static_assert(nDim == remove_cvref_t<SrcDesc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
+                          nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == DimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrcSliceOrigin(src_desc,
+                                                   src_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    template <typename SrcBuffer, typename DstBuffer>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.Run(src_desc, src_buf, dst_desc, dst_buf);
+        }
+    }
+
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(src_desc, step);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v6r1<SrcData,
+                                           DstData,
+                                           SrcDesc,
+                                           DstDesc,
+                                           ElementwiseOperation,
+                                           decltype(thread_slice_lengths),
+                                           DimAccessOrder,
+                                           VectorDim,
+                                           ScalarPerVector,
+                                           DstInMemOp,
+                                           ThreadTransferSrcResetCoordinateAfterRun,
+                                           ThreadTransferDstResetCoordinateAfterRun>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp
new file mode 100644
index 0000000..aa33fc0
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp
@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp"
+
+namespace ck {
+
+// this version does following things to avoid scratch memory issue
+// 1. Use StaticallyIndexedArray instead of C array for thread buffer
+// 2. It does not keep reference to tensor descriptor
+// 3. Run() does not construct new tensor coordinate
+template <typename ThreadGroup,
+          typename ElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename Src0Data,
+          typename Src1Data,
+          typename DstData,
+          typename Src0Desc,
+          typename Src1Desc,
+          typename DstDesc,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          bool ThreadTransferSrc0ResetCoordinateAfterRun,
+          bool ThreadTransferSrc1ResetCoordinateAfterRun,
+          bool ThreadTransferDstResetCoordinateAfterRun>
+struct ThreadGroupTensorSliceTransfer_v6r2
+{
+    static constexpr index_t nDim = remove_reference_t<Src0Desc>::GetNumOfDimension();
+
+    static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v6r2(const Src0Desc& src0_desc,
+                                                             const Index& src0_block_slice_origin,
+                                                             const Src1Desc& src1_desc,
+                                                             const Index& src1_block_slice_origin,
+                                                             const DstDesc& dst_desc,
+                                                             const Index& dst_block_slice_origin,
+                                                             const ElementwiseOperation& element_op)
+        : threadwise_transfer_(src0_desc,
+                               make_zero_multi_index<nDim>(),
+                               src1_desc,
+                               make_zero_multi_index<nDim>(),
+                               dst_desc,
+                               make_zero_multi_index<nDim>(),
+                               element_op)
+
+    {
+        static_assert(nDim == remove_cvref_t<Src0Desc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<Src1Desc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
+                          nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == DimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrc0SliceOrigin(
+                src0_desc, src0_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetSrc1SliceOrigin(
+                src1_desc, src1_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    template <typename Src0Buffer, typename Src1Buffer, typename DstBuffer>
+    __device__ void Run(const Src0Desc& src0_desc,
+                        const Src0Buffer& src0_buf,
+                        const Src1Desc& src1_desc,
+                        const Src1Buffer& src1_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.Run(src0_desc, src0_buf, src1_desc, src1_buf, dst_desc, dst_buf);
+        }
+    }
+
+    __device__ void MoveSrc0SliceWindow(const Src0Desc& src0_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrc0SliceWindow(src0_desc, step);
+        }
+    }
+
+    __device__ void MoveSrc1SliceWindow(const Src1Desc& src1_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrc1SliceWindow(src1_desc, step);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v6r2<Src0Data,
+                                           Src1Data,
+                                           DstData,
+                                           Src0Desc,
+                                           Src1Desc,
+                                           DstDesc,
+                                           ElementwiseOperation,
+                                           decltype(thread_slice_lengths),
+                                           DimAccessOrder,
+                                           VectorDim,
+                                           ScalarPerVector,
+                                           DstInMemOp,
+                                           ThreadTransferSrc0ResetCoordinateAfterRun,
+                                           ThreadTransferSrc1ResetCoordinateAfterRun,
+                                           ThreadTransferDstResetCoordinateAfterRun>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp
new file mode 100644
index 0000000..eb5f589
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp
@@ -0,0 +1,183 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp"
+
+namespace ck {
+
+// this version does following things to avoid scratch memory issue
+// 1. Use StaticallyIndexedArray instead of C array for thread buffer
+// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
+// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
+template <typename ThreadGroup,
+          typename ElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename Src0Data,
+          typename Src1Data,
+          typename Src2Data,
+          typename DstData,
+          typename Src0Desc,
+          typename Src1Desc,
+          typename Src2Desc,
+          typename DstDesc,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          bool ThreadTransferSrc0ResetCoordinateAfterRun,
+          bool ThreadTransferSrc1ResetCoordinateAfterRun,
+          bool ThreadTransferSrc2ResetCoordinateAfterRun,
+          bool ThreadTransferDstResetCoordinateAfterRun>
+struct ThreadGroupTensorSliceTransfer_v6r3
+{
+    static constexpr index_t nDim = remove_reference_t<Src0Desc>::GetNumOfDimension();
+
+    static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v6r3(const Src0Desc& src0_desc,
+                                                             const Index& src0_block_slice_origin,
+                                                             const Src1Desc& src1_desc,
+                                                             const Index& src1_block_slice_origin,
+                                                             const Src2Desc& src2_desc,
+                                                             const Index& src2_block_slice_origin,
+                                                             const DstDesc& dst_desc,
+                                                             const Index& dst_block_slice_origin,
+                                                             const ElementwiseOperation& element_op)
+        : threadwise_transfer_(src0_desc,
+                               make_zero_multi_index<nDim>(),
+                               src1_desc,
+                               make_zero_multi_index<nDim>(),
+                               src2_desc,
+                               make_zero_multi_index<nDim>(),
+                               dst_desc,
+                               make_zero_multi_index<nDim>(),
+                               element_op)
+
+    {
+        static_assert(nDim == remove_cvref_t<Src0Desc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<Src1Desc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<Src2Desc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
+                          nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == DimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(get_thread_local_1d_id()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrc0SliceOrigin(
+                src0_desc, src0_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetSrc1SliceOrigin(
+                src1_desc, src1_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetSrc2SliceOrigin(
+                src2_desc, src2_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    template <typename Src0Buffer, typename Src1Buffer, typename Src2Buffer, typename DstBuffer>
+    __device__ void Run(const Src0Desc& src0_desc,
+                        const Src0Buffer& src0_buf,
+                        const Src1Desc& src1_desc,
+                        const Src1Buffer& src1_buf,
+                        const Src2Desc& src2_desc,
+                        const Src2Buffer& src2_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.Run(
+                src0_desc, src0_buf, src1_desc, src1_buf, src2_desc, src2_buf, dst_desc, dst_buf);
+        }
+    }
+
+    __device__ void MoveSrc0SliceWindow(const Src0Desc& src0_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrc0SliceWindow(src0_desc, step);
+        }
+    }
+
+    __device__ void MoveSrc1SliceWindow(const Src1Desc& src1_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrc1SliceWindow(src1_desc, step);
+        }
+    }
+
+    __device__ void MoveSrc2SliceWindow(const Src2Desc& src2_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrc2SliceWindow(src2_desc, step);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v6r3<Src0Data,
+                                           Src1Data,
+                                           Src2Data,
+                                           DstData,
+                                           Src0Desc,
+                                           Src1Desc,
+                                           Src2Desc,
+                                           DstDesc,
+                                           ElementwiseOperation,
+                                           decltype(thread_slice_lengths),
+                                           DimAccessOrder,
+                                           VectorDim,
+                                           ScalarPerVector,
+                                           DstInMemOp,
+                                           ThreadTransferSrc0ResetCoordinateAfterRun,
+                                           ThreadTransferSrc1ResetCoordinateAfterRun,
+                                           ThreadTransferSrc2ResetCoordinateAfterRun,
+                                           ThreadTransferDstResetCoordinateAfterRun>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp
new file mode 100644
index 0000000..3bd7806
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp
@@ -0,0 +1,172 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp"
+
+namespace ck {
+
+// Thread-group level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//
+// Does following things to avoid scratch memory issue
+//   1. Pass tensor descritpors by reference (or tuple of references)
+//   2. Does not keep reference to tensor descriptor
+//   3. Does not construct new tensor coordinate when call Run()
+template <typename ThreadGroup,
+          typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          typename ThreadTransferSrcResetCoordinateAfterRunFlags,
+          typename ThreadTransferDstResetCoordinateAfterRunFlags>
+struct ThreadGroupTensorSliceTransfer_v7
+{
+    static constexpr index_t nDim =
+        remove_cvref_t<tuple_element_t<0, SrcDescs>>::GetNumOfDimension();
+
+    static constexpr index_t nSrc = remove_cvref_t<SrcDescs>::Size();
+    static constexpr index_t nDst = remove_cvref_t<DstDescs>::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v7(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_block_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_block_slice_origins,
+        const ElementwiseOperation& element_op)
+        : threadwise_transfer_(src_descs,
+                               StaticallyIndexedArray<Index, nSrc>{},
+                               dst_descs,
+                               StaticallyIndexedArray<Index, nDst>{},
+                               element_op)
+    {
+        static_assert(nSrc == SrcDatas::Size() && nSrc == SrcDescs::Size() &&
+                          nSrc == ThreadTransferSrcResetCoordinateAfterRunFlags::Size() &&
+                          nDst == DstDatas::Size() && nDst == DstDescs::Size() &&
+                          nDst == ThreadTransferDstResetCoordinateAfterRunFlags::Size(),
+                      "wrong!");
+
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            static_assert(
+                nDim == remove_cvref_t<tuple_element_t<i.value, SrcDescs>>::GetNumOfDimension(),
+                "wrong!");
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            static_assert(
+                nDim == remove_cvref_t<tuple_element_t<i.value, DstDescs>>::GetNumOfDimension(),
+                "wrong!");
+        });
+
+        static_assert(nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == DimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(get_thread_local_1d_id()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            const auto src_thread_slice_origins = generate_tuple(
+                [&](auto i) { return src_block_slice_origins[i] + thread_data_idx_begin; },
+                Number<nSrc>{});
+
+            const auto dst_thread_slice_origins = generate_tuple(
+                [&](auto i) { return dst_block_slice_origins[i] + thread_data_idx_begin; },
+                Number<nDst>{});
+
+            threadwise_transfer_.SetSrcSliceOrigins(src_descs, src_thread_slice_origins);
+            threadwise_transfer_.SetDstSliceOrigins(dst_descs, dst_thread_slice_origins);
+        }
+    }
+
+    template <typename SrcBuffers, typename DstBuffers>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.Run(src_descs, src_bufs, dst_descs, dst_bufs);
+        }
+    }
+
+    template <index_t ISrc>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDescs& src_descs, Number<ISrc> iSrc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(src_descs, iSrc, step);
+        }
+    }
+
+    template <index_t IDst>
+    __device__ void
+    MoveDstSliceWindow(const DstDescs& dst_descs, Number<IDst> iDst, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_descs, iDst, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v7<SrcDatas,
+                                         DstDatas,
+                                         SrcDescs,
+                                         DstDescs,
+                                         ElementwiseOperation,
+                                         DstInMemOps,
+                                         decltype(thread_slice_lengths),
+                                         DimAccessOrder,
+                                         VectorDim,
+                                         ScalarPerVector,
+                                         ThreadTransferSrcResetCoordinateAfterRunFlags,
+                                         ThreadTransferDstResetCoordinateAfterRunFlags>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp
new file mode 100644
index 0000000..a4a29f5
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp
@@ -0,0 +1,30 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct ConvolutionBackwardDataSpecialization
+{
+    Default,
+    Filter1x1Stride1Pad0,
+};
+
+inline std::string
+getConvBackwardDataSpecializationString(const ConvolutionBackwardDataSpecialization& s)
+{
+    switch(s)
+    {
+    case ConvolutionBackwardDataSpecialization::Default: return "Default";
+    case ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0:
+        return "FFilter1x1Stride1Pad0";
+    default: return "Unrecognized specialization!";
+    }
+}
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp
new file mode 100644
index 0000000..20b2a15
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp
@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct ConvolutionBackwardWeightSpecialization
+{
+    Default,
+    Filter1x1Stride1Pad0,
+    Filter1x1Pad0,
+    OddC,
+};
+
+inline std::string
+getConvBackwardWeightSpecializationString(const ConvolutionBackwardWeightSpecialization& s)
+{
+    switch(s)
+    {
+    case ConvolutionBackwardWeightSpecialization::Default: return "Default";
+    case ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0:
+        return "Filter1x1Stride1Pad0";
+    case ConvolutionBackwardWeightSpecialization::Filter1x1Pad0: return "Filter1x1Pad0";
+    case ConvolutionBackwardWeightSpecialization::OddC: return "OddC";
+    default: return "Unrecognized specialization!";
+    }
+}
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp
new file mode 100644
index 0000000..953ff1e
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <string>
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct ConvolutionForwardSpecialization
+{
+    Default,
+    Filter1x1Pad0,
+    Filter1x1Stride1Pad0,
+    OddC,
+};
+
+inline std::string getConvForwardSpecializationString(const ConvolutionForwardSpecialization& s)
+{
+    switch(s)
+    {
+    case ConvolutionForwardSpecialization::Default: return "Default";
+    case ConvolutionForwardSpecialization::Filter1x1Pad0: return "Filter1x1Pad0";
+    case ConvolutionForwardSpecialization::Filter1x1Stride1Pad0: return "Filter1x1Stride1Pad0";
+    case ConvolutionForwardSpecialization::OddC: return "OddC";
+    default: return "Unrecognized specialization!";
+    }
+}
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_base.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_base.hpp
new file mode 100644
index 0000000..8137098
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_base.hpp
@@ -0,0 +1,74 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cmath>
+#include <string>
+#include <sstream>
+
+#include "ck/stream_config.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+struct BaseArgument
+{
+    BaseArgument()                    = default;
+    BaseArgument(const BaseArgument&) = default;
+    BaseArgument& operator=(const BaseArgument&) = default;
+
+    virtual ~BaseArgument() {}
+
+    void* p_workspace_ = nullptr;
+};
+
+struct BaseInvoker
+{
+    BaseInvoker()                   = default;
+    BaseInvoker(const BaseInvoker&) = default;
+    BaseInvoker& operator=(const BaseInvoker&) = default;
+
+    virtual float Run(const BaseArgument*, const StreamConfig& = StreamConfig{})
+    {
+        return float{0};
+    }
+
+    virtual ~BaseInvoker() {}
+};
+
+struct BaseOperator
+{
+    BaseOperator()                    = default;
+    BaseOperator(const BaseOperator&) = default;
+    BaseOperator& operator=(const BaseOperator&) = default;
+
+    virtual bool IsSupportedArgument(const BaseArgument*) { return false; }
+    virtual std::string GetTypeString() const { return ""; }
+
+    virtual std::string GetTypeIdName() const { return typeid(*this).name(); }
+
+    virtual std::string GetTypeIdHashCode() const
+    {
+        std::ostringstream oss;
+
+        oss << std::hex << typeid(*this).hash_code();
+
+        return oss.str();
+    };
+
+    virtual size_t GetWorkSpaceSize(const BaseArgument*) const { return 0; }
+
+    virtual void SetWorkSpacePointer(BaseArgument* p_arg, void* p_workspace) const
+    {
+        assert(p_arg);
+        p_arg->p_workspace_ = p_workspace;
+    }
+
+    virtual ~BaseOperator() {}
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp
new file mode 100644
index 0000000..9fcd893
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp
@@ -0,0 +1,64 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceBatchedContractionMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_gs_ms_ns_lengths,
+                        const std::vector<index_t>& a_gs_ms_ks_strides,
+                        const std::vector<index_t>& b_gs_ns_ks_lengths,
+                        const std::vector<index_t>& b_gs_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                        const std::vector<index_t>& e_gs_ms_ns_lengths,
+                        const std::vector<index_t>& e_gs_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp
new file mode 100644
index 0000000..e755913
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceBatchedGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        ck::index_t BatchStrideA,
+                        ck::index_t BatchStrideB,
+                        ck::index_t BatchStrideC,
+                        ck::index_t Batch,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+using DeviceBatchedGemmPtr = std::unique_ptr<DeviceBatchedGemm<ALayout,
+                                                               BLayout,
+                                                               CLayout,
+                                                               ADataType,
+                                                               BDataType,
+                                                               CDataType,
+                                                               AElementwiseOperation,
+                                                               BElementwiseOperation,
+                                                               CElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp
new file mode 100644
index 0000000..acd779b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp
@@ -0,0 +1,50 @@
+#pragma once
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+struct BatchedGemmEPermuteDesc
+{
+    ck::index_t G0_, G1_, M_, N_;
+    ck::index_t stride_G0_, stride_G1_, stride_M_, stride_N_;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceBatchedGemmEPermute : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_e,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t stride_A,
+                        index_t stride_B,
+                        index_t batch_stride_A,
+                        index_t batch_stride_B,
+                        BatchedGemmEPermuteDesc batched_gemm_e_permute_desc,
+                        index_t BatchCount,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp
new file mode 100644
index 0000000..af68112
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename Acc0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceBatchedGemmGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b0,
+                        const void* p_b1,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t O,
+                        ck::index_t Batch,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB0,
+                        ck::index_t StrideB1,
+                        ck::index_t StrideC,
+                        ck::index_t BatchStrideA,
+                        ck::index_t BatchStrideB0,
+                        ck::index_t BatchStrideB1,
+                        ck::index_t BatchStrideC,
+                        AElementwiseOperation a_element_op,
+                        B0ElementwiseOperation b0_element_op,
+                        Acc0ElementwiseOperation acc0_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp
new file mode 100644
index 0000000..116e62c
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceBatchedGemmMultiD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static_assert(DsLayout::Size() == DsDataType::Size(), "wrong! inconsisiten NumDTensor");
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const std::array<const void*, NumDTensor>& p_ds,
+                        void* p_e,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t Batch,
+                        index_t StrideA,
+                        index_t StrideB,
+                        const std::array<ck::index_t, NumDTensor>& StrideDs,
+                        index_t StrideE,
+                        index_t BatchStrideA,
+                        index_t BatchStrideB,
+                        const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                        index_t BatchStrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp
new file mode 100644
index 0000000..eacc597
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp
@@ -0,0 +1,72 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename A0Layout,
+          typename B0Layout,
+          typename D0sLayout,
+          typename B1Layout,
+          typename D1sLayout,
+          typename E1Layout,
+          typename A0DataType,
+          typename B0DataType,
+          typename D0sDataType,
+          typename B1DataType,
+          typename D1sDataType,
+          typename E1DataType,
+          typename A0ElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename CDE0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CDE1ElementwiseOperation>
+struct DeviceBatchedGemmMultipleDGemmMultipleD : public BaseOperator
+{
+    static constexpr index_t NumD0Tensor = D0sDataType::Size();
+    static constexpr index_t NumD1Tensor = D1sDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a0,
+                        const void* p_b0,
+                        std::array<const void*, NumD0Tensor> p_d0s,
+                        const void* p_b1,
+                        std::array<const void*, NumD1Tensor> p_d1s,
+                        void* p_e1,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t O,
+                        ck::index_t Batch,
+                        ck::index_t StrideA0,
+                        ck::index_t StrideB0,
+                        std::array<ck::index_t, NumD0Tensor> StrideD0s,
+                        ck::index_t StrideB1,
+                        std::array<ck::index_t, NumD1Tensor> StrideD1s,
+                        ck::index_t StrideE1,
+                        ck::index_t BatchStrideA0,
+                        ck::index_t BatchStrideB0,
+                        std::array<ck::index_t, NumD0Tensor> BatchStrideD0s,
+                        ck::index_t BatchStrideB1,
+                        std::array<ck::index_t, NumD1Tensor> BatchStrideD1s,
+                        ck::index_t BatchStrideE1,
+                        A0ElementwiseOperation a0_element_op,
+                        B0ElementwiseOperation b0_element_op,
+                        CDE0ElementwiseOperation cde0_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CDE1ElementwiseOperation cde1_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
new file mode 100644
index 0000000..c1f85e5
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename Acc0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          bool MaskOutUpperTriangle> // TODO: enum for mask type
+struct DeviceBatchedGemmSoftmaxGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b0,
+                        const void* p_b1,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t O,
+                        ck::index_t Batch,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB0,
+                        ck::index_t StrideB1,
+                        ck::index_t StrideC,
+                        ck::index_t BatchStrideA,
+                        ck::index_t BatchStrideB0,
+                        ck::index_t BatchStrideB1,
+                        ck::index_t BatchStrideC,
+                        AElementwiseOperation a_element_op,
+                        B0ElementwiseOperation b0_element_op,
+                        Acc0ElementwiseOperation acc0_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp
new file mode 100644
index 0000000..ff55519
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp
@@ -0,0 +1,70 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasDataType,
+          typename Acc1BiasDataType,
+          typename AElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename Acc0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          MaskingSpecialization MaskingSpec>
+struct DeviceBatchedGemmSoftmaxGemmPermute : public BaseOperator
+{
+    static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
+    static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b0,
+        const void* p_b1,
+        void* p_c,
+        const std::array<void*, NumAcc0Bias> p_acc0_biases,
+        const std::array<void*, NumAcc1Bias> p_acc1_biases,
+        const std::vector<index_t>& a_gs_ms_ks_lengths,
+        const std::vector<index_t>& a_gs_ms_ks_strides,
+        const std::vector<index_t>& b_gs_ns_ks_lengths,
+        const std::vector<index_t>& b_gs_ns_ks_strides,
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+        const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+        const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+        const std::array<std::vector<index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+        const std::array<std::vector<index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+        AElementwiseOperation a_element_op,
+        B0ElementwiseOperation b0_element_op,
+        Acc0ElementwiseOperation acc0_element_op,
+        B1ElementwiseOperation b1_element_op,
+        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp
new file mode 100644
index 0000000..d39f3b7
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct DeviceBatchNormBwd : public BaseOperator
+{
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> xyLengths,
+                        const std::array<index_t, Rank> xStrides,
+                        const std::array<index_t, Rank> dyStrides,
+                        const std::array<index_t, Rank> dxStrides,
+                        const std::array<int, NumBatchNormReduceDim> reduceDims,
+                        const std::array<ck::index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                        const std::array<ck::index_t, NumInvariantDim> bnScaleStrides,
+                        const std::array<ck::index_t, NumInvariantDim> bnDscaleDbiasStrides,
+                        const std::array<ck::index_t, NumInvariantDim> bnMeanVarStrides,
+                        const void* p_x,
+                        const void* p_dy,
+                        const void* p_scale,
+                        const void* p_savedMean,
+                        const void* p_savedInvVar,
+                        double epsilon,
+                        const DyElementwiseOp dy_elementwise_op,
+                        void* p_dx,
+                        void* p_dscale,
+                        void* p_dbias) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+using DeviceBatchNormBwdPtr = std::unique_ptr<DeviceBatchNormBwd<XDataType,
+                                                                 DxDataType,
+                                                                 DyDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 DscaleDbiasDataType,
+                                                                 MeanVarDataType,
+                                                                 DyElementwiseOp,
+                                                                 Rank,
+                                                                 NumBatchNormReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp
new file mode 100644
index 0000000..aa93dd9
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp
@@ -0,0 +1,72 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct DeviceBatchNormFwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, Rank> xyLengths,
+        const std::array<index_t, Rank> xStrides,
+        const std::array<index_t, Rank> yStrides,
+        const std::array<int, NumBatchNormReduceDim> reduceDims,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+        const void* p_x,
+        const void* bnScale,
+        const void* bnBias,
+        double epsilon,
+        const YElementwiseOp y_elementwise_op,
+        void* p_y,
+        void* resultSaveMean,
+        void* resultSaveInvVariance,
+        double exponentialAverageFactor,
+        void* resultRunningMean,
+        void* resultRunningVariance) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+using DeviceBatchNormFwdPtr = std::unique_ptr<DeviceBatchNormFwd<XDataType,
+                                                                 YDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 BiasDataType,
+                                                                 MeanVarDataType,
+                                                                 YElementwiseOp,
+                                                                 Rank,
+                                                                 NumBatchNormReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp
new file mode 100644
index 0000000..8a00fd9
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp
@@ -0,0 +1,69 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct DeviceBatchNormInfer : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, Rank> xyLengths,
+        const std::array<index_t, Rank> xStrides,
+        const std::array<index_t, Rank> yStrides,
+        const std::array<int, NumBatchNormReduceDim> reduceDims,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+        const void* p_x,
+        const void* bnScale,
+        const void* bnBias,
+        double epsilon,
+        const YElementwiseOp y_elementwise_op,
+        const void* estimatedMean,
+        const void* estimatedInvVariance,
+        void* p_y) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+using DeviceBatchNormInferPtr = std::unique_ptr<DeviceBatchNormInfer<XDataType,
+                                                                     YDataType,
+                                                                     AccDataType,
+                                                                     ScaleDataType,
+                                                                     BiasDataType,
+                                                                     MeanVarDataType,
+                                                                     YElementwiseOp,
+                                                                     Rank,
+                                                                     NumBatchNormReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_cgemm.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_cgemm.hpp
new file mode 100644
index 0000000..aedae53
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_cgemm.hpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceCGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a_real,
+                                                              const void* p_a_imag,
+                                                              const void* p_b_real,
+                                                              const void* p_b_imag,
+                                                              void* p_c_real,
+                                                              void* p_c_imag,
+                                                              void* p_workspace,
+                                                              ck::index_t M,
+                                                              ck::index_t N,
+                                                              ck::index_t K,
+                                                              ck::index_t StrideA,
+                                                              ck::index_t StrideB,
+                                                              ck::index_t StrideC,
+                                                              AElementwiseOperation a_element_op,
+                                                              BElementwiseOperation b_element_op,
+                                                              CElementwiseOperation c_element_op,
+                                                              ck::index_t KBatch = 1) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+    virtual std::size_t GetWorkspaceSize(index_t MRaw,
+                                         index_t NRaw,
+                                         index_t KRaw,
+                                         index_t StrideA,
+                                         index_t StrideB,
+                                         index_t StrideC)     = 0;
+};
+
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+using DeviceCGemmPtr = std::unique_ptr<
+    DeviceCGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp
new file mode 100644
index 0000000..dbc525c
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp
@@ -0,0 +1,63 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceContractionMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_ms_ns_lengths,
+                        const std::vector<index_t>& a_ms_ks_strides,
+                        const std::vector<index_t>& b_ns_ks_lengths,
+                        const std::vector<index_t>& b_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides,
+                        const std::vector<index_t>& e_ms_ns_lengths,
+                        const std::vector<index_t>& e_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp
new file mode 100644
index 0000000..82054a3
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceConvBwdData : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* p_in,
+                        const void* p_wei,
+                        const void* p_out,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp
new file mode 100644
index 0000000..4b98810
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceConvFwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        const void* p_wei,
+                        void* p_out,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp
new file mode 100644
index 0000000..5a627de
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <iostream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceConvFwdBiasActivation : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        const void* p_wei,
+                        void* p_out,
+                        const void* p_bias,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+using DeviceConvFwdBiasActivationPtr =
+    std::unique_ptr<DeviceConvFwdBiasActivation<InElementwiseOperation,
+                                                WeiElementwiseOperation,
+                                                OutElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp
new file mode 100644
index 0000000..cc13930
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <iostream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceConvFwdBiasActivationAdd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        const void* p_wei,
+                        void* p_out,
+                        const void* p_bias,
+                        const void* p_resi,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+using DeviceConvFwdBiasActivationAddPtr =
+    std::unique_ptr<DeviceConvFwdBiasActivationAdd<InElementwiseOperation,
+                                                   WeiElementwiseOperation,
+                                                   OutElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_2d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_2d.hpp
new file mode 100644
index 0000000..23aada0
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_2d.hpp
@@ -0,0 +1,341 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise_base.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+template <typename InDataTypeTuple,
+          typename OutDataTypeTuple,
+          typename ElementwiseOperation,
+          index_t NumDim_m,
+          index_t NumDim_n,
+          index_t MPerThread,
+          index_t NPerThread,
+          typename InScalarPerVectorSeq,
+          typename OutScalarPerVectorSeq>
+struct DeviceElementwise : public DeviceElementwiseBase<InDataTypeTuple,
+                                                        OutDataTypeTuple,
+                                                        ElementwiseOperation,
+                                                        NumDim_m + NumDim_n>
+{
+    static constexpr index_t NumDim = NumDim_m + NumDim_n;
+
+    static constexpr int NumInput  = InDataTypeTuple::Size();
+    static constexpr int NumOutput = OutDataTypeTuple::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static_assert(NumInput == InScalarPerVectorSeq::Size() &&
+                      NumOutput == OutScalarPerVectorSeq::Size(),
+                  "Tuple size is inconsistent with the number of in/out!");
+
+    static auto GenerateInDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+
+                return static_cast<const DataType*>(nullptr);
+            },
+            Number<NumInput>{});
+    };
+
+    static auto GenerateOutDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+
+                return static_cast<DataType*>(nullptr);
+            },
+            Number<NumOutput>{});
+    };
+
+    using InDataTypePointerTuple  = decltype(GenerateInDataTypePointerTuple());
+    using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
+
+    template <typename Desc_MN>
+    static auto PadDescriptor_MN_2d(Desc_MN desc_mn,
+                                    index_t gridSize,
+                                    index_t blockSize,
+                                    index_t num_threads_m,
+                                    index_t num_threads_n)
+    {
+        std::ignore               = blockSize;
+        std::ignore               = gridSize;
+        const auto m              = desc_mn.GetLength(I0);
+        const auto n              = desc_mn.GetLength(I1);
+        const index_t loop_step_m = num_threads_m * MPerThread;
+        const index_t loop_step_n = num_threads_n * NPerThread;
+        const auto pad_m          = math::integer_least_multiple(m, loop_step_m) - m;
+        const auto pad_n          = math::integer_least_multiple(n, loop_step_n) - n;
+
+        const auto desc_mn_pad = transform_tensor_descriptor(
+            desc_mn,
+            make_tuple(make_right_pad_transform(m, pad_m), make_right_pad_transform(n, pad_n)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+        return desc_mn_pad;
+    }
+
+    static auto MakeDescriptor_MN(const std::array<index_t, NumDim>& lengths,
+                                  const std::array<index_t, NumDim>& stride,
+                                  index_t gridSize,
+                                  index_t blockSize,
+                                  index_t num_threads_m,
+                                  index_t num_threads_n)
+    {
+        auto tupleOfShape  = generate_tuple([&](auto I) { return lengths[I]; }, Number<NumDim>{});
+        auto tupleOfStride = generate_tuple([&](auto I) { return stride[I]; }, Number<NumDim>{});
+
+        // nd desc - [s0, s1, s2, ...]
+        const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
+
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDim_m, 1>::type();
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDim_m, NumDim_m + NumDim_n, 1>::type();
+
+        const auto mLengths = get_container_subset(tupleOfShape, mDimIds);
+        const auto nLengths = get_container_subset(tupleOfShape, nDimIds);
+
+        // merge nd to 2d desc - [s0 * s1 * ...]
+
+        if constexpr(NumDim > 2)
+        {
+            const auto desc_mn = transform_tensor_descriptor(
+                desc,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+                make_tuple(mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return PadDescriptor_MN_2d(desc_mn, gridSize, blockSize, num_threads_m, num_threads_n);
+        }
+        else
+            return PadDescriptor_MN_2d(desc, gridSize, blockSize, num_threads_m, num_threads_n);
+    }
+
+    template <index_t TupleSize>
+    static auto GenerateInOutGrid2dDescTuple(Number<TupleSize>)
+    {
+        return generate_tuple(
+            [&](auto) {
+                if constexpr(NumDim > 2)
+                {
+                    return MakeDescriptor_MN({1, 1}, {1, 1}, 1, 1, 1, 1);
+                }
+                else
+                {
+                    return MakeDescriptor_MN({1}, {1}, 1, 1, 1, 1);
+                };
+            },
+            Number<TupleSize>{});
+    };
+
+    using OutGrid2dDescTuple = decltype(GenerateInOutGrid2dDescTuple(Number<NumOutput>{}));
+    using InGrid2dDescTuple  = decltype(GenerateInOutGrid2dDescTuple(Number<NumInput>{}));
+
+    using GridwiseElementwise = GridwiseElementwise_2D<InGrid2dDescTuple,
+                                                       OutGrid2dDescTuple,
+                                                       InDataTypePointerTuple,
+                                                       OutDataTypePointerTuple,
+                                                       ElementwiseOperation,
+                                                       MPerThread,
+                                                       NPerThread,
+                                                       InScalarPerVectorSeq,
+                                                       OutScalarPerVectorSeq>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, NumDim> lengths,
+                 const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                 const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                 const std::array<const void*, NumInput> in_dev_buffers,
+                 const std::array<void*, NumOutput> out_dev_buffers,
+                 ElementwiseOperation elementwise_op)
+
+            : lengths_(lengths),
+              inStridesArray_(inStridesArray),
+              outStridesArray_(outStridesArray),
+              elementwise_op_(elementwise_op),
+              blockSize_(256),
+              gridSize_(120), // FIXME - Calculate the grid size by number of CU in the future
+              num_threads_m_((gridSize_ * blockSize_) / 16),
+              num_threads_n_(16)
+        {
+            static_assert(NumDim_m > 0, "");
+            static_assert(NumDim_n > 0, "");
+
+            in_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+                    return static_cast<const DataType*>(in_dev_buffers[I.value]);
+                },
+                Number<NumInput>{});
+
+            out_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+                    return static_cast<DataType*>(out_dev_buffers[I.value]);
+                },
+                Number<NumOutput>{});
+
+            in_grid_2d_desc_tuple_ = generate_tuple(
+                [&](auto I) {
+                    return MakeDescriptor_MN(lengths,
+                                             inStridesArray[I.value],
+                                             gridSize_,
+                                             blockSize_,
+                                             num_threads_m_,
+                                             num_threads_n_);
+                },
+                Number<NumInput>{});
+
+            out_grid_2d_desc_tuple_ = generate_tuple(
+                [&](auto I) {
+                    return MakeDescriptor_MN(lengths,
+                                             outStridesArray[I.value],
+                                             gridSize_,
+                                             blockSize_,
+                                             num_threads_m_,
+                                             num_threads_n_);
+                },
+                Number<NumOutput>{});
+        }
+
+        InDataTypePointerTuple in_dev_buffers_;
+        OutDataTypePointerTuple out_dev_buffers_;
+        InGrid2dDescTuple in_grid_2d_desc_tuple_;
+        OutGrid2dDescTuple out_grid_2d_desc_tuple_;
+
+        std::array<index_t, NumDim> lengths_;
+        std::array<std::array<index_t, NumDim>, NumInput> inStridesArray_;
+        std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray_;
+
+        ElementwiseOperation elementwise_op_;
+        index_t blockSize_;
+        index_t gridSize_;
+        index_t num_threads_m_;
+        index_t num_threads_n_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto kernel = kernel_elementwise_2d<GridwiseElementwise,
+                                                      InGrid2dDescTuple,
+                                                      OutGrid2dDescTuple,
+                                                      InDataTypePointerTuple,
+                                                      OutDataTypePointerTuple,
+                                                      ElementwiseOperation>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(arg.gridSize_),
+                                                        dim3(arg.blockSize_),
+                                                        0,
+                                                        arg.in_grid_2d_desc_tuple_,
+                                                        arg.out_grid_2d_desc_tuple_,
+                                                        arg.in_dev_buffers_,
+                                                        arg.out_dev_buffers_,
+                                                        arg.elementwise_op_,
+                                                        arg.num_threads_m_,
+                                                        arg.num_threads_n_);
+            return elapsed_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if(pArg == nullptr)
+            return false;
+
+        if(pArg->lengths_.back() % MPerThread != 0)
+            return false;
+
+        auto IsScalarPerVectorValid = [&](const std::array<index_t, NumDim>& lengths,
+                                          const std::array<index_t, NumDim>& strides,
+                                          index_t scalarPerVector,
+                                          index_t vectorDim) {
+            if(strides[vectorDim] == 1 &&
+               (lengths[vectorDim] % scalarPerVector == 0 ||
+                lengths[vectorDim] % scalarPerVector == lengths[vectorDim]))
+            {
+                return true;
+            }
+            if(strides[vectorDim] != 1 && scalarPerVector == strides[vectorDim])
+            {
+                return true;
+            }
+            return false;
+        };
+
+        bool valid = true;
+        static_for<0, NumInput, 1>{}([&](auto I) {
+            if(!IsScalarPerVectorValid(pArg->lengths_,
+                                       pArg->inStridesArray_[I.value],
+                                       InScalarPerVectorSeq::At(I),
+                                       NumDim_m - 1))
+                valid = false;
+        });
+
+        static_for<0, NumOutput, 1>{}([&](auto I) {
+            if(!IsScalarPerVectorValid(pArg->lengths_,
+                                       pArg->outStridesArray_[I.value],
+                                       OutScalarPerVectorSeq::At(I),
+                                       NumDim - 1))
+                valid = false;
+        });
+
+        return valid;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, NumDim> lengths,
+                        const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                        const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const std::array<void*, NumOutput> out_dev_buffers,
+                        ElementwiseOperation elementwise_op) override
+    {
+        return std::make_unique<Argument>(lengths,
+                                          inStridesArray,
+                                          outStridesArray,
+                                          in_dev_buffers,
+                                          out_dev_buffers,
+                                          elementwise_op);
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+}; // namespace device
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_base.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_base.hpp
new file mode 100644
index 0000000..728faf5
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_base.hpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+#include <array>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataTypeTuple,
+          typename OutDataTypeTuple,
+          typename ElementwiseOperation,
+          index_t NumDim>
+struct DeviceElementwiseBase : public BaseOperator
+{
+    static constexpr int NumInput  = InDataTypeTuple::Size();
+    static constexpr int NumOutput = OutDataTypeTuple::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, NumDim> lengths,
+                        const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                        const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const std::array<void*, NumOutput> out_dev_buffers,
+                        ElementwiseOperation elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+}; // namespace device
+
+template <typename InDataTypeTuple,
+          typename OutDataTypeTuple,
+          typename ElementwiseOperation,
+          index_t NumDim>
+using DeviceElementwiseBasePtr = std::unique_ptr<
+    DeviceElementwiseBase<InDataTypeTuple, OutDataTypeTuple, ElementwiseOperation, NumDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp
new file mode 100644
index 0000000..42f821f
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataTypeTuple,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          typename XElementwiseOperation,
+          typename YElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceElementwiseNormalization : public BaseOperator
+{
+    static constexpr int NumInput = InDataTypeTuple::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::array<std::vector<index_t>, NumInput> inStridesArray,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        AccDataType epsilon,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        XElementwiseOperation x_elementwise_op,
+                        YElementwiseOperation y_elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename InDataTypeTuple,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          typename XElementwiseOperation,
+          typename YElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+using DeviceElementwiseNormalizationPtr =
+    std::unique_ptr<DeviceElementwiseNormalization<InDataTypeTuple,
+                                                   GammaDataType,
+                                                   BetaDataType,
+                                                   AccDataType,
+                                                   YDataType,
+                                                   XElementwiseOperation,
+                                                   YElementwiseOperation,
+                                                   Rank,
+                                                   NumReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm.hpp
new file mode 100644
index 0000000..c0af6f8
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm.hpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp
new file mode 100644
index 0000000..4c2161e
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+struct DEGridDesc_M0_M1_M2_N0_N1
+{
+    ck::index_t M0_, M1_, M2_, N0_, N1_;
+    ck::index_t stride_M0_, stride_M1_, stride_M2_, stride_N0_, stride_N1_;
+};
+
+// input : A[M, K], B[K, N],
+// input : D[M, N], ...
+// output : E[M, N]
+// C = a_op(A) * b_op(B)
+// E = cde_op(C, D)
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGemmBiasCPermute : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const void* p_d,
+                        void* p_e,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        DEGridDesc_M0_M1_M2_N0_N1 d_gride_desc,
+                        DEGridDesc_M0_M1_M2_N0_N1 e_gride_desc,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
new file mode 100644
index 0000000..9113bb7
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// GEMM:
+//   input : A[M, K], B[K, N],
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGemmMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        ck::index_t StrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp
new file mode 100644
index 0000000..f4881e3
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp
@@ -0,0 +1,97 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// FIXME: DeviceGemmReduce type need to well define the problem
+// GEMM:
+//   input : A[AK0, M, AK1]
+//   input : B[AK0, N, AK1]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   output : R0[M], R1[M], ...
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+//   Q0 = reduce0(q_op0(E)), Q1 = reduce1(q_op0(E)), ...
+//   R0 = r_op0(Q0), R1 = r_op1(Q1), ...
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation>
+struct DeviceGemmMultipleDMultipleR : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    static constexpr index_t NumRTensor = RsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        std::array<void*, NumRTensor> p_rs,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        ck::index_t StrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op,
+                        QsElementwiseOperation qs_element_op,
+                        RsElementwiseOperation rs_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation>
+using DeviceGemmMultipleDMultipleRPtr =
+    std::unique_ptr<DeviceGemmMultipleDMultipleR<ALayout,
+                                                 BLayout,
+                                                 DELayout,
+                                                 ADataType,
+                                                 BDataType,
+                                                 DsDataType,
+                                                 EDataType,
+                                                 RsDataType,
+                                                 AElementwiseOperation,
+                                                 BElementwiseOperation,
+                                                 CDEElementwiseOperation,
+                                                 QsElementwiseOperation,
+                                                 RsElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
new file mode 100644
index 0000000..fcc088c
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// FIXME: DeviceGemmReduce type need to well define the problem
+template <ck::index_t NumDTensor, ck::index_t NumReduce>
+struct DeviceGemmReduce : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const void* p_bias,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_c,
+                        std::array<void*, NumReduce> p_reduces,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        std::array<void*, 3> gemm_element_ops,
+                        std::array<void*, NumDTensor> d_element_ops,
+                        std::array<void*, NumReduce> reduce_in_element_ops,
+                        std::array<void*, NumReduce> reduce_out_element_ops,
+                        ck::index_t BatchCount = 1) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <ck::index_t NumDTensor, ck::index_t NumReduce>
+using DeviceGemmReducePtr = std::unique_ptr<DeviceGemmReduce<NumDTensor, NumReduce>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
new file mode 100644
index 0000000..c701bff
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
@@ -0,0 +1,64 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemmSplitK : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                              const void* p_b,
+                                                              void* p_c,
+                                                              ck::index_t M,
+                                                              ck::index_t N,
+                                                              ck::index_t K,
+                                                              ck::index_t StrideA,
+                                                              ck::index_t StrideB,
+                                                              ck::index_t StrideC,
+                                                              AElementwiseOperation a_element_op,
+                                                              BElementwiseOperation b_element_op,
+                                                              CElementwiseOperation c_element_op,
+                                                              ck::index_t KBatch) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+using DeviceGemmSplitKPtr = std::unique_ptr<DeviceGemmSplitK<ALayout,
+                                                             BLayout,
+                                                             CLayout,
+                                                             ADataType,
+                                                             BDataType,
+                                                             CDataType,
+                                                             AElementwiseOperation,
+                                                             BElementwiseOperation,
+                                                             CElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp
new file mode 100644
index 0000000..173c613
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp
@@ -0,0 +1,72 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumDTensor>
+struct ContractionDesc
+{
+    std::vector<index_t> a_ms_ks_lengths;
+    std::vector<index_t> a_ms_ks_strides;
+
+    std::vector<index_t> b_ns_ks_lengths;
+    std::vector<index_t> b_ns_ks_strides;
+
+    std::array<std::vector<index_t>, NumDTensor> ds_ms_ns_lengths;
+    std::array<std::vector<index_t>, NumDTensor> ds_ms_ns_strides;
+
+    std::vector<index_t> e_ms_ns_lengths;
+    std::vector<index_t> e_ms_ns_strides;
+};
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGroupedContractionMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b_vec,
+                        std::vector<std::array<const void*, NumDTensor>> p_ds_vec,
+                        std::vector<void*> p_e_vec,
+                        std::vector<ContractionDesc<NumDTensor>> contraction_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp
new file mode 100644
index 0000000..3350aec
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Conv backward data multiple D:
+//   input : output image A[G, N, K, Ho, Wo]
+//   input : weight B[G, K, C, Y, X],
+//   input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
+//   output : input image E[G, N, C, Hi, Wi],
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGroupedConvBwdDataMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static_assert(NumDTensor == DsLayout::Size(), "wrong! Inconsistent NumDTensor");
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,                                                 // output image
+        const void* p_b,                                                 // weight
+        const std::array<const void*, NumDTensor>& p_ds,                 // bias
+        void* p_e,                                                       // input image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_lengths, // output image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_strides, // output image
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,  // weight
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,  // weight
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+            ds_g_n_k_wos_lengths, // bias
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+            ds_g_n_k_wos_strides,                                        // bias
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths, // input image
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_strides, // input image
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp
new file mode 100644
index 0000000..1258aed
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceGroupedConvBwdWeight : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        void* p_wei,
+                        const void* p_out,
+                        ck::index_t G,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+                        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+                        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+                        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+                        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+                        std::array<ck::index_t, NDimSpatial> input_left_pads,
+                        std::array<ck::index_t, NDimSpatial> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op,
+                        ck::index_t split_k) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp
new file mode 100644
index 0000000..644c7ee
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp
@@ -0,0 +1,55 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Convolution Forward:
+//   input : input image A[G, N, C, Hi, Wi],
+//   input : weight B[G, K, C, Y, X],
+//   output : output image E[G, N, K, Ho, Wo]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+template <index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceGroupedConvFwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,  // input image
+                        const void* p_wei, // weight
+                        void* p_out,       // output image
+                        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads,
+                        const InElementwiseOperation& in_element_op,
+                        const WeiElementwiseOperation& wei_element_op,
+                        const OutElementwiseOperation& out_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..3bb085c
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,959 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+template <index_t NumDTensor>
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_K0_M0_M1_K1,
+          typename BGridDesc_K0_N0_N1_K1,
+          typename DsGridDesc_M0_M10_M11_N0_N10_N11,
+          typename CGridDesc_M0_M10_M11_N0_N10_N11,
+          typename Block2CTileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_fwd_dl_multiple_d(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const index_t batch_count,
+            const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1,
+            const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1,
+            const DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11,
+            const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11,
+            const Block2CTileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__)||defined(__gfx926__)  || defined(__gfx1030__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
+
+    __shared__ ABDataType p_shared[shared_block_size];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor = DsGridDesc_M0_M10_M11_N0_N10_N11::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::Run(p_a_grid + a_batch_offset,
+                      p_b_grid + b_batch_offset,
+                      p_ds_grid_grp,
+                      p_e_grid + c_batch_offset,
+                      p_shared,
+                      a_element_op,
+                      b_element_op,
+                      cde_element_op,
+                      a_grid_desc_k0_m0_m1_k1,
+                      b_grid_desc_k0_n0_n1_k1,
+                      ds_grid_desc_m0_m10_m11_n0_n10_n11,
+                      e_grid_desc_m0_m10_m11_n0_n10_n11,
+                      block_2_ctile_map,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m0_m1_k1;
+    ignore = b_grid_desc_k0_n0_n1_k1;
+    ignore = ds_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = e_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+
+    compute_ptr_offset_of_batch.GetAPtrOffset(0);
+    compute_ptr_offset_of_batch.GetBPtrOffset(0);
+    compute_ptr_offset_of_batch.GetEPtrOffset(0);
+#endif
+}
+} // namespace
+
+//
+// @brief      Device Convolution operation.
+//
+// Supports:
+//  @li         Forward convolution with up to 3 spatial dimentions
+//  @li         Input tensor in GNWC data format
+//  @li         Weight tensor in GKXC data format
+//  @li         Output tensor in GNWK data format
+//
+// 1D:
+// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+// 2D:
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+// 3D:
+// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+//
+template <index_t NDimSpatial,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t K1,
+          index_t M1PerThread,
+          index_t N1PerThread,
+          index_t KPerThread,
+          typename M1N1ThreadClusterM1Xs,
+          typename M1N1ThreadClusterN1Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
+    : public DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                           ALayout,
+                                           BLayout,
+                                           DsLayout,
+                                           ELayout,
+                                           ADataType,
+                                           BDataType,
+                                           DsDataType,
+                                           EDataType,
+                                           AElementwiseOperation,
+                                           BElementwiseOperation,
+                                           CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_AK0_M_AK1(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                                  const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                                  const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                                  const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                                  const std::array<index_t, NDimSpatial>& input_left_pads,
+                                  const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        const auto M = in_gemmm_gemmk_desc.GetLength(I0);
+        const auto K = in_gemmm_gemmk_desc.GetLength(I1);
+
+        const auto AK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            in_gemmm_gemmk_desc,
+            make_tuple(make_unmerge_transform(make_tuple(AK0, K1)), make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_BK0_N_BK1(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        const auto N = wei_gemmn_gemmk_desc.GetLength(I0);
+        const auto K = wei_gemmn_gemmk_desc.GetLength(I1);
+
+        const auto BK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            wei_gemmn_gemmk_desc,
+            make_tuple(make_unmerge_transform(make_tuple(BK0, K1)), make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    template <typename ELay>
+    static auto
+    MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(ds_g_n_k_wos_lengths[i],
+                                                                  ds_g_n_k_wos_strides[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // desc for problem definition
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
+        MakeAGridDescriptor_AK0_M_AK1<ALayout>({}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeBGridDescriptor_BK0_N_BK1<BLayout>({}, {}))>;
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
+    using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDlMultipleD_km_kn_mn<BlockSize,
+                                         ADataType,
+                                         AccDataType,
+                                         DsDataType,
+                                         EDataType,
+                                         AElementwiseOperation,
+                                         BElementwiseOperation,
+                                         CDEElementwiseOperation,
+                                         InMemoryDataOperationEnum::Set,
+                                         AGridDesc_AK0_M_AK1,
+                                         BGridDesc_BK0_N_BK1,
+                                         EGridDesc_M_N,
+                                         MPerBlock,
+                                         NPerBlock,
+                                         K0PerBlock,
+                                         K1,
+                                         M1PerThread,
+                                         N1PerThread,
+                                         KPerThread,
+                                         M1N1ThreadClusterM1Xs,
+                                         M1N1ThreadClusterN1Xs,
+                                         ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                         ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                         ABlockTransferThreadClusterArrangeOrder,
+                                         ABlockTransferSrcAccessOrder,
+                                         ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                         ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                         ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                         BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                         BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                         BBlockTransferThreadClusterArrangeOrder,
+                                         BBlockTransferSrcAccessOrder,
+                                         BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                         BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                         BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                         CThreadTransferSrcDstAccessOrder,
+                                         CThreadTransferSrcDstVectorDim,
+                                         CThreadTransferDstScalarPerVector>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_AK0_M_AK1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_BK0_N_BK1{}));
+    using DsGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(DsGridDesc_M_N{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(EGridDesc_M_N{}));
+    using DefaultBlock2CTileMap =
+        decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(EGridDesc_M_N{}));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,
+                 const void* p_b,
+                 const std::array<const void*, NumDTensor>& p_ds,
+                 void* p_e,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CDEElementwiseOperation& cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              num_group_{a_g_n_c_wis_lengths[0]},
+              a_grid_desc_ak0_m_ak1_{
+                  DeviceOp::MakeAGridDescriptor_AK0_M_AK1<ALayout>(a_g_n_c_wis_lengths,
+                                                                   a_g_n_c_wis_strides,
+                                                                   b_g_k_c_xs_lengths,
+                                                                   b_g_k_c_xs_strides,
+                                                                   e_g_n_k_wos_lengths,
+                                                                   e_g_n_k_wos_strides,
+                                                                   conv_filter_strides,
+                                                                   conv_filter_dilations,
+                                                                   input_left_pads,
+                                                                   input_right_pads)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1<BLayout>(
+                  b_g_k_c_xs_lengths, b_g_k_c_xs_strides)},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides)},
+              a_grid_desc_k0_m0_m1_k1_{},
+              b_grid_desc_k0_n0_n1_k1_{},
+              ds_grid_desc_m0_m10_m11_n0_n10_n11_{},
+              e_grid_desc_m0_m10_m11_n0_n10_n11_{},
+              block_2_ctile_map_{},
+              compute_ptr_offset_of_batch_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
+              e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
+
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+
+                // D batch stride
+                compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
+
+                // D desc
+                ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
+                    ds_g_n_k_wos_lengths[i], ds_g_n_k_wos_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_ak0_m_ak1_, b_grid_desc_bk0_n_bk1_, e_grid_desc_m_n_))
+            {
+
+                a_grid_desc_k0_m0_m1_k1_ =
+                    GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_ak0_m_ak1_);
+                b_grid_desc_k0_n0_n1_k1_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_bk0_n_bk1_);
+                e_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(e_grid_desc_m_n_);
+
+                ds_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(ds_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[K0, M, K1]: " << a_grid_desc_ak0_m_ak1_ << std::endl;
+            std::cout << "B[K0, N, K1]: " << b_grid_desc_bk0_n_bk1_ << std::endl;
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+            std::cout << "num_group: " << num_group_ << std::endl;
+
+            std::cout << "A[k0, m0, m1, k1]: " << a_grid_desc_k0_m0_m1_k1_ << std::endl;
+            std::cout << "B[k0, n0, n1, k1]: " << b_grid_desc_k0_n0_n1_k1_ << std::endl;
+            std::cout << "A[m0, m10, m11, n0, n10, n11]: " << e_grid_desc_m0_m10_m11_n0_n10_n11_
+                      << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        index_t num_group_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_;
+        BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_;
+        DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11_;
+        CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        // block-to-e-tile map
+        DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config)
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(
+                   arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.e_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK has invalid setting");
+            }
+
+            const index_t grid_size =
+                GridwiseGemm::CalculateGridSize(arg.e_grid_desc_m_n_.GetLength(I0),
+                                                arg.e_grid_desc_m_n_.GetLength(I1)) *
+                arg.num_group_;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop,
+                                     auto has_double_tail_k_block_loop) {
+                constexpr bool has_main_loop   = has_main_k_block_loop.value;
+                constexpr bool has_double_loop = has_double_tail_k_block_loop;
+
+                const auto kernel = kernel_grouped_conv_fwd_dl_multiple_d<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_K0_M0_M1_K1,
+                    DeviceOp::BGridDesc_K0_N0_N1_K1,
+                    DeviceOp::DsGridDesc_M0_M10_M11_N0_N10_N11,
+                    DeviceOp::CGridDesc_M0_M10_M11_N0_N10_N11,
+                    DefaultBlock2CTileMap,
+                    ComputePtrOffsetOfStridedBatch<NumDTensor>,
+                    has_main_loop,
+                    has_double_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_k0_m0_m1_k1_,
+                                              arg.b_grid_desc_k0_n0_n1_k1_,
+                                              arg.ds_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.e_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.block_2_ctile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            const auto K0                    = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0);
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+            const bool has_double_tail_k_block_loop =
+                GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+            if(has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, true>{});
+            }
+            else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, false>{});
+            }
+            else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, false>{});
+            }
+            return 0;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx926" || ck::get_device_name() == "gfx1030"))
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    std::cout << "Filter1x1Stride1Pad0 check: XY_index = " << i << " X = " << X
+                              << " ConvStride = " << ConvStride << " LeftPad = " << LeftPad
+                              << " RightPad = " << RightPad << std::endl;
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    std::cout << "Filter1x1Stride1Pad0 check: XY_index = " << i << " X = " << X
+                              << " LeftPad = " << LeftPad << " RightPad = " << RightPad
+                              << std::endl;
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            auto srcVectorLengths = ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            auto srcVectorLengths = BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of E
+        if constexpr(is_same_v<ELayout, ctc::G_NW_K> || is_same_v<ELayout, ctc::G_NHW_K> ||
+                     is_same_v<ELayout, ctc::G_NDHW_K> || is_same_v<ELayout, ctc::GNWK> ||
+                     is_same_v<ELayout, ctc::GNHWK> || is_same_v<ELayout, ctc::GNDHWK> ||
+                     is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
+                     is_same_v<ELayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.e_g_n_k_wos_lengths_[2];
+
+            if(!(K % CThreadTransferDstScalarPerVector == 0 && CThreadTransferSrcDstVectorDim == 5))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+        // check Gridwise GEMM
+        return GridwiseGemm::CheckValidity(
+            arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.e_grid_desc_m_n_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_k_wos_lengths,
+                        ds_g_n_k_wos_strides,
+                        e_g_n_k_wos_lengths,
+                        e_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_k_wos_lengths,
+                                          ds_g_n_k_wos_strides,
+                                          e_g_n_k_wos_lengths,
+                                          e_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..acdd381
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,837 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA, index_t BatchStrideB, index_t BatchStrideC)
+        : BatchStrideA_(BatchStrideA), BatchStrideB_(BatchStrideB), BatchStrideC_(BatchStrideC)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideC_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    index_t BatchStrideC_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename CDataType,
+          typename AGridDesc_K0_M0_M1_K1,
+          typename BGridDesc_K0_N0_N1_K1,
+          typename CGridDesc_M0_M10_M11_N0_N10_N11,
+          typename Block2CTileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_fwd_dl(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            CDataType* __restrict__ p_c_grid,
+            const index_t batch_count,
+            const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1,
+            const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1,
+            const CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11,
+            const Block2CTileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__)|| defined(__gfx926__) || defined(__gfx1030__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
+
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
+
+    __shared__ ABDataType p_shared[shared_block_size];
+
+    GridwiseGemm::Run(p_a_grid + a_batch_offset,
+                      p_b_grid + b_batch_offset,
+                      p_c_grid + c_batch_offset,
+                      p_shared,
+                      a_grid_desc_k0_m0_m1_k1,
+                      b_grid_desc_k0_n0_n1_k1,
+                      c_grid_desc_m0_m10_m11_n0_n10_n11,
+                      block_2_ctile_map,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m0_m1_k1;
+    ignore = b_grid_desc_k0_n0_n1_k1;
+    ignore = c_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+
+    compute_ptr_offset_of_batch.GetAPtrOffset(0);
+    compute_ptr_offset_of_batch.GetBPtrOffset(0);
+    compute_ptr_offset_of_batch.GetCPtrOffset(0);
+#endif
+}
+
+} // namespace
+
+//
+// @brief      Device Convolution operation.
+//
+// Supports:
+//  @li         Forward convolution with up to 3 spatial dimentions
+//  @li         Input tensor in GNWC data format
+//  @li         Weight tensor in GKXC data format
+//  @li         Output tensor in GNWK data format
+//
+// 1D:
+// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+// 2D:
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+// 3D:
+// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+//
+template <
+    index_t NDimSpatial,
+    typename ADataType,
+    typename BDataType,
+    typename CDataType,
+    typename AccDataType,
+    typename ALayout,
+    typename BLayout,
+    typename CLayout,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    ConvolutionForwardSpecialization ConvForwardSpecialization,
+    GemmSpecialization GemmSpec,
+    index_t BlockSize,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t K0PerBlock,
+    index_t K1,
+    index_t M1PerThread,
+    index_t N1PerThread,
+    index_t KPerThread,
+    typename M1N1ThreadClusterM1Xs,
+    typename M1N1ThreadClusterN1Xs,
+    typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+    typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+    typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+    typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+    typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+    typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+    typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+    typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+    typename CThreadTransferSrcDstAccessOrder,
+    index_t CThreadTransferSrcDstVectorDim,
+    index_t CThreadTransferDstScalarPerVector,
+    enable_if_t<
+        is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+        bool> = false>
+struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimSpatial,
+                                                                           ALayout,
+                                                                           BLayout,
+                                                                           CLayout,
+                                                                           ADataType,
+                                                                           BDataType,
+                                                                           CDataType,
+                                                                           AElementwiseOperation,
+                                                                           BElementwiseOperation,
+                                                                           CElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_AK0_M_AK1(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                                  const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                                  const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                                  const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                                  const std::array<index_t, NDimSpatial>& input_left_pads,
+                                  const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        c_g_n_k_wos_lengths,
+                                                                        c_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        const auto M = in_gemmm_gemmk_desc.GetLength(I0);
+        const auto K = in_gemmm_gemmk_desc.GetLength(I1);
+
+        const auto AK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            in_gemmm_gemmk_desc,
+            make_tuple(make_unmerge_transform(make_tuple(AK0, K1)), make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_BK0_N_BK1(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        const auto N = wei_gemmn_gemmk_desc.GetLength(I0);
+        const auto K = wei_gemmn_gemmk_desc.GetLength(I1);
+
+        const auto BK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            wei_gemmn_gemmk_desc,
+            make_tuple(make_unmerge_transform(make_tuple(BK0, K1)), make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    template <typename CLay>
+    static auto
+    MakeCGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<CLay>(c_g_n_k_wos_lengths,
+                                                                        c_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    // desc for problem definition
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
+        MakeAGridDescriptor_AK0_M_AK1<ALayout>({}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeBGridDescriptor_BK0_N_BK1<BLayout>({}, {}))>;
+    using CGridDesc_M_N = remove_cvref_t<decltype(MakeCGridDescriptor_M_N<CLayout>({}, {}))>;
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDl_km_kn_mn_v1r3<BlockSize,
+                                     ADataType,
+                                     AccDataType,
+                                     CDataType,
+                                     InMemoryDataOperationEnum::Set,
+                                     AGridDesc_AK0_M_AK1,
+                                     BGridDesc_BK0_N_BK1,
+                                     CGridDesc_M_N,
+                                     MPerBlock,
+                                     NPerBlock,
+                                     K0PerBlock,
+                                     K1,
+                                     M1PerThread,
+                                     N1PerThread,
+                                     KPerThread,
+                                     M1N1ThreadClusterM1Xs,
+                                     M1N1ThreadClusterN1Xs,
+                                     ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterArrangeOrder,
+                                     ABlockTransferSrcAccessOrder,
+                                     ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                     ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                     ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                     BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterArrangeOrder,
+                                     BBlockTransferSrcAccessOrder,
+                                     BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                     BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                     BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                     CThreadTransferSrcDstAccessOrder,
+                                     CThreadTransferSrcDstVectorDim,
+                                     CThreadTransferDstScalarPerVector>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_AK0_M_AK1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_BK0_N_BK1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+    using DefaultBlock2CTileMap =
+        decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(CGridDesc_M_N{}));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,
+                 const void* p_b,
+                 void* p_c,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CElementwiseOperation& c_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_c_grid_{static_cast<CDataType*>(p_c)},
+              num_group_{a_g_n_c_wis_lengths[0]},
+              a_grid_desc_ak0_m_ak1_{
+                  DeviceOp::MakeAGridDescriptor_AK0_M_AK1<ALayout>(a_g_n_c_wis_lengths,
+                                                                   a_g_n_c_wis_strides,
+                                                                   b_g_k_c_xs_lengths,
+                                                                   b_g_k_c_xs_strides,
+                                                                   c_g_n_k_wos_lengths,
+                                                                   c_g_n_k_wos_strides,
+                                                                   conv_filter_strides,
+                                                                   conv_filter_dilations,
+                                                                   input_left_pads,
+                                                                   input_right_pads)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1<BLayout>(
+                  b_g_k_c_xs_lengths, b_g_k_c_xs_strides)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N<CLayout>(c_g_n_k_wos_lengths,
+                                                                          c_g_n_k_wos_strides)},
+              a_grid_desc_k0_m0_m1_k1_{},
+              b_grid_desc_k0_n0_n1_k1_{},
+              c_grid_desc_m0_m10_m11_n0_n10_n11_{},
+              block_2_ctile_map_{},
+              compute_ptr_offset_of_batch_{
+                  a_g_n_c_wis_strides[0], b_g_k_c_xs_strides[0], c_g_n_k_wos_strides[0]},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              c_g_n_k_wos_lengths_{c_g_n_k_wos_lengths},
+              c_g_n_k_wos_strides_{c_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideC_ = c_g_n_k_wos_strides[0];
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_ak0_m_ak1_, b_grid_desc_bk0_n_bk1_, c_grid_desc_m_n_))
+            {
+
+                a_grid_desc_k0_m0_m1_k1_ =
+                    GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_ak0_m_ak1_);
+                b_grid_desc_k0_n0_n1_k1_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_bk0_n_bk1_);
+                c_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(c_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[K0, M, K1]: " << a_grid_desc_ak0_m_ak1_ << std::endl;
+            std::cout << "B[K0, N, K1]: " << b_grid_desc_bk0_n_bk1_ << std::endl;
+            std::cout << "C[M, N]: " << c_grid_desc_m_n_ << std::endl;
+            std::cout << "num_group: " << num_group_ << std::endl;
+
+            std::cout << "A[k0, m0, m1, k1]: " << a_grid_desc_k0_m0_m1_k1_ << std::endl;
+            std::cout << "B[k0, n0, n1, k1]: " << b_grid_desc_k0_n0_n1_k1_ << std::endl;
+            std::cout << "A[m0, m10, m11, n0, n10, n11]: " << c_grid_desc_m0_m10_m11_n0_n10_n11_
+                      << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+
+        // tensor descriptors for problem definiton
+        index_t num_group_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_;
+        BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_;
+        CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        // block-to-e-tile map
+        DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<index_t, NDimSpatial + 3> c_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> c_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            // if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(
+                   arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.c_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK has invalid setting");
+            }
+
+            const index_t grid_size =
+                GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_.GetLength(I0),
+                                                arg.c_grid_desc_m_n_.GetLength(I1)) *
+                arg.num_group_;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop,
+                                     auto has_double_tail_k_block_loop) {
+                constexpr bool has_main_loop   = has_main_k_block_loop.value;
+                constexpr bool has_double_loop = has_double_tail_k_block_loop;
+
+                const auto kernel =
+                    kernel_grouped_conv_fwd_dl<GridwiseGemm,
+                                               ADataType, // TODO: distiguish A/B datatype
+                                               CDataType,
+                                               DeviceOp::AGridDesc_K0_M0_M1_K1,
+                                               DeviceOp::BGridDesc_K0_N0_N1_K1,
+                                               DeviceOp::CGridDesc_M0_M10_M11_N0_N10_N11,
+                                               DefaultBlock2CTileMap,
+                                               ComputePtrOffsetOfStridedBatch,
+                                               has_main_loop,
+                                               has_double_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_c_grid_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_k0_m0_m1_k1_,
+                                              arg.b_grid_desc_k0_n0_n1_k1_,
+                                              arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.block_2_ctile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            const auto K0                    = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0);
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+            const bool has_double_tail_k_block_loop =
+                GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+            if(has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, true>{});
+            }
+            else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, false>{});
+            }
+            else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(!(ck::get_device_name() == "gfx906" ||ck::get_device_name() == "gfx926" || ck::get_device_name() == "gfx1030"))
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    std::cout << "Filter1x1Stride1Pad0 check: i = " << i << " X = " << X
+                              << " ConvStride = " << ConvStride << " LeftPad = " << LeftPad
+                              << " RightPad = " << RightPad << std::endl;
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    std::cout << "Filter1x1Stride1Pad0 check: i = " << i << " X = " << X
+                              << " LeftPad = " << LeftPad << " RightPad = " << RightPad
+                              << std::endl;
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            auto srcVectorLengths = ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            auto srcVectorLengths = BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of C
+        if constexpr(is_same_v<CLayout, ctc::G_NW_K> || is_same_v<CLayout, ctc::G_NHW_K> ||
+                     is_same_v<CLayout, ctc::G_NDHW_K> || is_same_v<CLayout, ctc::GNWK> ||
+                     is_same_v<CLayout, ctc::GNHWK> || is_same_v<CLayout, ctc::GNDHWK> ||
+                     is_same_v<CLayout, ctc::NWGK> || is_same_v<CLayout, ctc::NHWGK> ||
+                     is_same_v<CLayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.c_g_n_k_wos_lengths_[2];
+
+            if(!(K % CThreadTransferDstScalarPerVector == 0 && CThreadTransferSrcDstVectorDim == 5))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+        // check Gridwise GEMM
+        return GridwiseGemm::CheckValidity(
+            arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.c_grid_desc_m_n_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             void* p_c,
+                             const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                             const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                             const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                             const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                             const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                             const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                             const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                             const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                             const std::array<index_t, NDimSpatial>& input_left_pads,
+                             const std::array<index_t, NDimSpatial>& input_right_pads,
+                             const AElementwiseOperation& a_element_op,
+                             const BElementwiseOperation& b_element_op,
+                             const CElementwiseOperation& c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        c_g_n_k_wos_lengths,
+                        c_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads,
+                        const AElementwiseOperation& a_element_op,
+                        const BElementwiseOperation& b_element_op,
+                        const CElementwiseOperation& c_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_c,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          c_g_n_k_wos_lengths,
+                                          c_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
new file mode 100644
index 0000000..1e2f819
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Convolution Forward:
+//   input : input image A[G, N, C, Hi, Wi],
+//   input : weight B[G, K, C, Y, X],
+//   input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
+//   output : output image E[G, N, K, Ho, Wo]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGroupedConvFwdMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static_assert(NumDTensor == DsLayout::Size(), "wrong! Inconsistent NumDTensor");
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a, // input image
+        const void* p_b, // weight
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e, // output image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp
new file mode 100644
index 0000000..181ee4b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp
@@ -0,0 +1,51 @@
+#pragma once
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+struct GemmDesc
+{
+    ck::index_t M_, N_, K_;
+    ck::index_t stride_A_, stride_B_, stride_C_;
+
+    std::vector<ck::index_t> stride_Ds_;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGroupedGemm : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static_assert(DsLayout::Size() == DsDataType::Size(), "wrong! inconsisiten NumDTensor");
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*>& p_a,
+                        std::vector<const void*>& p_b,
+                        std::vector<std::array<const void*, NumDTensor>>& p_ds,
+                        std::vector<void*>& p_e,
+                        std::vector<GemmDesc>& gemm_desc,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
new file mode 100644
index 0000000..b066a44
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
@@ -0,0 +1,75 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasDataType,
+          typename Acc1BiasDataType,
+          typename AElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename Acc0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          MaskingSpecialization MaskingSpec>
+struct DeviceGroupedGemmSoftmaxGemmPermute : public BaseOperator
+{
+    struct ProblemDesc
+    {
+        std::vector<index_t> a_gs_ms_ks_lengths;
+        std::vector<index_t> a_gs_ms_ks_strides;
+
+        std::vector<index_t> b0_gs_ns_ks_lengths;
+        std::vector<index_t> b0_gs_ns_ks_strides;
+
+        std::vector<index_t> b1_gs_os_ns_lengths;
+        std::vector<index_t> b1_gs_os_ns_strides;
+
+        std::vector<index_t> c_gs_ms_os_lengths;
+        std::vector<index_t> c_gs_ms_os_strides;
+
+        std::vector<std::vector<index_t>> acc0_biases_gs_ms_ns_lengths;
+        std::vector<std::vector<index_t>> acc0_biases_gs_ms_ns_strides;
+
+        std::vector<std::vector<index_t>> acc1_biases_gs_ms_os_lengths;
+        std::vector<std::vector<index_t>> acc1_biases_gs_ms_os_strides;
+    };
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b0_vec,
+                        std::vector<const void*> p_b1_vec,
+                        std::vector<void*> p_c_vec,
+                        std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                        std::vector<std::vector<const void*>> p_acc1_biases_vec,
+                        std::vector<ProblemDesc> problem_desc_vec,
+                        AElementwiseOperation a_element_op,
+                        B0ElementwiseOperation b0_element_op,
+                        Acc0ElementwiseOperation acc0_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
new file mode 100644
index 0000000..946a757
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
@@ -0,0 +1,881 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename GroupKernelArg,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_gemm_softmax_gemm_xdl_cshuffle_v1(
+            const void CK_CONSTANT_ADDRESS_SPACE* group_kernel_args,
+            const index_t group_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const AccElementwiseOperation acc_element_op,
+            const B1ElementwiseOperation b1_element_op,
+            const CElementwiseOperation c_element_op)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const auto arg_ptr = reinterpret_cast<const GroupKernelArg*>(
+        cast_pointer_to_generic_address_space(group_kernel_args));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+
+    while(
+        (!(block_id >= arg_ptr[group_id].block_start_ && block_id < arg_ptr[group_id].block_end_)))
+    {
+        if(block_id < arg_ptr[group_id].block_start_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    // per-group batch offset
+    const index_t num_blocks_per_batch = arg_ptr[group_id].num_blocks_per_batch_;
+    const index_t g_idx                = __builtin_amdgcn_readfirstlane(
+        (block_id - arg_ptr[group_id].block_start_) / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetBBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
+        arg_ptr[group_id].compute_base_ptr_of_batch_.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset  = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetCBasePtr(g_idx)));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        arg_ptr[group_id].p_a_grid_ + a_batch_offset,
+        arg_ptr[group_id].p_b_grid_ + b_batch_offset,
+        arg_ptr[group_id].p_b1_grid_ + b1_batch_offset,
+        arg_ptr[group_id].p_c_grid_ + c_batch_offset,
+        p_shared,
+        a_element_op,
+        b_element_op,
+        acc_element_op,
+        b1_element_op,
+        c_element_op,
+        arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
+        arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
+        arg_ptr[group_id].b1_grid_desc_bk0_n_bk1_,
+        arg_ptr[group_id].c_grid_desc_mblock_mperblock_nblock_nperblock_,
+        arg_ptr[group_id].block_2_ctile_map_,
+        arg_ptr[group_id].c0_matrix_mask_);
+#else
+    ignore = group_kernel_args;
+    ignore = group_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = acc_element_op;
+    ignore = b1_element_op;
+    ignore = c_element_op;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// Computes C = A * B0 * B1
+//              ^^^^^^ (Acc0)
+//              ^^^^^^^^^^^ (Acc1)
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO, // NumDimGemm1N
+          typename ADataType,
+          typename BDataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasDataType,
+          typename Acc1BiasDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization B1Spec,
+          TensorSpecialization CSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock, // Gemm0NPerBlock
+          index_t KPerBlock, // Gemm0KPerBlock
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t B1K1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          MaskingSpecialization MaskingSpec,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
+    : public DeviceGroupedGemmSoftmaxGemmPermute<NumDimG,
+                                                 NumDimM,
+                                                 NumDimN,
+                                                 NumDimK,
+                                                 NumDimO,
+                                                 ADataType,
+                                                 BDataType,
+                                                 B1DataType,
+                                                 CDataType,
+                                                 Acc0BiasDataType,
+                                                 Acc1BiasDataType,
+                                                 AElementwiseOperation,
+                                                 BElementwiseOperation,
+                                                 AccElementwiseOperation,
+                                                 B1ElementwiseOperation,
+                                                 CElementwiseOperation,
+                                                 MaskingSpec>
+{
+    static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0,
+                  "Number of dimension must be greater than 0");
+
+    static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
+    static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
+
+    // TODO ANT: implement bias combination
+    static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented");
+
+#if 0
+    // TODO ANT: use alias
+    static constexpr index_t NumDimGemm0M = NumDimM;
+    static constexpr index_t NumDimGemm0N = NumDimN;
+    static constexpr index_t NumDimGemm0K = NumDimK;
+    static constexpr index_t NumDimGemm1M = NumDimM;
+    static constexpr index_t NumDimGemm1N = NumDimO;
+    static constexpr index_t NumDimGemm1K = NumDimN;
+#endif
+
+    using DeviceOp    = DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle;
+    using ProblemDesc = typename DeviceGroupedGemmSoftmaxGemmPermute<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     NumDimO,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     B1DataType,
+                                                                     CDataType,
+                                                                     Acc0BiasDataType,
+                                                                     Acc1BiasDataType,
+                                                                     AElementwiseOperation,
+                                                                     BElementwiseOperation,
+                                                                     AccElementwiseOperation,
+                                                                     B1ElementwiseOperation,
+                                                                     CElementwiseOperation,
+                                                                     MaskingSpec>::ProblemDesc;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    using Transform = TransformBatchedContractionContractionToBatchedGemmGemm<
+        Sequence<NumDimG, NumDimM, NumDimN, NumDimK, NumDimO>,
+        Sequence<MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock>,
+        GemmSpec,
+        ASpec,
+        BSpec,
+        B1Spec,
+        CSpec>;
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                              const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        return Transform::MakeAGridDescriptor_AK0_M_AK1(
+            Transform::MakeAGridDescriptor_M_K(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec),
+            Number<AK1>{});
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                              const std::vector<index_t>& b_gs_ns_ks_strides_vec)
+    {
+        return Transform::MakeB0GridDescriptor_BK0_N_BK1(
+            Transform::MakeB0GridDescriptor_N_K(b_gs_ns_ks_lengths_vec, b_gs_ns_ks_strides_vec),
+            Number<BK1>{});
+    }
+
+    static auto
+    MakeB1GridDescriptor_BK0_N_BK1(const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                   const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides_vec)
+    {
+        return Transform::MakeB1GridDescriptor_BK0_N_BK1(
+            Transform::MakeB1GridDescriptor_N_K(b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                                b1_gs_gemm1ns_gemm1ks_strides_vec),
+            Number<B1K1>{});
+    }
+
+    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {}));
+    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {}));
+    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {}));
+    using CGridDesc_M_N        = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
+    using AGridDesc_G_M_K      = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
+    using BGridDesc_G_N_K      = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
+    using B1GridDesc_G_N_K     = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
+    using CGridDesc_G_M_N      = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
+
+    constexpr static auto make_MaskOutPredicate()
+    {
+        if constexpr(MaskingSpec == MaskingSpecialization::MaskDisabled)
+        {
+            return MaskDisabledPredicate{};
+        }
+        else if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+        {
+            return MaskOutUpperTrianglePredicate{};
+        }
+    }
+    using C0MatrixMask = C0MatrixMask_impl<decltype(make_MaskOutPredicate())>;
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k,
+                                     const BGridDesc_G_N_K& b_grid_desc_g_n_k,
+                                     const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
+                                     const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+            : a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
+              b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
+              b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
+              c_grid_desc_g_m_n_(c_grid_desc_g_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return a_grid_desc_g_m_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return b_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
+        {
+            return b1_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return c_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        private:
+        AGridDesc_G_M_K a_grid_desc_g_m_k_;
+        BGridDesc_G_N_K b_grid_desc_g_n_k_;
+        B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
+        CGridDesc_G_M_N c_grid_desc_g_m_n_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        AccElementwiseOperation,
+        B1ElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        B1GridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        AK1,
+        BK1,
+        B1K1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        Gemm1NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        true,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        true,
+        BBlockLdsExtraN,
+        B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B1BlockTransferThreadClusterArrangeOrder,
+        B1BlockTransferSrcAccessOrder,
+        B1BlockTransferSrcVectorDim,
+        B1BlockTransferSrcScalarPerVector,
+        B1BlockTransferDstScalarPerVector_BK1,
+        false,
+        B1BlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        Transform::matrix_padder.PadN,
+        MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle>;
+
+    using Block2CTileMap = OffsettedBlockToCTileMap<typename GridwiseGemm::DefaultBlock2CTileMap>;
+
+    struct GroupKernelArg
+    {
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        const B1DataType* p_b1_grid_;
+        CDataType* p_c_grid_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // batch & stride
+        index_t num_blocks_per_batch_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+
+        // check C0 masking and padding
+        C0MatrixMask c0_matrix_mask_;
+
+        // block-to-c-tile map
+        Block2CTileMap block_2_ctile_map_;
+
+        index_t block_start_, block_end_;
+    };
+
+    struct GroupDeviceArg
+    {
+        // lengths for the last dimensions of overall problem for sanity check of vector load/store
+        std::vector<index_t> raw_lengths_mz_nz_kz_gemm1nz_;
+
+        // strides for the last dimensions of each tensor for sanity check of vector load/store
+        std::vector<index_t> a_mz_kz_strides_;
+        std::vector<index_t> b_nz_kz_strides_;
+        std::vector<index_t> b1_nz_kz_strides_;
+        std::vector<index_t> c_mz_gemm1nz_strides_;
+
+        // for gridwise gemm check
+        CGridDesc_M_N c_grid_desc_m_n_;
+    };
+
+    // Argument
+    // FIXME: constness
+    struct Argument : public BaseArgument
+    {
+        Argument(std::vector<const void*> p_a_vec,
+                 std::vector<const void*> p_b_vec,
+                 std::vector<const void*> p_b1_vec,
+                 std::vector<void*> p_c_vec,
+                 std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                 std::vector<std::vector<const void*>> p_acc1_biases_vec,
+                 std::vector<ProblemDesc> problem_desc_vec,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 AccElementwiseOperation acc_element_op,
+                 B1ElementwiseOperation b1_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              b1_element_op_{b1_element_op},
+              c_element_op_{c_element_op}
+        {
+            // TODO ANT: implement bias addition
+            group_count_ = problem_desc_vec.size();
+
+            if(!(group_count_ == p_a_vec.size() && group_count_ == p_b_vec.size() &&
+                 group_count_ == p_b1_vec.size() && group_count_ == p_c_vec.size()))
+            {
+                throw std::runtime_error("wrong! group_count_ != a/b/b1/c_vec.size");
+            }
+
+            if(!(p_acc0_biases_vec.size() == p_acc1_biases_vec.size()))
+            {
+                throw std::runtime_error("wrong! acc0_bias_vec.size != acc1_bias_vec.size");
+            }
+
+            grid_size_ = 0;
+
+            for(std::size_t i = 0; i < group_count_; i++)
+            {
+                const auto p_a_grid  = static_cast<const ADataType*>(p_a_vec[i]);
+                const auto p_b_grid  = static_cast<const BDataType*>(p_b_vec[i]);
+                const auto p_b1_grid = static_cast<const B1DataType*>(p_b1_vec[i]);
+                const auto p_c_grid  = static_cast<CDataType*>(p_c_vec[i]);
+
+                const auto& problem_desc = problem_desc_vec[i];
+
+                const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
+                    problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
+                const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
+                    problem_desc.b0_gs_ns_ks_lengths, problem_desc.b0_gs_ns_ks_strides);
+                const auto b1_grid_desc_bk0_n_bk1 = MakeB1GridDescriptor_BK0_N_BK1(
+                    problem_desc.b1_gs_os_ns_lengths, problem_desc.b1_gs_os_ns_strides);
+                const auto c_grid_desc_m_n = Transform::MakeCGridDescriptor_M_N(
+                    problem_desc.c_gs_ms_os_lengths, problem_desc.c_gs_ms_os_strides);
+
+                const auto a_grid_desc_g_m_k = Transform::MakeAGridDescriptor_G_M_K(
+                    problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
+                const auto b_grid_desc_g_n_k = Transform::MakeB0GridDescriptor_G_N_K(
+                    problem_desc.b0_gs_ns_ks_lengths, problem_desc.b0_gs_ns_ks_strides);
+                const auto b1_grid_desc_g_n_k = Transform::MakeB1GridDescriptor_G_N_K(
+                    problem_desc.b1_gs_os_ns_lengths, problem_desc.b1_gs_os_ns_strides);
+                const auto c_grid_desc_g_m_n = Transform::MakeCGridDescriptor_G_M_N(
+                    problem_desc.c_gs_ms_os_lengths, problem_desc.c_gs_ms_os_strides);
+
+                const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n);
+
+                const index_t BlockStart     = grid_size_;
+                const auto block_2_ctile_map = Block2CTileMap(c_grid_desc_m_n, BlockStart);
+                const index_t batch_count    = c_grid_desc_g_m_n.GetLength(I0);
+                const index_t grid_size_grp =
+                    block_2_ctile_map.CalculateGridSize(c_grid_desc_m_n) * batch_count;
+                const index_t BlockEnd = grid_size_ + grid_size_grp;
+
+                // batch stride
+                const auto compute_base_ptr_of_batch = ComputeBasePtrOfStridedBatch(
+                    a_grid_desc_g_m_k, b_grid_desc_g_n_k, b1_grid_desc_g_n_k, c_grid_desc_g_m_n);
+
+                // C0 mask
+                const auto c0_matrix_mask = C0MatrixMask(b_grid_desc_g_n_k.GetLength(I1));
+
+                grid_size_ += grid_size_grp;
+
+                // for each group, make sure acc0_biases_gs_ms_ns_lengths.size() == NumAcc0Bias and
+                // so on
+                if(!(problem_desc.acc0_biases_gs_ms_ns_lengths.size() == NumAcc0Bias &&
+                     problem_desc.acc0_biases_gs_ms_ns_strides.size() == NumAcc0Bias &&
+                     problem_desc.acc1_biases_gs_ms_os_lengths.size() == NumAcc1Bias &&
+                     problem_desc.acc1_biases_gs_ms_os_strides.size() == NumAcc1Bias))
+                {
+                    throw std::runtime_error(
+                        "wrong! number of biases in function argument does not "
+                        "match that in template argument");
+                }
+
+                group_kernel_args_.push_back({p_a_grid,
+                                              p_b_grid,
+                                              p_b1_grid,
+                                              p_c_grid,
+                                              a_grid_desc_ak0_m_ak1,
+                                              b_grid_desc_bk0_n_bk1,
+                                              b1_grid_desc_bk0_n_bk1,
+                                              c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                              block_2_ctile_map.CalculateGridSize(c_grid_desc_m_n),
+                                              compute_base_ptr_of_batch,
+                                              c0_matrix_mask,
+                                              block_2_ctile_map,
+                                              BlockStart,
+                                              BlockEnd});
+
+                group_device_args_.push_back(
+                    {{problem_desc.a_gs_ms_ks_lengths[NumDimG + NumDimM - 1],
+                      problem_desc.b0_gs_ns_ks_lengths[NumDimG + NumDimN - 1],
+                      problem_desc.b0_gs_ns_ks_lengths[NumDimG + NumDimN + NumDimK - 1],
+                      problem_desc.b1_gs_os_ns_lengths[NumDimG + NumDimO - 1]},
+                     {problem_desc.a_gs_ms_ks_strides[NumDimG + NumDimM - 1],
+                      problem_desc.a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1]},
+                     {problem_desc.b0_gs_ns_ks_strides[NumDimG + NumDimN - 1],
+                      problem_desc.b0_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1]},
+                     {problem_desc.b1_gs_os_ns_strides[NumDimG + NumDimO - 1],
+                      problem_desc.b1_gs_os_ns_strides[NumDimG + NumDimO + NumDimN - 1]},
+                     {problem_desc.c_gs_ms_os_strides[NumDimG + NumDimM - 1],
+                      problem_desc.c_gs_ms_os_strides[NumDimG + NumDimM + NumDimO - 1]},
+                     c_grid_desc_m_n});
+            }
+        }
+
+        std::vector<GroupKernelArg> group_kernel_args_;
+        std::vector<GroupDeviceArg> group_device_args_;
+
+        std::size_t group_count_;
+        index_t grid_size_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!DeviceOp::IsSupportedArgument(arg))
+            {
+                throw std::runtime_error("wrong! unsupported argument");
+            }
+
+            bool all_has_main_k_block_loop  = true;
+            bool some_has_main_k_block_loop = false;
+            for(std::size_t i = 0; i < arg.group_count_; i++)
+            {
+                const auto K = arg.group_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                               arg.group_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2);
+                const bool y = GridwiseGemm::CalculateHasMainKBlockLoop(K);
+                all_has_main_k_block_loop &= y;
+                some_has_main_k_block_loop |= y;
+            }
+
+            hipGetErrorString(hipMemcpy(arg.p_workspace_,
+                                        arg.group_kernel_args_.data(),
+                                        arg.group_kernel_args_.size() * sizeof(GroupKernelArg),
+                                        hipMemcpyHostToDevice));
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel =
+                    kernel_grouped_gemm_softmax_gemm_xdl_cshuffle_v1<GridwiseGemm,
+                                                                     GroupKernelArg,
+                                                                     AElementwiseOperation,
+                                                                     BElementwiseOperation,
+                                                                     AccElementwiseOperation,
+                                                                     B1ElementwiseOperation,
+                                                                     CElementwiseOperation,
+                                                                     has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(arg.grid_size_),
+                    dim3(BlockSize),
+                    0,
+                    cast_pointer_to_constant_address_space(arg.p_workspace_),
+                    arg.group_count_,
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.acc_element_op_,
+                    arg.b1_element_op_,
+                    arg.c_element_op_);
+            };
+
+            // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
+            // to concern Gemm0's loop
+            if(all_has_main_k_block_loop)
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else if(!some_has_main_k_block_loop)
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+            else
+            {
+                throw std::runtime_error("wrong! all gemm problems have to simultaneously meet "
+                                         "has_main_k_block_loop or no_main_k_block_loop");
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        // TODO ANT: Check if tensor specialization & strides mismatch
+
+        bool all_has_main_k_block_loop  = true;
+        bool some_has_main_k_block_loop = false;
+
+        for(std::size_t i = 0; i < arg.group_count_; i++)
+        {
+            const auto& kernel_arg = arg.group_kernel_args_[i];
+            const auto& device_arg = arg.group_device_args_[i];
+
+            // Check if C permute dimension matches GEMM + GEMM shape
+            const index_t c_m       = device_arg.c_grid_desc_m_n_.GetLength(I0);
+            const index_t c_gemm1n  = device_arg.c_grid_desc_m_n_.GetLength(I1);
+            const index_t a_m       = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
+            const index_t b1_gemm1n = kernel_arg.b1_grid_desc_bk0_n_bk1_.GetLength(I1);
+            if(!(c_m == a_m && c_gemm1n == b1_gemm1n))
+            {
+                return false;
+            }
+
+            // Check if having main loop
+            const auto K = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                           kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+            const bool y = GridwiseGemm::CalculateHasMainKBlockLoop(K);
+            all_has_main_k_block_loop &= y;
+            some_has_main_k_block_loop |= y;
+
+            // Note: we need raw lengths since threadwise copy can not handle vector load when
+            // part of vector is out of bounds
+            const auto MzRaw      = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[0];
+            const auto NzRaw      = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[1];
+            const auto KzRaw      = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[2];
+            const auto Gemm1NzRaw = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[3];
+
+            // Check scalar per vector requirement
+            const auto a_extent_lowest  = ABlockTransferSrcVectorDim == 2 ? KzRaw : MzRaw;
+            const auto b_extent_lowest  = BBlockTransferSrcVectorDim == 2 ? KzRaw : NzRaw;
+            const auto b1_extent_lowest = B1BlockTransferSrcVectorDim == 2 ? NzRaw : Gemm1NzRaw;
+            const auto c_extent_lowest  = Gemm1NzRaw;
+
+            if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
+                 b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
+                 b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
+                 c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+
+            // Check vector load/store requirement
+            const auto a_stride_lowest = ABlockTransferSrcVectorDim == 2
+                                             ? device_arg.a_mz_kz_strides_[1]
+                                             : device_arg.a_mz_kz_strides_[0];
+            const auto b_stride_lowest = BBlockTransferSrcVectorDim == 2
+                                             ? device_arg.b_nz_kz_strides_[1]
+                                             : device_arg.b_nz_kz_strides_[0];
+            const auto b1_stride_lowest = B1BlockTransferSrcVectorDim == 2
+                                              ? device_arg.b1_nz_kz_strides_[1]
+                                              : device_arg.b1_nz_kz_strides_[0];
+            const auto c_stride_lowest =
+                device_arg.c_mz_gemm1nz_strides_[1]; // cshuffle assumes lowest dim in Gemm1Ns to be
+                                                     // contiguous
+
+            if(!(a_stride_lowest == 1 || b_stride_lowest == 1 || b1_stride_lowest == 1 ||
+                 c_stride_lowest == 1))
+            {
+                return false;
+            }
+
+            if(!GridwiseGemm::CheckValidity(kernel_arg.a_grid_desc_ak0_m_ak1_,
+                                            kernel_arg.b_grid_desc_bk0_n_bk1_,
+                                            kernel_arg.b1_grid_desc_bk0_n_bk1_,
+                                            device_arg.c_grid_desc_m_n_,
+                                            kernel_arg.block_2_ctile_map_))
+            {
+                return false;
+            }
+        }
+
+        // all gemm problems have to simultaneously meet has_main_k_block_loop or
+        // no_main_k_block_loop
+        if(!(all_has_main_k_block_loop || !some_has_main_k_block_loop))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*> p_a_vec,
+                             std::vector<const void*> p_b_vec,
+                             std::vector<const void*> p_b1_vec,
+                             std::vector<void*> p_c_vec,
+                             std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                             std::vector<std::vector<const void*>> p_acc1_biases_vec,
+                             std::vector<ProblemDesc> problem_desc_vec,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             AccElementwiseOperation acc_element_op,
+                             B1ElementwiseOperation b1_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a_vec,
+                        p_b_vec,
+                        p_b1_vec,
+                        p_c_vec,
+                        p_acc0_biases_vec,
+                        p_acc1_biases_vec,
+                        problem_desc_vec,
+                        a_element_op,
+                        b_element_op,
+                        acc_element_op,
+                        b1_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b_vec,
+                        std::vector<const void*> p_b1_vec,
+                        std::vector<void*> p_c_vec,
+                        std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                        std::vector<std::vector<const void*>> p_acc1_biases_vec,
+                        std::vector<ProblemDesc> problem_desc_vec,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        AccElementwiseOperation acc_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(p_a_vec,
+                                          p_b_vec,
+                                          p_b1_vec,
+                                          p_c_vec,
+                                          p_acc0_biases_vec,
+                                          p_acc1_biases_vec,
+                                          problem_desc_vec,
+                                          a_element_op,
+                                          b_element_op,
+                                          acc_element_op,
+                                          b1_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerBlock << ", "
+            << Gemm1NPerBlock << ", "
+            << Gemm1KPerBlock << ", "
+            << B1K1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << "ASpec" << getTensorSpecializationString(ASpec) << ", "
+            << "B0Spec" << getTensorSpecializationString(BSpec) << ", "
+            << "B1Spec" << getTensorSpecializationString(B1Spec) << ", "
+            << "CSpec" << getTensorSpecializationString(CSpec) << ", "
+            << getMaskingSpecializationString(MaskingSpec) << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->group_count_ * sizeof(GroupKernelArg);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp
new file mode 100644
index 0000000..93202e3
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include <array>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t Rank,
+          index_t NumReduceDim,
+          index_t NumReduction,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple>
+struct DeviceMultipleReduce : public BaseOperator
+{
+    static constexpr index_t NumInputDim  = Rank;
+    static constexpr index_t NumOutputDim = (Rank - NumReduceDim > 1) ? Rank - NumReduceDim : 1;
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, NumInputDim> inLengths,
+        const std::array<index_t, NumInputDim> inStrides,
+        const std::array<index_t, NumOutputDim> outLengths,
+        const std::array<std::array<index_t, NumOutputDim>, NumReduction> outStrides,
+        const std::array<int, NumReduceDim> reduceDims,
+        const std::array<const void*, NumReduction> alphas,
+        const std::array<const void*, NumReduction> betas,
+        const void* in_dev,
+        const std::array<void*, NumReduction> out_dev_buffers,
+        const InElementwiseOperationTuple in_elementwise_op_tuple,
+        const AccElementwiseOperationTuple acc_elementwise_op_tuple) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <index_t Rank,
+          index_t NumReduceDim,
+          index_t NumReduction,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple>
+using DeviceMultipleReducePtr = std::unique_ptr<DeviceMultipleReduce<Rank,
+                                                                     NumReduceDim,
+                                                                     NumReduction,
+                                                                     InElementwiseOperationTuple,
+                                                                     AccElementwiseOperationTuple>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_normalization.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_normalization.hpp
new file mode 100644
index 0000000..227c352
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_normalization.hpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          typename AccElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceNormalization : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::vector<index_t> xStrides,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        AccDataType epsilon,
+                        const void* p_x,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        void* p_savedMean,
+                        void* p_savedInvVar,
+                        AccElementwiseOperation acc_elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          typename AccElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+using DeviceNormalizationPtr = std::unique_ptr<DeviceNormalization<XDataType,
+                                                                   GammaDataType,
+                                                                   BetaDataType,
+                                                                   AccDataType,
+                                                                   YDataType,
+                                                                   AccElementwiseOperation,
+                                                                   Rank,
+                                                                   NumReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_permute.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_permute.hpp
new file mode 100644
index 0000000..baa9144
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_permute.hpp
@@ -0,0 +1,37 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <cmath>
+#include <memory>
+#include <type_traits>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumDim, typename InDataType, typename OutDataType, typename ElementwiseOperation>
+struct DevicePermute : BaseOperator
+{
+    using Lengths = std::array<index_t, NumDim>;
+    using Strides = Lengths;
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const Lengths& in_lengths,
+                        const Strides& in_strides,
+                        const Lengths& out_lengths,
+                        const Strides& out_strides,
+                        const void* in_dev_buffer,
+                        void* out_dev_buffer,
+                        ElementwiseOperation elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp
new file mode 100644
index 0000000..3b376c6
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <ck::ReduceTensorOp ReduceOpId>
+struct DevicePool2dFwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* in_dev,
+                        void* out_dev,
+                        void* out_indices_dev,
+                        ck::index_t N,
+                        ck::index_t C,
+                        std::array<ck::index_t, 2> input_spatial_lengths,
+                        std::array<ck::index_t, 2> window_spatial_lengths,
+                        std::array<ck::index_t, 2> output_spatial_lengths,
+                        std::array<ck::index_t, 2> window_strides,
+                        std::array<ck::index_t, 2> input_left_pads,
+                        std::array<ck::index_t, 2> input_right_pads) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <ck::ReduceTensorOp ReduceOpId>
+using DevicePool2dFwdPtr = std::unique_ptr<DevicePool2dFwd<ReduceOpId>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_reduce.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_reduce.hpp
new file mode 100644
index 0000000..15aeb8e
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_reduce.hpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t Rank,
+          index_t NumReduceDim,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation>
+struct DeviceReduce : public BaseOperator
+{
+    static constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> inLengths,
+                        const std::array<index_t, Rank> inStrides,
+                        const std::array<index_t, NumOutDim> outLengths,
+                        const std::array<index_t, NumOutDim> outStrides,
+                        const std::array<int, NumReduceDim> reduceDims,
+                        float alpha,
+                        float beta,
+                        const void* in_dev,
+                        const void* in_index_dev,
+                        void* out_dev,
+                        void* out_index_dev,
+                        const InElementwiseOperation in_elementwise_op,
+                        const AccElementwiseOperation acc_elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <index_t Rank,
+          index_t NumReduceDim,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation>
+using DeviceReducePtr = std::unique_ptr<
+    DeviceReduce<Rank, NumReduceDim, InElementwiseOperation, AccElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_softmax.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_softmax.hpp
new file mode 100644
index 0000000..676e081
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_softmax.hpp
@@ -0,0 +1,69 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          index_t Rank>
+struct DeviceSoftmax : public BaseOperator
+{
+    //
+    // @brief      Makes a pointer to Argument class.
+    //
+    // @param[in]  inLengths           Input tensor extent(s) from high to low dimension
+    // @param[in]  inStrides           Input tensor stride(s) from high to low dimension
+    // @param[in]  reduceDims          The dimension(s) the normalization operation is applied
+    // @param[in]  alpha               Typeless pointer in host memory storing the alpha scaling
+    //                                 value as type AccDataType
+    // @param[in]  beta                Typeless pointer in host memory storing the beta scaling
+    //                                 value as type AccDataType
+    // @param[in]  in_dev              Typeless const pointer in device memory storing the input
+    //                                 tensor
+    // @param      out_dev             Typeless pointer in device memory storing the output tensor
+    // @param[in]  in_elementwise_op   The input elementwise operation.
+    // @param[in]  acc_elementwise_op  The accumulation elementwise operation.
+    //
+    // @return     Unique pointer to the Argument class.
+    //
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> inLengths,
+                        const std::vector<index_t> inStrides,
+                        const std::vector<int> reduceDims,
+                        const void* alpha,
+                        const void* beta,
+                        const void* in_dev,
+                        void* out_dev,
+                        InElementwiseOp in_elementwise_op,
+                        AccElementwiseOp acc_elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+    virtual index_t GetRank() const                           = 0;
+    virtual index_t GetNumReduceDim() const                   = 0;
+};
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          index_t Rank>
+using DeviceSoftmaxPtr = std::unique_ptr<
+    DeviceSoftmax<InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp
new file mode 100644
index 0000000..f59e609
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceSplitKContractionMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_gs_ms_ns_lengths,
+                        const std::vector<index_t>& a_gs_ms_ks_strides,
+                        const std::vector<index_t>& b_gs_ns_ks_lengths,
+                        const std::vector<index_t>& b_gs_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                        const std::vector<index_t>& e_gs_ms_ns_lengths,
+                        const std::vector<index_t>& e_gs_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op,
+                        index_t split_k) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000..8eab1cd
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,1147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatDsPointer,
+          typename FloatE,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AKB_AK0_M_AK1,
+          typename BGridDesc_BKB_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_contraction_multiple_d_xdl_cshuffle(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatDsPointer p_ds_grid,
+            FloatE* __restrict__ p_e_grid,
+            const index_t batch_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const AGridDesc_AKB_AK0_M_AK1 a_grid_desc_akb_ak0_m_ak1,
+            const BGridDesc_BKB_BK0_N_BK1 b_grid_desc_bkb_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    FloatDsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_akb_ak0_m_ak1,
+                                                  b_grid_desc_bkb_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_akb_ak0_m_ak1;
+    ignore = b_grid_desc_bkb_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+    ignore = compute_ptr_offset_of_batch;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization DESpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceSplitKContractionMultipleD_Xdl_CShuffle
+    : public DeviceSplitKContractionMultipleD<NumDimG,
+                                              NumDimM,
+                                              NumDimN,
+                                              NumDimK,
+                                              ADataType,
+                                              BDataType,
+                                              DsDataType,
+                                              EDataType,
+                                              AElementwiseOperation,
+                                              BElementwiseOperation,
+                                              CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceSplitKContractionMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    // Assume: A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        assert(a_gs_ms_ks_lengths_vec.size() == NumDimG + NumDimM + NumDimK &&
+               a_gs_ms_ks_strides_vec.size() == NumDimG + NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto a_ms_ks_lengths = to_tuple(
+            a_gs_ms_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(
+            a_gs_ms_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(a_ms_ks_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(a_ms_ks_lengths, kDimIds);
+
+        if constexpr(ASpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto a_grid_desc_mraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(M, K),
+                make_tuple(a_ms_ks_strides[Number<NumDimM - 1>{}],
+                           a_ms_ks_strides[Number<NumDimM + NumDimK - 1>{}]));
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+        else
+        {
+            // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
+            const auto a_grid_desc_ms_ks =
+                make_naive_tensor_descriptor(a_ms_ks_lengths, a_ms_ks_strides);
+
+            // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
+            const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
+                a_grid_desc_ms_ks,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
+                make_tuple(mDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+    }
+
+    // Assume: B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                        const std::vector<index_t>& b_gs_ns_ks_strides_vec)
+    {
+        assert(b_gs_ns_ks_lengths_vec.size() == NumDimG + NumDimN + NumDimK &&
+               b_gs_ns_ks_strides_vec.size() == NumDimG + NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(
+            b_gs_ns_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(
+            b_gs_ns_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
+
+        // lengths for N0, N1, ...
+        const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
+
+        if constexpr(BSpec == TensorSpecialization::Packed)
+        {
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto b_grid_desc_nraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(N, K),
+                make_tuple(b_ns_ks_strides[Number<NumDimN - 1>{}],
+                           b_ns_ks_strides[Number<NumDimN + NumDimK - 1>{}]));
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+        else
+        {
+            // naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
+            const auto b_grid_desc_ns_ks =
+                make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
+
+            // transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
+            const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
+                b_grid_desc_ns_ks,
+                make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
+                make_tuple(nDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto e_ms_ns_lengths = to_tuple(
+            e_gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(
+            e_gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(M, N),
+                make_tuple(e_ms_ns_strides[Number<NumDimM - 1>{}],
+                           e_ms_ns_strides[Number<NumDimM + NumDimN - 1>{}]));
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+        else
+        {
+            // naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_ms_ns =
+                make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
+
+            // transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+            const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_ms_ns,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+                make_tuple(mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_G_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                          const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto e_gs_ms_ns_lengths =
+            to_tuple(e_gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_gs_ms_ns_strides =
+            to_tuple(e_gs_ms_ns_strides_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for G0, G1, ...
+        constexpr auto gDimIds = typename arithmetic_sequence_gen<0, NumDimG, 1>::type{};
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds =
+            typename arithmetic_sequence_gen<NumDimG, NumDimG + NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<NumDimG + NumDimM,
+                                                                  NumDimG + NumDimM + NumDimN,
+                                                                  1>::type{};
+
+        // lengths for G0, G1, ...
+        const auto gLengths = get_container_subset(e_gs_ms_ns_lengths, gDimIds);
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_gs_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_gs_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto G = container_reduce(gLengths, math::multiplies{}, Number<1>{});
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_g_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(G, M, N),
+                make_tuple(e_gs_ms_ns_strides[Number<NumDimG - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+        else
+        {
+            // naive tensor E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_gs_ms_ns =
+                make_naive_tensor_descriptor(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+            // transformed tensor E[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
+            // N2 * ...]
+            const auto e_grid_desc_g_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_gs_ms_ns,
+                make_tuple(make_merge_transform(gLengths),
+                           make_merge_transform(mLengths),
+                           make_merge_transform(nLengths)),
+                make_tuple(gDimIds, mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                         ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    static auto MakeDsGridDescriptor_G_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_G_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                           ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K({}, {}));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K({}, {}));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
+
+    using DsGridDesc_G_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_G_M_N({}, {}))>;
+    using EGridDesc_G_M_N  = decltype(MakeEGridDescriptor_G_M_N({}, {}));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t batch_stride_A,
+                                       index_t batch_stride_B,
+                                       DsGridDesc_G_M_N ds_grid_desc_g_m_n,
+                                       EGridDesc_G_M_N e_grid_desc_g_m_n)
+            : batch_stride_A_(batch_stride_A),
+              batch_stride_B_(batch_stride_B),
+              ds_grid_desc_g_m_n_(ds_grid_desc_g_m_n),
+              e_grid_desc_g_m_n_(e_grid_desc_g_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(batch_stride_A_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(batch_stride_B_);
+        }
+
+        __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+        {
+            std::array<long_index_t, NumDTensor> ds_offset;
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                ds_offset[i] = static_cast<long_index_t>(g_idx) *
+                               ds_grid_desc_g_m_n_[i].CalculateOffset(make_multi_index(1, 0, 0));
+            });
+
+            return ds_offset;
+        }
+
+        __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) *
+                   e_grid_desc_g_m_n_.CalculateOffset(make_multi_index(1, 0, 0));
+        }
+
+        private:
+        index_t batch_stride_A_;
+        index_t batch_stride_B_;
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmSplitKMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        DsGridDesc_M_N,
+        EGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // GridwiseGemm
+    using GridwiseGemmAtomicAdd = GridwiseGemmSplitKMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::AtomicAdd,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        DsGridDesc_M_N,
+        EGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AKB_AK0_M_AK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(AGridDesc_M_K{}, 1))>;
+    using BGridDesc_BKB_BK0_N_BK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(BGridDesc_N_K{}, 1))>;
+
+    using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 const std::vector<index_t>& a_gs_ms_ns_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op,
+                 index_t split_k)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{
+                  DeviceOp::MakeAGridDescriptor_M_K(a_gs_ms_ns_lengths, a_gs_ms_ks_strides)},
+              b_grid_desc_n_k_{
+                  DeviceOp::MakeBGridDescriptor_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{
+                  DeviceOp::MakeEGridDescriptor_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides)},
+              ds_grid_desc_g_m_n_{
+                  DeviceOp::MakeDsGridDescriptor_G_M_N(ds_gs_ms_ns_lengths, ds_gs_ms_ns_strides)},
+              e_grid_desc_g_m_n_{
+                  DeviceOp::MakeEGridDescriptor_G_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides)},
+              a_grid_desc_akb_ak0_m_ak1_{GridwiseGemm::MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(
+                  a_grid_desc_m_k_, split_k)},
+              b_grid_desc_bkb_bk0_n_bk1_{GridwiseGemm::MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(
+                  b_grid_desc_n_k_, split_k)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{
+                  GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_, split_k)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_mz_stride_{},
+              a_kz_stride_{},
+              b_nz_stride_{},
+              b_kz_stride_{},
+              ds_nz_stride_{},
+              e_nz_stride_{},
+              a_batch_stride_{a_gs_ms_ks_strides[NumDimG - 1]},
+              b_batch_stride_{b_gs_ns_ks_strides[NumDimG - 1]},
+              compute_ptr_offset_of_batch_{
+                  a_batch_stride_, b_batch_stride_, ds_grid_desc_g_m_n_, e_grid_desc_g_m_n_},
+              split_k_{split_k}
+        {
+            static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0, "");
+
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths[i],
+                                                                         ds_gs_ms_ns_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_akb_ak0_m_ak1_,
+                                           b_grid_desc_bkb_bk0_n_bk1_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+            }
+
+            // for sanity check of vector memory access
+            a_mz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM - 1];
+            a_kz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1];
+            b_nz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN - 1];
+            b_kz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1];
+
+            for(index_t i = 0; i < NumDTensor; ++i)
+            {
+                ds_nz_stride_[i] = ds_gs_ms_ns_strides[i][NumDimG + NumDimM + NumDimN - 1];
+            }
+
+            e_nz_stride_ = e_gs_ms_ns_strides[NumDimG + NumDimM + NumDimN - 1];
+
+            Print();
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_.GetLength(I0) << ", "
+                      << a_grid_desc_m_k_.GetLength(I1) << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_.GetLength(I0) << ", "
+                      << b_grid_desc_n_k_.GetLength(I1) << std::endl;
+
+            std::cout << "A[akb, ak0, m, ak1]: " << a_grid_desc_akb_ak0_m_ak1_.GetLength(I0) << ", "
+                      << a_grid_desc_akb_ak0_m_ak1_.GetLength(I1) << ", "
+                      << a_grid_desc_akb_ak0_m_ak1_.GetLength(I2) << ", "
+                      << a_grid_desc_akb_ak0_m_ak1_.GetLength(I3) << std::endl;
+            std::cout << "B[bkb, bk0, n, bk1]: " << b_grid_desc_bkb_bk0_n_bk1_.GetLength(I0) << ", "
+                      << b_grid_desc_bkb_bk0_n_bk1_.GetLength(I1) << ", "
+                      << b_grid_desc_bkb_bk0_n_bk1_.GetLength(I2) << ", "
+                      << b_grid_desc_bkb_bk0_n_bk1_.GetLength(I3) << std::endl;
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i].GetLength(I0) << ", "
+                          << ds_grid_desc_m_n_[i].GetLength(I1) << std::endl;
+            });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_.GetLength(I0) << ", "
+                      << e_grid_desc_m_n_.GetLength(I1) << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AKB_AK0_M_AK1 a_grid_desc_akb_ak0_m_ak1_;
+        BGridDesc_BKB_BK0_N_BK1 b_grid_desc_bkb_bk0_n_bk1_;
+        typename GridwiseGemm::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // Strides for the last M/N/K dimensions of A/B/Ds/E
+        //   for sanity check of vector load/store
+        index_t a_mz_stride_;
+        index_t a_kz_stride_;
+        index_t b_nz_stride_;
+        index_t b_kz_stride_;
+        std::array<index_t, NumDTensor> ds_nz_stride_;
+        index_t e_mz_stride_;
+        index_t e_nz_stride_;
+
+        index_t a_batch_stride_;
+        index_t b_batch_stride_;
+
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        index_t split_k_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_akb_ak0_m_ak1_,
+                                            arg.b_grid_desc_bkb_bk0_n_bk1_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
+            }
+
+            const index_t G = arg.e_grid_desc_g_m_n_.GetLength(I0);
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * G;
+
+            const auto K = arg.a_grid_desc_akb_ak0_m_ak1_.GetLength(I1) *
+                           arg.a_grid_desc_akb_ak0_m_ak1_.GetLength(I3);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_contraction_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AKB_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BKB_BK0_N_BK1,
+                    typename GridwiseGemm::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    ComputePtrOffsetOfStridedBatch,
+                    typename GridwiseGemm::DefaultBlock2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              G,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_akb_ak0_m_ak1_,
+                                              arg.b_grid_desc_bkb_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_etile_map_);
+            };
+
+            auto launch_kernel_atomic_add = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_contraction_multiple_d_xdl_cshuffle<
+                    GridwiseGemmAtomicAdd,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemmAtomicAdd::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AKB_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BKB_BK0_N_BK1,
+                    typename GridwiseGemmAtomicAdd::
+                        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemmAtomicAdd::
+                        EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    ComputePtrOffsetOfStridedBatch,
+                    typename GridwiseGemmAtomicAdd::DefaultBlock2ETileMap,
+                    has_main_loop>;
+
+                hipGetErrorString(hipMemset(
+                    arg.p_e_grid_,
+                    0,
+                    arg.e_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
+                        sizeof(EDataType)));
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              G,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_akb_ak0_m_ak1_,
+                                              arg.b_grid_desc_bkb_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                if(arg.split_k_ <= 1)
+                    return launch_kernel(integral_constant<bool, true>{});
+                else
+                    return launch_kernel_atomic_add(integral_constant<bool, true>{});
+            }
+            else
+            {
+                if(arg.split_k_ <= 1)
+                    return launch_kernel(integral_constant<bool, false>{});
+                else
+                    return launch_kernel_atomic_add(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_akb_ak0_m_ak1_,
+                                        arg.b_grid_desc_bkb_bk0_n_bk1_,
+                                        arg.ds_grid_desc_m_n_,
+                                        arg.e_grid_desc_m_n_,
+                                        arg.block_2_etile_map_))
+        {
+            return false;
+        }
+
+        // check vector access
+        static_assert((ABlockTransferSrcVectorDim == 2 || ABlockTransferSrcVectorDim == 3) &&
+                          (BBlockTransferSrcVectorDim == 2 || BBlockTransferSrcVectorDim == 3),
+                      "wrong!");
+
+        // vector memory access of A: could be on M or AK1 dimension
+        if constexpr(ABlockTransferSrcVectorDim == 2)
+        {
+            if(!(arg.a_mz_stride_ == 1 &&
+                 arg.a_grid_desc_akb_ak0_m_ak1_.GetLength(I2) % ABlockTransferSrcScalarPerVector ==
+                     0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.a_kz_stride_ == 1 &&
+                 arg.a_grid_desc_akb_ak0_m_ak1_.GetLength(I3) % ABlockTransferSrcScalarPerVector ==
+                     0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of B: could be on N or BK1 dimension
+        if constexpr(BBlockTransferSrcVectorDim == 2)
+        {
+            if(!(arg.b_nz_stride_ == 1 &&
+                 arg.b_grid_desc_bkb_bk0_n_bk1_.GetLength(I2) % BBlockTransferSrcScalarPerVector ==
+                     0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.b_kz_stride_ == 1 &&
+                 arg.b_grid_desc_bkb_bk0_n_bk1_.GetLength(I3) % BBlockTransferSrcScalarPerVector ==
+                     0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of Ds: always on NPerBlock dimension
+        bool valid_d_access = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            if(!(arg.ds_nz_stride_[i] == 1 &&
+                 arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_[i].GetLength(I3) %
+                         CDEBlockTransferScalarPerVector_NPerBlock ==
+                     0))
+            {
+                valid_d_access = false;
+            }
+        });
+
+        if(valid_d_access == false)
+        {
+            return false;
+        }
+
+        // vector memory access of E: always on NPerBlock dimension
+        if(!((arg.e_nz_stride_ == 1 &&
+              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
+                      CDEBlockTransferScalarPerVector_NPerBlock ==
+                  0) ||
+             CDEBlockTransferScalarPerVector_NPerBlock == 1))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto
+    MakeArgument(const void* p_a,
+                 const void* p_b,
+                 std::array<const void*, NumDTensor> p_ds,
+                 void* p_e,
+                 const std::vector<index_t>& a_gs_ms_ns_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op,
+                 index_t split_k)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_gs_ms_ns_lengths,
+                        a_gs_ms_ks_strides,
+                        b_gs_ns_ks_lengths,
+                        b_gs_ns_ks_strides,
+                        ds_gs_ms_ns_lengths,
+                        ds_gs_ms_ns_strides,
+                        e_gs_ms_ns_lengths,
+                        e_gs_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op,
+                        split_k};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_gs_ms_ns_lengths,
+                        const std::vector<index_t>& a_gs_ms_ks_strides,
+                        const std::vector<index_t>& b_gs_ns_ks_lengths,
+                        const std::vector<index_t>& b_gs_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                        const std::vector<index_t>& e_gs_ms_ns_lengths,
+                        const std::vector<index_t>& e_gs_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op,
+                        index_t split_k) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_gs_ms_ns_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b_gs_ns_ks_lengths,
+                                          b_gs_ns_ks_strides,
+                                          ds_gs_ms_ns_lengths,
+                                          ds_gs_ms_ns_strides,
+                                          e_gs_ms_ns_lengths,
+                                          e_gs_ms_ns_strides,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op,
+                                          split_k);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceSplitKContractionMultipleD_Xdl_CShuffle"
+            << "<"
+            << NumDimG << ", "
+            << NumDimM << ", "
+            << NumDimN << ", "
+            << NumDimK << ", "
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << ABlockTransferSrcVectorDim << ", "
+            << BBlockTransferSrcVectorDim
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/gemm_specialization.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
new file mode 100644
index 0000000..fc913e9
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct GemmSpecialization
+{
+    // Gemm
+    Default,
+    MPadding,
+    NPadding,
+    KPadding,
+    MNPadding,
+    MKPadding,
+    NKPadding,
+    MNKPadding,
+    // Gemm + Gemm
+    OPadding,
+    MOPadding,
+    NOPadding,
+    KOPadding,
+    MNOPadding,
+    MKOPadding,
+    NKOPadding,
+    MNKOPadding,
+};
+
+inline std::string getGemmSpecializationString(const GemmSpecialization& s)
+{
+    switch(s)
+    {
+    case GemmSpecialization::Default: return "Default";
+    case GemmSpecialization::MPadding: return "MPadding";
+    case GemmSpecialization::NPadding: return "NPadding";
+    case GemmSpecialization::KPadding: return "KPadding";
+    case GemmSpecialization::MNPadding: return "MNPadding";
+    case GemmSpecialization::MKPadding: return "MKPadding";
+    case GemmSpecialization::NKPadding: return "NKPadding";
+    case GemmSpecialization::MNKPadding: return "MNKPadding";
+    case GemmSpecialization::OPadding: return "OPadding";
+    case GemmSpecialization::MOPadding: return "MOPadding";
+    case GemmSpecialization::NOPadding: return "NOPadding";
+    case GemmSpecialization::KOPadding: return "KOPadding";
+    case GemmSpecialization::MNOPadding: return "MNOPadding";
+    case GemmSpecialization::MKOPadding: return "MKOPadding";
+    case GemmSpecialization::NKOPadding: return "NKOPadding";
+    case GemmSpecialization::MNKOPadding: return "MNKOPadding";
+    default: return "Unrecognized specialization!";
+    }
+}
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000..2237ad9
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,1040 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatDsPointer,
+          typename FloatE,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_contraction_multiple_d_xdl_cshuffle(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatDsPointer p_ds_grid,
+            FloatE* __restrict__ p_e_grid,
+            const index_t batch_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    FloatDsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+    ignore = compute_ptr_offset_of_batch;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+
+// NOTE: TensorSpecialization::Packed specialized tensor is "packed" in a sense that each inner
+// dimension in a dimension group (eg [G0, G1] in Gs, [M0, M1, M2] in Ms, etc.) are contiguous and
+// ordered. Not in a sense that the tensor [G0, G1, ..., M0, M1, ..., N0, N1...] can be permuted
+// while still being a contiguous, unpadded tensor. In other words, it merely degenerates into
+// TensorSpecialization::Default with NumDimG/M/N/K = 1
+//
+// Detail- Packed tensor satisfies
+//   stride_0 = 1
+//   stride_i = stride_{i - 1} * extent_{i - 1}
+// So tensor
+//   [G0, G1, G2, M, N]
+// transposed into tensor
+//   [G0, G2, G1, M, N]
+// with strides
+//   [G2 * G1 * M * N, G1 * M * N, M * N, N, 1]
+// is again a packed tensor. MakeGridDescriptor() currently just merges dimensions and ignores some
+// strides from input tensor extents so finer dimension information is lost. Merging dimensions is
+// essentially a degenerated case of TensorSpecialization::Default with NumDimG/M/N/K = 1.
+//
+// Might need to expose dimension order to the interface to fully support
+// TensorSpecialization::Packed in a traditional sense of "packed" tensor
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization DESpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
+    : public DeviceBatchedContractionMultipleD<NumDimG,
+                                               NumDimM,
+                                               NumDimN,
+                                               NumDimK,
+                                               ADataType,
+                                               BDataType,
+                                               DsDataType,
+                                               EDataType,
+                                               AElementwiseOperation,
+                                               BElementwiseOperation,
+                                               CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceBatchedContractionMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    // Assume: A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        assert(a_gs_ms_ks_lengths_vec.size() == NumDimG + NumDimM + NumDimK &&
+               a_gs_ms_ks_strides_vec.size() == NumDimG + NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto a_ms_ks_lengths = to_tuple(
+            a_gs_ms_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(
+            a_gs_ms_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(a_ms_ks_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(a_ms_ks_lengths, kDimIds);
+
+        if constexpr(ASpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto a_grid_desc_mraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(M, K),
+                make_tuple(a_ms_ks_strides[Number<NumDimM - 1>{}],
+                           a_ms_ks_strides[Number<NumDimM + NumDimK - 1>{}]));
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+        else
+        {
+            // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
+            const auto a_grid_desc_ms_ks =
+                make_naive_tensor_descriptor(a_ms_ks_lengths, a_ms_ks_strides);
+
+            // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
+            const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
+                a_grid_desc_ms_ks,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
+                make_tuple(mDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+    }
+
+    // Assume: B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                        const std::vector<index_t>& b_gs_ns_ks_strides_vec)
+    {
+        assert(b_gs_ns_ks_lengths_vec.size() == NumDimG + NumDimN + NumDimK &&
+               b_gs_ns_ks_strides_vec.size() == NumDimG + NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(
+            b_gs_ns_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(
+            b_gs_ns_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
+
+        // lengths for N0, N1, ...
+        const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
+
+        if constexpr(BSpec == TensorSpecialization::Packed)
+        {
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto b_grid_desc_nraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(N, K),
+                make_tuple(b_ns_ks_strides[Number<NumDimN - 1>{}],
+                           b_ns_ks_strides[Number<NumDimN + NumDimK - 1>{}]));
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+        else
+        {
+            // naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
+            const auto b_grid_desc_ns_ks =
+                make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
+
+            // transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
+            const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
+                b_grid_desc_ns_ks,
+                make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
+                make_tuple(nDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto e_ms_ns_lengths = to_tuple(
+            e_gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(
+            e_gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(M, N),
+                make_tuple(e_ms_ns_strides[Number<NumDimM - 1>{}],
+                           e_ms_ns_strides[Number<NumDimM + NumDimN - 1>{}]));
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+        else
+        {
+            // naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_ms_ns =
+                make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
+
+            // transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+            const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_ms_ns,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+                make_tuple(mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_G_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                          const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto e_gs_ms_ns_lengths =
+            to_tuple(e_gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_gs_ms_ns_strides =
+            to_tuple(e_gs_ms_ns_strides_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for G0, G1, ...
+        constexpr auto gDimIds = typename arithmetic_sequence_gen<0, NumDimG, 1>::type{};
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds =
+            typename arithmetic_sequence_gen<NumDimG, NumDimG + NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<NumDimG + NumDimM,
+                                                                  NumDimG + NumDimM + NumDimN,
+                                                                  1>::type{};
+
+        // lengths for G0, G1, ...
+        const auto gLengths = get_container_subset(e_gs_ms_ns_lengths, gDimIds);
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_gs_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_gs_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto G = container_reduce(gLengths, math::multiplies{}, Number<1>{});
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_g_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(G, M, N),
+                make_tuple(e_gs_ms_ns_strides[Number<NumDimG - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+        else
+        {
+            // naive tensor E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_gs_ms_ns =
+                make_naive_tensor_descriptor(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+            // transformed tensor E[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
+            // N2 * ...]
+            const auto e_grid_desc_g_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_gs_ms_ns,
+                make_tuple(make_merge_transform(gLengths),
+                           make_merge_transform(mLengths),
+                           make_merge_transform(nLengths)),
+                make_tuple(gDimIds, mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                         ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    static auto MakeDsGridDescriptor_G_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_G_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                           ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K({}, {}));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K({}, {}));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
+
+    using DsGridDesc_G_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_G_M_N({}, {}))>;
+    using EGridDesc_G_M_N  = decltype(MakeEGridDescriptor_G_M_N({}, {}));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t batch_stride_A,
+                                       index_t batch_stride_B,
+                                       DsGridDesc_G_M_N ds_grid_desc_g_m_n,
+                                       EGridDesc_G_M_N e_grid_desc_g_m_n)
+            : batch_stride_A_(batch_stride_A),
+              batch_stride_B_(batch_stride_B),
+              ds_grid_desc_g_m_n_(ds_grid_desc_g_m_n),
+              e_grid_desc_g_m_n_(e_grid_desc_g_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) * batch_stride_A_;
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) * batch_stride_B_;
+        }
+
+        __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+        {
+            std::array<long_index_t, NumDTensor> ds_offset;
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                ds_offset[i] = static_cast<long_index_t>(g_idx) *
+                               ds_grid_desc_g_m_n_[i].CalculateOffset(make_multi_index(1, 0, 0));
+            });
+
+            return ds_offset;
+        }
+
+        __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) *
+                   e_grid_desc_g_m_n_.CalculateOffset(make_multi_index(1, 0, 0));
+        }
+
+        private:
+        index_t batch_stride_A_;
+        index_t batch_stride_B_;
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 const std::vector<index_t>& a_gs_ms_ns_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{
+                  DeviceOp::MakeAGridDescriptor_M_K(a_gs_ms_ns_lengths, a_gs_ms_ks_strides)},
+              b_grid_desc_n_k_{
+                  DeviceOp::MakeBGridDescriptor_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{
+                  DeviceOp::MakeEGridDescriptor_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides)},
+              ds_grid_desc_g_m_n_{
+                  DeviceOp::MakeDsGridDescriptor_G_M_N(ds_gs_ms_ns_lengths, ds_gs_ms_ns_strides)},
+              e_grid_desc_g_m_n_{
+                  DeviceOp::MakeEGridDescriptor_G_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_mz_stride_{},
+              a_kz_stride_{},
+              b_nz_stride_{},
+              b_kz_stride_{},
+              ds_nz_stride_{},
+              e_nz_stride_{},
+              a_batch_stride_{a_gs_ms_ks_strides[NumDimG - 1]},
+              b_batch_stride_{b_gs_ns_ks_strides[NumDimG - 1]},
+              compute_ptr_offset_of_batch_{
+                  a_batch_stride_, b_batch_stride_, ds_grid_desc_g_m_n_, e_grid_desc_g_m_n_}
+        {
+            static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0, "");
+
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths[i],
+                                                                         ds_gs_ms_ns_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+            }
+
+            // for sanity check of vector memory access
+            a_mz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM - 1];
+            a_kz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1];
+            b_nz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN - 1];
+            b_kz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1];
+
+            for(index_t i = 0; i < NumDTensor; ++i)
+            {
+                ds_nz_stride_[i] = ds_gs_ms_ns_strides[i][NumDimG + NumDimM + NumDimN - 1];
+            }
+
+            e_nz_stride_ = e_gs_ms_ns_strides[NumDimG + NumDimM + NumDimN - 1];
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // Strides for the last M/N/K dimensions of A/B/Ds/E
+        //   for sanity check of vector load/store
+        index_t a_mz_stride_;
+        index_t a_kz_stride_;
+        index_t b_nz_stride_;
+        index_t b_kz_stride_;
+        std::array<index_t, NumDTensor> ds_nz_stride_;
+        index_t e_mz_stride_;
+        index_t e_nz_stride_;
+
+        index_t a_batch_stride_;
+        index_t b_batch_stride_;
+
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
+            }
+
+            const index_t G = arg.e_grid_desc_g_m_n_.GetLength(I0);
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * G;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_contraction_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    ComputePtrOffsetOfStridedBatch,
+                    DeviceOp::Block2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              G,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                        arg.b_grid_desc_n_k_,
+                                        arg.ds_grid_desc_m_n_,
+                                        arg.e_grid_desc_m_n_,
+                                        arg.block_2_etile_map_))
+        {
+            return false;
+        }
+
+        // check vector access
+        static_assert((ABlockTransferSrcVectorDim == 1 || ABlockTransferSrcVectorDim == 2) &&
+                          (BBlockTransferSrcVectorDim == 1 || BBlockTransferSrcVectorDim == 2),
+                      "wrong!");
+
+        // vector memory access of A: could be on M or AK1 dimension
+        if constexpr(ABlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.a_mz_stride_ == 1 &&
+                 arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.a_kz_stride_ == 1 &&
+                 arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of B: could be on N or BK1 dimension
+        if constexpr(BBlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.b_nz_stride_ == 1 &&
+                 arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.b_kz_stride_ == 1 &&
+                 arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of Ds: always on NPerBlock dimension
+        bool valid_d_access = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            if(!(arg.ds_nz_stride_[i] == 1 &&
+                 arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_[i].GetLength(I3) %
+                         CDEBlockTransferScalarPerVector_NPerBlock ==
+                     0))
+            {
+                valid_d_access = false;
+            }
+        });
+
+        if(valid_d_access == false)
+        {
+            return false;
+        }
+
+        // vector memory access of E: always on NPerBlock dimension
+        if(!((arg.e_nz_stride_ == 1 &&
+              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
+                      CDEBlockTransferScalarPerVector_NPerBlock ==
+                  0) ||
+             CDEBlockTransferScalarPerVector_NPerBlock == 1))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto
+    MakeArgument(const void* p_a,
+                 const void* p_b,
+                 std::array<const void*, NumDTensor> p_ds,
+                 void* p_e,
+                 const std::vector<index_t>& a_gs_ms_ns_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_gs_ms_ns_lengths,
+                        a_gs_ms_ks_strides,
+                        b_gs_ns_ks_lengths,
+                        b_gs_ns_ks_strides,
+                        ds_gs_ms_ns_lengths,
+                        ds_gs_ms_ns_strides,
+                        e_gs_ms_ns_lengths,
+                        e_gs_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_gs_ms_ns_lengths,
+                        const std::vector<index_t>& a_gs_ms_ks_strides,
+                        const std::vector<index_t>& b_gs_ns_ks_lengths,
+                        const std::vector<index_t>& b_gs_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                        const std::vector<index_t>& e_gs_ms_ns_lengths,
+                        const std::vector<index_t>& e_gs_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_gs_ms_ns_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b_gs_ns_ks_lengths,
+                                          b_gs_ns_ks_strides,
+                                          ds_gs_ms_ns_lengths,
+                                          ds_gs_ms_ns_strides,
+                                          e_gs_ms_ns_lengths,
+                                          e_gs_ms_ns_strides,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedContractionMultipleD_Xdl_CShuffle"
+            << "<"
+            << NumDimG << ", "
+            << NumDimM << ", "
+            << NumDimN << ", "
+            << NumDimK << ", "
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << ABlockTransferSrcVectorDim << ", "
+            << BBlockTransferSrcVectorDim
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
new file mode 100644
index 0000000..01f5e17
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
@@ -0,0 +1,683 @@
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
+\link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemmCPermute and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename EDataType,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_e_permute_xdl(const ABDataType* __restrict__ p_a_grid,
+                                          const ABDataType* __restrict__ p_b_grid,
+                                          EDataType* __restrict__ p_e_grid,
+                                          const index_t batch_count,
+                                          const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+                                          const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+                                          const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                                              e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                          const AElementwiseOperation a_element_op,
+                                          const BElementwiseOperation b_element_op,
+                                          const CDEElementwiseOperation cde_element_op,
+                                          const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+                                          const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  ck::Tuple<>{},
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ck::Tuple<>{},
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_etile_map;
+#endif
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumPrefetch,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<ALayout,
+                                                                       BLayout,
+                                                                       ELayout,
+                                                                       ADataType,
+                                                                       BDataType,
+                                                                       EDataType,
+                                                                       AElementwiseOperation,
+                                                                       BElementwiseOperation,
+                                                                       CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceBatchedGemmEPermuteXdl;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t stride_M, index_t stride_N)
+    {
+        const auto e_grid_desc_mraw_nraw =
+            make_naive_tensor_descriptor(make_tuple(MRaw, NRaw), make_tuple(stride_M, stride_N));
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeEGridDescriptor_G0_G1_M_N(index_t G0,
+                                              index_t G1,
+                                              index_t MRaw,
+                                              index_t NRaw,
+                                              index_t stride_G0,
+                                              index_t stride_G1,
+                                              index_t stride_M,
+                                              index_t stride_N)
+    {
+        const auto e_grid_desc_g0_g1_mraw_nraw = [&]() {
+            return make_naive_tensor_descriptor(
+                make_tuple(G0, G1, MRaw, NRaw),
+                make_tuple(stride_G0, stride_G1, stride_M, stride_N));
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(
+                e_grid_desc_g0_g1_mraw_nraw,
+                make_tuple(make_pass_through_transform(G0),
+                           make_pass_through_transform(G1),
+                           make_right_pad_transform(MRaw, MPad),
+                           make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                e_grid_desc_g0_g1_mraw_nraw,
+                make_tuple(make_pass_through_transform(G0),
+                           make_pass_through_transform(G1),
+                           make_right_pad_transform(MRaw, MPad),
+                           make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                e_grid_desc_g0_g1_mraw_nraw,
+                make_tuple(make_pass_through_transform(G0),
+                           make_pass_through_transform(G1),
+                           make_pass_through_transform(MRaw),
+                           make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return e_grid_desc_g0_g1_mraw_nraw;
+        }
+    }
+
+    using AGridDesc_M_K       = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K       = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using EGridDesc_M_N       = decltype(MakeEGridDescriptor_M_N(1, 1, 1, 1));
+    using EGridDesc_G0_G1_M_N = decltype(MakeEGridDescriptor_G0_G1_M_N(1, 1, 1, 1, 1, 1, 1, 1));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t Batchstride_A,
+                                       index_t Batchstride_B,
+                                       EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n)
+            : Batchstride_A_(Batchstride_A),
+              Batchstride_B_(Batchstride_B),
+              e_grid_desc_g0_g1_m_n_(e_grid_desc_g0_g1_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(Batchstride_A_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(Batchstride_B_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
+        {
+            const index_t G1 = e_grid_desc_g0_g1_m_n_.GetLength(I1);
+            index_t b0       = g_idx / G1;
+            index_t b1       = g_idx - b0 * G1; // g_idx % G1
+            return e_grid_desc_g0_g1_m_n_.CalculateOffset(make_multi_index(b0, b1, 0, 0));
+        }
+
+        private:
+        index_t Batchstride_A_;
+        index_t Batchstride_B_;
+        EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n_;
+    };
+
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>, // DsDataType,
+        EDataType,   // EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        Tuple<>,
+        EGridDesc_M_N,
+        NumPrefetch,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = decltype(
+        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}));
+    using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 EDataType* p_e_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t stride_A,
+                 index_t stride_B,
+                 index_t batch_stride_A,
+                 index_t batch_stride_B,
+                 BatchedGemmEPermuteDesc batched_gemm_e_permute_desc,
+                 index_t BatchCount,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_e_grid_{p_e_grid},
+              BatchCount_(BatchCount),
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(M, K, stride_A)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(K, N, stride_B)},
+              e_grid_desc_m_n_{
+                  DeviceOp::MakeEGridDescriptor_M_N(batched_gemm_e_permute_desc.M_,
+                                                    batched_gemm_e_permute_desc.N_,
+                                                    batched_gemm_e_permute_desc.stride_M_,
+                                                    batched_gemm_e_permute_desc.stride_N_)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              e_grid_desc_mblock_mperblock_nblock_nperblock{},
+              e_grid_desc_g0_g1_m_n_{
+                  DeviceOp::MakeEGridDescriptor_G0_G1_M_N(batched_gemm_e_permute_desc.G0_,
+                                                          batched_gemm_e_permute_desc.G1_,
+                                                          batched_gemm_e_permute_desc.M_,
+                                                          batched_gemm_e_permute_desc.N_,
+                                                          batched_gemm_e_permute_desc.stride_G0_,
+                                                          batched_gemm_e_permute_desc.stride_G1_,
+                                                          batched_gemm_e_permute_desc.stride_M_,
+                                                          batched_gemm_e_permute_desc.stride_N_)},
+              compute_ptr_offset_of_batch_{batch_stride_A, batch_stride_B, e_grid_desc_g0_g1_m_n_},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ck::Tuple<>{},
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            std::cout << "C[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        EDataType* p_e_grid_;
+
+        // batch count
+        index_t BatchCount_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock;
+        EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n_;
+
+        // for calculating Batch offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            ck::Tuple<>{},
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseBatchedGemmCPermute_km_kn_m0m1n0n1_xdlops_v2r3 has invalid "
+                    "setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.BatchCount_;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_batched_gemm_e_permute_xdl<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    EDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_AK0_M_AK1>,
+                    remove_reference_t<DeviceOp::BGridDesc_BK0_N_BK1>,
+                    typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    ComputePtrOffsetOfStridedBatch,
+                    remove_reference_t<Block2ETileMap>,
+                    has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_e_grid_,
+                                              arg.BatchCount_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           ck::Tuple<>{},
+                                           arg.e_grid_desc_m_n_,
+                                           arg.block_2_etile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             EDataType* p_e,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t stride_A,
+                             index_t stride_B,
+                             index_t batch_stride_A,
+                             index_t batch_stride_B,
+                             BatchedGemmEPermuteDesc batched_gemm_e_permute_desc,
+                             index_t BatchCount,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_e,
+                        M,
+                        N,
+                        K,
+                        stride_A,
+                        stride_B,
+                        batch_stride_A,
+                        batch_stride_B,
+                        batched_gemm_e_permute_desc,
+                        BatchCount,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_e,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t stride_A,
+                        index_t stride_B,
+                        index_t batch_stride_A,
+                        index_t batch_stride_B,
+                        BatchedGemmEPermuteDesc batched_gemm_e_permute_desc,
+                        index_t BatchCount,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<EDataType*>(p_e),
+                                          M,
+                                          N,
+                                          K,
+                                          stride_A,
+                                          stride_B,
+                                          batch_stride_A,
+                                          batch_stride_B,
+                                          batched_gemm_e_permute_desc,
+                                          BatchCount,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmEPermuteXdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
new file mode 100644
index 0000000..3b87e56
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
@@ -0,0 +1,747 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2CTileMap,
+          typename ComputeBasePtrOfStridedBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_gemm_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            const FloatAB* __restrict__ p_b1_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const AccElementwiseOperation acc_element_op,
+            const B1ElementwiseOperation b1_element_op,
+            const CElementwiseOperation c_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2CTileMap block_2_ctile_map,
+            const index_t batch_count,
+            const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetBBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetCBasePtr(g_idx)));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_b1_grid + b1_batch_offset,
+                                                  p_c_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  acc_element_op,
+                                                  b1_element_op,
+                                                  c_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  b1_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_b1_grid;
+    ignore = p_c_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = acc_element_op;
+    ignore = b1_element_op;
+    ignore = c_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = b1_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_ctile_map;
+    ignore = batch_count;
+    ignore = compute_base_ptr_of_batch;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// Computes C = A * B0 * B1
+//              ^^^^^^ (Acc0)
+//              ^^^^^^^^^^^ (Acc1)
+template <typename ALayout,
+          typename BLayout, // B0Layout
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename B1DataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock, // Gemm0NPerBlock
+          index_t KPerBlock, // Gemm0KPerBlock
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t B1K1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout,
+                                                                         BLayout,
+                                                                         B1Layout,
+                                                                         CLayout,
+                                                                         ADataType,
+                                                                         BDataType,
+                                                                         B1DataType,
+                                                                         CDataType,
+                                                                         AElementwiseOperation,
+                                                                         BElementwiseOperation,
+                                                                         AccElementwiseOperation,
+                                                                         B1ElementwiseOperation,
+                                                                         CElementwiseOperation>
+{
+    using DeviceOp = DeviceBatchedGemmGemm_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto matrix_padder =
+        GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
+            MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // Args: Gemm1KRaw, Gemm1NRaw, StrideB1
+    static auto MakeB1GridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b1_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b1_grid_desc_n_k = matrix_padder.PadB1Descriptor_N_K(b1_grid_desc_nraw_kraw);
+
+        const auto N = b1_grid_desc_n_k.GetLength(I0);
+        const auto K = b1_grid_desc_n_k.GetLength(I1);
+
+        const auto B1K0 = K / B1K1;
+
+        return transform_tensor_descriptor(
+            b1_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
+    }
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
+                                     index_t BatchStrideB,
+                                     index_t BatchStrideB1,
+                                     index_t BatchStrideC)
+            : BatchStrideA_(BatchStrideA),
+              BatchStrideB_(BatchStrideB),
+              BatchStrideB1_(BatchStrideB1),
+              BatchStrideC_(BatchStrideC)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB1_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideC_);
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        index_t BatchStrideB1_;
+        index_t BatchStrideC_;
+    };
+
+    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N        = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedGemmGemm_Xdl_CShuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        AccElementwiseOperation,
+        B1ElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        B1GridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        AK1,
+        BK1,
+        B1K1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        Gemm1NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        true,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        true,
+        BBlockLdsExtraN,
+        B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B1BlockTransferThreadClusterArrangeOrder,
+        B1BlockTransferSrcAccessOrder,
+        B1BlockTransferSrcVectorDim,
+        B1BlockTransferSrcScalarPerVector,
+        B1BlockTransferDstScalarPerVector_BK1,
+        false,
+        B1BlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 const B1DataType* p_b1_grid,
+                 CDataType* p_c_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t Gemm1NRaw, // = ORaw
+                 index_t Batch,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideB1,
+                 index_t StrideC,
+                 index_t BatchStrideA,
+                 index_t BatchStrideB,
+                 index_t BatchStrideB1,
+                 index_t BatchStrideC,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 AccElementwiseOperation acc_element_op,
+                 B1ElementwiseOperation b1_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_b1_grid_{p_b1_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              b1_grid_desc_bk0_n_bk1_{
+                  DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(NRaw, Gemm1NRaw, StrideB1)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, Gemm1NRaw, StrideC)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              b1_element_op_{b1_element_op},
+              c_element_op_{c_element_op},
+              batch_count_(Batch),
+              compute_base_ptr_of_batch_{BatchStrideA, BatchStrideB, BatchStrideB1, BatchStrideC},
+              raw_lengths_m_n_k_o_{MRaw, NRaw, KRaw, Gemm1NRaw}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           b1_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[AK0, M, AK1]: " << a_grid_desc_ak0_m_ak1_ << std::endl;
+            std::cout << "B0[BK0, N, BK1]: " << b_grid_desc_bk0_n_bk1_ << std::endl;
+            std::cout << "B1[BK0, N, BK1]: " << b1_grid_desc_bk0_n_bk1_ << std::endl;
+            std::cout << "C[M, N]: " << c_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        const B1DataType* p_b1_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        CElementwiseOperation c_element_op_;
+        index_t batch_count_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+
+        // For robust IsSupportedArgument() check
+        std::vector<index_t> raw_lengths_m_n_k_o_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!DeviceOp::IsSupportedArgument(arg))
+            {
+                throw std::runtime_error("wrong! unsupported argument");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.batch_count_;
+
+            // Gemm0_K
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_gemm_gemm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    AccElementwiseOperation,
+                    B1ElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::B1GridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    ComputeBasePtrOfStridedBatch,
+                    has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_b1_grid_,
+                                              arg.p_c_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.acc_element_op_,
+                                              arg.b1_element_op_,
+                                              arg.c_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.b1_grid_desc_bk0_n_bk1_,
+                                              arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_ctile_map_,
+                                              arg.batch_count_,
+                                              arg.compute_base_ptr_of_batch_);
+            };
+
+            // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
+            // to concern Gemm0's loop
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        // Note: we need raw lengths since threadwise copy can not handle vector load when part of
+        // vector is out of bounds
+        const auto MRaw      = arg.raw_lengths_m_n_k_o_[0];
+        const auto NRaw      = arg.raw_lengths_m_n_k_o_[1];
+        const auto KRaw      = arg.raw_lengths_m_n_k_o_[2];
+        const auto Gemm1NRaw = arg.raw_lengths_m_n_k_o_[3];
+
+        // Check scalar per vector requirement
+        const auto a_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, ALayout> ? KRaw : MRaw;
+        const auto b_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, BLayout> ? NRaw : KRaw;
+        const auto b1_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, B1Layout> ? Gemm1NRaw : NRaw;
+        const auto c_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, CLayout> ? Gemm1NRaw : MRaw;
+
+        if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
+             b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
+             b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
+             c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.b1_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             const B1DataType* p_b1,
+                             CDataType* p_c,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t Gemm1NRaw,
+                             index_t Batch,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideB1,
+                             index_t StrideC,
+                             index_t BatchStrideA,
+                             index_t BatchStrideB,
+                             index_t BatchStrideB1,
+                             index_t BatchStrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             AccElementwiseOperation acc_element_op,
+                             B1ElementwiseOperation b1_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,           p_b,          p_b1,         p_c,          MRaw,
+                        NRaw,          KRaw,         Gemm1NRaw,    Batch,        StrideA,
+                        StrideB,       StrideB1,     StrideC,      BatchStrideA, BatchStrideB,
+                        BatchStrideB1, BatchStrideC, a_element_op, b_element_op, acc_element_op,
+                        b1_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      const void* p_b1,
+                                                      void* p_c,
+                                                      index_t MRaw,
+                                                      index_t NRaw,
+                                                      index_t KRaw,
+                                                      index_t Gemm1NRaw,
+                                                      index_t Batch,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideB1,
+                                                      index_t StrideC,
+                                                      index_t BatchStrideA,
+                                                      index_t BatchStrideB,
+                                                      index_t BatchStrideB1,
+                                                      index_t BatchStrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      AccElementwiseOperation acc_element_op,
+                                                      B1ElementwiseOperation b1_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<const B1DataType*>(p_b1),
+                                          static_cast<CDataType*>(p_c),
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          Gemm1NRaw,
+                                          Batch,
+                                          StrideA,
+                                          StrideB,
+                                          StrideB1,
+                                          StrideC,
+                                          BatchStrideA,
+                                          BatchStrideB,
+                                          BatchStrideB1,
+                                          BatchStrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          acc_element_op,
+                                          b1_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmGemm_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerBlock << ", "
+            << Gemm1NPerBlock << ", "
+            << Gemm1KPerBlock << ", "
+            << B1K1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
new file mode 100644
index 0000000..e34c19b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
@@ -0,0 +1,716 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
+ * \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the
+ * computing of pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_xdl(const ABDataType* __restrict__ p_a_grid,
+                                const ABDataType* __restrict__ p_b_grid,
+                                DsPointer p_ds_grid,
+                                EDataType* __restrict__ p_e_grid,
+                                const index_t batch_count,
+                                const AElementwiseOperation a_element_op,
+                                const BElementwiseOperation b_element_op,
+                                const CDEElementwiseOperation cde_element_op,
+                                const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
+                                const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
+                                const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                                    ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                                    e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+                                const Block2ETileMap block_2_etile_map)
+{
+
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_etile_map;
+#endif
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceBatchedGemmMultiD_Xdl : public DeviceBatchedGemmMultiD<ALayout,
+                                                                    BLayout,
+                                                                    DsLayout,
+                                                                    ELayout,
+                                                                    ADataType,
+                                                                    BDataType,
+                                                                    DsDataType,
+                                                                    EDataType,
+                                                                    AElementwiseOperation,
+                                                                    BElementwiseOperation,
+                                                                    CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceBatchedGemmMultiD_Xdl;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                       index_t BatchStrideB,
+                                       std::array<ck::index_t, NumDTensor> BatchStrideDs,
+                                       index_t BatchStrideE)
+            : BatchStrideA_(BatchStrideA),
+              BatchStrideB_(BatchStrideB),
+              BatchStrideDs_(BatchStrideDs),
+              BatchStrideE_(BatchStrideE)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+        }
+
+        __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+        {
+            std::array<long_index_t, NumDTensor> ds_offset;
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                ds_offset[i] = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]);
+            });
+            return ds_offset;
+        }
+
+        __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideE_);
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        std::array<ck::index_t, NumDTensor> BatchStrideDs_;
+        index_t BatchStrideE_;
+    };
+
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t Batch,
+                 index_t StrideA,
+                 index_t StrideB,
+                 const std::array<ck::index_t, NumDTensor>& StrideDs,
+                 index_t StrideE,
+                 index_t BatchStrideA,
+                 index_t BatchStrideB,
+                 const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                 index_t BatchStrideE,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              Batch_(Batch),
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(MRaw, NRaw, StrideE)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              compute_ptr_offset_of_batch_{BatchStrideA, BatchStrideB, BatchStrideDs, BatchStrideE},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            // populate pointer, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) =
+                    DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaw, NRaw, StrideDs[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // Batch
+        index_t Batch_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // for calculating batch offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceBatchedGemmMultiD_Xdl::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.Batch_;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel =
+                    kernel_batched_gemm_xdl<GridwiseGemm,
+                                            ADataType, // TODO: distiguish A/B datatype
+                                            typename GridwiseGemm::DsGridPointer,
+                                            EDataType,
+                                            AElementwiseOperation,
+                                            BElementwiseOperation,
+                                            CDEElementwiseOperation,
+                                            DeviceOp::AGridDesc_AK0_M_AK1,
+                                            DeviceOp::BGridDesc_BK0_N_BK1,
+                                            DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                                            DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                                            ComputePtrOffsetOfStridedBatch,
+                                            Block2ETileMap,
+                                            has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.Batch_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.ds_grid_desc_m_n_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.block_2_etile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             const std::array<const void*, NumDTensor>& p_ds,
+                             void* p_e,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t Batch,
+                             index_t StrideA,
+                             index_t StrideB,
+                             const std::array<index_t, NumDTensor>& StrideDs,
+                             index_t StrideE,
+                             index_t BatchStrideA,
+                             index_t BatchStrideB,
+                             const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                             index_t BatchStrideE,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        M,
+                        N,
+                        K,
+                        Batch,
+                        StrideA,
+                        StrideB,
+                        StrideDs,
+                        StrideE,
+                        BatchStrideA,
+                        BatchStrideB,
+                        BatchStrideDs,
+                        BatchStrideE,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const std::array<const void*, NumDTensor>& p_ds,
+                        void* p_e,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t Batch,
+                        index_t StrideA,
+                        index_t StrideB,
+                        const std::array<ck::index_t, NumDTensor>& StrideDs,
+                        index_t StrideE,
+                        index_t BatchStrideA,
+                        index_t BatchStrideB,
+                        const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                        index_t BatchStrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          M,
+                                          N,
+                                          K,
+                                          Batch,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideE,
+                                          BatchStrideA,
+                                          BatchStrideB,
+                                          BatchStrideDs,
+                                          BatchStrideE,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmMultiD_Xdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << getGemmSpecializationString(GemmSpec)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000..19e2649
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,951 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename A0B0B1DataType,
+          typename D0sPointer,
+          typename D1sPointer,
+          typename E1DataType,
+          typename A0ElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename CDE0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CDE1ElementwiseOperation,
+          typename A0GridDesc_AK0_M_AK1,
+          typename B0GridDesc_BK0_N_BK1,
+          typename D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
+          typename B1GridDesc_BK0_N_BK1,
+          typename D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2E1TileMap,
+          typename ComputeBasePtrOfStridedBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_gemm_xdl_cshuffle_v1(
+            const A0B0B1DataType* __restrict__ p_a0_grid,
+            const A0B0B1DataType* __restrict__ p_b0_grid,
+            D0sPointer p_d0s_grid,
+            const A0B0B1DataType* __restrict__ p_b1_grid,
+            D1sPointer p_d1s_grid,
+            E1DataType* __restrict__ p_e1_grid,
+            const A0ElementwiseOperation a0_element_op,
+            const B0ElementwiseOperation b0_element_op,
+            const CDE0ElementwiseOperation cde0_element_op,
+            const B1ElementwiseOperation b1_element_op,
+            const CDE1ElementwiseOperation cde1_element_op,
+            const A0GridDesc_AK0_M_AK1 a0_grid_desc_ak0_m_ak1,
+            const B0GridDesc_BK0_N_BK1 b0_grid_desc_bk0_n_bk1,
+            const D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
+                d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+            const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
+            const D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                d1s_grid_desc_mblock_mperblock_nblock_nperblock,
+            const E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e1_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2E1TileMap block_2_e1tile_map,
+            const index_t batch_count,
+            const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetBBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetCBasePtr(g_idx)));
+
+    static_for<0, p_d0s_grid.Size(), 1>{}([&](auto In) {
+        const long_index_t d0_batch_offset = __builtin_amdgcn_readfirstlane(
+            static_cast<long_index_t>(compute_base_ptr_of_batch.GetD0BasePtr(g_idx, In)));
+        p_d0s_grid(In) = p_d0s_grid(In) + d0_batch_offset;
+    });
+
+    static_for<0, p_d1s_grid.Size(), 1>{}([&](auto In) {
+        const long_index_t d1_batch_offset = __builtin_amdgcn_readfirstlane(
+            static_cast<long_index_t>(compute_base_ptr_of_batch.GetD1BasePtr(g_idx, In)));
+        p_d1s_grid(In) = p_d1s_grid(In) + d1_batch_offset;
+    });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a0_grid + a_batch_offset,
+                                                  p_b0_grid + b_batch_offset,
+                                                  p_d0s_grid,
+                                                  p_b1_grid + b1_batch_offset,
+                                                  p_d1s_grid,
+                                                  p_e1_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a0_element_op,
+                                                  b0_element_op,
+                                                  cde0_element_op,
+                                                  b1_element_op,
+                                                  cde1_element_op,
+                                                  a0_grid_desc_ak0_m_ak1,
+                                                  b0_grid_desc_bk0_n_bk1,
+                                                  d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                                  b1_grid_desc_bk0_n_bk1,
+                                                  d1s_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e1_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_e1tile_map);
+#else
+    ignore = p_a0_grid;
+    ignore = p_b0_grid;
+    ignore = p_d0s_grid;
+    ignore = p_b1_grid;
+    ignore = p_d1s_grid;
+    ignore = p_e1_grid;
+    ignore = a0_element_op;
+    ignore = b0_element_op;
+    ignore = cde0_element_op;
+    ignore = b1_element_op;
+    ignore = cde1_element_op;
+    ignore = a0_grid_desc_ak0_m_ak1;
+    ignore = b0_grid_desc_bk0_n_bk1;
+    ignore = d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5;
+    ignore = b1_grid_desc_bk0_n_bk1;
+    ignore = d1s_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e1_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_e1tile_map;
+    ignore = batch_count;
+    ignore = compute_base_ptr_of_batch;
+#endif
+}
+
+// Computes C = A * B0 * B1
+//              ^^^^^^ (Acc0)
+//              ^^^^^^^^^^^ (Acc1)
+template <typename A0Layout,
+          typename B0Layout, // B0Layout
+          typename D0sLayout,
+          typename B1Layout,
+          typename D1sLayout,
+          typename E1Layout,
+          typename A0DataType,
+          typename B0DataType,
+          typename Acc0DataType,
+          typename D0sDataType,
+          typename B1DataType,
+          typename Acc1DataType,
+          typename C1ShuffleDataType,
+          typename D1sDataType,
+          typename E1DataType,
+          typename A0ElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename CDE0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CDE1ElementwiseOperation,
+          bool PadGemm0M,
+          bool PadGemm0N,
+          bool PadGemm0K,
+          bool PadGemm1N,
+          bool PadGemm1K,
+          index_t NumGemm0KPrefetchStage,
+          index_t BlockSize,
+          index_t Gemm0MPerBlock,
+          index_t Gemm0NPerBlock,
+          index_t Gemm0KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t A0K1,
+          index_t B0K1,
+          index_t B1K1,
+          index_t Gemm0MPerXdl,
+          index_t Gemm0NPerXdl,
+          index_t Gemm0MXdlPerWave,
+          index_t Gemm0NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename A0BlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename A0BlockTransferThreadClusterArrangeOrder,
+          typename A0BlockTransferSrcAccessOrder,
+          index_t A0BlockTransferSrcVectorDim,
+          index_t A0BlockTransferSrcScalarPerVector,
+          index_t A0BlockTransferDstScalarPerVector_AK1,
+          bool A0BlockLdsExtraM,
+          typename B0BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B0BlockTransferThreadClusterArrangeOrder,
+          typename B0BlockTransferSrcAccessOrder,
+          index_t B0BlockTransferSrcVectorDim,
+          index_t B0BlockTransferSrcScalarPerVector,
+          index_t B0BlockTransferDstScalarPerVector_BK1,
+          bool B0BlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1BlockLdsExtraN,
+          index_t C1ShuffleMXdlPerWavePerShuffle,
+          index_t C1ShuffleGemm0NXdlPerWavePerShuffle,
+          typename CDE1ShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDE1ShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+struct DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
+    : public DeviceBatchedGemmMultipleDGemmMultipleD<A0Layout,
+                                                     B0Layout,
+                                                     D0sLayout,
+                                                     B1Layout,
+                                                     D1sLayout,
+                                                     E1Layout,
+                                                     A0DataType,
+                                                     B0DataType,
+                                                     D0sDataType,
+                                                     B1DataType,
+                                                     D1sDataType,
+                                                     E1DataType,
+                                                     A0ElementwiseOperation,
+                                                     B0ElementwiseOperation,
+                                                     CDE0ElementwiseOperation,
+                                                     B1ElementwiseOperation,
+                                                     CDE1ElementwiseOperation>
+{
+    using DeviceOp = DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumD0Tensor = D0sDataType::Size();
+    static constexpr index_t NumD1Tensor = D1sDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+    static constexpr auto I8 = Number<8>{};
+    static constexpr auto I9 = Number<9>{};
+
+    static constexpr auto gemm0_padder =
+        GemmPadder_v2<PadGemm0M, PadGemm0N, PadGemm0K, index_t, index_t, index_t>{
+            Gemm0MPerBlock, Gemm0NPerBlock, Gemm0KPerBlock};
+
+    static constexpr auto gemm1_padder =
+        GemmPadder_v2<PadGemm0M, PadGemm1N, PadGemm1K, index_t, index_t, index_t>{
+            Gemm0MPerBlock, Gemm1NPerBlock, Gemm1KPerBlock};
+
+    // for Gemm0
+    static auto MakeA0GridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA0)
+    {
+        const auto a0_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, A0Layout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA0, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, A0Layout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA0));
+            }
+        }();
+
+        return gemm0_padder.PadADescriptor_M_K(a0_grid_desc_mraw_kraw);
+    }
+
+    // for Gemm0
+    static auto MakeB0GridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b0_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, B0Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, B0Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return gemm0_padder.PadBDescriptor_N_K(b0_grid_desc_nraw_kraw);
+    }
+
+    // for Gemm0
+    template <typename DLay>
+    static auto MakeD0GridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideD0)
+    {
+        const auto d0_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, DLay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideD0, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, DLay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideD0));
+            }
+        }();
+
+        return gemm0_padder.PadCDescriptor_M_N(d0_grid_desc_mraw_nraw);
+    }
+
+    // for Gemm1
+    static auto MakeB1GridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b1_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return gemm1_padder.PadBDescriptor_N_K(b1_grid_desc_nraw_kraw);
+    }
+
+    // for Gemm1
+    template <typename ELay>
+    static auto MakeE1GridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE1)
+    {
+        const auto e1_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE1, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE1));
+            }
+        }();
+
+        return gemm1_padder.PadCDescriptor_M_N(e1_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeD0sGridDescriptor_M_N(const std::array<index_t, NumD1Tensor>& MRaws,
+                                          const std::array<index_t, NumD1Tensor>& NRaws,
+                                          const std::array<index_t, NumD1Tensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, D0sLayout>>;
+
+                return DeviceOp::MakeD0GridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumD0Tensor>{});
+    }
+
+    static auto MakeD1sGridDescriptor_M_N(const std::array<index_t, NumD1Tensor>& MRaws,
+                                          const std::array<index_t, NumD1Tensor>& NRaws,
+                                          const std::array<index_t, NumD1Tensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, D1sLayout>>;
+
+                return DeviceOp::MakeE1GridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumD1Tensor>{});
+    }
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(index_t BatchStrideA0,
+                                     index_t BatchStrideB0,
+                                     std::array<index_t, NumD0Tensor> BatchStrideD0s,
+                                     index_t BatchStrideB1,
+                                     std::array<index_t, NumD1Tensor> BatchStrideD1s,
+                                     index_t BatchStrideE1)
+            : BatchStrideA0_(BatchStrideA0),
+              BatchStrideB0_(BatchStrideB0),
+              BatchStrideD0s_(BatchStrideD0s),
+              BatchStrideB1_(BatchStrideB1),
+              BatchStrideD1s_(BatchStrideD1s),
+              BatchStrideE1_(BatchStrideE1)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA0_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB0_);
+        }
+
+        template <index_t I>
+        __host__ __device__ constexpr long_index_t GetD0BasePtr(index_t g_idx,
+                                                                Number<I> d1_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideD0s_[d1_idx]);
+        }
+
+        __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB1_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideE1_);
+        }
+
+        template <index_t I>
+        __host__ __device__ constexpr auto GetD1BasePtr(index_t g_idx, Number<I> d1_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideD1s_[d1_idx]);
+        }
+
+        private:
+        index_t BatchStrideA0_;
+        index_t BatchStrideB0_;
+        std::array<index_t, NumD0Tensor> BatchStrideD0s_;
+        index_t BatchStrideB1_;
+        std::array<index_t, NumD1Tensor> BatchStrideD1s_;
+        index_t BatchStrideE1_;
+    };
+
+    using A0GridDesc_M_K  = decltype(MakeA0GridDescriptor_M_K(1, 1, 1));
+    using B0GridDesc_N_K  = decltype(MakeB0GridDescriptor_N_K(1, 1, 1));
+    using D0sGridDesc_M_N = remove_cvref_t<decltype(MakeD0sGridDescriptor_M_N({}, {}, {}))>;
+    using B1GridDesc_N_K  = decltype(MakeB1GridDescriptor_N_K(1, 1, 1));
+    using D1sGridDesc_M_N = remove_cvref_t<decltype(MakeD1sGridDescriptor_M_N({}, {}, {}))>;
+    using E1GridDesc_M_N  = decltype(MakeE1GridDescriptor_M_N<E1Layout>(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<
+        A0DataType, // TODO: distinguish A/B datatype
+        Acc0DataType,
+        D0sDataType,
+        Acc1DataType,
+        C1ShuffleDataType,
+        D1sDataType,
+        E1DataType,
+        A0ElementwiseOperation,
+        B0ElementwiseOperation,
+        CDE0ElementwiseOperation,
+        B1ElementwiseOperation,
+        CDE1ElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        A0GridDesc_M_K,
+        B0GridDesc_N_K,
+        D0sGridDesc_M_N,
+        B1GridDesc_N_K,
+        D1sGridDesc_M_N,
+        E1GridDesc_M_N,
+        NumGemm0KPrefetchStage,
+        BlockSize,
+        Gemm0MPerBlock,
+        Gemm0NPerBlock,
+        Gemm0KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        A0K1,
+        B0K1,
+        B1K1,
+        Gemm0MPerXdl,
+        Gemm0NPerXdl,
+        Gemm0MXdlPerWave,
+        Gemm0NXdlPerWave,
+        Gemm1NXdlPerWave,
+        A0BlockTransferThreadClusterLengths_AK0_M_AK1,
+        A0BlockTransferThreadClusterArrangeOrder,
+        A0BlockTransferSrcAccessOrder,
+        A0BlockTransferSrcVectorDim,
+        A0BlockTransferSrcScalarPerVector,
+        A0BlockTransferDstScalarPerVector_AK1,
+        true,
+        A0BlockLdsExtraM,
+        B0BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B0BlockTransferThreadClusterArrangeOrder,
+        B0BlockTransferSrcAccessOrder,
+        B0BlockTransferSrcVectorDim,
+        B0BlockTransferSrcScalarPerVector,
+        B0BlockTransferDstScalarPerVector_BK1,
+        true,
+        B0BlockLdsExtraN,
+        B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B1BlockTransferThreadClusterArrangeOrder,
+        B1BlockTransferSrcAccessOrder,
+        B1BlockTransferSrcVectorDim,
+        B1BlockTransferSrcScalarPerVector,
+        B1BlockTransferDstScalarPerVector_BK1,
+        false,
+        B1BlockLdsExtraN,
+        C1ShuffleMXdlPerWavePerShuffle,
+        C1ShuffleGemm0NXdlPerWavePerShuffle,
+        CDE1ShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDE1ShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using A0GridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultA0GridDescriptor_AK0_M_AK1(A0GridDesc_M_K{}))>;
+    using B0GridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultB0GridDescriptor_BK0_N_BK1(B0GridDesc_N_K{}))>;
+    using B1GridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultB1GridDescriptor_BK0_N_BK1(B1GridDesc_N_K{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const A0DataType* p_a0_grid,
+                 const B0DataType* p_b0_grid,
+                 std::array<const void*, NumD0Tensor> p_d0s_grid,
+                 const B1DataType* p_b1_grid,
+                 std::array<const void*, NumD1Tensor> p_d1s_grid,
+                 E1DataType* p_e1_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t Gemm1NRaw, // = ORaw
+                 index_t Batch,
+                 index_t StrideA0,
+                 index_t StrideB0,
+                 std::array<index_t, NumD0Tensor> StrideD0s,
+                 index_t StrideB1,
+                 std::array<index_t, NumD1Tensor> StrideD1s,
+                 index_t StrideE1,
+                 index_t BatchStrideA0,
+                 index_t BatchStrideB0,
+                 std::array<index_t, NumD0Tensor> BatchStrideD0s,
+                 index_t BatchStrideB1,
+                 std::array<index_t, NumD1Tensor> BatchStrideD1s,
+                 index_t BatchStrideE1,
+                 A0ElementwiseOperation a0_element_op,
+                 B0ElementwiseOperation b0_element_op,
+                 CDE0ElementwiseOperation cde0_element_op,
+                 B1ElementwiseOperation b1_element_op,
+                 CDE1ElementwiseOperation cde1_element_op)
+            : p_a0_grid_{p_a0_grid},
+              p_b0_grid_{p_b0_grid},
+              p_d0s_grid_{},
+              p_b1_grid_{p_b1_grid},
+              p_d1s_grid_{},
+              p_e1_grid_{p_e1_grid},
+              a0_grid_desc_m_k_{DeviceOp::MakeA0GridDescriptor_M_K(MRaw, KRaw, StrideA0)},
+              b0_grid_desc_n_k_{DeviceOp::MakeB0GridDescriptor_N_K(KRaw, NRaw, StrideB0)},
+              d0s_grid_desc_m_n_{},
+              b1_grid_desc_n_k_{DeviceOp::MakeB1GridDescriptor_N_K(NRaw, Gemm1NRaw, StrideB1)},
+              d1s_grid_desc_m_n_{},
+              e1_grid_desc_m_n_{
+                  DeviceOp::MakeE1GridDescriptor_M_N<E1Layout>(MRaw, Gemm1NRaw, StrideE1)},
+              a0_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultA0GridDescriptor_AK0_M_AK1(a0_grid_desc_m_k_)},
+              b0_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultB0GridDescriptor_BK0_N_BK1(b0_grid_desc_n_k_)},
+              d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_{},
+              b1_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultB1GridDescriptor_BK0_N_BK1(b1_grid_desc_n_k_)},
+              d1s_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e1_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_e1tile_map_{GridwiseGemm::MakeDefaultBlock2E1TileMap(e1_grid_desc_m_n_)},
+              a0_element_op_{a0_element_op},
+              b0_element_op_{b0_element_op},
+              cde0_element_op_{cde0_element_op},
+              b1_element_op_{b1_element_op},
+              cde1_element_op_{cde1_element_op},
+              batch_count_(Batch),
+              compute_base_ptr_of_batch_{BatchStrideA0,
+                                         BatchStrideB0,
+                                         BatchStrideD0s,
+                                         BatchStrideB1,
+                                         BatchStrideD1s,
+                                         BatchStrideE1}
+        {
+            std::cout << "a0_grid_desc_m_k_{" << a0_grid_desc_m_k_.GetLength(I0) << ", "
+                      << a0_grid_desc_m_k_.GetLength(I1) << "}" << std::endl;
+            std::cout << "b0_grid_desc_n_k_{" << b0_grid_desc_n_k_.GetLength(I0) << ", "
+                      << b0_grid_desc_n_k_.GetLength(I1) << "}" << std::endl;
+            std::cout << "d0s_grid_desc_m_n_[I0]{" << d0s_grid_desc_m_n_[I0].GetLength(I0) << ", "
+                      << d0s_grid_desc_m_n_[I0].GetLength(I1) << "}" << std::endl;
+            std::cout << "b1_grid_desc_n_k_{" << b1_grid_desc_n_k_.GetLength(I0) << ", "
+                      << b1_grid_desc_n_k_.GetLength(I1) << "}" << std::endl;
+            std::cout << "d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_{"
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I0) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I1) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I2) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I3) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I4) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I5) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I6) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I7) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I8) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I9) << "}"
+                      << std::endl;
+            std::cout << "e1_grid_desc_m_n_{" << e1_grid_desc_m_n_.GetLength(I0) << ", "
+                      << e1_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+            static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                using D0Layout   = remove_cvref_t<tuple_element_t<i.value, D0sLayout>>;
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+
+                // D0 pointer
+                p_d0s_grid_(i) = static_cast<const D0DataType*>(p_d0s_grid[i]);
+
+                // D0 desc
+                d0s_grid_desc_m_n_(i) =
+                    DeviceOp::MakeD0GridDescriptor_M_N<D0Layout>(MRaw, NRaw, StrideD0s[i]);
+            });
+
+            static_for<0, NumD1Tensor, 1>{}([&](auto i) {
+                using D1Layout   = remove_cvref_t<tuple_element_t<i.value, D1sLayout>>;
+                using D1DataType = remove_cvref_t<tuple_element_t<i.value, D1sDataType>>;
+
+                // D1 pointer
+                p_d1s_grid_(i) = static_cast<const D1DataType*>(p_d1s_grid[i]);
+
+                // D1 desc
+                d1s_grid_desc_m_n_(i) =
+                    DeviceOp::MakeE1GridDescriptor_M_N<D1Layout>(MRaw, Gemm1NRaw, StrideD1s[i]);
+            });
+
+            if(GridwiseGemm::CheckValidity(a0_grid_desc_m_k_,
+                                           b0_grid_desc_n_k_,
+                                           b1_grid_desc_n_k_,
+                                           e1_grid_desc_m_n_,
+                                           block_2_e1tile_map_))
+            {
+                e1_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeE1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e1_grid_desc_m_n_);
+
+                d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_ =
+                    GridwiseGemm::MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
+                        d0s_grid_desc_m_n_);
+
+                d1s_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeD1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        d1s_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        // pointers
+        const A0DataType* p_a0_grid_;
+        const B0DataType* p_b0_grid_;
+        typename GridwiseGemm::D0sGridPointer p_d0s_grid_;
+        const B1DataType* p_b1_grid_;
+        typename GridwiseGemm::D1sGridPointer p_d1s_grid_;
+        E1DataType* p_e1_grid_;
+
+        // tensor descriptors for problem definiton
+        A0GridDesc_M_K a0_grid_desc_m_k_;
+        B0GridDesc_N_K b0_grid_desc_n_k_;
+        D0sGridDesc_M_N d0s_grid_desc_m_n_;
+        B1GridDesc_N_K b1_grid_desc_n_k_;
+        D1sGridDesc_M_N d1s_grid_desc_m_n_;
+        E1GridDesc_M_N e1_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        A0GridDesc_AK0_M_AK1 a0_grid_desc_ak0_m_ak1_;
+        B0GridDesc_BK0_N_BK1 b0_grid_desc_bk0_n_bk1_;
+        typename GridwiseGemm::D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
+            d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        typename GridwiseGemm::D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            d1s_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e1_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e1-tile map
+        typename GridwiseGemm::DefaultBlock2E1TileMap block_2_e1tile_map_;
+
+        // element-wise op
+        A0ElementwiseOperation a0_element_op_;
+        B0ElementwiseOperation b0_element_op_;
+        CDE0ElementwiseOperation cde0_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        CDE1ElementwiseOperation cde1_element_op_;
+
+        // batch
+        index_t batch_count_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a0_grid_desc_m_k_,
+                                            arg.b0_grid_desc_n_k_,
+                                            arg.b1_grid_desc_n_k_,
+                                            arg.e1_grid_desc_m_n_,
+                                            arg.block_2_e1tile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_e1tile_map_.CalculateGridSize(arg.e1_grid_desc_m_n_) * arg.batch_count_;
+
+            // Gemm0_K
+            const auto K = arg.a0_grid_desc_m_k_.GetLength(I1);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_batched_gemm_gemm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    A0DataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::D0sGridPointer,
+                    typename GridwiseGemm::D1sGridPointer,
+                    E1DataType,
+                    A0ElementwiseOperation,
+                    B0ElementwiseOperation,
+                    CDE0ElementwiseOperation,
+                    B1ElementwiseOperation,
+                    CDE1ElementwiseOperation,
+                    DeviceOp::A0GridDesc_AK0_M_AK1,
+                    DeviceOp::B0GridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
+                    DeviceOp::B1GridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2E1TileMap,
+                    ComputeBasePtrOfStridedBatch,
+                    has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a0_grid_,
+                                              arg.p_b0_grid_,
+                                              arg.p_d0s_grid_,
+                                              arg.p_b1_grid_,
+                                              arg.p_d1s_grid_,
+                                              arg.p_e1_grid_,
+                                              arg.a0_element_op_,
+                                              arg.b0_element_op_,
+                                              arg.cde0_element_op_,
+                                              arg.b1_element_op_,
+                                              arg.cde1_element_op_,
+                                              arg.a0_grid_desc_ak0_m_ak1_,
+                                              arg.b0_grid_desc_bk0_n_bk1_,
+                                              arg.d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
+                                              arg.b1_grid_desc_bk0_n_bk1_,
+                                              arg.d1s_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e1_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_e1tile_map_,
+                                              arg.batch_count_,
+                                              arg.compute_base_ptr_of_batch_);
+            };
+
+            // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
+            // to concern Gemm0's loop
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a0_grid_desc_m_k_,
+                                           arg.b0_grid_desc_n_k_,
+                                           arg.b1_grid_desc_n_k_,
+                                           arg.e1_grid_desc_m_n_,
+                                           arg.block_2_e1tile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const A0DataType* p_a0,
+                             const B0DataType* p_b0,
+                             std::array<const void*, NumD0Tensor> p_d0s,
+                             const B1DataType* p_b1,
+                             std::array<const void*, NumD1Tensor> p_d1s,
+                             E1DataType* p_e1,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t Gemm1NRaw,
+                             index_t Batch,
+                             index_t StrideA0,
+                             index_t StrideB0,
+                             std::array<index_t, NumD0Tensor> StrideD0s,
+                             index_t StrideB1,
+                             std::array<index_t, NumD1Tensor> StrideD1s,
+                             index_t StrideE1,
+                             index_t BatchStrideA0,
+                             index_t BatchStrideB0,
+                             std::array<index_t, NumD0Tensor> BatchStrideD0s,
+                             index_t BatchStrideB1,
+                             std::array<index_t, NumD1Tensor> BatchStrideD1s,
+                             index_t BatchStrideE1,
+                             A0ElementwiseOperation a0_element_op,
+                             B0ElementwiseOperation b0_element_op,
+                             CDE0ElementwiseOperation cde0_element_op,
+                             B1ElementwiseOperation b1_element_op,
+                             CDE1ElementwiseOperation cde1_element_op)
+    {
+        return Argument{p_a0,          p_b0,
+                        p_d0s,         p_b1,
+                        p_d1s,         p_e1,
+                        MRaw,          NRaw,
+                        KRaw,          Gemm1NRaw,
+                        Batch,         StrideA0,
+                        StrideB0,      StrideD0s,
+                        StrideB1,      StrideD1s,
+                        StrideE1,      BatchStrideA0,
+                        BatchStrideB0, BatchStrideD0s,
+                        BatchStrideB1, BatchStrideD1s,
+                        BatchStrideE1, a0_element_op,
+                        b0_element_op, cde0_element_op,
+                        b1_element_op, cde1_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a0,
+                        const void* p_b0,
+                        std::array<const void*, NumD0Tensor> p_d0s,
+                        const void* p_b1,
+                        std::array<const void*, NumD1Tensor> p_d1s,
+                        void* p_e1,
+                        index_t MRaw,
+                        index_t NRaw,
+                        index_t KRaw,
+                        index_t Gemm1NRaw,
+                        index_t Batch,
+                        index_t StrideA0,
+                        index_t StrideB0,
+                        std::array<ck::index_t, NumD0Tensor> StrideD0s,
+                        index_t StrideB1,
+                        std::array<ck::index_t, NumD1Tensor> StrideD1s,
+                        index_t StrideE1,
+                        index_t BatchStrideA0,
+                        index_t BatchStrideB0,
+                        std::array<ck::index_t, NumD0Tensor> BatchStrideD0s,
+                        index_t BatchStrideB1,
+                        std::array<ck::index_t, NumD1Tensor> BatchStrideD1s,
+                        index_t BatchStrideE1,
+                        A0ElementwiseOperation a0_element_op,
+                        B0ElementwiseOperation b0_element_op,
+                        CDE0ElementwiseOperation cde0_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CDE1ElementwiseOperation cde1_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const A0DataType*>(p_a0),
+                                          static_cast<const B0DataType*>(p_b0),
+                                          p_d0s,
+                                          static_cast<const B1DataType*>(p_b1),
+                                          p_d1s,
+                                          static_cast<E1DataType*>(p_e1),
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          Gemm1NRaw,
+                                          Batch,
+                                          StrideA0,
+                                          StrideB0,
+                                          StrideD0s,
+                                          StrideB1,
+                                          StrideD1s,
+                                          StrideE1,
+                                          BatchStrideA0,
+                                          BatchStrideB0,
+                                          BatchStrideD0s,
+                                          BatchStrideB1,
+                                          BatchStrideD1s,
+                                          BatchStrideE1,
+                                          a0_element_op,
+                                          b0_element_op,
+                                          cde0_element_op,
+                                          b1_element_op,
+                                          cde1_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << Gemm0MPerBlock << ", "
+            << Gemm0NPerBlock << ", "
+            << Gemm0KPerBlock << ", "
+            << A0K1 << ", "
+            << B0K1 << ", "
+            << B1K1 << ", "
+            << Gemm0MPerXdl << ", "
+            << Gemm0NPerXdl << ", "
+            << Gemm0MXdlPerWave << ", "
+            << Gemm0NXdlPerWave << ", "
+            << Gemm1NXdlPerWave << "> ";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp
new file mode 100644
index 0000000..3c5fdbd
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp
@@ -0,0 +1,1001 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename ReduceGridDescriptor_MBlock_MPerBlock,
+          typename ComputeBasePrtOfBatch,
+          typename Block2CTileMap,
+          bool HasMainK0BlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_reduce_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            ReducePtrsGlobal p_reduces_grid,
+            const index_t batch_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const ReduceInElementwiseOperations reduce_in_element_ops,
+            const ReduceAccElementwiseOperations reduce_out_element_ops,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
+            const ComputeBasePrtOfBatch compute_base_ptr_of_batch_,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch_.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch_.GetBBasePtr(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch_.GetCBasePtr(g_idx)));
+
+    static_for<0, p_reduces_grid.Size(), 1>{}([&](auto In) {
+        const long_index_t d_batch_offset = __builtin_amdgcn_readfirstlane(
+            static_cast<long_index_t>(compute_base_ptr_of_batch_.GetDBasePtr(g_idx, In)));
+        p_reduces_grid(In) = p_reduces_grid(In) + d_batch_offset;
+    });
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid + a_batch_offset,
+                                                   p_b_grid + b_batch_offset,
+                                                   p_c_grid + c_batch_offset,
+                                                   p_reduces_grid,
+                                                   p_shared,
+                                                   a_element_op,
+                                                   b_element_op,
+                                                   c_element_op,
+                                                   reduce_in_element_ops,
+                                                   reduce_out_element_ops,
+                                                   a_grid_desc_ak0_m_ak1,
+                                                   b_grid_desc_bk0_n_bk1,
+                                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                   reduce_grid_desc_mblock_mperblock,
+                                                   block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_reduces_grid;
+    ignore = batch_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = reduce_in_element_ops;
+    ignore = reduce_out_element_ops;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = reduce_grid_desc_mblock_mperblock;
+    ignore = compute_base_ptr_of_batch_;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+// Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
+// version currently has compiler issues with register spill which further causes validation
+// failures.
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename ReduceAccDataType,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename ReduceOperations,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          typename ReduceGlobalMemoryDataOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+          index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<0, ReduceOperations::Size()>
+{
+    using DeviceOp = DeviceBatchedGemmReduce_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(M)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_right_pad_transform(MRaw, MPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(N)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_right_pad_transform(NRaw, NPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    // assume D is packed tensor
+    static auto MakeReduceGridDescriptor_M(index_t MRaw)
+    {
+        const auto d_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(MRaw));
+
+        const auto M    = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto MPad = M - MRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M
+            return transform_tensor_descriptor(d_grid_desc_mraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad)),
+                                               make_tuple(Sequence<0>{}),
+                                               make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad M
+            return d_grid_desc_mraw;
+        }
+    }
+
+    using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N       = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+    using ReduceGridDesc_M    = decltype(MakeReduceGridDescriptor_M(1));
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
+                                     index_t BatchStrideB,
+                                     index_t BatchStrideC,
+                                     index_t BatchStrideD)
+            : BatchStrideA_(BatchStrideA),
+              BatchStrideB_(BatchStrideB),
+              BatchStrideC_(BatchStrideC),
+              BatchStrideD_(BatchStrideD)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideC_);
+        }
+
+        template <index_t I>
+        __host__ __device__ constexpr long_index_t GetDBasePtr(index_t g_idx,
+                                                               Number<I> reduction_idx) const
+        {
+            // TODO - Support sequence of StrideD in MakeArgument()
+            (void)reduction_idx;
+            return g_idx * static_cast<long_index_t>(BatchStrideD_);
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        index_t BatchStrideC_;
+        index_t BatchStrideD_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        ReduceAccDataType,
+        ReducePtrsGlobal,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        ReduceOperations,
+        ReduceInElementwiseOperations,
+        ReduceAccElementwiseOperations,
+        InMemoryDataOperationEnum::Set,
+        ReduceGlobalMemoryDataOperation,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        ReduceGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+        CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+        CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+        LoopSched>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 ReducePtrsGlobal p_reduces_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op,
+                 ReduceInElementwiseOperations reduce_in_element_ops,
+                 ReduceAccElementwiseOperations reduce_out_element_ops,
+                 index_t Batch)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              p_reduces_grid_{p_reduces_grid},
+              Batch_(Batch),
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC)},
+              reduce_grid_desc_m_{DeviceOp::MakeReduceGridDescriptor_M(MRaw)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              reduce_grid_desc_mblock_mperblock_{},
+              compute_base_ptr_of_batch_{
+                  type_convert<index_t>(a_grid_desc_ak0_m_ak1_.GetElementSpaceSize()),
+                  type_convert<index_t>(b_grid_desc_bk0_n_bk1_.GetElementSpaceSize()),
+                  type_convert<index_t>(c_grid_desc_m_n_.GetElementSpaceSize()),
+                  type_convert<index_t>(reduce_grid_desc_m_.GetElementSpaceSize())},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              reduce_in_element_ops_{reduce_in_element_ops},
+              reduce_out_element_ops_{reduce_out_element_ops}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+
+                reduce_grid_desc_mblock_mperblock_ =
+                    GridwiseGemm::MakeReduceGridDescriptor_MBlock_MPerBlock(reduce_grid_desc_m_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        ReducePtrsGlobal p_reduces_grid_;
+        index_t Batch_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        ReduceGridDesc_M reduce_grid_desc_m_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock
+            reduce_grid_desc_mblock_mperblock_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+        ReduceInElementwiseOperations reduce_in_element_ops_;
+        ReduceAccElementwiseOperations reduce_out_element_ops_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << "arg.Batch_ = " << arg.Batch_ << std::endl;
+
+                std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_bk0_n_bk1_{"
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+                std::cout << "arg.reduce_grid_desc_m_{ " << arg.reduce_grid_desc_m_.GetLength(I0) << "}"
+                          << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                            arg.b_grid_desc_bk0_n_bk1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.Batch_;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float elapsed_time = 0.0f;
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_batched_gemm_reduce_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    ReducePtrsGlobal,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    ReduceInElementwiseOperations,
+                    ReduceAccElementwiseOperations,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock,
+                    ComputeBasePtrOfStridedBatch,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    true>;
+
+                elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_reduces_grid_,
+                                           arg.Batch_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.reduce_in_element_ops_,
+                                           arg.reduce_out_element_ops_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.reduce_grid_desc_mblock_mperblock_,
+                                           arg.compute_base_ptr_of_batch_,
+                                           arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_batched_gemm_reduce_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    ReducePtrsGlobal,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    ReduceInElementwiseOperations,
+                    ReduceAccElementwiseOperations,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock,
+                    ComputeBasePtrOfStridedBatch,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    false>;
+
+                elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_reduces_grid_,
+                                           arg.Batch_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.reduce_in_element_ops_,
+                                           arg.reduce_out_element_ops_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.reduce_grid_desc_mblock_mperblock_,
+                                           arg.compute_base_ptr_of_batch_,
+                                           arg.block_2_ctile_map_);
+            }
+
+            return elapsed_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        auto casted_p_arg = dynamic_cast<const Argument*>(p_arg);
+        if(casted_p_arg == nullptr)
+        {
+            return false;
+        }
+        else
+        {
+            return IsSupportedArgument(*casted_p_arg);
+        }
+    }
+
+    static constexpr int NumReduce = ReduceOperations::Size();
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             const void* p_bias,
+                             std::array<const void*, 0> p_ds,
+                             void* p_c,
+                             std::array<void*, NumReduce> p_reduces,
+                             ck::index_t M,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t StrideA,
+                             ck::index_t StrideB,
+                             ck::index_t StrideC,
+                             std::array<ck::index_t, 0> StrideDs,
+                             std::array<void*, 3> gemm_element_ops,
+                             std::array<void*, 0> d_element_ops,
+                             std::array<void*, NumReduce> reduce_in_element_op,
+                             std::array<void*, NumReduce> reduce_out_element_op,
+                             index_t Batch)
+    {
+        (void)p_bias;
+        (void)p_ds;
+        (void)StrideDs;
+        (void)d_element_ops;
+
+        ReducePtrsGlobal reduce_tuple = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReducePtrsGlobal{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return static_cast<T*>(p_reduces[I]);
+            },
+            Number<NumReduce>{});
+
+        ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceInElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_in_element_op[I]));
+            },
+            Number<NumReduce>{});
+        ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceAccElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_out_element_op[I]));
+            },
+            Number<NumReduce>{});
+
+        AElementwiseOperation a_element_op =
+            *(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
+        BElementwiseOperation b_element_op =
+            *(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
+        CElementwiseOperation c_element_op =
+            *(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
+
+        return Argument{static_cast<const ADataType*>(p_a),
+                        static_cast<const BDataType*>(p_b),
+                        static_cast<CDataType*>(p_c),
+                        reduce_tuple,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op,
+                        reduce_in_element_ops,
+                        reduce_out_element_ops,
+                        Batch};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const void* p_bias,
+                        std::array<const void*, 0> p_ds,
+                        void* p_c,
+                        std::array<void*, NumReduce> p_reduces,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        std::array<ck::index_t, 0> StrideDs,
+                        std::array<void*, 3> gemm_element_ops,
+                        std::array<void*, 0> d_element_ops,
+                        std::array<void*, NumReduce> reduce_in_element_op,
+                        std::array<void*, NumReduce> reduce_out_element_op,
+                        index_t Batch = 1) override
+    {
+        (void)p_bias;
+        (void)p_ds;
+        (void)StrideDs;
+        (void)d_element_ops;
+
+        ReducePtrsGlobal reduce_tuple = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReducePtrsGlobal{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return static_cast<T*>(p_reduces[I]);
+            },
+            Number<NumReduce>{});
+
+        ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceInElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_in_element_op[I]));
+            },
+            Number<NumReduce>{});
+        ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceAccElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_out_element_op[I]));
+            },
+            Number<NumReduce>{});
+
+        AElementwiseOperation a_element_op =
+            *(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
+        BElementwiseOperation b_element_op =
+            *(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
+        CElementwiseOperation c_element_op =
+            *(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
+
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          reduce_tuple,
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op,
+                                          reduce_in_element_ops,
+                                          reduce_out_element_ops,
+                                          Batch);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmReduce_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
new file mode 100644
index 0000000..5baa0f8
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
@@ -0,0 +1,859 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2CTileMap,
+          typename ComputeBasePtrOfStridedBatch,
+          typename C0MatrixMask,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_softmax_gemm_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            const FloatAB* __restrict__ p_b1_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const AccElementwiseOperation acc_element_op,
+            const B1ElementwiseOperation b1_element_op,
+            const CElementwiseOperation c_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2CTileMap block_2_ctile_map,
+            const index_t batch_count,
+            const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
+            const C0MatrixMask c0_matrix_mask)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetBBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetCBasePtr(g_idx)));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_b1_grid + b1_batch_offset,
+                                                  p_c_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  acc_element_op,
+                                                  b1_element_op,
+                                                  c_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  b1_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_ctile_map,
+                                                  c0_matrix_mask);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_b1_grid;
+    ignore = p_c_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = acc_element_op;
+    ignore = b1_element_op;
+    ignore = c_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = b1_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_ctile_map;
+    ignore = batch_count;
+    ignore = compute_base_ptr_of_batch;
+    ignore = c0_matrix_mask;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// Computes C = A * B0 * B1
+//              ^^^^^^ (Acc0)
+//              ^^^^^^^^^^^ (Acc1)
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO, // NumDimGemm1N
+          typename ADataType,
+          typename BDataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasDataType,
+          typename Acc1BiasDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization B1Spec,
+          TensorSpecialization CSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock, // Gemm0NPerBlock
+          index_t KPerBlock, // Gemm0KPerBlock
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t B1K1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          MaskingSpecialization MaskingSpec,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
+    : public DeviceBatchedGemmSoftmaxGemmPermute<NumDimG,
+                                                 NumDimM,
+                                                 NumDimN,
+                                                 NumDimK,
+                                                 NumDimO,
+                                                 ADataType,
+                                                 BDataType,
+                                                 B1DataType,
+                                                 CDataType,
+                                                 Acc0BiasDataType,
+                                                 Acc1BiasDataType,
+                                                 AElementwiseOperation,
+                                                 BElementwiseOperation,
+                                                 AccElementwiseOperation,
+                                                 B1ElementwiseOperation,
+                                                 CElementwiseOperation,
+                                                 MaskingSpec>
+{
+    static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0,
+                  "Number of dimension must be greater than 0");
+
+    static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
+    static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
+
+    // TODO ANT: implement bias combination
+    static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented");
+
+#if 0
+    // TODO ANT: use alias
+    static constexpr index_t NumDimGemm0M = NumDimM;
+    static constexpr index_t NumDimGemm0N = NumDimN;
+    static constexpr index_t NumDimGemm0K = NumDimK;
+    static constexpr index_t NumDimGemm1M = NumDimM;
+    static constexpr index_t NumDimGemm1N = NumDimO;
+    static constexpr index_t NumDimGemm1K = NumDimN;
+#endif
+
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    using Transform = TransformBatchedContractionContractionToBatchedGemmGemm<
+        Sequence<NumDimG, NumDimM, NumDimN, NumDimK, NumDimO>,
+        Sequence<MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock>,
+        GemmSpec,
+        ASpec,
+        BSpec,
+        B1Spec,
+        CSpec>;
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                              const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        return Transform::MakeAGridDescriptor_AK0_M_AK1(
+            Transform::MakeAGridDescriptor_M_K(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec),
+            Number<AK1>{});
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                              const std::vector<index_t>& b_gs_ns_ks_strides_vec)
+    {
+        return Transform::MakeB0GridDescriptor_BK0_N_BK1(
+            Transform::MakeB0GridDescriptor_N_K(b_gs_ns_ks_lengths_vec, b_gs_ns_ks_strides_vec),
+            Number<BK1>{});
+    }
+
+    static auto
+    MakeB1GridDescriptor_BK0_N_BK1(const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                   const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides_vec)
+    {
+        return Transform::MakeB1GridDescriptor_BK0_N_BK1(
+            Transform::MakeB1GridDescriptor_N_K(b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                                b1_gs_gemm1ns_gemm1ks_strides_vec),
+            Number<B1K1>{});
+    }
+
+    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {}));
+    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {}));
+    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {}));
+    using CGridDesc_M_N        = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
+    using AGridDesc_G_M_K      = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
+    using BGridDesc_G_N_K      = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
+    using B1GridDesc_G_N_K     = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
+    using CGridDesc_G_M_N      = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
+
+    constexpr static auto make_MaskOutPredicate()
+    {
+        if constexpr(MaskingSpec == MaskingSpecialization::MaskDisabled)
+        {
+            return MaskDisabledPredicate{};
+        }
+        else if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+        {
+            return MaskOutUpperTrianglePredicate{};
+        }
+    }
+    using C0MatrixMask = C0MatrixMask_impl<decltype(make_MaskOutPredicate())>;
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k,
+                                     const BGridDesc_G_N_K& b_grid_desc_g_n_k,
+                                     const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
+                                     const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+            : a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
+              b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
+              b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
+              c_grid_desc_g_m_n_(c_grid_desc_g_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return a_grid_desc_g_m_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return b_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
+        {
+            return b1_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return c_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        private:
+        AGridDesc_G_M_K a_grid_desc_g_m_k_;
+        BGridDesc_G_N_K b_grid_desc_g_n_k_;
+        B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
+        CGridDesc_G_M_N c_grid_desc_g_m_n_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        AccElementwiseOperation,
+        B1ElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        B1GridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        AK1,
+        BK1,
+        B1K1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        Gemm1NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        true,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        true,
+        BBlockLdsExtraN,
+        B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B1BlockTransferThreadClusterArrangeOrder,
+        B1BlockTransferSrcAccessOrder,
+        B1BlockTransferSrcVectorDim,
+        B1BlockTransferSrcScalarPerVector,
+        B1BlockTransferDstScalarPerVector_BK1,
+        false,
+        B1BlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        Transform::matrix_padder.PadN,
+        MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle>;
+
+    // Argument
+    // FIXME: constness
+    struct Argument : public BaseArgument
+    {
+        Argument(
+            const ADataType* p_a_grid,
+            const BDataType* p_b_grid,
+            const B1DataType* p_b1_grid,
+            CDataType* p_c_grid,
+            const std::array<void*, NumAcc0Bias> p_acc0_biases,
+            const std::array<void*, NumAcc1Bias> p_acc1_biases,
+            const std::vector<index_t>& a_gs_ms_ks_lengths,
+            const std::vector<index_t>& a_gs_ms_ks_strides,
+            const std::vector<index_t>& b_gs_ns_ks_lengths,
+            const std::vector<index_t>& b_gs_ns_ks_strides,
+            const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+            const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+            const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+            const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+            const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
+            const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
+            const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+                acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+            const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+                acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+            AElementwiseOperation a_element_op,
+            BElementwiseOperation b_element_op,
+            AccElementwiseOperation acc_element_op,
+            B1ElementwiseOperation b1_element_op,
+            CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_b1_grid_{p_b1_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_ak0_m_ak1_{
+                  DeviceOp::MakeAGridDescriptor_AK0_M_AK1(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
+              b_grid_desc_bk0_n_bk1_{
+                  DeviceOp::MakeBGridDescriptor_BK0_N_BK1(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
+              b1_grid_desc_bk0_n_bk1_{DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(
+                  b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
+              c_grid_desc_m_n_{Transform::MakeCGridDescriptor_M_N(c_gs_ms_gemm1ns_lengths,
+                                                                  c_gs_ms_gemm1ns_strides)},
+              a_grid_desc_g_m_k_{
+                  Transform::MakeAGridDescriptor_G_M_K(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
+              b_grid_desc_g_n_k_{
+                  Transform::MakeB0GridDescriptor_G_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
+              b1_grid_desc_g_n_k_{Transform::MakeB1GridDescriptor_G_N_K(
+                  b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
+              c_grid_desc_g_m_n_{Transform::MakeCGridDescriptor_G_M_N(c_gs_ms_gemm1ns_lengths,
+                                                                      c_gs_ms_gemm1ns_strides)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              b1_element_op_{b1_element_op},
+              c_element_op_{c_element_op},
+              c0_matrix_mask_{b_grid_desc_g_n_k_.GetLength(I1)},
+              raw_lengths_mz_nz_kz_gemm1nz_{a_gs_ms_ks_lengths[NumDimG + NumDimM - 1],
+                                            b_gs_ns_ks_lengths[NumDimG + NumDimN - 1],
+                                            b_gs_ns_ks_lengths[NumDimG + NumDimN + NumDimK - 1],
+                                            b1_gs_gemm1ns_gemm1ks_lengths[NumDimG + NumDimO - 1]},
+              a_mz_kz_strides_{a_gs_ms_ks_strides[NumDimG + NumDimM - 1],
+                               a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1]},
+              b_nz_kz_strides_{b_gs_ns_ks_strides[NumDimG + NumDimN - 1],
+                               b_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1]},
+              b1_nz_kz_strides_{b1_gs_gemm1ns_gemm1ks_strides[NumDimG + NumDimO - 1],
+                                b1_gs_gemm1ns_gemm1ks_strides[NumDimG + NumDimO + NumDimN - 1]},
+              c_mz_gemm1nz_strides_{c_gs_ms_gemm1ns_strides[NumDimG + NumDimM - 1],
+                                    c_gs_ms_gemm1ns_strides[NumDimG + NumDimM + NumDimO - 1]},
+              batch_count_{c_grid_desc_g_m_n_.GetLength(I0)},
+              compute_base_ptr_of_batch_{
+                  a_grid_desc_g_m_k_, b_grid_desc_g_n_k_, b1_grid_desc_g_n_k_, c_grid_desc_g_m_n_}
+        {
+            // TODO ANT: implement bias addition
+            ignore = p_acc0_biases;
+            ignore = p_acc1_biases;
+            ignore = acc0_biases_gs_ms_ns_lengths;
+            ignore = acc0_biases_gs_ms_ns_strides;
+            ignore = acc1_biases_gs_ms_gemm1ns_lengths;
+            ignore = acc1_biases_gs_ms_gemm1ns_strides;
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           b1_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "a_grid_desc_g_m_k_: " << a_grid_desc_g_m_k_.GetLength(I0) << ", "
+                      << a_grid_desc_g_m_k_.GetLength(I1) << ", "
+                      << a_grid_desc_g_m_k_.GetLength(I2) << '\n';
+            // a_grid_desc_g_m_k_.Print();
+            std::cout << "b_grid_desc_g_n_k_: " << b_grid_desc_g_n_k_.GetLength(I0) << ", "
+                      << b_grid_desc_g_n_k_.GetLength(I1) << ", "
+                      << b_grid_desc_g_n_k_.GetLength(I2) << '\n';
+            // b_grid_desc_g_n_k_.Print();
+            std::cout << "b1_grid_desc_g_n_k_: " << b1_grid_desc_g_n_k_.GetLength(I0) << ", "
+                      << b1_grid_desc_g_n_k_.GetLength(I1) << ", "
+                      << b1_grid_desc_g_n_k_.GetLength(I2) << '\n';
+            // b1_grid_desc_g_n_k_.Print();
+            std::cout << "c_grid_desc_g_m_n_: " << c_grid_desc_g_m_n_.GetLength(I0) << ", "
+                      << c_grid_desc_g_m_n_.GetLength(I1) << ", "
+                      << c_grid_desc_g_m_n_.GetLength(I2) << '\n';
+            // c_grid_desc_g_m_n_.Print();
+        }
+
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        const B1DataType* p_b1_grid_;
+        CDataType* p_c_grid_;
+
+        // tensor descriptor
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        AGridDesc_G_M_K a_grid_desc_g_m_k_;
+        BGridDesc_G_N_K b_grid_desc_g_n_k_;
+        B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
+        CGridDesc_G_M_N c_grid_desc_g_m_n_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-c-tile map
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        CElementwiseOperation c_element_op_;
+
+        // check C0 masking and padding
+        C0MatrixMask c0_matrix_mask_;
+
+        // For robust IsSupportedArgument() check
+        std::vector<index_t> raw_lengths_mz_nz_kz_gemm1nz_;
+        std::vector<index_t> a_mz_kz_strides_;
+        std::vector<index_t> b_nz_kz_strides_;
+        std::vector<index_t> b1_nz_kz_strides_;
+        std::vector<index_t> c_mz_gemm1nz_strides_;
+
+        index_t batch_count_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!DeviceOp::IsSupportedArgument(arg))
+            {
+                throw std::runtime_error("wrong! unsupported argument");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.batch_count_;
+
+            // Gemm0_K
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_batched_gemm_softmax_gemm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    AccElementwiseOperation,
+                    B1ElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::B1GridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    ComputeBasePtrOfStridedBatch,
+                    C0MatrixMask,
+                    has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_b1_grid_,
+                                              arg.p_c_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.acc_element_op_,
+                                              arg.b1_element_op_,
+                                              arg.c_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.b1_grid_desc_bk0_n_bk1_,
+                                              arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_ctile_map_,
+                                              arg.batch_count_,
+                                              arg.compute_base_ptr_of_batch_,
+                                              arg.c0_matrix_mask_);
+            };
+
+            // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
+            // to concern Gemm0's loop
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+#if 0
+        arg.Print();
+#endif
+
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        // TODO ANT: Check if tensor specialization & strides mismatch
+
+        // Check if C permute dimension matches GEMM + GEMM shape
+        const index_t c_g       = arg.c_grid_desc_g_m_n_.GetLength(I0); // unpadded
+        const index_t c_m       = arg.c_grid_desc_m_n_.GetLength(I0);
+        const index_t c_gemm1n  = arg.c_grid_desc_m_n_.GetLength(I1);
+        const index_t a_m       = arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
+        const index_t b1_gemm1n = arg.b1_grid_desc_bk0_n_bk1_.GetLength(I1);
+
+        if(!(c_g == arg.batch_count_ && c_m == a_m && c_gemm1n == b1_gemm1n))
+        {
+            return false;
+        }
+
+        // Note: we need raw lengths since threadwise copy can not handle vector load when part of
+        // vector is out of bounds
+        // Note: need lowest dim in Ms/Ns/Ks/Os, not merged M/N/K/O
+        const auto MzRaw      = arg.raw_lengths_mz_nz_kz_gemm1nz_[0];
+        const auto NzRaw      = arg.raw_lengths_mz_nz_kz_gemm1nz_[1];
+        const auto KzRaw      = arg.raw_lengths_mz_nz_kz_gemm1nz_[2];
+        const auto Gemm1NzRaw = arg.raw_lengths_mz_nz_kz_gemm1nz_[3];
+
+        // Check scalar per vector requirement
+        const auto a_extent_lowest  = ABlockTransferSrcVectorDim == 2 ? KzRaw : MzRaw;
+        const auto b_extent_lowest  = BBlockTransferSrcVectorDim == 2 ? KzRaw : NzRaw;
+        const auto b1_extent_lowest = B1BlockTransferSrcVectorDim == 2 ? NzRaw : Gemm1NzRaw;
+        const auto c_extent_lowest  = Gemm1NzRaw;
+
+        if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
+             b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
+             b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
+             c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // Check vector load/store requirement
+        const auto a_stride_lowest =
+            ABlockTransferSrcVectorDim == 2 ? arg.a_mz_kz_strides_[1] : arg.a_mz_kz_strides_[0];
+        const auto b_stride_lowest =
+            BBlockTransferSrcVectorDim == 2 ? arg.b_nz_kz_strides_[1] : arg.b_nz_kz_strides_[0];
+        const auto b1_stride_lowest =
+            B1BlockTransferSrcVectorDim == 2 ? arg.b1_nz_kz_strides_[1] : arg.b1_nz_kz_strides_[0];
+        const auto c_stride_lowest =
+            arg.c_mz_gemm1nz_strides_[1]; // cshuffle assumes lowest dim in Gemm1Ns to be contiguous
+
+        if(!(a_stride_lowest == 1 || b_stride_lowest == 1 || b1_stride_lowest == 1 ||
+             c_stride_lowest == 1))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.b1_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const ADataType* p_a,
+        const BDataType* p_b,
+        const B1DataType* p_b1,
+        CDataType* p_c,
+        const std::array<void*, NumAcc0Bias> p_acc0_biases,
+        const std::array<void*, NumAcc1Bias> p_acc1_biases,
+        const std::vector<index_t>& a_gs_ms_ks_lengths,
+        const std::vector<index_t>& a_gs_ms_ks_strides,
+        const std::vector<index_t>& b_gs_ns_ks_lengths,
+        const std::vector<index_t>& b_gs_ns_ks_strides,
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+        const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+        const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+        AElementwiseOperation a_element_op,
+        BElementwiseOperation b_element_op,
+        AccElementwiseOperation acc_element_op,
+        B1ElementwiseOperation b1_element_op,
+        CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_b1,
+                        p_c,
+                        p_acc0_biases,
+                        p_acc1_biases,
+                        a_gs_ms_ks_lengths,
+                        a_gs_ms_ks_strides,
+                        b_gs_ns_ks_lengths,
+                        b_gs_ns_ks_strides,
+                        b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+                        b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+                        c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+                        c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+                        acc0_biases_gs_ms_ns_lengths,
+                        acc0_biases_gs_ms_ns_strides,
+                        acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+                        acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+                        a_element_op,
+                        b_element_op,
+                        acc_element_op,
+                        b1_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    // FIXME: constness
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const void* p_b1,
+        void* p_c,
+        const std::array<void*, NumAcc0Bias> p_acc0_biases,
+        const std::array<void*, NumAcc1Bias> p_acc1_biases,
+        const std::vector<index_t>& a_gs_ms_ks_lengths,
+        const std::vector<index_t>& a_gs_ms_ks_strides,
+        const std::vector<index_t>& b_gs_ns_ks_lengths,
+        const std::vector<index_t>& b_gs_ns_ks_strides,
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+        const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+        const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+        AElementwiseOperation a_element_op,
+        BElementwiseOperation b_element_op,
+        AccElementwiseOperation acc_element_op,
+        B1ElementwiseOperation b1_element_op,
+        CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<const B1DataType*>(p_b1),
+                                          static_cast<CDataType*>(p_c),
+                                          p_acc0_biases, // cast in struct Argument
+                                          p_acc1_biases, // cast in struct Argument
+                                          a_gs_ms_ks_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b_gs_ns_ks_lengths,
+                                          b_gs_ns_ks_strides,
+                                          b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+                                          b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+                                          c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+                                          c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+                                          acc0_biases_gs_ms_ns_lengths,
+                                          acc0_biases_gs_ms_ns_strides,
+                                          acc1_biases_gs_ms_gemm1ns_lengths,
+                                          acc1_biases_gs_ms_gemm1ns_strides,
+                                          a_element_op,
+                                          b_element_op,
+                                          acc_element_op,
+                                          b1_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerBlock << ", "
+            << Gemm1NPerBlock << ", "
+            << Gemm1KPerBlock << ", "
+            << B1K1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << "ASpec" << getTensorSpecializationString(ASpec) << ", "
+            << "B0Spec" << getTensorSpecializationString(BSpec) << ", "
+            << "B1Spec" << getTensorSpecializationString(B1Spec) << ", "
+            << "CSpec" << getTensorSpecializationString(CSpec) << ", "
+            << getMaskingSpecializationString(MaskingSpec) << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
new file mode 100644
index 0000000..1f21f2d
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
@@ -0,0 +1,771 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2CTileMap,
+          typename ComputeBasePtrOfStridedBatch,
+          typename C0MatrixMask,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_softmax_gemm_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            const FloatAB* __restrict__ p_b1_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const AccElementwiseOperation acc_element_op,
+            const B1ElementwiseOperation b1_element_op,
+            const CElementwiseOperation c_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2CTileMap block_2_ctile_map,
+            const index_t batch_count,
+            const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
+            const C0MatrixMask c0_matrix_mask)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetBBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetCBasePtr(g_idx)));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_b1_grid + b1_batch_offset,
+                                                  p_c_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  acc_element_op,
+                                                  b1_element_op,
+                                                  c_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  b1_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_ctile_map,
+                                                  c0_matrix_mask);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_b1_grid;
+    ignore = p_c_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = acc_element_op;
+    ignore = b1_element_op;
+    ignore = c_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = b1_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_ctile_map;
+    ignore = batch_count;
+    ignore = compute_base_ptr_of_batch;
+    ignore = c0_matrix_mask;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// Computes C = A * B0 * B1
+//              ^^^^^^ (Acc0)
+//              ^^^^^^^^^^^ (Acc1)
+
+// When using NPadding as GemmSpecialization, AccElementwiseOperation should be set to
+// ScaleAndResetNaNToMinusInfinity.
+// if !isNan(AccElement)
+//     AccElement *= scale
+// else
+//     AccElement = -INFINITY
+// Otherwise, result may be wrong.
+
+template <typename ALayout,
+          typename BLayout, // B0Layout
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename B1DataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock, // Gemm0NPerBlock
+          index_t KPerBlock, // Gemm0KPerBlock
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t B1K1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          bool MaskOutUpperTriangle,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
+    : public DeviceBatchedGemmSoftmaxGemm<ALayout,
+                                          BLayout,
+                                          B1Layout,
+                                          CLayout,
+                                          ADataType,
+                                          BDataType,
+                                          B1DataType,
+                                          CDataType,
+                                          AElementwiseOperation,
+                                          BElementwiseOperation,
+                                          AccElementwiseOperation,
+                                          B1ElementwiseOperation,
+                                          CElementwiseOperation,
+                                          MaskOutUpperTriangle>
+{
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto matrix_padder =
+        GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
+            MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // Args: Gemm1KRaw, Gemm1NRaw, StrideB1
+    static auto MakeB1GridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b1_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b1_grid_desc_n_k = matrix_padder.PadB1Descriptor_N_K(b1_grid_desc_nraw_kraw);
+
+        const auto N = b1_grid_desc_n_k.GetLength(I0);
+        const auto K = b1_grid_desc_n_k.GetLength(I1);
+
+        const auto B1K0 = K / B1K1;
+
+        return transform_tensor_descriptor(
+            b1_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
+    }
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
+                                     index_t BatchStrideB,
+                                     index_t BatchStrideB1,
+                                     index_t BatchStrideC)
+            : BatchStrideA_(BatchStrideA),
+              BatchStrideB_(BatchStrideB),
+              BatchStrideB1_(BatchStrideB1),
+              BatchStrideC_(BatchStrideC)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB1_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideC_);
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        index_t BatchStrideB1_;
+        index_t BatchStrideC_;
+    };
+
+    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N        = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    using C0MatrixMask = conditional_t<MaskOutUpperTriangle,
+                                       C0MatrixMask_impl<MaskOutUpperTrianglePredicate>,
+                                       C0MatrixMask_impl<MaskDisabledPredicate>>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        AccElementwiseOperation,
+        B1ElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        B1GridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        AK1,
+        BK1,
+        B1K1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        Gemm1NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        true,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        true,
+        BBlockLdsExtraN,
+        B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B1BlockTransferThreadClusterArrangeOrder,
+        B1BlockTransferSrcAccessOrder,
+        B1BlockTransferSrcVectorDim,
+        B1BlockTransferSrcScalarPerVector,
+        B1BlockTransferDstScalarPerVector_BK1,
+        false,
+        B1BlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        matrix_padder.PadN,
+        MaskOutUpperTriangle>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 const B1DataType* p_b1_grid,
+                 CDataType* p_c_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t Gemm1NRaw, // = ORaw
+                 index_t Batch,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideB1,
+                 index_t StrideC,
+                 index_t BatchStrideA,
+                 index_t BatchStrideB,
+                 index_t BatchStrideB1,
+                 index_t BatchStrideC,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 AccElementwiseOperation acc_element_op,
+                 B1ElementwiseOperation b1_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_b1_grid_{p_b1_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              b1_grid_desc_bk0_n_bk1_{
+                  DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(NRaw, Gemm1NRaw, StrideB1)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, Gemm1NRaw, StrideC)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              b1_element_op_{b1_element_op},
+              c_element_op_{c_element_op},
+              batch_count_(Batch),
+              compute_base_ptr_of_batch_{BatchStrideA, BatchStrideB, BatchStrideB1, BatchStrideC},
+              c0_matrix_mask_{NRaw},
+              raw_lengths_m_n_k_o_{MRaw, NRaw, KRaw, Gemm1NRaw}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           b1_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        const B1DataType* p_b1_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        CElementwiseOperation c_element_op_;
+        index_t batch_count_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+
+        // check C0 masking and padding
+        C0MatrixMask c0_matrix_mask_;
+
+        // For robust IsSupportedArgument() check
+        std::vector<index_t> raw_lengths_m_n_k_o_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                            arg.b_grid_desc_bk0_n_bk1_,
+                                            arg.b1_grid_desc_bk0_n_bk1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.batch_count_;
+
+            // Gemm0_K
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_batched_gemm_softmax_gemm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    AccElementwiseOperation,
+                    B1ElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::B1GridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    ComputeBasePtrOfStridedBatch,
+                    C0MatrixMask,
+                    has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_b1_grid_,
+                                              arg.p_c_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.acc_element_op_,
+                                              arg.b1_element_op_,
+                                              arg.c_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.b1_grid_desc_bk0_n_bk1_,
+                                              arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_ctile_map_,
+                                              arg.batch_count_,
+                                              arg.compute_base_ptr_of_batch_,
+                                              arg.c0_matrix_mask_);
+            };
+
+            // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
+            // to concern Gemm0's loop
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        // Note: we need raw lengths since threadwise copy can not handle vector load when part of
+        // vector is out of bounds
+        const auto MRaw      = arg.raw_lengths_m_n_k_o_[0];
+        const auto NRaw      = arg.raw_lengths_m_n_k_o_[1];
+        const auto KRaw      = arg.raw_lengths_m_n_k_o_[2];
+        const auto Gemm1NRaw = arg.raw_lengths_m_n_k_o_[3];
+
+        // Check scalar per vector requirement
+        const auto a_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, ALayout> ? KRaw : MRaw;
+        const auto b_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, BLayout> ? NRaw : KRaw;
+        const auto b1_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, B1Layout> ? Gemm1NRaw : NRaw;
+        const auto c_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, CLayout> ? Gemm1NRaw : MRaw;
+
+        if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
+             b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
+             b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
+             c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.b1_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             const B1DataType* p_b1,
+                             CDataType* p_c,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t Gemm1NRaw,
+                             index_t Batch,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideB1,
+                             index_t StrideC,
+                             index_t BatchStrideA,
+                             index_t BatchStrideB,
+                             index_t BatchStrideB1,
+                             index_t BatchStrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             AccElementwiseOperation acc_element_op,
+                             B1ElementwiseOperation b1_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,           p_b,          p_b1,         p_c,          MRaw,
+                        NRaw,          KRaw,         Gemm1NRaw,    Batch,        StrideA,
+                        StrideB,       StrideB1,     StrideC,      BatchStrideA, BatchStrideB,
+                        BatchStrideB1, BatchStrideC, a_element_op, b_element_op, acc_element_op,
+                        b1_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      const void* p_b1,
+                                                      void* p_c,
+                                                      index_t MRaw,
+                                                      index_t NRaw,
+                                                      index_t KRaw,
+                                                      index_t Gemm1NRaw,
+                                                      index_t Batch,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideB1,
+                                                      index_t StrideC,
+                                                      index_t BatchStrideA,
+                                                      index_t BatchStrideB,
+                                                      index_t BatchStrideB1,
+                                                      index_t BatchStrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      AccElementwiseOperation acc_element_op,
+                                                      B1ElementwiseOperation b1_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<const B1DataType*>(p_b1),
+                                          static_cast<CDataType*>(p_c),
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          Gemm1NRaw,
+                                          Batch,
+                                          StrideA,
+                                          StrideB,
+                                          StrideB1,
+                                          StrideC,
+                                          BatchStrideA,
+                                          BatchStrideB,
+                                          BatchStrideB1,
+                                          BatchStrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          acc_element_op,
+                                          b1_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerBlock << ", "
+            << Gemm1NPerBlock << ", "
+            << Gemm1KPerBlock << ", "
+            << B1K1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
new file mode 100644
index 0000000..5ea3296
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
@@ -0,0 +1,668 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2CTileMap Block2CTileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2CTileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_xdlops_v2r3(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const index_t batch_count,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_c_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector,
+          ck::index_t NumGemmKPrefetchStage = 1,
+          ck::LoopScheduler LoopSched       = make_default_loop_scheduler(),
+          ck::PipelineVersion PipelineVer   = ck::PipelineVersion::v1>
+struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
+                                                       BLayout,
+                                                       CLayout,
+                                                       ADataType,
+                                                       BDataType,
+                                                       CDataType,
+                                                       AElementwiseOperation,
+                                                       BElementwiseOperation,
+                                                       CElementwiseOperation>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+
+        const auto a_grid_desc_k0_mp_k1 =
+            transform_tensor_descriptor(a_grid_desc_m_k,
+                                        make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                                                   make_right_pad_transform(M, PadM)),
+                                        make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                        make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        return a_grid_desc_k0_mp_k1;
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+        const auto b_grid_desc_k0_np_k1 =
+            transform_tensor_descriptor(b_grid_desc_k_n,
+                                        make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                                                   make_right_pad_transform(N, PadN)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        return b_grid_desc_k0_np_k1;
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+        const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+        const auto c_grid_desc_mp_np = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return c_grid_desc_mp_np;
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using CGridDesc_M_N     = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                       index_t BatchStrideB,
+                                       index_t BatchStrideC)
+            : BatchStrideA_(BatchStrideA), BatchStrideB_(BatchStrideB), BatchStrideC_(BatchStrideC)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideC_);
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        index_t BatchStrideC_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<BlockSize,
+                                                ADataType, // TODO: distinguish A/B datatype
+                                                AccDataType,
+                                                CDataType,
+                                                InMemoryDataOperationEnum::Set,
+                                                AGridDesc_K0_M_K1,
+                                                BGridDesc_K0_N_K1,
+                                                CGridDesc_M_N,
+                                                AElementwiseOperation,
+                                                BElementwiseOperation,
+                                                CElementwiseOperation,
+                                                MPerBlock,
+                                                NPerBlock,
+                                                K0PerBlock,
+                                                MPerXDL,
+                                                NPerXDL,
+                                                K1,
+                                                MXdlPerWave,
+                                                NXdlPerWave,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ABlockTransferSrcAccessOrder,
+                                                ABlockTransferSrcVectorDim,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                false, // AThreadTransferSrcResetCoordinateAfterRun,
+                                                ABlockLdsAddExtraM,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                BBlockTransferSrcAccessOrder,
+                                                BBlockTransferSrcVectorDim,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                false, // BThreadTransferSrcResetCoordinateAfterRun,
+                                                BBlockLdsAddExtraN,
+                                                Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
+                                                CThreadTransferSrcDstVectorDim,
+                                                CThreadTransferDstScalarPerVector,
+                                                NumGemmKPrefetchStage,
+                                                LoopSched,
+                                                PipelineVer>;
+
+    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
+    using Block2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 index_t BatchStrideA,
+                 index_t BatchStrideB,
+                 index_t BatchStrideC,
+                 index_t Batch,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              Batch_(Batch),
+              a_grid_desc_k0_m_k1_{
+                  DeviceBatchedGemmXdl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA)},
+              b_grid_desc_k0_n_k1_{
+                  DeviceBatchedGemmXdl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB)},
+              c_grid_desc_m_n_{DeviceBatchedGemmXdl::MakeCGridDescriptor_M_N(M, N, StrideC)},
+              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
+              compute_ptr_offset_of_batch_{BatchStrideA, BatchStrideB, BatchStrideC},
+              block_2_ctile_map_{
+                  GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01)},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              kraw_{K}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
+                                           b_grid_desc_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        index_t Batch_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+        Block2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+        index_t kraw_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceBatchedGemmXdl::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{" << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseBatchedGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.Batch_;
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_batched_gemm_xdlops_v2r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceBatchedGemmXdl::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceBatchedGemmXdl::BGridDesc_K0_N_K1>,
+                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    ComputePtrOffsetOfStridedBatch,
+                    remove_reference_t<Block2CTileMap>,
+                    true>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.Batch_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.compute_ptr_offset_of_batch_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_batched_gemm_xdlops_v2r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceBatchedGemmXdl::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceBatchedGemmXdl::BGridDesc_K0_N_K1>,
+                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    ComputePtrOffsetOfStridedBatch,
+                    remove_reference_t<Block2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.Batch_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.compute_ptr_offset_of_batch_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(arg.kraw_ % K1 != 0)
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             index_t BatchStrideA,
+                             index_t BatchStrideB,
+                             index_t BatchStrideC,
+                             index_t Batch,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        BatchStrideA,
+                        BatchStrideB,
+                        BatchStrideC,
+                        Batch,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      index_t BatchStrideA,
+                                                      index_t BatchStrideB,
+                                                      index_t BatchStrideC,
+                                                      index_t Batch,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          BatchStrideA,
+                                          BatchStrideB,
+                                          BatchStrideC,
+                                          Batch,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceBatchedGemmXdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock
+            << ">"
+            << " NumGemmKPrefetchStage: "
+            << NumGemmKPrefetchStage << ", "
+            << "LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp
new file mode 100644
index 0000000..ab16a75
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp
@@ -0,0 +1,874 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/welford_helper.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim,
+          bool UseMultiblockInK,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XDyDxVectorDim,
+          index_t XSrcVectorSize,
+          index_t DySrcVectorSize,
+          index_t DxDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t DscaleDbiasDstVectorSize,
+          index_t MeanVarSrcVectorSize>
+struct DeviceBatchNormBwdImpl : public DeviceBatchNormBwd<XDataType,
+                                                          DxDataType,
+                                                          DyDataType,
+                                                          AccDataType,
+                                                          ScaleDataType,
+                                                          DscaleDbiasDataType,
+                                                          MeanVarDataType,
+                                                          DyElementwiseOp,
+                                                          Rank,
+                                                          NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "Invalid thread cluster size assignments!");
+
+    static_assert((XDyDxVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0 &&
+                   MThreadSliceSize % DySrcVectorSize == 0 &&
+                   MThreadSliceSize % DxDstVectorSize == 0) ||
+                      (XDyDxVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0 &&
+                       KThreadSliceSize % DySrcVectorSize == 0 &&
+                       KThreadSliceSize % DxDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeXY2dDescriptor(const std::array<index_t, Rank>& xyLengths,
+                                   const std::array<index_t, Rank>& xyStrides,
+                                   int blkGroupSize,
+                                   int numBlockTileIteration)
+    {
+        const auto tupleXYLengths =
+            generate_tuple([&](auto I) { return xyLengths[I]; }, Number<Rank>{});
+        const auto tupleXYStrides =
+            generate_tuple([&](auto I) { return xyStrides[I]; }, Number<Rank>{});
+
+        const auto raw_grid_desc = make_naive_tensor_descriptor(tupleXYLengths, tupleXYStrides);
+
+        const auto grid_desc_m_k = [&]() {
+            using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+            using ReduceDims    = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+            const auto reduceDimLengths =
+                generate_tuple([&](auto I) { return xyLengths[NumInvariantDim + I]; },
+                               Number<NumBatchNormReduceDim>{});
+            const auto invariantDimLengths =
+                generate_tuple([&](auto I) { return xyLengths[I]; }, Number<NumInvariantDim>{});
+
+            return transform_tensor_descriptor(raw_grid_desc,
+                                               make_tuple(make_merge_transform(invariantDimLengths),
+                                                          make_merge_transform(reduceDimLengths)),
+                                               make_tuple(InvariantDims{}, ReduceDims{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }();
+
+        const auto invariantLength = grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = grid_desc_m_k.GetLength(Number<1>{});
+
+        const int workSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto kPad = workSizePerBlock * blkGroupSize - reduceLength;
+
+        auto grid_desc_m_k_padded =
+            transform_tensor_descriptor(grid_desc_m_k,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_right_pad_transform(reduceLength, kPad)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_k_padded);
+    };
+
+    static auto MakeMultiblockFirstReduceOutputMG2dDescriptor(int invariantLength, int blkGroupSize)
+    {
+        const auto grid_desc_m_g =
+            make_naive_tensor_descriptor_packed(make_tuple(invariantLength, blkGroupSize));
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_g_padded =
+            transform_tensor_descriptor(grid_desc_m_g,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_pass_through_transform(blkGroupSize)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_g_padded);
+    };
+
+    static auto MakeMultiblockFinalReduceInputMK2dDescriptor(int invariantLength, int blkGroupSize)
+    {
+        const auto reduceLength = blkGroupSize;
+        const auto grid_desc_m_k =
+            make_naive_tensor_descriptor_packed(make_tuple(invariantLength, reduceLength));
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto kPad =
+            math::integer_least_multiple(reduceLength, KThreadClusterSize) - reduceLength;
+
+        auto grid_desc_m_k_padded =
+            transform_tensor_descriptor(grid_desc_m_k,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_right_pad_transform(reduceLength, kPad)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_k_padded);
+    };
+
+    static auto
+    MakeScaleBiasMeanVar1dDescriptor(const std::array<index_t, NumInvariantDim>& lengths,
+                                     const std::array<index_t, NumInvariantDim>& strides)
+    {
+        const auto tupleLengths =
+            generate_tuple([&](auto I) { return lengths[I]; }, Number<NumInvariantDim>{});
+        const auto tupleStrides =
+            generate_tuple([&](auto I) { return strides[I]; }, Number<NumInvariantDim>{});
+
+        auto raw_grid_desc = make_naive_tensor_descriptor(tupleLengths, tupleStrides);
+
+        auto grid_desc_m = transform_tensor_descriptor(
+            raw_grid_desc,
+            make_tuple(make_merge_transform(tupleLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = grid_desc_m.GetLength(Number<0>{});
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_padded =
+            transform_tensor_descriptor(grid_desc_m,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return (grid_desc_m_padded);
+    };
+
+    using XYGridDesc_M_K      = decltype(MakeXY2dDescriptor({1}, {1}, 1, 1));
+    using ScaleBiasGridDesc_M = decltype(MakeScaleBiasMeanVar1dDescriptor({1}, {1}));
+    using MeanVarGridDesc_M   = ScaleBiasGridDesc_M;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> dyStrides,
+                 const std::array<index_t, Rank> dxStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<ck::index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                 const std::array<ck::index_t, NumInvariantDim> bnScaleStrides,
+                 const std::array<ck::index_t, NumInvariantDim> bnDscaleDbiasStrides,
+                 const std::array<ck::index_t, NumInvariantDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const DyDataType* p_dy,
+                 const ScaleDataType* p_scale,
+                 const MeanVarDataType* p_savedMean,
+                 const MeanVarDataType* p_savedInvVar,
+                 const DyElementwiseOp dy_elementwise_op,
+                 double epsilon,
+                 DxDataType* p_dx,
+                 DscaleDbiasDataType* p_dscale,
+                 DscaleDbiasDataType* p_dbias)
+            : bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnDscaleDbiasStrides_(bnDscaleDbiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              p_dy_(p_dy),
+              p_scale_(p_scale),
+              p_savedMean_(p_savedMean),
+              p_savedInvVar_(p_savedInvVar),
+              dy_elementwise_op_(dy_elementwise_op),
+              p_dx_(p_dx),
+              p_dscale_(p_dscale),
+              p_dbias_(p_dbias)
+        {
+            xyLengths_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xyLengths, reduceDims);
+            xStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xStrides, reduceDims);
+            dyStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(dyStrides, reduceDims);
+            dxStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(dxStrides, reduceDims);
+
+            std::tie(invariant_length, reduce_length) =
+                get_2d_lengths<Rank, NumBatchNormReduceDim>(xyLengths_);
+
+            epsilon_ = type_convert<AccDataType>(epsilon);
+
+            haveSavedMeanInvVar_ = (p_savedMean_ != nullptr && p_savedInvVar_ != nullptr);
+
+            if(UseMultiblockInK)
+            {
+                int iterations = 1;
+                while(true)
+                {
+                    int testBlkGroupSize = (reduce_length + (K_BlockTileSize * iterations) - 1) /
+                                           (K_BlockTileSize * iterations);
+
+                    // we want the blkGroupSize be not more than 128
+                    if(testBlkGroupSize <= 128)
+                        break;
+
+                    iterations++;
+                };
+
+                blkGroupSize = (reduce_length + (K_BlockTileSize * iterations) - 1) /
+                               (K_BlockTileSize * iterations);
+
+                numBlockTileIteration = iterations;
+            }
+            else
+            {
+                blkGroupSize          = 1;
+                numBlockTileIteration = (reduce_length + K_BlockTileSize - 1) / K_BlockTileSize;
+            };
+
+            gridSize = (invariant_length + M_BlockTileSize - 1) / M_BlockTileSize * blkGroupSize;
+
+            x_grid_desc_m_k =
+                MakeXY2dDescriptor(xyLengths_, xStrides_, blkGroupSize, numBlockTileIteration);
+            dy_grid_desc_m_k =
+                MakeXY2dDescriptor(xyLengths_, dyStrides_, blkGroupSize, numBlockTileIteration);
+            dx_grid_desc_m_k =
+                MakeXY2dDescriptor(xyLengths_, dxStrides_, blkGroupSize, numBlockTileIteration);
+            scale_grid_desc_m =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnScaleStrides);
+            dscale_dbias_grid_desc_m =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnDscaleDbiasStrides);
+            mean_var_grid_desc_m =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnMeanVarStrides);
+        }
+
+        AccDataType epsilon_;
+
+        bool haveSavedMeanInvVar_;
+
+        std::array<index_t, Rank> xyLengths_;
+        std::array<index_t, Rank> xStrides_;
+        std::array<index_t, Rank> dyStrides_;
+        std::array<index_t, Rank> dxStrides_;
+
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnDscaleDbiasStrides_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides_;
+
+        const XDataType* p_x_;
+        const DyDataType* p_dy_;
+        const ScaleDataType* p_scale_;
+        const MeanVarDataType* p_savedMean_;
+        const MeanVarDataType* p_savedInvVar_;
+        const DyElementwiseOp dy_elementwise_op_;
+        DxDataType* p_dx_;
+        DscaleDbiasDataType* p_dscale_;
+        DscaleDbiasDataType* p_dbias_;
+
+        long_index_t invariant_length;
+        long_index_t reduce_length;
+
+        int blkGroupSize;
+        int numBlockTileIteration;
+        size_t gridSize;
+
+        XYGridDesc_M_K x_grid_desc_m_k;
+        XYGridDesc_M_K dy_grid_desc_m_k;
+        XYGridDesc_M_K dx_grid_desc_m_k;
+        ScaleBiasGridDesc_M scale_grid_desc_m;
+        ScaleBiasGridDesc_M dscale_dbias_grid_desc_m;
+        MeanVarGridDesc_M mean_var_grid_desc_m;
+
+        void* workspace_mean;
+        void* workspace_variance;
+        void* workspace_count;
+
+        void* workspace_savedMean;
+        void* workspace_savedInvVar;
+
+        void* workspace_reduce_dscale;
+        void* workspace_reduce_dbias;
+    };
+
+    size_t GetWorkSpaceSize(const BaseArgument* pArg) const override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        size_t workspace_size = 0;
+
+        if(UseMultiblockInK && pArg_->blkGroupSize > 1)
+        {
+            // workspace for the partial reduced result for dscale
+            workspace_size +=
+                pArg_->invariant_length * pArg_->blkGroupSize * sizeof(DscaleDbiasDataType) + 64;
+
+            // workspace for the partial reduced result for dbias
+            workspace_size +=
+                pArg_->invariant_length * pArg_->blkGroupSize * sizeof(DscaleDbiasDataType) + 64;
+
+            if(!pArg_->haveSavedMeanInvVar_)
+            {
+                // workspace for welford intermediate mean
+                workspace_size +=
+                    pArg_->invariant_length * pArg_->blkGroupSize * sizeof(MeanVarDataType) + 64;
+
+                // workspace for welford intermediate variance
+                workspace_size +=
+                    pArg_->invariant_length * pArg_->blkGroupSize * sizeof(MeanVarDataType) + 64;
+
+                // workspace for welford intermediate count
+                workspace_size +=
+                    pArg_->invariant_length * pArg_->blkGroupSize * sizeof(int32_t) + 64;
+
+                // workspace for welford result mean
+                workspace_size += pArg_->invariant_length * sizeof(MeanVarDataType) + 64;
+
+                // workspace for welford result inv_variance
+                workspace_size += pArg_->invariant_length * sizeof(MeanVarDataType) + 64;
+            };
+        }
+
+        return (workspace_size);
+    };
+
+    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    {
+        Argument* pArg_ = dynamic_cast<Argument*>(pArg);
+
+        pArg_->p_workspace_ = p_workspace;
+
+        index_t space_sz;
+
+        // setup buffer for the partial reduced result for dscale
+        pArg_->workspace_reduce_dscale = pArg_->p_workspace_;
+
+        space_sz = pArg_->invariant_length * pArg_->blkGroupSize * sizeof(DscaleDbiasDataType);
+        space_sz = math::integer_least_multiple(space_sz, 64);
+
+        // setup buffer for the partial reduced result for dbias
+        pArg_->workspace_reduce_dbias =
+            reinterpret_cast<char*>(pArg_->workspace_reduce_dscale) + space_sz;
+
+        if(UseMultiblockInK && pArg_->blkGroupSize > 1)
+        {
+            space_sz = pArg_->invariant_length * pArg_->blkGroupSize * sizeof(DscaleDbiasDataType);
+            space_sz = math::integer_least_multiple(space_sz, 64);
+
+            // setup buffer for welford intermediate mean
+            pArg_->workspace_mean =
+                reinterpret_cast<char*>(pArg_->workspace_reduce_dbias) + space_sz;
+
+            space_sz = pArg_->invariant_length * pArg_->blkGroupSize * sizeof(MeanVarDataType);
+            space_sz = math::integer_least_multiple(space_sz, 64);
+
+            // setup buffer for welford intermediate varirance
+            pArg_->workspace_variance = reinterpret_cast<char*>(pArg_->workspace_mean) + space_sz;
+
+            space_sz = pArg_->invariant_length * pArg_->blkGroupSize * sizeof(MeanVarDataType);
+            space_sz = math::integer_least_multiple(space_sz, 64);
+
+            // setup buffer for welford intermediate count
+            pArg_->workspace_count = reinterpret_cast<char*>(pArg_->workspace_variance) + space_sz;
+
+            space_sz = pArg_->invariant_length * pArg_->blkGroupSize * sizeof(int32_t);
+            space_sz = math::integer_least_multiple(space_sz, 64);
+
+            // setup buffer for welford result mean
+            pArg_->workspace_savedMean = reinterpret_cast<char*>(pArg_->workspace_count) + space_sz;
+
+            space_sz = pArg_->invariant_length * sizeof(MeanVarDataType);
+            space_sz = math::integer_least_multiple(space_sz, 64);
+
+            // setup buffer for welford result inv_variance
+            pArg_->workspace_savedInvVar =
+                reinterpret_cast<char*>(pArg_->workspace_savedMean) + space_sz;
+        };
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float avg_time = 0;
+
+            const auto mean_var_count_grid_desc_m_g =
+                DeviceBatchNormBwdImpl::MakeMultiblockFirstReduceOutputMG2dDescriptor(
+                    arg.invariant_length, arg.blkGroupSize);
+
+            const auto dscale_dbias_grid_desc_m_g =
+                DeviceBatchNormBwdImpl::MakeMultiblockFirstReduceOutputMG2dDescriptor(
+                    arg.invariant_length, arg.blkGroupSize);
+
+            const auto mean_var_count_grid_desc_m_k =
+                DeviceBatchNormBwdImpl::MakeMultiblockFinalReduceInputMK2dDescriptor(
+                    arg.invariant_length, arg.blkGroupSize);
+
+            const auto dscale_dbias_grid_desc_m_k =
+                DeviceBatchNormBwdImpl::MakeMultiblockFinalReduceInputMK2dDescriptor(
+                    arg.invariant_length, arg.blkGroupSize);
+
+            using MeanVarCountGridDesc_M_G = decltype(mean_var_count_grid_desc_m_g);
+            using MeanVarCountGridDesc_M_K = decltype(mean_var_count_grid_desc_m_k);
+            using DscaleDbiasGridDesc_M_G  = decltype(dscale_dbias_grid_desc_m_g);
+            using DscaleDbiasGridDesc_M_K  = decltype(dscale_dbias_grid_desc_m_k);
+
+            using GridwiseWelfordSecondHalfReduceFirstHalf_ =
+                GridwiseWelfordSecondHalfReduceFirstHalf<XDataType,
+                                                         DyDataType,
+                                                         AccDataType,
+                                                         ScaleDataType,
+                                                         DscaleDbiasDataType,
+                                                         MeanVarDataType,
+                                                         DyElementwiseOp,
+                                                         XYGridDesc_M_K,
+                                                         MeanVarGridDesc_M,
+                                                         MeanVarCountGridDesc_M_K,
+                                                         DscaleDbiasGridDesc_M_G,
+                                                         BlockSize,
+                                                         MThreadClusterSize,
+                                                         KThreadClusterSize,
+                                                         MThreadSliceSize,
+                                                         KThreadSliceSize,
+                                                         XDyDxVectorDim,
+                                                         XSrcVectorSize,
+                                                         DySrcVectorSize,
+                                                         MeanVarSrcVectorSize>;
+
+            using GridwiseReduceSecondHalfBatchNormBwdFinal_ =
+                GridwiseReduceSecondHalfBatchNormBackwardFinal<XDataType,
+                                                               DyDataType,
+                                                               DxDataType,
+                                                               AccDataType,
+                                                               ScaleDataType,
+                                                               DscaleDbiasDataType,
+                                                               MeanVarDataType,
+                                                               DyElementwiseOp,
+                                                               XYGridDesc_M_K,
+                                                               DscaleDbiasGridDesc_M_K,
+                                                               MeanVarGridDesc_M,
+                                                               ScaleBiasGridDesc_M,
+                                                               BlockSize,
+                                                               MThreadClusterSize,
+                                                               KThreadClusterSize,
+                                                               MThreadSliceSize,
+                                                               KThreadSliceSize,
+                                                               XDyDxVectorDim,
+                                                               XSrcVectorSize,
+                                                               DySrcVectorSize,
+                                                               DxDstVectorSize,
+                                                               ScaleSrcVectorSize,
+                                                               DscaleDbiasDstVectorSize,
+                                                               MeanVarSrcVectorSize>;
+
+            if(UseMultiblockInK && arg.blkGroupSize > 1)
+            {
+                using GetReduceCountPerThreadFunctor =
+                    GetReduceCountPerThreadForMultiblockWelford<K_BlockTileSize, KThreadSliceSize>;
+
+                GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
+                    arg.blkGroupSize, arg.numBlockTileIteration, arg.reduce_length);
+
+                if(!arg.haveSavedMeanInvVar_)
+                {
+                    using GridwiseMultiblockWelfordFirstHalf_ =
+                        GridwiseMultiblockWelfordFirstHalf<XDataType,
+                                                           AccDataType,
+                                                           MeanVarDataType,
+                                                           XYGridDesc_M_K,
+                                                           MeanVarCountGridDesc_M_G,
+                                                           GetReduceCountPerThreadFunctor,
+                                                           BlockSize,
+                                                           MThreadClusterSize,
+                                                           KThreadClusterSize,
+                                                           MThreadSliceSize,
+                                                           KThreadSliceSize,
+                                                           XDyDxVectorDim,
+                                                           XSrcVectorSize>;
+
+                    const auto kern_multiblock_welford_first_half =
+                        kernel_multiblock_welford_first_half<GridwiseMultiblockWelfordFirstHalf_,
+                                                             XDataType,
+                                                             MeanVarDataType,
+                                                             XYGridDesc_M_K,
+                                                             MeanVarCountGridDesc_M_G,
+                                                             GetReduceCountPerThreadFunctor>;
+
+                    avg_time += launch_and_time_kernel(
+                        stream_config,
+                        kern_multiblock_welford_first_half,
+                        dim3(arg.gridSize),
+                        dim3(BlockSize),
+                        0,
+                        arg.x_grid_desc_m_k,
+                        mean_var_count_grid_desc_m_g,
+                        get_reduce_count_per_thread,
+                        arg.numBlockTileIteration,
+                        arg.p_x_,
+                        static_cast<MeanVarDataType*>(arg.workspace_mean),
+                        static_cast<MeanVarDataType*>(arg.workspace_variance),
+                        static_cast<int32_t*>(arg.workspace_count));
+                };
+
+                const auto kern_welford_second_half_reduce_first_half =
+                    kernel_welford_second_half_reduce_first_half<
+                        GridwiseWelfordSecondHalfReduceFirstHalf_,
+                        XDataType,
+                        DyDataType,
+                        AccDataType,
+                        ScaleDataType,
+                        DscaleDbiasDataType,
+                        MeanVarDataType,
+                        DyElementwiseOp,
+                        XYGridDesc_M_K,
+                        MeanVarGridDesc_M,
+                        MeanVarCountGridDesc_M_K,
+                        DscaleDbiasGridDesc_M_G>;
+
+                const auto kern_reduce_second_half_batchnorm_backward_final =
+                    kernel_reduce_second_half_batchnorm_backward_final<
+                        GridwiseReduceSecondHalfBatchNormBwdFinal_,
+                        XDataType,
+                        DyDataType,
+                        DxDataType,
+                        ScaleDataType,
+                        DscaleDbiasDataType,
+                        MeanVarDataType,
+                        DyElementwiseOp,
+                        XYGridDesc_M_K,
+                        DscaleDbiasGridDesc_M_K,
+                        MeanVarGridDesc_M,
+                        ScaleBiasGridDesc_M>;
+
+                index_t numDscaleDbiasBlockTileIteration =
+                    (arg.blkGroupSize + KThreadClusterSize - 1) / KThreadClusterSize;
+
+                avg_time += launch_and_time_kernel(
+                    stream_config,
+                    kern_welford_second_half_reduce_first_half,
+                    dim3(arg.gridSize),
+                    dim3(BlockSize),
+                    0,
+                    arg.x_grid_desc_m_k,
+                    arg.dy_grid_desc_m_k,
+                    arg.mean_var_grid_desc_m,
+                    mean_var_count_grid_desc_m_k,
+                    dscale_dbias_grid_desc_m_g,
+                    arg.blkGroupSize,
+                    arg.numBlockTileIteration,
+                    numDscaleDbiasBlockTileIteration,
+                    arg.epsilon_,
+                    arg.haveSavedMeanInvVar_,
+                    arg.haveSavedMeanInvVar_ ? arg.p_savedMean_ : nullptr,
+                    arg.haveSavedMeanInvVar_ ? arg.p_savedInvVar_ : nullptr,
+                    arg.haveSavedMeanInvVar_
+                        ? nullptr
+                        : static_cast<const MeanVarDataType*>(arg.workspace_mean),
+                    arg.haveSavedMeanInvVar_
+                        ? nullptr
+                        : static_cast<const MeanVarDataType*>(arg.workspace_variance),
+                    arg.haveSavedMeanInvVar_ ? nullptr
+                                             : static_cast<const int32_t*>(arg.workspace_count),
+                    arg.dy_elementwise_op_,
+                    arg.haveSavedMeanInvVar_
+                        ? nullptr
+                        : static_cast<MeanVarDataType*>(arg.workspace_savedMean),
+                    arg.haveSavedMeanInvVar_
+                        ? nullptr
+                        : static_cast<MeanVarDataType*>(arg.workspace_savedInvVar),
+                    arg.p_x_,
+                    arg.p_dy_,
+                    static_cast<DscaleDbiasDataType*>(arg.workspace_reduce_dscale),
+                    static_cast<DscaleDbiasDataType*>(arg.workspace_reduce_dbias));
+
+                avg_time += launch_and_time_kernel(
+                    stream_config,
+                    kern_reduce_second_half_batchnorm_backward_final,
+                    dim3(arg.gridSize),
+                    dim3(BlockSize),
+                    0,
+                    arg.x_grid_desc_m_k,
+                    arg.dy_grid_desc_m_k,
+                    arg.dx_grid_desc_m_k,
+                    dscale_dbias_grid_desc_m_k,
+                    arg.mean_var_grid_desc_m,
+                    arg.scale_grid_desc_m,
+                    arg.dscale_dbias_grid_desc_m,
+                    arg.blkGroupSize,
+                    arg.reduce_length,
+                    arg.numBlockTileIteration,
+                    numDscaleDbiasBlockTileIteration,
+                    static_cast<const DscaleDbiasDataType*>(arg.workspace_reduce_dscale),
+                    static_cast<const DscaleDbiasDataType*>(arg.workspace_reduce_dbias),
+                    arg.haveSavedMeanInvVar_
+                        ? arg.p_savedMean_
+                        : static_cast<const MeanVarDataType*>(arg.workspace_savedMean),
+                    arg.haveSavedMeanInvVar_
+                        ? arg.p_savedInvVar_
+                        : static_cast<const MeanVarDataType*>(arg.workspace_savedInvVar),
+                    arg.p_x_,
+                    arg.p_dy_,
+                    arg.p_scale_,
+                    arg.dy_elementwise_op_,
+                    arg.p_dx_,
+                    arg.p_dscale_,
+                    arg.p_dbias_);
+            }
+            else
+            {
+                using GetReduceCountPerThreadFunctor =
+                    GetReduceCountPerThreadForBlockwiseWelford<K_BlockTileSize, KThreadSliceSize>;
+
+                GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
+                    arg.numBlockTileIteration, arg.reduce_length);
+
+                using GridwiseBatchNormBackwardWithBlockwiseWelford_ =
+                    GridwiseBatchNormBackwardWithBlockwiseWelford<XDataType,
+                                                                  DyDataType,
+                                                                  DxDataType,
+                                                                  AccDataType,
+                                                                  ScaleDataType,
+                                                                  DscaleDbiasDataType,
+                                                                  MeanVarDataType,
+                                                                  DyElementwiseOp,
+                                                                  XYGridDesc_M_K,
+                                                                  ScaleBiasGridDesc_M,
+                                                                  MeanVarGridDesc_M,
+                                                                  GetReduceCountPerThreadFunctor,
+                                                                  BlockSize,
+                                                                  MThreadClusterSize,
+                                                                  KThreadClusterSize,
+                                                                  MThreadSliceSize,
+                                                                  KThreadSliceSize,
+                                                                  XDyDxVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  DySrcVectorSize,
+                                                                  DxDstVectorSize,
+                                                                  ScaleSrcVectorSize,
+                                                                  DscaleDbiasDstVectorSize,
+                                                                  MeanVarSrcVectorSize>;
+
+                const auto kern_batchnorm_bwd = kernel_batchnorm_backward_with_blockwise_welford<
+                    GridwiseBatchNormBackwardWithBlockwiseWelford_,
+                    XDataType,
+                    DyDataType,
+                    DxDataType,
+                    AccDataType,
+                    ScaleDataType,
+                    DscaleDbiasDataType,
+                    MeanVarDataType,
+                    DyElementwiseOp,
+                    XYGridDesc_M_K,
+                    ScaleBiasGridDesc_M,
+                    MeanVarGridDesc_M,
+                    GetReduceCountPerThreadFunctor>;
+
+                avg_time += launch_and_time_kernel(stream_config,
+                                                   kern_batchnorm_bwd,
+                                                   dim3(arg.gridSize),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.x_grid_desc_m_k,
+                                                   arg.dy_grid_desc_m_k,
+                                                   arg.dx_grid_desc_m_k,
+                                                   arg.scale_grid_desc_m,
+                                                   arg.dscale_dbias_grid_desc_m,
+                                                   arg.mean_var_grid_desc_m,
+                                                   get_reduce_count_per_thread,
+                                                   arg.reduce_length,
+                                                   arg.numBlockTileIteration,
+                                                   arg.epsilon_,
+                                                   arg.p_x_,
+                                                   arg.p_dy_,
+                                                   arg.p_scale_,
+                                                   arg.haveSavedMeanInvVar_,
+                                                   arg.p_savedMean_,
+                                                   arg.p_savedInvVar_,
+                                                   arg.dy_elementwise_op_,
+                                                   arg.p_dx_,
+                                                   arg.p_dscale_,
+                                                   arg.p_dbias_);
+            };
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* pArg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(pArg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* pArg) override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        if constexpr(XDyDxVectorDim == 0)
+        {
+            if(pArg_->xStrides_[NumInvariantDim - 1] != 1 ||
+               pArg_->dyStrides_[NumInvariantDim - 1] != 1 ||
+               pArg_->dxStrides_[NumInvariantDim - 1] != 1)
+                return false;
+
+            if(pArg_->xyLengths_[NumInvariantDim - 1] % XSrcVectorSize != 0 ||
+               pArg_->xyLengths_[NumInvariantDim - 1] % DySrcVectorSize != 0 ||
+               pArg_->xyLengths_[NumInvariantDim - 1] % DxDstVectorSize != 0)
+                return false;
+        }
+        else
+        {
+            if(pArg_->xStrides_[Rank - 1] != 1 || pArg_->dyStrides_[Rank - 1] != 1 ||
+               pArg_->dxStrides_[Rank - 1] != 1)
+                return false;
+
+            if(pArg_->xyLengths_[Rank - 1] % XSrcVectorSize != 0 ||
+               pArg_->xyLengths_[Rank - 1] % DySrcVectorSize != 0 ||
+               pArg_->xyLengths_[Rank - 1] % DxDstVectorSize != 0)
+                return false;
+        };
+
+        if(pArg_->bnScaleStrides_[NumInvariantDim - 1] != 1 && ScaleSrcVectorSize != 1)
+            return false;
+
+        if(pArg_->bnDscaleDbiasStrides_[NumInvariantDim - 1] != 1 && DscaleDbiasDstVectorSize != 1)
+            return false;
+
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % ScaleSrcVectorSize != 0)
+            return false;
+
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % DscaleDbiasDstVectorSize != 0)
+            return false;
+
+        if(pArg_->haveSavedMeanInvVar_)
+        {
+            if(pArg_->bnMeanVarStrides_[NumInvariantDim - 1] != 1 && MeanVarSrcVectorSize != 1)
+                return false;
+
+            if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % MeanVarSrcVectorSize != 0)
+                return false;
+        };
+
+        bool is_valid = true;
+
+        static_for<0, NumInvariantDim, 1>{}([&](auto I) {
+            if(pArg_->xyLengths_[I] != pArg_->bnScaleBiasMeanVarLengths_[I])
+                is_valid = false;
+        });
+
+        if(!is_valid)
+            return false;
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> xyLengths,
+                        const std::array<index_t, Rank> xStrides,
+                        const std::array<index_t, Rank> dyStrides,
+                        const std::array<index_t, Rank> dxStrides,
+                        const std::array<int, NumBatchNormReduceDim> reduceDims,
+                        const std::array<ck::index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                        const std::array<ck::index_t, NumInvariantDim> bnScaleStrides,
+                        const std::array<ck::index_t, NumInvariantDim> bnDscaleDbiasStrides,
+                        const std::array<ck::index_t, NumInvariantDim> bnMeanVarStrides,
+                        const void* p_x,
+                        const void* p_dy,
+                        const void* p_scale,
+                        const void* p_savedMean,
+                        const void* p_savedInvVar,
+                        double epsilon,
+                        const DyElementwiseOp dy_elementwise_op,
+                        void* p_dx,
+                        void* p_dscale,
+                        void* p_dbias) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          dyStrides,
+                                          dxStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnDscaleDbiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const DyDataType*>(p_dy),
+                                          static_cast<const ScaleDataType*>(p_scale),
+                                          static_cast<const MeanVarDataType*>(p_savedMean),
+                                          static_cast<const MeanVarDataType*>(p_savedInvVar),
+                                          dy_elementwise_op,
+                                          epsilon,
+                                          static_cast<DxDataType*>(p_dx),
+                                          static_cast<DscaleDbiasDataType*>(p_dscale),
+                                          static_cast<DscaleDbiasDataType*>(p_dbias));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchNormBwdImpl<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "XDyDxVectorDim_" << XDyDxVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_scale_" << ScaleSrcVectorSize << "_bias_" << DscaleDbiasDstVectorSize << "_mean_var_" << MeanVarSrcVectorSize << "_Dx_" << DxDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+}; // namespace device
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp
new file mode 100644
index 0000000..5a16ff7
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp
@@ -0,0 +1,718 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/device/welford_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim,
+          bool UseMultiblockInK,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcYDstVectorDim,
+          index_t XSrcVectorSize,
+          index_t YDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t BiasSrcVectorSize,
+          index_t MeanVarSrcDstVectorSize>
+struct DeviceBatchNormFwdImpl : public DeviceBatchNormFwd<XDataType,
+                                                          YDataType,
+                                                          AccDataType,
+                                                          ScaleDataType,
+                                                          BiasDataType,
+                                                          MeanVarDataType,
+                                                          YElementwiseOp,
+                                                          Rank,
+                                                          NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "Invalid thread cluster size assignments!");
+
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeXY2dDescriptor(const std::array<index_t, Rank>& xyLengths,
+                                   const std::array<index_t, Rank>& xyStrides,
+                                   int blkGroupSize,
+                                   int numBlockTileIteration)
+    {
+        const auto tupleXYLengths =
+            generate_tuple([&](auto I) { return xyLengths[I]; }, Number<Rank>{});
+        const auto tupleXYStrides =
+            generate_tuple([&](auto I) { return xyStrides[I]; }, Number<Rank>{});
+
+        const auto raw_grid_desc = make_naive_tensor_descriptor(tupleXYLengths, tupleXYStrides);
+
+        const auto grid_desc_m_k = [&]() {
+            using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+            using ReduceDims    = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+            const auto reduceDimLengths =
+                generate_tuple([&](auto I) { return xyLengths[NumInvariantDim + I]; },
+                               Number<NumBatchNormReduceDim>{});
+            const auto invariantDimLengths =
+                generate_tuple([&](auto I) { return xyLengths[I]; }, Number<NumInvariantDim>{});
+
+            return transform_tensor_descriptor(raw_grid_desc,
+                                               make_tuple(make_merge_transform(invariantDimLengths),
+                                                          make_merge_transform(reduceDimLengths)),
+                                               make_tuple(InvariantDims{}, ReduceDims{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }();
+
+        const auto invariantLength = grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = grid_desc_m_k.GetLength(Number<1>{});
+
+        const int workSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto kPad = workSizePerBlock * blkGroupSize - reduceLength;
+
+        auto grid_desc_m_k_padded =
+            transform_tensor_descriptor(grid_desc_m_k,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_right_pad_transform(reduceLength, kPad)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_k_padded);
+    };
+
+    static auto MakeMeanVarCountOutputMG2dDescriptor(int invariantLength, int blkGroupSize)
+    {
+        const auto grid_desc_m_g =
+            make_naive_tensor_descriptor_packed(make_tuple(invariantLength, blkGroupSize));
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_g_padded =
+            transform_tensor_descriptor(grid_desc_m_g,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_pass_through_transform(blkGroupSize)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_g_padded);
+    };
+
+    static auto MakeMeanVarCountInputMK2dDescriptor(int invariantLength, int blkGroupSize)
+    {
+        const auto reduceLength = blkGroupSize;
+        const auto grid_desc_m_k =
+            make_naive_tensor_descriptor_packed(make_tuple(invariantLength, reduceLength));
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto kPad =
+            math::integer_least_multiple(reduceLength, KThreadClusterSize) - reduceLength;
+
+        auto grid_desc_m_k_padded =
+            transform_tensor_descriptor(grid_desc_m_k,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_right_pad_transform(reduceLength, kPad)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_k_padded);
+    };
+
+    static auto
+    MakeScaleBiasMeanVar1dDescriptor(const std::array<index_t, NumInvariantDim>& lengths,
+                                     const std::array<index_t, NumInvariantDim>& strides)
+    {
+        const auto tupleLengths =
+            generate_tuple([&](auto I) { return lengths[I]; }, Number<NumInvariantDim>{});
+        const auto tupleStrides =
+            generate_tuple([&](auto I) { return strides[I]; }, Number<NumInvariantDim>{});
+
+        auto raw_grid_desc = make_naive_tensor_descriptor(tupleLengths, tupleStrides);
+
+        auto grid_desc_m = transform_tensor_descriptor(
+            raw_grid_desc,
+            make_tuple(make_merge_transform(tupleLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = grid_desc_m.GetLength(Number<0>{});
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_padded =
+            transform_tensor_descriptor(grid_desc_m,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return (grid_desc_m_padded);
+    };
+
+    using XYGridDesc_M_K             = decltype(MakeXY2dDescriptor({1}, {1}, 1, 1));
+    using ScaleBiasMeanVarGridDesc_M = decltype(MakeScaleBiasMeanVar1dDescriptor({1}, {1}));
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> yStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const ScaleDataType* p_scale,
+                 const BiasDataType* p_bias,
+                 const YElementwiseOp y_elementwise_op,
+                 double epsilon,
+                 YDataType* p_y,
+                 MeanVarDataType* resultSaveMean,
+                 MeanVarDataType* resultSaveInvVariance,
+                 double averageFactor,
+                 MeanVarDataType* resultRunningMean,
+                 MeanVarDataType* resultRunningVariance)
+            : bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnBiasStrides_(bnBiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              p_scale_(p_scale),
+              p_bias_(p_bias),
+              y_elementwise_op_(y_elementwise_op),
+              p_y_(p_y),
+              resultSaveMean_(resultSaveMean),
+              resultSaveInvVariance_(resultSaveInvVariance),
+              resultRunningMean_(resultRunningMean),
+              resultRunningVariance_(resultRunningVariance)
+        {
+            xyLengths_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xyLengths, reduceDims);
+            xStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xStrides, reduceDims);
+            yStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(yStrides, reduceDims);
+
+            std::tie(invariant_length_, reduce_length_) =
+                get_2d_lengths<Rank, NumBatchNormReduceDim>(xyLengths_);
+
+            epsilon_       = type_convert<AccDataType>(epsilon);
+            averageFactor_ = type_convert<AccDataType>(averageFactor);
+
+            updateMovingAverage_ =
+                (resultRunningMean != nullptr && resultRunningVariance != nullptr);
+            saveMeanInvVariance_ = (resultSaveMean != nullptr && resultSaveInvVariance_ != nullptr);
+
+            if(UseMultiblockInK)
+            {
+                int iterations = 1;
+                while(true)
+                {
+                    int testBlkGroupSize = (reduce_length_ + (K_BlockTileSize * iterations) - 1) /
+                                           (K_BlockTileSize * iterations);
+
+                    // we want the blkGroupSize be not more than 128
+                    if(testBlkGroupSize <= 128)
+                        break;
+
+                    iterations++;
+                };
+
+                blkGroupSize_ = (reduce_length_ + (K_BlockTileSize * iterations) - 1) /
+                                (K_BlockTileSize * iterations);
+
+                numBlockTileIteration_ = iterations;
+            }
+            else
+            {
+                blkGroupSize_          = 1;
+                numBlockTileIteration_ = (reduce_length_ + K_BlockTileSize - 1) / K_BlockTileSize;
+            };
+
+            gridSize_ = (invariant_length_ + M_BlockTileSize - 1) / M_BlockTileSize * blkGroupSize_;
+
+            x_grid_desc_m_k_ =
+                MakeXY2dDescriptor(xyLengths_, xStrides_, blkGroupSize_, numBlockTileIteration_);
+            y_grid_desc_m_k_ =
+                MakeXY2dDescriptor(xyLengths_, yStrides_, blkGroupSize_, numBlockTileIteration_);
+            scale_grid_desc_m_ =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnScaleStrides_);
+            bias_grid_desc_m_ =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnBiasStrides_);
+            mean_var_grid_desc_m_ =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnMeanVarStrides_);
+        }
+
+        AccDataType epsilon_;
+        AccDataType averageFactor_;
+
+        bool updateMovingAverage_;
+        bool saveMeanInvVariance_;
+
+        std::array<index_t, Rank> xyLengths_;
+        std::array<index_t, Rank> xStrides_;
+        std::array<index_t, Rank> yStrides_;
+
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides_;
+
+        const XDataType* p_x_;
+        const ScaleDataType* p_scale_;
+        const BiasDataType* p_bias_;
+        const YElementwiseOp y_elementwise_op_;
+        YDataType* p_y_;
+
+        MeanVarDataType* resultSaveMean_;
+        MeanVarDataType* resultSaveInvVariance_;
+
+        MeanVarDataType* resultRunningMean_;
+        MeanVarDataType* resultRunningVariance_;
+
+        long_index_t invariant_length_;
+        long_index_t reduce_length_;
+
+        int blkGroupSize_;
+        int numBlockTileIteration_;
+        size_t gridSize_;
+
+        XYGridDesc_M_K x_grid_desc_m_k_;
+        XYGridDesc_M_K y_grid_desc_m_k_;
+        ScaleBiasMeanVarGridDesc_M scale_grid_desc_m_;
+        ScaleBiasMeanVarGridDesc_M bias_grid_desc_m_;
+        ScaleBiasMeanVarGridDesc_M mean_var_grid_desc_m_;
+
+        void* workspace_mean_;
+        void* workspace_variance_;
+        void* workspace_count_;
+    };
+
+    size_t GetWorkSpaceSize(const BaseArgument* pArg) const override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        size_t workspace_size = 0;
+
+        if(UseMultiblockInK && pArg_->blkGroupSize_ > 1)
+        {
+            // workspace for welford intermediate mean
+            workspace_size +=
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType) + 64;
+
+            // workspace for welford intermediate variance
+            workspace_size +=
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType) + 64;
+
+            // workspace for welford intermediate count
+            workspace_size +=
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(int32_t) + 64;
+        }
+
+        return (workspace_size);
+    };
+
+    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    {
+        Argument* pArg_ = dynamic_cast<Argument*>(pArg);
+
+        pArg_->p_workspace_ = p_workspace;
+
+        if(UseMultiblockInK && pArg_->blkGroupSize_ > 1)
+        {
+
+            // setup buffer used for intermediate welford mean
+            pArg_->workspace_mean_ = static_cast<char*>(pArg_->p_workspace_);
+
+            index_t mean_space_sz =
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType);
+
+            mean_space_sz = math::integer_least_multiple(mean_space_sz, 64);
+
+            // setup buffer used for intermediate welford varirance
+            pArg_->workspace_variance_ =
+                reinterpret_cast<char*>(pArg_->workspace_mean_) + mean_space_sz;
+
+            index_t variance_space_sz =
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType);
+
+            variance_space_sz = math::integer_least_multiple(variance_space_sz, 64);
+
+            // setup buffer used for intermediate welfor count
+            pArg_->workspace_count_ =
+                reinterpret_cast<char*>(pArg_->workspace_variance_) + variance_space_sz;
+        };
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float avg_time = 0;
+
+            if(UseMultiblockInK && arg.blkGroupSize_ > 1)
+            {
+                using GetReduceCountPerThreadFunctor =
+                    GetReduceCountPerThreadForMultiblockWelford<K_BlockTileSize, KThreadSliceSize>;
+
+                GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
+                    arg.blkGroupSize_, arg.numBlockTileIteration_, arg.reduce_length_);
+
+                const auto mean_var_count_grid_desc_m_g =
+                    DeviceBatchNormFwdImpl::MakeMeanVarCountOutputMG2dDescriptor(
+                        arg.invariant_length_, arg.blkGroupSize_);
+
+                const auto mean_var_count_grid_desc_m_k =
+                    DeviceBatchNormFwdImpl::MakeMeanVarCountInputMK2dDescriptor(
+                        arg.invariant_length_, arg.blkGroupSize_);
+
+                using MeanVarCountGridDesc_M_G = decltype(mean_var_count_grid_desc_m_g);
+                using MeanVarCountGridDesc_M_K = decltype(mean_var_count_grid_desc_m_k);
+
+                using GridwiseMultiblockWelfordFirstHalf_ =
+                    GridwiseMultiblockWelfordFirstHalf<XDataType,
+                                                       AccDataType,
+                                                       MeanVarDataType,
+                                                       XYGridDesc_M_K,
+                                                       MeanVarCountGridDesc_M_G,
+                                                       GetReduceCountPerThreadFunctor,
+                                                       BlockSize,
+                                                       MThreadClusterSize,
+                                                       KThreadClusterSize,
+                                                       MThreadSliceSize,
+                                                       KThreadSliceSize,
+                                                       XSrcYDstVectorDim,
+                                                       XSrcVectorSize>;
+
+                using GridwiseWelfordSecondHalfBatchNormForwardFinal_ =
+                    GridwiseWelfordSecondHalfBatchNormForwardFinal<XDataType,
+                                                                   YDataType,
+                                                                   AccDataType,
+                                                                   ScaleDataType,
+                                                                   BiasDataType,
+                                                                   MeanVarDataType,
+                                                                   YElementwiseOp,
+                                                                   XYGridDesc_M_K,
+                                                                   MeanVarCountGridDesc_M_K,
+                                                                   ScaleBiasMeanVarGridDesc_M,
+                                                                   ScaleBiasMeanVarGridDesc_M,
+                                                                   BlockSize,
+                                                                   MThreadClusterSize,
+                                                                   KThreadClusterSize,
+                                                                   MThreadSliceSize,
+                                                                   KThreadSliceSize,
+                                                                   XSrcYDstVectorDim,
+                                                                   XSrcVectorSize,
+                                                                   YDstVectorSize,
+                                                                   ScaleSrcVectorSize,
+                                                                   BiasSrcVectorSize,
+                                                                   MeanVarSrcDstVectorSize>;
+
+                index_t numMeanVarCountBlockTileIteration =
+                    (arg.blkGroupSize_ + KThreadClusterSize - 1) / KThreadClusterSize;
+
+                const auto kern_multiblock_welford_first_half =
+                    kernel_multiblock_welford_first_half<GridwiseMultiblockWelfordFirstHalf_,
+                                                         XDataType,
+                                                         MeanVarDataType,
+                                                         XYGridDesc_M_K,
+                                                         MeanVarCountGridDesc_M_G,
+                                                         GetReduceCountPerThreadFunctor>;
+
+                const auto kern_welford_second_half_batchnorm_forward_final =
+                    kernel_welford_second_half_batchnorm_forward_final<
+                        GridwiseWelfordSecondHalfBatchNormForwardFinal_,
+                        XDataType,
+                        YDataType,
+                        AccDataType,
+                        ScaleDataType,
+                        BiasDataType,
+                        MeanVarDataType,
+                        YElementwiseOp,
+                        XYGridDesc_M_K,
+                        MeanVarCountGridDesc_M_K,
+                        ScaleBiasMeanVarGridDesc_M,
+                        ScaleBiasMeanVarGridDesc_M>;
+
+                avg_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kern_multiblock_welford_first_half,
+                                           dim3(arg.gridSize_),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.x_grid_desc_m_k_,
+                                           mean_var_count_grid_desc_m_g,
+                                           get_reduce_count_per_thread,
+                                           arg.numBlockTileIteration_,
+                                           arg.p_x_,
+                                           static_cast<MeanVarDataType*>(arg.workspace_mean_),
+                                           static_cast<MeanVarDataType*>(arg.workspace_variance_),
+                                           static_cast<int32_t*>(arg.workspace_count_));
+
+                avg_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kern_welford_second_half_batchnorm_forward_final,
+                                           dim3(arg.gridSize_),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.x_grid_desc_m_k_,
+                                           arg.y_grid_desc_m_k_,
+                                           mean_var_count_grid_desc_m_k,
+                                           arg.scale_grid_desc_m_,
+                                           arg.bias_grid_desc_m_,
+                                           arg.mean_var_grid_desc_m_,
+                                           arg.blkGroupSize_,
+                                           arg.numBlockTileIteration_,
+                                           numMeanVarCountBlockTileIteration,
+                                           arg.epsilon_,
+                                           static_cast<MeanVarDataType*>(arg.workspace_mean_),
+                                           static_cast<MeanVarDataType*>(arg.workspace_variance_),
+                                           static_cast<int32_t*>(arg.workspace_count_),
+                                           arg.p_x_,
+                                           arg.p_scale_,
+                                           arg.p_bias_,
+                                           arg.y_elementwise_op_,
+                                           arg.p_y_,
+                                           arg.updateMovingAverage_,
+                                           arg.averageFactor_,
+                                           arg.resultRunningMean_,
+                                           arg.resultRunningVariance_,
+                                           arg.saveMeanInvVariance_,
+                                           arg.resultSaveMean_,
+                                           arg.resultSaveInvVariance_);
+            }
+            else
+            {
+                using GetReduceCountPerThreadFunctor =
+                    GetReduceCountPerThreadForBlockwiseWelford<K_BlockTileSize, KThreadSliceSize>;
+
+                GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
+                    arg.numBlockTileIteration_, arg.reduce_length_);
+
+                using GridwiseBatchNormForwardWithBlockwiseWelford_ =
+                    GridwiseBatchNormForwardWithBlockwiseWelford<XDataType,
+                                                                 YDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 BiasDataType,
+                                                                 MeanVarDataType,
+                                                                 YElementwiseOp,
+                                                                 XYGridDesc_M_K,
+                                                                 ScaleBiasMeanVarGridDesc_M,
+                                                                 ScaleBiasMeanVarGridDesc_M,
+                                                                 GetReduceCountPerThreadFunctor,
+                                                                 BlockSize,
+                                                                 MThreadClusterSize,
+                                                                 KThreadClusterSize,
+                                                                 MThreadSliceSize,
+                                                                 KThreadSliceSize,
+                                                                 XSrcYDstVectorDim,
+                                                                 XSrcVectorSize,
+                                                                 YDstVectorSize,
+                                                                 ScaleSrcVectorSize,
+                                                                 BiasSrcVectorSize,
+                                                                 MeanVarSrcDstVectorSize>;
+
+                const auto kern_batchnorm_fwd = kernel_batchnorm_forward_with_blockwise_welford<
+                    GridwiseBatchNormForwardWithBlockwiseWelford_,
+                    XDataType,
+                    YDataType,
+                    AccDataType,
+                    ScaleDataType,
+                    BiasDataType,
+                    MeanVarDataType,
+                    YElementwiseOp,
+                    XYGridDesc_M_K,
+                    ScaleBiasMeanVarGridDesc_M,
+                    ScaleBiasMeanVarGridDesc_M,
+                    GetReduceCountPerThreadFunctor>;
+
+                avg_time += launch_and_time_kernel(stream_config,
+                                                   kern_batchnorm_fwd,
+                                                   dim3(arg.gridSize_),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.x_grid_desc_m_k_,
+                                                   arg.y_grid_desc_m_k_,
+                                                   arg.scale_grid_desc_m_,
+                                                   arg.bias_grid_desc_m_,
+                                                   arg.mean_var_grid_desc_m_,
+                                                   get_reduce_count_per_thread,
+                                                   arg.numBlockTileIteration_,
+                                                   arg.epsilon_,
+                                                   arg.p_x_,
+                                                   arg.p_scale_,
+                                                   arg.p_bias_,
+                                                   arg.y_elementwise_op_,
+                                                   arg.p_y_,
+                                                   arg.updateMovingAverage_, // true or false
+                                                   arg.averageFactor_,
+                                                   arg.resultRunningMean_,
+                                                   arg.resultRunningVariance_,
+                                                   arg.saveMeanInvVariance_, // true or false
+                                                   arg.resultSaveMean_,
+                                                   arg.resultSaveInvVariance_);
+            };
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* pArg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(pArg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* pArg) override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        if constexpr(XSrcYDstVectorDim == 0)
+        {
+            if(pArg_->xStrides_[NumInvariantDim - 1] != 1 ||
+               pArg_->yStrides_[NumInvariantDim - 1] != 1)
+                return false;
+
+            if(pArg_->xyLengths_[NumInvariantDim - 1] % XSrcVectorSize != 0 ||
+               pArg_->xyLengths_[NumInvariantDim - 1] % YDstVectorSize != 0)
+                return false;
+        }
+        else
+        {
+            if(pArg_->xStrides_[Rank - 1] != 1 || pArg_->yStrides_[Rank - 1] != 1)
+                return false;
+
+            if(pArg_->xyLengths_[Rank - 1] % XSrcVectorSize != 0 ||
+               pArg_->xyLengths_[Rank - 1] % YDstVectorSize != 0)
+                return false;
+        };
+
+        if(pArg_->bnScaleStrides_[NumInvariantDim - 1] != 1 && ScaleSrcVectorSize != 1)
+            return false;
+        if(pArg_->bnBiasStrides_[NumInvariantDim - 1] != 1 && BiasSrcVectorSize != 1)
+            return false;
+
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % ScaleSrcVectorSize != 0)
+            return false;
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % BiasSrcVectorSize != 0)
+            return false;
+
+        if(pArg_->bnMeanVarStrides_[NumInvariantDim - 1] != 1 && MeanVarSrcDstVectorSize != 1)
+            return false;
+
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % MeanVarSrcDstVectorSize != 0)
+            return false;
+
+        bool is_valid = true;
+
+        static_for<0, NumInvariantDim, 1>{}([&](auto I) {
+            if(pArg_->xyLengths_[I] != pArg_->bnScaleBiasMeanVarLengths_[I])
+                is_valid = false;
+        });
+
+        if(!is_valid)
+            return false;
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, Rank> xyLengths,
+        const std::array<index_t, Rank> xStrides,
+        const std::array<index_t, Rank> yStrides,
+        const std::array<int, NumBatchNormReduceDim> reduceDims,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+        const void* p_x,
+        const void* p_scale,
+        const void* p_bias,
+        double epsilon,
+        const YElementwiseOp y_elementwise_op,
+        void* p_y,
+        void* resultSaveMean,
+        void* resultSaveInvVariance,
+        double averageFactor,
+        void* resultRunningMean,
+        void* resultRunningVariance) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnBiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const ScaleDataType*>(p_scale),
+                                          static_cast<const BiasDataType*>(p_bias),
+                                          y_elementwise_op,
+                                          epsilon,
+                                          static_cast<YDataType*>(p_y),
+                                          static_cast<MeanVarDataType*>(resultSaveMean),
+                                          static_cast<MeanVarDataType*>(resultSaveInvVariance),
+                                          averageFactor,
+                                          static_cast<MeanVarDataType*>(resultRunningMean),
+                                          static_cast<MeanVarDataType*>(resultRunningVariance));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchNormFwdImpl<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "XSrcYDstVectorDim_" << XSrcYDstVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_scale_" << ScaleSrcVectorSize << "_bias_" << BiasSrcVectorSize << "_mean_var_" << MeanVarSrcDstVectorSize << "_Y" << YDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp
new file mode 100644
index 0000000..2997845
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp
@@ -0,0 +1,948 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_cgemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <
+    typename ALayout,
+    typename BLayout,
+    typename CLayout,
+    typename ADataType,
+    typename BDataType,
+    typename CDataType,
+    typename GemmAccDataType,
+    typename CShuffleDataType,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    GemmSpecialization GemmSpec,
+    index_t NumGemmKPrefetchStage,
+    index_t BlockSize,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t KPerBlock,
+    index_t AK1,
+    index_t BK1,
+    index_t MPerXDL,
+    index_t NPerXDL,
+    index_t MXdlPerWave,
+    index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    index_t ABlockTransferSrcVectorDim,
+    index_t ABlockTransferSrcScalarPerVector,
+    index_t ABlockTransferDstScalarPerVector_AK1,
+    bool ABlockLdsExtraM,
+    typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    index_t BBlockTransferSrcVectorDim,
+    index_t BBlockTransferSrcScalarPerVector,
+    index_t BBlockTransferDstScalarPerVector_BK1,
+    bool BBlockLdsExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+    index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+    LoopScheduler LoopSched = make_default_loop_scheduler(),
+    enable_if_t<
+        is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+        bool> = false>
+struct DeviceCGemm_4Gemm_Xdl_CShuffle
+    : public DeviceCGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
+{
+    using DeviceOp = DeviceCGemm_4Gemm_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto MPerThread       = Number<4>{};
+    static constexpr auto AScalarPerVector = Number<4>{};
+    static constexpr auto BScalarPerVector = Number<4>{};
+    static constexpr auto CScalarPerVector = Number<4>{};
+
+    template <typename Desc_M>
+    static auto PadDescriptor_M_1d(Desc_M desc_m, index_t gridSize, index_t blockSize)
+    {
+        const auto M            = desc_m.GetLength(I0);
+        const index_t loop_step = gridSize * blockSize * MPerThread;
+        const auto pad          = math::integer_least_multiple(M, loop_step) - M;
+        const auto desc_m_pad =
+            transform_tensor_descriptor(desc_m,
+                                        make_tuple(make_right_pad_transform(M, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return desc_m_pad;
+    }
+
+    static auto MakeDescriptor_M(const std::vector<index_t>& lengths,
+                                 const std::vector<index_t>& strides,
+                                 index_t gridSize,
+                                 index_t blockSize)
+    {
+        auto tupleOfShape  = generate_tuple([&](auto I) { return lengths[I]; }, Number<2>{});
+        auto tupleOfStride = generate_tuple([&](auto I) { return strides[I]; }, Number<2>{});
+
+        // nd desc - [s0, s1, s2, ...]
+        const auto desc   = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
+        const auto desc_m = transform_tensor_descriptor(
+            desc,
+            make_tuple(make_merge_transform(tupleOfShape)),
+            make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<2>{})),
+            make_tuple(Sequence<0>{}));
+
+        return PadDescriptor_M_1d(desc_m, gridSize, blockSize);
+    }
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(M)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_right_pad_transform(MRaw, MPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(N)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_right_pad_transform(NRaw, NPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N       = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+    using CGridDesc_M         = decltype(MakeDescriptor_M({1, 1}, {1, 1}, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid_real,
+                 const ADataType* p_a_grid_imag,
+                 const BDataType* p_b_grid_real,
+                 const BDataType* p_b_grid_imag,
+                 CDataType* p_c_grid_real,
+                 CDataType* p_c_grid_imag,
+                 CDataType* p_workspace,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_real_{p_a_grid_real},
+              p_a_grid_imag_{p_a_grid_imag},
+              p_b_grid_real_{p_b_grid_real},
+              p_b_grid_imag_{p_b_grid_imag},
+              p_c_grid_real_{p_c_grid_real},
+              p_c_grid_imag_{p_c_grid_imag},
+              p_aux_grid_{p_workspace},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+            }
+
+            const index_t grid_size = block_2_ctile_map_.CalculateGridSize(c_grid_desc_m_n_);
+
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                c_grid_desc_m_ =
+                    DeviceOp::MakeDescriptor_M({MRaw, NRaw}, {StrideC, I1}, grid_size, BlockSize);
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                c_grid_desc_m_ =
+                    DeviceOp::MakeDescriptor_M({MRaw, NRaw}, {I1, StrideC}, grid_size, BlockSize);
+            }
+
+            p_aux_2_grid_ = p_workspace + c_grid_desc_m_n_.GetElementSpaceSize();
+        }
+
+        //  private:
+        const ADataType* p_a_grid_real_;
+        const ADataType* p_a_grid_imag_;
+        const BDataType* p_b_grid_real_;
+        const BDataType* p_b_grid_imag_;
+        CDataType* p_c_grid_real_;
+        CDataType* p_c_grid_imag_;
+        CDataType* p_aux_grid_;
+        CDataType* p_aux_2_grid_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_M c_grid_desc_m_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                            arg.b_grid_desc_bk0_n_bk1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            using Add      = ck::tensor_operation::element_wise::Add;
+            using Subtract = ck::tensor_operation::element_wise::Subtract;
+
+            using GridwiseBinAdd =
+                GridwiseElementwise_1D<Tuple<CGridDesc_M, CGridDesc_M>,
+                                       Tuple<CGridDesc_M>,
+                                       Tuple<const CDataType*, const CDataType*>,
+                                       Tuple<CDataType*>,
+                                       Add,
+                                       MPerThread,
+                                       Sequence<AScalarPerVector, BScalarPerVector>,
+                                       Sequence<CScalarPerVector>>;
+
+            using GridwiseBinSubtract =
+                GridwiseElementwise_1D<Tuple<CGridDesc_M, CGridDesc_M>,
+                                       Tuple<CGridDesc_M>,
+                                       Tuple<const CDataType*, const CDataType*>,
+                                       Tuple<CDataType*>,
+                                       Subtract,
+                                       MPerThread,
+                                       Sequence<AScalarPerVector, BScalarPerVector>,
+                                       Sequence<CScalarPerVector>>;
+
+            const auto add_kernel = kernel_elementwise_1d<GridwiseBinAdd,
+                                                          Tuple<CGridDesc_M, CGridDesc_M>,
+                                                          Tuple<CGridDesc_M>,
+                                                          Tuple<const CDataType*, const CDataType*>,
+                                                          Tuple<CDataType*>,
+                                                          Add>;
+
+            const auto subtract_kernel =
+                kernel_elementwise_1d<GridwiseBinSubtract,
+                                      Tuple<CGridDesc_M, CGridDesc_M>,
+                                      Tuple<CGridDesc_M>,
+                                      Tuple<const CDataType*, const CDataType*>,
+                                      Tuple<CDataType*>,
+                                      Subtract>;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    true>;
+
+                ave_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_real_,
+                                           arg.p_b_grid_real_,
+                                           arg.p_aux_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+
+                ave_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_imag_,
+                                           arg.p_b_grid_imag_,
+                                           arg.p_aux_2_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+
+                // c_real = aux - aux_2
+                ave_time += launch_and_time_kernel(
+                    stream_config,
+                    subtract_kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    make_tuple(arg.c_grid_desc_m_, arg.c_grid_desc_m_),
+                    make_tuple(arg.c_grid_desc_m_),
+                    make_tuple(const_cast<const CDataType*>(arg.p_aux_grid_),
+                               const_cast<const CDataType*>(arg.p_aux_2_grid_)),
+                    make_tuple(arg.p_c_grid_real_),
+                    Subtract{});
+
+                ave_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_real_,
+                                           arg.p_b_grid_imag_,
+                                           arg.p_aux_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+
+                ave_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_imag_,
+                                           arg.p_b_grid_real_,
+                                           arg.p_aux_2_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+
+                // c_imag = aux + aux_2
+                ave_time += launch_and_time_kernel(
+                    stream_config,
+                    add_kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    make_tuple(arg.c_grid_desc_m_, arg.c_grid_desc_m_),
+                    make_tuple(arg.c_grid_desc_m_),
+                    make_tuple(const_cast<const CDataType*>(arg.p_aux_grid_),
+                               const_cast<const CDataType*>(arg.p_aux_2_grid_)),
+                    make_tuple(arg.p_c_grid_imag_),
+                    Add{});
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    false>;
+
+                ave_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_real_,
+                                           arg.p_b_grid_real_,
+                                           arg.p_aux_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+
+                ave_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_imag_,
+                                           arg.p_b_grid_imag_,
+                                           arg.p_aux_2_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+
+                // c_real = aux - aux_2
+                ave_time += launch_and_time_kernel(
+                    stream_config,
+                    subtract_kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    make_tuple(arg.c_grid_desc_m_, arg.c_grid_desc_m_),
+                    make_tuple(arg.c_grid_desc_m_),
+                    make_tuple(const_cast<const CDataType*>(arg.p_aux_grid_),
+                               const_cast<const CDataType*>(arg.p_aux_2_grid_)),
+                    make_tuple(arg.p_c_grid_real_),
+                    Subtract{});
+
+                ave_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_real_,
+                                           arg.p_b_grid_imag_,
+                                           arg.p_aux_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+
+                ave_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_imag_,
+                                           arg.p_b_grid_real_,
+                                           arg.p_aux_2_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+
+                // c_imag = aux + aux_2
+                ave_time += launch_and_time_kernel(
+                    stream_config,
+                    add_kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    make_tuple(arg.c_grid_desc_m_, arg.c_grid_desc_m_),
+                    make_tuple(arg.c_grid_desc_m_),
+                    make_tuple(const_cast<const CDataType*>(arg.p_aux_grid_),
+                               const_cast<const CDataType*>(arg.p_aux_2_grid_)),
+                    make_tuple(arg.p_c_grid_imag_),
+                    Add{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a_real,
+                             const ADataType* p_a_imag,
+                             const BDataType* p_b_real,
+                             const BDataType* p_b_imag,
+                             CDataType* p_c_real,
+                             CDataType* p_c_imag,
+                             CDataType* p_workspace,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a_real,
+                        p_a_imag,
+                        p_b_real,
+                        p_b_imag,
+                        p_c_real,
+                        p_c_imag,
+                        p_workspace,
+                        MRaw,
+                        NRaw,
+                        KRaw,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a_real,
+                                                      const void* p_a_imag,
+                                                      const void* p_b_real,
+                                                      const void* p_b_imag,
+                                                      void* p_c_real,
+                                                      void* p_c_imag,
+                                                      void* p_workspace,
+                                                      index_t MRaw,
+                                                      index_t NRaw,
+                                                      index_t KRaw,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op,
+                                                      index_t /* KBatch */ = 1) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a_real),
+                                          static_cast<const ADataType*>(p_a_imag),
+                                          static_cast<const BDataType*>(p_b_real),
+                                          static_cast<const BDataType*>(p_b_imag),
+                                          static_cast<CDataType*>(p_c_real),
+                                          static_cast<CDataType*>(p_c_imag),
+                                          static_cast<CDataType*>(p_workspace),
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceCGemm_4Gemm_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    std::size_t GetWorkspaceSize(index_t MRaw,
+                                 index_t NRaw,
+                                 [[maybe_unused]] index_t KRaw,
+                                 [[maybe_unused]] index_t StrideA,
+                                 [[maybe_unused]] index_t StrideB,
+                                 index_t StrideC) override
+    {
+        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC);
+
+        return 2 * sizeof(CDataType) * c_grid_desc_m_n.GetElementSpaceSize();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000..72c6d0b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,779 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatDsPointer,
+          typename FloatE,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_contraction_multiple_d_xdl_cshuffle(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatDsPointer p_ds_grid,
+            FloatE* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_ds_grid,
+                                                  p_e_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceContractionMultipleD_Xdl_CShuffle
+    : public DeviceContractionMultipleD<NumDimM,
+                                        NumDimN,
+                                        NumDimK,
+                                        ADataType,
+                                        BDataType,
+                                        DsDataType,
+                                        EDataType,
+                                        AElementwiseOperation,
+                                        BElementwiseOperation,
+                                        CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceContractionMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    // Assume: A[M0, M1, M2, ..., K0, K1, K2, ...]
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_ms_ks_strides_vec)
+    {
+        assert(a_ms_ks_lengths_vec.size() == NumDimM + NumDimK &&
+               a_ms_ks_strides_vec.size() == NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto a_ms_ns_lengths = to_tuple(a_ms_ks_lengths_vec, Number<NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(a_ms_ks_strides_vec, Number<NumDimM + NumDimK>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(a_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(a_ms_ns_lengths, kDimIds);
+
+        // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
+        const auto a_grid_desc_ms_ks =
+            make_naive_tensor_descriptor(a_ms_ns_lengths, a_ms_ks_strides);
+
+        // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
+        const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
+            a_grid_desc_ms_ks,
+            make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
+            make_tuple(mDimIds, kDimIds),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    // Assume: B[N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_ns_ks_lengths_vec,
+                                        const std::vector<index_t>& b_ns_ks_strides_vec)
+    {
+        assert(b_ns_ks_lengths_vec.size() == NumDimN + NumDimK &&
+               b_ns_ks_strides_vec.size() == NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(b_ns_ks_lengths_vec, Number<NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(b_ns_ks_strides_vec, Number<NumDimN + NumDimK>{});
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
+
+        // lengths for N0, N1, ...
+        const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
+
+        // naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
+        const auto b_grid_desc_ns_ks =
+            make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
+
+        // transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
+        const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
+            b_grid_desc_ns_ks,
+            make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
+            make_tuple(nDimIds, kDimIds),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    // assume E[M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& e_ms_ns_strides_vec)
+    {
+        assert(e_ms_ns_lengths_vec.size() == NumDimM + NumDimN &&
+               e_ms_ns_strides_vec.size() == NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto e_ms_ns_lengths = to_tuple(e_ms_ns_lengths_vec, Number<NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(e_ms_ns_strides_vec, Number<NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
+
+        // naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
+        const auto e_grid_desc_ms_ns =
+            make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
+
+        // transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+        const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
+            e_grid_desc_ms_ns,
+            make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+            make_tuple(mDimIds, nDimIds),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_M_N(ds_ms_ns_lengths_vec[i],
+                                                         ds_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K({}, {}));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K({}, {}));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 const std::vector<index_t>& a_ms_ns_lengths,
+                 const std::vector<index_t>& a_ms_ks_strides,
+                 const std::vector<index_t>& b_ns_ks_lengths,
+                 const std::vector<index_t>& b_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides,
+                 const std::vector<index_t>& e_ms_ns_lengths,
+                 const std::vector<index_t>& e_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(a_ms_ns_lengths, a_ms_ks_strides)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(b_ns_ks_lengths, b_ns_ks_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N(e_ms_ns_lengths, e_ms_ns_strides)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_mz_stride_{},
+              a_kz_stride_{},
+              b_nz_stride_{},
+              b_kz_stride_{},
+              ds_nz_stride_{},
+              e_nz_stride_{}
+        {
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) =
+                    DeviceOp::MakeEGridDescriptor_M_N(ds_ms_ns_lengths[i], ds_ms_ns_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+            }
+
+            // for sanity check of vector memory access
+            a_mz_stride_ = a_ms_ks_strides[NumDimM - 1];
+            a_kz_stride_ = a_ms_ks_strides[NumDimM + NumDimK - 1];
+
+            b_nz_stride_ = b_ns_ks_strides[NumDimN - 1];
+            b_kz_stride_ = b_ns_ks_strides[NumDimN + NumDimK - 1];
+
+            for(index_t i = 0; i < NumDTensor; ++i)
+            {
+                ds_nz_stride_[i] = ds_ms_ns_strides[i][NumDimM + NumDimN - 1];
+            }
+
+            e_nz_stride_ = e_ms_ns_strides[NumDimM + NumDimN - 1];
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // Strides for the last M/N/K dimensions of A/B/Ds/E
+        //   for sanity check of vector load/store
+        index_t a_mz_stride_;
+        index_t a_kz_stride_;
+        index_t b_nz_stride_;
+        index_t b_kz_stride_;
+        std::array<index_t, NumDTensor> ds_nz_stride_;
+        index_t e_mz_stride_;
+        index_t e_nz_stride_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_contraction_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::Block2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                        arg.b_grid_desc_n_k_,
+                                        arg.ds_grid_desc_m_n_,
+                                        arg.e_grid_desc_m_n_,
+                                        arg.block_2_etile_map_))
+        {
+            return false;
+        }
+
+        // check vector access
+        static_assert((ABlockTransferSrcVectorDim == 1 || ABlockTransferSrcVectorDim == 2) &&
+                          (BBlockTransferSrcVectorDim == 1 || BBlockTransferSrcVectorDim == 2),
+                      "wrong!");
+
+        // vector memory access of A: could be on M or AK1 dimension
+        if constexpr(ABlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.a_mz_stride_ == 1 &&
+                 arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.a_kz_stride_ == 1 &&
+                 arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of B: could be on N or BK1 dimension
+        if constexpr(BBlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.b_nz_stride_ == 1 &&
+                 arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.b_kz_stride_ == 1 &&
+                 arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of Ds: always on NPerBlock dimension
+        bool valid_d_access = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            if(!(arg.ds_nz_stride_[i] == 1 &&
+                 arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_[i].GetLength(I3) %
+                         CDEBlockTransferScalarPerVector_NPerBlock ==
+                     0))
+            {
+                valid_d_access = false;
+            }
+        });
+
+        if(valid_d_access == false)
+        {
+            return false;
+        }
+
+        // vector memory access of E: always on NPerBlock dimension
+        if(!(arg.e_nz_stride_ == 1 &&
+             arg.e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
+                     CDEBlockTransferScalarPerVector_NPerBlock ==
+                 0))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_e,
+                             const std::vector<index_t>& a_ms_ns_lengths,
+                             const std::vector<index_t>& a_ms_ks_strides,
+                             const std::vector<index_t>& b_ns_ks_lengths,
+                             const std::vector<index_t>& b_ns_ks_strides,
+                             const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths,
+                             const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides,
+                             const std::vector<index_t>& e_ms_ns_lengths,
+                             const std::vector<index_t>& e_ms_ns_strides,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_ms_ns_lengths,
+                        a_ms_ks_strides,
+                        b_ns_ks_lengths,
+                        b_ns_ks_strides,
+                        ds_ms_ns_lengths,
+                        ds_ms_ns_strides,
+                        e_ms_ns_lengths,
+                        e_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_ms_ns_lengths,
+                        const std::vector<index_t>& a_ms_ks_strides,
+                        const std::vector<index_t>& b_ns_ks_lengths,
+                        const std::vector<index_t>& b_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides,
+                        const std::vector<index_t>& e_ms_ns_lengths,
+                        const std::vector<index_t>& e_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_ms_ns_lengths,
+                                          a_ms_ks_strides,
+                                          b_ns_ks_lengths,
+                                          b_ns_ks_strides,
+                                          ds_ms_ns_lengths,
+                                          ds_ms_ns_strides,
+                                          e_ms_ns_lengths,
+                                          e_ms_ns_strides,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceContractionMultipleD_Xdl_CShuffle"
+            << "<"
+            << NumDimM << ", "
+            << NumDimN << ", "
+            << NumDimK << ", "
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << ABlockTransferSrcVectorDim << ", "
+            << BBlockTransferSrcVectorDim
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..4760422
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,787 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_weight.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXdl,
+          ck::index_t NPerXdl,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
+struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvBwdWeight<2,
+                                 ck::tensor_layout::convolution::NHWC,
+                                 ck::tensor_layout::convolution::KYXC,
+                                 ck::tensor_layout::convolution::NHWK,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementwiseOperation,
+                                 WeiElementwiseOperation,
+                                 OutElementwiseOperation>
+{
+    static constexpr ck::index_t NDimSpatial = 2;
+
+    using DeviceOp =
+        DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = OutDataType;
+    using BDataType = InDataType;
+    using CDataType = WeiDataType;
+
+    using AElementwiseOperation = OutElementwiseOperation;
+    using BElementwiseOperation = InElementwiseOperation;
+    using CElementwiseOperation = WeiElementwiseOperation;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static constexpr auto N1Number = K1Number;
+
+    // Bytes per 32 lds bank: 32 * 4 bytes
+    static constexpr auto BankLength = 128;
+    static constexpr auto ElePerBank = BankLength / sizeof(ADataType);
+
+    // M1 & M0
+    static constexpr auto ABlockLdsM1PerBlock = ElePerBank / K1;
+    static constexpr auto ABlockLdsM0PerBlock = MPerBlock / ABlockLdsM1PerBlock;
+    static constexpr auto ABlockLdsM1Padding  = 4;
+
+    // N1 & N0
+    static constexpr auto BBlockLdsN1PerBlock = ElePerBank / K1;
+    static constexpr auto BBlockLdsN0PerBlock = NPerBlock / BBlockLdsN1PerBlock;
+    static constexpr auto BBlockLdsN1Padding  = 4;
+
+    static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        ck::index_t N,
+        ck::index_t K,
+        ck::index_t C,
+        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+        std::array<ck::index_t, NDimSpatial> input_left_pads,
+        std::array<ck::index_t, NDimSpatial> input_right_pads,
+        ck::index_t batch_k)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmKTotal = N * Ho * Wo;
+        const index_t GemmM      = K;
+        const index_t GemmN      = C * X * Y;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        // A: output tensor
+        const index_t N0          = N / N1Number;
+        const index_t GemmK0Total = N0 * Ho * Wo;
+
+        const index_t GemmK0S =
+            math::integer_divide_ceil(GemmK0Total, K0PerBlock * GemmKBatch) * K0PerBlock;
+        const index_t GemmK0Pad = GemmKBatch * GemmK0S;
+        const auto out_n_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Ho * Wo, K));
+
+        const auto out_n0_ho_wo_k_n1_grid_desc =
+            transform_tensor_descriptor(out_n_ho_wo_k_grid_desc,
+                                        make_tuple(make_unmerge_transform(make_tuple(N0, N1Number)),
+                                                   make_pass_through_transform(Ho * Wo),
+                                                   make_pass_through_transform(K)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0, 3>{}, Sequence<1>{}, Sequence<2>{}));
+
+        const auto out_gemmk0total_gemmm_gemmk1_grid_desc =
+            transform_tensor_descriptor(out_n0_ho_wo_k_n1_grid_desc,
+                                        make_tuple(make_merge_transform(make_tuple(N0, Ho * Wo)),
+                                                   make_pass_through_transform(K),
+                                                   make_pass_through_transform(N1Number)),
+                                        make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        const auto out_gemmk0pad_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+            out_gemmk0total_gemmm_gemmk1_grid_desc,
+            make_tuple(make_right_pad_transform(GemmK0Total, GemmK0Pad - GemmK0Total),
+                       make_pass_through_transform(GemmM),
+                       make_pass_through_transform(N1Number)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+            out_gemmk0pad_gemmm_gemmk1_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0)),
+                       make_pass_through_transform(GemmM),
+                       make_pass_through_transform(N1Number)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        // B: input tensor
+        const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+            in_n_hi_wi_c_grid_desc,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                       make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+            in_n_hip_wip_c_grid_desc,
+            make_tuple(
+                make_pass_through_transform(N),
+                make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+        const auto in_n0_y_ho_x_wo_c_n1_grid_desc =
+            transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                        make_tuple(make_unmerge_transform(make_tuple(N0, N1Number)),
+                                                   make_pass_through_transform(Y),
+                                                   make_pass_through_transform(Ho),
+                                                   make_pass_through_transform(X),
+                                                   make_pass_through_transform(Wo),
+                                                   make_pass_through_transform(C)),
+                                        make_tuple(Sequence<0>{},
+                                                   Sequence<1>{},
+                                                   Sequence<2>{},
+                                                   Sequence<3>{},
+                                                   Sequence<4>{},
+                                                   Sequence<5>{}),
+                                        make_tuple(Sequence<0, 6>{},
+                                                   Sequence<1>{},
+                                                   Sequence<2>{},
+                                                   Sequence<3>{},
+                                                   Sequence<4>{},
+                                                   Sequence<5>{}));
+
+        const auto in_gemmk0total_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+            in_n0_y_ho_x_wo_c_n1_grid_desc,
+            make_tuple(make_merge_transform(make_tuple(N0, Ho, Wo)),
+                       make_merge_transform(make_tuple(Y, X, C)),
+                       make_pass_through_transform(N1Number)),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}, Sequence<6>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        const auto in_gemmk0pad_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+            in_gemmk0total_gemmn_gemmk1_grid_desc,
+            make_tuple(make_right_pad_transform(GemmK0Total, GemmK0Pad - GemmK0Total),
+                       make_pass_through_transform(GemmN),
+                       make_pass_through_transform(N1Number)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+            in_gemmk0pad_gemmn_gemmk1_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0)),
+                       make_pass_through_transform(GemmN),
+                       make_pass_through_transform(N1Number)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        // C: weight tensor
+        const auto wei_gemmm_gemmn_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+        return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                          in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                          wei_gemmm_gemmn_grid_desc);
+    }
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, 1));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXdl,
+        NPerXdl,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        ABlockLdsM1PerBlock,
+        ABlockLdsM0PerBlock,
+        ABlockLdsM1Padding,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        BBlockLdsN1PerBlock,
+        BBlockLdsN0PerBlock,
+        BBlockLdsN1Padding,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXdl,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        true,
+        true>;
+
+    using GridwiseGemmAtomicAdd = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::AtomicAdd,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXdl,
+        NPerXdl,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        ABlockLdsM1PerBlock,
+        ABlockLdsM0PerBlock,
+        ABlockLdsM1Padding,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        BBlockLdsN1PerBlock,
+        BBlockLdsN0PerBlock,
+        BBlockLdsN1Padding,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXdl,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        true,
+        true>;
+    // Argument
+    using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        decltype(GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}));
+
+    using Block2CTileMap =
+        decltype(GridwiseGemm::MakeCBlockClusterAdaptor(CGridDesc_M_N{}, 1, 1, 1));
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+                 std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+                 std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+                 std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+                 std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+                 std::array<ck::index_t, NDimSpatial> input_left_pads,
+                 std::array<ck::index_t, NDimSpatial> input_right_pads,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op,
+                 ck::index_t split_k)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_in_grid},
+              p_c_grid_{p_wei_grid},
+              a_grid_desc_kbatch_k0_m_k1_{},
+              b_grid_desc_kbatch_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{out_element_op},
+              b_element_op_{in_element_op},
+              c_element_op_{wei_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              output_spatial_lengths_{output_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads},
+              k_batch_{split_k}
+        {
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
+                                                                          K,
+                                                                          C,
+                                                                          input_spatial_lengths,
+                                                                          filter_spatial_lengths,
+                                                                          output_spatial_lengths,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads,
+                                                                          k_batch_);
+
+            a_grid_desc_kbatch_k0_m_k1_ = descs[I0];
+            b_grid_desc_kbatch_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_            = descs[I2];
+
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_kbatch_k0_m_k1_,
+                                           b_grid_desc_kbatch_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n_);
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_kbatch_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_kbatch_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        Block2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        InElementwiseOperation a_element_op_;
+        OutElementwiseOperation b_element_op_;
+        WeiElementwiseOperation c_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+        std::array<index_t, NDimSpatial> output_spatial_lengths_;
+        std::array<index_t, NDimSpatial> filter_spatial_lengths_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+        index_t k_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        void ShowInfo(const Argument& arg)
+        {
+            std::cout << "arg.a_grid_desc_kbatch_k0_m_k1_{"
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I2) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.b_grid_desc_kbatch_k0_n_k1_{"
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I0) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I1) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I2) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                      << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+        }
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            ShowInfo(arg);
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                            arg.b_grid_desc_kbatch_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight has invalid setting");
+            }
+            const auto kbatch = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0);
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K0 = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+
+            const auto Run = [&](const auto& kernel) {
+                hipGetErrorString(hipMemset(
+                    arg.p_c_grid_,
+                    0,
+                    arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
+                        sizeof(CDataType)));
+
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.a_grid_desc_kbatch_k0_m_k1_,
+                                           arg.b_grid_desc_kbatch_k0_n_k1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.block_2_ctile_map_);
+            };
+
+            if(has_main_k0_block_loop)
+            {
+                if(kbatch == 1)
+                {
+                    const auto kernel = kernel_gemm_xdlops_bwd_weight<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        OutElementwiseOperation,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        remove_reference_t<DeviceOp::Block2CTileMap>,
+                        true>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_bwd_weight<
+                        GridwiseGemmAtomicAdd,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        OutElementwiseOperation,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        remove_reference_t<DeviceOp::Block2CTileMap>,
+                        true>;
+
+                    Run(kernel);
+                }
+            }
+            else
+            {
+                if(kbatch == 1)
+                {
+                    const auto kernel = kernel_gemm_xdlops_bwd_weight<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        OutElementwiseOperation,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        remove_reference_t<DeviceOp::Block2CTileMap>,
+                        false>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_bwd_weight<
+                        GridwiseGemmAtomicAdd,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        OutElementwiseOperation,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        remove_reference_t<DeviceOp::Block2CTileMap>,
+                        false>;
+
+                    Run(kernel);
+                }
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_K_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // unmerge N to N0 and N1, where N1 equals to K1
+        if(!(arg.Conv_N_ % K1 == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_C_ % CBlockTransferScalarPerVector_NWaveNPerXdl == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                           arg.b_grid_desc_kbatch_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             WeiDataType* p_wei_grid,
+                             const OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+                             std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+                             std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+                             std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+                             std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+                             std::array<ck::index_t, NDimSpatial> input_left_pads,
+                             std::array<ck::index_t, NDimSpatial> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op,
+                             ck::index_t split_k)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op,
+                        split_k};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        void* p_wei_grid,
+                        const void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+                        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+                        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+                        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+                        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+                        std::array<ck::index_t, NDimSpatial> input_left_pads,
+                        std::array<ck::index_t, NDimSpatial> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op,
+                        ck::index_t split_k) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<WeiDataType*>(p_wei_grid),
+                                          static_cast<const OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op,
+                                          split_k);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..ca79b93
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,835 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionBackwardDataSpecialization ConvBackwardDataSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXdl,
+          ck::index_t NPerXdl,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector>
+struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvBwdData<2,
+                               ck::tensor_layout::convolution::NHWC,
+                               ck::tensor_layout::convolution::KYXC,
+                               ck::tensor_layout::convolution::NHWK,
+                               InDataType,
+                               WeiDataType,
+                               OutDataType,
+                               InElementwiseOperation,
+                               WeiElementwiseOperation,
+                               OutElementwiseOperation>
+{
+    using DeviceOp = DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = OutDataType;
+    using BDataType = WeiDataType;
+    using CDataType = InDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr index_t NDimSpatial = 2;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static_assert((K1 % ABlockTransferThreadClusterLengths_K0_M_K1{}[I2]) %
+                      ABlockTransferSrcScalarPerVector ==
+                  0);
+    static_assert((NPerBlock / BBlockTransferThreadClusterLengths_K0_N_K1{}[I1]) %
+                      BBlockTransferSrcScalarPerVector ==
+                  0);
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    index_t i_ytilde,
+                                                    index_t i_xtilde)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const auto K0 = K / K1;
+
+        const auto out_n_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K));
+        const auto wei_k_y_x_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Ho * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_ho_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Ho, I0, I0),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                out_n_hop_wop_k_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(YDot, HTilde),
+                                         make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_slice_transform(YDot, I0, YDotSlice),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_slice_transform(XDot, I0, XDotSlice),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_unmerge_transform(make_tuple(K0, K1))),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5, 6>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                wei_k_y_x_c_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_embed_transform(make_tuple(YDot, YTilde),
+                                                make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                           make_embed_transform(make_tuple(XDot, XTilde),
+                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto wei_k0_k1_ydotslice_xdotslice_c_grid_desc =
+                transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_slice_transform(YDot, I0, YDotSlice),
+                                                       make_slice_transform(XDot, I0, XDotSlice),
+                                                       make_freeze_transform(i_ytilde),
+                                                       make_freeze_transform(i_xtilde),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{},
+                                                       Sequence<1>{},
+                                                       Sequence<3>{},
+                                                       Sequence<2>{},
+                                                       Sequence<4>{},
+                                                       Sequence<5>{}),
+                                            make_tuple(Sequence<0, 1>{},
+                                                       Sequence<2>{},
+                                                       Sequence<3>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<4>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(YTilde, HTilde),
+                                                make_tuple(ConvDilationH, ConvStrideH)),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_ytilde),
+                           make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<>{},
+                           Sequence<2>{},
+                           Sequence<3>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_htildeslice_wtildeslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, 0, 0));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXdl,
+        NPerXdl,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
+        7,                                // CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_in_grid},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{out_element_op},
+              b_element_op_{wei_element_op},
+              c_element_op_{in_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              input_spatial_lengths_{input_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              output_spatial_lengths_{output_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            const index_t ConvStrideH = conv_filter_strides[0];
+            const index_t ConvStrideW = conv_filter_strides[1];
+
+            const index_t ConvDilationH = conv_filter_dilations[0];
+            const index_t ConvDilationW = conv_filter_dilations[1];
+
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+            {
+                for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+                {
+                    // check slice is valid
+                    const index_t Y      = filter_spatial_lengths_[0];
+                    const index_t X      = filter_spatial_lengths_[1];
+                    const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                    const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                    if(YDotSlice * XDotSlice <= 0)
+                    {
+                        continue;
+                    }
+
+                    const auto descs = DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        i_ytilde,
+                        i_xtilde);
+                    a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                    b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                    c_grid_desc_m_n_container_.push_back(descs[I2]);
+
+                    auto block_2_ctile_map =
+                        GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2], M01, N01);
+
+                    if(GridwiseGemm::CheckValidity(
+                           descs[I0], descs[I1], descs[I2], block_2_ctile_map))
+                    {
+                        c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_.push_back(
+                            GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(descs[I2]));
+
+                        block_2_ctile_map_container_.push_back(block_2_ctile_map);
+                    }
+                }
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        std::vector<AGridDesc_K0_M_K1> a_grid_desc_k0_m_k1_container_;
+        std::vector<BGridDesc_K0_N_K1> b_grid_desc_k0_n_k1_container_;
+        std::vector<CGridDesc_M_N> c_grid_desc_m_n_container_;
+        std::vector<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>
+            c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_;
+        std::vector<typename GridwiseGemm::DefaultBlock2CTileMap> block_2_ctile_map_container_;
+        index_t M01_;
+        index_t N01_;
+        OutElementwiseOperation a_element_op_;
+        WeiElementwiseOperation b_element_op_;
+        InElementwiseOperation c_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+
+        std::vector<ck::index_t> input_spatial_lengths_;
+        std::vector<ck::index_t> filter_spatial_lengths_;
+        std::vector<ck::index_t> output_spatial_lengths_;
+        std::vector<ck::index_t> conv_filter_strides_;
+        std::vector<ck::index_t> conv_filter_dilations_;
+        std::vector<ck::index_t> input_left_pads_;
+        std::vector<ck::index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float ave_time = 0;
+            for(size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+            {
+#if 0
+                {
+                    std::cout << "arg.a_grid_desc_k0_m_k1_container_{"
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) << ", "
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I1) << ", "
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.b_grid_desc_k0_n_k1_container_{"
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I0) << ", "
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I1) << ", "
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I2) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.c_grid_desc_m_n_container_{ "
+                              << arg.c_grid_desc_m_n_container_[i].GetLength(I0) << ", "
+                              << arg.c_grid_desc_m_n_container_[i].GetLength(I1) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I0)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I1)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I2)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I3)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I4)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I5)
+                              << " ) " << std::endl;
+                }
+#endif
+
+                if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
+                                                arg.b_grid_desc_k0_n_k1_container_[i],
+                                                arg.c_grid_desc_m_n_container_[i],
+                                                arg.block_2_ctile_map_container_[i]))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
+                }
+
+                const index_t grid_size = arg.block_2_ctile_map_container_[i].CalculateGridSize(
+                    arg.c_grid_desc_m_n_container_[i]);
+
+                const auto K = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) *
+                               arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2);
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r3<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<
+                            typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                        OutElementwiseOperation,
+                        WeiElementwiseOperation,
+                        InElementwiseOperation,
+                        remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                        true>;
+
+                    ave_time += launch_and_time_kernel(
+                        stream_config,
+                        kernel,
+                        dim3(grid_size),
+                        dim3(BlockSize),
+                        0,
+                        arg.p_a_grid_,
+                        arg.p_b_grid_,
+                        arg.p_c_grid_,
+                        arg.a_grid_desc_k0_m_k1_container_[i],
+                        arg.b_grid_desc_k0_n_k1_container_[i],
+                        arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i],
+                        arg.a_element_op_,
+                        arg.b_element_op_,
+                        arg.c_element_op_,
+                        arg.block_2_ctile_map_container_[i]);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r3<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<
+                            typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                        OutElementwiseOperation,
+                        WeiElementwiseOperation,
+                        InElementwiseOperation,
+                        remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                        false>;
+
+                    ave_time += launch_and_time_kernel(
+                        stream_config,
+                        kernel,
+                        dim3(grid_size),
+                        dim3(BlockSize),
+                        0,
+                        arg.p_a_grid_,
+                        arg.p_b_grid_,
+                        arg.p_c_grid_,
+                        arg.a_grid_desc_k0_m_k1_container_[i],
+                        arg.b_grid_desc_k0_n_k1_container_[i],
+                        arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i],
+                        arg.a_element_op_,
+                        arg.b_element_op_,
+                        arg.c_element_op_,
+                        arg.block_2_ctile_map_container_[i]);
+                }
+            }
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 pad = 0 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 1 &&
+             arg.Conv_K_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_C_ % CThreadTransferDstScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        for(std::size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
+                                            arg.b_grid_desc_k0_n_k1_container_[i],
+                                            arg.c_grid_desc_m_n_container_[i],
+                                            arg.block_2_ctile_map_container_[i]))
+            {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             const OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* p_in_grid,
+                        const void* p_wei_grid,
+                        const void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<const OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..4749665
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,968 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] =
+//     activate(in[N, Hi, Wi, C] * wei[K, Y, X, C] + bias[K]) + residual[N, Ho, Wo, K]
+template <
+    typename InDataType,
+    typename WeiDataType,
+    typename OutDataType,
+    typename AccDataType,
+    typename InElementwiseOperation,
+    typename WeiElementwiseOperation,
+    typename OutElementwiseOperation,
+    ConvolutionForwardSpecialization ConvForwardSpecialization,
+    ck::index_t BlockSize,
+    ck::index_t MPerBlock,
+    ck::index_t NPerBlock,
+    ck::index_t K0PerBlock,
+    ck::index_t K1,
+    ck::index_t MPerXDL,
+    ck::index_t NPerXDL,
+    ck::index_t MXdlPerWave,
+    ck::index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_K0_M_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    ck::index_t ABlockTransferSrcVectorDim,
+    ck::index_t ABlockTransferSrcScalarPerVector,
+    ck::index_t ABlockTransferDstScalarPerVector_K1,
+    bool ABlockLdsAddExtraM,
+    typename BBlockTransferThreadClusterLengths_K0_N_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    ck::index_t BBlockTransferSrcVectorDim,
+    ck::index_t BBlockTransferSrcScalarPerVector,
+    ck::index_t BBlockTransferDstScalarPerVector_K1,
+    bool BBlockLdsAddExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+    index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
+struct
+    DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvFwdBiasActivationAdd<InElementwiseOperation,
+                                            WeiElementwiseOperation,
+                                            OutElementwiseOperation>
+{
+    using DeviceOp =
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    // TODO make it support any # of spatial dimensions
+    static constexpr index_t NDimSpatial = 2;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmMRaw = N * Ho * Wo;
+        const index_t GemmN    = K;
+
+        const auto GemmM    = math::integer_least_multiple(GemmMRaw, MPerBlock);
+        const auto GemmMPad = GemmM - GemmMRaw;
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        { // 1x1, stride=1, pad=0
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_gemmmraw_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, C));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmmraw_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            // C1: residual tensor: assume same layout as output tensor
+            const auto resi_grid_desc_gemmm_gemmn = out_gemmm_gemmn_grid_desc;
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn,
+                              resi_grid_desc_gemmm_gemmn);
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        { // 1x1, pad=0
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_n_ho_wo_c_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_merge_transform(make_tuple(N, Ho, Wo))),
+                make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            // C1: residual tensor: assume same layout as output tensor
+            const auto resi_grid_desc_gemmm_gemmn = out_gemmm_gemmn_grid_desc;
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn,
+                              resi_grid_desc_gemmm_gemmn);
+        }
+        else if constexpr(ConvForwardSpecialization == ConvolutionForwardSpecialization::OddC)
+        { // C = odd value
+            const index_t GemmKRaw = Y * X * C;
+            const index_t GemmK = math::integer_least_multiple(GemmKRaw, K0PerBlock * GemmK1Number);
+            const index_t GemmKPad = GemmK - GemmKRaw;
+            const index_t GemmK0   = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmkraw_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
+                in_gemmkraw_gemmmraw_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKRaw, GemmKPad),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_right_pad_transform(GemmKRaw, GemmKPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            // C1: residual tensor: assume same layout as output tensor
+            const auto resi_grid_desc_gemmm_gemmn = out_gemmm_gemmn_grid_desc;
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn,
+                              resi_grid_desc_gemmm_gemmn);
+        }
+        else
+        {
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmk_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmmraw_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmMRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            // C1: residual tensor: assume same layout as output tensor
+            const auto resi_grid_desc_gemmm_gemmn = out_gemmm_gemmn_grid_desc;
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn,
+                              resi_grid_desc_gemmm_gemmn);
+        }
+    }
+
+    using GridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(GridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(GridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(GridDescs{}[I2])>;
+    using C0GridDesc_M_N    = remove_cvref_t<decltype(GridDescs{}[I3])>;
+    using C1GridDesc_M_N    = remove_cvref_t<decltype(GridDescs{}[I4])>;
+
+    using Block2CTileMap = BlockToCTileMap_M00_N0_M01<MPerBlock, NPerBlock, CGridDesc_M_N>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        C0GridDesc_M_N,
+        C1GridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,                 // ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // BBlockTransferSrcAccessOrder,
+        2,                 // BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+        CBlockTransferScalarPerVector_NWaveNPerXdl>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 OutDataType* p_out_grid,
+                 const OutDataType* p_bias_grid,
+                 const OutDataType* p_resi_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_in_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_out_grid},
+              p_c0_grid_{p_bias_grid},
+              p_c1_grid_{p_resi_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c0_grid_desc_m_n_{},
+              c1_grid_desc_m_n_{},
+              c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
+              c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
+              c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
+              block_2_ctile_map_{},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              input_spatial_lengths_{input_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              output_spatial_lengths_{output_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
+                                                                          K,
+                                                                          C,
+                                                                          input_spatial_lengths,
+                                                                          filter_spatial_lengths,
+                                                                          output_spatial_lengths,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+            c0_grid_desc_m_n_    = descs[I3];
+            c1_grid_desc_m_n_    = descs[I4];
+
+            block_2_ctile_map_ = Block2CTileMap{c_grid_desc_m_n_};
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
+                                           b_grid_desc_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
+                    GridwiseGemm::
+                        MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                            c_grid_desc_m_n_);
+
+                c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
+                    GridwiseGemm::
+                        MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                            c0_grid_desc_m_n_);
+
+                c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
+                    GridwiseGemm::
+                        MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                            c1_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        const CDataType* p_c0_grid_;
+        const CDataType* p_c1_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        C0GridDesc_M_N c0_grid_desc_m_n_;
+        C1GridDesc_M_N c1_grid_desc_m_n_;
+        typename GridwiseGemm::
+            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
+        typename GridwiseGemm::
+            C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
+        typename GridwiseGemm::
+            C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
+        Block2CTileMap block_2_ctile_map_;
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+        std::vector<index_t> input_spatial_lengths_;
+        std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> output_spatial_lengths_;
+        std::vector<index_t> conv_filter_strides_;
+        std::vector<index_t> conv_filter_dilations_;
+        std::vector<index_t> input_left_pads_;
+        std::vector<index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << DeviceOp{}.GetTypeString() << std::endl;
+                std::cout << "N " << arg.Conv_N_ << ", "
+                          << "K " << arg.Conv_K_ << ", "
+                          << "C " << arg.Conv_C_ << ", " << std::endl;
+                std::cout << "Y X " << arg.filter_spatial_lengths_[0] << ", "
+                          << arg.filter_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Hi Wi " << arg.input_spatial_lengths_[0] << ", "
+                          << arg.input_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Ho Wo " << arg.output_spatial_lengths_[0] << ", "
+                          << arg.output_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Strides " << arg.conv_filter_strides_[0] << ", "
+                          << arg.conv_filter_strides_[1] << ", " << std::endl;
+                std::cout << "Dilations " << arg.conv_filter_dilations_[0] << ", "
+                          << arg.conv_filter_dilations_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_left_pads_[0] << ", "
+                          << arg.input_left_pads_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_right_pads_[0] << ", "
+                          << arg.input_right_pads_[1] << ", " << std::endl;
+            }
+
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+                std::cout << "arg.c0_grid_desc_m_n_{ " << arg.c0_grid_desc_m_n_.GetLength(I0)
+                          << ", " << arg.c0_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+                std::cout << "arg.c1_grid_desc_m_n_{ " << arg.c1_grid_desc_m_n_.GetLength(I0)
+                          << ", " << arg.c1_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r3 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    Block2CTileMap,
+                    true>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_c_grid_,
+                    arg.p_c0_grid_,
+                    arg.p_c1_grid_,
+                    arg.a_grid_desc_k0_m_k1_,
+                    arg.b_grid_desc_k0_n_k1_,
+                    arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.in_element_op_,
+                    arg.wei_element_op_,
+                    arg.out_element_op_,
+                    arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    Block2CTileMap,
+                    false>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_c_grid_,
+                    arg.p_c0_grid_,
+                    arg.p_c1_grid_,
+                    arg.a_grid_desc_k0_m_k1_,
+                    arg.b_grid_desc_k0_n_k1_,
+                    arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.in_element_op_,
+                    arg.wei_element_op_,
+                    arg.out_element_op_,
+                    arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_C_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_K_ % CBlockTransferScalarPerVector_NWaveNPerXdl == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             OutDataType* p_out_grid,
+                             const OutDataType* p_bias_grid,
+                             const OutDataType* p_resi_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        p_bias_grid,
+                        p_resi_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        const void* p_wei_grid,
+                        void* p_out_grid,
+                        const void* p_bias_grid,
+                        const void* p_resi_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<OutDataType*>(p_out_grid),
+                                          static_cast<const OutDataType*>(p_bias_grid),
+                                          static_cast<const OutDataType*>(p_resi_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..bafbfe4
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,925 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] =
+//     activate(in[N, Hi, Wi, C] * wei[K, Y, X, C] + bias[K])
+template <
+    typename InDataType,
+    typename WeiDataType,
+    typename OutDataType,
+    typename AccDataType,
+    typename InElementwiseOperation,
+    typename WeiElementwiseOperation,
+    typename OutElementwiseOperation,
+    InMemoryDataOperationEnum OutGlobalMemoryDataOperation,
+    ConvolutionForwardSpecialization ConvForwardSpecialization,
+    ck::index_t BlockSize,
+    ck::index_t MPerBlock,
+    ck::index_t NPerBlock,
+    ck::index_t K0PerBlock,
+    ck::index_t K1,
+    ck::index_t MPerXDL,
+    ck::index_t NPerXDL,
+    ck::index_t MXdlPerWave,
+    ck::index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_K0_M_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    ck::index_t ABlockTransferSrcVectorDim,
+    ck::index_t ABlockTransferSrcScalarPerVector,
+    ck::index_t ABlockTransferDstScalarPerVector_K1,
+    bool ABlockLdsAddExtraM,
+    typename BBlockTransferThreadClusterLengths_K0_N_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    ck::index_t BBlockTransferSrcVectorDim,
+    ck::index_t BBlockTransferSrcScalarPerVector,
+    ck::index_t BBlockTransferDstScalarPerVector_K1,
+    bool BBlockLdsAddExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+    index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
+struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvFwdBiasActivation<InElementwiseOperation,
+                                         WeiElementwiseOperation,
+                                         OutElementwiseOperation>
+{
+    using DeviceOp =
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    // TODO make it support any # of spatial dimensions
+    static constexpr index_t NDimSpatial = 2;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmMRaw = N * Ho * Wo;
+        const index_t GemmN    = K;
+
+        const auto GemmM    = math::integer_least_multiple(GemmMRaw, MPerBlock);
+        const auto GemmMPad = GemmM - GemmMRaw;
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        { // 1x1, stride=1, pad=0
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_gemmmraw_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, C));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmmraw_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn);
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        { // 1x1, pad=0
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_n_ho_wo_c_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_merge_transform(make_tuple(N, Ho, Wo))),
+                make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn);
+        }
+        else if constexpr(ConvForwardSpecialization == ConvolutionForwardSpecialization::OddC)
+        { // C = odd value
+            const index_t GemmKRaw = Y * X * C;
+            const index_t GemmK = math::integer_least_multiple(GemmKRaw, K0PerBlock * GemmK1Number);
+            const index_t GemmKPad = GemmK - GemmKRaw;
+            const index_t GemmK0   = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmkraw_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
+                in_gemmkraw_gemmmraw_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKRaw, GemmKPad),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_right_pad_transform(GemmKRaw, GemmKPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn);
+        }
+        else
+        {
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmk_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmmraw_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmMRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn);
+        }
+    }
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+    using C0GridDesc_M_N    = remove_cvref_t<decltype(ABCGridDescs{}[I3])>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        OutGlobalMemoryDataOperation,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        C0GridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,                 // ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // BBlockTransferSrcAccessOrder,
+        2,                 // BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+        CBlockTransferScalarPerVector_NWaveNPerXdl>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 OutDataType* p_out_grid,
+                 const OutDataType* p_bias_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_in_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_out_grid},
+              p_c0_grid_{p_bias_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c0_grid_desc_m_n_{},
+              c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
+              c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              input_spatial_lengths_{input_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              output_spatial_lengths_{output_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
+                                                                          K,
+                                                                          C,
+                                                                          input_spatial_lengths,
+                                                                          filter_spatial_lengths,
+                                                                          output_spatial_lengths,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+            c0_grid_desc_m_n_    = descs[I3];
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
+                                           b_grid_desc_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
+                    GridwiseGemm::
+                        MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                            c_grid_desc_m_n_);
+
+                c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
+                    GridwiseGemm::
+                        MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                            c0_grid_desc_m_n_);
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        const CDataType* p_c0_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        C0GridDesc_M_N c0_grid_desc_m_n_;
+        typename GridwiseGemm::
+            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
+        typename GridwiseGemm::
+            C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+        std::vector<index_t> input_spatial_lengths_;
+        std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> output_spatial_lengths_;
+        std::vector<index_t> conv_filter_strides_;
+        std::vector<index_t> conv_filter_dilations_;
+        std::vector<index_t> input_left_pads_;
+        std::vector<index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << DeviceOp{}.GetTypeString() << std::endl;
+                std::cout << "N " << arg.Conv_N_ << ", "
+                          << "K " << arg.Conv_K_ << ", "
+                          << "C " << arg.Conv_C_ << ", " << std::endl;
+                std::cout << "Y X " << arg.filter_spatial_lengths_[0] << ", "
+                          << arg.filter_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Hi Wi " << arg.input_spatial_lengths_[0] << ", "
+                          << arg.input_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Ho Wo " << arg.output_spatial_lengths_[0] << ", "
+                          << arg.output_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Strides " << arg.conv_filter_strides_[0] << ", "
+                          << arg.conv_filter_strides_[1] << ", " << std::endl;
+                std::cout << "Dilations " << arg.conv_filter_dilations_[0] << ", "
+                          << arg.conv_filter_dilations_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_left_pads_[0] << ", "
+                          << arg.input_left_pads_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_right_pads_[0] << ", "
+                          << arg.input_right_pads_[1] << ", " << std::endl;
+            }
+
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+                std::cout << "arg.c0_grid_desc_m_n_{ " << arg.c0_grid_desc_m_n_.GetLength(I0)
+                          << ", " << arg.c0_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r2 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r2<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    true>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_c_grid_,
+                    arg.p_c0_grid_,
+                    arg.a_grid_desc_k0_m_k1_,
+                    arg.b_grid_desc_k0_n_k1_,
+                    arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.in_element_op_,
+                    arg.wei_element_op_,
+                    arg.out_element_op_,
+                    arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r2<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_c_grid_,
+                    arg.p_c0_grid_,
+                    arg.a_grid_desc_k0_m_k1_,
+                    arg.b_grid_desc_k0_n_k1_,
+                    arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.in_element_op_,
+                    arg.wei_element_op_,
+                    arg.out_element_op_,
+                    arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_C_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_K_ % CBlockTransferScalarPerVector_NWaveNPerXdl == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             OutDataType* p_out_grid,
+                             const OutDataType* p_bias_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        p_bias_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        const void* p_wei_grid,
+                        void* p_out_grid,
+                        const void* p_bias_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<OutDataType*>(p_out_grid),
+                                          static_cast<const OutDataType*>(p_bias_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..6a6d24b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,893 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <
+    typename InDataType,
+    typename WeiDataType,
+    typename OutDataType,
+    typename AccDataType,
+    typename InElementwiseOperation,
+    typename WeiElementwiseOperation,
+    typename OutElementwiseOperation,
+    ConvolutionForwardSpecialization ConvForwardSpecialization,
+    ck::index_t BlockSize,
+    ck::index_t MPerBlock,
+    ck::index_t NPerBlock,
+    ck::index_t K0PerBlock,
+    ck::index_t K1,
+    ck::index_t MPerXdl,
+    ck::index_t NPerXdl,
+    ck::index_t MXdlPerWave,
+    ck::index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_K0_M_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    ck::index_t ABlockTransferSrcVectorDim,
+    ck::index_t ABlockTransferSrcScalarPerVector,
+    ck::index_t ABlockTransferDstScalarPerVector_K1,
+    bool ABlockLdsAddExtraM,
+    typename BBlockTransferThreadClusterLengths_K0_N_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    ck::index_t BBlockTransferSrcVectorDim,
+    ck::index_t BBlockTransferSrcScalarPerVector,
+    ck::index_t BBlockTransferDstScalarPerVector_K1,
+    bool BBlockLdsAddExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+    index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
+struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvFwd<2,
+                           ck::tensor_layout::convolution::NHWC,
+                           ck::tensor_layout::convolution::KYXC,
+                           ck::tensor_layout::convolution::NHWK,
+                           InDataType,
+                           WeiDataType,
+                           OutDataType,
+                           InElementwiseOperation,
+                           WeiElementwiseOperation,
+                           OutElementwiseOperation>
+{
+    using DeviceOp = DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr index_t NDimSpatial = 2;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmMRaw = N * Ho * Wo;
+        const index_t GemmN    = K;
+
+        const auto GemmM    = math::integer_least_multiple(GemmMRaw, MPerBlock);
+        const auto GemmMPad = GemmM - GemmMRaw;
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        { // 1x1, stride=1, pad=0
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_gemmmraw_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, C));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmmraw_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        { // 1x1, pad=0
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_n_ho_wo_c_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_merge_transform(make_tuple(N, Ho, Wo))),
+                make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+        else if constexpr(ConvForwardSpecialization == ConvolutionForwardSpecialization::OddC)
+        { // C = odd value
+            const index_t GemmKRaw = Y * X * C;
+            const index_t GemmK = math::integer_least_multiple(GemmKRaw, K0PerBlock * GemmK1Number);
+            const index_t GemmKPad = GemmK - GemmKRaw;
+            const index_t GemmK0   = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmkraw_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
+                in_gemmkraw_gemmmraw_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKRaw, GemmKPad),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_right_pad_transform(GemmKRaw, GemmKPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmk_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmmraw_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmMRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+    }
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    using Block2CTileMap = BlockToCTileMap_M00_N0_M01<MPerBlock, NPerBlock, CGridDesc_M_N>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType, // TODO: Add ShuffleType for DeviceConv2d
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock * K1,
+        K1, // AK1
+        K1, // BK1
+        MPerXdl,
+        NPerXdl,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,                 // ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // BBlockTransferSrcAccessOrder,
+        2,                 // BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+        CBlockTransferScalarPerVector_NWaveNPerXdl>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_in_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_out_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
+              block_2_ctile_map_{},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              input_spatial_lengths_{input_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              output_spatial_lengths_{output_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
+                                                                          K,
+                                                                          C,
+                                                                          input_spatial_lengths,
+                                                                          filter_spatial_lengths,
+                                                                          output_spatial_lengths,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+
+            block_2_ctile_map_ = Block2CTileMap{c_grid_desc_m_n_};
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
+                                           b_grid_desc_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
+                    GridwiseGemm::
+                        MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                            c_grid_desc_m_n_);
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::
+            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
+        Block2CTileMap block_2_ctile_map_;
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+        std::vector<index_t> input_spatial_lengths_;
+        std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> output_spatial_lengths_;
+        std::vector<index_t> conv_filter_strides_;
+        std::vector<index_t> conv_filter_dilations_;
+        std::vector<index_t> input_left_pads_;
+        std::vector<index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << DeviceOp{}.GetTypeString() << std::endl;
+                std::cout << "N " << arg.Conv_N_ << ", "
+                          << "K " << arg.Conv_K_ << ", "
+                          << "C " << arg.Conv_C_ << ", " << std::endl;
+                std::cout << "Y X " << arg.filter_spatial_lengths_[0] << ", "
+                          << arg.filter_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Hi Wi " << arg.input_spatial_lengths_[0] << ", "
+                          << arg.input_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Ho Wo " << arg.output_spatial_lengths_[0] << ", "
+                          << arg.output_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Strides " << arg.conv_filter_strides_[0] << ", "
+                          << arg.conv_filter_strides_[1] << ", " << std::endl;
+                std::cout << "Dilations " << arg.conv_filter_dilations_[0] << ", "
+                          << arg.conv_filter_dilations_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_left_pads_[0] << ", "
+                          << arg.input_left_pads_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_right_pads_[0] << ", "
+                          << arg.input_right_pads_[1] << ", " << std::endl;
+            }
+
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+                std::cout
+                    << "arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_"
+                       "nwavenperxdl_{ "
+                    << arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_
+                           .GetLength(I0)
+                    << ", "
+                    << arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_
+                           .GetLength(I1)
+                    << ", "
+                    << arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_
+                           .GetLength(I2)
+                    << ", "
+                    << arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_
+                           .GetLength(I3)
+                    << ", "
+                    << arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_
+                           .GetLength(I4)
+                    << ", "
+                    << arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_
+                           .GetLength(I5)
+                    << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    Block2CTileMap,
+                    true>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_c_grid_,
+                    arg.a_grid_desc_k0_m_k1_,
+                    arg.b_grid_desc_k0_n_k1_,
+                    arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.in_element_op_,
+                    arg.wei_element_op_,
+                    arg.out_element_op_,
+                    arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    Block2CTileMap,
+                    false>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_c_grid_,
+                    arg.a_grid_desc_k0_m_k1_,
+                    arg.b_grid_desc_k0_n_k1_,
+                    arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.in_element_op_,
+                    arg.wei_element_op_,
+                    arg.out_element_op_,
+                    arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_C_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_K_ % CBlockTransferScalarPerVector_NWaveNPerXdl == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        const void* p_wei_grid,
+                        void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000..5821e06
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,733 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector>
+struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvFwd<2,
+                           ck::tensor_layout::convolution::NHWC,
+                           ck::tensor_layout::convolution::KYXC,
+                           ck::tensor_layout::convolution::NHWK,
+                           InDataType,
+                           WeiDataType,
+                           OutDataType,
+                           InElementwiseOperation,
+                           WeiElementwiseOperation,
+                           OutElementwiseOperation>
+{
+    using DeviceOp = DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr index_t NDimSpatial = 2;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmMRaw = N * Ho * Wo;
+        const index_t GemmN    = K;
+        const index_t GemmK    = Y * X * C;
+
+        const auto GemmMPad = math::integer_least_multiple(GemmMRaw, MPerBlock) - GemmMRaw;
+
+        assert(GemmK % GemmK1Number == 0);
+
+        const index_t GemmK0 = GemmK / GemmK1Number;
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: input tensor
+            const auto in_gemmmraw_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, C));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmmraw_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_n_ho_wo_c_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_merge_transform(make_tuple(N, Ho, Wo))),
+                make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmk_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmmraw_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmMRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+    }
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,                 // ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // BBlockTransferSrcAccessOrder,
+        2,                 // BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
+        7,                                // CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_in_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_out_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
+                                                                          K,
+                                                                          C,
+                                                                          input_spatial_lengths,
+                                                                          filter_spatial_lengths,
+                                                                          output_spatial_lengths,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
+                                           b_grid_desc_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
+            c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+        std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> conv_filter_strides_;
+        std::vector<index_t> input_left_pads_;
+        std::vector<index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    true>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.in_element_op_,
+                                                  arg.wei_element_op_,
+                                                  arg.out_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.in_element_op_,
+                                                  arg.wei_element_op_,
+                                                  arg.out_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_C_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_K_ % CThreadTransferDstScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        const void* p_wei_grid,
+                        void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp
new file mode 100644
index 0000000..f69d8f1
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp
@@ -0,0 +1,268 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef DEVICE_CONV3D_FWD_NAIVE_HPP
+#define DEVICE_CONV3D_FWD_NAIVE_HPP
+
+#include <iostream>
+#include <memory>
+#include <sstream>
+#include "conv_util.hpp"
+#include "device.hpp"
+#include "device_conv_fwd.hpp"
+#include "common_header.hpp"
+#include "naive_conv_fwd.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// specialization for #D conv: in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+template <typename InDataType,
+          typename WeiDataType, // WeiDataType must be the same as InDataType
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceConv3dFwdNaive_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K
+    : public DeviceConvFwd<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>
+
+{
+    using DeviceOp = DeviceConv3dFwdNaive_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in,
+                 const WeiDataType* p_wei,
+                 OutDataType* p_out,
+                 const index_t N,
+                 const index_t K,
+                 const index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : params_{3,
+                      N,
+                      K,
+                      C,
+                      filter_spatial_lengths,
+                      input_spatial_lengths,
+                      conv_filter_strides,
+                      conv_filter_dilations,
+                      input_left_pads,
+                      input_right_pads},
+              out_spatial_lengths_{output_spatial_lengths},
+              p_in_{p_in},
+              p_wei_{p_wei},
+              p_out_{p_out},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+
+        {
+        }
+
+        //  private:
+        utils::conv::ConvParams params_;
+        std::vector<index_t> out_spatial_lengths_;
+
+        const InDataType* p_in_;
+        const WeiDataType* p_wei_;
+        OutDataType* p_out_;
+
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto naive_conv3d_fwd =
+                ref::naive_conv_fwd_ndhwc_kzyxc_ndhwk<InDataType,
+                                                      WeiDataType,
+                                                      OutDataType,
+                                                      AccDataType,
+                                                      InElementwiseOperation,
+                                                      WeiElementwiseOperation,
+                                                      OutElementwiseOperation>;
+
+            float ave_time = launch_and_time_kernel(stream_config,
+                                                    naive_conv3d_fwd,
+                                                    dim3(256),
+                                                    dim3(256),
+                                                    0,
+                                                    arg.p_in_,
+                                                    arg.p_wei_,
+                                                    arg.p_out_,
+                                                    arg.N_,
+                                                    arg.K_,
+                                                    arg.C_,
+                                                    arg.in_spatial_lengths_[0],
+                                                    arg.in_spatial_lengths_[1],
+                                                    arg.in_spatial_lengths_[2],
+                                                    arg.filter_spatial_lengths_[0],
+                                                    arg.filter_spatial_lengths_[1],
+                                                    arg.filter_spatial_lengths_[2],
+                                                    arg.out_spatial_lengths_[0],
+                                                    arg.out_spatial_lengths_[1],
+                                                    arg.out_spatial_lengths_[2],
+                                                    arg.conv_filter_strides_[0],
+                                                    arg.conv_filter_strides_[1],
+                                                    arg.conv_filter_strides_[2],
+                                                    arg.conv_filter_dilations_[0],
+                                                    arg.conv_filter_dilations_[1],
+                                                    arg.conv_filter_dilations_[2],
+                                                    arg.in_left_pads_[0],
+                                                    arg.in_left_pads_[1],
+                                                    arg.in_left_pads_[2]);
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        std::vector<index_t> out_spatial_lengths = arg.params_.GetOutputSpatialLengths();
+
+        bool out_lengths_are_consistent = out_spatial_lengths[0] == arg.out_spatial_lengths_[0] &&
+                                          out_spatial_lengths[1] == arg.out_spatial_lengths_[1] &&
+                                          out_spatial_lengths[2] == arg.out_spatial_lengths_[2];
+        return out_lengths_are_consistent;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in,
+                             const WeiDataType* p_wei,
+                             OutDataType* p_out,
+                             const index_t N,
+                             const index_t K,
+                             const index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in,
+                        p_wei,
+                        p_out,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        const void* p_wei,
+                        void* p_out,
+                        const index_t N,
+                        const index_t K,
+                        const index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in),
+                                          static_cast<const WeiDataType*>(p_wei),
+                                          static_cast<OutDataType*>(p_out),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv3dFwdNaive_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K<>";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp
new file mode 100644
index 0000000..f950538
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp
@@ -0,0 +1,642 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef DEVICE_CONV3D_FWD_XDL_HPP
+#define DEVICE_CONV3D_FWD_XDL_HPP
+
+#include <iostream>
+#include <memory>
+#include <sstream>
+#include "device.hpp"
+#include "device_conv_fwd.hpp"
+#include "common_header.hpp"
+#include "tensor_layout.hpp"
+#include "convolution_forward_specialization.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp"
+#include "gridwise_gemm_xdlops_v2r3.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/*
+ * \see \link impl/device_batched_gemm_xdl.hpp kernel_batched_gemm_xdlops_v2r3() \endlink.
+ */
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v2r3_for_conv3d(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const index_t num_batches,
+            const index_t a_batch_stride,
+            const index_t b_batch_stride,
+            const index_t c_batch_stride,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / num_batches);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset =
+        __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(a_batch_stride) * g_idx);
+    const long_index_t b_batch_offset =
+        __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(b_batch_stride) * g_idx);
+    const long_index_t c_batch_offset =
+        __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(c_batch_stride) * g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_c_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  block_2_ctile_map);
+
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = num_batches;
+    ignore = a_batch_stride;
+    ignore = b_batch_stride;
+    ignore = c_batch_stride;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// specialization for #D conv: in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+template <typename InDataType,
+          typename WeiDataType, // WeiDataType must be the same as InDataType
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector>
+struct DeviceConv3dFwdXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K
+    : public DeviceConvFwd<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>
+
+{
+    using DeviceOp = DeviceConv3dFwdXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    /*
+     * \brief Split the number of batches, \p N, into N = B * N1, such that the memory
+     * space of input and output tensors stays with the value range of index_t, and each subbatch
+     * can be dealed with GridwiseGemm.
+     */
+    static index_t GetMaxAllowableSubBatchSize(const index_t N,
+                                               const index_t K,
+                                               const index_t C,
+                                               std::vector<ck::index_t> input_spatial_lengths,
+                                               std::vector<ck::index_t> output_spatial_lengths)
+    {
+        const index_t Di = input_spatial_lengths[0];
+        const index_t Hi = input_spatial_lengths[1];
+        const index_t Wi = input_spatial_lengths[2];
+
+        const index_t Do = output_spatial_lengths[0];
+        const index_t Ho = output_spatial_lengths[1];
+        const index_t Wo = output_spatial_lengths[2];
+
+        // N1 should satisfy that
+        //   1) N % N1 = 0;
+        //   2) N1 * (Do * Ho * Wo * K) < (2^31 - 1)
+        //   3) N1 * (Di * Hi * Wi * C) < (2^31 - 1)
+        //
+        // Do NOT confuse (B, N1) in this function with (B, N1) in gridewise GEMM.
+        auto N1 = N + 1;
+
+        const auto stride =
+            math::max(long_index_t(Do) * Ho * Wo * K, long_index_t(Di) * Hi * Wi * C);
+        const index_t max_stride = NumericLimits<index_t>::Max();
+
+        for(index_t n0 = 1; n0 <= N; ++n0)
+        {
+            index_t n1 = N / n0;
+            if(n0 * n1 == N && long_index_t(n1) * long_index_t(stride) < max_stride)
+            {
+                N1 = n1;
+                break;
+            }
+        }
+
+        const auto B = N / N1;
+        if(B * N1 != N)
+        {
+            throw std::runtime_error(__func__ +
+                                     std::string(": failed to find num_subbatches for conv3d.\n"));
+        }
+
+        return N1;
+    }
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(const index_t N,
+                                                    const index_t K,
+                                                    const index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        assert(input_spatial_lengths.size() > 2);
+        assert(filter_spatial_lengths.size() > 2);
+        assert(conv_filter_strides.size() > 2);
+        assert(conv_filter_dilations.size() > 2);
+        assert(input_left_pads.size() > 2);
+        assert(input_right_pads.size() > 2);
+
+        const index_t Di = input_spatial_lengths[0];
+        const index_t Hi = input_spatial_lengths[1];
+        const index_t Wi = input_spatial_lengths[2];
+        const index_t Z  = filter_spatial_lengths[0];
+        const index_t Y  = filter_spatial_lengths[1];
+        const index_t X  = filter_spatial_lengths[2];
+
+        const index_t Do = output_spatial_lengths[0];
+        const index_t Ho = output_spatial_lengths[1];
+        const index_t Wo = output_spatial_lengths[2];
+
+        static_assert(ConvForwardSpecialization == ConvolutionForwardSpecialization::Default,
+                      "Wrong! This specialization not implemented!");
+
+        const auto in_desc_n_di_hi_wi_c =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+        const auto wei_desc_k_z_y_x_c =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Z, Y, X, C));
+        const auto out_desc_n_do_ho_wo_k =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Do, Ho, Wo, K));
+
+        const auto descs = transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk_pad(
+            in_desc_n_di_hi_wi_c,
+            wei_desc_k_z_y_x_c,
+            out_desc_n_do_ho_wo_k,
+            make_tuple(conv_filter_strides[0], conv_filter_strides[1], conv_filter_strides[2]),
+            make_tuple(
+                conv_filter_dilations[0], conv_filter_dilations[1], conv_filter_dilations[2]),
+            make_tuple(input_left_pads[0], input_left_pads[1], input_left_pads[2]),
+            make_tuple(input_right_pads[0], input_right_pads[1], input_right_pads[2]),
+            Number<K1>{});
+
+        return descs;
+    }
+
+    using ABCGridDescs = remove_cvref_t<decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}))>;
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        InDataType,
+        AccDataType,
+        OutDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
+        7,
+        CThreadTransferDstScalarPerVector>;
+
+    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
+    using Block2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in,
+                 const WeiDataType* p_wei,
+                 OutDataType* p_out,
+                 const index_t N,
+                 const index_t K,
+                 const index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 index_t M01,
+                 index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_in},
+              p_b_grid_{p_wei},
+              p_c_grid_{p_out},
+              M01_{M01},
+              N01_{N01},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+            const index_t subbatch_size =
+                GetMaxAllowableSubBatchSize(N, K, C, input_spatial_lengths, output_spatial_lengths);
+            num_subbatches_ = N / subbatch_size;
+
+            const auto descs =
+                MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(subbatch_size,
+                                                                K,
+                                                                C,
+                                                                input_spatial_lengths,
+                                                                filter_spatial_lengths,
+                                                                output_spatial_lengths,
+                                                                conv_filter_strides,
+                                                                conv_filter_dilations,
+                                                                input_left_pads,
+                                                                input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+
+            a_batch_stride_ = a_grid_desc_k0_m_k1_.GetElementSpaceSize();
+            b_batch_stride_ = 0;
+            c_batch_stride_ = c_grid_desc_m_n_.GetElementSpaceSize();
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
+                                           b_grid_desc_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const InDataType* p_a_grid_;
+        const WeiDataType* p_b_grid_;
+        OutDataType* p_c_grid_;
+        index_t num_subbatches_;
+        index_t a_batch_stride_;
+        index_t b_batch_stride_;
+        index_t c_batch_stride_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
+        Block2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            {
+                std::cout << "num_batches_of_GEMM = " << arg.num_subbatches_ << std::endl;
+                std::cout << "a_grid_desc_k0_m_k1{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "b_grid_desc_k0_n_k1{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "c_grid_desc_m_n{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) *
+                arg.num_subbatches_;
+
+            const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+            if(has_main_k0_block_loop)
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3_for_conv3d<
+                    GridwiseGemm,
+                    InDataType,
+                    OutDataType,
+                    remove_reference_t<AGridDesc_K0_M_K1>,
+                    remove_reference_t<BGridDesc_K0_N_K1>,
+                    remove_reference_t<CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<Block2CTileMap>,
+                    true>;
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.num_subbatches_,
+                                                  arg.a_batch_stride_,
+                                                  arg.b_batch_stride_,
+                                                  arg.c_batch_stride_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.in_element_op_,
+                                                  arg.wei_element_op_,
+                                                  arg.out_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3_for_conv3d<
+                    GridwiseGemm,
+                    InDataType,
+                    OutDataType,
+                    remove_reference_t<AGridDesc_K0_M_K1>,
+                    remove_reference_t<BGridDesc_K0_N_K1>,
+                    remove_reference_t<CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<Block2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.num_subbatches_,
+                                                  arg.a_batch_stride_,
+                                                  arg.b_batch_stride_,
+                                                  arg.c_batch_stride_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.in_element_op_,
+                                                  arg.wei_element_op_,
+                                                  arg.out_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in,
+                             const WeiDataType* p_wei,
+                             OutDataType* p_out,
+                             const index_t N,
+                             const index_t K,
+                             const index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in,
+                        p_wei,
+                        p_out,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        const void* p_wei,
+                        void* p_out,
+                        const index_t N,
+                        const index_t K,
+                        const index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in),
+                                          static_cast<const WeiDataType*>(p_wei),
+                                          static_cast<OutDataType*>(p_out),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv3dFwdXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp
new file mode 100644
index 0000000..1dc39d9
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp
@@ -0,0 +1,1583 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionBackwardDataSpecialization ConvBackwardDataSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          index_t M1PerThread,
+          index_t N1PerThread,
+          index_t KPerThread,
+          typename M1N1ThreadClusterM1Xs,
+          typename M1N1ThreadClusterN1Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct DeviceConvNdBwdDataNwcKxcNwk_Dl
+    : public DeviceConvBwdData<
+          NDimSpatial,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::NWC,
+                                        ck::tensor_layout::convolution::NHWC,
+                                        ck::tensor_layout::convolution::NDHWC>>,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::KXC,
+                                        ck::tensor_layout::convolution::KYXC,
+                                        ck::tensor_layout::convolution::KZYXC>>,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::NWK,
+                                        ck::tensor_layout::convolution::NHWK,
+                                        ck::tensor_layout::convolution::NDHWK>>,
+          InDataType,
+          WeiDataType,
+          OutDataType,
+          InElementwiseOperation,
+          WeiElementwiseOperation,
+          OutElementwiseOperation>
+{
+    using DeviceOp = DeviceConvNdBwdDataNwcKxcNwk_Dl;
+
+    using ADataType = OutDataType;
+    using BDataType = WeiDataType;
+    using CDataType = InDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    std::vector<ck::index_t> tildes)
+    {
+        using namespace ck;
+
+        index_t i_xtilde = tildes[0];
+
+        const index_t Wi            = input_spatial_lengths[0];
+        const index_t Wo            = output_spatial_lengths[0];
+        const index_t X             = filter_spatial_lengths[0];
+        const index_t InLeftPadW    = input_left_pads[0];
+        const index_t InRightPadW   = input_right_pads[0];
+        const index_t ConvStrideW   = conv_filter_strides[0];
+        const index_t ConvDilationW = conv_filter_dilations[0];
+
+        const auto K0 = K / K1;
+
+        const auto in_n_wi_c_grid_desc = make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto out_n_wo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Wo, K));
+            const auto wei_k_x_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, X, C));
+
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto out_n_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                out_n_wop_k_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto out_n_xdotslice_wtildeslice_k0_k1_grid_desc = transform_tensor_descriptor(
+                out_n_xdot_wtilde_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_slice_transform(XDot, I0, XDotSlice),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(XDotSlice, K0)),
+                           make_merge_transform(make_tuple(N, WTildeSlice)),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3>{}, Sequence<0, 2>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                wei_k_x_c_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_embed_transform(make_tuple(XDot, XTilde),
+                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto wei_k0_k1_xdotslice_c_grid_desc = transform_tensor_descriptor(
+                wei_k_xdot_xtilde_c_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                           make_slice_transform(XDot, I0, XDotSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<>{}, Sequence<3>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 0>{}, Sequence<3>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+                in_n_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_n_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                in_n_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_wtildeslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, WTildeSlice)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    std::vector<ck::index_t> tildes)
+    {
+        using namespace ck;
+
+        index_t i_ytilde = tildes[0];
+        index_t i_xtilde = tildes[1];
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const auto K0 = K / K1;
+
+        const auto out_n_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K));
+        const auto wei_k_y_x_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Ho * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_ho_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Ho, I0, I0),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                out_n_hop_wop_k_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(YDot, HTilde),
+                                         make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_slice_transform(YDot, I0, YDotSlice),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_slice_transform(XDot, I0, XDotSlice),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_unmerge_transform(make_tuple(K0, K1))),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5, 6>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                wei_k_y_x_c_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_embed_transform(make_tuple(YDot, YTilde),
+                                                make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                           make_embed_transform(make_tuple(XDot, XTilde),
+                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto wei_k0_k1_ydotslice_xdotslice_c_grid_desc =
+                transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_slice_transform(YDot, I0, YDotSlice),
+                                                       make_slice_transform(XDot, I0, XDotSlice),
+                                                       make_freeze_transform(i_ytilde),
+                                                       make_freeze_transform(i_xtilde),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{},
+                                                       Sequence<1>{},
+                                                       Sequence<3>{},
+                                                       Sequence<2>{},
+                                                       Sequence<4>{},
+                                                       Sequence<5>{}),
+                                            make_tuple(Sequence<0, 1>{},
+                                                       Sequence<2>{},
+                                                       Sequence<3>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<4>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(YTilde, HTilde),
+                                                make_tuple(ConvDilationH, ConvStrideH)),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_ytilde),
+                           make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<>{},
+                           Sequence<2>{},
+                           Sequence<3>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_htildeslice_wtildeslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    std::vector<ck::index_t> tildes)
+    {
+        using namespace ck;
+
+        const index_t i_ztilde = tildes[0];
+        const index_t i_ytilde = tildes[1];
+        const index_t i_xtilde = tildes[2];
+
+        const index_t Di = input_spatial_lengths[0];
+        const index_t Hi = input_spatial_lengths[1];
+        const index_t Wi = input_spatial_lengths[2];
+
+        const index_t Do = output_spatial_lengths[0];
+        const index_t Ho = output_spatial_lengths[1];
+        const index_t Wo = output_spatial_lengths[2];
+
+        const index_t Z = filter_spatial_lengths[0];
+        const index_t Y = filter_spatial_lengths[1];
+        const index_t X = filter_spatial_lengths[2];
+
+        const index_t InLeftPadD = input_left_pads[0];
+        const index_t InLeftPadH = input_left_pads[1];
+        const index_t InLeftPadW = input_left_pads[2];
+
+        const index_t InRightPadD = input_right_pads[0];
+        const index_t InRightPadH = input_right_pads[1];
+        const index_t InRightPadW = input_right_pads[2];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        const index_t ConvDilationD = conv_filter_dilations[0];
+        const index_t ConvDilationH = conv_filter_dilations[1];
+        const index_t ConvDilationW = conv_filter_dilations[2];
+
+        const auto K0 = K / K1;
+
+        const auto out_n_do_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Do, Ho, Wo, K));
+        const auto wei_k_z_y_x_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Z, Y, X, C));
+        const auto in_n_di_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Do * Ho * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Do), make_tuple(I1, ConvStrideD)),
+                           make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{},
+                           Sequence<3>{},
+                           Sequence<5>{},
+                           Sequence<0, 2, 4, 6>{},
+                           Sequence<7>{}),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto ZDot = math::integer_divide_ceil(Z, ZTilde);
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto DTilde =
+                Do + math::integer_divide_ceil(ConvDilationD * (Z - I1), ConvStrideD);
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IDTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadD - ConvDilationD * (ZTilde - I1)), ConvStrideD);
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IDTildeSliceEnd = math::min(
+                DTilde, math::integer_divide_ceil(InLeftPadD + Di - I1, ConvStrideD) + I1);
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto DTildeSlice = IDTildeSliceEnd - IDTildeSliceBegin;
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_dop_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_do_ho_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Do, I0, I0),
+                           make_pad_transform(Ho, I0, I0),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_dop_hop_wop_k_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(N),
+                        make_embed_transform(make_tuple(ZDot, DTilde),
+                                             make_tuple(-ConvDilationD / GcdStrideDilationD, I1)),
+                        make_embed_transform(make_tuple(YDot, HTilde),
+                                             make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                        make_embed_transform(make_tuple(XDot, WTilde),
+                                             make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                        make_pass_through_transform(K)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+            const auto
+                out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
+                    transform_tensor_descriptor(
+                        out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_slice_transform(ZDot, I0, ZDotSlice),
+                                   make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                                   make_slice_transform(YDot, I0, YDotSlice),
+                                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                                   make_slice_transform(XDot, I0, XDotSlice),
+                                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                                   make_unmerge_transform(make_tuple(K0, K1))),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{},
+                                   Sequence<6>{},
+                                   Sequence<7>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{},
+                                   Sequence<6>{},
+                                   Sequence<7, 8>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(
+                    make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, K0)),
+                    make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                    make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}, Sequence<8>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc =
+                transform_tensor_descriptor(
+                    wei_k_z_y_x_c_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(K),
+                        make_embed_transform(make_tuple(ZDot, ZTilde),
+                                             make_tuple(ConvStrideD / GcdStrideDilationD, I1)),
+                        make_embed_transform(make_tuple(YDot, YTilde),
+                                             make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                        make_embed_transform(make_tuple(XDot, XTilde),
+                                             make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                        make_pass_through_transform(C)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+            const auto wei_k0_k1_zdotslice_ydotslice_xdotslice_c_grid_desc =
+                transform_tensor_descriptor(wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_slice_transform(ZDot, I0, ZDotSlice),
+                                                       make_slice_transform(YDot, I0, YDotSlice),
+                                                       make_slice_transform(XDot, I0, XDotSlice),
+                                                       make_freeze_transform(i_ztilde),
+                                                       make_freeze_transform(i_ytilde),
+                                                       make_freeze_transform(i_xtilde),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{},
+                                                       Sequence<1>{},
+                                                       Sequence<3>{},
+                                                       Sequence<5>{},
+                                                       Sequence<2>{},
+                                                       Sequence<4>{},
+                                                       Sequence<6>{},
+                                                       Sequence<7>{}),
+                                            make_tuple(Sequence<0, 1>{},
+                                                       Sequence<2>{},
+                                                       Sequence<3>{},
+                                                       Sequence<4>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<5>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_zdotslice_ydotslice_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 3, 4, 0>{}, Sequence<5>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc =
+                transform_tensor_descriptor(
+                    in_n_dip_hip_wip_c_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_embed_transform(make_tuple(ZTilde, DTilde),
+                                                    make_tuple(ConvDilationD, ConvStrideD)),
+                               make_embed_transform(make_tuple(YTilde, HTilde),
+                                                    make_tuple(ConvDilationH, ConvStrideH)),
+                               make_embed_transform(make_tuple(XTilde, WTilde),
+                                                    make_tuple(ConvDilationW, ConvStrideW)),
+                               make_pass_through_transform(C)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+            const auto in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc =
+                transform_tensor_descriptor(
+                    in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_freeze_transform(i_ztilde),
+                               make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                               make_freeze_transform(i_ytilde),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_freeze_transform(i_xtilde),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{},
+                               Sequence<6>{},
+                               Sequence<7>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<>{},
+                               Sequence<1>{},
+                               Sequence<>{},
+                               Sequence<2>{},
+                               Sequence<>{},
+                               Sequence<3>{},
+                               Sequence<4>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc,
+                make_tuple(
+                    make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<1>(
+            1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1}, {0});
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<2>(
+            1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {0, 0});
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<3>(1,
+                                                                  1,
+                                                                  1,
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {0, 0, 0});
+    }
+
+    using ABCGridDescs = decltype(GetABCGridDesc<NDimSpatial>());
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDl_km_kn_mn_v1r3<BlockSize,
+                                     ADataType,
+                                     AccDataType,
+                                     CDataType,
+                                     InMemoryDataOperationEnum::Set,
+                                     AGridDesc_K0_M_K1,
+                                     BGridDesc_K0_N_K1,
+                                     CGridDesc_M_N,
+                                     MPerBlock,
+                                     NPerBlock,
+                                     K0PerBlock,
+                                     K1,
+                                     M1PerThread,
+                                     N1PerThread,
+                                     KPerThread,
+                                     M1N1ThreadClusterM1Xs,
+                                     M1N1ThreadClusterN1Xs,
+                                     ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterArrangeOrder,
+                                     ABlockTransferSrcAccessOrder,
+                                     ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                     ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                     ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                     BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterArrangeOrder,
+                                     BBlockTransferSrcAccessOrder,
+                                     BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                     BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                     BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                     CThreadTransferSrcDstAccessOrder,
+                                     CThreadTransferSrcDstVectorDim,
+                                     CThreadTransferDstScalarPerVector>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+    using DefaultBlock2CTileMap =
+        decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(CGridDesc_M_N{}));
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_in_grid},
+              a_element_op_{out_element_op},
+              b_element_op_{wei_element_op},
+              c_element_op_{in_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              input_spatial_lengths_{input_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              output_spatial_lengths_{output_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            CreateABCDesc<NDimSpatial>();
+        }
+
+        template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+        void CreateABCDesc()
+        {
+            const index_t ConvStrideW     = conv_filter_strides_[0];
+            const index_t ConvDilationW   = conv_filter_dilations_[0];
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+            const auto XTilde             = ConvStrideW / GcdStrideDilationW;
+
+            const index_t X = filter_spatial_lengths_[0];
+
+            for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+            {
+                // check slice is valid
+                const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                if(XDotSlice <= 0)
+                {
+                    continue;
+                }
+
+                const auto descs =
+                    DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                        Conv_N_,
+                        Conv_K_,
+                        Conv_C_,
+                        input_spatial_lengths_,
+                        filter_spatial_lengths_,
+                        output_spatial_lengths_,
+                        conv_filter_strides_,
+                        conv_filter_dilations_,
+                        input_left_pads_,
+                        input_right_pads_,
+                        {i_xtilde});
+                a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                c_grid_desc_m_n_container_.push_back(descs[I2]);
+
+                if(GridwiseGemm::CheckValidity(descs[I0], descs[I1], descs[I2]))
+                {
+                    a_grid_desc_k0_m0_m1_k1_container_.push_back(
+                        GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(descs[I0]));
+                    b_grid_desc_k0_n0_n1_k1_container_.push_back(
+                        GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(descs[I1]));
+                    c_grid_desc_m0_m10_m11_n0_n10_n11_container_.push_back(
+                        GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(descs[I2]));
+
+                    block_2_ctile_map_container_.push_back(
+                        GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2]));
+                }
+            }
+        }
+        template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+        void CreateABCDesc()
+        {
+            const index_t ConvStrideH = conv_filter_strides_[0];
+            const index_t ConvStrideW = conv_filter_strides_[1];
+
+            const index_t ConvDilationH = conv_filter_dilations_[0];
+            const index_t ConvDilationW = conv_filter_dilations_[1];
+
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const index_t Y = filter_spatial_lengths_[0];
+            const index_t X = filter_spatial_lengths_[1];
+            for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+            {
+                for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+                {
+                    // check slice is valid
+                    const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                    const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                    if(YDotSlice * XDotSlice <= 0)
+                    {
+                        continue;
+                    }
+
+                    const auto descs =
+                        DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                            Conv_N_,
+                            Conv_K_,
+                            Conv_C_,
+                            input_spatial_lengths_,
+                            filter_spatial_lengths_,
+                            output_spatial_lengths_,
+                            conv_filter_strides_,
+                            conv_filter_dilations_,
+                            input_left_pads_,
+                            input_right_pads_,
+                            {i_ytilde, i_xtilde});
+                    a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                    b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                    c_grid_desc_m_n_container_.push_back(descs[I2]);
+
+                    if(GridwiseGemm::CheckValidity(descs[I0], descs[I1], descs[I2]))
+                    {
+                        a_grid_desc_k0_m0_m1_k1_container_.push_back(
+                            GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(descs[I0]));
+                        b_grid_desc_k0_n0_n1_k1_container_.push_back(
+                            GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(descs[I1]));
+                        c_grid_desc_m0_m10_m11_n0_n10_n11_container_.push_back(
+                            GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(descs[I2]));
+
+                        block_2_ctile_map_container_.push_back(
+                            GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2]));
+                    }
+                }
+            }
+        }
+        template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+        void CreateABCDesc()
+        {
+            const index_t ConvStrideD = conv_filter_strides_[0];
+            const index_t ConvStrideH = conv_filter_strides_[1];
+            const index_t ConvStrideW = conv_filter_strides_[2];
+
+            const index_t ConvDilationD = conv_filter_dilations_[0];
+            const index_t ConvDilationH = conv_filter_dilations_[1];
+            const index_t ConvDilationW = conv_filter_dilations_[2];
+
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const index_t Z = filter_spatial_lengths_[0];
+            const index_t Y = filter_spatial_lengths_[1];
+            const index_t X = filter_spatial_lengths_[2];
+            for(index_t i_ztilde = 0; i_ztilde < ZTilde; ++i_ztilde)
+            {
+                for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+                {
+                    for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+                    {
+                        // check slice is valid
+                        const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
+                        const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                        const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                        if(ZDotSlice * YDotSlice * XDotSlice <= 0)
+                        {
+                            continue;
+                        }
+
+                        const auto descs =
+                            DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                                Conv_N_,
+                                Conv_K_,
+                                Conv_C_,
+                                input_spatial_lengths_,
+                                filter_spatial_lengths_,
+                                output_spatial_lengths_,
+                                conv_filter_strides_,
+                                conv_filter_dilations_,
+                                input_left_pads_,
+                                input_right_pads_,
+                                {i_ztilde, i_ytilde, i_xtilde});
+                        a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                        b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                        c_grid_desc_m_n_container_.push_back(descs[I2]);
+
+                        if(GridwiseGemm::CheckValidity(descs[I0], descs[I1], descs[I2]))
+                        {
+                            a_grid_desc_k0_m0_m1_k1_container_.push_back(
+                                GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(descs[I0]));
+                            b_grid_desc_k0_n0_n1_k1_container_.push_back(
+                                GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(descs[I1]));
+                            c_grid_desc_m0_m10_m11_n0_n10_n11_container_.push_back(
+                                GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(descs[I2]));
+
+                            block_2_ctile_map_container_.push_back(
+                                GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2]));
+                        }
+                    }
+                }
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        std::vector<AGridDesc_K0_M_K1> a_grid_desc_k0_m_k1_container_;
+        std::vector<BGridDesc_K0_N_K1> b_grid_desc_k0_n_k1_container_;
+        std::vector<CGridDesc_M_N> c_grid_desc_m_n_container_;
+
+        std::vector<AGridDesc_K0_M0_M1_K1> a_grid_desc_k0_m0_m1_k1_container_;
+        std::vector<BGridDesc_K0_N0_N1_K1> b_grid_desc_k0_n0_n1_k1_container_;
+        std::vector<CGridDesc_M0_M10_M11_N0_N10_N11> c_grid_desc_m0_m10_m11_n0_n10_n11_container_;
+
+        std::vector<DefaultBlock2CTileMap> block_2_ctile_map_container_;
+
+        // element-wise op
+        OutElementwiseOperation a_element_op_;
+        WeiElementwiseOperation b_element_op_;
+        InElementwiseOperation c_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+
+        std::vector<ck::index_t> input_spatial_lengths_;
+        std::vector<ck::index_t> filter_spatial_lengths_;
+        std::vector<ck::index_t> output_spatial_lengths_;
+        std::vector<ck::index_t> conv_filter_strides_;
+        std::vector<ck::index_t> conv_filter_dilations_;
+        std::vector<ck::index_t> input_left_pads_;
+        std::vector<ck::index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float ave_time = 0;
+            for(size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+            {
+                {
+                    std::cout << "arg.a_grid_desc_k0_m_k1_container_{"
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) << ", "
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I1) << ", "
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.b_grid_desc_k0_n_k1_container_{"
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I0) << ", "
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I1) << ", "
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I2) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.c_grid_desc_m_n_container_{ "
+                              << arg.c_grid_desc_m_n_container_[i].GetLength(I0) << ", "
+                              << arg.c_grid_desc_m_n_container_[i].GetLength(I1) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_( "
+                              << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I0)
+                              << ", "
+                              << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I1)
+                              << ", "
+                              << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I2)
+                              << ", "
+                              << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I3)
+                              << ", "
+                              << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I4)
+                              << ", "
+                              << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I5)
+                              << " ) " << std::endl;
+                }
+
+                if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
+                                                arg.b_grid_desc_k0_n_k1_container_[i],
+                                                arg.c_grid_desc_m_n_container_[i]))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
+                }
+
+                const index_t grid_size = arg.block_2_ctile_map_container_[i].CalculateGridSize(
+                    arg.c_grid_desc_m_n_container_[i]);
+
+                auto launch_kernel = [&](auto has_main_k_block_loop,
+                                         auto has_double_tail_k_block_loop) {
+                    constexpr bool has_main_loop   = has_main_k_block_loop.value;
+                    constexpr bool has_double_loop = has_double_tail_k_block_loop;
+
+                    const auto kernel = kernel_gemm_dl_v1r3<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M0_M1_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N0_N1_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_M0_M10_M11_N0_N10_N11>,
+                        remove_reference_t<DeviceOp::DefaultBlock2CTileMap>,
+                        has_main_loop,
+                        has_double_loop>;
+
+                    ave_time +=
+                        launch_and_time_kernel(stream_config,
+                                               kernel,
+                                               dim3(grid_size),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.p_a_grid_,
+                                               arg.p_b_grid_,
+                                               arg.p_c_grid_,
+                                               arg.a_grid_desc_k0_m0_m1_k1_container_[i],
+                                               arg.b_grid_desc_k0_n0_n1_k1_container_[i],
+                                               arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i],
+                                               arg.block_2_ctile_map_container_[i]);
+                };
+
+                const auto K0 = arg.a_grid_desc_k0_m0_m1_k1_container_[i].GetLength(I0);
+                const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+                const bool has_double_tail_k_block_loop =
+                    GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+                if(has_main_k_block_loop && has_double_tail_k_block_loop)
+                {
+                    launch_kernel(integral_constant<bool, true>{}, integral_constant<bool, true>{});
+                }
+                else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+                {
+                    launch_kernel(integral_constant<bool, true>{},
+                                  integral_constant<bool, false>{});
+                }
+                else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+                {
+                    launch_kernel(integral_constant<bool, false>{},
+                                  integral_constant<bool, true>{});
+                }
+                else
+                {
+                    launch_kernel(integral_constant<bool, false>{},
+                                  integral_constant<bool, false>{});
+                }
+            }
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        // check device
+        if(!(ck::get_device_name() == "gfx906"||ck::get_device_name() == "gfx926" || ck::get_device_name() == "gfx1030"))
+        {
+            return false;
+        }
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 pad = 0 conv
+            for(int i = 0; i < NDimSpatial; i++)
+            {
+                if(!(arg.filter_spatial_lengths_[i] == 1 && arg.conv_filter_strides_[i] == 1 &&
+                     arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // matrix A
+        {
+            auto srcVectorLengths = ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t K = arg.Conv_K_;
+
+            if(K % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+
+        // matrix B
+        {
+            auto srcLoadLenghts   = BBlockTransferThreadSliceLengths_K0_N0_N1_K1{};
+            auto srcVectorLengths = BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1{};
+            if(srcVectorLengths[I0] != 1 || srcVectorLengths[I3] != 1)
+            {
+                return false;
+            }
+            if(srcLoadLenghts[I1] % srcVectorLengths[I1] != 0 ||
+               srcLoadLenghts[I2] % srcVectorLengths[I2] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.Conv_K_;
+
+            if(C % (srcVectorLengths[I1] * srcVectorLengths[I2]) != 0)
+            {
+                return false;
+            }
+        }
+        // vector store C matrix into global memory
+        if(!(arg.Conv_C_ % CThreadTransferDstScalarPerVector == 0))
+        {
+            std::cout << "Not surpport,because: arg.Conv_C_ % CThreadTransferDstScalarPerVector = "
+                      << arg.Conv_C_ % CThreadTransferDstScalarPerVector << std::endl;
+            return false;
+        }
+
+        // Gridwise GEMM size
+        for(std::size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
+                                            arg.b_grid_desc_k0_n_k1_container_[i],
+                                            arg.c_grid_desc_m_n_container_[i]))
+            {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             const OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* p_in_grid,
+                        const void* p_wei_grid,
+                        const void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<const OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConvNdBwdDataNwcKxcNwk_Dl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock
+            << ">";
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0){
+
+            str<< " Filter1x1Stride1Pad0";
+        }
+
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp
new file mode 100644
index 0000000..e10e374
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp
@@ -0,0 +1,1568 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionBackwardDataSpecialization ConvBackwardDataSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXdl,
+          ck::index_t NPerXdl,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector>
+struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
+    : public DeviceConvBwdData<
+          NDimSpatial,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::NWC,
+                                        ck::tensor_layout::convolution::NHWC,
+                                        ck::tensor_layout::convolution::NDHWC>>,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::KXC,
+                                        ck::tensor_layout::convolution::KYXC,
+                                        ck::tensor_layout::convolution::KZYXC>>,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::NWK,
+                                        ck::tensor_layout::convolution::NHWK,
+                                        ck::tensor_layout::convolution::NDHWK>>,
+          InDataType,
+          WeiDataType,
+          OutDataType,
+          InElementwiseOperation,
+          WeiElementwiseOperation,
+          OutElementwiseOperation>
+{
+    using DeviceOp = DeviceConvNdBwdDataNwcKxcNwk_Xdl;
+
+    using ADataType = OutDataType;
+    using BDataType = WeiDataType;
+    using CDataType = InDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    static_assert((K1 % ABlockTransferThreadClusterLengths_K0_M_K1{}[I2]) %
+                      ABlockTransferSrcScalarPerVector ==
+                  0);
+    static_assert((NPerBlock / BBlockTransferThreadClusterLengths_K0_N_K1{}[I1]) %
+                      BBlockTransferSrcScalarPerVector ==
+                  0);
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    std::vector<ck::index_t> tildes)
+    {
+        using namespace ck;
+
+        index_t i_xtilde = tildes[0];
+
+        const index_t Wi            = input_spatial_lengths[0];
+        const index_t Wo            = output_spatial_lengths[0];
+        const index_t X             = filter_spatial_lengths[0];
+        const index_t InLeftPadW    = input_left_pads[0];
+        const index_t InRightPadW   = input_right_pads[0];
+        const index_t ConvStrideW   = conv_filter_strides[0];
+        const index_t ConvDilationW = conv_filter_dilations[0];
+
+        const auto K0 = K / K1;
+
+        const auto in_n_wi_c_grid_desc = make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto out_n_wo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Wo, K));
+            const auto wei_k_x_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, X, C));
+
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto out_n_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                out_n_wop_k_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto out_n_xdotslice_wtildeslice_k0_k1_grid_desc = transform_tensor_descriptor(
+                out_n_xdot_wtilde_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_slice_transform(XDot, I0, XDotSlice),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(XDotSlice, K0)),
+                           make_merge_transform(make_tuple(N, WTildeSlice)),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3>{}, Sequence<0, 2>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                wei_k_x_c_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_embed_transform(make_tuple(XDot, XTilde),
+                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto wei_k0_k1_xdotslice_c_grid_desc = transform_tensor_descriptor(
+                wei_k_xdot_xtilde_c_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                           make_slice_transform(XDot, I0, XDotSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<>{}, Sequence<3>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 0>{}, Sequence<3>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+                in_n_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_n_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                in_n_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_wtildeslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, WTildeSlice)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    std::vector<ck::index_t> tildes)
+    {
+        using namespace ck;
+
+        index_t i_ytilde = tildes[0];
+        index_t i_xtilde = tildes[1];
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const auto K0 = K / K1;
+
+        const auto out_n_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K));
+        const auto wei_k_y_x_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Ho * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_ho_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Ho, I0, I0),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                out_n_hop_wop_k_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(YDot, HTilde),
+                                         make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_slice_transform(YDot, I0, YDotSlice),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_slice_transform(XDot, I0, XDotSlice),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_unmerge_transform(make_tuple(K0, K1))),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5, 6>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                wei_k_y_x_c_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_embed_transform(make_tuple(YDot, YTilde),
+                                                make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                           make_embed_transform(make_tuple(XDot, XTilde),
+                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto wei_k0_k1_ydotslice_xdotslice_c_grid_desc =
+                transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_slice_transform(YDot, I0, YDotSlice),
+                                                       make_slice_transform(XDot, I0, XDotSlice),
+                                                       make_freeze_transform(i_ytilde),
+                                                       make_freeze_transform(i_xtilde),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{},
+                                                       Sequence<1>{},
+                                                       Sequence<3>{},
+                                                       Sequence<2>{},
+                                                       Sequence<4>{},
+                                                       Sequence<5>{}),
+                                            make_tuple(Sequence<0, 1>{},
+                                                       Sequence<2>{},
+                                                       Sequence<3>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<4>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(YTilde, HTilde),
+                                                make_tuple(ConvDilationH, ConvStrideH)),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_ytilde),
+                           make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<>{},
+                           Sequence<2>{},
+                           Sequence<3>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_htildeslice_wtildeslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    std::vector<ck::index_t> tildes)
+    {
+        using namespace ck;
+
+        const index_t i_ztilde = tildes[0];
+        const index_t i_ytilde = tildes[1];
+        const index_t i_xtilde = tildes[2];
+
+        const index_t Di = input_spatial_lengths[0];
+        const index_t Hi = input_spatial_lengths[1];
+        const index_t Wi = input_spatial_lengths[2];
+
+        const index_t Do = output_spatial_lengths[0];
+        const index_t Ho = output_spatial_lengths[1];
+        const index_t Wo = output_spatial_lengths[2];
+
+        const index_t Z = filter_spatial_lengths[0];
+        const index_t Y = filter_spatial_lengths[1];
+        const index_t X = filter_spatial_lengths[2];
+
+        const index_t InLeftPadD = input_left_pads[0];
+        const index_t InLeftPadH = input_left_pads[1];
+        const index_t InLeftPadW = input_left_pads[2];
+
+        const index_t InRightPadD = input_right_pads[0];
+        const index_t InRightPadH = input_right_pads[1];
+        const index_t InRightPadW = input_right_pads[2];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        const index_t ConvDilationD = conv_filter_dilations[0];
+        const index_t ConvDilationH = conv_filter_dilations[1];
+        const index_t ConvDilationW = conv_filter_dilations[2];
+
+        const auto K0 = K / K1;
+
+        const auto out_n_do_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Do, Ho, Wo, K));
+        const auto wei_k_z_y_x_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Z, Y, X, C));
+        const auto in_n_di_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Do * Ho * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Do), make_tuple(I1, ConvStrideD)),
+                           make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{},
+                           Sequence<3>{},
+                           Sequence<5>{},
+                           Sequence<0, 2, 4, 6>{},
+                           Sequence<7>{}),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto ZDot = math::integer_divide_ceil(Z, ZTilde);
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto DTilde =
+                Do + math::integer_divide_ceil(ConvDilationD * (Z - I1), ConvStrideD);
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IDTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadD - ConvDilationD * (ZTilde - I1)), ConvStrideD);
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IDTildeSliceEnd = math::min(
+                DTilde, math::integer_divide_ceil(InLeftPadD + Di - I1, ConvStrideD) + I1);
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto DTildeSlice = IDTildeSliceEnd - IDTildeSliceBegin;
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_dop_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_do_ho_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Do, I0, I0),
+                           make_pad_transform(Ho, I0, I0),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_dop_hop_wop_k_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(N),
+                        make_embed_transform(make_tuple(ZDot, DTilde),
+                                             make_tuple(-ConvDilationD / GcdStrideDilationD, I1)),
+                        make_embed_transform(make_tuple(YDot, HTilde),
+                                             make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                        make_embed_transform(make_tuple(XDot, WTilde),
+                                             make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                        make_pass_through_transform(K)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+            const auto
+                out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
+                    transform_tensor_descriptor(
+                        out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_slice_transform(ZDot, I0, ZDotSlice),
+                                   make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                                   make_slice_transform(YDot, I0, YDotSlice),
+                                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                                   make_slice_transform(XDot, I0, XDotSlice),
+                                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                                   make_unmerge_transform(make_tuple(K0, K1))),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{},
+                                   Sequence<6>{},
+                                   Sequence<7>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{},
+                                   Sequence<6>{},
+                                   Sequence<7, 8>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(
+                    make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, K0)),
+                    make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                    make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}, Sequence<8>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc =
+                transform_tensor_descriptor(
+                    wei_k_z_y_x_c_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(K),
+                        make_embed_transform(make_tuple(ZDot, ZTilde),
+                                             make_tuple(ConvStrideD / GcdStrideDilationD, I1)),
+                        make_embed_transform(make_tuple(YDot, YTilde),
+                                             make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                        make_embed_transform(make_tuple(XDot, XTilde),
+                                             make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                        make_pass_through_transform(C)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+            const auto wei_k0_k1_zdotslice_ydotslice_xdotslice_c_grid_desc =
+                transform_tensor_descriptor(wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_slice_transform(ZDot, I0, ZDotSlice),
+                                                       make_slice_transform(YDot, I0, YDotSlice),
+                                                       make_slice_transform(XDot, I0, XDotSlice),
+                                                       make_freeze_transform(i_ztilde),
+                                                       make_freeze_transform(i_ytilde),
+                                                       make_freeze_transform(i_xtilde),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{},
+                                                       Sequence<1>{},
+                                                       Sequence<3>{},
+                                                       Sequence<5>{},
+                                                       Sequence<2>{},
+                                                       Sequence<4>{},
+                                                       Sequence<6>{},
+                                                       Sequence<7>{}),
+                                            make_tuple(Sequence<0, 1>{},
+                                                       Sequence<2>{},
+                                                       Sequence<3>{},
+                                                       Sequence<4>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<5>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_zdotslice_ydotslice_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 3, 4, 0>{}, Sequence<5>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc =
+                transform_tensor_descriptor(
+                    in_n_dip_hip_wip_c_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_embed_transform(make_tuple(ZTilde, DTilde),
+                                                    make_tuple(ConvDilationD, ConvStrideD)),
+                               make_embed_transform(make_tuple(YTilde, HTilde),
+                                                    make_tuple(ConvDilationH, ConvStrideH)),
+                               make_embed_transform(make_tuple(XTilde, WTilde),
+                                                    make_tuple(ConvDilationW, ConvStrideW)),
+                               make_pass_through_transform(C)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+            const auto in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc =
+                transform_tensor_descriptor(
+                    in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_freeze_transform(i_ztilde),
+                               make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                               make_freeze_transform(i_ytilde),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_freeze_transform(i_xtilde),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{},
+                               Sequence<6>{},
+                               Sequence<7>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<>{},
+                               Sequence<1>{},
+                               Sequence<>{},
+                               Sequence<2>{},
+                               Sequence<>{},
+                               Sequence<3>{},
+                               Sequence<4>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc,
+                make_tuple(
+                    make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<1>(
+            1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1}, {0});
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<2>(
+            1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {0, 0});
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<3>(1,
+                                                                  1,
+                                                                  1,
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {0, 0, 0});
+    }
+
+    using ABCGridDescs = decltype(GetABCGridDesc<NDimSpatial>());
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXdl,
+        NPerXdl,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
+        7,                                // CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_in_grid},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{out_element_op},
+              b_element_op_{wei_element_op},
+              c_element_op_{in_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              input_spatial_lengths_{input_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              output_spatial_lengths_{output_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            CreateABCDesc<NDimSpatial>();
+        }
+
+        template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+        void CreateABCDesc()
+        {
+            const index_t ConvStrideW     = conv_filter_strides_[0];
+            const index_t ConvDilationW   = conv_filter_dilations_[0];
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+            const auto XTilde             = ConvStrideW / GcdStrideDilationW;
+
+            const index_t X = filter_spatial_lengths_[0];
+
+            for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+            {
+                // check slice is valid
+                const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                if(XDotSlice <= 0)
+                {
+                    continue;
+                }
+
+                const auto descs =
+                    DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                        Conv_N_,
+                        Conv_K_,
+                        Conv_C_,
+                        input_spatial_lengths_,
+                        filter_spatial_lengths_,
+                        output_spatial_lengths_,
+                        conv_filter_strides_,
+                        conv_filter_dilations_,
+                        input_left_pads_,
+                        input_right_pads_,
+                        {i_xtilde});
+                a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                c_grid_desc_m_n_container_.push_back(descs[I2]);
+
+                auto block_2_ctile_map =
+                    GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2], M01_, N01_);
+
+                if(GridwiseGemm::CheckValidity(descs[I0], descs[I1], descs[I2], block_2_ctile_map))
+                {
+                    c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_.push_back(
+                        GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(descs[I2]));
+
+                    block_2_ctile_map_container_.push_back(block_2_ctile_map);
+                }
+            }
+        }
+        template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+        void CreateABCDesc()
+        {
+            const index_t ConvStrideH = conv_filter_strides_[0];
+            const index_t ConvStrideW = conv_filter_strides_[1];
+
+            const index_t ConvDilationH = conv_filter_dilations_[0];
+            const index_t ConvDilationW = conv_filter_dilations_[1];
+
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const index_t Y = filter_spatial_lengths_[0];
+            const index_t X = filter_spatial_lengths_[1];
+            for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+            {
+                for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+                {
+                    // check slice is valid
+                    const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                    const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                    if(YDotSlice * XDotSlice <= 0)
+                    {
+                        continue;
+                    }
+
+                    const auto descs =
+                        DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                            Conv_N_,
+                            Conv_K_,
+                            Conv_C_,
+                            input_spatial_lengths_,
+                            filter_spatial_lengths_,
+                            output_spatial_lengths_,
+                            conv_filter_strides_,
+                            conv_filter_dilations_,
+                            input_left_pads_,
+                            input_right_pads_,
+                            {i_ytilde, i_xtilde});
+                    a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                    b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                    c_grid_desc_m_n_container_.push_back(descs[I2]);
+
+                    auto block_2_ctile_map =
+                        GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2], M01_, N01_);
+
+                    if(GridwiseGemm::CheckValidity(
+                           descs[I0], descs[I1], descs[I2], block_2_ctile_map))
+                    {
+                        c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_.push_back(
+                            GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(descs[I2]));
+
+                        block_2_ctile_map_container_.push_back(block_2_ctile_map);
+                    }
+                }
+            }
+        }
+        template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+        void CreateABCDesc()
+        {
+            const index_t ConvStrideD = conv_filter_strides_[0];
+            const index_t ConvStrideH = conv_filter_strides_[1];
+            const index_t ConvStrideW = conv_filter_strides_[2];
+
+            const index_t ConvDilationD = conv_filter_dilations_[0];
+            const index_t ConvDilationH = conv_filter_dilations_[1];
+            const index_t ConvDilationW = conv_filter_dilations_[2];
+
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const index_t Z = filter_spatial_lengths_[0];
+            const index_t Y = filter_spatial_lengths_[1];
+            const index_t X = filter_spatial_lengths_[2];
+            for(index_t i_ztilde = 0; i_ztilde < ZTilde; ++i_ztilde)
+            {
+                for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+                {
+                    for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+                    {
+                        // check slice is valid
+                        const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
+                        const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                        const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                        if(ZDotSlice * YDotSlice * XDotSlice <= 0)
+                        {
+                            continue;
+                        }
+
+                        const auto descs =
+                            DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                                Conv_N_,
+                                Conv_K_,
+                                Conv_C_,
+                                input_spatial_lengths_,
+                                filter_spatial_lengths_,
+                                output_spatial_lengths_,
+                                conv_filter_strides_,
+                                conv_filter_dilations_,
+                                input_left_pads_,
+                                input_right_pads_,
+                                {i_ztilde, i_ytilde, i_xtilde});
+                        a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                        b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                        c_grid_desc_m_n_container_.push_back(descs[I2]);
+
+                        auto block_2_ctile_map =
+                            GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2], M01_, N01_);
+
+                        if(GridwiseGemm::CheckValidity(
+                               descs[I0], descs[I1], descs[I2], block_2_ctile_map))
+                        {
+                            c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_.push_back(
+                                GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(
+                                    descs[I2]));
+
+                            block_2_ctile_map_container_.push_back(block_2_ctile_map);
+                        }
+                    }
+                }
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        std::vector<AGridDesc_K0_M_K1> a_grid_desc_k0_m_k1_container_;
+        std::vector<BGridDesc_K0_N_K1> b_grid_desc_k0_n_k1_container_;
+        std::vector<CGridDesc_M_N> c_grid_desc_m_n_container_;
+        std::vector<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>
+            c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_;
+        std::vector<typename GridwiseGemm::DefaultBlock2CTileMap> block_2_ctile_map_container_;
+        index_t M01_;
+        index_t N01_;
+        OutElementwiseOperation a_element_op_;
+        WeiElementwiseOperation b_element_op_;
+        InElementwiseOperation c_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+
+        std::vector<ck::index_t> input_spatial_lengths_;
+        std::vector<ck::index_t> filter_spatial_lengths_;
+        std::vector<ck::index_t> output_spatial_lengths_;
+        std::vector<ck::index_t> conv_filter_strides_;
+        std::vector<ck::index_t> conv_filter_dilations_;
+        std::vector<ck::index_t> input_left_pads_;
+        std::vector<ck::index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float ave_time = 0;
+            for(size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+            {
+                {
+                    std::cout << "arg.a_grid_desc_k0_m_k1_container_{"
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) << ", "
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I1) << ", "
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.b_grid_desc_k0_n_k1_container_{"
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I0) << ", "
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I1) << ", "
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I2) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.c_grid_desc_m_n_container_{ "
+                              << arg.c_grid_desc_m_n_container_[i].GetLength(I0) << ", "
+                              << arg.c_grid_desc_m_n_container_[i].GetLength(I1) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I0)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I1)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I2)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I3)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I4)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I5)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I6)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I7)
+                              << " ) " << std::endl;
+                }
+
+                if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
+                                                arg.b_grid_desc_k0_n_k1_container_[i],
+                                                arg.c_grid_desc_m_n_container_[i],
+                                                arg.block_2_ctile_map_container_[i]))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
+                }
+
+                const index_t grid_size = arg.block_2_ctile_map_container_[i].CalculateGridSize(
+                    arg.c_grid_desc_m_n_container_[i]);
+
+                const auto K = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) *
+                               arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2);
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r3<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<
+                            typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                        OutElementwiseOperation,
+                        WeiElementwiseOperation,
+                        InElementwiseOperation,
+                        remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                        true>;
+
+                    ave_time += launch_and_time_kernel(
+                        stream_config,
+                        kernel,
+                        dim3(grid_size),
+                        dim3(BlockSize),
+                        0,
+                        arg.p_a_grid_,
+                        arg.p_b_grid_,
+                        arg.p_c_grid_,
+                        arg.a_grid_desc_k0_m_k1_container_[i],
+                        arg.b_grid_desc_k0_n_k1_container_[i],
+                        arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i],
+                        arg.a_element_op_,
+                        arg.b_element_op_,
+                        arg.c_element_op_,
+                        arg.block_2_ctile_map_container_[i]);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r3<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<
+                            typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                        OutElementwiseOperation,
+                        WeiElementwiseOperation,
+                        InElementwiseOperation,
+                        remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                        false>;
+
+                    ave_time += launch_and_time_kernel(
+                        stream_config,
+                        kernel,
+                        dim3(grid_size),
+                        dim3(BlockSize),
+                        0,
+                        arg.p_a_grid_,
+                        arg.p_b_grid_,
+                        arg.p_c_grid_,
+                        arg.a_grid_desc_k0_m_k1_container_[i],
+                        arg.b_grid_desc_k0_n_k1_container_[i],
+                        arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i],
+                        arg.a_element_op_,
+                        arg.b_element_op_,
+                        arg.c_element_op_,
+                        arg.block_2_ctile_map_container_[i]);
+                }
+            }
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 pad = 0 conv
+            for(int i = 0; i < NDimSpatial; i++)
+            {
+                if(!(arg.filter_spatial_lengths_[i] == 1 && arg.conv_filter_strides_[i] == 1 &&
+                     arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 1 &&
+             arg.Conv_K_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_C_ % CThreadTransferDstScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        for(std::size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
+                                            arg.b_grid_desc_k0_n_k1_container_[i],
+                                            arg.c_grid_desc_m_n_container_[i],
+                                            arg.block_2_ctile_map_container_[i]))
+            {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             const OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* p_in_grid,
+                        const void* p_wei_grid,
+                        const void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<const OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConvNdBwdDataNwcKxcNwk_Xdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock
+            << ">";
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0){
+
+            str<< " Filter1x1Stride1Pad0";
+        }
+
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise.hpp
new file mode 100644
index 0000000..8e62880
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise.hpp
@@ -0,0 +1,304 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise_base.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataTypeTuple,
+          typename OutDataTypeTuple,
+          typename ElementwiseOperation,
+          index_t NumDim,
+          index_t MPerThread,
+          typename InScalarPerVectorSeq,
+          typename OutScalarPerVectorSeq>
+struct DeviceElementwise
+    : public DeviceElementwiseBase<InDataTypeTuple, OutDataTypeTuple, ElementwiseOperation, NumDim>
+{
+    static constexpr int NumInput  = InDataTypeTuple::Size();
+    static constexpr int NumOutput = OutDataTypeTuple::Size();
+
+    static_assert(NumInput == InScalarPerVectorSeq::Size() &&
+                      NumOutput == OutScalarPerVectorSeq::Size(),
+                  "Tuple size is inconsistent with the number of in/out!");
+
+    static auto GenerateInDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+
+                return static_cast<const DataType*>(nullptr);
+            },
+            Number<NumInput>{});
+    };
+
+    static auto GenerateOutDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+
+                return static_cast<DataType*>(nullptr);
+            },
+            Number<NumOutput>{});
+    };
+
+    using InDataTypePointerTuple  = decltype(GenerateInDataTypePointerTuple());
+    using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
+
+    template <typename Desc_M>
+    static auto PadDescriptor_M_1d(Desc_M desc_m, index_t gridSize, index_t blockSize)
+    {
+        constexpr auto I0 = Number<0>{};
+
+        const auto m            = desc_m.GetLength(I0);
+        const index_t loop_step = gridSize * blockSize * MPerThread;
+        const auto pad          = math::integer_least_multiple(m, loop_step) - m;
+        const auto desc_m_pad =
+            transform_tensor_descriptor(desc_m,
+                                        make_tuple(make_right_pad_transform(m, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return desc_m_pad;
+    }
+
+    static auto MakeDescriptor_M(const std::array<index_t, NumDim>& lengths,
+                                 const std::array<index_t, NumDim>& stride,
+                                 index_t gridSize,
+                                 index_t blockSize)
+    {
+        auto tupleOfShape  = generate_tuple([&](auto I) { return lengths[I]; }, Number<NumDim>{});
+        auto tupleOfStride = generate_tuple([&](auto I) { return stride[I]; }, Number<NumDim>{});
+
+        // nd desc - [s0, s1, s2, ...]
+        const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
+
+        // merge nd to 1d desc - [s0 * s1 * ...]
+        if constexpr(NumDim > 1)
+        {
+            const auto desc_m = transform_tensor_descriptor(
+                desc,
+                make_tuple(make_merge_transform(tupleOfShape)),
+                make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<NumDim>{})),
+                make_tuple(Sequence<0>{}));
+
+            return PadDescriptor_M_1d(desc_m, gridSize, blockSize);
+        }
+        else
+            return PadDescriptor_M_1d(desc, gridSize, blockSize);
+    }
+
+    template <index_t TupleSize>
+    static auto GenerateInOutGrid1dDescTuple(Number<TupleSize>)
+    {
+        return generate_tuple(
+            [&](auto) {
+                if constexpr(NumDim > 1)
+                {
+                    return MakeDescriptor_M({1, 1}, {1, 1}, 1, 1);
+                }
+                else
+                {
+                    return MakeDescriptor_M({1}, {1}, 1, 1);
+                };
+            },
+            Number<TupleSize>{});
+    };
+
+    using InGrid1dDescTuple  = decltype(GenerateInOutGrid1dDescTuple(Number<NumInput>{}));
+    using OutGrid1dDescTuple = decltype(GenerateInOutGrid1dDescTuple(Number<NumOutput>{}));
+
+    using GridwiseElementwise = GridwiseElementwise_1D<InGrid1dDescTuple,
+                                                       OutGrid1dDescTuple,
+                                                       InDataTypePointerTuple,
+                                                       OutDataTypePointerTuple,
+                                                       ElementwiseOperation,
+                                                       MPerThread,
+                                                       InScalarPerVectorSeq,
+                                                       OutScalarPerVectorSeq>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, NumDim> lengths,
+                 const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                 const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                 const std::array<const void*, NumInput> in_dev_buffers,
+                 const std::array<void*, NumOutput> out_dev_buffers,
+                 ElementwiseOperation elementwise_op)
+
+            : lengths_(lengths),
+              inStridesArray_(inStridesArray),
+              outStridesArray_(outStridesArray),
+              elementwise_op_(elementwise_op),
+              blockSize_(256),
+              gridSize_(120) // FIXME - Calculate the grid size by number of CU in the future
+        {
+            in_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+                    return static_cast<const DataType*>(in_dev_buffers[I.value]);
+                },
+                Number<NumInput>{});
+
+            out_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+                    return static_cast<DataType*>(out_dev_buffers[I.value]);
+                },
+                Number<NumOutput>{});
+
+            in_grid_1d_desc_tuple_ = generate_tuple(
+                [&](auto I) {
+                    return MakeDescriptor_M(
+                        lengths, inStridesArray[I.value], gridSize_, blockSize_);
+                },
+                Number<NumInput>{});
+
+            out_grid_1d_desc_tuple_ = generate_tuple(
+                [&](auto I) {
+                    return MakeDescriptor_M(
+                        lengths, outStridesArray[I.value], gridSize_, blockSize_);
+                },
+                Number<NumOutput>{});
+        }
+
+        InDataTypePointerTuple in_dev_buffers_;
+        OutDataTypePointerTuple out_dev_buffers_;
+        InGrid1dDescTuple in_grid_1d_desc_tuple_;
+        OutGrid1dDescTuple out_grid_1d_desc_tuple_;
+
+        std::array<index_t, NumDim> lengths_;
+        std::array<std::array<index_t, NumDim>, NumInput> inStridesArray_;
+        std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray_;
+
+        ElementwiseOperation elementwise_op_;
+        index_t blockSize_;
+        index_t gridSize_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto kernel = kernel_elementwise_1d<GridwiseElementwise,
+                                                      InGrid1dDescTuple,
+                                                      OutGrid1dDescTuple,
+                                                      InDataTypePointerTuple,
+                                                      OutDataTypePointerTuple,
+                                                      ElementwiseOperation>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(arg.gridSize_),
+                                                        dim3(arg.blockSize_),
+                                                        0,
+                                                        arg.in_grid_1d_desc_tuple_,
+                                                        arg.out_grid_1d_desc_tuple_,
+                                                        arg.in_dev_buffers_,
+                                                        arg.out_dev_buffers_,
+                                                        arg.elementwise_op_);
+            return elapsed_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(arg.lengths_.back() % MPerThread != 0)
+            return false;
+
+        auto IsScalarPerVectorValid = [&](const std::array<index_t, NumDim>& lengths,
+                                          const std::array<index_t, NumDim>& strides,
+                                          index_t scalarPerVector) {
+            if(strides.back() == 1 && lengths.back() % scalarPerVector == 0)
+                return true;
+
+            if(strides.back() != 1 && scalarPerVector == 1)
+                return true;
+
+            return false;
+        };
+
+        bool valid = true;
+        static_for<0, NumInput, 1>{}([&](auto I) {
+            if(!IsScalarPerVectorValid(
+                   arg.lengths_, arg.inStridesArray_[I.value], InScalarPerVectorSeq::At(I)))
+                valid = false;
+        });
+
+        static_for<0, NumOutput, 1>{}([&](auto I) {
+            if(!IsScalarPerVectorValid(
+                   arg.lengths_, arg.outStridesArray_[I.value], OutScalarPerVectorSeq::At(I)))
+                valid = false;
+        });
+
+        return valid;
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto
+    MakeArgument(const std::array<index_t, NumDim> lengths,
+                 const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                 const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                 const std::array<const void*, NumInput> in_dev_buffers,
+                 const std::array<void*, NumOutput> out_dev_buffers,
+                 ElementwiseOperation elementwise_op)
+    {
+        return Argument{lengths,
+                        inStridesArray,
+                        outStridesArray,
+                        in_dev_buffers,
+                        out_dev_buffers,
+                        elementwise_op};
+    }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, NumDim> lengths,
+                        const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                        const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const std::array<void*, NumOutput> out_dev_buffers,
+                        ElementwiseOperation elementwise_op) override
+    {
+        return std::make_unique<Argument>(lengths,
+                                          inStridesArray,
+                                          outStridesArray,
+                                          in_dev_buffers,
+                                          out_dev_buffers,
+                                          elementwise_op);
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+}; // namespace device
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp
new file mode 100644
index 0000000..8ffc5ef
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp
@@ -0,0 +1,592 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+// X = Elementwise(input1, input2, input3, ...)
+// Y = Normalization(X, beta, gamma)
+namespace ck {
+template <typename GridwiseElementwiseReduction,
+          typename InDataTypePointerTuple, // Datatype tuple of inputs
+          typename XDataType,              // Datatype of X
+          typename GammaDataType,          // Datatype of Gamma
+          typename BetaDataType,           // Datatype of Beta
+          typename YDataType,              // Datatype of Y
+          typename AccDataType,            // AccDatatype
+          typename XElementwiseOperation,  // Operation of input
+          typename YElementwiseOperation,  // Operation of output of normalization
+          typename InGrid2dDescTuple,      // Descriptor tuple of inputs
+          typename GridDesc_M_K>           // Descriptor of inputs, Gamma, Beta
+__global__ void kernel_elementwise_layernorm(
+    const InGrid2dDescTuple in_grid_2d_desc_tuple,          // Descriptor tuple of inputs
+    const GridDesc_M_K x_grid_desc_m_k,                     // Descriptor of X
+    const GridDesc_M_K gamma_grid_desc_m_k,                 // Descriptor of gamma
+    const GridDesc_M_K beta_grid_desc_m_k,                  // Descriptor of beta
+    const GridDesc_M_K y_grid_desc_m_k,                     // Descriptor of Y
+    index_t num_k_block_tile_iteration,                     //
+    AccDataType epsilon,                                    // Datatype of epsilon
+    const InDataTypePointerTuple p_in_global_tuple,         // Ptr tuple of input matrixs
+    const GammaDataType* const __restrict__ p_gamma_global, // Ptr of gamma
+    const BetaDataType* const __restrict__ p_beta_global,   // Ptr of beta
+    YDataType* const __restrict__ p_y_global,               // Ptr of y
+    const XElementwiseOperation x_elementwise_op,           // Operation of input
+    const YElementwiseOperation y_elementwise_op)           // Operation of output of normalization
+{
+    extern __shared__ XDataType p_x_lds[];
+    GridwiseElementwiseReduction::Run(in_grid_2d_desc_tuple,      // Descriptor tuple of inputs
+                                      x_grid_desc_m_k,            // Descriptor of X
+                                      gamma_grid_desc_m_k,        // Descriptor of Gamma
+                                      beta_grid_desc_m_k,         // Descriptor of Beta
+                                      y_grid_desc_m_k,            // Descriptor of Y
+                                      num_k_block_tile_iteration, //
+                                      epsilon,                    // epsilon
+                                      p_in_global_tuple,          // Ptr tuple of inputs
+                                      p_x_lds,                    // Ptr of X
+                                      p_gamma_global,             // Ptr of gamma
+                                      p_beta_global,              // Ptr of beta
+                                      p_y_global,                 // Ptr of Y
+                                      x_elementwise_op,           // Operation of input
+                                      y_elementwise_op); // Operation of output of normalization
+};
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Y = LayerNorm(A + B, Beta, Gamma)
+template <typename InDataTypeTuple,       // Datatype of inputs
+          typename GammaDataType,         // Datatype of gamma
+          typename BetaDataType,          // Datatype of beta
+          typename AccDataType,           //
+          typename YDataType,             //
+          typename XElementwiseOperation, //
+          typename YElementwiseOperation, //
+          index_t Rank,                   //
+          index_t NumReduceDim,           //
+          index_t BlockSize,              //
+          index_t MThreadClusterSize,     // Num of threads in a block on M direction
+          index_t KThreadClusterSize,     // Num of threads in a block on N direction
+          index_t MThreadSliceSize,       // Each thread calculate rows
+          index_t KThreadSliceSize,       // Each thread calculate columns
+          index_t XYSrcVectorDim,         // Dimension to do reduce
+          index_t XSrcVectorSize,         // Size to fetch source x
+          index_t GammaSrcVectorDim,      // Dimension for gamma to do reduce
+          index_t GammaSrcVectorSize,     // Size to fetch source gamma
+          index_t BetaSrcVectorDim,       // Dimension for beta to do reduce
+          index_t BetaSrcVectorSize,      // Size to fetch source beta
+          index_t YDstVectorSize>         // Size to write destination Y
+struct DeviceElementwiseNormalizationImpl
+    : public DeviceElementwiseNormalization<InDataTypeTuple,
+                                            GammaDataType,
+                                            BetaDataType,
+                                            AccDataType,
+                                            YDataType,
+                                            XElementwiseOperation,
+                                            YElementwiseOperation,
+                                            Rank,
+                                            NumReduceDim>
+{
+    static constexpr int NumInput = InDataTypeTuple::Size();
+
+    using XDataType = YDataType;
+
+    static_assert(
+        (KThreadSliceSize % GammaSrcVectorSize == 0),
+        "Invalid thread slice sizes and/or gamma vector sizes configuration, please check!");
+
+    static_assert(
+        (KThreadSliceSize % BetaSrcVectorSize == 0),
+        "Invalid thread slice sizes and/or beta vector sizes configuration, please check!");
+
+    static constexpr index_t M_BlockTileSize =
+        MThreadClusterSize * MThreadSliceSize; // num of rows calculated in a block
+    static constexpr index_t K_BlockTileSize =
+        KThreadClusterSize * KThreadSliceSize; // num of columns calculated in a block
+
+    static auto GenerateInDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+                return static_cast<const DataType*>(nullptr);
+            },
+            Number<NumInput>{});
+    };
+
+    using InDataTypePointerTuple = decltype(GenerateInDataTypePointerTuple());
+
+    static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
+                                    const std::vector<index_t>& inStrides,
+                                    int blkGroupSize,
+                                    int numBlockTileIteration)
+    {
+        constexpr index_t NumInvariantDim  = Rank - NumReduceDim;
+        static constexpr index_t numSrcDim = Rank;
+        static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+        const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
+        const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, numSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, ReduceDims{});
+                const auto invariantDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, InvariantDims{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    template <index_t TupleSize>
+    static auto GenerateSrcGrid2dDescTuple(Number<TupleSize>)
+    {
+        return generate_tuple([&](auto) { return MakeSrc2dDescriptor({1}, {1}, 1, 1); },
+                              Number<TupleSize>{});
+    };
+
+    using InGrid2dDescTuple = decltype(GenerateSrcGrid2dDescTuple(Number<NumInput>{}));
+
+    using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
+
+    using GridwiseReduceLayernormGeneric =
+        GridwiseElementwiseLayernormWelfordVariance_mk_to_mk<InDataTypePointerTuple,
+                                                             XDataType,
+                                                             GammaDataType,
+                                                             BetaDataType,
+                                                             YDataType,
+                                                             AccDataType,
+                                                             XElementwiseOperation,
+                                                             YElementwiseOperation,
+                                                             InGrid2dDescTuple,
+                                                             GridDesc_M_K,
+                                                             BlockSize,
+                                                             MThreadClusterSize,
+                                                             KThreadClusterSize,
+                                                             MThreadSliceSize,
+                                                             KThreadSliceSize,
+                                                             XYSrcVectorDim,
+                                                             XSrcVectorSize,
+                                                             GammaSrcVectorDim,
+                                                             GammaSrcVectorSize,
+                                                             BetaSrcVectorDim,
+                                                             BetaSrcVectorSize,
+                                                             XYSrcVectorDim,
+                                                             YDstVectorSize,
+                                                             false>;
+
+    using GridwiseReduceLayernormSweepOnce =
+        GridwiseElementwiseLayernormWelfordVariance_mk_to_mk<InDataTypePointerTuple,
+                                                             XDataType,
+                                                             GammaDataType,
+                                                             BetaDataType,
+                                                             YDataType,
+                                                             AccDataType,
+                                                             XElementwiseOperation,
+                                                             YElementwiseOperation,
+                                                             InGrid2dDescTuple,
+                                                             GridDesc_M_K,
+                                                             BlockSize,
+                                                             MThreadClusterSize,
+                                                             KThreadClusterSize,
+                                                             MThreadSliceSize,
+                                                             KThreadSliceSize,
+                                                             XYSrcVectorDim,
+                                                             XSrcVectorSize,
+                                                             GammaSrcVectorDim,
+                                                             GammaSrcVectorSize,
+                                                             BetaSrcVectorDim,
+                                                             BetaSrcVectorSize,
+                                                             XYSrcVectorDim,
+                                                             YDstVectorSize,
+                                                             true>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::vector<index_t> lengths,
+                 const std::array<std::vector<index_t>, NumInput> inStridesArray,
+                 const std::vector<index_t> gammaStrides,
+                 const std::vector<index_t> betaStrides,
+                 const std::vector<index_t> yStrides,
+                 const std::vector<index_t> reduceDims,
+                 XElementwiseOperation x_elementwise_op,
+                 YElementwiseOperation y_elementwise_op,
+                 AccDataType epsilon,
+                 const std::array<const void*, NumInput> in_dev_buffers,
+                 const GammaDataType* p_gamma,
+                 const BetaDataType* p_beta,
+                 YDataType* p_y)
+            : epsilon_(epsilon),
+              p_gamma_(p_gamma),
+              p_beta_(p_beta),
+              p_y_(p_y),
+              x_elementwise_op_(x_elementwise_op),
+              y_elementwise_op_(y_elementwise_op)
+        {
+
+            Lengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(lengths, reduceDims);
+            for(int i = 0; i < NumInput; i++)
+            {
+                inStridesArray_[i] =
+                    shuffle_tensor_dimensions<Rank, NumReduceDim>(inStridesArray[i], reduceDims);
+            }
+
+            yStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
+            xStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
+
+            gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims);
+            betaStrides_  = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims);
+
+            in_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+                    return static_cast<const DataType*>(in_dev_buffers[I.value]);
+                },
+                Number<NumInput>{});
+
+            long_index_t invariant_total_length;
+            long_index_t reduce_total_length;
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(Lengths_);
+
+            blkGroupSize_          = 1;
+            numBlockTileIteration_ = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+
+            gridSize_ = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                        M_BlockTileSize * blkGroupSize_;
+
+            in_grid_2d_desc_tuple_ = generate_tuple(
+                [&](auto I) {
+                    return MakeSrc2dDescriptor(
+                        Lengths_, inStridesArray_[I.value], blkGroupSize_, numBlockTileIteration_);
+                },
+                Number<NumInput>{});
+
+            x_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, xStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            gamma_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, gammaStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            beta_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, betaStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            y_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, yStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            sweep_once_ =
+                x_grid_desc_m_k_.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
+
+            if(!sweep_once_) // if not sweep once, compute memory size for matrix X in lds for
+                             // store Intermediate results
+            {
+                int block_TileSize = M_BlockTileSize * reduce_total_length;
+                x_lds_size_        = block_TileSize * sizeof(XDataType);
+            }
+            else
+                x_lds_size_ = 0;
+        }
+
+        AccDataType epsilon_;
+
+        InDataTypePointerTuple in_dev_buffers_;
+        const GammaDataType* p_gamma_;
+        const BetaDataType* p_beta_;
+        YDataType* p_y_;
+
+        std::vector<index_t> Lengths_;
+        std::array<std::vector<index_t>, NumInput> inStridesArray_;
+        std::vector<index_t> xStrides_;
+        std::vector<index_t> gammaStrides_;
+        std::vector<index_t> betaStrides_;
+        std::vector<index_t> yStrides_;
+
+        XElementwiseOperation x_elementwise_op_;
+        YElementwiseOperation y_elementwise_op_;
+
+        int blkGroupSize_;
+        int numBlockTileIteration_;
+        size_t gridSize_;
+
+        InGrid2dDescTuple in_grid_2d_desc_tuple_;
+        GridDesc_M_K x_grid_desc_m_k_;
+        GridDesc_M_K gamma_grid_desc_m_k_;
+        GridDesc_M_K beta_grid_desc_m_k_;
+        GridDesc_M_K y_grid_desc_m_k_;
+        bool sweep_once_;
+        int x_lds_size_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto kernel_main =
+                arg.sweep_once_ ? kernel_elementwise_layernorm<GridwiseReduceLayernormSweepOnce,
+                                                               InDataTypePointerTuple,
+                                                               XDataType,
+                                                               GammaDataType,
+                                                               BetaDataType,
+                                                               YDataType,
+                                                               AccDataType,
+                                                               XElementwiseOperation,
+                                                               YElementwiseOperation,
+                                                               InGrid2dDescTuple,
+                                                               GridDesc_M_K>
+                                : kernel_elementwise_layernorm<GridwiseReduceLayernormGeneric,
+                                                               InDataTypePointerTuple,
+                                                               XDataType,
+                                                               GammaDataType,
+                                                               BetaDataType,
+                                                               YDataType,
+                                                               AccDataType,
+                                                               XElementwiseOperation,
+                                                               YElementwiseOperation,
+                                                               InGrid2dDescTuple,
+                                                               GridDesc_M_K>;
+
+            float avg_time = 0;
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize_),
+                                               dim3(BlockSize),
+                                               arg.x_lds_size_,
+                                               arg.in_grid_2d_desc_tuple_,
+                                               arg.x_grid_desc_m_k_,
+                                               arg.gamma_grid_desc_m_k_,
+                                               arg.beta_grid_desc_m_k_,
+                                               arg.y_grid_desc_m_k_,
+                                               arg.numBlockTileIteration_,
+                                               arg.epsilon_,
+                                               arg.in_dev_buffers_,
+                                               arg.p_gamma_,
+                                               arg.p_beta_,
+                                               arg.p_y_,
+                                               arg.x_elementwise_op_,
+                                               arg.y_elementwise_op_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
+
+        constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+        if constexpr(XYSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return false;
+            }
+            else
+            {
+                for(int i = 0; i < NumInput; i++)
+                {
+                    if(p_arg_->inStridesArray_[i][NumInvariantDim - 1] != 1)
+                        return false;
+                }
+
+                if(p_arg_->inStridesArray_[0][NumInvariantDim - 1] != 1 &&
+                   p_arg_->inStridesArray_[1][NumInvariantDim - 1] != 1)
+                    return false;
+
+                if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
+                    return false;
+            };
+        }
+        else
+        {
+            for(int i = 0; i < NumInput; i++)
+            {
+                if(p_arg_->inStridesArray_[i][Rank - 1] != 1)
+                    return false;
+            }
+
+            if(p_arg_->Lengths_[Rank - 1] % XSrcVectorSize != 0)
+                return false;
+        };
+
+        if(p_arg_->Lengths_[Rank - 1] % YDstVectorSize != 0)
+        {
+            return false;
+        }
+
+        auto IsScalarPerVectorValid = [](bool isLastDimensionCoalesced, int scalarPerVector) {
+            bool ret = true;
+
+            if(!isLastDimensionCoalesced)
+                ret = scalarPerVector == 1;
+            else
+                ret = KThreadSliceSize % scalarPerVector == 0;
+
+            return ret;
+        };
+
+        if(!IsScalarPerVectorValid(p_arg_->gammaStrides_.back() == 1, GammaSrcVectorSize))
+            return false;
+
+        if(!IsScalarPerVectorValid(p_arg_->betaStrides_.back() == 1, BetaSrcVectorSize))
+            return false;
+
+        // if fastest dim is not reduced
+        if constexpr(XYSrcVectorDim == 0) //
+        {
+            if(p_arg_->gammaStrides_[NumInvariantDim - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return (false);
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->gammaStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return (false);
+        }
+
+        // if fastest dim is not reduced
+        if constexpr(XYSrcVectorDim == 0)
+        {
+            if(p_arg_->betaStrides_[NumInvariantDim - 1] != 1)
+                return (false);
+
+            if(p_arg_->invariant_lowest_length % BetaSrcVectorSize != 0)
+                return (false);
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->betaStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % BetaSrcVectorSize != 0)
+                return (false);
+        }
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::array<std::vector<index_t>, NumInput> inStridesArray,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        AccDataType epsilon,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        XElementwiseOperation x_elementwise_op,
+                        YElementwiseOperation y_elementwise_op) override
+    {
+        return std::make_unique<Argument>(lengths,
+                                          inStridesArray,
+                                          gammaStrides,
+                                          betaStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          x_elementwise_op,
+                                          y_elementwise_op,
+                                          epsilon,
+                                          in_dev_buffers,
+                                          static_cast<const GammaDataType*>(p_gamma),
+                                          static_cast<const BetaDataType*>(p_beta),
+                                          static_cast<YDataType*>(p_y));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceElementwiseNormalizationImpl<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "XYSrcVectorDim_" << XYSrcVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_Gamma" << GammaSrcVectorSize << "_Beta" << BetaSrcVectorSize << "_Y" << YDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp
new file mode 100644
index 0000000..b9a64e8
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp
@@ -0,0 +1,875 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
+// version currently has compiler issues with register spill which further causes validation
+// failures.
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename BiasDataType,
+          typename D0DataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename ReduceAccDataType,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename D0ElementwiseOperation,
+          typename ReduceOperations,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          typename ReduceGlobalMemoryDataOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+          index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGemmBiasAddReduce_Xdl_CShuffle : public DeviceGemmReduce<1, ReduceOperations::Size()>
+{
+    using DeviceOp = DeviceGemmBiasAddReduce_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(M)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_right_pad_transform(MRaw, MPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(N)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_right_pad_transform(NRaw, NPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    // assume D is packed tensor
+    static auto MakeReduceGridDescriptor_M(index_t MRaw)
+    {
+        const auto d_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(MRaw));
+
+        const auto M    = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto MPad = M - MRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M
+            return transform_tensor_descriptor(d_grid_desc_mraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad)),
+                                               make_tuple(Sequence<0>{}),
+                                               make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad M
+            return d_grid_desc_mraw;
+        }
+    }
+
+    using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N       = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+    using C0GridDesc_M_N      = decltype(MakeCGridDescriptor_M_N(1, 1, 0));
+    using C1GridDesc_M_N      = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+    using ReduceGridDesc_M    = decltype(MakeReduceGridDescriptor_M(1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmBiasAddReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        BiasDataType,
+        D0DataType,
+        ReduceAccDataType,
+        ReducePtrsGlobal,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        D0ElementwiseOperation,
+        ReduceOperations,
+        ReduceInElementwiseOperations,
+        ReduceAccElementwiseOperations,
+        InMemoryDataOperationEnum::Set,
+        ReduceGlobalMemoryDataOperation,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        C0GridDesc_M_N,
+        C1GridDesc_M_N,
+        ReduceGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+        CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+        CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+        LoopSched>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 const BiasDataType* p_bias_grid,
+                 const D0DataType* p_d0_grid,
+                 ReducePtrsGlobal p_reduces_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 index_t StrideC1,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op,
+                 D0ElementwiseOperation d0_element_op,
+                 ReduceInElementwiseOperations reduce_in_element_ops,
+                 ReduceAccElementwiseOperations reduce_out_element_ops)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              p_bias_grid_{p_bias_grid},
+              p_d0_grid_{p_d0_grid},
+              p_reduces_grid_{p_reduces_grid},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC)},
+              c0_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, 0)},
+              c1_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC1)},
+              reduce_grid_desc_m_{DeviceOp::MakeReduceGridDescriptor_M(MRaw)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              c0_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              c1_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              reduce_grid_desc_mblock_mperblock_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              d0_element_op_{d0_element_op},
+              reduce_in_element_ops_{reduce_in_element_ops},
+              reduce_out_element_ops_{reduce_out_element_ops}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+
+                c0_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c0_grid_desc_m_n_);
+
+                c1_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c1_grid_desc_m_n_);
+
+                reduce_grid_desc_mblock_mperblock_ =
+                    GridwiseGemm::MakeReduceGridDescriptor_MBlock_MPerBlock(reduce_grid_desc_m_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        const BiasDataType* p_bias_grid_;
+        const D0DataType* p_d0_grid_;
+        ReducePtrsGlobal p_reduces_grid_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        C0GridDesc_M_N c0_grid_desc_m_n_;
+        C1GridDesc_M_N c1_grid_desc_m_n_;
+        ReduceGridDesc_M reduce_grid_desc_m_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c0_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c1_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock
+            reduce_grid_desc_mblock_mperblock_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+        D0ElementwiseOperation d0_element_op_;
+        ReduceInElementwiseOperations reduce_in_element_ops_;
+        ReduceAccElementwiseOperations reduce_out_element_ops_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                            arg.b_grid_desc_bk0_n_bk1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float elapsed_time = 0.0f;
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_bias_add_reduce_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    BiasDataType,
+                    D0DataType,
+                    ReducePtrsGlobal,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    D0ElementwiseOperation,
+                    ReduceInElementwiseOperations,
+                    ReduceAccElementwiseOperations,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    true>;
+
+                elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_bias_grid_,
+                                           arg.p_d0_grid_,
+                                           arg.p_reduces_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.d0_element_op_,
+                                           arg.reduce_in_element_ops_,
+                                           arg.reduce_out_element_ops_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.c0_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.c1_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.reduce_grid_desc_mblock_mperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_bias_add_reduce_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    BiasDataType,
+                    D0DataType,
+                    ReducePtrsGlobal,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    D0ElementwiseOperation,
+                    ReduceInElementwiseOperations,
+                    ReduceAccElementwiseOperations,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    false>;
+
+                elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_bias_grid_,
+                                           arg.p_d0_grid_,
+                                           arg.p_reduces_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.d0_element_op_,
+                                           arg.reduce_in_element_ops_,
+                                           arg.reduce_out_element_ops_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.c0_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.c1_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.reduce_grid_desc_mblock_mperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+
+            return elapsed_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static constexpr int NumReduce = ReduceOperations::Size();
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             const void* p_bias,
+                             std::array<const void*, 1> p_ds,
+                             void* p_c,
+                             std::array<void*, NumReduce> p_reduces,
+                             ck::index_t M,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t StrideA,
+                             ck::index_t StrideB,
+                             ck::index_t StrideC,
+                             std::array<ck::index_t, 1> StrideDs,
+                             std::array<void*, 3> gemm_element_ops,
+                             std::array<void*, 1> d_element_ops,
+                             std::array<void*, NumReduce> reduce_in_element_op,
+                             std::array<void*, NumReduce> reduce_out_element_op)
+    {
+        ReducePtrsGlobal reduce_tuple = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReducePtrsGlobal{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return static_cast<T*>(p_reduces[I]);
+            },
+            Number<NumReduce>{});
+
+        ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceInElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_in_element_op[I]));
+            },
+            Number<NumReduce>{});
+        ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceAccElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_out_element_op[I]));
+            },
+            Number<NumReduce>{});
+
+        AElementwiseOperation a_element_op =
+            *(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
+        BElementwiseOperation b_element_op =
+            *(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
+        CElementwiseOperation c_element_op =
+            *(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
+        D0ElementwiseOperation d_element_op =
+            *(static_cast<D0ElementwiseOperation*>(d_element_ops[0]));
+
+        return Argument{static_cast<const ADataType*>(p_a),
+                        static_cast<const BDataType*>(p_b),
+                        static_cast<CDataType*>(p_c),
+                        static_cast<const BiasDataType*>(p_bias),
+                        static_cast<const D0DataType*>(p_ds[0]),
+                        reduce_tuple,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        StrideDs[0],
+                        a_element_op,
+                        b_element_op,
+                        c_element_op,
+                        d_element_op,
+                        reduce_in_element_ops,
+                        reduce_out_element_ops};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const void* p_bias,
+                        std::array<const void*, 1> p_ds,
+                        void* p_c,
+                        std::array<void*, NumReduce> p_reduces,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        std::array<ck::index_t, 1> StrideDs,
+                        std::array<void*, 3> gemm_element_ops,
+                        std::array<void*, 1> d_element_ops,
+                        std::array<void*, NumReduce> reduce_in_element_op,
+                        std::array<void*, NumReduce> reduce_out_element_op,
+                        index_t /* KBatch */ = 1) override
+    {
+        ReducePtrsGlobal reduce_tuple = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReducePtrsGlobal{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return static_cast<T*>(p_reduces[I]);
+            },
+            Number<NumReduce>{});
+
+        ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceInElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_in_element_op[I]));
+            },
+            Number<NumReduce>{});
+        ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceAccElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_out_element_op[I]));
+            },
+            Number<NumReduce>{});
+
+        AElementwiseOperation a_element_op =
+            *(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
+        BElementwiseOperation b_element_op =
+            *(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
+        CElementwiseOperation c_element_op =
+            *(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
+        D0ElementwiseOperation d_element_op =
+            *(static_cast<D0ElementwiseOperation*>(d_element_ops[0]));
+
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          static_cast<const BiasDataType*>(p_bias),
+                                          static_cast<const D0DataType*>(p_ds[0]),
+                                          reduce_tuple,
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          StrideDs[0],
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op,
+                                          d_element_op,
+                                          reduce_in_element_ops,
+                                          reduce_out_element_ops);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmBiasAddReduce_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_e_permute_xdl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_e_permute_xdl.hpp
new file mode 100644
index 0000000..1914068
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_e_permute_xdl.hpp
@@ -0,0 +1,572 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatDsPointer,
+          typename FloatE,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_bias_e_permute(const FloatAB* __restrict__ p_a_grid,
+                                   const FloatAB* __restrict__ p_b_grid,
+                                   FloatDsPointer p_ds_grid,
+                                   FloatE* __restrict__ p_e_grid,
+                                   const AElementwiseOperation a_element_op,
+                                   const BElementwiseOperation b_element_op,
+                                   const CDEElementwiseOperation cde_element_op,
+                                   const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+                                   const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+                                   const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                                       ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                   const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                                       e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                   const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_ds_grid,
+                                                  p_e_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// input : A[M, K], or A[K, N]
+// input : B[K, N], or A[N, K]
+// input : D0[M, N], D1[M, N], ...
+// output : E[M, N]
+// C = a_op(A) * b_op(B)
+// E = cde_op(C, D0, D1, ...)
+template <typename ALayout,
+          typename BLayout,
+          typename CDELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGemmBiasEPermute_Xdl : public DeviceGemmBiasCPermute<AElementwiseOperation,
+                                                                  BElementwiseOperation,
+                                                                  CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGemmBiasEPermute_Xdl;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static constexpr index_t NumDTensor = 1;
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    static auto MakeEGridDescriptor_M_N(DEGridDesc_M0_M1_M2_N0_N1 d_e_grid_desc)
+    {
+        index_t M0 = d_e_grid_desc.M0_;
+        index_t M1 = d_e_grid_desc.M1_;
+        index_t M2 = d_e_grid_desc.M2_;
+        index_t N0 = d_e_grid_desc.N0_;
+        index_t N1 = d_e_grid_desc.N1_;
+
+        index_t stride_M0 = d_e_grid_desc.stride_M0_;
+        index_t stride_M1 = d_e_grid_desc.stride_M1_;
+        index_t stride_M2 = d_e_grid_desc.stride_M2_;
+        index_t stride_N0 = d_e_grid_desc.stride_N0_;
+        index_t stride_N1 = d_e_grid_desc.stride_N1_;
+
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            const auto e_grid_desc_m0_m1_m2_n0_n1 = make_naive_tensor_descriptor(
+                make_tuple(M0, M1, M2, N0, N1),
+                make_tuple(stride_M0, stride_M1, stride_M2, stride_N0, stride_N1));
+
+            return transform_tensor_descriptor(
+                e_grid_desc_m0_m1_m2_n0_n1,
+                make_tuple(make_merge_transform(make_tuple(M0, M1, M2)),
+                           make_merge_transform(make_tuple(N0, N1))),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    using AGridDesc_M_K = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N(DEGridDesc_M0_M1_M2_N0_N1{}));
+
+    using DsGridDesc_M_N = Tuple<EGridDesc_M_N>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<DDataType>,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+
+    using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 const void* p_d_grid,
+                 void* p_e_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 DEGridDesc_M0_M1_M2_N0_N1 d_grid_desc,
+                 DEGridDesc_M0_M1_M2_N0_N1 e_grid_desc,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N(e_grid_desc)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+
+            if(MRaw != d_grid_desc.M0_ * d_grid_desc.M1_ * d_grid_desc.M2_)
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            if(NRaw != d_grid_desc.N0_ * d_grid_desc.N1_)
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            // populate pointer, desc for Ds
+            // D pointer
+            p_ds_grid_(I0) = static_cast<const DDataType*>(p_d_grid);
+
+            // D desc
+            ds_grid_desc_m_n_(I0) = DeviceOp::MakeEGridDescriptor_M_N(d_grid_desc);
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_(I0) =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_[I0]);
+            }
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        typename GridwiseGemm::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_gemm_bias_e_permute<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.ds_grid_desc_m_n_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.block_2_etile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             const void* p_d,
+                             void* p_e,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t StrideA,
+                             index_t StrideB,
+                             DEGridDesc_M0_M1_M2_N0_N1 d_grid_desc,
+                             DEGridDesc_M0_M1_M2_N0_N1 e_grid_desc,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_d,
+                        p_e,
+                        MRaw,
+                        NRaw,
+                        KRaw,
+                        StrideA,
+                        StrideB,
+                        d_grid_desc,
+                        e_grid_desc,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const void* p_d,
+                        void* p_e,
+                        index_t MRaw,
+                        index_t NRaw,
+                        index_t KRaw,
+                        index_t StrideA,
+                        index_t StrideB,
+                        DEGridDesc_M0_M1_M2_N0_N1 d_grid_desc,
+                        DEGridDesc_M0_M1_M2_N0_N1 e_grid_desc,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_d,
+                                          p_e,
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          StrideA,
+                                          StrideB,
+                                          d_grid_desc,
+                                          e_grid_desc,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmBiasEPermute_Xdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp
new file mode 100644
index 0000000..b08a418
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp
@@ -0,0 +1,595 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <
+    typename ADataType,
+    typename BDataType,
+    typename CDataType,
+    typename AccDataType,
+    typename ALayout,
+    typename BLayout,
+    typename CLayout,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    GemmSpecialization GemmSpec,
+    index_t BlockSize,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t K0PerBlock,
+    index_t K1,
+    index_t M1PerThread,
+    index_t N1PerThread,
+    index_t KPerThread,
+    typename M1N1ThreadClusterM1Xs,
+    typename M1N1ThreadClusterN1Xs,
+    typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+    typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+    typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+    typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+    typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+    typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+    typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+    typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+    typename CThreadTransferSrcDstAccessOrder,
+    index_t CThreadTransferSrcDstVectorDim,
+    index_t CThreadTransferDstScalarPerVector,
+    enable_if_t<
+        is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+        bool> = false>
+struct DeviceGemmDl : public DeviceGemm<ALayout,
+                                        BLayout,
+                                        CLayout,
+                                        ADataType,
+                                        BDataType,
+                                        CDataType,
+                                        AElementwiseOperation,
+                                        BElementwiseOperation,
+                                        CElementwiseOperation>
+
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using CGridDesc_M_N     = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDl_km_kn_mn_v1r3<BlockSize,
+                                     ADataType,
+                                     AccDataType,
+                                     CDataType,
+                                     InMemoryDataOperationEnum::Set,
+                                     AGridDesc_K0_M_K1,
+                                     BGridDesc_K0_N_K1,
+                                     CGridDesc_M_N,
+                                     MPerBlock,
+                                     NPerBlock,
+                                     K0PerBlock,
+                                     K1,
+                                     M1PerThread,
+                                     N1PerThread,
+                                     KPerThread,
+                                     M1N1ThreadClusterM1Xs,
+                                     M1N1ThreadClusterN1Xs,
+                                     ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterArrangeOrder,
+                                     ABlockTransferSrcAccessOrder,
+                                     ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                     ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                     ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                     BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterArrangeOrder,
+                                     BBlockTransferSrcAccessOrder,
+                                     BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                     BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                     BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                     CThreadTransferSrcDstAccessOrder,
+                                     CThreadTransferSrcDstVectorDim,
+                                     CThreadTransferDstScalarPerVector>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+    using DefaultBlock2CTileMap =
+        decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(CGridDesc_M_N{}));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_k0_m0_m1_k1_{},
+              b_grid_desc_k0_n0_n1_k1_{},
+              c_grid_desc_m0_m10_m11_n0_n10_n11_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+            a_grid_desc_k0_m_k1_ = DeviceGemmDl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+            b_grid_desc_k0_n_k1_ = DeviceGemmDl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+            c_grid_desc_m_n_     = DeviceGemmDl::MakeCGridDescriptor_M_N(M, N, StrideC);
+
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_))
+            {
+                a_grid_desc_k0_m0_m1_k1_ =
+                    GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_k0_m_k1_);
+                b_grid_desc_k0_n0_n1_k1_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_k0_n_k1_);
+                c_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(c_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+
+        AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_;
+        BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_;
+        CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // TODO: unused, but may be useful in future.
+        index_t M01_;
+        index_t N01_;
+
+        // TODO: unused since gridwise_gemm_dl_v1r3 does NOT support prologue for the time being.
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceGemmDl::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            {
+                std::cout << "arg.a_grid_desc_k0_m0_m1_k1_{"
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n0_n1_k1_{"
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+
+            if(!GridwiseGemm::CheckValidity(
+                   arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdl_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size = GridwiseGemm::CalculateGridSize(
+                arg.c_grid_desc_m_n_.GetLength(I0), arg.c_grid_desc_m_n_.GetLength(I1));
+
+            const auto K0                    = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0);
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+            const bool has_double_tail_k_block_loop =
+                GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+            float ave_time = 0;
+
+            if(has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                const auto kernel =
+                    kernel_gemm_dl_v1r3<GridwiseGemm,
+                                        ADataType,
+                                        CDataType,
+                                        remove_reference_t<AGridDesc_K0_M0_M1_K1>,
+                                        remove_reference_t<BGridDesc_K0_N0_N1_K1>,
+                                        remove_reference_t<CGridDesc_M0_M10_M11_N0_N10_N11>,
+                                        remove_reference_t<DefaultBlock2CTileMap>,
+                                        true,
+                                        true>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m0_m1_k1_,
+                                                  arg.b_grid_desc_k0_n0_n1_k1_,
+                                                  arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+            {
+                const auto kernel =
+                    kernel_gemm_dl_v1r3<GridwiseGemm,
+                                        ADataType,
+                                        CDataType,
+                                        remove_reference_t<AGridDesc_K0_M0_M1_K1>,
+                                        remove_reference_t<BGridDesc_K0_N0_N1_K1>,
+                                        remove_reference_t<CGridDesc_M0_M10_M11_N0_N10_N11>,
+                                        remove_reference_t<DefaultBlock2CTileMap>,
+                                        true,
+                                        false>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m0_m1_k1_,
+                                                  arg.b_grid_desc_k0_n0_n1_k1_,
+                                                  arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                const auto kernel =
+                    kernel_gemm_dl_v1r3<GridwiseGemm,
+                                        ADataType,
+                                        CDataType,
+                                        remove_reference_t<AGridDesc_K0_M0_M1_K1>,
+                                        remove_reference_t<BGridDesc_K0_N0_N1_K1>,
+                                        remove_reference_t<CGridDesc_M0_M10_M11_N0_N10_N11>,
+                                        remove_reference_t<DefaultBlock2CTileMap>,
+                                        false,
+                                        true>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m0_m1_k1_,
+                                                  arg.b_grid_desc_k0_n0_n1_k1_,
+                                                  arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel =
+                    kernel_gemm_dl_v1r3<GridwiseGemm,
+                                        ADataType,
+                                        CDataType,
+                                        remove_reference_t<AGridDesc_K0_M0_M1_K1>,
+                                        remove_reference_t<BGridDesc_K0_N0_N1_K1>,
+                                        remove_reference_t<CGridDesc_M0_M10_M11_N0_N10_N11>,
+                                        remove_reference_t<DefaultBlock2CTileMap>,
+                                        false,
+                                        false>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m0_m1_k1_,
+                                                  arg.b_grid_desc_k0_n0_n1_k1_,
+                                                  arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(ck::get_device_name() == "gfx906" ||ck::get_device_name() == "gfx926" || ck::get_device_name() == "gfx1030")
+        {
+            return GridwiseGemm::CheckValidity(
+                arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_);
+        }
+        else
+        {
+            return false;
+        }
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmDl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << M1PerThread << ", "
+            << N1PerThread << ", "
+            << KPerThread
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
new file mode 100644
index 0000000..f1fb4ab
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
@@ -0,0 +1,682 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatDsPointer,
+          typename FloatE,
+          typename FloatRsPointer,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename RsGridDescriptor_MBlock_MPerBlock,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_multiple_d_multiple_r_xdl_cshuffle(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatDsPointer p_ds_grid,
+            FloatE* __restrict__ p_e_grid,
+            FloatRsPointer p_rs_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const QsElementwiseOperation qs_element_op,
+            const RsElementwiseOperation rs_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const RsGridDescriptor_MBlock_MPerBlock rs_grid_desc_mblock_mperblock,
+            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_ds_grid,
+                                                  p_e_grid,
+                                                  p_rs_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  qs_element_op,
+                                                  rs_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  rs_grid_desc_mblock_mperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = p_rs_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = qs_element_op;
+    ignore = rs_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = rs_grid_desc_mblock_mperblock;
+    ignore = block_2_etile_map;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// GEMM:
+//   input : A[AK0, M, AK1]
+//   input : B[AK0, N, AK1]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   output : R0[M], R1[M], ...
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+//   Q0 = reduce0(q_op0(E)), Q1 = reduce1(q_op0(E)), ...
+//   R0 = r_op0(Q0), R1 = r_op1(Q1), ...
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename ADataType,
+          typename BDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ReduceAccDataType,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation,
+          typename ThreadReduceOperations,
+          typename RsGlobalMemoryDataOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
+          index_t CDEReduceThreadTransferScalarPerVector_NPerBlock,
+          index_t RThreadTransferDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+    : public DeviceGemmMultipleDMultipleR<ALayout,
+                                          BLayout,
+                                          DELayout,
+                                          ADataType,
+                                          BDataType,
+                                          DsDataType,
+                                          EDataType,
+                                          RsDataType,
+                                          AElementwiseOperation,
+                                          BElementwiseOperation,
+                                          CDEElementwiseOperation,
+                                          QsElementwiseOperation,
+                                          RsElementwiseOperation>
+{
+    using DeviceOp = DeviceGemmMultipleDMultipleR_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    static constexpr index_t NumRTensor = RsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, DELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, DELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    // assume D is packed tensor
+    static auto MakeRGridDescriptor_M(index_t MRaw)
+    {
+        const auto r_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(MRaw));
+
+        const auto M    = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto MPad = M - MRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M
+            return transform_tensor_descriptor(r_grid_desc_mraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad)),
+                                               make_tuple(Sequence<0>{}),
+                                               make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad M
+            return r_grid_desc_mraw;
+        }
+    }
+
+    using AGridDesc_M_K = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N(1, 1, 1));
+    using RGridDesc_M   = decltype(MakeRGridDescriptor_M(1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleDMultipleR_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        ReduceAccDataType,
+        RsDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        QsElementwiseOperation,
+        RsElementwiseOperation,
+        ThreadReduceOperations,
+        InMemoryDataOperationEnum::Set,
+        RsGlobalMemoryDataOperation,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        EGridDesc_M_N,
+        RGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
+        CDEReduceThreadTransferScalarPerVector_NPerBlock,
+        RThreadTransferDstScalarPerVector_MPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+
+    using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 std::array<void*, NumRTensor> p_rs_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 std::array<index_t, NumDTensor> StrideDs,
+                 index_t StrideE,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op,
+                 QsElementwiseOperation qs_element_op,
+                 RsElementwiseOperation rs_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{}, // FIXME
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              p_rs_grid_{}, // FIXME
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N(MRaw, NRaw, StrideE)},
+              r_grid_desc_m_{DeviceOp::MakeRGridDescriptor_M(MRaw)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              rs_grid_desc_mblock_mperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              qs_element_op_{qs_element_op},
+              rs_element_op_{rs_element_op}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           e_grid_desc_m_n_,
+                                           r_grid_desc_m_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                static_for<0, NumDTensor, 1>{}([&](auto i) {
+                    using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                    p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                    const auto d_grid_desc_m_n =
+                        DeviceOp::MakeEGridDescriptor_M_N(MRaw, NRaw, StrideDs[i]);
+
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock_(i) =
+                        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            d_grid_desc_m_n);
+                });
+
+                static_for<0, NumRTensor, 1>{}([&](auto i) {
+                    using RDataType = remove_cvref_t<tuple_element_t<i.value, RsDataType>>;
+
+                    p_rs_grid_(i) = static_cast<RDataType*>(p_rs_grid[i]);
+
+                    rs_grid_desc_mblock_mperblock_(i) =
+                        GridwiseGemm::MakeRGridDescriptor_MBlock_MPerBlock(r_grid_desc_m_);
+                });
+            }
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+        typename GridwiseGemm::RsGridPointer p_rs_grid_;
+
+        // tensor descriptors
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+        RGridDesc_M r_grid_desc_m_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        StaticallyIndexedArray<
+            typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+            NumDTensor>
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_; // FIXME: Ds desc may be of different
+                                                             // type from E
+        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        StaticallyIndexedArray<typename GridwiseGemm::RGridDescriptor_MBlock_MPerBlock, NumRTensor>
+            rs_grid_desc_mblock_mperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+        QsElementwiseOperation qs_element_op_;
+        RsElementwiseOperation rs_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.r_grid_desc_m_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_gemm_multiple_d_multiple_r_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    typename GridwiseGemm::RsGridPointer,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    QsElementwiseOperation,
+                    RsElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    ck::StaticallyIndexedArray<
+                        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                        NumDTensor>,
+                    typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    ck::StaticallyIndexedArray<
+                        typename GridwiseGemm::RGridDescriptor_MBlock_MPerBlock,
+                        NumRTensor>,
+                    typename GridwiseGemm::DefaultBlock2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.p_rs_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.qs_element_op_,
+                                              arg.rs_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.rs_grid_desc_mblock_mperblock_,
+                                              arg.block_2_etile_map_);
+            };
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.r_grid_desc_m_,
+                                           arg.block_2_etile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_e,
+                             std::array<void*, NumRTensor> p_rs,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t StrideA,
+                             index_t StrideB,
+                             std::array<index_t, NumDTensor> StrideDs,
+                             index_t StrideE,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op,
+                             QsElementwiseOperation qs_element_op,
+                             RsElementwiseOperation rs_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        p_rs,
+                        MRaw,
+                        NRaw,
+                        KRaw,
+                        StrideA,
+                        StrideB,
+                        StrideDs,
+                        StrideE,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op,
+                        qs_element_op,
+                        rs_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      std::array<const void*, NumDTensor> p_ds,
+                                                      void* p_e,
+                                                      std::array<void*, NumRTensor> p_rs,
+                                                      index_t MRaw,
+                                                      index_t NRaw,
+                                                      index_t KRaw,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      std::array<index_t, NumDTensor> StrideDs,
+                                                      index_t StrideE,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CDEElementwiseOperation cde_element_op,
+                                                      QsElementwiseOperation qs_element_op,
+                                                      RsElementwiseOperation rs_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          p_rs,
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideE,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op,
+                                          qs_element_op,
+                                          rs_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmMultipleDMultipleR_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << getGemmSpecializationString(GemmSpec)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000..3830e1c
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,698 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_multiple_d_xdl_cshuffle(const ABDataType* __restrict__ p_a_grid,
+                                            const ABDataType* __restrict__ p_b_grid,
+                                            DsPointer p_ds_grid,
+                                            EDataType* __restrict__ p_e_grid,
+                                            const AElementwiseOperation a_element_op,
+                                            const BElementwiseOperation b_element_op,
+                                            const CDEElementwiseOperation cde_element_op,
+                                            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+                                            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+                                            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                                                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                                                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_ds_grid,
+                                                  p_e_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
+                                                                     BLayout,
+                                                                     DsLayout,
+                                                                     ELayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     DsDataType,
+                                                                     EDataType,
+                                                                     AElementwiseOperation,
+                                                                     BElementwiseOperation,
+                                                                     CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGemmMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // desc for problem definition
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        PipelineVer>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 std::array<index_t, NumDTensor> StrideDs,
+                 index_t StrideE,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(MRaw, NRaw, StrideE)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              MRaw_{MRaw},
+              NRaw_{NRaw},
+              KRaw_{KRaw}
+        {
+            // populate pointer, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) =
+                    DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaw, NRaw, StrideDs[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // for checking vector load/store
+        index_t MRaw_;
+        index_t NRaw_;
+        index_t KRaw_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_gemm_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::Block2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        // check vector load/store
+        {
+            using Row = ck::tensor_layout::gemm::RowMajor;
+            using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+            // check vector load of A
+            if constexpr(is_same_v<ALayout, Row> && ABlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<ALayout, Col> && ABlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.MRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector laod of B
+            if constexpr(is_same_v<BLayout, Col> && BBlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<BLayout, Row> && BBlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.NRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector load of Ds
+            // only support RowMajor for now
+            bool all_valid = true;
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                if constexpr(!is_same_v<DLayout, Row>)
+                {
+                    all_valid = false;
+                }
+            });
+
+            if(!all_valid)
+            {
+                return false;
+            }
+
+            // check vector store of E
+            // only support RowMajor for now
+            if constexpr(is_same_v<ELayout, Row>)
+            {
+                if(arg.NRaw_ % CDEBlockTransferScalarPerVector_NPerBlock != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.ds_grid_desc_m_n_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.block_2_etile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_e,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t StrideA,
+                             index_t StrideB,
+                             std::array<index_t, NumDTensor> StrideDs,
+                             index_t StrideE,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        MRaw,
+                        NRaw,
+                        KRaw,
+                        StrideA,
+                        StrideB,
+                        StrideDs,
+                        StrideE,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        index_t MRaw,
+                        index_t NRaw,
+                        index_t KRaw,
+                        index_t StrideA,
+                        index_t StrideB,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        index_t StrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideE,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceGemmMultipleD_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << getGemmSpecializationString(GemmSpec)
+            << ">"
+            << " LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp
new file mode 100644
index 0000000..cf19083
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp
@@ -0,0 +1,835 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
+// version currently has compiler issues with register spill which further causes validation
+// failures.
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename ReduceAccDataType,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename ReduceOperations,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          typename ReduceGlobalMemoryDataOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+          index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<0, ReduceOperations::Size()>
+{
+    using DeviceOp = DeviceGemmReduce_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(M)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_right_pad_transform(MRaw, MPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(N)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_right_pad_transform(NRaw, NPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    // assume Reduce is packed tensor
+    static auto MakeReduceGridDescriptor_M(index_t MRaw)
+    {
+        const auto d_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(MRaw));
+
+        const auto M    = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto MPad = M - MRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M
+            return transform_tensor_descriptor(d_grid_desc_mraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad)),
+                                               make_tuple(Sequence<0>{}),
+                                               make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad M
+            return d_grid_desc_mraw;
+        }
+    }
+
+    using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N       = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+    using ReduceGridDesc_M    = decltype(MakeReduceGridDescriptor_M(1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        ReduceAccDataType,
+        ReducePtrsGlobal,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        ReduceOperations,
+        ReduceInElementwiseOperations,
+        ReduceAccElementwiseOperations,
+        InMemoryDataOperationEnum::Set,
+        ReduceGlobalMemoryDataOperation,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        ReduceGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+        CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+        CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+        LoopSched>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 ReducePtrsGlobal p_reduces_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op,
+                 ReduceInElementwiseOperations reduce_in_element_ops,
+                 ReduceAccElementwiseOperations reduce_out_element_ops)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              p_reduces_grid_{p_reduces_grid},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC)},
+              reduce_grid_desc_m_{DeviceOp::MakeReduceGridDescriptor_M(MRaw)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              reduce_grid_desc_mblock_mperblock_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              reduce_in_element_ops_{reduce_in_element_ops},
+              reduce_out_element_ops_{reduce_out_element_ops}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+
+                reduce_grid_desc_mblock_mperblock_ =
+                    GridwiseGemm::MakeReduceGridDescriptor_MBlock_MPerBlock(reduce_grid_desc_m_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        ReducePtrsGlobal p_reduces_grid_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        ReduceGridDesc_M reduce_grid_desc_m_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock
+            reduce_grid_desc_mblock_mperblock_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+        ReduceInElementwiseOperations reduce_in_element_ops_;
+        ReduceAccElementwiseOperations reduce_out_element_ops_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_bk0_n_bk1_{"
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+                std::cout << "arg.reduce_grid_desc_m_{ " << arg.reduce_grid_desc_m_.GetLength(I0) << "}"
+                          << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                            arg.b_grid_desc_bk0_n_bk1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float elapsed_time = 0.0f;
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_reduce_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    ReducePtrsGlobal,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    ReduceInElementwiseOperations,
+                    ReduceAccElementwiseOperations,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    true>;
+
+                elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_reduces_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.reduce_in_element_ops_,
+                                           arg.reduce_out_element_ops_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.reduce_grid_desc_mblock_mperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_reduce_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    ReducePtrsGlobal,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    ReduceInElementwiseOperations,
+                    ReduceAccElementwiseOperations,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    false>;
+
+                elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_reduces_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.reduce_in_element_ops_,
+                                           arg.reduce_out_element_ops_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.reduce_grid_desc_mblock_mperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+
+            return elapsed_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static constexpr int NumReduce = ReduceOperations::Size();
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             const void* p_bias,
+                             std::array<const void*, 0> p_ds,
+                             void* p_c,
+                             std::array<void*, NumReduce> p_reduces,
+                             ck::index_t M,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t StrideA,
+                             ck::index_t StrideB,
+                             ck::index_t StrideC,
+                             std::array<ck::index_t, 0> StrideDs,
+                             std::array<void*, 3> gemm_element_ops,
+                             std::array<void*, 0> d_element_ops,
+                             std::array<void*, NumReduce> reduce_in_element_op,
+                             std::array<void*, NumReduce> reduce_out_element_op)
+    {
+        (void)p_bias;
+        (void)p_ds;
+        (void)StrideDs;
+        (void)d_element_ops;
+
+        ReducePtrsGlobal reduce_tuple = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReducePtrsGlobal{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return static_cast<T*>(p_reduces[I]);
+            },
+            Number<NumReduce>{});
+
+        ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceInElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_in_element_op[I]));
+            },
+            Number<NumReduce>{});
+        ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceAccElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_out_element_op[I]));
+            },
+            Number<NumReduce>{});
+
+        AElementwiseOperation a_element_op =
+            *(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
+        BElementwiseOperation b_element_op =
+            *(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
+        CElementwiseOperation c_element_op =
+            *(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
+
+        return Argument{static_cast<const ADataType*>(p_a),
+                        static_cast<const BDataType*>(p_b),
+                        static_cast<CDataType*>(p_c),
+                        reduce_tuple,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op,
+                        reduce_in_element_ops,
+                        reduce_out_element_ops};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const void* p_bias,
+                        std::array<const void*, 0> p_ds,
+                        void* p_c,
+                        std::array<void*, NumReduce> p_reduces,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        std::array<ck::index_t, 0> StrideDs,
+                        std::array<void*, 3> gemm_element_ops,
+                        std::array<void*, 0> d_element_ops,
+                        std::array<void*, NumReduce> reduce_in_element_op,
+                        std::array<void*, NumReduce> reduce_out_element_op,
+                        ck::index_t = 1) override
+    {
+        (void)p_bias;
+        (void)p_ds;
+        (void)StrideDs;
+        (void)d_element_ops;
+
+        ReducePtrsGlobal reduce_tuple = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReducePtrsGlobal{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return static_cast<T*>(p_reduces[I]);
+            },
+            Number<NumReduce>{});
+
+        ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceInElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_in_element_op[I]));
+            },
+            Number<NumReduce>{});
+        ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceAccElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_out_element_op[I]));
+            },
+            Number<NumReduce>{});
+
+        AElementwiseOperation a_element_op =
+            *(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
+        BElementwiseOperation b_element_op =
+            *(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
+        CElementwiseOperation c_element_op =
+            *(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
+
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          reduce_tuple,
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op,
+                                          reduce_in_element_ops,
+                                          reduce_out_element_ops);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmReduce_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
new file mode 100644
index 0000000..21bb36b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
@@ -0,0 +1,570 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector,
+          ck::index_t NumPrefetch         = 1,
+          ck::LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
+struct DeviceGemmXdl : public DeviceGemm<ALayout,
+                                         BLayout,
+                                         CLayout,
+                                         ADataType,
+                                         BDataType,
+                                         CDataType,
+                                         AElementwiseOperation,
+                                         BElementwiseOperation,
+                                         CElementwiseOperation>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using CGridDesc_M_N     = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder,
+        CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector,
+        NumPrefetch,
+        LoopSched,
+        PipelineVer>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              kraw_{K}
+        {
+            a_grid_desc_k0_m_k1_ = DeviceGemmXdl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+            b_grid_desc_k0_n_k1_ = DeviceGemmXdl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+            c_grid_desc_m_n_     = DeviceGemmXdl::MakeCGridDescriptor_M_N(M, N, StrideC);
+
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
+                                           b_grid_desc_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
+            c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+        index_t kraw_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceGemmXdl::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceGemmXdl::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceGemmXdl::BGridDesc_K0_N_K1>,
+                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    true>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceGemmXdl::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceGemmXdl::BGridDesc_K0_N_K1>,
+                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(ck::get_device_name() == "gfx908")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else if(ck::get_device_name() == "gfx90a")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        if(arg.kraw_ % K1 != 0)
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceGemmXdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave
+            << ">"
+            << " NumPrefetch: "
+            << NumPrefetch << ", "
+            << "LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
new file mode 100644
index 0000000..cc8c8d4
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
@@ -0,0 +1,700 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
+// version currently has compiler issues with register spill which further causes validation
+// failures.
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
+                                                   BLayout,
+                                                   CLayout,
+                                                   ADataType,
+                                                   BDataType,
+                                                   CDataType,
+                                                   AElementwiseOperation,
+                                                   BElementwiseOperation,
+                                                   CElementwiseOperation>
+{
+    using DeviceOp = DeviceGemm_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(M)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_right_pad_transform(MRaw, MPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(N)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_right_pad_transform(NRaw, NPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N       = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        PipelineVer>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              kraw_{KRaw}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+        index_t kraw_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_bk0_n_bk1_{"
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                            arg.b_grid_desc_bk0_n_bk1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    true>;
+
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    false>;
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        if((arg.kraw_ % AK1 != 0 || arg.kraw_ % BK1 != 0) &&
+           !(GemmSpec == GemmSpecialization::MKPadding ||
+             GemmSpec == GemmSpecialization::NKPadding ||
+             GemmSpec == GemmSpecialization::MNKPadding ||
+             GemmSpec == GemmSpecialization::KPadding))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        MRaw,
+                        NRaw,
+                        KRaw,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t MRaw,
+                                                      index_t NRaw,
+                                                      index_t KRaw,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceGemm_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1
+            << ">"
+            << " LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp
new file mode 100644
index 0000000..875623d
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp
@@ -0,0 +1,773 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// The GEMM + Layernorm implementation is a specialized kernel which allows fusing both layers
+// together given the condition GEMM extents N of MNK is spanned by a single workgroup. For example,
+// a kernel configured with NPerBlock = 128 allows to operate on all GEMM sizes if N <= 128
+//
+// Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
+// version currently has compiler issues with register spill which further causes validation
+// failures.
+//
+// D = Layernorm(acc_element_op(A * B + broadcast(bias)) + add) * broadcast(gamma) + broadcast(beta)
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename C0DataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename ReduceAccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGemmLayerNorm_Xdl_CShuffle : public BaseOperator
+{
+    using DeviceOp = DeviceGemmLayerNorm_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(M)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_right_pad_transform(MRaw, MPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(N)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_right_pad_transform(NRaw, NPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    static auto MakeGridDescriptor_N(index_t NRaw)
+    {
+        const auto grid_desc_nraw = make_naive_tensor_descriptor_packed(make_tuple(NRaw));
+
+        const auto N    = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad N
+            return transform_tensor_descriptor(grid_desc_nraw,
+                                               make_tuple(make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}),
+                                               make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad N
+            return grid_desc_nraw;
+        }
+    }
+
+    using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N       = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+    using C0GridDesc_N        = decltype(MakeGridDescriptor_N(1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmLayernorm_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        C0DataType,
+        ReduceAccDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        AccElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        C0GridDesc_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+        CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using Block2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 const C0DataType* p_c0_grid_add,
+                 const C0DataType* p_c0_grid_bias,
+                 const C0DataType* p_c0_grid_gamma,
+                 const C0DataType* p_c0_grid_beta,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 AccElementwiseOperation acc_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              p_c0_grid_bias_{p_c0_grid_bias},
+              p_c0_grid_add_{p_c0_grid_add},
+              p_c0_grid_gamma_{p_c0_grid_gamma},
+              p_c0_grid_beta_{p_c0_grid_beta},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC)},
+              c0_grid_desc_n_{MakeGridDescriptor_N(NRaw)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              c0_grid_desc_nblock_nperblock_{},
+              block_2_ctile_map_{Block2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              c_element_op_{c_element_op}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+
+                c0_grid_desc_nblock_nperblock_ =
+                    GridwiseGemm::MakeC0GridDescriptor_NBlock_NPerBlock(c0_grid_desc_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        const C0DataType* p_c0_grid_bias_;
+        const C0DataType* p_c0_grid_add_;
+        const C0DataType* p_c0_grid_gamma_;
+        const C0DataType* p_c0_grid_beta_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        C0GridDesc_N c0_grid_desc_n_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::C0GridDescriptor_NBlock_NPerBlock c0_grid_desc_nblock_nperblock_;
+        Block2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_bk0_n_bk1_{"
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                            arg.b_grid_desc_bk0_n_bk1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_layernorm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    C0DataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    AccElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C0GridDescriptor_NBlock_NPerBlock,
+                    Block2CTileMap,
+                    true>;
+
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_c0_grid_bias_,
+                                           arg.p_c0_grid_add_,
+                                           arg.p_c0_grid_gamma_,
+                                           arg.p_c0_grid_beta_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.acc_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.c0_grid_desc_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_layernorm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    C0DataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    AccElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C0GridDescriptor_NBlock_NPerBlock,
+                    Block2CTileMap,
+                    false>;
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_c0_grid_bias_,
+                                           arg.p_c0_grid_add_,
+                                           arg.p_c0_grid_gamma_,
+                                           arg.p_c0_grid_beta_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.acc_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.c0_grid_desc_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             const C0DataType* p_c0_bias,
+                             const C0DataType* p_c0_add,
+                             const C0DataType* p_c0_gamma,
+                             const C0DataType* p_c0_beta,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             AccElementwiseOperation acc_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        p_c0_bias,
+                        p_c0_add,
+                        p_c0_gamma,
+                        p_c0_beta,
+                        MRaw,
+                        NRaw,
+                        KRaw,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        a_element_op,
+                        b_element_op,
+                        acc_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      const void* p_c0_bias,
+                                                      const void* p_c0_add,
+                                                      const void* p_c0_gamma,
+                                                      const void* p_c0_beta,
+                                                      index_t MRaw,
+                                                      index_t NRaw,
+                                                      index_t KRaw,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      AccElementwiseOperation acc_element_op,
+                                                      CElementwiseOperation c_element_op,
+                                                      index_t /* KBatch */ = 1)
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          static_cast<const C0DataType*>(p_c0_bias),
+                                          static_cast<const C0DataType*>(p_c0_add),
+                                          static_cast<const C0DataType*>(p_c0_gamma),
+                                          static_cast<const C0DataType*>(p_c0_beta),
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          acc_element_op,
+                                          c_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmLayerNorm_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp
new file mode 100644
index 0000000..42cabce
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp
@@ -0,0 +1,523 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp"
+
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockBufferSize,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector>
+struct DeviceGemmXdlSkipBLds : public DeviceGemm<ALayout,
+                                                 BLayout,
+                                                 CLayout,
+                                                 ADataType,
+                                                 BDataType,
+                                                 CDataType,
+                                                 AElementwiseOperation,
+                                                 BElementwiseOperation,
+                                                 CElementwiseOperation>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+    static_assert(BBlockBufferSize >= 2);
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using CGridDesc_M_N     = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_skip_b_lds_v1<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferSrcScalarPerVector,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockBufferSize,
+        Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder,
+        CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+            a_grid_desc_k0_m_k1_ =
+                DeviceGemmXdlSkipBLds::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+            b_grid_desc_k0_n_k1_ =
+                DeviceGemmXdlSkipBLds::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+            c_grid_desc_m_n_ = DeviceGemmXdlSkipBLds::MakeCGridDescriptor_M_N(M, N, StrideC);
+
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
+            {
+                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
+
+                block_2_ctile_map_ =
+                    GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+
+                b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_K1_K2_N0_N1_N2_N3_K3(b_grid_desc_k0_n_k1_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3
+            b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3_;
+        typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
+            c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceGemmXdlSkipBLds::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.M01_,
+                                            arg.N01_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+
+            if(has_main_k0_block_loop)
+            {
+                const auto kernel = kernel_gemm_xdlops_skip_b_lds_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceGemmXdlSkipBLds::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceGemmXdlSkipBLds::BGridDesc_K0_N_K1>,
+                    remove_reference_t<typename GridwiseGemm::BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3>,
+                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    true>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_skip_b_lds_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceGemmXdlSkipBLds::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceGemmXdlSkipBLds::BGridDesc_K0_N_K1>,
+                    remove_reference_t<typename GridwiseGemm::BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3>,
+                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.M01_,
+                                           arg.N01_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmXdlSkipBLds"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
new file mode 100644
index 0000000..5051518
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
@@ -0,0 +1,650 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXDL>
+struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
+                                                             BLayout,
+                                                             CLayout,
+                                                             ADataType,
+                                                             BDataType,
+                                                             CDataType,
+                                                             AElementwiseOperation,
+                                                             BElementwiseOperation,
+                                                             CElementwiseOperation>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+
+    static auto
+    MakeAGridDescriptor_KBatch_K0_M_K1(index_t M, index_t K, index_t StrideA, int KBatch, int KPad)
+    {
+        assert(KPad % (K1 * KBatch) == 0);
+
+        const index_t K0 = KPad / (K1 * KBatch);
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    static auto
+    MakeBGridDescriptor_KBatch_K0_N_K1(index_t K, index_t N, index_t StrideB, int KBatch, int KPad)
+    {
+        assert(KPad % (K1 * KBatch) == 0);
+
+        const index_t K0 = KPad / (K1 * KBatch);
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+            b_grid_desc_k_n,
+            make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto GetKPad(index_t K, index_t KBatch)
+    {
+        const index_t K0   = math::integer_divide_ceil(K, K1 * K0PerBlock * KBatch) * K0PerBlock;
+        const index_t KPad = KBatch * K0 * K1;
+        return KPad;
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_KBatch_K0_M_K1(1, 1, 1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_KBatch_K0_N_K1(1, 1, 1, 1, 1));
+    using CGridDesc_M_N     = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMRepeatPerShuffle,
+        CShuffleNRepeatPerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXDL,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
+
+    // GridwiseGemm
+    using GridwiseGemmAtomicAdd = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::AtomicAdd,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMRepeatPerShuffle,
+        CShuffleNRepeatPerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXDL,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
+
+    using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        decltype(GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}));
+
+    using Block2CTileMap = typename GridwiseGemm::CBlockClusterAdaptor;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op,
+                 index_t k_batch)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_kbatch_k0_m_k1_{},
+              b_grid_desc_kbatch_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              k_batch_{k_batch}
+        {
+            int KPad = DeviceGemmXdlSplitKCShuffle::GetKPad(K, k_batch_);
+
+            a_grid_desc_kbatch_k0_m_k1_ =
+                DeviceGemmXdlSplitKCShuffle::MakeAGridDescriptor_KBatch_K0_M_K1(
+                    M, K, StrideA, k_batch_, KPad);
+            b_grid_desc_kbatch_k0_n_k1_ =
+                DeviceGemmXdlSplitKCShuffle::MakeBGridDescriptor_KBatch_K0_N_K1(
+                    K, N, StrideB, k_batch_, KPad);
+            c_grid_desc_m_n_ = DeviceGemmXdlSplitKCShuffle::MakeCGridDescriptor_M_N(M, N, StrideC);
+
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_kbatch_k0_m_k1_,
+                                           b_grid_desc_kbatch_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_kbatch_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_kbatch_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        Block2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+        index_t k_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceGemmXdlSplitKCShuffle::Argument;
+
+        void ShowInfo(const Argument& arg)
+        {
+            std::cout << "arg.a_grid_desc_kbatch_k0_m_k1_{"
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I2) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.b_grid_desc_kbatch_k0_n_k1_{"
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I0) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I1) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I2) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                      << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+        }
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            ShowInfo(arg);
+
+            const auto kbatch = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0);
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                            arg.b_grid_desc_kbatch_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K0 = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+
+            const auto Run = [&](const auto& kernel) {
+                hipGetErrorString(hipMemset(
+                    arg.p_c_grid_,
+                    0,
+                    arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
+                        sizeof(CDataType)));
+
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.a_grid_desc_kbatch_k0_m_k1_,
+                                           arg.b_grid_desc_kbatch_k0_n_k1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.block_2_ctile_map_);
+            };
+
+            if(has_main_k0_block_loop)
+            {
+                if(kbatch == 1)
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r4r2<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::
+                                               CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        AElementwiseOperation,
+                        BElementwiseOperation,
+                        CElementwiseOperation,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::Block2CTileMap>,
+                        true>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r4r2<
+                        GridwiseGemmAtomicAdd,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::
+                                               CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        AElementwiseOperation,
+                        BElementwiseOperation,
+                        CElementwiseOperation,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::Block2CTileMap>,
+                        true>;
+
+                    Run(kernel);
+                }
+            }
+            else
+            {
+                if(kbatch == 1)
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r4r2<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::
+                                               CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        AElementwiseOperation,
+                        BElementwiseOperation,
+                        CElementwiseOperation,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::Block2CTileMap>,
+                        false>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r4r2<
+                        GridwiseGemmAtomicAdd,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::
+                                               CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        AElementwiseOperation,
+                        BElementwiseOperation,
+                        CElementwiseOperation,
+                        remove_reference_t<DeviceGemmXdlSplitKCShuffle::Block2CTileMap>,
+                        false>;
+
+                    Run(kernel);
+                }
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                           arg.b_grid_desc_kbatch_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op,
+                             index_t KBatch)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op,
+                        KBatch};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op,
+                                                      ck::index_t KBatch = 1) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op,
+                                          KBatch);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmXdlSplitKCShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000..03d9e26
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,907 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename ContractionMultiDKernelArg,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_contraction_multiple_d_xdl_cshuffle(
+            const void CK_CONSTANT_ADDRESS_SPACE* contraction_args,
+            const index_t group_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const auto contraction_arg_ptr = reinterpret_cast<const ContractionMultiDKernelArg*>(
+        cast_pointer_to_generic_address_space(contraction_args));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+
+    while((!(block_id >= contraction_arg_ptr[group_id].block_start_ &&
+             block_id < contraction_arg_ptr[group_id].block_end_)) &&
+          left <= right)
+    {
+        if(block_id < contraction_arg_ptr[group_id].block_start_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        contraction_arg_ptr[group_id].p_a_grid_,
+        contraction_arg_ptr[group_id].p_b_grid_,
+        contraction_arg_ptr[group_id].p_ds_grid_,
+        contraction_arg_ptr[group_id].p_e_grid_,
+        p_shared,
+        a_element_op,
+        b_element_op,
+        cde_element_op,
+        contraction_arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
+        contraction_arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
+        contraction_arg_ptr[group_id].ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+        contraction_arg_ptr[group_id].e_grid_desc_mblock_mperblock_nblock_nperblock_,
+        contraction_arg_ptr[group_id].block_2_etile_map_);
+#else
+    ignore = contraction_args;
+    ignore = group_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization DESpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedContractionMultipleD_Xdl_CShuffle
+    : public DeviceGroupedContractionMultipleD<NumDimM,
+                                               NumDimN,
+                                               NumDimK,
+                                               ADataType,
+                                               BDataType,
+                                               DsDataType,
+                                               EDataType,
+                                               AElementwiseOperation,
+                                               BElementwiseOperation,
+                                               CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedContractionMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    // Assume: A[M0, M1, M2, ..., K0, K1, K2, ...]
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_ms_ks_strides_vec)
+    {
+        assert(a_ms_ks_lengths_vec.size() == NumDimM + NumDimK &&
+               a_ms_ks_strides_vec.size() == NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto a_ms_ks_lengths = to_tuple(a_ms_ks_lengths_vec, Number<NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(a_ms_ks_strides_vec, Number<NumDimM + NumDimK>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(a_ms_ks_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(a_ms_ks_lengths, kDimIds);
+
+        if constexpr(ASpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto a_grid_desc_mraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(M, K),
+                make_tuple(a_ms_ks_strides[Number<NumDimM - 1>{}],
+                           a_ms_ks_strides[Number<NumDimM + NumDimK - 1>{}]));
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+        else
+        {
+            // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
+            const auto a_grid_desc_ms_ks =
+                make_naive_tensor_descriptor(a_ms_ks_lengths, a_ms_ks_strides);
+
+            // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
+            const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
+                a_grid_desc_ms_ks,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
+                make_tuple(mDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+    }
+
+    // Assume: B[N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_ns_ks_lengths_vec,
+                                        const std::vector<index_t>& b_ns_ks_strides_vec)
+    {
+        assert(b_ns_ks_lengths_vec.size() == NumDimN + NumDimK &&
+               b_ns_ks_strides_vec.size() == NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(b_ns_ks_lengths_vec, Number<NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(b_ns_ks_strides_vec, Number<NumDimN + NumDimK>{});
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
+
+        // lengths for N0, N1, ...
+        const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
+
+        if constexpr(BSpec == TensorSpecialization::Packed)
+        {
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto b_grid_desc_nraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(N, K),
+                make_tuple(b_ns_ks_strides[Number<NumDimN - 1>{}],
+                           b_ns_ks_strides[Number<NumDimN + NumDimK - 1>{}]));
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+        else
+        {
+            // naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
+            const auto b_grid_desc_ns_ks =
+                make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
+
+            // transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
+            const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
+                b_grid_desc_ns_ks,
+                make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
+                make_tuple(nDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+    }
+
+    // assume E[M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& e_ms_ns_strides_vec)
+    {
+        assert(e_ms_ns_lengths_vec.size() == NumDimM + NumDimN &&
+               e_ms_ns_strides_vec.size() == NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto e_ms_ns_lengths = to_tuple(e_ms_ns_lengths_vec, Number<NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(e_ms_ns_strides_vec, Number<NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(M, N),
+                make_tuple(e_ms_ns_strides[Number<NumDimM - 1>{}],
+                           e_ms_ns_strides[Number<NumDimM + NumDimN - 1>{}]));
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+        else
+        {
+            // naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_ms_ns =
+                make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
+
+            // transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+            const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_ms_ns,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+                make_tuple(mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_M_N(ds_ms_ns_lengths_vec[i],
+                                                         ds_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K({}, {}));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K({}, {}));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+
+    struct GroupedContractionBlock2ETileMap
+    {
+        // block-to-e-tile map
+        using Block2ETileMap =
+            remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+        GroupedContractionBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n,
+                                         ck::index_t BlockStart)
+        {
+            default_block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n);
+            block_start_               = BlockStart;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            return default_block_2_etile_map_.CalculateBottomIndex(
+                make_multi_index(idx_top[I0] - block_start_));
+        }
+
+        // it's actually E-Tile
+        template <typename CTileIdx, typename CTileDim>
+        __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                                 const CTileDim& c_tile_dim) const
+        {
+            return default_block_2_etile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+        }
+
+        __host__ bool CheckValidity(const EGridDesc_M_N& e_grid_desc_m_n) const
+        {
+            return default_block_2_etile_map_.CheckValidity(e_grid_desc_m_n);
+        }
+
+        Block2ETileMap default_block_2_etile_map_;
+        ck::index_t block_start_;
+    };
+
+    struct ContractionMultiDKernelArg
+    {
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // lock-to-e-tile map
+        GroupedContractionBlock2ETileMap block_2_etile_map_;
+
+        ck::index_t block_start_, block_end_;
+    };
+
+    struct ContractionMultiDDeviceArg
+    {
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // Strides for the last M/N/K dimensions of A/B/Ds/E
+        //   for sanity check of vector load/store
+        index_t a_mz_stride_;
+        index_t a_kz_stride_;
+        index_t b_nz_stride_;
+        index_t b_kz_stride_;
+        std::array<index_t, NumDTensor> ds_nz_stride_;
+        // index_t e_mz_stride_;
+        index_t e_nz_stride_;
+    };
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(std::vector<const void*> p_a_vec,
+                 std::vector<const void*> p_b_vec,
+                 std::vector<std::array<const void*, NumDTensor>> p_ds_vec,
+                 std::vector<void*> p_e_vec,
+                 std::vector<ContractionDesc<NumDTensor>> contraction_descs,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            group_count_ = contraction_descs.size();
+
+            if(!(group_count_ == p_a_vec.size() && group_count_ == p_b_vec.size() &&
+                 group_count_ == p_e_vec.size()))
+            {
+                throw std::runtime_error("wrong! group_count_ != a/b/e_vec.size");
+            }
+
+            contraction_multi_d_kernel_args_.reserve(group_count_);
+
+            grid_size_ = 0;
+
+            for(std::size_t i = 0; i < group_count_; i++)
+            {
+                const auto p_a_grid = static_cast<const ADataType*>(p_a_vec[i]);
+                const auto p_b_grid = static_cast<const BDataType*>(p_b_vec[i]);
+                const auto p_e_grid = static_cast<EDataType*>(p_e_vec[i]);
+
+                const auto a_grid_desc_m_k = DeviceOp::MakeAGridDescriptor_M_K(
+                    contraction_descs[i].a_ms_ks_lengths, contraction_descs[i].a_ms_ks_strides);
+                const auto b_grid_desc_n_k = DeviceOp::MakeBGridDescriptor_N_K(
+                    contraction_descs[i].b_ns_ks_lengths, contraction_descs[i].b_ns_ks_strides);
+
+                DsGridDesc_M_N ds_grid_desc_m_n;
+                typename GridwiseGemm::DsGridPointer p_ds_grid;
+
+                // populate pointer, batch stride, desc for Ds
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
+
+                    // D pointer
+                    p_ds_grid(j) = static_cast<const DDataType*>(p_ds_vec[i][j]);
+
+                    // D desc
+                    ds_grid_desc_m_n(j) =
+                        DeviceOp::MakeEGridDescriptor_M_N(contraction_descs[i].ds_ms_ns_lengths[j],
+                                                          contraction_descs[i].ds_ms_ns_strides[j]);
+                });
+
+                const auto e_grid_desc_m_n = DeviceOp::MakeEGridDescriptor_M_N(
+                    contraction_descs[i].e_ms_ns_lengths, contraction_descs[i].e_ms_ns_strides);
+
+                const auto a_grid_desc_ak0_m_ak1 =
+                    GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k);
+                const auto b_grid_desc_bk0_n_bk1 =
+                    GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k);
+
+                const auto ds_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n);
+                const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n);
+
+                const index_t grid_size_grp =
+                    GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n)
+                        .CalculateGridSize(e_grid_desc_m_n);
+
+                const index_t BlockStart = grid_size_;
+                const index_t BlockEnd   = grid_size_ + grid_size_grp;
+
+                grid_size_ += grid_size_grp;
+
+                const auto block_2_etile_map =
+                    GroupedContractionBlock2ETileMap(e_grid_desc_m_n, BlockStart);
+
+                // for sanity check of vector memory access
+                const index_t a_mz_stride = contraction_descs[i].a_ms_ks_strides[NumDimM - 1];
+                const index_t a_kz_stride =
+                    contraction_descs[i].a_ms_ks_strides[NumDimM + NumDimK - 1];
+
+                const index_t b_nz_stride = contraction_descs[i].b_ns_ks_strides[NumDimN - 1];
+                const index_t b_kz_stride =
+                    contraction_descs[i].b_ns_ks_strides[NumDimN + NumDimK - 1];
+
+                std::array<index_t, NumDTensor> ds_nz_stride;
+                for(index_t j = 0; j < NumDTensor; ++j)
+                {
+                    ds_nz_stride[j] =
+                        contraction_descs[i].ds_ms_ns_strides[j][NumDimM + NumDimN - 1];
+                }
+
+                const index_t e_nz_stride =
+                    contraction_descs[i].e_ms_ns_strides[NumDimM + NumDimN - 1];
+
+                if(GridwiseGemm::CheckValidity(a_grid_desc_m_k,
+                                               b_grid_desc_n_k,
+                                               ds_grid_desc_m_n,
+                                               e_grid_desc_m_n,
+                                               block_2_etile_map))
+                {
+                    contraction_multi_d_kernel_args_.push_back(
+                        {p_a_grid,
+                         p_b_grid,
+                         p_ds_grid,
+                         p_e_grid,
+                         a_grid_desc_ak0_m_ak1,
+                         b_grid_desc_bk0_n_bk1,
+                         ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                         e_grid_desc_mblock_mperblock_nblock_nperblock,
+                         block_2_etile_map,
+                         BlockStart,
+                         BlockEnd});
+
+                    contraction_multi_d_device_args_.push_back({a_grid_desc_m_k,
+                                                                b_grid_desc_n_k,
+                                                                ds_grid_desc_m_n,
+                                                                e_grid_desc_m_n,
+                                                                a_mz_stride,
+                                                                a_kz_stride,
+                                                                b_nz_stride,
+                                                                b_kz_stride,
+                                                                ds_nz_stride,
+                                                                e_nz_stride});
+                }
+            }
+        }
+
+        std::vector<ContractionMultiDKernelArg> contraction_multi_d_kernel_args_;
+        std::vector<ContractionMultiDDeviceArg> contraction_multi_d_device_args_;
+
+        std::size_t group_count_;
+        index_t grid_size_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            bool has_main_k_block_loop = true;
+
+            for(std::size_t i = 0; i < arg.group_count_; i++)
+            {
+                const auto K =
+                    arg.contraction_multi_d_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                    arg.contraction_multi_d_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(K) != has_main_k_block_loop)
+                {
+                    throw std::runtime_error("wrong! not all gemm has_main_k_block_loop");
+                }
+            }
+
+            hipGetErrorString(hipMemcpy(arg.p_workspace_,
+                                        arg.contraction_multi_d_kernel_args_.data(),
+                                        arg.contraction_multi_d_kernel_args_.size() *
+                                            sizeof(ContractionMultiDKernelArg),
+                                        hipMemcpyHostToDevice));
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel =
+                    kernel_grouped_contraction_multiple_d_xdl_cshuffle<GridwiseGemm,
+                                                                       ContractionMultiDKernelArg,
+                                                                       AElementwiseOperation,
+                                                                       BElementwiseOperation,
+                                                                       CDEElementwiseOperation,
+                                                                       has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(arg.grid_size_),
+                    dim3(BlockSize),
+                    0,
+                    cast_pointer_to_constant_address_space(arg.p_workspace_),
+                    arg.group_count_,
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.cde_element_op_);
+            };
+
+            if(has_main_k_block_loop)
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        for(std::size_t i = 0; i < arg.group_count_; i++)
+        {
+            const auto a_grid_desc_m_k_ = arg.contraction_multi_d_device_args_[i].a_grid_desc_m_k_;
+            const auto b_grid_desc_n_k_ = arg.contraction_multi_d_device_args_[i].b_grid_desc_n_k_;
+            const auto ds_grid_desc_m_n_ =
+                arg.contraction_multi_d_device_args_[i].ds_grid_desc_m_n_;
+            const auto e_grid_desc_m_n_ = arg.contraction_multi_d_device_args_[i].e_grid_desc_m_n_;
+            const auto a_grid_desc_ak0_m_ak1_ =
+                arg.contraction_multi_d_kernel_args_[i].a_grid_desc_ak0_m_ak1_;
+            const auto b_grid_desc_bk0_n_bk1_ =
+                arg.contraction_multi_d_kernel_args_[i].b_grid_desc_bk0_n_bk1_;
+            const auto ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                arg.contraction_multi_d_kernel_args_[i]
+                    .ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+            const auto e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                arg.contraction_multi_d_kernel_args_[i]
+                    .e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+            const auto block_2_etile_map_ =
+                arg.contraction_multi_d_kernel_args_[i].block_2_etile_map_;
+
+            const auto a_mz_stride_  = arg.contraction_multi_d_device_args_[i].a_mz_stride_;
+            const auto a_kz_stride_  = arg.contraction_multi_d_device_args_[i].a_kz_stride_;
+            const auto b_nz_stride_  = arg.contraction_multi_d_device_args_[i].b_nz_stride_;
+            const auto b_kz_stride_  = arg.contraction_multi_d_device_args_[i].b_kz_stride_;
+            const auto ds_nz_stride_ = arg.contraction_multi_d_device_args_[i].ds_nz_stride_;
+            const auto e_nz_stride_  = arg.contraction_multi_d_device_args_[i].e_nz_stride_;
+
+            if(!GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                            b_grid_desc_n_k_,
+                                            ds_grid_desc_m_n_,
+                                            e_grid_desc_m_n_,
+                                            block_2_etile_map_))
+            {
+                return false;
+            }
+
+            // check vector access
+            static_assert((ABlockTransferSrcVectorDim == 1 || ABlockTransferSrcVectorDim == 2) &&
+                              (BBlockTransferSrcVectorDim == 1 || BBlockTransferSrcVectorDim == 2),
+                          "wrong!");
+
+            // vector memory access of A: could be on M or AK1 dimension
+            if constexpr(ABlockTransferSrcVectorDim == 1)
+            {
+                if(!(a_mz_stride_ == 1 &&
+                     a_grid_desc_ak0_m_ak1_.GetLength(I1) % ABlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                if(!(a_kz_stride_ == 1 &&
+                     a_grid_desc_ak0_m_ak1_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
+            }
+
+            // vector memory access of B: could be on N or BK1 dimension
+            if constexpr(BBlockTransferSrcVectorDim == 1)
+            {
+                if(!(b_nz_stride_ == 1 &&
+                     b_grid_desc_bk0_n_bk1_.GetLength(I1) % BBlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                if(!(b_kz_stride_ == 1 &&
+                     b_grid_desc_bk0_n_bk1_.GetLength(I2) % BBlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
+            }
+
+            // vector memory access of Ds: always on NPerBlock dimension
+            bool valid_d_access = true;
+
+            static_for<0, NumDTensor, 1>{}([&](auto j) {
+                if(!(ds_nz_stride_[j] == 1 &&
+                     ds_grid_desc_mblock_mperblock_nblock_nperblock_[j].GetLength(I3) %
+                             CDEBlockTransferScalarPerVector_NPerBlock ==
+                         0))
+                {
+                    valid_d_access = false;
+                }
+            });
+
+            if(valid_d_access == false)
+            {
+                return false;
+            }
+
+            // vector memory access of E: always on NPerBlock dimension
+            if(!(e_nz_stride_ == 1 && e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
+                                              CDEBlockTransferScalarPerVector_NPerBlock ==
+                                          0))
+            {
+                return false;
+            }
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*> p_a_vec,
+                             std::vector<const void*> p_b_vec,
+                             std::vector<std::array<const void*, NumDTensor>> p_ds_vec,
+                             std::vector<void*> p_e_vec,
+                             std::vector<ContractionDesc<NumDTensor>> contraction_descs,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a_vec,
+                        p_b_vec,
+                        p_ds_vec,
+                        p_e_vec,
+                        contraction_descs,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b_vec,
+                        std::vector<std::array<const void*, NumDTensor>> p_ds_vec,
+                        std::vector<void*> p_e_vec,
+                        std::vector<ContractionDesc<NumDTensor>> contraction_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a_vec,
+                                          p_b_vec,
+                                          p_ds_vec,
+                                          p_e_vec,
+                                          contraction_descs,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedContractionMultipleD_Xdl_CShuffle"
+            << "<"
+            << NumDimM << ", "
+            << NumDimN << ", "
+            << NumDimK << ", "
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << ABlockTransferSrcVectorDim << ", "
+            << BBlockTransferSrcVectorDim
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->group_count_ *
+               sizeof(ContractionMultiDKernelArg);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000..682aba0
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
@@ -0,0 +1,1015 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+template <index_t NumDTensor>
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_bwd_data_multiple_d_xdl_cshuffle(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const index_t batch_count,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            const Block2ETileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+} // namespace
+
+// Conv backward data multiple D:
+//   input : output image A: [G, N, K, Ho, Wo]
+//   input : weight B: [G, K, C, Y, X],
+//   input : D0, D1, ... : [G, N, K, Ho, Wo]
+//   output : input image E: [G, N, C, Hi, Wi]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+template <index_t NDimSpatial,
+          typename ALayout,   // output image
+          typename BLayout,   // weight
+          typename DsLayout,  // bias
+          typename ELayout,   // input image
+          typename ADataType, // output image
+          typename BDataType, // weight
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,       // bias
+          typename EDataType,        // input image
+          typename AElementwiseOp,   // output image
+          typename BElementwiseOp,   // weight
+          typename CDEElementwiseOp, // C, bias, and input image
+          ConvolutionBackwardDataSpecialization ConvBackwardDataSpecialization,
+          bool DoPadGemmM,
+          bool DoPadGemmN,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
+    : public DeviceGroupedConvBwdDataMultipleD<NDimSpatial,
+                                               ALayout,    // output image
+                                               BLayout,    // weight
+                                               DsLayout,   // bias
+                                               ELayout,    // input image
+                                               ADataType,  // output image
+                                               BDataType,  // weight
+                                               DsDataType, // bias
+                                               EDataType,  // input image
+                                               AElementwiseOp,
+                                               BElementwiseOp,
+                                               CDEElementwiseOp>
+{
+    // FIXME
+    static_assert(NDimSpatial == 2, "wrong! only implemented for 2D now");
+
+    using DeviceOp = DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    // TODO make A/B datatype different
+    using ABDataType = ADataType;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto transform_conv_to_gemm =
+        TransformConvBwdDataToGemm_v1<NDimSpatial,
+                                      ConvBackwardDataSpecialization,
+                                      AK1,
+                                      BK1,
+                                      MPerBlock,
+                                      NPerBlock,
+                                      DoPadGemmM,
+                                      DoPadGemmN>{};
+
+    static auto GetDummyABDsEGridDescriptor()
+    {
+        const std::array<index_t, NDimSpatial + 3> dummy_tensor_lengths = {1};
+        const std::array<index_t, NDimSpatial + 3> dummy_tensor_strides = {1};
+        const std::array<index_t, NDimSpatial> dummy_spatial_lengths    = {1};
+
+        const auto a_grid_desc_ak0_m_ak1 =
+            transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
+                dummy_tensor_lengths,
+                dummy_tensor_strides,
+                dummy_tensor_lengths,
+                dummy_tensor_strides,
+                dummy_tensor_lengths,
+                dummy_tensor_strides,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths);
+
+        const auto b_grid_desc_bk0_n_bk1 =
+            transform_conv_to_gemm.template MakeBDescriptor_BK0_N_BK1<BLayout>(
+                dummy_tensor_lengths,
+                dummy_tensor_strides,
+                dummy_tensor_lengths,
+                dummy_tensor_strides,
+                dummy_tensor_lengths,
+                dummy_tensor_strides,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths);
+
+        const auto ds_grid_desc_m_n = generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return transform_conv_to_gemm.template MakeCDescriptor_M_N<DLayout>(
+                    dummy_tensor_lengths,
+                    dummy_tensor_strides,
+                    dummy_tensor_lengths,
+                    dummy_tensor_strides,
+                    dummy_tensor_lengths,
+                    dummy_tensor_strides,
+                    dummy_spatial_lengths,
+                    dummy_spatial_lengths,
+                    dummy_spatial_lengths,
+                    dummy_spatial_lengths,
+                    dummy_spatial_lengths);
+            },
+            Number<NumDTensor>{});
+
+        const auto e_grid_desc_m_n =
+            transform_conv_to_gemm.template MakeCDescriptor_M_N<ELayout>(dummy_tensor_lengths,
+                                                                         dummy_tensor_strides,
+                                                                         dummy_tensor_lengths,
+                                                                         dummy_tensor_strides,
+                                                                         dummy_tensor_lengths,
+                                                                         dummy_tensor_strides,
+                                                                         dummy_spatial_lengths,
+                                                                         dummy_spatial_lengths,
+                                                                         dummy_spatial_lengths,
+                                                                         dummy_spatial_lengths,
+                                                                         dummy_spatial_lengths);
+
+        return make_tuple(
+            a_grid_desc_ak0_m_ak1, b_grid_desc_bk0_n_bk1, ds_grid_desc_m_n, e_grid_desc_m_n);
+    }
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOp,
+        BElementwiseOp,
+        CDEElementwiseOp,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    template <typename Desc_K0_M_K1>
+    static auto transform_k0_m_k1_to_m_k(const Desc_K0_M_K1& desc_k0_m_k1)
+    {
+        const auto grid_desc_m_k = transform_tensor_descriptor(
+            desc_k0_m_k1,
+            make_tuple(make_pass_through_transform(desc_k0_m_k1.GetLength(I1)),
+                       make_merge_transform(
+                           make_tuple(desc_k0_m_k1.GetLength(I0), desc_k0_m_k1.GetLength(I2)))),
+            make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return grid_desc_m_k;
+    }
+
+    // desc
+    using ABDsEGridDesc = decltype(GetDummyABDsEGridDescriptor());
+
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<tuple_element_t<0, ABDsEGridDesc>>;
+    using BGridDesc_BK0_N_BK1 = remove_cvref_t<tuple_element_t<1, ABDsEGridDesc>>;
+    using DsGridDesc_M_N      = remove_cvref_t<tuple_element_t<2, ABDsEGridDesc>>;
+    using EGridDesc_M_N       = remove_cvref_t<tuple_element_t<3, ABDsEGridDesc>>;
+
+    using AGridDesc_M_K = decltype(transform_k0_m_k1_to_m_k(AGridDesc_AK0_M_AK1{}));
+    using BGridDesc_N_K = decltype(transform_k0_m_k1_to_m_k(BGridDesc_BK0_N_BK1{}));
+
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = decltype(
+        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}));
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = decltype(
+        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}));
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,                                 // output image
+                 const void* p_b,                                 // weight
+                 const std::array<const void*, NumDTensor>& p_ds, // bias
+                 void* p_e,                                       // input image
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_c_wis_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOp& a_element_op,
+                 const BElementwiseOp& b_element_op,
+                 const CDEElementwiseOp& cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              num_group_{a_g_n_k_wos_lengths[0]},
+              num_gemm_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_g_n_k_wos_lengths_{a_g_n_k_wos_lengths},
+              a_g_n_k_wos_strides_{a_g_n_k_wos_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              ds_g_n_c_wis_lengths_{ds_g_n_c_wis_lengths},
+              ds_g_n_c_wis_strides_{ds_g_n_c_wis_strides},
+              e_g_n_c_wis_lengths_{e_g_n_c_wis_lengths},
+              e_g_n_c_wis_strides_{e_g_n_c_wis_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // populate Ds pointer
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+            });
+
+            // A/B/Ds/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_k_wos_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_c_wis_strides[0];
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_c_wis_strides[i][0];
+            });
+
+            // problem definition
+            const index_t Y = b_g_k_c_xs_lengths[3];
+            const index_t X = b_g_k_c_xs_lengths[4];
+
+            const index_t ConvStrideH = conv_filter_strides_[0];
+            const index_t ConvStrideW = conv_filter_strides_[1];
+
+            const index_t ConvDilationH = conv_filter_dilations_[0];
+            const index_t ConvDilationW = conv_filter_dilations_[1];
+
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            // number of GEMM
+            num_gemm_ = YTilde * XTilde;
+
+            for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+            {
+                for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+                {
+                    // check slice is valid
+                    const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                    const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+                    if(YDotSlice * XDotSlice <= 0)
+                    {
+                        continue;
+                    }
+
+                    const auto a_grid_desc_ak0_m_ak1 =
+                        transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
+                            a_g_n_k_wos_lengths,
+                            a_g_n_k_wos_strides,
+                            b_g_k_c_xs_lengths,
+                            b_g_k_c_xs_strides,
+                            e_g_n_c_wis_lengths,
+                            e_g_n_c_wis_strides,
+                            conv_filter_strides,
+                            conv_filter_dilations,
+                            input_left_pads,
+                            input_right_pads,
+                            {i_ytilde, i_xtilde});
+
+                    const auto b_grid_desc_bk0_n_bk1 =
+                        transform_conv_to_gemm.template MakeBDescriptor_BK0_N_BK1<BLayout>(
+                            a_g_n_k_wos_lengths,
+                            a_g_n_k_wos_strides,
+                            b_g_k_c_xs_lengths,
+                            b_g_k_c_xs_strides,
+                            e_g_n_c_wis_lengths,
+                            e_g_n_c_wis_strides,
+                            conv_filter_strides,
+                            conv_filter_dilations,
+                            input_left_pads,
+                            input_right_pads,
+                            {i_ytilde, i_xtilde});
+
+                    DsGridDesc_M_N ds_grid_desc_m_n;
+
+                    // populate Ds desc
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                        ds_grid_desc_m_n(i) =
+                            transform_conv_to_gemm.template MakeCDescriptor_M_N<DLayout>(
+                                a_g_n_k_wos_lengths,
+                                a_g_n_k_wos_strides,
+                                b_g_k_c_xs_lengths,
+                                b_g_k_c_xs_strides,
+                                ds_g_n_c_wis_lengths[i],
+                                ds_g_n_c_wis_strides[i],
+                                conv_filter_strides,
+                                conv_filter_dilations,
+                                input_left_pads,
+                                input_right_pads,
+                                {i_ytilde, i_xtilde});
+                    });
+
+                    const auto e_grid_desc_m_n =
+                        transform_conv_to_gemm.template MakeCDescriptor_M_N<ELayout>(
+                            a_g_n_k_wos_lengths,
+                            a_g_n_k_wos_strides,
+                            b_g_k_c_xs_lengths,
+                            b_g_k_c_xs_strides,
+                            e_g_n_c_wis_lengths,
+                            e_g_n_c_wis_strides,
+                            conv_filter_strides,
+                            conv_filter_dilations,
+                            input_left_pads,
+                            input_right_pads,
+                            {i_ytilde, i_xtilde});
+
+                    // desc for problem definition
+                    const auto a_grid_desc_m_k = transform_k0_m_k1_to_m_k(a_grid_desc_ak0_m_ak1);
+                    const auto b_grid_desc_n_k = transform_k0_m_k1_to_m_k(b_grid_desc_bk0_n_bk1);
+
+                    a_grid_desc_m_k_container_.push_back(a_grid_desc_m_k);
+                    b_grid_desc_n_k_container_.push_back(b_grid_desc_n_k);
+                    ds_grid_desc_m_n_container_.push_back(ds_grid_desc_m_n);
+                    e_grid_desc_m_n_container_.push_back(e_grid_desc_m_n);
+
+                    // desc for blockwise copy
+                    a_grid_desc_ak0_m_ak1_container_.push_back(a_grid_desc_ak0_m_ak1);
+                    b_grid_desc_bk0_n_bk1_container_.push_back(b_grid_desc_bk0_n_bk1);
+
+                    // block-to-e-tile-map
+                    auto block_2_etile_map =
+                        GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n);
+
+                    block_2_etile_map_container_.push_back(block_2_etile_map);
+
+                    if(GridwiseGemm::CheckValidity(a_grid_desc_m_k,
+                                                   b_grid_desc_n_k,
+                                                   ds_grid_desc_m_n,
+                                                   e_grid_desc_m_n,
+                                                   block_2_etile_map))
+                    {
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock_container_.push_back(
+                            GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                                ds_grid_desc_m_n));
+
+                        e_grid_desc_mblock_mperblock_nblock_nperblock_container_.push_back(
+                            GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                                e_grid_desc_m_n));
+                    }
+                }
+            }
+        }
+
+        void Print() const
+        {
+            for(index_t i = 0; i < num_gemm_; i++)
+            {
+                std::cout << "a_grid_desc_ak0_m_ak1_container_"
+                          << a_grid_desc_ak0_m_ak1_container_[i] << std::endl;
+
+                std::cout << "b_grid_desc_bk0_n_bk1_container_"
+                          << b_grid_desc_bk0_n_bk1_container_[i] << std::endl;
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    std::cout << "ds_grid_desc_mblock_mperblock_nblock_nperblock_container_"
+                              << ds_grid_desc_mblock_mperblock_nblock_nperblock_container_[i][j]
+                              << std::endl;
+                });
+
+                std::cout << "e_grid_desc_mblock_mperblock_nblock_nperblock_container_"
+                          << e_grid_desc_mblock_mperblock_nblock_nperblock_container_[i]
+                          << std::endl;
+            }
+        }
+
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptor for problem definition
+        index_t num_group_;
+        index_t num_gemm_;
+        std::vector<AGridDesc_M_K> a_grid_desc_m_k_container_;
+        std::vector<BGridDesc_N_K> b_grid_desc_n_k_container_;
+        std::vector<DsGridDesc_M_N> ds_grid_desc_m_n_container_;
+        std::vector<EGridDesc_M_N> e_grid_desc_m_n_container_;
+
+        // tensor descriptor for block-wise copy
+        std::vector<AGridDesc_AK0_M_AK1> a_grid_desc_ak0_m_ak1_container_;
+        std::vector<BGridDesc_BK0_N_BK1> b_grid_desc_bk0_n_bk1_container_;
+        std::vector<DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_container_;
+        std::vector<EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>
+            e_grid_desc_mblock_mperblock_nblock_nperblock_container_;
+
+        // block-to-e-tile map
+        std::vector<Block2ETileMap> block_2_etile_map_container_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOp a_element_op_;
+        BElementwiseOp b_element_op_;
+        CDEElementwiseOp cde_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_c_wis_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            float ave_time = 0;
+
+            for(index_t i = 0; i < arg.num_gemm_; i++)
+            {
+                if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_container_[i],
+                                                arg.b_grid_desc_n_k_container_[i],
+                                                arg.ds_grid_desc_m_n_container_[i],
+                                                arg.e_grid_desc_m_n_container_[i],
+                                                arg.block_2_etile_map_container_[i]))
+                {
+                    throw std::runtime_error("wrong! device_op has invalid setting");
+                }
+
+                const index_t grid_size = arg.block_2_etile_map_container_[i].CalculateGridSize(
+                                              arg.e_grid_desc_m_n_container_[i]) *
+                                          arg.num_group_;
+
+                const auto GemmK = arg.a_grid_desc_m_k_container_[i].GetLength(I1);
+
+                auto launch_kernel = [&](auto has_main_k_block_loop) {
+                    constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                    const auto kernel = kernel_grouped_conv_bwd_data_multiple_d_xdl_cshuffle<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        typename GridwiseGemm::DsGridPointer,
+                        EDataType,
+                        AElementwiseOp,
+                        BElementwiseOp,
+                        CDEElementwiseOp,
+                        DeviceOp::AGridDesc_AK0_M_AK1,
+                        DeviceOp::BGridDesc_BK0_N_BK1,
+                        DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                        DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                        Block2ETileMap,
+                        ComputePtrOffsetOfStridedBatch<NumDTensor>,
+                        has_main_loop>;
+
+                    return launch_and_time_kernel(
+                        stream_config,
+                        kernel,
+                        dim3(grid_size),
+                        dim3(BlockSize),
+                        0,
+                        arg.p_a_grid_,
+                        arg.p_b_grid_,
+                        arg.p_ds_grid_,
+                        arg.p_e_grid_,
+                        arg.a_element_op_,
+                        arg.b_element_op_,
+                        arg.cde_element_op_,
+                        arg.a_g_n_k_wos_lengths_[0], // Group count
+                        arg.a_grid_desc_ak0_m_ak1_container_[i],
+                        arg.b_grid_desc_bk0_n_bk1_container_[i],
+                        arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_container_[i],
+                        arg.e_grid_desc_mblock_mperblock_nblock_nperblock_container_[i],
+                        arg.block_2_etile_map_container_[i],
+                        arg.compute_ptr_offset_of_batch_);
+                };
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(GemmK))
+                {
+                    ave_time += launch_kernel(integral_constant<bool, true>{});
+                }
+                else
+                {
+                    ave_time += launch_kernel(integral_constant<bool, false>{});
+                }
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        const index_t ConvK = arg.b_g_k_c_xs_lengths_[1];
+        const index_t ConvC = arg.b_g_k_c_xs_lengths_[2];
+
+        // Specifialization
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 pad = 0 conv
+            for(int i = 0; i < NDimSpatial; i++)
+            {
+                if(!(arg.b_g_k_c_xs_lengths_[3 + i] == 1 && arg.conv_filter_strides_[i] == 1 &&
+                     arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // vector load for A matrix from global memory to LDS
+        if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNHWK>)
+        {
+            if(!(ABlockTransferSrcVectorDim == 2 && ConvK % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // vector load for B matrix from global memory to LDS
+        if constexpr(is_same_v<BLayout, tensor_layout::convolution::GKYXC>)
+        {
+            if(!(BBlockTransferSrcVectorDim == 1 && ConvC % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // vector store for Ds
+        bool ds_valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+            if constexpr(is_same_v<DLayout, tensor_layout::convolution::GNHWC> ||
+                         is_same_v<DLayout, tensor_layout::convolution::NHWGC> ||
+                         is_same_v<DLayout, tensor_layout::convolution::G_NHW_C> ||
+                         is_same_v<DLayout, tensor_layout::convolution::GC> ||
+                         is_same_v<DLayout, tensor_layout::convolution::G_C>)
+            {
+                // vector load D matrix from global memory
+                if(!(ConvC % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+                {
+                    ds_valid = false;
+                }
+            }
+            else
+            {
+                ds_valid = false;
+            }
+        });
+
+        if(!ds_valid)
+        {
+            return false;
+        }
+
+        // vector store for E
+        if constexpr(is_same_v<ELayout, tensor_layout::convolution::GNHWC>)
+        {
+            // vector store C matrix into global memory
+            if(!(ConvC % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        for(std::size_t i = 0; i < arg.a_grid_desc_ak0_m_ak1_container_.size(); i++)
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_container_[i],
+                                            arg.b_grid_desc_n_k_container_[i],
+                                            arg.ds_grid_desc_m_n_container_[i],
+                                            arg.e_grid_desc_m_n_container_[i],
+                                            arg.block_2_etile_map_container_[i]))
+            {
+                return false;
+            }
+        }
+
+        return true;
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto
+    MakeArgument(const void* p_a,                                                 // output image
+                 const void* p_b,                                                 // weight
+                 const std::array<const void*, NumDTensor>& p_ds,                 // bias
+                 void* p_e,                                                       // input image
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_lengths, // output image
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_strides, // output image
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,  // weight
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,  // weight
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_c_wis_lengths, // bias
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_c_wis_strides,                                        // bias
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths, // input image
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_strides, // input image
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOp& a_element_op,
+                 const BElementwiseOp& b_element_op,
+                 const CDEElementwiseOp& cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_g_n_k_wos_lengths,
+                        a_g_n_k_wos_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_c_wis_lengths,
+                        ds_g_n_c_wis_strides,
+                        e_g_n_c_wis_lengths,
+                        e_g_n_c_wis_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,                                                 // output image
+        const void* p_b,                                                 // weight
+        const std::array<const void*, NumDTensor>& p_ds,                 // bias
+        void* p_e,                                                       // input image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_lengths, // output image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_strides, // output image
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,  // weight
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,  // weight
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+            ds_g_n_c_wis_lengths, // bias
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+            ds_g_n_c_wis_strides,                                        // bias
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths, // input image
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_strides, // input image
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOp& a_element_op,
+        const BElementwiseOp& b_element_op,
+        const CDEElementwiseOp& cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_g_n_k_wos_lengths,
+                                          a_g_n_k_wos_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_c_wis_lengths,
+                                          ds_g_n_c_wis_strides,
+                                          e_g_n_c_wis_lengths,
+                                          e_g_n_c_wis_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << getConvBackwardDataSpecializationString(ConvBackwardDataSpecialization)
+            << ">";
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp
new file mode 100644
index 0000000..d9e7b54
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp
@@ -0,0 +1,1244 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+struct ComputePtrOffsetOfStridedBatch
+{
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideC_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    index_t BatchStrideC_;
+};
+
+} // namespace
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename AGridDesc_B_K0_M_K1,
+          typename BGridDesc_B_K0_N_K1,
+          typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2CTileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_xdlops_bwd_weight(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const index_t batch_count,
+            const AGridDesc_B_K0_M_K1 a_b_k0_m_k1_grid_desc,
+            const BGridDesc_B_K0_N_K1 b_b_k0_n_k1_grid_desc,
+            const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2CTileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
+
+    __shared__ FloatAB p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB)];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_c_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a_b_k0_m_k1_grid_desc,
+                                                  b_b_k0_n_k1_grid_desc,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_b_k0_m_k1_grid_desc;
+    ignore = b_b_k0_n_k1_grid_desc;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = batch_count;
+    ignore = block_2_ctile_map;
+    ignore = compute_ptr_offset_of_batch;
+
+    compute_ptr_offset_of_batch.GetAPtrOffset(0);
+    compute_ptr_offset_of_batch.GetBPtrOffset(0);
+    compute_ptr_offset_of_batch.GetCPtrOffset(0);
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionBackwardWeightSpecialization ConvBackwardWeightSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXdl,
+          ck::index_t NPerXdl,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
+struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle
+    : public DeviceGroupedConvBwdWeight<
+          NDimSpatial,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                        ck::tensor_layout::convolution::GNHWC,
+                                        ck::tensor_layout::convolution::GNDHWC>>,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                        ck::tensor_layout::convolution::GKYXC,
+                                        ck::tensor_layout::convolution::GKZYXC>>,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                        ck::tensor_layout::convolution::GNHWK,
+                                        ck::tensor_layout::convolution::GNDHWK>>,
+          InDataType,
+          WeiDataType,
+          OutDataType,
+          InElementwiseOperation,
+          WeiElementwiseOperation,
+          OutElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle;
+
+    using ADataType = OutDataType;
+    using BDataType = InDataType;
+    using CDataType = WeiDataType;
+
+    using AElementwiseOperation = OutElementwiseOperation;
+    using BElementwiseOperation = InElementwiseOperation;
+    using CElementwiseOperation = WeiElementwiseOperation;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    // Bytes per 32 lds bank: 32 * 4 bytes
+    static constexpr auto BankLength = 128;
+    static constexpr auto ElePerBank = BankLength / sizeof(ADataType);
+
+    // M1 & M0
+    static constexpr auto ABlockLdsM1PerBlock = ElePerBank / K1;
+    static constexpr auto ABlockLdsM0PerBlock = MPerBlock / ABlockLdsM1PerBlock;
+    static constexpr auto ABlockLdsM1Padding  = 4;
+
+    // N1 & N0
+    static constexpr auto BBlockLdsN1PerBlock = ElePerBank / K1;
+    static constexpr auto BBlockLdsN0PerBlock = NPerBlock / BBlockLdsN1PerBlock;
+    static constexpr auto BBlockLdsN1Padding  = 4;
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        ck::index_t N,
+        ck::index_t K,
+        ck::index_t C,
+        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+        std::array<ck::index_t, NDimSpatial> input_left_pads,
+        std::array<ck::index_t, NDimSpatial> input_right_pads,
+        ck::index_t batch_k)
+    {
+        using namespace ck;
+
+        const index_t Wi            = input_spatial_lengths[0];
+        const index_t Wo            = output_spatial_lengths[0];
+        const index_t X             = filter_spatial_lengths[0];
+        const index_t ConvStrideW   = conv_filter_strides[0];
+        const index_t ConvDilationW = conv_filter_dilations[0];
+        const index_t InLeftPadW    = input_left_pads[0];
+        const index_t InRightPadW   = input_right_pads[0];
+
+        const index_t GemmKTotal = N * Wo;
+        const index_t GemmM      = K;
+        const index_t GemmN      = C * X;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
+
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K));
+
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_gemmktotal_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wi, C));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K));
+            const auto in_n_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmktotal_gemmn_grid_desc =
+                transform_tensor_descriptor(in_n_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(X, C)),
+                                                       make_merge_transform(make_tuple(N, Wo))),
+                                            make_tuple(Sequence<1, 3>{}, Sequence<0, 2>{}),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        ck::index_t N,
+        ck::index_t K,
+        ck::index_t C,
+        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+        std::array<ck::index_t, NDimSpatial> input_left_pads,
+        std::array<ck::index_t, NDimSpatial> input_right_pads,
+        ck::index_t batch_k)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmKTotal = N * Ho * Wo;
+        const index_t GemmM      = K;
+        const index_t GemmN      = C * X * Y;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
+
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_gemmktotal_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Hi * Wi, C));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmktotal_gemmn_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        ck::index_t N,
+        ck::index_t K,
+        ck::index_t C,
+        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+        std::array<ck::index_t, NDimSpatial> input_left_pads,
+        std::array<ck::index_t, NDimSpatial> input_right_pads,
+        ck::index_t batch_k)
+    {
+        using namespace ck;
+
+        const index_t Di = input_spatial_lengths[0];
+        const index_t Hi = input_spatial_lengths[1];
+        const index_t Wi = input_spatial_lengths[2];
+
+        const index_t Do = output_spatial_lengths[0];
+        const index_t Ho = output_spatial_lengths[1];
+        const index_t Wo = output_spatial_lengths[2];
+
+        const index_t Z = filter_spatial_lengths[0];
+        const index_t Y = filter_spatial_lengths[1];
+        const index_t X = filter_spatial_lengths[2];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        const index_t ConvDilationD = conv_filter_dilations[0];
+        const index_t ConvDilationH = conv_filter_dilations[1];
+        const index_t ConvDilationW = conv_filter_dilations[2];
+
+        const index_t InLeftPadD = input_left_pads[0];
+        const index_t InLeftPadH = input_left_pads[1];
+        const index_t InLeftPadW = input_left_pads[2];
+
+        const index_t InRightPadD = input_right_pads[0];
+        const index_t InRightPadH = input_right_pads[1];
+        const index_t InRightPadW = input_right_pads[2];
+
+        const index_t GemmKTotal = N * Do * Ho * Wo;
+        const index_t GemmM      = K;
+        const index_t GemmN      = C * Z * X * Y;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
+
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
+
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_gemmktotal_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Di * Hi * Wi, C));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
+            const auto in_n_di_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_dip_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmktotal_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(Z, Y, X, C)),
+                           make_merge_transform(make_tuple(N, Do, Ho, Wo))),
+                make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+    } // function end
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<1>(
+            1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1}, 1);
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<2>(
+            1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, 1);
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<3>(1,
+                                                                  1,
+                                                                  1,
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  1);
+    }
+
+    // type convert descs
+    template <typename Desc_M0>
+    static auto PadDescriptor_M0_1d(Desc_M0 desc_m0, index_t gridSize, index_t blockSize)
+    {
+        const auto m0           = desc_m0.GetLength(I0);
+        const index_t loop_step = gridSize * blockSize * 4;
+        const auto pad          = math::integer_least_multiple(m0, loop_step) - m0;
+        const auto desc_m0_pad =
+            transform_tensor_descriptor(desc_m0,
+                                        make_tuple(make_right_pad_transform(m0, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return desc_m0_pad;
+    }
+
+    template <index_t Dim>
+    static auto MakeDescriptor_M0(const std::array<index_t, Dim>& shape,
+                                  const std::array<index_t, Dim>& stride,
+                                  index_t gridSize,
+                                  index_t blockSize)
+    {
+        auto tupleOfShape  = generate_tuple([&](auto I) { return shape[I]; }, Number<Dim>{});
+        auto tupleOfStride = generate_tuple([&](auto I) { return stride[I]; }, Number<Dim>{});
+
+        // nd desc - [s0, s1, s2, ...]
+        const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
+
+        // merge nd to 1d desc - [s0 * s1 * ...]
+        if constexpr(Dim > 1)
+        {
+            const auto desc_m0 = transform_tensor_descriptor(
+                desc,
+                make_tuple(make_merge_transform(tupleOfShape)),
+                make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<Dim>{})),
+                make_tuple(Sequence<0>{}));
+
+            return PadDescriptor_M0_1d(desc_m0, gridSize, blockSize);
+        }
+        else
+            return PadDescriptor_M0_1d(desc, gridSize, blockSize);
+    }
+
+    using GridDesc_M0 = decltype(MakeDescriptor_M0<1>({1}, {1}, 1, 1));
+
+    using ABCGridDescs = decltype(GetABCGridDesc<NDimSpatial>());
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::AtomicAdd,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXdl,
+        NPerXdl,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        ABlockLdsM1PerBlock,
+        ABlockLdsM0PerBlock,
+        ABlockLdsM1Padding,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        BBlockLdsN1PerBlock,
+        BBlockLdsN0PerBlock,
+        BBlockLdsN1Padding,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXdl,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        true,
+        true>;
+
+    // Argument
+    using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        decltype(GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}));
+
+    using Block2CTileMap =
+        decltype(GridwiseGemm::MakeCBlockClusterAdaptor(CGridDesc_M_N{}, 1, 1, 1));
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 ck::index_t G,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+                 std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+                 std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+                 std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+                 std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+                 std::array<ck::index_t, NDimSpatial> input_left_pads,
+                 std::array<ck::index_t, NDimSpatial> input_right_pads,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op,
+                 ck::index_t split_k)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_in_grid},
+              p_c_grid_{p_wei_grid},
+              a_grid_desc_kbatch_k0_m_k1_{},
+              b_grid_desc_kbatch_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{},
+              compute_ptr_offset_of_batch_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{out_element_op},
+              b_element_op_{in_element_op},
+              c_element_op_{wei_element_op},
+              Conv_G_{G},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              output_spatial_lengths_{output_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads},
+              k_batch_{split_k}
+        {
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                    N,
+                    K,
+                    C,
+                    input_spatial_lengths,
+                    filter_spatial_lengths,
+                    output_spatial_lengths,
+                    conv_filter_strides,
+                    conv_filter_dilations,
+                    input_left_pads,
+                    input_right_pads,
+                    k_batch_);
+
+            a_grid_desc_kbatch_k0_m_k1_ = descs[I0];
+            b_grid_desc_kbatch_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_            = descs[I2];
+
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
+
+            // A/B/C Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ =
+                N * K *
+                std::accumulate(begin(output_spatial_lengths),
+                                end(output_spatial_lengths),
+                                index_t{1},
+                                std::multiplies<>{});
+            compute_ptr_offset_of_batch_.BatchStrideB_ =
+                N * C *
+                std::accumulate(begin(input_spatial_lengths),
+                                end(input_spatial_lengths),
+                                index_t{1},
+                                std::multiplies<>{});
+            compute_ptr_offset_of_batch_.BatchStrideC_ =
+                K * C *
+                std::accumulate(begin(filter_spatial_lengths),
+                                end(filter_spatial_lengths),
+                                index_t{1},
+                                std::multiplies<>{});
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_kbatch_k0_m_k1_,
+                                           b_grid_desc_kbatch_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n_);
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_kbatch_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_kbatch_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock c_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        Block2CTileMap block_2_ctile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        index_t M01_;
+        index_t N01_;
+
+        InElementwiseOperation a_element_op_;
+        OutElementwiseOperation b_element_op_;
+        WeiElementwiseOperation c_element_op_;
+
+        // for checking IsSupportedArgument()
+        index_t Conv_G_;
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+        std::array<ck::index_t, NDimSpatial> output_spatial_lengths_;
+        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths_;
+        std::array<ck::index_t, NDimSpatial> conv_filter_strides_;
+        std::array<ck::index_t, NDimSpatial> input_left_pads_;
+        std::array<ck::index_t, NDimSpatial> input_right_pads_;
+        index_t k_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        void ShowInfo(const Argument& arg)
+        {
+            std::cout << "arg.a_grid_desc_kbatch_k0_m_k1_{"
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I2) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.b_grid_desc_kbatch_k0_n_k1_{"
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I0) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I1) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I2) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.c_grid_desc_m_n_{" << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                      << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+        }
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                            arg.b_grid_desc_kbatch_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.Conv_G_;
+
+            const auto K0 = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_batched_gemm_xdlops_bwd_weight<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    OutElementwiseOperation,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                    remove_reference_t<DeviceOp::Block2CTileMap>,
+                    ComputePtrOffsetOfStridedBatch,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_c_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.c_element_op_,
+                                              arg.Conv_G_,
+                                              arg.a_grid_desc_kbatch_k0_m_k1_,
+                                              arg.b_grid_desc_kbatch_k0_n_k1_,
+                                              arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_ctile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            if(has_main_k0_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 pad = 0 conv
+            for(int i = 0; i < NDimSpatial; i++)
+            {
+                if(!(arg.filter_spatial_lengths_[i] == 1 && arg.conv_filter_strides_[i] == 1 &&
+                     arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_K_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_C_ % CBlockTransferScalarPerVector_NWaveNPerXdl == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                           arg.b_grid_desc_kbatch_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             WeiDataType* p_wei_grid,
+                             const OutDataType* p_out_grid,
+                             ck::index_t G,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+                             std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+                             std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+                             std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+                             std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+                             std::array<ck::index_t, NDimSpatial> input_left_pads,
+                             std::array<ck::index_t, NDimSpatial> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op,
+                             ck::index_t split_k)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        G,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op,
+                        split_k};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        void* p_wei_grid,
+                        const void* p_out_grid,
+                        ck::index_t G,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+                        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+                        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+                        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+                        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+                        std::array<ck::index_t, NDimSpatial> input_left_pads,
+                        std::array<ck::index_t, NDimSpatial> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op,
+                        ck::index_t split_k) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<WeiDataType*>(p_wei_grid),
+                                          static_cast<const OutDataType*>(p_out_grid),
+                                          G,
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op,
+                                          split_k);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvBackwardWeightSpecializationString(ConvBackwardWeightSpecialization)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp
new file mode 100644
index 0000000..03185d5
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Grouped Convolution Forward:
+//   input : input image A[G, N, C, Hi, Wi],
+//   input : weight B[G, K, C, Y, X],
+//   input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
+//   output : output image E[G, N, K, Ho, Wo]
+//   output : R0[G, N, Ho, Wo], R1[G, N, Ho, Wo], ...
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+//   Q0 = reduce0(q_op0(E)), Q1 = reduce1(q_op0(E)), ...
+//   R0 = r_op0(Q0), R1 = r_op1(Q1), ...
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename RLayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation>
+struct DeviceGroupedConvFwdMultipleDMultipleR : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    static constexpr index_t NumRTensor = RsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        std::array<void*, NumRTensor> p_rs,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const QsElementwiseOperation& qs_element_op,
+        const RsElementwiseOperation& rs_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
new file mode 100644
index 0000000..8b54ee4
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
@@ -0,0 +1,1105 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+template <index_t NumDTensor, index_t NumRTensor>
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE,
+                                   Array<ck::index_t, NumRTensor> BatchStrideRs)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE),
+          BatchStrideRs_(BatchStrideRs)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    __host__ __device__ constexpr auto GetRsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumRTensor> rs_offset;
+        static_for<0, NumRTensor, 1>{}(
+            [&](auto i) { rs_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideRs_[i]); });
+        return rs_offset;
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+    Array<ck::index_t, NumRTensor> BatchStrideRs_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
+ * \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the
+ * computing of pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename RsPointer,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename RsGridDescriptor_MBlock_MPerBlock,
+          typename Block2ETileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batch_gemm_multiple_d_xdl_cshuffle(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            RsPointer p_rs_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const QsElementwiseOperation qs_element_op,
+            const RsElementwiseOperation rs_element_op,
+            const index_t batch_count,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            const RsGridDescriptor_MBlock_MPerBlock rs_grid_desc_mblock_mperblock,
+            const Block2ETileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+    const auto rs_batch_offset = compute_ptr_offset_of_batch.GetRsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    RsPointer p_rs_grid_grp;
+
+    static constexpr index_t NumRTensor = RsGridDescriptor_MBlock_MPerBlock::Size();
+
+    static_for<0, NumRTensor, 1>{}(
+        [&](auto i) { p_rs_grid_grp(i) = p_rs_grid[i] + rs_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_rs_grid_grp,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  qs_element_op,
+                                                  rs_element_op,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                                  rs_grid_desc_mblock_mperblock,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = p_rs_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = rs_grid_desc_mblock_mperblock;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = qs_element_op;
+    ignore = rs_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+} // namespace
+
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename RLayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ReduceAccDataType,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation,
+          typename ThreadReduceOperations,
+          typename RsGlobalMemoryDataOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          index_t RThreadTransferDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
+    : public DeviceGroupedConvFwdMultipleDMultipleR<NDimSpatial,
+                                                    ALayout,
+                                                    BLayout,
+                                                    DELayout,
+                                                    RLayout,
+                                                    ADataType,
+                                                    BDataType,
+                                                    DsDataType,
+                                                    EDataType,
+                                                    RsDataType,
+                                                    AElementwiseOperation,
+                                                    BElementwiseOperation,
+                                                    CDEElementwiseOperation,
+                                                    RsElementwiseOperation,
+                                                    QsElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    static constexpr index_t NumRTensor = RsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                            const std::array<index_t, NDimSpatial>& input_left_pads,
+                            const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        return in_gemmm_gemmk_desc;
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        return wei_gemmn_gemmk_desc;
+    }
+
+    template <typename ELay>
+    static auto
+    MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    template <typename Descriptor>
+    static auto GetPaddedRGridDescriptor(Descriptor descriptor, index_t MRaw)
+    {
+        const auto M    = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto MPad = M - MRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M
+            return transform_tensor_descriptor(
+                descriptor,
+                make_tuple(make_right_pad_transform(descriptor, MPad)),
+                make_tuple(Sequence<0>{}),
+                make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad M
+            return descriptor;
+        }
+    }
+
+    template <typename RLay,
+              typename std::enable_if<is_same_v<RLay, tensor_layout::convolution::GNW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::GNHW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::GNDHW>,
+                                      bool>::type = false>
+    static auto
+    MakeRGridDescriptor_M(const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+                          const std::array<index_t, NDimSpatial + 2>& /* r_g_n_wos_strides */)
+    {
+        const index_t N = r_g_n_wos_lengths[1];
+
+        const index_t NHoWo =
+            N * ck::accumulate_n<index_t>(
+                    r_g_n_wos_lengths.begin() + 2, NDimSpatial, 1, std::multiplies<>());
+
+        const auto r_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(NHoWo));
+
+        return GetPaddedRGridDescriptor(r_grid_desc_mraw, NHoWo);
+    }
+
+    template <typename RLay,
+              typename std::enable_if<is_same_v<RLay, tensor_layout::convolution::G_NW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::G_NHW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::G_NDHW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::NWG> ||
+                                          is_same_v<RLay, tensor_layout::convolution::NHWG> ||
+                                          is_same_v<RLay, tensor_layout::convolution::NDHWG>,
+                                      bool>::type = false>
+    static auto MakeRGridDescriptor_M(const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+                                      const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides)
+    {
+        const index_t N = r_g_n_wos_lengths[1];
+
+        const index_t WoStride = r_g_n_wos_strides[NDimSpatial + 2];
+
+        const index_t NHoWo =
+            N * ck::accumulate_n<index_t>(
+                    r_g_n_wos_lengths.begin() + 2, NDimSpatial, 1, std::multiplies<>());
+
+        const auto r_grid_desc_mraw =
+            make_naive_tensor_descriptor(make_tuple(NHoWo), make_tuple(WoStride));
+
+        return GetPaddedRGridDescriptor(r_grid_desc_mraw, NHoWo);
+    }
+
+    using AGridDesc_M_K = remove_cvref_t<decltype(
+        MakeAGridDescriptor_M_K<ALayout>({}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_N_K = remove_cvref_t<decltype(MakeBGridDescriptor_N_K<BLayout>({}, {}))>;
+    using EGridDesc_M_N = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<DELayout>({}, {}))>;
+    using RGridDesc_M   = remove_cvref_t<decltype(MakeRGridDescriptor_M<RLayout>({}, {}))>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleDMultipleR_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        ReduceAccDataType,
+        RsDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        QsElementwiseOperation,
+        RsElementwiseOperation,
+        ThreadReduceOperations,
+        InMemoryDataOperationEnum::Set,
+        RsGlobalMemoryDataOperation,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        EGridDesc_M_N,
+        RGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        RThreadTransferDstScalarPerVector_MPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+
+    using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,
+                 const void* p_b,
+                 const std::array<const void*, NumDTensor>& p_ds,
+                 void* p_e,
+                 std::array<void*, NumRTensor> p_rs,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CDEElementwiseOperation& cde_element_op,
+                 const QsElementwiseOperation& qs_element_op,
+                 const RsElementwiseOperation& rs_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              p_rs_grid_{}, // FIXME
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K<ALayout>(a_g_n_c_wis_lengths,
+                                                                          a_g_n_c_wis_strides,
+                                                                          b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides,
+                                                                          e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K<BLayout>(b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<DELayout>(e_g_n_k_wos_lengths,
+                                                                           e_g_n_k_wos_strides)},
+              r_grid_desc_m_{
+                  DeviceOp::MakeRGridDescriptor_M<RLayout>(r_g_n_wos_lengths, r_g_n_wos_strides)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              rs_grid_desc_mblock_mperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              compute_ptr_offset_of_batch_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              qs_element_op_{qs_element_op},
+              rs_element_op_{rs_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              ds_g_n_k_wos_lengths_{ds_g_n_k_wos_lengths},
+              ds_g_n_k_wos_strides_{ds_g_n_k_wos_strides},
+              e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
+              e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           e_grid_desc_m_n_,
+                                           r_grid_desc_m_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                // populate pointer, batch stride, desc for Ds
+                static_for<0, NumDTensor, 1>{}([&](auto i) {
+                    using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                    // D pointer
+                    p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+
+                    // D batch stride
+                    compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
+
+                    // D desc
+                    ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DELayout>(
+                        ds_g_n_k_wos_lengths[i], ds_g_n_k_wos_strides[i]);
+
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock_(i) =
+                        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            ds_grid_desc_m_n_(i));
+                });
+
+                // populate pointer for Rs
+                static_for<0, NumRTensor, 1>{}([&](auto i) {
+                    using RDataType = remove_cvref_t<tuple_element_t<i.value, RsDataType>>;
+
+                    // R pointer
+                    p_rs_grid_(i) = static_cast<RDataType*>(p_rs[i]);
+
+                    rs_grid_desc_mblock_mperblock_(i) =
+                        GridwiseGemm::MakeRGridDescriptor_MBlock_MPerBlock(r_grid_desc_m_);
+                });
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+        typename GridwiseGemm::RsGridPointer p_rs_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        EGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+        RGridDesc_M r_grid_desc_m_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        StaticallyIndexedArray<
+            typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+            NumDTensor>
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_; // FIXME: Ds desc may be of different
+                                                             // type from E
+        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        StaticallyIndexedArray<typename GridwiseGemm::RGridDescriptor_MBlock_MPerBlock, NumRTensor>
+            rs_grid_desc_mblock_mperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        ComputePtrOffsetOfStridedBatch<NumDTensor, NumRTensor> compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+        QsElementwiseOperation qs_element_op_;
+        RsElementwiseOperation rs_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.r_grid_desc_m_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) *
+                arg.a_g_n_c_wis_lengths_[0]; // Group count
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_batch_gemm_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    typename GridwiseGemm::RsGridPointer,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    QsElementwiseOperation,
+                    RsElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    ck::StaticallyIndexedArray<
+                        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                        NumDTensor>,
+                    typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    ck::StaticallyIndexedArray<
+                        typename GridwiseGemm::RGridDescriptor_MBlock_MPerBlock,
+                        NumRTensor>,
+                    Block2ETileMap,
+                    ComputePtrOffsetOfStridedBatch<NumDTensor, NumRTensor>,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.p_rs_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.qs_element_op_,
+                                              arg.rs_element_op_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.rs_grid_desc_mblock_mperblock_,
+                                              arg.block_2_etile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(get_device_name() == "gfx908")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else if(get_device_name() == "gfx90a")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        //  check vector access of Ds
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            // FIXME: layout
+            if constexpr(is_same_v<DELayout, ctc::G_NW_K> || is_same_v<DELayout, ctc::G_NHW_K> ||
+                         is_same_v<DELayout, ctc::G_NDHW_K> || is_same_v<DELayout, ctc::GNWK> ||
+                         is_same_v<DELayout, ctc::GNHWK> || is_same_v<DELayout, ctc::GNDHWK> ||
+                         is_same_v<DELayout, ctc::NWGK> || is_same_v<DELayout, ctc::NHWGK> ||
+                         is_same_v<DELayout, ctc::NDHWGK>)
+            {
+                const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
+
+                if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+                {
+                    valid = false;
+                }
+            }
+            else
+            {
+                valid = false;
+            }
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check vector access of E
+        if constexpr(is_same_v<DELayout, ctc::G_NW_K> || is_same_v<DELayout, ctc::G_NHW_K> ||
+                     is_same_v<DELayout, ctc::G_NDHW_K> || is_same_v<DELayout, ctc::GNWK> ||
+                     is_same_v<DELayout, ctc::GNHWK> || is_same_v<DELayout, ctc::GNDHWK> ||
+                     is_same_v<DELayout, ctc::NWGK> || is_same_v<DELayout, ctc::NHWGK> ||
+                     is_same_v<DELayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.e_g_n_k_wos_lengths_[2];
+
+            if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of R
+        if constexpr(!(is_same_v<RLayout, ctc::G_NW> || is_same_v<RLayout, ctc::G_NHW> ||
+                       is_same_v<RLayout, ctc::G_NDHW> || is_same_v<RLayout, ctc::GNW> ||
+                       is_same_v<RLayout, ctc::GNHW> || is_same_v<RLayout, ctc::GNDHW> ||
+                       is_same_v<RLayout, ctc::NWG> || is_same_v<RLayout, ctc::NHWG> ||
+                       is_same_v<RLayout, ctc::NDHWG>))
+        {
+            return false;
+        }
+
+        // check Gridwise GEMM
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.r_grid_desc_m_,
+                                           arg.block_2_etile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        std::array<void*, NumRTensor> p_rs,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const QsElementwiseOperation& qs_element_op,
+        const RsElementwiseOperation& rs_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        p_rs,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_k_wos_lengths,
+                        ds_g_n_k_wos_strides,
+                        e_g_n_k_wos_lengths,
+                        e_g_n_k_wos_strides,
+                        r_g_n_wos_lengths,
+                        r_g_n_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op,
+                        qs_element_op,
+                        rs_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        std::array<void*, NumRTensor> p_rs,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const QsElementwiseOperation& qs_element_op,
+        const RsElementwiseOperation& rs_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          p_rs,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_k_wos_lengths,
+                                          ds_g_n_k_wos_strides,
+                                          e_g_n_k_wos_lengths,
+                                          e_g_n_k_wos_strides,
+                                          r_g_n_wos_lengths,
+                                          r_g_n_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op,
+                                          qs_element_op,
+                                          rs_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvFwdMultipleD_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000..bb7a2f8
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,952 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+template <index_t NumDTensor>
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
+ * \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the
+ * computing of pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_fwd_multiple_d_xdl_cshuffle(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const index_t batch_count,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            const Block2ETileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+} // namespace
+
+//
+// @brief      Device Convolution operation.
+//
+// Supports:
+//  @li         Forward convolution with up to 3 spatial dimentions
+//  @li         Input tensor in GNWC data format
+//  @li         Weight tensor in GKXC data format
+//  @li         Output tensor in GNWK data format
+//
+// 1D:
+// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+// 2D:
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+// 3D:
+// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+//
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
+    : public DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                           ALayout,
+                                           BLayout,
+                                           DsLayout,
+                                           ELayout,
+                                           ADataType,
+                                           BDataType,
+                                           DsDataType,
+                                           EDataType,
+                                           AElementwiseOperation,
+                                           BElementwiseOperation,
+                                           CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                            const std::array<index_t, NDimSpatial>& input_left_pads,
+                            const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        return in_gemmm_gemmk_desc;
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        return wei_gemmn_gemmk_desc;
+    }
+
+    template <typename ELay>
+    static auto
+    MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(ds_g_n_k_wos_lengths[i],
+                                                                  ds_g_n_k_wos_strides[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // desc for problem definition
+    using AGridDesc_M_K  = remove_cvref_t<decltype(
+        MakeAGridDescriptor_M_K<ALayout>({}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_N_K  = remove_cvref_t<decltype(MakeBGridDescriptor_N_K<BLayout>({}, {}))>;
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
+    using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,
+                 const void* p_b,
+                 const std::array<const void*, NumDTensor>& p_ds,
+                 void* p_e,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CDEElementwiseOperation& cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              num_group_{a_g_n_c_wis_lengths[0]},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K<ALayout>(a_g_n_c_wis_lengths,
+                                                                          a_g_n_c_wis_strides,
+                                                                          b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides,
+                                                                          e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K<BLayout>(b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              compute_ptr_offset_of_batch_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              ds_g_n_k_wos_lengths_{ds_g_n_k_wos_lengths},
+              ds_g_n_k_wos_strides_{ds_g_n_k_wos_strides},
+              e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
+              e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
+
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+
+                // D batch stride
+                compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
+
+                // D desc
+                ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
+                    ds_g_n_k_wos_lengths[i], ds_g_n_k_wos_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        index_t num_group_;
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.num_group_;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_grouped_conv_fwd_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Block2ETileMap,
+                    ComputePtrOffsetOfStridedBatch<NumDTensor>,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_etile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(get_device_name() == "gfx908")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else if(get_device_name() == "gfx90a")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        //  check vector access of Ds
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+            // FIXME: layout
+            if constexpr(is_same_v<DLayout, ctc::G_NW_K> || is_same_v<DLayout, ctc::G_NHW_K> ||
+                         is_same_v<DLayout, ctc::G_NDHW_K> || is_same_v<DLayout, ctc::GNWK> ||
+                         is_same_v<DLayout, ctc::GNHWK> || is_same_v<DLayout, ctc::GNDHWK> ||
+                         is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
+                         is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::GK> ||
+                         is_same_v<DLayout, ctc::G_K>)
+            {
+                const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
+
+                if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+                {
+                    valid = false;
+                }
+            }
+            else
+            {
+                valid = false;
+            }
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check vector access of E
+        if constexpr(is_same_v<ELayout, ctc::G_NW_K> || is_same_v<ELayout, ctc::G_NHW_K> ||
+                     is_same_v<ELayout, ctc::G_NDHW_K> || is_same_v<ELayout, ctc::GNWK> ||
+                     is_same_v<ELayout, ctc::GNHWK> || is_same_v<ELayout, ctc::GNDHWK> ||
+                     is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
+                     is_same_v<ELayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.e_g_n_k_wos_lengths_[2];
+
+            if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check Gridwise GEMM
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.ds_grid_desc_m_n_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.block_2_etile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_k_wos_lengths,
+                        ds_g_n_k_wos_strides,
+                        e_g_n_k_wos_lengths,
+                        e_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_k_wos_lengths,
+                                          ds_g_n_k_wos_strides,
+                                          e_g_n_k_wos_lengths,
+                                          e_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvFwdMultipleD_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
new file mode 100644
index 0000000..aabcc73
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
@@ -0,0 +1,677 @@
+#pragma once
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename GemmDesc,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_gemm_xdl(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
+                                const index_t group_count,
+                                const AElementwiseOperation a_element_op,
+                                const BElementwiseOperation b_element_op,
+                                const CDEElementwiseOperation c_element_op)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const auto gemm_desc_ptr =
+        reinterpret_cast<const GemmDesc*>(cast_pointer_to_generic_address_space(gemm_descs_const));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+    while((!(block_id >= gemm_desc_ptr[group_id].BlockStart_ &&
+             block_id < gemm_desc_ptr[group_id].BlockEnd_)) &&
+          left <= right)
+    {
+        if(block_id < gemm_desc_ptr[group_id].BlockStart_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        gemm_desc_ptr[group_id].a_ptr_,
+        gemm_desc_ptr[group_id].b_ptr_,
+        gemm_desc_ptr[group_id].ds_ptr_,
+        gemm_desc_ptr[group_id].e_ptr_,
+        p_shared,
+        a_element_op,
+        b_element_op,
+        c_element_op,
+        gemm_desc_ptr[group_id].a_grid_desc_ak0_m_ak1_,
+        gemm_desc_ptr[group_id].b_grid_desc_bk0_n_bk1_,
+        gemm_desc_ptr[group_id].ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+        gemm_desc_ptr[group_id].e_grid_desc_mblock_mperblock_nblock_nperblock_,
+        gemm_desc_ptr[group_id].block_2_etile_map_);
+#else
+    ignore = gemm_descs_const;
+    ignore = group_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+#endif
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t NumPrefetch,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t KPerBlock,
+          ck::index_t AK1,
+          ck::index_t BK1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
+                                                        BLayout,
+                                                        DsLayout,
+                                                        ELayout,
+                                                        ADataType,
+                                                        BDataType,
+                                                        DsDataType,
+                                                        EDataType,
+                                                        AElementwiseOperation,
+                                                        BElementwiseOperation,
+                                                        CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedGemm_Xdl;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumPrefetch, // NumGemmKPrefetchStage
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+
+    struct GroupedGemmBlock2ETileMap
+    {
+        using Block2ETileMap =
+            remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+        GroupedGemmBlock2ETileMap()
+        {
+            block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{});
+            BlockStart_        = -1;
+        }
+
+        GroupedGemmBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n, ck::index_t BlockStart)
+        {
+            block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n);
+            BlockStart_        = BlockStart;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            return block_2_etile_map_.CalculateBottomIndex(
+                make_multi_index(idx_top[I0] - BlockStart_));
+        }
+
+        // it's actually E-Tile
+        template <typename CTileIdx, typename CTileDim>
+        __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                                 const CTileDim& c_tile_dim) const
+        {
+            return block_2_etile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+        }
+
+        __host__ bool CheckValidity(const EGridDesc_M_N& e_grid_desc_m_n) const
+        {
+            return block_2_etile_map_.CheckValidity(e_grid_desc_m_n);
+        }
+
+        Block2ETileMap block_2_etile_map_;
+        ck::index_t BlockStart_;
+    };
+
+    struct GemmBiasTransKernelArg
+    {
+        // pointers
+        const ADataType* a_ptr_;
+        const BDataType* b_ptr_;
+        typename GridwiseGemm::DsGridPointer ds_ptr_;
+        EDataType* e_ptr_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        GroupedGemmBlock2ETileMap block_2_etile_map_;
+        ck::index_t BlockStart_, BlockEnd_;
+    };
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(std::vector<const void*>& p_As,
+                 std::vector<const void*>& p_Bs,
+                 std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                 std::vector<void*>& p_Es,
+                 std::vector<GemmDesc>& gemm_descs,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation c_element_op)
+            : a_element_op_{a_element_op}, b_element_op_{b_element_op}, c_element_op_{c_element_op}
+        {
+            grid_size_ = 0;
+
+            group_count_ = ck::type_convert<ck::index_t>(gemm_descs.size());
+
+            if(!(group_count_ == ck::type_convert<ck::index_t>(p_As.size()) &&
+                 group_count_ == ck::type_convert<ck::index_t>(p_Bs.size()) &&
+                 group_count_ == ck::type_convert<ck::index_t>(p_Es.size())))
+            {
+                throw std::runtime_error("wrong! group_count_ != p_As/b/c.size");
+            }
+
+            gemm_desc_kernel_arg_.reserve(group_count_);
+
+            skipped_group_count_ = 0;
+
+            for(std::size_t i = 0; i < gemm_descs.size(); i++)
+            {
+                const index_t M = gemm_descs[i].M_;
+                const index_t N = gemm_descs[i].N_;
+                const index_t K = gemm_descs[i].K_;
+
+                if(M == 0)
+                {
+                    skipped_group_count_++;
+                    continue;
+                }
+
+                const index_t StrideA = gemm_descs[i].stride_A_;
+                const index_t StrideB = gemm_descs[i].stride_B_;
+                const index_t StrideC = gemm_descs[i].stride_C_;
+
+                // pointer
+                typename GridwiseGemm::DsGridPointer p_ds_grid{};
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
+
+                    p_ds_grid(j) = static_cast<const DDataType*>(p_Ds[i][j]);
+                });
+
+                // tensor descriptors for problem definiton
+                const auto a_grid_desc_m_k = DeviceOp::MakeAGridDescriptor_M_K(M, K, StrideA);
+                const auto b_grid_desc_n_k = DeviceOp::MakeBGridDescriptor_N_K(K, N, StrideB);
+
+                DsGridDesc_M_N ds_grid_desc_m_n;
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+                    ds_grid_desc_m_n(j) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
+                        M, N, gemm_descs[i].stride_Ds_[j]);
+                });
+
+                const auto e_grid_desc_m_n =
+                    DeviceOp::MakeEGridDescriptor_M_N<ELayout>(M, N, StrideC);
+
+                // tensor descriptors for block/thread-wise copy
+                const auto a_grid_desc_ak0_m_ak1 =
+                    GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k);
+
+                const auto b_grid_desc_bk0_n_bk1 =
+                    GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k);
+
+                const index_t grid_size_grp =
+                    GroupedGemmBlock2ETileMap(e_grid_desc_m_n, 0)
+                        .block_2_etile_map_.CalculateGridSize(e_grid_desc_m_n);
+
+                const index_t BlockStart = grid_size_;
+                const index_t BlockEnd   = grid_size_ + grid_size_grp;
+
+                grid_size_ += grid_size_grp;
+
+                // block-to-e-tile map
+                const auto block_2_etile_map =
+                    GroupedGemmBlock2ETileMap(e_grid_desc_m_n, BlockStart);
+
+                if(GridwiseGemm::CheckValidity(a_grid_desc_m_k,
+                                               b_grid_desc_n_k,
+                                               ds_grid_desc_m_n,
+                                               e_grid_desc_m_n,
+                                               block_2_etile_map))
+                {
+                    // tensor descriptors for block/thread-wise copy
+                    DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+                    static_for<0, NumDTensor, 1>{}([&](auto j) {
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                            GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                                ds_grid_desc_m_n[j]);
+                    });
+
+                    const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+                        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            e_grid_desc_m_n);
+
+                    gemm_desc_kernel_arg_.push_back(
+                        GemmBiasTransKernelArg{static_cast<const ADataType*>(p_As[i]),
+                                               static_cast<const BDataType*>(p_Bs[i]),
+                                               p_ds_grid,
+                                               static_cast<EDataType*>(p_Es[i]),
+                                               a_grid_desc_m_k,
+                                               b_grid_desc_n_k,
+                                               ds_grid_desc_m_n,
+                                               e_grid_desc_m_n,
+                                               a_grid_desc_ak0_m_ak1,
+                                               b_grid_desc_bk0_n_bk1,
+                                               ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                               e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                               block_2_etile_map,
+                                               BlockStart,
+                                               BlockEnd});
+                }
+            }
+        }
+
+        //  private:
+        index_t group_count_;
+        index_t skipped_group_count_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation c_element_op_;
+
+        std::vector<GemmBiasTransKernelArg> gemm_desc_kernel_arg_;
+
+        index_t grid_size_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            bool has_main_k_block_loop = true;
+
+            for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
+            {
+                std::cout << "group: " << i << " arg.a_grid_desc_ak0_m_ak1_{"
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I1)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2)
+                          << "}";
+
+                std::cout << ", arg.b_grid_desc_bk0_n_bk1_{"
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_bk0_n_bk1_.GetLength(I0)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_bk0_n_bk1_.GetLength(I1)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_bk0_n_bk1_.GetLength(I2)
+                          << "}";
+
+                std::cout << ", arg.e_grid_desc_m_n_{ "
+                          << arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_.GetLength(I1) << "}"
+                          << std::endl;
+
+                if(!GridwiseGemm::CheckValidity(arg.gemm_desc_kernel_arg_[i].a_grid_desc_m_k_,
+                                                arg.gemm_desc_kernel_arg_[i].b_grid_desc_n_k_,
+                                                arg.gemm_desc_kernel_arg_[i].ds_grid_desc_m_n_,
+                                                arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_,
+                                                arg.gemm_desc_kernel_arg_[i].block_2_etile_map_))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+                }
+
+                const auto K = arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                               arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(K) != has_main_k_block_loop)
+                {
+                    throw std::runtime_error("wrong! not all gemm has_main_k_block_loop");
+                }
+            }
+
+            hipGetErrorString(
+                hipMemcpy(arg.p_workspace_,
+                          arg.gemm_desc_kernel_arg_.data(),
+                          arg.gemm_desc_kernel_arg_.size() * sizeof(GemmBiasTransKernelArg),
+                          hipMemcpyHostToDevice));
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_grouped_gemm_xdl<GridwiseGemm,
+                                                            GemmBiasTransKernelArg,
+                                                            AElementwiseOperation,
+                                                            BElementwiseOperation,
+                                                            CDEElementwiseOperation,
+                                                            has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(arg.grid_size_),
+                    dim3(BlockSize),
+                    0,
+                    cast_pointer_to_constant_address_space(arg.p_workspace_),
+                    arg.gemm_desc_kernel_arg_.size(),
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.c_element_op_);
+            };
+
+            if(has_main_k_block_loop)
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if((ck::type_convert<ck::index_t>(arg.gemm_desc_kernel_arg_.size()) +
+            arg.skipped_group_count_) != arg.group_count_)
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*>& p_As,
+                             std::vector<const void*>& p_Bs,
+                             std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                             std::vector<void*>& p_Es,
+                             std::vector<GemmDesc> gemm_descs,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation c_element_op)
+    {
+        return Argument{
+            p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*>& p_As,
+                        std::vector<const void*>& p_Bs,
+                        std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                        std::vector<void*>& p_Es,
+                        std::vector<GemmDesc>& gemm_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(
+            p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedGemm_Xdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->group_count_ * sizeof(GemmBiasTransKernelArg);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp
new file mode 100644
index 0000000..dbeeb98
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp
@@ -0,0 +1,595 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_multiple_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumReduction,
+          typename InDataType,
+          typename AccDataType,
+          typename OutDataTypeTuple,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          typename OutDstVectorSizeSeq>
+struct DeviceMultipleReduceMultiBlock : public DeviceMultipleReduce<Rank,
+                                                                    NumReduceDim,
+                                                                    NumReduction,
+                                                                    InElementwiseOperationTuple,
+                                                                    AccElementwiseOperationTuple>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "Invalid thread cluster size assignments!");
+
+    static_assert((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                      (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(NumReduction == OutDataTypeTuple::Size() &&
+                      NumReduction == InElementwiseOperationTuple::Size() &&
+                      NumReduction == AccElementwiseOperationTuple::Size() &&
+                      NumReduction == OutDstVectorSizeSeq::Size(),
+                  "All tuple should have the same size as the number of Reductions!");
+
+    static_assert(sequence_all_of(OutDstVectorSizeSeq{},
+                                  [](auto vectorSize) {
+                                      return (MThreadSliceSize % vectorSize == 0);
+                                  }),
+                  "The OutDstVectorSize should completely divide the MThreadSliceSize!");
+
+    static constexpr bool CheckDataTypeTuple()
+    {
+        bool flag = true;
+
+        static_for<0, NumReduction, 1>{}([&](auto I) {
+            using OutDataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+            flag =
+                flag && ck::reduce::InMemoryDataOperatonSupportedOnDataType<OutMemoryDataOperation,
+                                                                            OutDataType>::value;
+        });
+
+        return flag;
+    };
+
+    static_assert(CheckDataTypeTuple(),
+                  "The OutDataType must support the specified OutMemoryDataOperation!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumInputDim  = Rank;
+    static constexpr index_t NumOutputDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim    = (NumInvariantDim == 0);
+
+    // So far, only AtomicAdd is considered, other Atomic Operation like AtomicMax can be added
+    // later
+    static constexpr bool use_multiblock =
+        (OutMemoryDataOperation == InMemoryDataOperationEnum::AtomicAdd);
+
+    static_assert(
+        ReduceOperation::IsCompatibleInMemoryDataOperation(OutMemoryDataOperation),
+        "The reduction accumulation operation must be compatible with the OutMemoryDataOperation!");
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto GenerateOutDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+
+                return static_cast<DataType*>(nullptr);
+            },
+            Number<NumReduction>{});
+    };
+
+    using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
+
+    static auto MakeSrc2dDescriptor(const std::array<index_t, NumInputDim>& inLengths,
+                                    const std::array<index_t, NumInputDim>& inStrides,
+                                    int blkGroupSize,
+                                    int numBlockTileIteration)
+    {
+        const auto tupleSrcLengths =
+            generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInputDim>{});
+        const auto tupleSrcStrides =
+            generate_tuple([&](auto I) { return inStrides[I]; }, Number<NumInputDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, NumInputDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths = generate_tuple(
+                    [&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
+                const auto invariantDimLengths =
+                    generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    static auto MakeDst1dDescriptor(const std::array<index_t, NumOutputDim>& outLengths,
+                                    const std::array<index_t, NumOutputDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumOutputDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumOutputDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumOutputDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto outPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto out_grid_desc_m_padded = transform_tensor_descriptor(
+            out_grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, outPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    static auto GenerateOutGrid1dDescTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                (void)I;
+                return MakeDst1dDescriptor(std::array<index_t, NumOutputDim>{},
+                                           std::array<index_t, NumOutputDim>{});
+            },
+            Number<NumReduction>{});
+    };
+
+    using InGridDesc_M_K      = decltype(MakeSrc2dDescriptor(
+        std::array<index_t, NumInputDim>{}, std::array<index_t, NumInputDim>{}, 1, 1));
+    using OutGridDesc_M_Tuple = decltype(GenerateOutGrid1dDescTuple());
+
+    static auto MakeDst1dDescriptorForBufferSet(const std::array<index_t, NumOutputDim>& outLengths,
+                                                const std::array<index_t, NumOutputDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumOutputDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumOutputDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumOutputDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto length = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto pad = math::integer_least_multiple(length, BlockSize) - length;
+
+        auto out_grid_desc_m_padded =
+            transform_tensor_descriptor(out_grid_desc_m,
+                                        make_tuple(make_right_pad_transform(length, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    static auto GenerateOutGrid1dDescTuple_2()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                (void)I;
+                return MakeDst1dDescriptorForBufferSet(std::array<index_t, NumOutputDim>{},
+                                                       std::array<index_t, NumOutputDim>{});
+            },
+            Number<NumReduction>{});
+    };
+
+    using OutGridDesc_M_Tuple_2 = decltype(GenerateOutGrid1dDescTuple_2());
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, NumInputDim>& inLengths,
+                 const std::array<index_t, NumInputDim>& inStrides,
+                 const std::array<index_t, NumOutputDim>& outLengths,
+                 const std::array<std::array<index_t, NumOutputDim>, NumReduction>& outStridesArray,
+                 const std::array<int, NumReduceDim>& reduceDims,
+                 const std::array<const void*, NumReduction>& alphas,
+                 const std::array<const void*, NumReduction>& betas,
+                 const void* in_dev,
+                 const std::array<void*, NumReduction>& out_dev_buffers,
+                 const InElementwiseOperationTuple in_elementwise_op_tuple,
+                 const AccElementwiseOperationTuple acc_elementwise_op_tuple)
+            : outLengths_{outLengths},
+              outStridesArray_{outStridesArray},
+              in_elementwise_op_tuple_{in_elementwise_op_tuple},
+              acc_elementwise_op_tuple_{acc_elementwise_op_tuple}
+        {
+            inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
+
+            for(size_t i = 0; i < NumReduction; i++)
+            {
+                alpha_values_(i) = *static_cast<const AccDataType*>(alphas[i]);
+                beta_values_(i)  = *static_cast<const AccDataType*>(betas[i]);
+            };
+
+            in_dev_ = static_cast<const InDataType*>(in_dev);
+
+            out_dev_buffers_ = generate_tuple(
+                [&](auto iR) {
+                    using OutDataTypePointer =
+                        remove_cvref_t<decltype(OutDataTypePointerTuple{}[iR])>;
+                    using OutDataType = remove_cvref_t<remove_pointer_t<OutDataTypePointer>>;
+                    return static_cast<OutDataType*>(out_dev_buffers[iR]);
+                },
+                Number<NumReduction>{});
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            if constexpr(use_multiblock)
+            {
+
+                int iterations = 1;
+                while(true)
+                {
+                    int testBlkGroupSize =
+                        (reduce_total_length + (K_BlockTileSize * iterations) - 1) /
+                        (K_BlockTileSize * iterations);
+
+                    // we want the blkGroupSize be not more than 128
+                    if(testBlkGroupSize <= 128)
+                        break;
+
+                    iterations++;
+                };
+
+                blkGroupSize = (reduce_total_length + (K_BlockTileSize * iterations) - 1) /
+                               (K_BlockTileSize * iterations);
+
+                numBlockTileIteration = iterations;
+            }
+            else
+            {
+                blkGroupSize = 1;
+                numBlockTileIteration =
+                    (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+            };
+
+            in_grid_desc_m_k =
+                MakeSrc2dDescriptor(inLengths_, inStrides_, blkGroupSize, numBlockTileIteration);
+
+            out_grid_desc_m_tuple = generate_tuple(
+                [&](auto I) { return MakeDst1dDescriptor(outLengths, outStridesArray[I]); },
+                Number<NumReduction>{});
+
+            out_grid_desc_m_tuple_2 = generate_tuple(
+                [&](auto I) {
+                    return MakeDst1dDescriptorForBufferSet(outLengths, outStridesArray[I]);
+                },
+                Number<NumReduction>{});
+
+            gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                       M_BlockTileSize * blkGroupSize;
+
+            gridSize_pre =
+                math::integer_least_multiple(invariant_total_length, BlockSize) / BlockSize;
+        }
+
+        std::array<index_t, NumInputDim> inLengths_;
+        std::array<index_t, NumInputDim> inStrides_;
+
+        std::array<index_t, NumOutputDim> outLengths_;
+        std::array<std::array<index_t, NumOutputDim>, NumReduction> outStridesArray_;
+
+        Array<AccDataType, NumReduction> alpha_values_;
+        Array<AccDataType, NumReduction> beta_values_;
+
+        const InDataType* in_dev_;
+        OutDataTypePointerTuple out_dev_buffers_;
+
+        InGridDesc_M_K in_grid_desc_m_k;
+        OutGridDesc_M_Tuple out_grid_desc_m_tuple;
+        OutGridDesc_M_Tuple_2 out_grid_desc_m_tuple_2;
+
+        InElementwiseOperationTuple in_elementwise_op_tuple_;
+        AccElementwiseOperationTuple acc_elementwise_op_tuple_;
+
+        long_index_t invariant_total_length;
+        long_index_t reduce_total_length;
+
+        int blkGroupSize;
+        int numBlockTileIteration;
+        size_t gridSize;
+
+        size_t gridSize_pre;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            using GridwiseMultipleReduce =
+                GridwiseMultipleReduction_mk_to_m_multiblock<NumReduction,
+                                                             InDataType,
+                                                             OutDataTypePointerTuple,
+                                                             AccDataType,
+                                                             InGridDesc_M_K,
+                                                             OutGridDesc_M_Tuple,
+                                                             ReduceOperation,
+                                                             InElementwiseOperationTuple,
+                                                             AccElementwiseOperationTuple,
+                                                             OutMemoryDataOperation,
+                                                             PropagateNan,
+                                                             BlockSize,
+                                                             MThreadClusterSize,
+                                                             KThreadClusterSize,
+                                                             MThreadSliceSize,
+                                                             KThreadSliceSize,
+                                                             InSrcVectorDim,
+                                                             InSrcVectorSize,
+                                                             OutDstVectorSizeSeq>;
+
+            const auto kernel_main =
+                kernel_multiple_reduce_multiblock<GridwiseMultipleReduce,
+                                                  NumReduction,
+                                                  InDataType,
+                                                  OutDataTypePointerTuple,
+                                                  AccDataType,
+                                                  InGridDesc_M_K,
+                                                  OutGridDesc_M_Tuple,
+                                                  InElementwiseOperationTuple,
+                                                  AccElementwiseOperationTuple>;
+
+            float avg_time = 0;
+
+            if constexpr(use_multiblock)
+            {
+                auto identity_values = generate_tuple(
+                    [&](auto iR) {
+                        using OutDataType = remove_cvref_t<decltype(OutDataTypeTuple{}[iR])>;
+                        return ck::reduce::GetIdentityValueForInMemoryDataOperation<OutDataType>(
+                            OutMemoryDataOperation);
+                    },
+                    Number<NumReduction>{});
+
+                const auto kernel_pre = kernel_multiple_buffer_set_value<OutGridDesc_M_Tuple_2,
+                                                                         NumReduction,
+                                                                         BlockSize,
+                                                                         OutDataTypePointerTuple,
+                                                                         OutDataTypeTuple>;
+
+                avg_time += launch_and_time_kernel(stream_config,
+                                                   kernel_pre,
+                                                   dim3(arg.gridSize_pre),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.out_grid_desc_m_tuple_2,
+                                                   arg.out_dev_buffers_,
+                                                   identity_values);
+            };
+
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.in_grid_desc_m_k,
+                                               arg.out_grid_desc_m_tuple,
+                                               arg.in_elementwise_op_tuple_,
+                                               arg.acc_elementwise_op_tuple_,
+                                               arg.blkGroupSize,
+                                               arg.numBlockTileIteration,
+                                               arg.alpha_values_,
+                                               arg.in_dev_,
+                                               arg.beta_values_,
+                                               arg.out_dev_buffers_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if constexpr(use_multiblock)
+        {
+            for(size_t i = 0; i < pArg->beta_values_.Size(); i++)
+                if(pArg->beta_values_[i] != 0.0f)
+                    return (false);
+        };
+
+        if constexpr(InSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return (false);
+            }
+            else
+            {
+                if(pArg->inStrides_[NumInvariantDim - 1] != 1 && InSrcVectorSize != 1)
+                    return (false);
+
+                if(pArg->inLengths_[NumInvariantDim - 1] % InSrcVectorSize != 0)
+                    return (false);
+            };
+        }
+        else
+        {
+            if(pArg->inStrides_[Rank - 1] != 1 && InSrcVectorSize != 1)
+                return (false);
+
+            if(pArg->inLengths_[Rank - 1] % InSrcVectorSize != 0)
+                return (false);
+        };
+        // To improve
+        bool valid = true;
+        static_for<0, NumReduction, 1>{}([&](auto I) {
+            if(pArg->outStridesArray_[I.value][NumOutputDim - 1] != 1 &&
+               OutDstVectorSizeSeq::At(I) != 1)
+                valid = false;
+
+            if(pArg->outLengths_[NumOutputDim - 1] % OutDstVectorSizeSeq::At(I) != 0)
+                valid = false;
+        });
+
+        if(!valid)
+            return (false);
+
+        if constexpr(use_multiblock)
+        {
+            // blkGroupSize of 1 should be handled by Blockwise path using
+            // InMemoryDataOperationEnum::Set
+            if(pArg->blkGroupSize == 1)
+                return (false);
+
+            // This is very strong restriction, but needed to avoid some failure
+            if(pArg->outLengths_[NumOutputDim - 1] % M_BlockTileSize != 0)
+                return (false);
+        }
+        else
+        {
+            // cases with very small reduce_total_length should be handled by ThreadWise kernel
+            if(pArg->reduce_total_length / KThreadSliceSize < 2)
+                return (false);
+        };
+
+        return (true);
+    };
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, NumInputDim> inLengths,
+        const std::array<index_t, NumInputDim> inStrides,
+        const std::array<index_t, NumOutputDim> outLengths,
+        const std::array<std::array<index_t, NumOutputDim>, NumReduction> outStridesArray,
+        const std::array<int, NumReduceDim> reduceDims,
+        const std::array<const void*, NumReduction> alphas,
+        const std::array<const void*, NumReduction> betas,
+        const void* in_dev,
+        const std::array<void*, NumReduction> out_dev_buffers,
+        const InElementwiseOperationTuple in_elementwise_op_tuple,
+        const AccElementwiseOperationTuple acc_elementwise_op_tuple) override
+    {
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStridesArray,
+                                          reduceDims,
+                                          alphas,
+                                          betas,
+                                          in_dev,
+                                          out_dev_buffers,
+                                          in_elementwise_op_tuple,
+                                          acc_elementwise_op_tuple);
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << (OutMemoryDataOperation == InMemoryDataOperationEnum::Set? "DeviceMultipleReduceBlockWise<" : "DeviceMultipleReduceMultiBlock<") << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "InSrcVectorDim_" << InSrcVectorDim << "_InSrcVectorSize_" << InSrcVectorSize << ",";
+        str << "OutDstVectorSize"; 
+        static_for<0, OutDstVectorSizeSeq::Size(), 1>{}([&](auto I) {str << "_" << OutDstVectorSizeSeq::At(I); }); 
+        str << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp
new file mode 100644
index 0000000..ff8465e
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp
@@ -0,0 +1,422 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_multiple_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumReduction,
+          typename InDataType,
+          typename AccDataType,
+          typename OutDataTypeTuple,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          typename OutDstVectorSizeSeq>
+struct DeviceMultipleReduceThreadWise : public DeviceMultipleReduce<Rank,
+                                                                    NumReduceDim,
+                                                                    NumReduction,
+                                                                    InElementwiseOperationTuple,
+                                                                    AccElementwiseOperationTuple>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static_assert((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                      (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(NumReduction == OutDataTypeTuple::Size() &&
+                      NumReduction == InElementwiseOperationTuple::Size() &&
+                      NumReduction == AccElementwiseOperationTuple::Size() &&
+                      NumReduction == OutDstVectorSizeSeq::Size(),
+                  "All tuple should have the same size as the number of Reductions!");
+
+    static_assert(sequence_all_of(OutDstVectorSizeSeq{},
+                                  [](auto vectorSize) {
+                                      return (MThreadSliceSize % vectorSize == 0);
+                                  }),
+                  "The OutDstVectorSize should completely divide the MThreadSliceSize!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumInputDim  = Rank;
+    static constexpr index_t NumOutputDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim    = (NumInvariantDim == 0);
+
+    static constexpr index_t M_BlockTileSize = BlockSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = 1 * KThreadSliceSize;
+
+    static auto GenerateOutDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+
+                return static_cast<DataType*>(nullptr);
+            },
+            Number<NumReduction>{});
+    };
+
+    using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
+
+    static auto MakeSrc2dDescriptor(const std::array<index_t, NumInputDim>& inLengths,
+                                    const std::array<index_t, NumInputDim>& inStrides)
+    {
+        const auto tupleSrcLengths =
+            generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInputDim>{});
+        const auto tupleSrcStrides =
+            generate_tuple([&](auto I) { return inStrides[I]; }, Number<NumInputDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, NumInputDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths = generate_tuple(
+                    [&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
+                const auto invariantDimLengths =
+                    generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K =
+            math::integer_least_multiple(reduceLength, K_BlockTileSize) - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    static auto MakeDst1dDescriptor(const std::array<index_t, NumOutputDim>& outLengths,
+                                    const std::array<index_t, NumOutputDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumOutputDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumOutputDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumOutputDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto outPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto out_grid_desc_m_padded = transform_tensor_descriptor(
+            out_grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, outPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    static auto GenerateOutGrid1dDescTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                (void)I;
+                return MakeDst1dDescriptor(std::array<index_t, NumOutputDim>{},
+                                           std::array<index_t, NumOutputDim>{});
+            },
+            Number<NumReduction>{});
+    };
+
+    using InGridDesc_M_K      = decltype(MakeSrc2dDescriptor(std::array<index_t, NumInputDim>{},
+                                                        std::array<index_t, NumInputDim>{}));
+    using OutGridDesc_M_Tuple = decltype(GenerateOutGrid1dDescTuple());
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, NumInputDim>& inLengths,
+                 const std::array<index_t, NumInputDim>& inStrides,
+                 const std::array<index_t, NumOutputDim>& outLengths,
+                 const std::array<std::array<index_t, NumOutputDim>, NumReduction>& outStridesArray,
+                 const std::array<int, NumReduceDim>& reduceDims,
+                 const std::array<const void*, NumReduction>& alphas,
+                 const std::array<const void*, NumReduction>& betas,
+                 const void* in_dev,
+                 const std::array<void*, NumReduction>& out_dev_buffers,
+                 const InElementwiseOperationTuple in_elementwise_op_tuple,
+                 const AccElementwiseOperationTuple acc_elementwise_op_tuple)
+            : outLengths_{outLengths},
+              outStridesArray_{outStridesArray},
+              in_elementwise_op_tuple_{in_elementwise_op_tuple},
+              acc_elementwise_op_tuple_{acc_elementwise_op_tuple}
+        {
+            inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
+
+            for(size_t i = 0; i < NumReduction; i++)
+            {
+                alpha_values_(i) = *static_cast<const AccDataType*>(alphas[i]);
+                beta_values_(i)  = *static_cast<const AccDataType*>(betas[i]);
+            };
+
+            in_dev_ = static_cast<const InDataType*>(in_dev);
+
+            out_dev_buffers_ = generate_tuple(
+                [&](auto iR) {
+                    using OutDataTypePointer =
+                        remove_cvref_t<decltype(OutDataTypePointerTuple{}[iR])>;
+                    using OutDataType = remove_cvref_t<remove_pointer_t<OutDataTypePointer>>;
+                    return static_cast<OutDataType*>(out_dev_buffers[iR]);
+                },
+                Number<NumReduction>{});
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            in_grid_desc_m_k = MakeSrc2dDescriptor(inLengths_, inStrides_);
+
+            out_grid_desc_m_tuple = generate_tuple(
+                [&](auto I) { return MakeDst1dDescriptor(outLengths, outStridesArray[I]); },
+                Number<NumReduction>{});
+
+            gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                       M_BlockTileSize;
+        }
+
+        std::array<index_t, NumInputDim> inLengths_;
+        std::array<index_t, NumInputDim> inStrides_;
+
+        std::array<index_t, NumOutputDim> outLengths_;
+        std::array<std::array<index_t, NumOutputDim>, NumReduction> outStridesArray_;
+
+        Array<AccDataType, NumReduction> alpha_values_;
+        Array<AccDataType, NumReduction> beta_values_;
+
+        const InDataType* in_dev_;
+        OutDataTypePointerTuple out_dev_buffers_;
+
+        InGridDesc_M_K in_grid_desc_m_k;
+        OutGridDesc_M_Tuple out_grid_desc_m_tuple;
+
+        InElementwiseOperationTuple in_elementwise_op_tuple_;
+        AccElementwiseOperationTuple acc_elementwise_op_tuple_;
+
+        long_index_t invariant_total_length;
+        long_index_t reduce_total_length;
+
+        size_t gridSize;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            using GridwiseMultipleReduce =
+                GridwiseMultipleReduction_mk_to_m_threadwise<NumReduction,
+                                                             InDataType,
+                                                             OutDataTypePointerTuple,
+                                                             AccDataType,
+                                                             InGridDesc_M_K,
+                                                             OutGridDesc_M_Tuple,
+                                                             ReduceOperation,
+                                                             InElementwiseOperationTuple,
+                                                             AccElementwiseOperationTuple,
+                                                             InMemoryDataOperationEnum::Set,
+                                                             PropagateNan,
+                                                             BlockSize,
+                                                             MThreadSliceSize,
+                                                             KThreadSliceSize,
+                                                             InSrcVectorDim,
+                                                             InSrcVectorSize,
+                                                             OutDstVectorSizeSeq>;
+
+            const auto kernel_main =
+                kernel_multiple_reduce_threadwise<GridwiseMultipleReduce,
+                                                  NumReduction,
+                                                  InDataType,
+                                                  OutDataTypePointerTuple,
+                                                  AccDataType,
+                                                  InGridDesc_M_K,
+                                                  OutGridDesc_M_Tuple,
+                                                  InElementwiseOperationTuple,
+                                                  AccElementwiseOperationTuple>;
+
+            float avg_time = 0;
+
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.in_grid_desc_m_k,
+                                               arg.out_grid_desc_m_tuple,
+                                               arg.in_elementwise_op_tuple_,
+                                               arg.acc_elementwise_op_tuple_,
+                                               arg.alpha_values_,
+                                               arg.in_dev_,
+                                               arg.beta_values_,
+                                               arg.out_dev_buffers_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if constexpr(InSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return (false);
+            }
+            else
+            {
+                if(pArg->inStrides_[NumInvariantDim - 1] != 1 && InSrcVectorSize != 1)
+                    return (false);
+
+                if(pArg->inLengths_[NumInvariantDim - 1] % InSrcVectorSize != 0)
+                    return (false);
+            };
+        }
+        else
+        {
+            if(pArg->inStrides_[Rank - 1] != 1 && InSrcVectorSize != 1)
+                return (false);
+
+            if(pArg->inLengths_[Rank - 1] % InSrcVectorSize != 0)
+                return (false);
+        };
+
+        // To improve
+        bool valid = true;
+        static_for<0, NumReduction, 1>{}([&](auto I) {
+            if(pArg->outStridesArray_[I.value][NumOutputDim - 1] != 1 &&
+               OutDstVectorSizeSeq::At(I) != 1)
+                valid = false;
+
+            if(pArg->outLengths_[NumOutputDim - 1] % OutDstVectorSizeSeq::At(I) != 0)
+                valid = false;
+        });
+
+        if(!valid)
+            return (false);
+
+        return (true);
+    };
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, NumInputDim> inLengths,
+        const std::array<index_t, NumInputDim> inStrides,
+        const std::array<index_t, NumOutputDim> outLengths,
+        const std::array<std::array<index_t, NumOutputDim>, NumReduction> outStridesArray,
+        const std::array<int, NumReduceDim> reduceDims,
+        const std::array<const void*, NumReduction> alphas,
+        const std::array<const void*, NumReduction> betas,
+        const void* in_dev,
+        const std::array<void*, NumReduction> out_dev_buffers,
+        const InElementwiseOperationTuple in_elementwise_op_tuple,
+        const AccElementwiseOperationTuple acc_elementwise_op_tuple) override
+    {
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStridesArray,
+                                          reduceDims,
+                                          alphas,
+                                          betas,
+                                          in_dev,
+                                          out_dev_buffers,
+                                          in_elementwise_op_tuple,
+                                          acc_elementwise_op_tuple);
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceMultipleReduceThreadwise<" << BlockSize << ",";
+        str << "M_C" << BlockSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << 1 << "_S" << KThreadSliceSize << ",";
+        str << "InSrcVectorDim_" << InSrcVectorDim << "_InSrcVectorSize_" << InSrcVectorSize << ",";
+        str << "OutDstVectorSize"; 
+        static_for<0, OutDstVectorSizeSeq::Size(), 1>{}([&](auto I) {str << "_" << OutDstVectorSizeSeq::At(I); }); 
+        str << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp
new file mode 100644
index 0000000..47d9df8
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp
@@ -0,0 +1,476 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+template <typename GridwiseReduction,
+          typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename AccElementwiseOperation,
+          typename GridDesc_M_K>
+__global__ void kernel_normalization(const GridDesc_M_K x_grid_desc_m_k,
+                                     const GridDesc_M_K gamma_grid_desc_m_k,
+                                     const GridDesc_M_K beta_grid_desc_m_k,
+                                     const GridDesc_M_K y_grid_desc_m_k,
+                                     index_t num_k_block_tile_iteration,
+                                     AccDataType epsilon,
+                                     const XDataType* const __restrict__ p_x_global,
+                                     const GammaDataType* const __restrict__ p_gamma_global,
+                                     const BetaDataType* const __restrict__ p_beta_global,
+                                     YDataType* const __restrict__ p_y_global,
+                                     const AccElementwiseOperation acc_elementwise_op)
+{
+    GridwiseReduction::Run(x_grid_desc_m_k,
+                           gamma_grid_desc_m_k,
+                           beta_grid_desc_m_k,
+                           y_grid_desc_m_k,
+                           num_k_block_tile_iteration,
+                           epsilon,
+                           p_x_global,
+                           p_gamma_global,
+                           p_beta_global,
+                           p_y_global,
+                           acc_elementwise_op);
+};
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Y = Normalization(X, Beta, Gamma)
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          typename AccElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XYSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorSize>
+struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
+                                                            GammaDataType,
+                                                            BetaDataType,
+                                                            AccDataType,
+                                                            YDataType,
+                                                            AccElementwiseOperation,
+                                                            Rank,
+                                                            NumReduceDim>
+{
+    static_assert(
+        ((GammaSrcVectorDim == 0 && MThreadSliceSize % GammaSrcVectorSize == 0) ||
+         (GammaSrcVectorDim == 1 && KThreadSliceSize % GammaSrcVectorSize == 0)),
+        "Invalid thread slice sizes and/or gamma vector sizes configuration, please check!");
+
+    static_assert(
+        ((BetaSrcVectorDim == 0 && MThreadSliceSize % BetaSrcVectorSize == 0) ||
+         (BetaSrcVectorDim == 1 && KThreadSliceSize % BetaSrcVectorSize == 0)),
+        "Invalid thread slice sizes and/or beta vector sizes configuration, please check!");
+
+    using PassThrough = tensor_operation::element_wise::PassThrough;
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
+                                    const std::vector<index_t>& inStrides,
+                                    int blkGroupSize,
+                                    int numBlockTileIteration)
+    {
+        constexpr index_t NumInvariantDim  = Rank - NumReduceDim;
+        static constexpr index_t numSrcDim = Rank;
+        static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+        const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
+        const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, numSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, ReduceDims{});
+                const auto invariantDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, InvariantDims{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
+
+    using GridwiseReduceLayernormGeneric =
+        GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
+                                                      GammaDataType,
+                                                      BetaDataType,
+                                                      YDataType,
+                                                      AccDataType,
+                                                      AccElementwiseOperation,
+                                                      GridDesc_M_K,
+                                                      BlockSize,
+                                                      MThreadClusterSize,
+                                                      KThreadClusterSize,
+                                                      MThreadSliceSize,
+                                                      KThreadSliceSize,
+                                                      XYSrcVectorDim,
+                                                      XSrcVectorSize,
+                                                      GammaSrcVectorDim,
+                                                      GammaSrcVectorSize,
+                                                      BetaSrcVectorDim,
+                                                      BetaSrcVectorSize,
+                                                      XYSrcVectorDim,
+                                                      YDstVectorSize,
+                                                      false>;
+    using GridwiseNormalizationSweepOnce =
+        GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
+                                                      GammaDataType,
+                                                      BetaDataType,
+                                                      YDataType,
+                                                      AccDataType,
+                                                      AccElementwiseOperation,
+                                                      GridDesc_M_K,
+                                                      BlockSize,
+                                                      MThreadClusterSize,
+                                                      KThreadClusterSize,
+                                                      MThreadSliceSize,
+                                                      KThreadSliceSize,
+                                                      XYSrcVectorDim,
+                                                      XSrcVectorSize,
+                                                      GammaSrcVectorDim,
+                                                      GammaSrcVectorSize,
+                                                      BetaSrcVectorDim,
+                                                      BetaSrcVectorSize,
+                                                      XYSrcVectorDim,
+                                                      YDstVectorSize,
+                                                      true>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::vector<index_t> lengths,
+                 const std::vector<index_t> xStrides,
+                 const std::vector<index_t> gammaStrides,
+                 const std::vector<index_t> betaStrides,
+                 const std::vector<index_t> yStrides,
+                 const std::vector<index_t> reduceDims,
+                 AccElementwiseOperation acc_elementwise_op,
+                 AccDataType epsilon,
+                 const XDataType* p_x,
+                 const GammaDataType* p_gamma,
+                 const BetaDataType* p_beta,
+                 YDataType* p_y)
+            : epsilon_(epsilon),
+              p_x_(p_x),
+              p_gamma_(p_gamma),
+              p_beta_(p_beta),
+              p_y_(p_y),
+              acc_elementwise_op_(acc_elementwise_op)
+        {
+            Lengths_      = shuffle_tensor_dimensions<Rank, NumReduceDim>(lengths, reduceDims);
+            xStrides_     = shuffle_tensor_dimensions<Rank, NumReduceDim>(xStrides, reduceDims);
+            yStrides_     = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
+            gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims);
+            betaStrides_  = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims);
+
+            long_index_t invariant_total_length;
+            long_index_t reduce_total_length;
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(Lengths_);
+
+            blkGroupSize_          = 1;
+            numBlockTileIteration_ = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+
+            gridSize_ = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                        M_BlockTileSize * blkGroupSize_;
+
+            x_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, xStrides_, blkGroupSize_, numBlockTileIteration_);
+            gamma_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, gammaStrides_, blkGroupSize_, numBlockTileIteration_);
+            beta_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, betaStrides_, blkGroupSize_, numBlockTileIteration_);
+            y_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, yStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            isSweeponce_ =
+                x_grid_desc_m_k_.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
+        }
+
+        AccDataType epsilon_;
+
+        const XDataType* p_x_;
+        const GammaDataType* p_gamma_;
+        const BetaDataType* p_beta_;
+        YDataType* p_y_;
+
+        std::vector<index_t> Lengths_;
+        std::vector<index_t> xStrides_;
+        std::vector<index_t> gammaStrides_;
+        std::vector<index_t> betaStrides_;
+        std::vector<index_t> yStrides_;
+
+        AccElementwiseOperation acc_elementwise_op_;
+
+        int blkGroupSize_;
+        int numBlockTileIteration_;
+        size_t gridSize_;
+
+        GridDesc_M_K x_grid_desc_m_k_;
+        GridDesc_M_K gamma_grid_desc_m_k_;
+        GridDesc_M_K beta_grid_desc_m_k_;
+        GridDesc_M_K y_grid_desc_m_k_;
+        bool isSweeponce_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto kernel_main = arg.isSweeponce_
+                                         ? kernel_normalization<GridwiseNormalizationSweepOnce,
+                                                                XDataType,
+                                                                GammaDataType,
+                                                                BetaDataType,
+                                                                YDataType,
+                                                                AccDataType,
+                                                                AccElementwiseOperation,
+                                                                GridDesc_M_K>
+                                         : kernel_normalization<GridwiseReduceLayernormGeneric,
+                                                                XDataType,
+                                                                GammaDataType,
+                                                                BetaDataType,
+                                                                YDataType,
+                                                                AccDataType,
+                                                                AccElementwiseOperation,
+                                                                GridDesc_M_K>;
+
+            float avg_time = 0;
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize_),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.x_grid_desc_m_k_,
+                                               arg.gamma_grid_desc_m_k_,
+                                               arg.beta_grid_desc_m_k_,
+                                               arg.y_grid_desc_m_k_,
+                                               arg.numBlockTileIteration_,
+                                               arg.epsilon_,
+                                               arg.p_x_,
+                                               arg.p_gamma_,
+                                               arg.p_beta_,
+                                               arg.p_y_,
+                                               arg.acc_elementwise_op_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
+
+        constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+        if constexpr(XYSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return false;
+            }
+            else
+            {
+                if(p_arg_->xStrides_[NumInvariantDim - 1] != 1)
+                    return false;
+
+                if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
+                    return false;
+            };
+        }
+        else
+        {
+            if(p_arg_->xStrides_[Rank - 1] != 1)
+                return false;
+
+            if(p_arg_->Lengths_[Rank - 1] % XSrcVectorSize != 0)
+                return false;
+        };
+
+        if(p_arg_->Lengths_[Rank - 1] % YDstVectorSize != 0)
+        {
+            return false;
+        }
+
+        // if fastest dim is not reduced
+        if constexpr(GammaSrcVectorDim == 0)
+        {
+            if(p_arg_->gammaStrides_[NumInvariantDim - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return (false);
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->gammaStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return (false);
+        }
+
+        // if fastest dim is not reduced
+        if constexpr(BetaSrcVectorDim == 0)
+        {
+            if(p_arg_->betaStrides_[NumInvariantDim - 1] != 1)
+                return (false);
+
+            if(p_arg_->invariant_lowest_length % BetaSrcVectorSize != 0)
+                return (false);
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->betaStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % BetaSrcVectorSize != 0)
+                return (false);
+        }
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::vector<index_t> xStrides,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        AccDataType epsilon,
+                        const void* p_x,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        void* p_saveMean,
+                        void* p_saveInvVar,
+                        AccElementwiseOperation acc_elementwise_op) override
+    {
+        // TODO
+        // Optional cache of the intermediate results (mean and InvVariance) during the
+        // forward pass could speedup in the backward
+        ignore = p_saveMean;
+        ignore = p_saveInvVar;
+
+        return std::make_unique<Argument>(lengths,
+                                          xStrides,
+                                          gammaStrides,
+                                          betaStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          acc_elementwise_op,
+                                          epsilon,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const GammaDataType*>(p_gamma),
+                                          static_cast<const BetaDataType*>(p_beta),
+                                          static_cast<YDataType*>(p_y));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceNormalizationImpl<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "XYSrcVectorDim_" << XYSrcVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_Gamma" << GammaSrcVectorSize << "_Beta" << BetaSrcVectorSize << "_Y" << YDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp
new file mode 100644
index 0000000..7b96373
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp
@@ -0,0 +1,282 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <memory>
+#include <utility>
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_permute.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_permute.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Swap last 2 dimensions
+// input shape: [d[0], d[1], d[2], ..., d[NumDim-3], d[NumDim-2], d[NumDim-1]]
+//                                                                ^^^^^^^^^^^
+// output shape: [d[0], d[1], d[2], ..., d[NumDim-3], d[NumDim-1], d[NumDim-2]]
+//                                                    ^^^^^^^^^^^
+template <index_t NumDim,
+          typename InDataType,
+          typename OutDataType,
+          typename ElementwiseOperation,
+          index_t BlockSize,
+          index_t NPerBlock,
+          index_t HPerBlock,
+          index_t WPerBlock,
+          index_t InBlockLdsExtraW,
+          typename InBlockTransferThreadClusterLengths,
+          typename InBlockTransferThreadClusterArrangeOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector>
+struct DevicePermuteImpl : DevicePermute<NumDim, InDataType, OutDataType, ElementwiseOperation>
+{
+    using BaseType = DevicePermute<NumDim, InDataType, OutDataType, ElementwiseOperation>;
+    using typename BaseType::Lengths;
+    using typename BaseType::Strides;
+
+    static_assert(3 <= NumDim, "Only accept at least 3D dimension tensor");
+    static_assert((NumDim - 2) <= SrcVectorDim && SrcVectorDim < NumDim);
+    static_assert((NumDim - 2) <= DstVectorDim && DstVectorDim < NumDim);
+    static_assert(SrcVectorDim != DstVectorDim);
+
+    template <index_t N = NumDim>
+    static auto ConvertArrayToTuple(const std::array<index_t, NumDim>& array)
+    {
+        static_assert(1 <= N && N <= NumDim);
+
+        return generate_tuple([&](auto I) { return array[I]; }, Number<N>{});
+    }
+
+    static auto MakeDescriptor_N_H_W(const Lengths& lengths, const Strides& stride)
+    {
+        // create nd descriptor, shape: [d[0], d[1], d[2], ..., d[NumDim-3], d[NumDim-2],
+        // d[NumDim-1]]
+        const auto desc =
+            make_naive_tensor_descriptor(ConvertArrayToTuple(lengths), ConvertArrayToTuple(stride));
+
+        // merge nd to 3d descriptor, shape: [(d[0] * d[1] * d[2] * ... * d[NumDim-3]), d[NumDim-2],
+        // d[NumDim-1]]
+        //                                   => [N, H, W]
+        const index_t H       = *std::next(rbegin(lengths));
+        const index_t W       = *rbegin(lengths);
+        const auto desc_n_h_w = transform_tensor_descriptor(
+            desc,
+            make_tuple(make_merge_transform(ConvertArrayToTuple<NumDim - 2>(lengths)),
+                       make_pass_through_transform(H),
+                       make_pass_through_transform(W)),
+            make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<NumDim - 2>{}),
+                       Sequence<NumDim - 2>{},
+                       Sequence<NumDim - 1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        return PadTensorDescriptor(
+            desc_n_h_w, make_tuple(NPerBlock, HPerBlock, WPerBlock), Sequence<true, true, true>{});
+    }
+
+    using InGridDesc  = decltype(MakeDescriptor_N_H_W({1, 1}, {1, 1}));
+    using OutGridDesc = InGridDesc;
+
+    using GridwisePermute = GridwisePermute<
+        InGridDesc,
+        OutGridDesc,
+        InDataType,
+        OutDataType,
+        ElementwiseOperation,
+        BlockSize,
+        NPerBlock,
+        HPerBlock,
+        WPerBlock,
+        InBlockLdsExtraW,
+        InBlockTransferThreadClusterLengths,
+        InBlockTransferThreadClusterArrangeOrder,
+        SrcVectorDim - (NumDim - 3), // calculate new SrcVectorDim for the merged descriptor
+        DstVectorDim - (NumDim - 3), // calculate new DstVectorDim for the merged descriptor
+        SrcScalarPerVector,
+        DstScalarPerVector>;
+
+    using Block2TileMap = typename GridwisePermute::DefaultBlock2TileMap;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const Lengths& in_lengths,
+                 const Strides& in_strides,
+                 const Lengths& out_lengths,
+                 const Strides& out_strides,
+                 const void* in_dev_buffer,
+                 void* out_dev_buffer,
+                 ElementwiseOperation elementwise_op)
+            : in_dev_buffer_(static_cast<const InDataType*>(in_dev_buffer)),
+              out_dev_buffer_(static_cast<OutDataType*>(out_dev_buffer)),
+              in_grid_desc_(MakeDescriptor_N_H_W(in_lengths, in_strides)),
+              out_grid_desc_(MakeDescriptor_N_H_W(out_lengths, out_strides)),
+              in_lengths_(in_lengths),
+              in_strides_(in_strides),
+              out_lengths_(out_lengths),
+              out_strides_(out_strides),
+              elementwise_op_(elementwise_op),
+              block_2_tile_map_(GridwisePermute::MakeDefaultBlock2TileMap(in_grid_desc_))
+        {
+        }
+
+        const InDataType* in_dev_buffer_;
+        OutDataType* out_dev_buffer_;
+        InGridDesc in_grid_desc_;
+        OutGridDesc out_grid_desc_;
+
+        Lengths in_lengths_;
+        Strides in_strides_;
+        Lengths out_lengths_;
+        Strides out_strides_;
+
+        ElementwiseOperation elementwise_op_;
+
+        Block2TileMap block_2_tile_map_;
+    };
+
+    struct Invoker : BaseInvoker
+    {
+        static float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const index_t grid_size = arg.block_2_tile_map_.CalculateGridSize(arg.in_grid_desc_);
+
+            const auto kernel = kernel_nd_permute<GridwisePermute,
+                                                  InGridDesc,
+                                                  OutGridDesc,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  ElementwiseOperation,
+                                                  Block2TileMap>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(grid_size),
+                                                        dim3(BlockSize),
+                                                        0,
+                                                        arg.in_grid_desc_,
+                                                        arg.out_grid_desc_,
+                                                        arg.in_dev_buffer_,
+                                                        arg.out_dev_buffer_,
+                                                        arg.elementwise_op_,
+                                                        arg.block_2_tile_map_);
+            return elapsed_time;
+        }
+
+        float Run(const BaseArgument* arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override final
+        {
+            const auto* const argument = dynamic_cast<const Argument*>(arg);
+            if(!argument)
+            {
+                return NAN;
+            }
+
+            return Run(*argument, stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        constexpr auto GetPaddedLength = [](index_t length, index_t tile_length) {
+            return math::integer_divide_ceil(length, tile_length) * tile_length;
+        };
+
+        constexpr auto IsScalarPerVectorValid =
+            [](index_t length, index_t stride, index_t scalar_per_vector) {
+                if(stride == 1 && length % scalar_per_vector == 0)
+                {
+                    return true;
+                }
+                else if(stride != 1 && scalar_per_vector == 1)
+                {
+                    return true;
+                }
+
+                return false;
+            };
+
+        return IsScalarPerVectorValid(arg.in_lengths_[SrcVectorDim],
+                                      arg.in_strides_[SrcVectorDim],
+                                      SrcScalarPerVector) &&
+               IsScalarPerVectorValid(
+                   GetPaddedLength(arg.in_lengths_[SrcVectorDim],
+                                   (SrcVectorDim == NumDim - 2 ? HPerBlock : WPerBlock)),
+                   arg.in_strides_[SrcVectorDim],
+                   SrcScalarPerVector) &&
+               IsScalarPerVectorValid(arg.out_lengths_[DstVectorDim],
+                                      arg.out_strides_[DstVectorDim],
+                                      DstScalarPerVector) &&
+               IsScalarPerVectorValid(
+                   GetPaddedLength(arg.out_lengths_[DstVectorDim],
+                                   (DstVectorDim == NumDim - 2 ? HPerBlock : WPerBlock)),
+                   arg.in_strides_[DstVectorDim],
+                   DstScalarPerVector) &&
+               GridwisePermute::CheckValidity(arg.in_grid_desc_, arg.out_grid_desc_);
+    };
+
+    // override methods inherited from 'BaseOperator'
+    bool IsSupportedArgument(const BaseArgument* arg) override final
+    {
+        const auto* const argument = dynamic_cast<const Argument*>(arg);
+        if(!argument)
+        {
+            return false;
+        }
+
+        return IsSupportedArgument(*argument);
+    }
+
+    // override methods inherited from 'DevicePermute'
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const Lengths& in_lengths,
+                        const Strides& in_strides,
+                        const Lengths& out_lengths,
+                        const Strides& out_strides,
+                        const void* in_dev_buffer,
+                        void* out_dev_buffer,
+                        ElementwiseOperation elementwise_op) override final
+    {
+        return std::make_unique<Argument>(in_lengths,
+                                          in_strides,
+                                          out_lengths,
+                                          out_strides,
+                                          in_dev_buffer,
+                                          out_dev_buffer,
+                                          elementwise_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override final
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    // other constructor methods
+    template <typename... Args>
+    static std::enable_if_t<std::is_constructible_v<Argument, Args...>, Argument>
+    MakeArgument(Args&&... args) noexcept(std::is_nothrow_constructible_v<Argument, Args...>)
+    {
+        return Argument{std::forward<Args>(args)...};
+    }
+
+    static std::enable_if_t<std::is_default_constructible_v<Invoker>, Invoker>
+    MakeInvoker() noexcept(std::is_nothrow_default_constructible_v<Invoker>)
+    {
+        return Invoker{};
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp
new file mode 100644
index 0000000..bfde40c
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp
@@ -0,0 +1,327 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          bool OuputIndex,
+          ck::index_t BlockSize,
+          ck::index_t ReduceMThreadClusterSize,
+          ck::index_t ReduceKThreadClusterSize,
+          ck::index_t ReduceMThreadSliceSize,
+          ck::index_t ReduceKThreadSliceSize,
+          ck::index_t InSrcOutDstVectorSize>
+struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd<ReduceOpId>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    using IndexDataType = int32_t;
+
+    using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+
+    using InElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+
+    using AccElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+    static constexpr index_t InSrcOutDstVectorDim =
+        0; // for NHWC, the dim C is the vector Dim for both input and output in memory, which is
+           // not reduced.
+
+    static constexpr ck::index_t ReduceM_BlockTileSize =
+        ReduceMThreadClusterSize * ReduceMThreadSliceSize;
+    static constexpr ck::index_t ReduceK_BlockTileSize =
+        ReduceKThreadClusterSize * ReduceKThreadSliceSize;
+
+    static auto MakeABGridDescriptor_A_M_K_B_M(ck::index_t N,
+                                               ck::index_t C,
+                                               std::array<ck::index_t, 2> input_spatial_lengths,
+                                               std::array<ck::index_t, 2> window_spatial_lengths,
+                                               std::array<ck::index_t, 2> output_spatial_lengths,
+                                               std::array<ck::index_t, 2> window_strides,
+                                               std::array<ck::index_t, 2> input_left_pads,
+                                               std::array<ck::index_t, 2> input_right_pads)
+    {
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = window_spatial_lengths[0];
+        const index_t X = window_spatial_lengths[1];
+
+        const index_t ConvStrideH = window_strides[0];
+        const index_t ConvStrideW = window_strides[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t ReduceMRaw = N * Ho * Wo * C;
+        const index_t ReduceMPad =
+            math::integer_least_multiple(ReduceMRaw, ReduceM_BlockTileSize) - ReduceMRaw;
+
+        const index_t ReduceKRaw = Y * X;
+        const index_t ReduceKPad =
+            math::integer_least_multiple(ReduceKRaw, ReduceK_BlockTileSize) - ReduceKRaw;
+
+        // A[ReduceM, ReduceK]
+        const auto in_grid_desc_n_hi_wi_c =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        const auto in_grid_desc_n_hip_wip_c = transform_tensor_descriptor(
+            in_grid_desc_n_hi_wi_c,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                       make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        const auto in_grid_desc_n_y_ho_x_wo_c = transform_tensor_descriptor(
+            in_grid_desc_n_hip_wip_c,
+            make_tuple(make_pass_through_transform(N),
+                       make_embed_transform(make_tuple(Y, Ho), make_tuple(I1, ConvStrideH)),
+                       make_embed_transform(make_tuple(X, Wo), make_tuple(I1, ConvStrideW)),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+        const auto in_grid_desc_reducemraw_reducekraw =
+            transform_tensor_descriptor(in_grid_desc_n_y_ho_x_wo_c,
+                                        make_tuple(make_merge_transform(make_tuple(N, Ho, Wo, C)),
+                                                   make_merge_transform(make_tuple(Y, X))),
+                                        make_tuple(Sequence<0, 2, 4, 5>{}, Sequence<1, 3>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto in_grid_desc_reducem_reducek = transform_tensor_descriptor(
+            in_grid_desc_reducemraw_reducekraw,
+            make_tuple(make_right_pad_transform(ReduceMRaw, ReduceMPad),
+                       make_right_pad_transform(ReduceKRaw, ReduceKPad)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        // B[ReduceM]
+        const auto out_grid_desc_reducemraw =
+            make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo * C));
+
+        const auto out_grid_desc_reducem = transform_tensor_descriptor(
+            out_grid_desc_reducemraw,
+            make_tuple(make_right_pad_transform(ReduceMRaw, ReduceMPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+
+        return make_tuple(in_grid_desc_reducem_reducek, out_grid_desc_reducem);
+    }
+
+    using ABGridDescs = decltype(
+        MakeABGridDescriptor_A_M_K_B_M(1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_M_K = remove_cvref_t<decltype(ABGridDescs{}[I0])>;
+    using BGridDesc_M   = remove_cvref_t<decltype(ABGridDescs{}[I1])>;
+
+    // TODO
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_dev,
+                 OutDataType* p_out_dev,
+                 int* p_out_indices_dev,
+                 ck::index_t N,
+                 ck::index_t C,
+                 std::array<ck::index_t, 2>& input_spatial_lengths,
+                 std::array<ck::index_t, 2>& window_spatial_lengths,
+                 std::array<ck::index_t, 2>& output_spatial_lengths,
+                 std::array<ck::index_t, 2>& window_strides,
+                 std::array<ck::index_t, 2>& input_left_pads,
+                 std::array<ck::index_t, 2>& input_right_pads)
+            : p_in_dev_{p_in_dev},
+              p_out_dev_{p_out_dev},
+              p_out_indices_dev_{p_out_indices_dev},
+              a_grid_desc_m_k_{},
+              b_grid_desc_m_{}
+        {
+            const auto descs = MakeABGridDescriptor_A_M_K_B_M(N,
+                                                              C,
+                                                              input_spatial_lengths,
+                                                              window_spatial_lengths,
+                                                              output_spatial_lengths,
+                                                              window_strides,
+                                                              input_left_pads,
+                                                              input_right_pads);
+
+            a_grid_desc_m_k_ = descs[I0];
+            b_grid_desc_m_   = descs[I1];
+
+            invariant_lowest_length_ = C;
+            reduce_lowest_length_    = window_spatial_lengths[1];
+
+            int32_t reduceLength = window_spatial_lengths[0] * window_spatial_lengths[1];
+
+            std::tie(in_element_op_, acc_element_op_) =
+                reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(reduceLength);
+        }
+
+        const InDataType* p_in_dev_;
+        OutDataType* p_out_dev_;
+        int* p_out_indices_dev_;
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_M b_grid_desc_m_;
+        InElementwiseOperation in_element_op_;
+        AccElementwiseOperation acc_element_op_;
+
+        // for checking vector load/store
+        ck::index_t invariant_lowest_length_;
+        ck::index_t reduce_lowest_length_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            using gridwise_reduce =
+                GridwiseReduction_mk_to_m_threadwise<InDataType,
+                                                     OutDataType,
+                                                     AccDataType,
+                                                     IndexDataType,
+                                                     AGridDesc_M_K,
+                                                     BGridDesc_M,
+                                                     ReduceOperation,
+                                                     InElementwiseOperation,
+                                                     AccElementwiseOperation,
+                                                     InMemoryDataOperationEnum::Set,
+                                                     false, // propagate_nan
+                                                     BlockSize,
+                                                     ReduceMThreadSliceSize,
+                                                     ReduceKThreadSliceSize,
+                                                     InSrcOutDstVectorDim,
+                                                     InSrcOutDstVectorSize,
+                                                     InSrcOutDstVectorSize>;
+
+            const auto kernel = kernel_reduce_threadwise<gridwise_reduce,
+                                                         OuputIndex,
+                                                         false, // don't have index input
+                                                         InDataType,
+                                                         OutDataType,
+                                                         AccDataType,
+                                                         IndexDataType,
+                                                         AGridDesc_M_K,
+                                                         BGridDesc_M,
+                                                         InElementwiseOperation,
+                                                         AccElementwiseOperation>;
+
+            ck::index_t ReduceM = arg.a_grid_desc_m_k_.GetLength(I0);
+
+            const index_t grid_size = (ReduceM / ReduceM_BlockTileSize);
+
+            return launch_and_time_kernel(stream_config,
+                                          kernel,
+                                          dim3(grid_size),
+                                          dim3(BlockSize),
+                                          0,
+                                          arg.a_grid_desc_m_k_,
+                                          arg.b_grid_desc_m_,
+                                          arg.in_element_op_,
+                                          arg.acc_element_op_,
+                                          float(1),
+                                          arg.p_in_dev_,
+                                          nullptr,
+                                          float(0),
+                                          arg.p_out_dev_,
+                                          arg.p_out_indices_dev_);
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if(pArg->invariant_lowest_length_ % InSrcOutDstVectorSize != 0)
+        {
+            return (false);
+        }
+
+        return (true);
+    }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_dev,
+                        void* p_out_dev,
+                        void* p_out_indices_dev,
+                        ck::index_t N,
+                        ck::index_t C,
+                        std::array<ck::index_t, 2> input_spatial_lengths,
+                        std::array<ck::index_t, 2> window_spatial_lengths,
+                        std::array<ck::index_t, 2> output_spatial_lengths,
+                        std::array<ck::index_t, 2> window_strides,
+                        std::array<ck::index_t, 2> input_left_pads,
+                        std::array<ck::index_t, 2> input_right_pads) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_dev),
+                                          static_cast<OutDataType*>(p_out_dev),
+                                          static_cast<int*>(p_out_indices_dev),
+                                          N,
+                                          C,
+                                          input_spatial_lengths,
+                                          window_spatial_lengths,
+                                          output_spatial_lengths,
+                                          window_strides,
+                                          input_left_pads,
+                                          input_right_pads);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<" << BlockSize << ",";
+        str << "M_C" << ReduceMThreadClusterSize << "_S" << ReduceMThreadSliceSize << ",";
+        str << "K_C" << ReduceKThreadClusterSize << "_S" << ReduceKThreadSliceSize << ",";
+        str <<"InSrcOutDstVectorSize_" << InSrcOutDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp
new file mode 100644
index 0000000..5dc051b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp
@@ -0,0 +1,142 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <cassert>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// here, inLengths[] is already shuffled so that lengths of invariant dims are included before those
+// of reduce dims
+template <index_t Rank, int NumReduceDim>
+std::pair<long_index_t, long_index_t> get_2d_lengths(const std::vector<index_t>& inLengths)
+{
+    static_assert(Rank <= 6, "bigger Rank size not supported!");
+
+    long_index_t invariant_total_length = 1;
+    long_index_t reduce_total_length    = 1;
+
+    constexpr int NumInvariantDim = Rank - NumReduceDim;
+
+    for(int i = NumInvariantDim; i < Rank; i++)
+        reduce_total_length *= inLengths[i];
+
+    for(int i = 0; i < NumInvariantDim; i++)
+        invariant_total_length *= inLengths[i];
+
+    return std::make_pair(invariant_total_length, reduce_total_length);
+};
+
+template <index_t Rank, int NumReduceDim>
+std::pair<long_index_t, long_index_t> get_2d_lengths(const std::array<index_t, Rank>& inLengths)
+{
+    static_assert(Rank <= 6, "bigger Rank size not supported!");
+
+    long_index_t invariant_total_length = 1;
+    long_index_t reduce_total_length    = 1;
+
+    constexpr int NumInvariantDim = Rank - NumReduceDim;
+
+    for(int i = NumInvariantDim; i < Rank; i++)
+        reduce_total_length *= inLengths[i];
+
+    for(int i = 0; i < NumInvariantDim; i++)
+        invariant_total_length *= inLengths[i];
+
+    return std::make_pair(invariant_total_length, reduce_total_length);
+};
+
+// helper functions using variadic template arguments
+template <index_t... Ns>
+auto make_tuple_from_array_and_index_seq(const std::vector<index_t>& lengths, Sequence<Ns...>)
+{
+    return make_tuple(static_cast<index_t>(lengths[Ns])...);
+};
+
+template <index_t arraySize>
+auto make_tuple_from_array(const std::vector<index_t>& lengths, Number<arraySize>)
+{
+    static_assert(arraySize >= 1 && arraySize <= 6, "The tensor should have 1 to 6 dimensions");
+
+    constexpr auto index_seq = typename arithmetic_sequence_gen<0, arraySize, 1>::type{};
+
+    return make_tuple_from_array_and_index_seq(lengths, index_seq);
+};
+
+template <index_t Rank, index_t NumReduceDim>
+std::vector<index_t> shuffle_tensor_dimensions(const std::vector<index_t>& origLengthsStrides,
+                                               const std::vector<int>& reduceDims)
+{
+    std::vector<index_t> newLengthsStrides;
+
+    assert(Rank == origLengthsStrides.size() && NumReduceDim == reduceDims.size());
+
+    int reduceFlag = 0;
+
+    // flag the bits for the reduceDims
+    for(int i = 0; i < NumReduceDim; i++)
+    {
+        reduceFlag |= 1 << reduceDims[i];
+    };
+
+    // collect invariant dimensions
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) == 0)
+        {
+            newLengthsStrides.push_back(origLengthsStrides[i]);
+        };
+
+    // collect reduce dimensions
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) > 0)
+        {
+            newLengthsStrides.push_back(origLengthsStrides[i]);
+        };
+
+    return newLengthsStrides;
+};
+
+template <index_t Rank, index_t NumReduceDim>
+std::array<index_t, Rank>
+shuffle_tensor_dimensions(const std::array<index_t, Rank>& origLengthsStrides,
+                          const std::array<int, NumReduceDim>& reduceDims)
+{
+    std::array<index_t, Rank> newLengthsStrides;
+
+    int reduceFlag = 0;
+
+    // flag the bits for the reduceDims
+    for(int i = 0; i < NumReduceDim; i++)
+    {
+        reduceFlag |= 1 << reduceDims[i];
+    };
+
+    // collect invariant dimensions
+    int pos = 0;
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) == 0)
+        {
+            newLengthsStrides[pos++] = origLengthsStrides[i];
+        };
+
+    // collect reduce dimensions
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) > 0)
+        {
+            newLengthsStrides[pos++] = origLengthsStrides[i];
+        };
+
+    return newLengthsStrides;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
new file mode 100644
index 0000000..93855eb
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
@@ -0,0 +1,543 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <array>
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          bool OutputIndex,
+          bool HaveIndexInputIfOutputIndex,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize>
+struct DeviceReduceMultiBlock
+    : public DeviceReduce<Rank, NumReduceDim, InElementwiseOperation, AccElementwiseOperation>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "Invalid thread cluster size assignments!");
+
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)) &&
+                      (MThreadSliceSize % OutDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    using IndexDataType = int32_t;
+
+    static constexpr bool HaveIndexInput = OutputIndex && HaveIndexInputIfOutputIndex;
+
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumSrcDim = Rank;
+    static constexpr index_t NumDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+    // So far, only AtomicAdd is considered, other Atomic Operation like AtomicMax can be added
+    // later
+    static constexpr bool use_multiblock =
+        (OutMemoryDataOperation == InMemoryDataOperationEnum::AtomicAdd);
+
+    static_assert(ck::reduce::InMemoryDataOperatonSupportedOnDataType<OutMemoryDataOperation,
+                                                                      OutDataType>::value,
+                  "The OutDataType must support the specified OutMemoryDataOperation!");
+
+    static_assert(!use_multiblock || (use_multiblock && !OutputIndex),
+                  "MultiBlock reduction can only be used when outputing index is not required");
+
+    static_assert(
+        ReduceOperation::IsCompatibleInMemoryDataOperation(OutMemoryDataOperation),
+        "The reduction accumulation operation must be compatible with the OutMemoryDataOperation!");
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeSrc2dDescriptor(const std::array<index_t, Rank>& inLengths,
+                                    const std::array<index_t, Rank>& inStrides,
+                                    int blkGroupSize,
+                                    int numBlockTileIteration)
+    {
+        const auto tupleSrcLengths =
+            generate_tuple([&](auto I) { return inLengths[I]; }, Number<Rank>{});
+        const auto tupleSrcStrides =
+            generate_tuple([&](auto I) { return inStrides[I]; }, Number<Rank>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, NumSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths = generate_tuple(
+                    [&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
+                const auto invariantDimLengths =
+                    generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    static auto MakeDst1dDescriptor(const std::array<index_t, NumDstDim>& outLengths,
+                                    const std::array<index_t, NumDstDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumDstDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumDstDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumDstDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto outPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto out_grid_desc_m_padded = transform_tensor_descriptor(
+            out_grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, outPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    static auto MakeDst1dDescriptorForBufferSet(const std::array<index_t, NumDstDim>& outLengths,
+                                                const std::array<index_t, NumDstDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumDstDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumDstDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumDstDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto length = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto pad = math::integer_least_multiple(length, BlockSize) - length;
+
+        auto out_grid_desc_m_padded =
+            transform_tensor_descriptor(out_grid_desc_m,
+                                        make_tuple(make_right_pad_transform(length, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> inLengths,
+                 const std::array<index_t, Rank> inStrides,
+                 const std::array<index_t, NumDstDim> outLengths,
+                 const std::array<index_t, NumDstDim> outStrides,
+                 const std::array<int, NumReduceDim> reduceDims,
+                 float alpha,
+                 float beta,
+                 const InDataType* in_dev,
+                 const IndexDataType* in_index_dev,
+                 OutDataType* out_dev,
+                 IndexDataType* out_index_dev,
+                 const InElementwiseOperation in_elementwise_op,
+                 const AccElementwiseOperation acc_elementwise_op)
+            : outLengths_{outLengths},
+              outStrides_{outStrides},
+              in_dev_{in_dev},
+              in_index_dev_{in_index_dev},
+              out_dev_{out_dev},
+              out_index_dev_{out_index_dev},
+              in_elementwise_op_{in_elementwise_op},
+              acc_elementwise_op_{acc_elementwise_op}
+        {
+            if(Rank != inLengths.size() || Rank != inStrides.size() ||
+               NumReduceDim != reduceDims.size())
+            {
+                throw std::runtime_error(
+                    "One of inLengths/inStrides/reduceDims has invalid size!"
+                    "\nExpected size inLengths: " +
+                    std::to_string(Rank) + ", inStrides: " + std::to_string(Rank) +
+                    ", reduceDims: " + std::to_string(NumReduceDim) +
+                    "\nBut have inLengths: " + std::to_string(inLengths.size()) +
+                    ", inStrides: " + std::to_string(inStrides.size()) +
+                    ", reduceDims: " + std::to_string(reduceDims.size()));
+            }
+
+            for(std::size_t i = 0; i < reduceDims.size(); ++i)
+            {
+                if(reduceDims[i] < 0 || reduceDims[i] >= Rank)
+                {
+                    throw std::runtime_error("Provided reduce dimension exceed input tensor Rank!"
+                                             "\nHave reduceDims[" +
+                                             std::to_string(i) +
+                                             "]: " + std::to_string(reduceDims[i]));
+                }
+            }
+
+            inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
+
+            alpha_ = type_convert<AccDataType>(alpha);
+            beta_  = type_convert<AccDataType>(beta);
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            if constexpr(NumInvariantDim == 0)
+                invariant_lowest_length = 1;
+            else
+                invariant_lowest_length = inLengths_[NumInvariantDim - 1];
+
+            reduce_lowest_length = inLengths_[Rank - 1];
+
+            if constexpr(use_multiblock)
+            {
+
+                int iterations = 1;
+                while(true)
+                {
+                    int testBlkGroupSize =
+                        (reduce_total_length + (K_BlockTileSize * iterations) - 1) /
+                        (K_BlockTileSize * iterations);
+
+                    // we want the blkGroupSize be not more than 128
+                    if(testBlkGroupSize <= 128)
+                        break;
+
+                    iterations++;
+                };
+
+                blkGroupSize = (reduce_total_length + (K_BlockTileSize * iterations) - 1) /
+                               (K_BlockTileSize * iterations);
+
+                numBlockTileIteration = iterations;
+            }
+            else
+            {
+                blkGroupSize = 1;
+                numBlockTileIteration =
+                    (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+            };
+
+            gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                       M_BlockTileSize * blkGroupSize;
+
+            gridSize_pre =
+                math::integer_least_multiple(invariant_total_length, BlockSize) / BlockSize;
+        }
+
+        std::array<index_t, Rank> inLengths_;
+        std::array<index_t, Rank> inStrides_;
+        std::array<index_t, NumDstDim> outLengths_;
+        std::array<index_t, NumDstDim> outStrides_;
+
+        AccDataType alpha_;
+        AccDataType beta_;
+
+        const InDataType* in_dev_;
+        const IndexDataType* in_index_dev_;
+        OutDataType* out_dev_;
+        IndexDataType* out_index_dev_;
+
+        InElementwiseOperation in_elementwise_op_;
+        AccElementwiseOperation acc_elementwise_op_;
+
+        index_t invariant_lowest_length;
+        index_t reduce_lowest_length;
+        long_index_t invariant_total_length;
+        long_index_t reduce_total_length;
+
+        int blkGroupSize;
+        int numBlockTileIteration;
+        size_t gridSize;
+
+        size_t gridSize_pre;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto in_grid_desc_m_k = DeviceReduceMultiBlock::MakeSrc2dDescriptor(
+                arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.numBlockTileIteration);
+            const auto out_grid_desc_m =
+                DeviceReduceMultiBlock::MakeDst1dDescriptor(arg.outLengths_, arg.outStrides_);
+            const auto out_grid_desc_m_2 = DeviceReduceMultiBlock::MakeDst1dDescriptorForBufferSet(
+                arg.outLengths_, arg.outStrides_);
+
+            using InGridDesc_M_K  = decltype(in_grid_desc_m_k);
+            using OutGridDesc_M   = decltype(out_grid_desc_m);
+            using OutGridDesc_M_2 = decltype(out_grid_desc_m_2);
+
+            using GridwiseReduce = GridwiseReduction_mk_to_m_multiblock<InDataType,
+                                                                        OutDataType,
+                                                                        AccDataType,
+                                                                        IndexDataType,
+                                                                        InGridDesc_M_K,
+                                                                        OutGridDesc_M,
+                                                                        ReduceOperation,
+                                                                        InElementwiseOperation,
+                                                                        AccElementwiseOperation,
+                                                                        OutMemoryDataOperation,
+                                                                        PropagateNan,
+                                                                        BlockSize,
+                                                                        MThreadClusterSize,
+                                                                        KThreadClusterSize,
+                                                                        MThreadSliceSize,
+                                                                        KThreadSliceSize,
+                                                                        InSrcVectorDim,
+                                                                        InSrcVectorSize,
+                                                                        OutDstVectorSize>;
+
+            const auto kernel_main = kernel_reduce_multiblock<GridwiseReduce,
+                                                              OutputIndex,
+                                                              HaveIndexInput,
+                                                              InDataType,
+                                                              OutDataType,
+                                                              AccDataType,
+                                                              int32_t,
+                                                              InGridDesc_M_K,
+                                                              OutGridDesc_M,
+                                                              InElementwiseOperation,
+                                                              AccElementwiseOperation>;
+
+            float avg_time = 0;
+
+            if constexpr(use_multiblock)
+            {
+                const auto identityVal =
+                    ck::reduce::GetIdentityValueForInMemoryDataOperation<OutDataType>(
+                        OutMemoryDataOperation);
+
+                const auto kernel_pre =
+                    kernel_buffer_set_value<BlockSize, OutDataType, OutGridDesc_M_2>;
+
+                avg_time += launch_and_time_kernel(stream_config,
+                                                   kernel_pre,
+                                                   dim3(arg.gridSize_pre),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   out_grid_desc_m_2,
+                                                   arg.out_dev_,
+                                                   identityVal);
+            };
+
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize),
+                                               dim3(BlockSize),
+                                               0,
+                                               in_grid_desc_m_k,
+                                               out_grid_desc_m,
+                                               arg.in_elementwise_op_,
+                                               arg.acc_elementwise_op_,
+                                               arg.blkGroupSize,
+                                               arg.numBlockTileIteration,
+                                               arg.alpha_,
+                                               arg.in_dev_,
+                                               arg.in_index_dev_,
+                                               arg.beta_,
+                                               arg.out_dev_,
+                                               arg.out_index_dev_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    static bool IsSupportedArgument(const Argument* pArg)
+    {
+        if constexpr(use_multiblock)
+        {
+            if(static_cast<float>(pArg->beta_) != 0.0f)
+                return (false);
+        };
+
+        if constexpr(InSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return (false);
+            }
+            else
+            {
+                if(pArg->inStrides_[NumInvariantDim - 1] != 1)
+                    return (false);
+
+                if(pArg->invariant_lowest_length % InSrcVectorSize != 0)
+                    return (false);
+            };
+        }
+        else
+        {
+            if(pArg->inStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(pArg->reduce_lowest_length % InSrcVectorSize != 0)
+                return (false);
+        };
+
+        // To improve
+        if(pArg->invariant_lowest_length % OutDstVectorSize != 0)
+            return (false);
+
+        if constexpr(use_multiblock)
+        {
+            // blkGroupSize of 1 should be handled by Blockwise path using
+            // InMemoryDataOperationEnum::Set
+            if(pArg->blkGroupSize == 1)
+                return (false);
+
+            // This is very strong restriction, but needed to avoid some failure
+            if(pArg->invariant_lowest_length % M_BlockTileSize != 0)
+                return (false);
+        }
+        else
+        {
+            // cases with very small reduce_total_length should be handled by ThreadWise kernel
+            // if(pArg->reduce_total_length / KThreadSliceSize < 2)
+            //     return (false);
+        };
+
+        return (true);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(dynamic_cast<const Argument*>(p_arg));
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> inLengths,
+                        const std::array<index_t, Rank> inStrides,
+                        const std::array<index_t, NumDstDim> outLengths,
+                        const std::array<index_t, NumDstDim> outStrides,
+                        const std::array<int, NumReduceDim> reduceDims,
+                        float alpha,
+                        float beta,
+                        const void* in_dev,
+                        const void* in_index_dev,
+                        void* out_dev,
+                        void* out_index_dev,
+                        const InElementwiseOperation in_elementwise_op,
+                        const AccElementwiseOperation acc_elementwise_op) override
+    {
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStrides,
+                                          reduceDims,
+                                          alpha,
+                                          beta,
+                                          static_cast<const InDataType*>(in_dev),
+                                          static_cast<const IndexDataType*>(in_index_dev),
+                                          static_cast<OutDataType*>(out_dev),
+                                          static_cast<IndexDataType*>(out_index_dev),
+                                          in_elementwise_op,
+                                          acc_elementwise_op);
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << (OutMemoryDataOperation == InMemoryDataOperationEnum::Set? "DeviceReduceBlockWise<" : "DeviceReduceMultiBlock<") << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "InSrcVectorDim_" << InSrcVectorDim << "_InSrcVectorSize_" << InSrcVectorSize << "_OutDstVectorSize_" << OutDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
new file mode 100644
index 0000000..05e14f0
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
@@ -0,0 +1,382 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <array>
+
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          bool PropagateNan,
+          bool OutputIndex,
+          bool HaveIndexInputIfOutputIndex,
+          index_t BlockSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize>
+struct DeviceReduceThreadWise
+    : public DeviceReduce<Rank, NumReduceDim, InElementwiseOperation, AccElementwiseOperation>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)) &&
+                      (MThreadSliceSize % OutDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    using IndexDataType = int32_t;
+
+    static constexpr bool HaveIndexInput = OutputIndex && HaveIndexInputIfOutputIndex;
+
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumSrcDim = Rank;
+    static constexpr index_t NumDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+    static constexpr index_t M_BlockTileSize = BlockSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = 1 * KThreadSliceSize;
+
+    static auto MakeSrc2dDescriptor(const std::array<index_t, Rank>& inLengths,
+                                    const std::array<index_t, Rank>& inStrides)
+    {
+        const auto tupleSrcLengths =
+            generate_tuple([&](auto I) { return inLengths[I]; }, Number<Rank>{});
+        const auto tupleSrcStrides =
+            generate_tuple([&](auto I) { return inStrides[I]; }, Number<Rank>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, NumSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths = generate_tuple(
+                    [&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
+                const auto invariantDimLengths =
+                    generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K =
+            math::integer_least_multiple(reduceLength, K_BlockTileSize) - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    static auto MakeDst1dDescriptor(const std::array<index_t, NumDstDim>& outLengths,
+                                    const std::array<index_t, NumDstDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumDstDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumDstDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumDstDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto outPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto out_grid_desc_m_padded = transform_tensor_descriptor(
+            out_grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, outPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> inLengths,
+                 const std::array<index_t, Rank> inStrides,
+                 const std::array<index_t, NumDstDim> outLengths,
+                 const std::array<index_t, NumDstDim> outStrides,
+                 const std::array<int, NumReduceDim> reduceDims,
+                 float alpha,
+                 float beta,
+                 const InDataType* in_dev,
+                 OutDataType* out_dev,
+                 IndexDataType* out_index_dev,
+                 const InElementwiseOperation in_elementwise_op,
+                 const AccElementwiseOperation acc_elementwise_op)
+            : outLengths_{outLengths},
+              outStrides_{outStrides},
+              in_dev_{in_dev},
+              out_dev_{out_dev},
+              out_index_dev_{out_index_dev},
+              in_elementwise_op_{in_elementwise_op},
+              acc_elementwise_op_{acc_elementwise_op}
+        {
+            inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
+
+            alpha_ = type_convert<AccDataType>(alpha);
+            beta_  = type_convert<AccDataType>(beta);
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            if constexpr(NumInvariantDim == 0)
+                invariant_lowest_length = 1;
+            else
+                invariant_lowest_length = inLengths_[NumInvariantDim - 1];
+
+            reduce_lowest_length = inLengths_[Rank - 1];
+
+            numBlockTileIteration = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+
+            gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                       M_BlockTileSize;
+        }
+
+        std::array<index_t, Rank> inLengths_;
+        std::array<index_t, Rank> inStrides_;
+        std::array<index_t, NumDstDim> outLengths_;
+        std::array<index_t, NumDstDim> outStrides_;
+
+        AccDataType alpha_;
+        AccDataType beta_;
+
+        const InDataType* in_dev_;
+        OutDataType* out_dev_;
+        IndexDataType* out_index_dev_;
+
+        InElementwiseOperation in_elementwise_op_;
+        AccElementwiseOperation acc_elementwise_op_;
+
+        index_t invariant_lowest_length;
+        index_t reduce_lowest_length;
+        long_index_t invariant_total_length;
+        long_index_t reduce_total_length;
+
+        int numBlockTileIteration;
+        size_t gridSize;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto in_grid_desc_m_k =
+                DeviceReduceThreadWise::MakeSrc2dDescriptor(arg.inLengths_, arg.inStrides_);
+            const auto out_grid_desc_m =
+                DeviceReduceThreadWise::MakeDst1dDescriptor(arg.outLengths_, arg.outStrides_);
+            using InGridDesc_M_K = decltype(in_grid_desc_m_k);
+            using OutGridDesc_M  = decltype(out_grid_desc_m);
+
+            float avg_time = 0;
+
+            using GridwiseReduce =
+                GridwiseReduction_mk_to_m_threadwise<InDataType,
+                                                     OutDataType,
+                                                     AccDataType,
+                                                     IndexDataType,
+                                                     InGridDesc_M_K,
+                                                     OutGridDesc_M,
+                                                     ReduceOperation,
+                                                     InElementwiseOperation,
+                                                     AccElementwiseOperation,
+                                                     InMemoryDataOperationEnum::Set,
+                                                     PropagateNan,
+                                                     BlockSize,
+                                                     MThreadSliceSize,
+                                                     KThreadSliceSize,
+                                                     InSrcVectorDim,
+                                                     InSrcVectorSize,
+                                                     OutDstVectorSize>;
+
+            const auto kernel = kernel_reduce_threadwise<GridwiseReduce,
+                                                         OutputIndex,
+                                                         HaveIndexInput,
+                                                         InDataType,
+                                                         OutDataType,
+                                                         AccDataType,
+                                                         IndexDataType,
+                                                         InGridDesc_M_K,
+                                                         OutGridDesc_M,
+                                                         InElementwiseOperation,
+                                                         AccElementwiseOperation>;
+
+            avg_time = launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(arg.gridSize),
+                                              dim3(BlockSize),
+                                              0,
+                                              in_grid_desc_m_k,
+                                              out_grid_desc_m,
+                                              arg.in_elementwise_op_,
+                                              arg.acc_elementwise_op_,
+                                              arg.alpha_,
+                                              arg.in_dev_,
+                                              nullptr,
+                                              arg.beta_,
+                                              arg.out_dev_,
+                                              arg.out_index_dev_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if constexpr(InSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return (false);
+            }
+            else
+            {
+                if(pArg->inStrides_[NumInvariantDim - 1] != 1)
+                    return (false);
+
+                if(pArg->invariant_lowest_length % InSrcVectorSize != 0)
+                    return (false);
+            };
+        }
+        else
+        {
+            if(pArg->inStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(pArg->reduce_lowest_length % InSrcVectorSize != 0)
+                return (false);
+        };
+
+        // To improve
+        if(pArg->invariant_lowest_length % OutDstVectorSize != 0)
+            return (false);
+
+        // cases with big reduce_total_length should be handled by Blockwise kernel
+        if(pArg->reduce_total_length / KThreadSliceSize >= 32)
+            return (false);
+
+        return (true);
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> inLengths,
+                        const std::array<index_t, Rank> inStrides,
+                        const std::array<index_t, NumDstDim> outLengths,
+                        const std::array<index_t, NumDstDim> outStrides,
+                        const std::array<int, NumReduceDim> reduceDims,
+                        float alpha,
+                        float beta,
+                        const void* in_dev,
+                        const void* in_index_dev,
+                        void* out_dev,
+                        void* out_index_dev,
+                        const InElementwiseOperation in_elementwise_op,
+                        const AccElementwiseOperation acc_elementwise_op) override
+    {
+        (void)in_index_dev;
+
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStrides,
+                                          reduceDims,
+                                          alpha,
+                                          beta,
+                                          static_cast<const InDataType*>(in_dev),
+                                          static_cast<OutDataType*>(out_dev),
+                                          static_cast<IndexDataType*>(out_index_dev),
+                                          in_elementwise_op,
+                                          acc_elementwise_op);
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceReduceThreadWise<" << BlockSize << ",";
+        str << "M_C" << BlockSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << 1 << "_S" << KThreadSliceSize << ",";
+        str << "InSrcVectorDim_" << InSrcVectorDim << "_InSrcVectorSize_" << InSrcVectorSize << "_OutDstVectorSize_" << OutDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp
new file mode 100644
index 0000000..8630a2c
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp
@@ -0,0 +1,423 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          index_t Rank,
+          index_t NumReduceDim,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize>
+struct DeviceSoftmaxImpl : public DeviceSoftmax<InDataType,
+                                                AccDataType,
+                                                OutDataType,
+                                                InElementwiseOp,
+                                                AccElementwiseOp,
+                                                Rank>
+{
+    static constexpr index_t kRank            = Rank;
+    static constexpr index_t kNumReduceDim    = NumReduceDim;
+    static constexpr index_t kNumInvariantDim = Rank - NumReduceDim;
+
+    virtual index_t GetRank() const override { return kRank; }
+
+    virtual index_t GetNumReduceDim() const override { return kNumReduceDim; }
+
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumSrcDim = Rank;
+    static constexpr index_t NumDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
+                                    const std::vector<index_t>& inStrides,
+                                    int blkGroupSize,
+                                    int numBlockTileIteration)
+    {
+        const auto tupleSrcLengths =
+            generate_tuple([&](auto I) { return inLengths[I]; }, Number<Rank>{});
+        const auto tupleSrcStrides =
+            generate_tuple([&](auto I) { return inStrides[I]; }, Number<Rank>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, NumSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths = generate_tuple(
+                    [&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
+                const auto invariantDimLengths =
+                    generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
+
+    using GridwiseSoftmaxGeneric = GridwiseSoftmax_mk_to_mk<InDataType,
+                                                            OutDataType,
+                                                            AccDataType,
+                                                            GridDesc_M_K,
+                                                            BlockSize,
+                                                            MThreadClusterSize,
+                                                            KThreadClusterSize,
+                                                            MThreadSliceSize,
+                                                            KThreadSliceSize,
+                                                            InSrcVectorDim,
+                                                            InSrcVectorSize,
+                                                            OutDstVectorSize,
+                                                            false>;
+
+    using GridwiseSoftmaxSweepOnce = GridwiseSoftmax_mk_to_mk<InDataType,
+                                                              OutDataType,
+                                                              AccDataType,
+                                                              GridDesc_M_K,
+                                                              BlockSize,
+                                                              MThreadClusterSize,
+                                                              KThreadClusterSize,
+                                                              MThreadSliceSize,
+                                                              KThreadSliceSize,
+                                                              InSrcVectorDim,
+                                                              InSrcVectorSize,
+                                                              OutDstVectorSize,
+                                                              true>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::vector<index_t> inLengths,
+                 const std::vector<index_t> inStrides,
+                 const std::vector<index_t> reduceDims,
+                 AccDataType alpha,
+                 AccDataType beta,
+                 const InDataType* in_dev,
+                 OutDataType* out_dev,
+                 InElementwiseOp in_elementwise_op,
+                 AccElementwiseOp acc_elementwise_op)
+            : alpha_{alpha},
+              beta_{beta},
+              in_dev_{in_dev},
+              out_dev_{out_dev},
+              in_elementwise_op_{in_elementwise_op},
+              acc_elementwise_op_{acc_elementwise_op}
+        {
+            if(Rank != inLengths.size() || Rank != inStrides.size() ||
+               NumReduceDim != reduceDims.size())
+            {
+                throw std::runtime_error(
+                    "One of inLengths/inStrides/reduceDims has invalid size!"
+                    "\nExpected size inLengths: " +
+                    std::to_string(Rank) + ", inStrides: " + std::to_string(Rank) +
+                    ", reduceDims: " + std::to_string(NumReduceDim) +
+                    "\nBut have inLengths: " + std::to_string(inLengths.size()) +
+                    ", inStrides: " + std::to_string(inStrides.size()) +
+                    ", reduceDims: " + std::to_string(reduceDims.size()));
+            }
+
+            for(std::size_t i = 0; i < reduceDims.size(); ++i)
+            {
+                if(reduceDims[i] < 0 || reduceDims[i] >= Rank)
+                {
+                    throw std::runtime_error("Provided reduce dimension exceed input tensor Rank!"
+                                             "\nHave reduceDims[" +
+                                             std::to_string(i) +
+                                             "]: " + std::to_string(reduceDims[i]));
+                }
+            }
+
+            inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
+
+            long_index_t invariant_total_length;
+            long_index_t reduce_total_length;
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            if constexpr(NumInvariantDim == 0)
+                invariant_lowest_length_ = 1;
+            else
+                invariant_lowest_length_ = inLengths_[NumInvariantDim - 1];
+
+            blkGroupSize          = 1;
+            numBlockTileIteration = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+
+            gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                       M_BlockTileSize * blkGroupSize;
+        }
+
+        std::vector<index_t> inLengths_;
+        std::vector<index_t> inStrides_;
+
+        AccDataType alpha_;
+        AccDataType beta_;
+
+        const InDataType* in_dev_;
+        OutDataType* out_dev_;
+
+        InElementwiseOp in_elementwise_op_;
+        AccElementwiseOp acc_elementwise_op_;
+
+        index_t invariant_lowest_length_;
+
+        int blkGroupSize;
+        int numBlockTileIteration;
+        size_t gridSize;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto in_grid_desc_m_k = DeviceSoftmaxImpl::MakeSrc2dDescriptor(
+                arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.numBlockTileIteration);
+            const auto out_grid_desc_m_k = DeviceSoftmaxImpl::MakeSrc2dDescriptor(
+                arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.numBlockTileIteration);
+
+            bool sweep_once =
+                in_grid_desc_m_k.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
+
+            const auto kernel_main = sweep_once ? kernel_softmax<GridwiseSoftmaxSweepOnce,
+                                                                 InDataType,
+                                                                 OutDataType,
+                                                                 AccDataType,
+                                                                 GridDesc_M_K>
+                                                : kernel_softmax<GridwiseSoftmaxGeneric,
+                                                                 InDataType,
+                                                                 OutDataType,
+                                                                 AccDataType,
+                                                                 GridDesc_M_K>;
+
+            float avg_time = 0;
+
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize),
+                                               dim3(BlockSize),
+                                               0,
+                                               in_grid_desc_m_k,
+                                               out_grid_desc_m_k,
+                                               arg.blkGroupSize,
+                                               arg.numBlockTileIteration,
+                                               arg.alpha_,
+                                               arg.in_dev_,
+                                               arg.beta_,
+                                               arg.out_dev_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(InSrcVectorDim == 0)
+        {
+            if constexpr(kNumInvariantDim == 0)
+            {
+                return false;
+            }
+            else
+            {
+                if(arg.inStrides_[kNumInvariantDim - 1] != 1 && InSrcVectorSize != 1)
+                {
+                    return false;
+                }
+                if(arg.invariant_lowest_length_ % InSrcVectorSize != 0)
+                {
+                    return false;
+                }
+            }
+        }
+        else
+        {
+            if(arg.inStrides_[Rank - 1] != 1 && InSrcVectorSize != 1)
+            {
+                return false;
+            }
+            if(arg.inLengths_[Rank - 1] % InSrcVectorSize != 0)
+            {
+                return false;
+            }
+        }
+
+        // To improve
+        if(kNumInvariantDim > 0 && arg.invariant_lowest_length_ % OutDstVectorSize != 0)
+        {
+            return false;
+        }
+
+        if(arg.inLengths_[Rank - 1] % OutDstVectorSize != 0)
+        {
+            return false;
+        }
+
+        return true;
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const std::vector<index_t> inLengths,
+                             const std::vector<index_t> inStrides,
+                             const std::vector<int> reduceDims,
+                             const AccDataType alpha,
+                             const AccDataType beta,
+                             const InDataType* in_dev,
+                             OutDataType* out_dev,
+                             InElementwiseOp in_elementwise_op,
+                             AccElementwiseOp acc_elementwise_op)
+    {
+        return Argument{inLengths,
+                        inStrides,
+                        reduceDims,
+                        alpha,
+                        beta,
+                        in_dev,
+                        out_dev,
+                        in_elementwise_op,
+                        acc_elementwise_op};
+    };
+
+    //
+    // @brief      Makes a pointer to Argument class.
+    //
+    // @param[in]  inLengths           Input tensor extent(s) from high to low dimension
+    // @param[in]  inStrides           Input tensor stride(s) from high to low dimension
+    // @param[in]  reduceDims          The dimension(s) the normalization operation is applied
+    // @param[in]  alpha               Typeless pointer in host memory storing the alpha scaling
+    //                                 value as type AccDataType
+    // @param[in]  beta                Typeless pointer in host memory storing the beta scaling
+    //                                 value as type AccDataType
+    // @param[in]  in_dev              Typeless const pointer in device memory storing the input
+    //                                 tensor
+    // @param      out_dev             Typeless pointer in device memory storing the output tensor
+    // @param[in]  in_elementwise_op   The input elementwise operation.
+    // @param[in]  acc_elementwise_op  The accumulation elementwise operation.
+    //
+    // @return     Unique pointer to the Argument class.
+    //
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const std::vector<index_t> inLengths,
+                                                      const std::vector<index_t> inStrides,
+                                                      const std::vector<int> reduceDims,
+                                                      const void* alpha,
+                                                      const void* beta,
+                                                      const void* in_dev,
+                                                      void* out_dev,
+                                                      InElementwiseOp in_elementwise_op,
+                                                      AccElementwiseOp acc_elementwise_op) override
+    {
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          reduceDims,
+                                          *static_cast<const AccDataType*>(alpha),
+                                          *static_cast<const AccDataType*>(beta),
+                                          static_cast<const InDataType*>(in_dev),
+                                          static_cast<OutDataType*>(out_dev),
+                                          in_elementwise_op,
+                                          acc_elementwise_op);
+    };
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceReduceSoftmax<" 
+            << Rank << "," << NumReduceDim << "," << BlockSize << ","
+            << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ","
+            << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ","
+            << "InSrcVectorDim_" << InSrcVectorDim 
+            << "_InSrcVectorSize_" << InSrcVectorSize 
+            << "_OutDstVectorSize_" << OutDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_sparse_embedding3_forward_layernorm.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_sparse_embedding3_forward_layernorm.hpp
new file mode 100644
index 0000000..1f2b46e
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_sparse_embedding3_forward_layernorm.hpp
@@ -0,0 +1,210 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_sparse_embedding3_forward_layernorm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename EmbType,
+          typename IndexType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename OutType,
+          ck::index_t BlockSize,
+          ck::index_t DimClusterSize,
+          ck::index_t RowClusterSize,
+          ck::index_t DimPerBlock,
+          ck::index_t RowPerBlock,
+          ck::index_t DimThreadSize,
+          ck::index_t RowVectorSize>
+struct DeviceSparseEmbedding3ForwardLayernorm : public BaseOperator
+{
+
+    static auto MakeOutputDescriptor(const index_t index_length, const index_t rows)
+    {
+        return make_naive_tensor_descriptor_packed(make_tuple(index_length, rows));
+    }
+
+    struct Argument : public BaseArgument
+    {
+        Argument(OutType* p_out,
+                 const EmbType* p_emb_a,
+                 const EmbType* p_emb_b,
+                 const EmbType* p_emb_c,
+                 const IndexType* p_index_a,
+                 const IndexType* p_index_b,
+                 const IndexType* p_index_c,
+                 const GammaDataType* p_gamma,
+                 const BetaDataType* p_beta,
+                 const ck::index_t NumRows,
+                 const ck::index_t EmbeddingDim,
+                 const ck::index_t IndexLength,
+                 const AccDataType epsilon)
+            : p_out_(p_out),
+              p_emb_a_(p_emb_a),
+              p_emb_b_(p_emb_b),
+              p_emb_c_(p_emb_c),
+              p_index_a_(p_index_a),
+              p_index_b_(p_index_b),
+              p_index_c_(p_index_c),
+              p_gamma_(p_gamma),
+              p_beta_(p_beta),
+              NumRows_(NumRows),
+              EmbeddingDim_(EmbeddingDim),
+              IndexLength_(IndexLength),
+              epsilon_(epsilon)
+        {
+            grid_size_ = (IndexLength + DimClusterSize - 1) / DimClusterSize;
+        }
+
+        OutType* p_out_;
+        const EmbType* p_emb_a_;
+        const EmbType* p_emb_b_;
+        const EmbType* p_emb_c_;
+        const IndexType* p_index_a_;
+        const IndexType* p_index_b_;
+        const IndexType* p_index_c_;
+        const GammaDataType* p_gamma_;
+        const BetaDataType* p_beta_;
+        ck::index_t NumRows_;
+        ck::index_t EmbeddingDim_;
+        ck::index_t IndexLength_;
+        AccDataType epsilon_;
+
+        size_t grid_size_;
+    };
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(void* p_out,
+                                                              const void* p_emb_a,
+                                                              const void* p_emb_b,
+                                                              const void* p_emb_c,
+                                                              const void* p_index_a,
+                                                              const void* p_index_b,
+                                                              const void* p_index_c,
+                                                              const void* p_gamma,
+                                                              const void* p_beta,
+                                                              ck::index_t NumRows,
+                                                              ck::index_t EmbeddingDim,
+                                                              ck::index_t IndexLength,
+                                                              const AccDataType epsilon)
+    {
+        return std::make_unique<Argument>(reinterpret_cast<OutType*>(p_out),
+                                          reinterpret_cast<const EmbType*>(p_emb_a),
+                                          reinterpret_cast<const EmbType*>(p_emb_b),
+                                          reinterpret_cast<const EmbType*>(p_emb_c),
+                                          reinterpret_cast<const IndexType*>(p_index_a),
+                                          reinterpret_cast<const IndexType*>(p_index_b),
+                                          reinterpret_cast<const IndexType*>(p_index_c),
+                                          reinterpret_cast<const GammaDataType*>(p_gamma),
+                                          reinterpret_cast<const BetaDataType*>(p_beta),
+                                          NumRows,
+                                          EmbeddingDim,
+                                          IndexLength,
+                                          epsilon);
+    }
+
+    using GridwiseSparseEmbedding =
+        GridwiseSparseEmbedding3ForwardLayernorm<EmbType,
+                                                 IndexType,
+                                                 GammaDataType,
+                                                 BetaDataType,
+                                                 AccDataType,
+                                                 OutType,
+                                                 decltype(MakeOutputDescriptor(1, 1)),
+                                                 BlockSize,
+                                                 DimClusterSize,
+                                                 RowClusterSize,
+                                                 DimPerBlock,
+                                                 RowPerBlock,
+                                                 DimThreadSize,
+                                                 RowVectorSize>;
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            auto out_desc = MakeOutputDescriptor(arg.IndexLength_, arg.EmbeddingDim_);
+            const auto kernel_main =
+                kernel_sparse_embedding3_forward_layernorm<GridwiseSparseEmbedding,
+                                                           EmbType,
+                                                           IndexType,
+                                                           GammaDataType,
+                                                           BetaDataType,
+                                                           AccDataType,
+                                                           OutType,
+                                                           decltype(out_desc)>;
+            float avg_time = 0;
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.grid_size_),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.p_out_,
+                                               arg.p_emb_a_,
+                                               arg.p_emb_b_,
+                                               arg.p_emb_c_,
+                                               arg.p_index_a_,
+                                               arg.p_index_b_,
+                                               arg.p_index_c_,
+                                               arg.p_gamma_,
+                                               arg.p_beta_,
+                                               out_desc,
+                                               arg.epsilon_);
+
+            return (avg_time);
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    static bool IsSupportedArgument(const Argument* p_arg)
+    {
+        return (RowPerBlock == p_arg->EmbeddingDim_) && (p_arg->NumRows_ % DimPerBlock == 0);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(dynamic_cast<const Argument*>(p_arg));
+    }
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>();
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceSparseEmbedding3ForwardLayernorm_"<< BlockSize << "_" <<
+            DimClusterSize << "x" << RowClusterSize << "_" <<
+            DimPerBlock << "x" << RowPerBlock << "_" <<
+            DimThreadSize << "x" << RowVectorSize;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/masking_specialization.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/masking_specialization.hpp
new file mode 100644
index 0000000..ea0f589
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/masking_specialization.hpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct MaskingSpecialization
+{
+    MaskDisabled,
+    MaskOutUpperTriangle
+};
+
+inline std::string getMaskingSpecializationString(const MaskingSpecialization& s)
+{
+    switch(s)
+    {
+    case MaskingSpecialization::MaskDisabled: return "MaskDisabled";
+    case MaskingSpecialization::MaskOutUpperTriangle: return "MaskOutUpperTriangle";
+    default: return "Unrecognized specialization!";
+    }
+}
+
+struct MaskDisabledPredicate
+{
+    __host__ __device__ constexpr bool operator()(index_t /*m*/, index_t /*n*/) const
+    {
+        return false;
+    };
+
+    __host__ __device__ constexpr bool
+        IsTileSkippable(index_t /*m*/, index_t /*n*/, index_t /*m_tile*/, index_t /*n_tile*/) const
+    {
+        return false;
+    }
+};
+
+struct MaskOutUpperTrianglePredicate
+{
+    __host__ __device__ constexpr bool operator()(index_t m, index_t n) const { return n > m; }
+
+    __host__ __device__ constexpr bool
+    IsTileSkippable(index_t m, index_t n, index_t m_tile, index_t /*n_tile*/) const
+    {
+        return operator()(m + m_tile - 1, n);
+    }
+};
+
+// to track the points which need to be set to -inf on C0
+// Note: no need to reset M padding value, because they will not be stored out.
+template <typename MaskOutPredicate>
+struct C0MatrixMask_impl
+{
+    C0MatrixMask_impl(index_t NRaw) : NRaw_(NRaw), predicate_(MaskOutPredicate{}) {}
+
+    __host__ __device__ constexpr bool IsNOutOfBound(/*index_t m, */ index_t n) const
+    {
+        return n >= NRaw_;
+    }
+
+    __host__ __device__ constexpr bool IsMaskedElement(index_t m, index_t n) const
+    {
+        return predicate_(m, n) || IsNOutOfBound(n);
+    }
+
+    __host__ __device__ constexpr bool
+    IsTileSkippable(index_t m, index_t n, index_t m_tile, index_t n_tile) const
+    {
+        return predicate_.IsTileSkippable(m, n, m_tile, n_tile);
+    }
+
+    private:
+    // index_t MRaw_;
+    index_t NRaw_;
+    MaskOutPredicate predicate_;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/matrix_padder.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/matrix_padder.hpp
new file mode 100644
index 0000000..70e61bc
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/matrix_padder.hpp
@@ -0,0 +1,382 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename TensorDesc,
+          typename TileLengths, // Tuple<...>
+          typename DoPads>      // Sequence<bool, bool, ...>
+__host__ __device__ constexpr auto
+PadTensorDescriptor(const TensorDesc& desc, const TileLengths& tile_lengths, DoPads)
+{
+    constexpr index_t num_dim = DoPads::Size();
+
+    static_assert(num_dim == TileLengths::Size() && num_dim == TensorDesc::GetNumOfDimension(),
+                  "wrong! inconsistent # of dimensions");
+
+    // transforms
+    const auto transforms = generate_tuple(
+        [&](auto idim) {
+            const auto MRaw = desc.GetLength(idim);
+
+            const auto MPerTile = tile_lengths[idim];
+
+            const auto M = math::integer_divide_ceil(MRaw, MPerTile) * MPerTile;
+
+            const auto MPad = M - MRaw;
+
+            const bool DoPadM = DoPads::At(idim);
+
+            const auto MTransform = conditional_expr<DoPadM>(make_right_pad_transform(MRaw, MPad),
+                                                             make_pass_through_transform(MRaw));
+
+            return MTransform;
+        },
+        Number<num_dim>{});
+
+    // lower dimension Id
+    const auto lower_dimss =
+        generate_tuple([&](auto idim) { return Sequence<idim.value>{}; }, Number<num_dim>{});
+
+    // upper dimension Id
+    const auto upper_dimss = lower_dimss;
+
+    return transform_tensor_descriptor(desc, transforms, lower_dimss, upper_dimss);
+}
+
+// M/N/K/OPerTileType could be index_t or Number<>
+template <GemmSpecialization GemmSpec,
+          typename MPerTileType,
+          typename NPerTileType,
+          typename KPerTileType,
+          typename OPerTileType>
+struct GemmGemmPadder
+{
+    // TODO: hard to scale; use mask instead
+    static constexpr bool PadM =
+        GemmSpec == GemmSpecialization::MPadding || GemmSpec == GemmSpecialization::MNPadding ||
+        GemmSpec == GemmSpecialization::MKPadding || GemmSpec == GemmSpecialization::MNKPadding ||
+        GemmSpec == GemmSpecialization::MOPadding || GemmSpec == GemmSpecialization::MNOPadding ||
+        GemmSpec == GemmSpecialization::MKOPadding || GemmSpec == GemmSpecialization::MNKOPadding;
+    static constexpr bool PadN =
+        GemmSpec == GemmSpecialization::NPadding || GemmSpec == GemmSpecialization::MNPadding ||
+        GemmSpec == GemmSpecialization::NKPadding || GemmSpec == GemmSpecialization::MNKPadding ||
+        GemmSpec == GemmSpecialization::NOPadding || GemmSpec == GemmSpecialization::MNOPadding ||
+        GemmSpec == GemmSpecialization::NKOPadding || GemmSpec == GemmSpecialization::MNKOPadding;
+    static constexpr bool PadK =
+        GemmSpec == GemmSpecialization::KPadding || GemmSpec == GemmSpecialization::MKPadding ||
+        GemmSpec == GemmSpecialization::NKPadding || GemmSpec == GemmSpecialization::MNKPadding ||
+        GemmSpec == GemmSpecialization::KOPadding || GemmSpec == GemmSpecialization::MKOPadding ||
+        GemmSpec == GemmSpecialization::NKOPadding || GemmSpec == GemmSpecialization::MNKOPadding;
+    static constexpr bool PadO =
+        GemmSpec == GemmSpecialization::OPadding || GemmSpec == GemmSpecialization::MOPadding ||
+        GemmSpec == GemmSpecialization::NOPadding || GemmSpec == GemmSpecialization::KOPadding ||
+        GemmSpec == GemmSpecialization::MNOPadding || GemmSpec == GemmSpecialization::MKOPadding ||
+        GemmSpec == GemmSpecialization::NKOPadding || GemmSpec == GemmSpecialization::MNKOPadding;
+
+    // A[M, K]
+    template <typename ADesc_MRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadADescriptor_M_K(const ADesc_MRaw_KRaw& a_desc_mraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            a_desc_mraw_kraw, make_tuple(MPerTile_, KPerTile_), Sequence<PadM, PadK>{});
+    }
+
+    // B[K, N]
+    template <typename BDesc_NRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadBDescriptor_N_K(const BDesc_NRaw_KRaw& b_desc_nraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            b_desc_nraw_kraw, make_tuple(NPerTile_, KPerTile_), Sequence<PadN, PadK>{});
+    }
+
+    // B1[Gemm1N, Gemm1K] = B1[O, N]
+    template <typename B1Desc_NRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadB1Descriptor_N_K(const B1Desc_NRaw_KRaw& b1_desc_nraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            b1_desc_nraw_kraw, make_tuple(OPerTile_, NPerTile_), Sequence<PadO, PadN>{});
+    }
+
+    // C[M, Gemm1N] = C[M, O]
+    template <typename CDesc_MRaw_NRaw>
+    __host__ __device__ constexpr auto
+    PadCDescriptor_M_N(const CDesc_MRaw_NRaw& c_desc_mraw_nraw) const
+    {
+        return PadTensorDescriptor(
+            c_desc_mraw_nraw, make_tuple(MPerTile_, OPerTile_), Sequence<PadM, PadO>{});
+    }
+
+    MPerTileType MPerTile_;
+    NPerTileType NPerTile_;
+    KPerTileType KPerTile_;
+    OPerTileType OPerTile_;
+};
+
+// M/N/KPerTileType could be index_t or Number<>
+template <GemmSpecialization GemmSpec,
+          typename MPerTileType,
+          typename NPerTileType,
+          typename KPerTileType>
+struct GemmPadder
+{
+    static constexpr bool PadM =
+        (GemmSpec == GemmSpecialization::MPadding || GemmSpec == GemmSpecialization::MNPadding ||
+         GemmSpec == GemmSpecialization::MKPadding || GemmSpec == GemmSpecialization::MNKPadding);
+    static constexpr bool PadN =
+        (GemmSpec == GemmSpecialization::NPadding || GemmSpec == GemmSpecialization::MNPadding ||
+         GemmSpec == GemmSpecialization::NKPadding || GemmSpec == GemmSpecialization::MNKPadding);
+    static constexpr bool PadK =
+        (GemmSpec == GemmSpecialization::KPadding || GemmSpec == GemmSpecialization::MKPadding ||
+         GemmSpec == GemmSpecialization::NKPadding || GemmSpec == GemmSpecialization::MNKPadding);
+
+    template <typename ADesc_MRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadADescriptor_M_K(const ADesc_MRaw_KRaw& a_desc_mraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            a_desc_mraw_kraw, make_tuple(MPerTile_, KPerTile_), Sequence<PadM, PadK>{});
+    }
+
+    template <typename BDesc_NRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadBDescriptor_N_K(const BDesc_NRaw_KRaw& b_desc_nraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            b_desc_nraw_kraw, make_tuple(NPerTile_, KPerTile_), Sequence<PadN, PadK>{});
+    }
+
+    template <typename CDesc_MRaw_NRaw>
+    __host__ __device__ constexpr auto
+    PadCDescriptor_M_N(const CDesc_MRaw_NRaw& c_desc_mraw_nraw) const
+    {
+        return PadTensorDescriptor(
+            c_desc_mraw_nraw, make_tuple(MPerTile_, NPerTile_), Sequence<PadM, PadN>{});
+    }
+
+    MPerTileType MPerTile_;
+    NPerTileType NPerTile_;
+    KPerTileType KPerTile_;
+};
+
+// Alias of GemmPadder; to deprecate
+template <GemmSpecialization GemmSpec,
+          typename MPerTileType,
+          typename NPerTileType,
+          typename KPerTileType>
+struct MatrixPadder : public GemmPadder<GemmSpec, MPerTileType, NPerTileType, KPerTileType>
+{
+};
+
+// M/N/KPerTileType could be index_t or Number<>
+template <bool PadM,
+          bool PadN,
+          bool PadK,
+          typename MPerTileType,
+          typename NPerTileType,
+          typename KPerTileType>
+struct GemmPadder_v2
+{
+    template <typename ADesc_MRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadADescriptor_M_K(const ADesc_MRaw_KRaw& a_desc_mraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            a_desc_mraw_kraw, make_tuple(MPerTile_, KPerTile_), Sequence<PadM, PadK>{});
+    }
+
+    template <typename BDesc_NRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadBDescriptor_N_K(const BDesc_NRaw_KRaw& b_desc_nraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            b_desc_nraw_kraw, make_tuple(NPerTile_, KPerTile_), Sequence<PadN, PadK>{});
+    }
+
+    template <typename CDesc_MRaw_NRaw>
+    __host__ __device__ constexpr auto
+    PadCDescriptor_M_N(const CDesc_MRaw_NRaw& c_desc_mraw_nraw) const
+    {
+        return PadTensorDescriptor(
+            c_desc_mraw_nraw, make_tuple(MPerTile_, NPerTile_), Sequence<PadM, PadN>{});
+    }
+
+    MPerTileType MPerTile_;
+    NPerTileType NPerTile_;
+    KPerTileType KPerTile_;
+};
+
+// M/N/KPerTileType could be index_t or Number<>
+template <bool PadM,
+          bool PadN,
+          bool PadK,
+          typename MPerTileType,
+          typename NPerTileType,
+          typename KPerTileType>
+struct MatrixPadder_v2
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    template <typename ADesc_MRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadADescriptor_M_K(const ADesc_MRaw_KRaw& a_desc_mraw_kraw) const
+    {
+        const auto MRaw = a_desc_mraw_kraw.GetLength(I0);
+        const auto KRaw = a_desc_mraw_kraw.GetLength(I1);
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerTile_) * MPerTile_;
+        const auto K = math::integer_divide_ceil(KRaw, KPerTile_) * KPerTile_;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(PadM && PadK)
+        {
+            // pad both M and K
+            return transform_tensor_descriptor(a_desc_mraw_kraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(KRaw, KPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(PadM && (!PadK))
+        {
+            // pad M, but not K
+            return transform_tensor_descriptor(
+                a_desc_mraw_kraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(KRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr((!PadM) && PadK)
+        {
+            // pad K, but not M
+            return transform_tensor_descriptor(
+                a_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or K
+            return a_desc_mraw_kraw;
+        }
+    }
+
+    template <typename BDesc_NRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadBDescriptor_N_K(const BDesc_NRaw_KRaw& b_desc_nraw_kraw) const
+    {
+        const auto NRaw = b_desc_nraw_kraw.GetLength(I0);
+        const auto KRaw = b_desc_nraw_kraw.GetLength(I1);
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerTile_) * NPerTile_;
+        const auto K = math::integer_divide_ceil(KRaw, KPerTile_) * KPerTile_;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(PadN && PadK)
+        {
+            // pad both N and K
+            return transform_tensor_descriptor(b_desc_nraw_kraw,
+                                               make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                          make_right_pad_transform(KRaw, KPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(PadN && (!PadK))
+        {
+            // pad N, but not K
+            return transform_tensor_descriptor(
+                b_desc_nraw_kraw,
+                make_tuple(make_right_pad_transform(NRaw, NPad), make_pass_through_transform(KRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr((!PadN) && PadK)
+        {
+            // pad K, but not N
+            return transform_tensor_descriptor(
+                b_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad N or K
+            return b_desc_nraw_kraw;
+        }
+    }
+
+    template <typename CDesc_MRaw_NRaw>
+    __host__ __device__ constexpr auto
+    PadCDescriptor_M_N(const CDesc_MRaw_NRaw& c_desc_mraw_nraw) const
+    {
+        const auto MRaw = c_desc_mraw_nraw.GetLength(I0);
+        const auto NRaw = c_desc_mraw_nraw.GetLength(I1);
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerTile_) * MPerTile_;
+        const auto N = math::integer_divide_ceil(NRaw, NPerTile_) * NPerTile_;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(PadM && PadN)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(PadM && (!PadN))
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr((!PadM) && PadN)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_desc_mraw_nraw;
+        }
+    }
+
+    MPerTileType MPerTile_;
+    NPerTileType NPerTile_;
+    KPerTileType KPerTile_;
+};
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp
new file mode 100644
index 0000000..d353183
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp
@@ -0,0 +1,186 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+// FIXME: can it be replaced with ck::Tuple?
+#include <tuple>
+
+namespace ck {
+
+// The templated struct reduce_binary_operator maps the enum Ids of binary operators to their
+// respective functor classes.
+// The boolean member "indexable" are also provided in reduce_binary_operactor for
+// easier checking by the upper-layer codes in the kernels.
+
+template <ReduceTensorOp Op>
+struct reduce_binary_operator;
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::ADD>
+{
+    using opType = reduce::Add;
+
+    static constexpr bool indexable = false;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::MUL>
+{
+    using opType = reduce::Mul;
+
+    static constexpr bool indexable = false;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::MIN>
+{
+    using opType = reduce::Min;
+
+    static constexpr bool indexable = true;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::MAX>
+{
+    using opType = reduce::Max;
+
+    static constexpr bool indexable = true;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::AMAX>
+{
+    using opType = reduce::AMax;
+
+    static constexpr bool indexable = true;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::AVG>
+{
+    using opType = reduce::Add;
+
+    static constexpr bool indexable = false;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::NORM1>
+{
+    using opType = reduce::Add;
+
+    static constexpr bool indexable = false;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::NORM2>
+{
+    using opType = reduce::Add;
+
+    static constexpr bool indexable = false;
+};
+
+// The templated struct reduce_unary_operator maps the enum Ids of Reduce operators to two unary
+// functor classes.
+// The two unary functors are called before and afer the Reduction is executed respectively
+template <ReduceTensorOp Op, bool IsFirstReduce, bool IsLastReduce>
+struct reduce_unary_operator
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::PassThrough;
+    using AccElementwiseOperation = tensor_operation::element_wise::PassThrough;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
+};
+
+template <bool IsFirstReduce>
+struct reduce_unary_operator<ReduceTensorOp::AVG, IsFirstReduce, true>
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::PassThrough;
+    using AccElementwiseOperation = tensor_operation::element_wise::UnaryDivide;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{reduceLength});
+    };
+};
+
+template <bool IsLastReduce>
+struct reduce_unary_operator<ReduceTensorOp::NORM1, true, IsLastReduce>
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::UnaryAbs;
+    using AccElementwiseOperation = tensor_operation::element_wise::PassThrough;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
+};
+
+template <bool IsLastReduce>
+struct reduce_unary_operator<ReduceTensorOp::AMAX, true, IsLastReduce>
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::UnaryAbs;
+    using AccElementwiseOperation = tensor_operation::element_wise::PassThrough;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
+};
+
+template <>
+struct reduce_unary_operator<ReduceTensorOp::NORM2, true, false>
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::UnarySquare;
+    using AccElementwiseOperation = tensor_operation::element_wise::PassThrough;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
+};
+
+template <>
+struct reduce_unary_operator<ReduceTensorOp::NORM2, true, true>
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::UnarySquare;
+    using AccElementwiseOperation = tensor_operation::element_wise::UnarySqrt;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
+};
+
+template <>
+struct reduce_unary_operator<ReduceTensorOp::NORM2, false, true>
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::PassThrough;
+    using AccElementwiseOperation = tensor_operation::element_wise::UnarySqrt;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/tensor_layout.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/tensor_layout.hpp
new file mode 100644
index 0000000..b444274
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/tensor_layout.hpp
@@ -0,0 +1,417 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_layout {
+
+struct BaseTensorLayout
+{
+};
+
+namespace gemm {
+
+struct RowMajor : public BaseTensorLayout
+{
+    static constexpr const char* name = "RowMajor";
+};
+
+struct ColumnMajor : public BaseTensorLayout
+{
+    static constexpr const char* name = "ColumnMajor";
+};
+} // namespace gemm
+
+namespace convolution {
+
+// input tensor
+// packed NCW/NCHW/NCDHW
+struct NCW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NCW";
+};
+
+struct NCHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NCHW";
+};
+
+struct NCDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NCDHW";
+};
+
+// packed GNCW/GNCHW/GNCDHW
+struct GNCW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNCW";
+};
+
+struct GNCHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNCHW";
+};
+
+struct GNCDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNCDHW";
+};
+
+// input tensor
+// packed NWC/NHWC/NDHWC
+struct NWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWC";
+};
+
+struct NHWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWC";
+};
+
+struct NDHWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NDHWC";
+};
+
+// input tensor
+// packed GNWC/GNHWC/GNDHWC
+struct GNWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNWC";
+};
+
+struct GNHWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNHWC";
+};
+
+struct GNDHWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNDHWC";
+};
+
+// for input bias
+struct GC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GC";
+};
+
+// input tensor
+// packed NWGC/NHWGC/NDHWGC
+struct NWGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWGC";
+};
+
+struct NHWGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWGC";
+};
+
+struct NDHWGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NDHWGC";
+};
+
+// input tensor
+// strided layout
+struct G_NW_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NW_C";
+};
+
+struct G_NHW_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NHW_C";
+};
+
+struct G_NDHW_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NDHW_C";
+};
+
+// for input bias
+struct G_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_C";
+};
+
+// weight tensor
+// packed KCX/KCYX/KCZYX
+struct KCX : public BaseTensorLayout
+{
+    static constexpr const char* name = "KCX";
+};
+
+struct KCYX : public BaseTensorLayout
+{
+    static constexpr const char* name = "KCYX";
+};
+
+struct KCZYX : public BaseTensorLayout
+{
+    static constexpr const char* name = "KCZYX";
+};
+
+// weight tensor
+// packed KCX/KCYX/KCZYX
+struct GKCX : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKCX";
+};
+
+struct GKCYX : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKCYX";
+};
+
+struct GKCZYX : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKCZYX";
+};
+
+// weight tensor
+// packed KXC/KYXC/KZYXC
+struct KXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KXC";
+};
+
+struct KYXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KYXC";
+};
+
+struct KZYXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KZYXC";
+};
+
+// weight tensor
+// packed GKXC/GKYXC/GKZYXC
+struct GKXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKXC";
+};
+
+struct GKYXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKYXC";
+};
+
+struct GKZYXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKZYXC";
+};
+
+// weight tensor
+// packed KXGC/KYXGC/KZYXGC
+struct KXGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KXGC";
+};
+
+struct KYXGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KYXGC";
+};
+
+struct KZYXGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KZYXGC";
+};
+
+// weight tensor
+// strided
+struct G_K_X_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_K_X_C";
+};
+
+struct G_K_YX_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_K_YX_C";
+};
+
+struct G_K_ZYX_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_K_ZYX_C";
+};
+
+// output tensor
+// packed NKW/NKHW/NKDHW
+struct NKW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NKW";
+};
+
+struct NKHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NKHW";
+};
+
+struct NKDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NKDHW";
+};
+
+// output tensor
+// packed GNKW/GNKHW/GNKDHW
+struct GNKW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNKW";
+};
+
+struct GNKHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNKHW";
+};
+
+struct GNKDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNKDHW";
+};
+
+// output tensor
+// packed NWK/NHWK/NDHWK
+struct NWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWK";
+};
+
+struct NHWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWK";
+};
+
+struct NDHWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NDHWK";
+};
+
+// output tensor
+// packed GNWK/GNHWK/GNDHWK
+struct GNWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNWK";
+};
+
+struct GNHWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNHWK";
+};
+
+struct GNDHWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNDHWK";
+};
+
+// for output bias
+struct GK : public BaseTensorLayout
+{
+    static constexpr const char* name = "GK";
+};
+
+// output tensor
+// packed NWGK/NHWGK/NDHWGK
+struct NWGK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWGK";
+};
+
+struct NHWGK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWGK";
+};
+
+struct NDHWGK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NDHWGK";
+};
+
+// output tensor
+// strided layout
+struct G_NW_K : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NW_K";
+};
+
+struct G_NHW_K : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NHW_K";
+};
+
+struct G_NDHW_K : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NDHW_K";
+};
+
+// for output bias
+struct G_K : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_K";
+};
+
+// K-reduced output tensor (packed)
+struct GNW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNW";
+};
+
+struct GNHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNHW";
+};
+
+struct GNDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNDHW";
+};
+
+// K-reduced output tensor (packed)
+struct NWG : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWG";
+};
+
+struct NHWG : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWG";
+};
+
+struct NDHWG : public BaseTensorLayout
+{
+    static constexpr const char* name = "NDHWG";
+};
+
+// K-reduced output tensor (strided)
+struct G_NW : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NW";
+};
+
+struct G_NHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NHW";
+};
+
+struct G_NDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NDHW";
+};
+
+} // namespace convolution
+
+template <
+    typename Layout,
+    typename std::enable_if<std::is_base_of<BaseTensorLayout, Layout>::value, bool>::type = false>
+std::ostream& operator<<(std::ostream& os, const Layout&)
+{
+    os << Layout::name;
+    return os;
+}
+
+} // namespace tensor_layout
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/tensor_specialization.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/tensor_specialization.hpp
new file mode 100644
index 0000000..0ec0df2
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/tensor_specialization.hpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct TensorSpecialization
+{
+    Default,
+    Packed
+};
+
+inline std::string getTensorSpecializationString(const TensorSpecialization& s)
+{
+    switch(s)
+    {
+    case TensorSpecialization::Default: return "Default";
+    case TensorSpecialization::Packed: return "Packed";
+    default: return "Unrecognized specialization!";
+    }
+}
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/welford_helper.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/welford_helper.hpp
new file mode 100644
index 0000000..6c909b7
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/welford_helper.hpp
@@ -0,0 +1,89 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t K_BlockTileSize, index_t KThreadSliceSize>
+struct GetReduceCountPerThreadForBlockwiseWelford
+{
+    GetReduceCountPerThreadForBlockwiseWelford(index_t numBlockTileIteration,
+                                               long_index_t reduce_length)
+        : numBlockTileIteration_{numBlockTileIteration}
+    {
+        count_in_last_tile_ = reduce_length % K_BlockTileSize;
+    };
+
+    __device__ index_t operator()(index_t thread_k_cluster_id) const
+    {
+        if(count_in_last_tile_ == 0)
+            return (KThreadSliceSize * numBlockTileIteration_);
+        else
+        {
+            index_t num_complete_slice  = count_in_last_tile_ / KThreadSliceSize;
+            index_t count_in_last_slice = count_in_last_tile_ % KThreadSliceSize;
+
+            if(thread_k_cluster_id < num_complete_slice)
+                return (KThreadSliceSize * numBlockTileIteration_);
+            else if(thread_k_cluster_id == num_complete_slice)
+                return (KThreadSliceSize * (numBlockTileIteration_ - 1) + count_in_last_slice);
+            else
+                return (KThreadSliceSize * (numBlockTileIteration_ - 1));
+        };
+    };
+
+    index_t numBlockTileIteration_;
+    index_t count_in_last_tile_;
+};
+
+template <index_t K_BlockTileSize, index_t KThreadSliceSize>
+struct GetReduceCountPerThreadForMultiblockWelford
+{
+    GetReduceCountPerThreadForMultiblockWelford(index_t blkGroupSize,
+                                                index_t numBlockTileIteration,
+                                                long_index_t reduce_length)
+        : blkGroupSize_(blkGroupSize), numBlockTileIteration_{numBlockTileIteration}
+    {
+        last_block_reduce_length_ =
+            reduce_length - K_BlockTileSize * numBlockTileIteration_ * (blkGroupSize_ - 1);
+        numBlockTileIterationByLastBlock_ =
+            (last_block_reduce_length_ + K_BlockTileSize - 1) / K_BlockTileSize;
+    };
+
+    __device__ index_t operator()(index_t block_local_id, index_t thread_k_cluster_id) const
+    {
+        if(last_block_reduce_length_ == K_BlockTileSize * numBlockTileIteration_ ||
+           block_local_id < blkGroupSize_ - 1)
+            return (KThreadSliceSize * numBlockTileIteration_);
+
+        index_t count_in_last_tile = last_block_reduce_length_ % K_BlockTileSize;
+
+        if(count_in_last_tile == 0)
+            return (KThreadSliceSize * numBlockTileIterationByLastBlock_);
+        else
+        {
+            index_t num_complete_slice = count_in_last_tile / KThreadSliceSize;
+
+            if(thread_k_cluster_id < num_complete_slice)
+                return (KThreadSliceSize * numBlockTileIterationByLastBlock_);
+            else if(thread_k_cluster_id == num_complete_slice)
+                return (KThreadSliceSize * (numBlockTileIterationByLastBlock_ - 1) +
+                        count_in_last_tile);
+            else
+                return (KThreadSliceSize * (numBlockTileIterationByLastBlock_ - 1));
+        };
+    };
+
+    index_t blkGroupSize_;
+    index_t numBlockTileIteration_;
+
+    index_t last_block_reduce_length_;
+    index_t numBlockTileIterationByLastBlock_;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
new file mode 100644
index 0000000..a4053b1
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
@@ -0,0 +1,286 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace element_wise {
+
+struct Add
+{
+    template <typename Y, typename X0, typename X1>
+    __host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float>(float& y, const float& x0, const float& x1) const
+    {
+        y = x0 + x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double>(double& y, const double& x0, const double& x1) const
+    {
+        y = x0 + x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float>(float& y, const float& x0, const half_t& x1) const
+    {
+        y = x0 + type_convert<half_t>(x1);
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const float& x0, const half_t& x1) const
+    {
+        y = type_convert<half_t>(x0) + x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        y = x0 + x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<bhalf_t>(bhalf_t& y, const bhalf_t& x0, const bhalf_t& x1) const
+    {
+        const float x1_tmp = ck::type_convert<float>(x0);
+        const float x2_tmp = ck::type_convert<float>(x1);
+        const float y_tmp  = x1_tmp + x2_tmp;
+        y                  = ck::type_convert<bhalf_t>(y_tmp);
+    }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<int8_t>(int8_t& y, const int8_t& x0, const int8_t& x1) const
+    {
+        y = x0 + x1;
+    };
+};
+
+struct Subtract
+{
+    template <typename T>
+    __host__ __device__ constexpr void operator()(T& y, const T& x0, const T& x1) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float>(float& y, const float& x0, const float& x1) const
+    {
+        y = x0 - x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double>(double& y, const double& x0, const double& x1) const
+    {
+        y = x0 - x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        y = x0 - x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<bhalf_t>(bhalf_t& y, const bhalf_t& x0, const bhalf_t& x1) const
+    {
+        const float x1_tmp = ck::type_convert<float>(x0);
+        const float x2_tmp = ck::type_convert<float>(x1);
+        const float y_tmp  = x1_tmp - x2_tmp;
+        y                  = ck::type_convert<bhalf_t>(y_tmp);
+    }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<int8_t>(int8_t& y, const int8_t& x0, const int8_t& x1) const
+    {
+        y = x0 - x1;
+    };
+};
+
+struct Bilinear
+{
+    Bilinear(float alpha, float beta) : alpha_(alpha), beta_(beta){};
+
+    template <typename Y, typename X0, typename X1>
+    __host__ __device__ constexpr void operator()(Y&, const X0&, const X1&) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float, float, float>(float& y, const float& x0, const float& x1) const
+    {
+        y = alpha_ * x0 + beta_ * x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, half_t, half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        y = type_convert<half_t>(alpha_) * x0 + type_convert<half_t>(beta_) * x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, float, half_t>(half_t& y, const float& x0, const half_t& x1) const
+    {
+        y = type_convert<half_t>(alpha_ * x0 + beta_ * ck::type_convert<float>(x1));
+    };
+
+    float alpha_;
+    float beta_;
+};
+
+struct AddRelu
+{
+    template <typename Y, typename X0, typename X1>
+    __host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float, float, float>(float& y, const float& x0, const float& x1) const
+    {
+        const float a = x0 + x1;
+        y             = a > 0.0f ? a : 0.0f;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double, double, double>(double& y, const double& x0, const double& x1) const
+    {
+        const double a = x0 + x1;
+        y              = a > 0.0 ? a : 0.0;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, half_t, half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        const half_t a = x0 + x1;
+        y              = a > type_convert<half_t>(0.0f) ? a : type_convert<half_t>(0.0f);
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, float, half_t>(half_t& y, const float& x0, const half_t& x1) const
+    {
+        const float a = x0 + x1;
+        y             = a > type_convert<half_t>(0.0f) ? a : type_convert<half_t>(0.0f);
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float, float, half_t>(float& y, const float& x0, const half_t& x1) const
+    {
+        const float a = x0 + type_convert<float>(x1);
+        y             = a > 0.0f ? a : 0.0f;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<int, int, int8_t>(int& y, const int& x0, const int8_t& x1) const
+    {
+        const int8_t a = x0 + x1;
+        y              = a > 0 ? a : 0;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<int8_t, int8_t, int8_t>(int8_t& y, const int8_t& x0, const int8_t& x1) const
+    {
+        const int8_t a = x0 + x1;
+        y              = a > 0 ? a : 0;
+    };
+};
+
+struct AddHardswish
+{
+    template <typename T>
+    __host__ __device__ constexpr void operator()(T& y, const T& x0, const T& x1) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float>(float& y, const float& x0, const float& x1) const
+    {
+        float a = x0 + x1;
+        float b = a + float{3};
+        float c = (b > 0) * (b > 6.0f ? 6.0f : b) * a * 0.166667f;
+        y       = c;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double>(double& y, const double& x0, const double& x1) const
+    {
+        double a = x0 + x1;
+        double b = a + 3.0;
+        double c = (b > 0) * (b > 6.0 ? 6.0 : b) * a * 0.166667;
+        y        = c;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        float a = x0 + x1;
+        float b = a + 3.0f;
+        float c = (b > 0) * (b > 6.0f ? 6.0f : b) * a * 0.166667f;
+        y       = c;
+    };
+};
+
+// C = A * B
+// E = FastGelu(C + D)
+struct AddFastGelu
+{
+    // Fast GeLU
+    // https://paperswithcode.com/method/gelu
+    // y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
+    __host__ __device__ static constexpr float GetFastGeLU(float x)
+    {
+        const float u   = 2.f * x * (0.035677f * x * x + 0.797885f);
+        const float emu = exp(-u);
+        const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
+        return x * cdf;
+    }
+
+    template <typename T>
+    static inline constexpr bool is_valid_param_type_v =
+        std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
+        std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>;
+
+    template <typename E, typename C, typename D>
+    __host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const
+    {
+        static_assert(is_valid_param_type_v<E> && is_valid_param_type_v<C> &&
+                      is_valid_param_type_v<D>);
+
+        const float y = GetFastGeLU(type_convert<float>(c) + type_convert<float>(d));
+
+        e = type_convert<E>(y);
+    }
+
+    template <typename D>
+    __host__ __device__ constexpr void operator()(float& e, const float& c, const D& d) const
+    {
+        static_assert(is_valid_param_type_v<D>);
+
+        e = GetFastGeLU(c + type_convert<float>(d));
+    }
+};
+
+} // namespace element_wise
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
new file mode 100644
index 0000000..b66107a
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
@@ -0,0 +1,304 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/quantization_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace element_wise {
+
+// Need to ensure compiler will fail if there is no matching candidate, instead of compiler
+// siliently do implicit type conversion
+//
+// Method 1:
+//
+// struct ExampleElementwiseOp
+// {
+//     template<typename Y, typename X>
+//     __host__ __device__ constexpr void
+//     operator()(Y&, const X) const;
+//
+//     template<>
+//     __host__ __device__ constexpr void
+//     operator()<half_t, half_t>(half_t& y, const half_t& x) const
+//     {
+//     }
+// };
+//
+// Method 2:
+//
+// template <typename Y, typename X>
+// struct ExampleElementwiseOp;
+//
+// template <>
+// struct ExampleElementwiseOp<float, ck::bhalf_t>
+// {
+//     __host__ __device__ void operator()(float& y, ck::bhalf_t& x) const
+//     {
+//     }
+// };
+
+struct AddReluAdd
+{
+    template <typename Y, typename X0, typename X1, typename X2>
+    __host__ __device__ constexpr void operator()(Y&, const X0&, const X1&, const X2&) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
+        half_t& y, const half_t& x0, const half_t& x1, const half_t& x2) const
+    {
+        half_t a = x0 + x1;
+        half_t b = a > 0 ? a : 0;
+        y        = b + x2;
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float>(float& y,
+                                                                              const float& x0,
+                                                                              const float& x1,
+                                                                              const float& x2) const
+    {
+        float a = x0 + x1;
+        float b = a > 0 ? a : 0;
+        float c = b + x2;
+        y       = c;
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, float, half_t, half_t>(
+        half_t& y, const float& x0, const half_t& x1, const half_t& x2) const
+    {
+        float a = x0 + x1;
+        float b = a > 0 ? a : 0;
+        float c = b + x2;
+        y       = c;
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<bhalf_t, float, bhalf_t, bhalf_t>(
+        bhalf_t& y, const float& x0, const bhalf_t& x1, const bhalf_t& x2) const
+    {
+        float a = x0 + x1;
+        float b = a > 0 ? a : 0;
+        float c = b + x2;
+        y       = c;
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<int8_t, int8_t, int8_t, int8_t>(
+        int8_t& y, const int8_t& x0, const int8_t& x1, const int8_t& x2) const
+    {
+        int32_t a = x0 + x1;
+        int32_t b = a > 0 ? a : 0;
+        int32_t c = b + x2;
+        y         = c;
+    }
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    template <>
+    __host__ __device__ constexpr void operator()<int4_t, int8_t, int4_t, int4_t>(
+        int4_t& y, const int8_t& x0, const int4_t& x1, const int4_t& x2) const
+    {
+        int32_t a = x0 + x1;
+        int32_t b = a > 0 ? a : 0;
+        int32_t c = b + x2;
+        y         = c;
+    }
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+};
+
+struct AddHardswishAdd
+{
+    template <typename Y, typename X0, typename X1, typename X2>
+    __host__ __device__ constexpr void operator()(Y&, const X0&, const X1&, const X2&) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float>(float& y,
+                                                                              const float& x0,
+                                                                              const float& x1,
+                                                                              const float& x2) const
+    {
+        float a = x0 + x1;
+        float b = a + float{3};
+        float c = (b > 0) * (b > float{6} ? float{6} : b) * a * float{0.166667};
+        float d = c + x2;
+        y       = d;
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
+        half_t& y, const half_t& x0, const half_t& x1, const half_t& x2) const
+    {
+        float a = x0 + x1;
+        float b = a + float{3};
+        float c = (b > 0) * (b > float{6} ? float{6} : b) * a * float{0.166667};
+        float d = c + x2;
+        y       = d;
+    }
+};
+
+// C = A * B
+// E = C + D0 + D1
+struct AddAdd
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ void operator()(E& e, const C& c, const D0& d0, const D1& d1) const
+    {
+        // Only support floating so far
+        static_assert(is_same<E, half_t>::value || is_same<E, float>::value ||
+                          is_same<E, double>::value,
+                      "Data type is not supported by this operation!");
+
+        static_assert(is_same<C, half_t>::value || is_same<C, float>::value ||
+                          is_same<C, double>::value,
+                      "Data type is not supported by this operation!");
+
+        static_assert(is_same<D0, half_t>::value || is_same<D0, float>::value ||
+                          is_same<D0, double>::value,
+                      "Data type is not supported by this operation!");
+
+        static_assert(is_same<D1, half_t>::value || is_same<D1, float>::value ||
+                          is_same<D1, double>::value,
+                      "Data type is not supported by this operation!");
+
+        const C y = c + type_convert<C>(d0) + type_convert<C>(d1);
+        e         = type_convert<E>(y);
+    }
+};
+
+// C = A * B
+// E = FastGelu(C + D0 + D1)
+struct AddAddFastGelu
+{
+    // Fast GeLU
+    // https://paperswithcode.com/method/gelu
+    // y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
+    __host__ __device__ static constexpr float GetFastGeLU(float x)
+    {
+        const float u   = 2.f * x * (0.035677f * x * x + 0.797885f);
+        const float emu = exp(-u);
+        const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
+        return x * cdf;
+    }
+
+    template <typename T>
+    static inline constexpr bool is_valid_param_type_v =
+        std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
+        std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        || std::is_same_v<T, ck::int4_t>
+#endif
+        ;
+
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1) const
+    {
+        static_assert(is_valid_param_type_v<E> && is_valid_param_type_v<C> &&
+                      is_valid_param_type_v<D0> && is_valid_param_type_v<D1>);
+
+        const float y =
+            GetFastGeLU(type_convert<float>(c) + type_convert<float>(d0) + type_convert<float>(d1));
+
+        e = type_convert<E>(y);
+    }
+};
+
+struct Normalize
+{
+    // FIXME: is double absolutely necessary?
+    Normalize(double epsilon = 1e-4) : epsilon_(epsilon) {}
+
+    template <typename T1, typename T2, typename T3>
+    __host__ __device__ constexpr void operator()(T1& y,
+                                                  const T1& x,
+                                                  const T2& mean,
+                                                  const T2& mean_square,
+                                                  const T3& gamma,
+                                                  const T3& beta) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, float, half_t>(half_t& y,
+                                                                         const half_t& x,
+                                                                         const float& mean,
+                                                                         const float& mean_square,
+                                                                         const half_t& gamma,
+                                                                         const half_t& beta) const
+    {
+        using ck::math::sqrt;
+
+        float variance = mean_square - (mean * mean);
+
+        float tmp_x     = type_convert<float>(x);
+        float tmp_gamma = type_convert<float>(gamma);
+        float tmp_beta  = type_convert<float>(beta);
+
+        float tmp_y =
+            ((tmp_x - mean) / sqrt(variance + type_convert<float>(epsilon_))) * tmp_gamma +
+            tmp_beta;
+
+        y = type_convert<half_t>(tmp_y);
+    };
+
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float>(float& y,
+                                                                       const float& x,
+                                                                       const float& mean,
+                                                                       const float& mean_square,
+                                                                       const float& gamma,
+                                                                       const float& beta) const
+    {
+        using ck::math::sqrt;
+
+        float variance = mean_square - (mean * mean);
+        y = ((x - mean) / sqrt(variance + type_convert<float>(epsilon_))) * gamma + beta;
+    };
+
+    template <>
+    __host__ __device__ constexpr void operator()<double, double, double>(double& y,
+                                                                          const double& x,
+                                                                          const double& mean,
+                                                                          const double& mean_square,
+                                                                          const double& gamma,
+                                                                          const double& beta) const
+    {
+        using ck::math::sqrt;
+
+        double variance = mean_square - (mean * mean);
+        y               = ((x - mean) / sqrt(variance + epsilon_)) * gamma + beta;
+    };
+
+    // FIXME: is double absolutely necessary?
+    double epsilon_;
+};
+
+template <typename Y, typename X>
+struct UnaryTypeConvert;
+
+template <>
+struct UnaryTypeConvert<float, ck::bhalf_t>
+{
+    __host__ __device__ void operator()(float& y, ck::bhalf_t& x) const
+    {
+        y = ck::type_convert<float, ck::bhalf_t>(x);
+    }
+};
+
+template <>
+struct UnaryTypeConvert<ck::bhalf_t, float>
+{
+    __host__ __device__ void operator()(ck::bhalf_t& y, float& x) const
+    {
+        y = ck::type_convert<ck::bhalf_t, float>(x);
+    }
+};
+
+} // namespace element_wise
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/quantization_operation.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/quantization_operation.hpp
new file mode 100644
index 0000000..3f2c2f8
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/quantization_operation.hpp
@@ -0,0 +1,124 @@
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace element_wise {
+
+// For Activation function which is piecewise linear function, such as relu, leaky relu ...etc
+template <typename Activation>
+struct Activation_Mul_Clamp
+{
+    Activation_Mul_Clamp(float requantScale, Activation activationOp)
+        : requantScale_(requantScale), activationOp_(activationOp)
+    {
+    }
+
+    __host__ __device__ constexpr void operator()(int8_t& y, const int32_t& x) const
+    {
+        float x_fp32 = ck::type_convert<float>(x);
+        activationOp_(x_fp32, x_fp32);
+        float y_fp32 = math::clamp(requantScale_ * x_fp32, -128.f, 127.f);
+        y            = ck::type_convert<int8_t>(y_fp32);
+    }
+
+    __host__ __device__ constexpr void operator()(float& y, const int32_t& x) const
+    {
+        // We might type_convert to int8 after lambda in someplace
+        float x_fp32 = ck::type_convert<float>(x);
+        activationOp_(x_fp32, x_fp32);
+        y = math::clamp(requantScale_ * x_fp32, -128.f, 127.f);
+    }
+
+    float requantScale_;
+    Activation activationOp_;
+};
+
+// Conv Perchannel quantization + Activation function which is piecewise linear function, such as
+// relu, leaky relu ...etc
+template <typename Activation>
+struct Activation_Mul2_Clamp
+{
+    Activation_Mul2_Clamp(Activation activationOp) : activationOp_(activationOp) {}
+
+    __host__ __device__ constexpr void
+    operator()(int8_t& y, const int32_t& x, const float& requantScale) const
+    {
+        float y_fp32 = ck::type_convert<float>(x);
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(requantScale * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int8_t>(y_fp32);
+    }
+
+    Activation activationOp_;
+};
+
+// For Activation function which is piecewise linear function, such as relu, leaky relu ...etc
+template <typename Activation>
+struct Add_Activation_Mul_Clamp
+{
+    Add_Activation_Mul_Clamp(float requantScale, Activation activationOp)
+        : requantScale_(requantScale), activationOp_(activationOp)
+    {
+    }
+
+    __host__ __device__ constexpr void
+    operator()(int8_t& y, const int32_t& x, const int32_t& bias) const
+    {
+        float y_fp32 = ck::type_convert<float>(x + bias);
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(requantScale_ * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int8_t>(y_fp32);
+    }
+
+    float requantScale_;
+    Activation activationOp_;
+};
+
+// Conv Perchannel quantization + Activation function which is piecewise linear function, such as
+// relu, leaky relu ...etc
+template <typename Activation>
+struct Add_Activation_Mul2_Clamp
+{
+    Add_Activation_Mul2_Clamp(Activation activationOp) : activationOp_(activationOp) {}
+
+    __host__ __device__ constexpr void
+    operator()(int8_t& y, const int32_t& x, const int32_t& bias, const float& requantScale) const
+    {
+        float y_fp32 = ck::type_convert<float>(x + bias);
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(requantScale * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int8_t>(y_fp32);
+    }
+
+    Activation activationOp_;
+};
+
+// For Activation function which is non piecewise linear function, such as TanH, Sigmoid ...etc
+template <typename Activation>
+struct Add_Mul_Activation_Mul_Clamp
+{
+    Add_Mul_Activation_Mul_Clamp(float requantScale1, float requantScale2, Activation activationOp)
+        : requantScale1_(requantScale1), requantScale2_(requantScale2), activationOp_(activationOp)
+    {
+    }
+
+    __host__ __device__ constexpr void
+    operator()(int8_t& y, const int32_t& x, const int32_t& bias) const
+    {
+        float y_fp32 = ck::type_convert<float>(x + bias);
+        y_fp32       = requantScale1_ * y_fp32;
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(requantScale2_ * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int8_t>(y_fp32);
+    }
+
+    float requantScale1_;
+    float requantScale2_;
+    Activation activationOp_;
+};
+
+} // namespace element_wise
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
new file mode 100644
index 0000000..fbdfe92
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
@@ -0,0 +1,268 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/utility/math_v2.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace element_wise {
+
+struct PassThrough
+{
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ __device__ void operator()<double, double>(double& y, const double& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<float, float>(float& y, const float& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<half_t, half_t>(half_t& y, const half_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<bhalf_t, bhalf_t>(bhalf_t& y, const bhalf_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<int32_t, int32_t>(int32_t& y, const int32_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<bhalf_t, float>(bhalf_t& y, const float& x) const
+    {
+        y = type_convert<bhalf_t>(x);
+    }
+
+    template <>
+    __host__ __device__ void operator()<int8_t, int8_t>(int8_t& y, const int8_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<int8_t, int32_t>(int8_t& y, const int32_t& x) const
+    {
+        y = type_convert<int8_t>(x);
+    }
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    template <>
+    __host__ __device__ void operator()<int4_t, int4_t>(int4_t& y, const int4_t& x) const
+    {
+        y = x;
+    }
+#endif
+};
+
+struct UnaryConvert
+{
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
+    {
+        y = type_convert<Y>(x);
+    }
+};
+
+struct Scale
+{
+    __host__ __device__ Scale(float scale) : scale_(scale) {}
+
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ __device__ void operator()<float, float>(float& y, const float& x) const
+    {
+        y = scale_ * x;
+    };
+
+    float scale_;
+};
+
+struct ScaleAndResetNaNToMinusInfinity
+{
+    __host__ __device__ ScaleAndResetNaNToMinusInfinity(float scale) : scale_(scale) {}
+
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ __device__ void operator()<float, float>(float& y, const float& x) const
+    {
+        y = ck::math::isnan(x) ? -ck::NumericLimits<float>::Infinity() : scale_ * x;
+    };
+
+    float scale_;
+};
+
+struct UnaryDivide
+{
+    __host__ __device__ UnaryDivide(const int32_t divider = 1) : divider_(divider) {}
+
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, int32_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = x / type_convert<T>(divider_);
+    };
+
+    int32_t divider_ = 1;
+};
+
+struct UnarySquare
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same_v<T, float> || is_same_v<T, double> || is_same_v<T, int32_t> ||
+                          is_same_v<T, int8_t>
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+                          || is_same_v<T, int4_t>
+#endif
+                      ,
+                      "Data type is not supported by this operation!");
+        y = x * x;
+    };
+};
+
+struct UnaryAbs
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = ck::math::abs(x);
+    };
+};
+
+struct UnarySqrt
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value,
+                      "Data type is not supported by this operation!");
+
+        y = ck::math::sqrt(x);
+    };
+};
+
+struct Relu
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "Data type is not supported by this operation!");
+        y = x > 0 ? x : 0;
+    }
+
+    template <>
+    __host__ __device__ void operator()(bhalf_t& y, const bhalf_t& x) const
+    {
+        float x_f32 = ck::type_convert<float>(x);
+        float y_f32 = x_f32 > 0 ? x_f32 : 0;
+        y           = ck::type_convert<bhalf_t>(y_f32);
+    }
+};
+
+// Y = FastGelu(X)
+struct FastGelu
+{
+    // Fast GeLU
+    // https://paperswithcode.com/method/gelu
+    // y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
+    __host__ __device__ static constexpr float GetFastGeLU(float x)
+    {
+        const float u   = 2.f * x * (0.035677f * x * x + 0.797885f);
+        const float emu = exp(-u);
+        const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
+        return x * cdf;
+    }
+
+    template <typename T>
+    static inline constexpr bool is_valid_param_type_v =
+        std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
+        std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        || std::is_same_v<T, ck::int4_t>
+#endif
+        ;
+
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
+    {
+        static_assert(is_valid_param_type_v<Y> && is_valid_param_type_v<X>);
+
+        const float tmp_y = GetFastGeLU(type_convert<float>(x));
+        y                 = type_convert<Y>(tmp_y);
+    }
+};
+
+// https://paperswithcode.com/method/gelu
+// y = 0.5*x*(1+erf(x/sqrt(2)))
+struct Gelu
+{
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ __device__ void operator()<float, float>(float& y, const float& x) const
+    {
+        y = 0.5f * x * (1.f + erf(float(0.70710678118f * x)));
+    }
+
+    template <>
+    __host__ __device__ void operator()<ck::half_t, ck::half_t>(ck::half_t& y,
+                                                                const ck::half_t& x) const
+    {
+        y = ck::half_t(0.5) * x * (ck::half_t(1) + ck::half_t(erf(float(0.70710678118f * x))));
+    }
+};
+
+struct Sigmoid
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, ck::half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = 1 / (ck::type_convert<T>(1) + exp(-x));
+    };
+
+    int32_t divider_ = 1;
+};
+
+} // namespace element_wise
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp
new file mode 100644
index 0000000..a72a4ee
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp
@@ -0,0 +1,498 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseReduceSecondHalfBatchNormBackwardFinal_,
+          typename XDataType,
+          typename DyDataType,
+          typename DxDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename DscaleDbiasGridDesc_M_K,
+          typename MeanVarGridDesc_M,
+          typename ScaleBiasGridDesc_M>
+__global__ void kernel_reduce_second_half_batchnorm_backward_final(
+    const XYGridDesc_M_K x_grid_desc_m_k,
+    const XYGridDesc_M_K dy_grid_desc_m_k,
+    const XYGridDesc_M_K dx_grid_desc_m_k,
+    const DscaleDbiasGridDesc_M_K dscale_dbias_grid_desc_m_k,
+    const MeanVarGridDesc_M mean_var_grid_desc_m,
+    const ScaleBiasGridDesc_M scale_grid_desc_m,
+    const ScaleBiasGridDesc_M bias_grid_desc_m,
+    index_t blkgroup_size,
+    long_index_t reduce_size,
+    index_t num_xy_k_block_tile_iteration,
+    index_t num_dscale_dbias_k_block_tile_iteration,
+    const DscaleDbiasDataType* const __restrict__ p_reduce_dscale,
+    const DscaleDbiasDataType* const __restrict__ p_reduce_dbias,
+    const MeanVarDataType* const __restrict__ p_mean,
+    const MeanVarDataType* const __restrict__ p_inv_var,
+    const XDataType* const __restrict__ p_x,
+    const DyDataType* const __restrict__ p_dy,
+    const ScaleDataType* const __restrict__ p_scale,
+    const DyElementwiseOp dy_elementwise_op,
+    DxDataType* const __restrict__ p_dx,
+    DscaleDbiasDataType* const __restrict__ p_dscale,
+    DscaleDbiasDataType* const __restrict__ p_dbias)
+{
+    GridwiseReduceSecondHalfBatchNormBackwardFinal_::Run(x_grid_desc_m_k,
+                                                         dy_grid_desc_m_k,
+                                                         dx_grid_desc_m_k,
+                                                         dscale_dbias_grid_desc_m_k,
+                                                         mean_var_grid_desc_m,
+                                                         scale_grid_desc_m,
+                                                         bias_grid_desc_m,
+                                                         blkgroup_size,
+                                                         reduce_size,
+                                                         num_xy_k_block_tile_iteration,
+                                                         num_dscale_dbias_k_block_tile_iteration,
+                                                         p_reduce_dscale,
+                                                         p_reduce_dbias,
+                                                         p_mean,
+                                                         p_inv_var,
+                                                         p_x,
+                                                         p_dy,
+                                                         p_scale,
+                                                         dy_elementwise_op,
+                                                         p_dx,
+                                                         p_dscale,
+                                                         p_dbias);
+};
+
+template <typename XDataType,
+          typename DyDataType,
+          typename DxDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename DscaleDbiasGridDesc_M_K,
+          typename MeanVarGridDesc_M,
+          typename ScaleBiasGridDesc_M,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XDyDxVectorDim,
+          index_t XSrcVectorSize,
+          index_t DySrcVectorSize,
+          index_t DxDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t DscaleDbiasDstVectorSize,
+          index_t MeanVarSrcVectorSize>
+struct GridwiseReduceSecondHalfBatchNormBackwardFinal
+{
+    static_assert((XDyDxVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0 &&
+                   MThreadSliceSize % DySrcVectorSize == 0 &&
+                   MThreadSliceSize % DxDstVectorSize == 0) ||
+                      (XDyDxVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0 &&
+                       KThreadSliceSize % DySrcVectorSize == 0 &&
+                       KThreadSliceSize % DxDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XDyDxVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_1 = decltype(
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using BlockwiseReduce = PartitionedBlockwiseReduction<AccDataType,
+                                                          BlockSize,
+                                                          ThreadClusterLengths_M_K,
+                                                          ThreadClusterArrangeOrder,
+                                                          ck::reduce::Add,
+                                                          false>;
+
+    using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                 ThreadReduceSrcDesc_M_1,
+                                                 ThreadReduceDstDesc_M,
+                                                 ck::reduce::Add,
+                                                 false>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    // clang-format off
+    // Two of the steps of Multiblock BatchNorm Backward
+    // Step 1: Second half of Reduction: dbias = sum(dy), dscale = sum(dy * (x-mean) * inv-variance)
+    // Step 2: calculating dx = 1/reduce_size * inv-variance * scale * (reduce_size * dy - dbias - dscale * (x - mean) * inv-variance)) elementwise-ly
+    // clang-format on
+    __device__ static void Run(const XYGridDesc_M_K& x_grid_desc_m_k,
+                               const XYGridDesc_M_K& dy_grid_desc_m_k,
+                               const XYGridDesc_M_K& dx_grid_desc_m_k,
+                               const DscaleDbiasGridDesc_M_K& dscale_dbias_grid_desc_m_k,
+                               const MeanVarGridDesc_M& mean_var_grid_desc_m,
+                               const ScaleBiasGridDesc_M& scale_grid_desc_m,
+                               const ScaleBiasGridDesc_M& dscale_dbias_grid_desc_m,
+                               index_t blkgroup_size,
+                               long_index_t reduce_size,
+                               index_t num_xy_k_block_tile_iteration,
+                               index_t num_dscale_dbias_k_block_tile_iteration,
+                               const DscaleDbiasDataType* const __restrict__ p_reduce_dscale,
+                               const DscaleDbiasDataType* const __restrict__ p_reduce_dbias,
+                               const MeanVarDataType* const __restrict__ p_mean,
+                               const MeanVarDataType* const __restrict__ p_inv_var,
+                               const XDataType* const __restrict__ p_x,
+                               const DyDataType* const __restrict__ p_dy,
+                               const ScaleDataType* const __restrict__ p_scale,
+                               const DyElementwiseOp dy_elementwise_op,
+                               DxDataType* const __restrict__ p_dx,
+                               DscaleDbiasDataType* const __restrict__ p_dscale,
+                               DscaleDbiasDataType* const __restrict__ p_dbias)
+    {
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * 1, true>
+            reduce_dscale_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * 1, true>
+            reduce_dbias_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> dscale_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> dbias_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            dy_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            dx_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            inv_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> scale_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / blkgroup_size;
+        const index_t block_local_id  = block_global_id % blkgroup_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M           = Sequence<MThreadSliceSize>;
+        using ThreadBufferLengths_M_1         = Sequence<MThreadSliceSize, 1>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m_1 = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<1>{}));
+
+        // clang-format off
+        // Step 1: do final reduction of dbias = sum(dy), dscale = sum(dy * (x-mean) * inv-variance)
+        // clang-format on
+
+        auto threadwise_dscale_dbias_load_m_k =
+            ThreadwiseTensorSliceTransfer_v2<DscaleDbiasDataType,
+                                             AccDataType,
+                                             DscaleDbiasGridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             Sequence<0, 1>,
+                                             1,
+                                             1,
+                                             1,
+                                             true>(
+                dscale_dbias_grid_desc_m_k,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * 1));
+
+        auto threadwise_dscale_dbias_store_m =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               DscaleDbiasDataType,
+                                               decltype(thread_buffer_desc_m),
+                                               ScaleBiasGridDesc_M,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M,
+                                               Sequence<0>,
+                                               0,
+                                               DscaleDbiasDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                dscale_dbias_grid_desc_m,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize),
+                PassThroughOp{});
+
+        const auto reduce_dscale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_reduce_dscale, dscale_dbias_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto reduce_dbias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_reduce_dbias, dscale_dbias_grid_desc_m_k.GetElementSpaceSize());
+
+        auto dscale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dscale, dscale_dbias_grid_desc_m.GetElementSpaceSize());
+
+        auto dbias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dbias, dscale_dbias_grid_desc_m.GetElementSpaceSize());
+
+        constexpr auto dscale_dbias_thread_copy_step_m_k =
+            make_multi_index(0, KThreadClusterSize * 1);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            dscale_thread_buf(I) = type_convert<AccDataType>(0.0f);
+            dbias_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_dscale_dbias_k_block_tile_iteration;
+            ++reducedTiles)
+        {
+            threadwise_dscale_dbias_load_m_k.Run(dscale_dbias_grid_desc_m_k,
+                                                 reduce_dscale_global_buf,
+                                                 thread_buffer_desc_m_1,
+                                                 make_tuple(I0, I0),
+                                                 reduce_dscale_thread_buf);
+
+            threadwise_dscale_dbias_load_m_k.Run(dscale_dbias_grid_desc_m_k,
+                                                 reduce_dbias_global_buf,
+                                                 thread_buffer_desc_m_1,
+                                                 make_tuple(I0, I0),
+                                                 reduce_dbias_thread_buf);
+
+            ThreadwiseReduce::Reduce(reduce_dscale_thread_buf, dscale_thread_buf);
+            ThreadwiseReduce::Reduce(reduce_dbias_thread_buf, dbias_thread_buf);
+
+            threadwise_dscale_dbias_load_m_k.MoveSrcSliceWindow(dscale_dbias_grid_desc_m_k,
+                                                                dscale_dbias_thread_copy_step_m_k);
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            BlockwiseReduce::Reduce(reduce_work_buf, dscale_thread_buf(I));
+            block_sync_lds();
+            BlockwiseReduce::Reduce(reduce_work_buf, dbias_thread_buf(I));
+        });
+
+        threadwise_dscale_dbias_store_m.Run(thread_buffer_desc_m,
+                                            make_tuple(I0),
+                                            dscale_thread_buf,
+                                            dscale_dbias_grid_desc_m,
+                                            dscale_global_buf);
+
+        threadwise_dscale_dbias_store_m.Run(thread_buffer_desc_m,
+                                            make_tuple(I0),
+                                            dbias_thread_buf,
+                                            dscale_dbias_grid_desc_m,
+                                            dbias_global_buf);
+
+        // clang-format off
+        // Step 2: calculate dx = 1/N * inv-variance * scale * (N * dy - dbias - dscale * (x - mean) * inv-variance)
+        // clang-format on
+
+        const index_t workSizePerBlock = K_BlockTileSize * num_xy_k_block_tile_iteration;
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XYGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XDyDxVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             workSizePerBlock * block_local_id +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dy_load = ThreadwiseTensorSliceTransfer_v2<DyDataType,
+                                                                   AccDataType,
+                                                                   XYGridDesc_M_K,
+                                                                   decltype(thread_buffer_desc_m_k),
+                                                                   ThreadBufferLengths_M_K,
+                                                                   ThreadBufferDimAccessOrder,
+                                                                   XDyDxVectorDim,
+                                                                   DySrcVectorSize,
+                                                                   1,
+                                                                   true>(
+            dy_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             workSizePerBlock * block_local_id +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dx_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               DxDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               XYGridDesc_M_K,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               XDyDxVectorDim,
+                                               DxDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                dx_grid_desc_m_k,
+                make_multi_index(
+                    blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                    workSizePerBlock * block_local_id + thread_k_cluster_id * KThreadSliceSize),
+                PassThroughOp{});
+
+        auto threadwise_scale_load =
+            ThreadwiseTensorSliceTransfer_v2<ScaleDataType,
+                                             AccDataType,
+                                             ScaleBiasGridDesc_M,
+                                             decltype(thread_buffer_desc_m),
+                                             ThreadBufferLengths_M,
+                                             Sequence<0>,
+                                             0,
+                                             ScaleSrcVectorSize,
+                                             1,
+                                             true>(
+                scale_grid_desc_m,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize));
+
+        auto threadwise_mean_var_load =
+            ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                             AccDataType,
+                                             MeanVarGridDesc_M,
+                                             decltype(thread_buffer_desc_m),
+                                             ThreadBufferLengths_M,
+                                             Sequence<0>,
+                                             0,
+                                             MeanVarSrcVectorSize,
+                                             1,
+                                             true>(
+                mean_var_grid_desc_m,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize));
+
+        const auto x_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto dy_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dy, dy_grid_desc_m_k.GetElementSpaceSize());
+
+        auto dx_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dx, dx_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto scale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_scale, scale_grid_desc_m.GetElementSpaceSize());
+
+        const auto mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_mean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+        const auto inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_inv_var, mean_var_grid_desc_m.GetElementSpaceSize());
+
+        threadwise_scale_load.Run(scale_grid_desc_m,
+                                  scale_global_buf,
+                                  thread_buffer_desc_m,
+                                  make_tuple(I0),
+                                  scale_thread_buf);
+
+        threadwise_mean_var_load.Run(mean_var_grid_desc_m,
+                                     mean_global_buf,
+                                     thread_buffer_desc_m,
+                                     make_tuple(I0),
+                                     mean_thread_buf);
+
+        threadwise_mean_var_load.Run(mean_var_grid_desc_m,
+                                     inv_var_global_buf,
+                                     thread_buffer_desc_m,
+                                     make_tuple(I0),
+                                     inv_var_thread_buf);
+
+        constexpr auto xy_thread_copy_step_m_k = make_multi_index(0, K_BlockTileSize);
+
+        AccDataType inv_reduce_size =
+            type_convert<AccDataType>(1.0) / type_convert<AccDataType>(reduce_size);
+
+        for(index_t reducedTiles = 0; reducedTiles < num_xy_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_dy_load.Run(dy_grid_desc_m_k,
+                                   dy_global_buf,
+                                   thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   dy_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                AccDataType multiplier =
+                    inv_reduce_size * inv_var_thread_buf[iM] * scale_thread_buf[iM];
+
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    dy_elementwise_op(dy_thread_buf(Number<offset>{}),
+                                      dy_thread_buf[Number<offset>{}]);
+
+                    AccDataType norm_x = (x_thread_buf[Number<offset>{}] - mean_thread_buf[iM]) *
+                                         inv_var_thread_buf[iM];
+
+                    AccDataType tmpVal = norm_x * dscale_thread_buf[iM];
+
+                    dx_thread_buf(Number<offset>{}) =
+                        multiplier *
+                        (type_convert<AccDataType>(reduce_size) * dy_thread_buf[Number<offset>{}] -
+                         dbias_thread_buf[iM] - tmpVal);
+                });
+            });
+
+            threadwise_dx_store.Run(thread_buffer_desc_m_k,
+                                    make_tuple(I0, I0),
+                                    dx_thread_buf,
+                                    dx_grid_desc_m_k,
+                                    dx_global_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, xy_thread_copy_step_m_k);
+            threadwise_dy_load.MoveSrcSliceWindow(dy_grid_desc_m_k, xy_thread_copy_step_m_k);
+            threadwise_dx_store.MoveDstSliceWindow(dx_grid_desc_m_k, xy_thread_copy_step_m_k);
+        }
+    };
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp
new file mode 100644
index 0000000..08cb0dd
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp
@@ -0,0 +1,261 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseMultiblockWelfordFirstHalf_,
+          typename XDataType,
+          typename MeanVarDataType,
+          typename XGridDesc_M_K,
+          typename MeanVarCountGridDesc_M_G,
+          typename GetReduceCountPerThreadFunctor>
+__global__ void kernel_multiblock_welford_first_half(
+    const XGridDesc_M_K x_grid_desc_m_k,
+    const MeanVarCountGridDesc_M_G mean_var_count_grid_desc_m_g,
+    const GetReduceCountPerThreadFunctor get_reduce_count_per_thread,
+    index_t num_k_block_tile_iteration,
+    const XDataType* const __restrict__ p_x,
+    MeanVarDataType* const p_welford_mean,
+    MeanVarDataType* const p_welford_variance,
+    int32_t* const p_welford_count)
+{
+    GridwiseMultiblockWelfordFirstHalf_::Run(x_grid_desc_m_k,
+                                             mean_var_count_grid_desc_m_g,
+                                             get_reduce_count_per_thread,
+                                             num_k_block_tile_iteration,
+                                             p_x,
+                                             p_welford_mean,
+                                             p_welford_variance,
+                                             p_welford_count);
+};
+
+template <typename XDataType,
+          typename AccDataType,
+          typename MeanVarDataType,
+          typename XGridDesc_M_K,
+          typename MeanVarCountGridDesc_M_G,
+          typename GetReduceCountPerThreadFunctor,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcCountSrcVectorDim,
+          index_t XSrcCountSrcVectorSize>
+struct GridwiseMultiblockWelfordFirstHalf
+{
+    static_assert((XSrcCountSrcVectorDim == 0 && MThreadSliceSize % XSrcCountSrcVectorSize == 0) ||
+                      (XSrcCountSrcVectorDim == 1 &&
+                       KThreadSliceSize % XSrcCountSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XSrcCountSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder,
+                                              false>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    // clang-format off
+    // First half of the Multiblock Welford method to calculate mean and variance, used by both batchnorm-forward and batchnorm-backward.
+    // clang-format on
+    __device__ static void Run(const XGridDesc_M_K& x_grid_desc_m_k,
+                               const MeanVarCountGridDesc_M_G& mean_var_count_grid_desc_m_g,
+                               const GetReduceCountPerThreadFunctor& get_reduce_count_per_thread,
+                               index_t num_k_block_tile_iteration,
+                               const XDataType* const __restrict__ p_x,
+                               MeanVarDataType* const p_welford_mean,
+                               MeanVarDataType* const p_welford_variance,
+                               int32_t* const p_welford_count)
+    {
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
+            welford_count_thread_buf;
+
+        const index_t blkgroup_size = mean_var_count_grid_desc_m_g.GetLength(I1);
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / blkgroup_size;
+        const index_t block_local_id  = block_global_id % blkgroup_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M_1 = Sequence<MThreadSliceSize, 1>;
+
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m_1 = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<1>{}));
+
+        const index_t reduceSizePerBlock = K_BlockTileSize * num_k_block_tile_iteration;
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcCountSrcVectorDim,
+                                                                  XSrcCountSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             block_local_id * reduceSizePerBlock +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_welford_mean_var_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               MeanVarDataType,
+                                               decltype(thread_buffer_desc_m_1),
+                                               MeanVarCountGridDesc_M_G,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_1,
+                                               Sequence<0, 1>,
+                                               1,
+                                               1,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                mean_var_count_grid_desc_m_g,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 block_local_id),
+                PassThroughOp{});
+
+        auto threadwise_welford_count_store =
+            ThreadwiseTensorSliceTransfer_v1r3<int32_t,
+                                               int32_t,
+                                               decltype(thread_buffer_desc_m_1),
+                                               MeanVarCountGridDesc_M_G,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_1,
+                                               Sequence<0, 1>,
+                                               1,
+                                               1,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                mean_var_count_grid_desc_m_g,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 block_local_id),
+                PassThroughOp{});
+
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileSize);
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        auto welford_mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_mean, mean_var_count_grid_desc_m_g.GetElementSpaceSize());
+
+        auto welford_var_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_variance, mean_var_count_grid_desc_m_g.GetElementSpaceSize());
+
+        auto welford_count_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_count, mean_var_count_grid_desc_m_g.GetElementSpaceSize());
+
+        auto threadwise_welford = ThreadwiseWelford();
+        threadwise_welford.max_count_ =
+            get_reduce_count_per_thread(block_local_id, thread_k_cluster_id);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            welford_mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
+            welford_var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_val_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+            threadwise_welford.Run(x_thread_buf, welford_mean_thread_buf, welford_var_thread_buf);
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            welford_count_thread_buf(I) = threadwise_welford.cur_count_;
+            BlockwiseWelford::Run(
+                welford_mean_thread_buf(I), welford_var_thread_buf(I), welford_count_thread_buf(I));
+        });
+
+        if(thread_k_cluster_id == 0)
+        {
+            threadwise_welford_mean_var_store.Run(thread_buffer_desc_m_1,
+                                                  make_tuple(I0, I0),
+                                                  welford_mean_thread_buf,
+                                                  mean_var_count_grid_desc_m_g,
+                                                  welford_mean_global_val_buf);
+
+            threadwise_welford_mean_var_store.Run(thread_buffer_desc_m_1,
+                                                  make_tuple(I0, I0),
+                                                  welford_var_thread_buf,
+                                                  mean_var_count_grid_desc_m_g,
+                                                  welford_var_global_val_buf);
+
+            threadwise_welford_count_store.Run(thread_buffer_desc_m_1,
+                                               make_tuple(I0, I0),
+                                               welford_count_thread_buf,
+                                               mean_var_count_grid_desc_m_g,
+                                               welford_count_global_val_buf);
+        };
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final.hpp
new file mode 100644
index 0000000..548d7fd
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final.hpp
@@ -0,0 +1,571 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseWelfordSecondHalfBatchNormForwardFinal_,
+          typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename MeanVarCountGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M>
+__global__ void kernel_welford_second_half_batchnorm_forward_final(
+    const XYGridDesc_M_K x_grid_desc_m_k,
+    const XYGridDesc_M_K y_grid_desc_m_k,
+    const MeanVarCountGridDesc_M_K mean_var_count_grid_desc_m_k,
+    const ScaleBiasGridDesc_M scale_grid_desc_m,
+    const ScaleBiasGridDesc_M bias_grid_desc_m,
+    const MeanVarGridDesc_M mean_var_grid_desc_m,
+    index_t blkgroup_size,
+    index_t num_xy_k_block_tile_iteration,
+    index_t num_mean_var_count_k_block_tile_iteration,
+    AccDataType epsilon,
+    const MeanVarDataType* const __restrict__ p_in_welford_mean,
+    const MeanVarDataType* const __restrict__ p_in_welford_variance,
+    const int32_t* const __restrict__ p_in_welford_count,
+    const XDataType* const __restrict__ p_x,
+    const ScaleDataType* const __restrict__ p_scale,
+    const BiasDataType* const __restrict__ p_bias,
+    const YElementwiseOp y_elementwise_op,
+    YDataType* const __restrict__ p_y,
+    bool updateMovingAverage,
+    AccDataType averageFactor,
+    MeanVarDataType* const __restrict__ resultRunningMean,
+    MeanVarDataType* const __restrict__ resultRunningVariance,
+    bool saveMeanInvVariance,
+    MeanVarDataType* const __restrict__ resultSaveMean,
+    MeanVarDataType* const __restrict__ resultSaveInvVariance)
+{
+    GridwiseWelfordSecondHalfBatchNormForwardFinal_::Run(x_grid_desc_m_k,
+                                                         y_grid_desc_m_k,
+                                                         mean_var_count_grid_desc_m_k,
+                                                         scale_grid_desc_m,
+                                                         bias_grid_desc_m,
+                                                         mean_var_grid_desc_m,
+                                                         blkgroup_size,
+                                                         num_xy_k_block_tile_iteration,
+                                                         num_mean_var_count_k_block_tile_iteration,
+                                                         epsilon,
+                                                         p_in_welford_mean,
+                                                         p_in_welford_variance,
+                                                         p_in_welford_count,
+                                                         p_x,
+                                                         p_scale,
+                                                         p_bias,
+                                                         y_elementwise_op,
+                                                         p_y,
+                                                         updateMovingAverage,
+                                                         averageFactor,
+                                                         resultRunningMean,
+                                                         resultRunningVariance,
+                                                         saveMeanInvVariance,
+                                                         resultSaveMean,
+                                                         resultSaveInvVariance);
+};
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename MeanVarCountGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcYDstVectorDim,
+          index_t XSrcVectorSize,
+          index_t YDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t BiasSrcVectorSize,
+          index_t MeanVarSrcDstVectorSize>
+struct GridwiseWelfordSecondHalfBatchNormForwardFinal
+{
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XSrcYDstVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_1 = decltype(
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelfordMerge<AccDataType, ThreadReduceSrcDesc_M_1, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    __device__ static void Run(const XYGridDesc_M_K& x_grid_desc_m_k,
+                               const XYGridDesc_M_K& y_grid_desc_m_k,
+                               const MeanVarCountGridDesc_M_K& mean_var_count_grid_desc_m_k,
+                               const ScaleBiasGridDesc_M& scale_grid_desc_m,
+                               const ScaleBiasGridDesc_M& bias_grid_desc_m,
+                               const MeanVarGridDesc_M& mean_var_grid_desc_m,
+                               index_t blkgroup_size,
+                               index_t num_xy_k_block_tile_iteration,
+                               index_t num_mean_var_count_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               const MeanVarDataType* const __restrict__ p_in_welford_mean,
+                               const MeanVarDataType* const __restrict__ p_in_welford_variance,
+                               const int32_t* const __restrict__ p_in_welford_count,
+                               const XDataType* const __restrict__ p_x,
+                               const ScaleDataType* const __restrict__ p_scale,
+                               const BiasDataType* const __restrict__ p_bias,
+                               const YElementwiseOp y_elementwise_op,
+                               YDataType* const __restrict__ p_y,
+                               bool updateMovingAverage,
+                               AccDataType averageFactor,
+                               MeanVarDataType* const __restrict__ resultRunningMean,
+                               MeanVarDataType* const __restrict__ resultRunningVariance,
+                               bool saveMeanInvVariance,
+                               MeanVarDataType* const __restrict__ resultSaveMean,
+                               MeanVarDataType* const __restrict__ resultSaveInvVariance)
+
+    {
+        using ck::math::sqrt;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * 1, true>
+            in_welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * 1, true>
+            in_welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize * 1, true>
+            in_welford_count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
+            welford_count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            y_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> scale_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> bias_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / blkgroup_size;
+        const index_t block_local_id  = block_global_id % blkgroup_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M           = Sequence<MThreadSliceSize>;
+        using ThreadBufferLengths_M_1         = Sequence<MThreadSliceSize, 1>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m_1 = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<1>{}));
+
+        auto threadwise_mean_var_load_m_k =
+            ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                             AccDataType,
+                                             MeanVarCountGridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             Sequence<0, 1>,
+                                             1,
+                                             1,
+                                             1,
+                                             true>(
+                mean_var_count_grid_desc_m_k,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * 1));
+
+        auto threadwise_count_load_m_k =
+            ThreadwiseTensorSliceTransfer_v2<int32_t,
+                                             int32_t,
+                                             MeanVarCountGridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             Sequence<0, 1>,
+                                             1,
+                                             1,
+                                             1,
+                                             true>(
+                mean_var_count_grid_desc_m_k,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * 1));
+
+        const auto welford_mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_welford_mean, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto welford_var_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_welford_variance, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto welford_count_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_welford_count, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
+
+        constexpr auto mean_var_count_thread_copy_step_m_k =
+            make_multi_index(0, KThreadClusterSize * 1);
+
+        // Step 1: do final welford reduction to get mean and variance
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            welford_mean_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+            welford_var_thread_buf(I)   = type_convert<AccDataType>(0.0f);
+            welford_count_thread_buf(I) = 0;
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_mean_var_count_k_block_tile_iteration;
+            ++reducedTiles)
+        {
+            threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                             welford_mean_global_val_buf,
+                                             thread_buffer_desc_m_1,
+                                             make_tuple(I0, I0),
+                                             in_welford_mean_thread_buf);
+
+            threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                             welford_var_global_val_buf,
+                                             thread_buffer_desc_m_1,
+                                             make_tuple(I0, I0),
+                                             in_welford_var_thread_buf);
+
+            threadwise_count_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                          welford_count_global_val_buf,
+                                          thread_buffer_desc_m_1,
+                                          make_tuple(I0, I0),
+                                          in_welford_count_thread_buf);
+
+            ThreadwiseWelford::Run(in_welford_mean_thread_buf,
+                                   in_welford_var_thread_buf,
+                                   in_welford_count_thread_buf,
+                                   welford_mean_thread_buf,
+                                   welford_var_thread_buf,
+                                   welford_count_thread_buf);
+
+            threadwise_mean_var_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
+                                                            mean_var_count_thread_copy_step_m_k);
+            threadwise_count_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
+                                                         mean_var_count_thread_copy_step_m_k);
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            BlockwiseWelford::Run(
+                welford_mean_thread_buf(I), welford_var_thread_buf(I), welford_count_thread_buf(I));
+        });
+
+        // Step 2: do normalization and output y
+
+        const index_t workSizePerBlock = K_BlockTileSize * num_xy_k_block_tile_iteration;
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XYGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcYDstVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             workSizePerBlock * block_local_id +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               XYGridDesc_M_K,
+                                               YElementwiseOp,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               XSrcYDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                y_grid_desc_m_k,
+                make_multi_index(
+                    blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                    workSizePerBlock * block_local_id + thread_k_cluster_id * KThreadSliceSize),
+                y_elementwise_op);
+
+        auto threadwise_scale_load =
+            ThreadwiseTensorSliceTransfer_v2<ScaleDataType,
+                                             AccDataType,
+                                             ScaleBiasGridDesc_M,
+                                             decltype(thread_buffer_desc_m),
+                                             ThreadBufferLengths_M,
+                                             Sequence<0>,
+                                             0,
+                                             ScaleSrcVectorSize,
+                                             1,
+                                             true>(
+                scale_grid_desc_m,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize));
+
+        auto threadwise_bias_load = ThreadwiseTensorSliceTransfer_v2<BiasDataType,
+                                                                     AccDataType,
+                                                                     ScaleBiasGridDesc_M,
+                                                                     decltype(thread_buffer_desc_m),
+                                                                     ThreadBufferLengths_M,
+                                                                     Sequence<0>,
+                                                                     0,
+                                                                     BiasSrcVectorSize,
+                                                                     1,
+                                                                     true>(
+            bias_grid_desc_m,
+            make_multi_index(blkgroup_id * M_BlockTileSize +
+                             thread_m_cluster_id * MThreadSliceSize));
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto scale_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_scale, scale_grid_desc_m.GetElementSpaceSize());
+
+        const auto bias_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_bias, bias_grid_desc_m.GetElementSpaceSize());
+
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y, y_grid_desc_m_k.GetElementSpaceSize());
+
+        threadwise_scale_load.Run(scale_grid_desc_m,
+                                  scale_global_val_buf,
+                                  thread_buffer_desc_m,
+                                  make_tuple(I0),
+                                  scale_thread_buf);
+
+        threadwise_bias_load.Run(bias_grid_desc_m,
+                                 bias_global_val_buf,
+                                 thread_buffer_desc_m,
+                                 make_tuple(I0),
+                                 bias_thread_buf);
+
+        constexpr auto xy_thread_copy_step_m_k = make_multi_index(0, K_BlockTileSize);
+
+        for(index_t workTiles = 0; workTiles < num_xy_k_block_tile_iteration; ++workTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_val_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                AccDataType multiplier =
+                    scale_thread_buf[iM] / sqrt(welford_var_thread_buf[iM] + epsilon);
+
+                AccDataType fused_mean_bias =
+                    bias_thread_buf[iM] - welford_mean_thread_buf[iM] * multiplier;
+
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    y_thread_buf(Number<offset>{}) =
+                        x_thread_buf[Number<offset>{}] * multiplier + fused_mean_bias;
+                });
+            });
+
+            threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   y_thread_buf,
+                                   y_grid_desc_m_k,
+                                   y_global_val_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, xy_thread_copy_step_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, xy_thread_copy_step_m_k);
+        }
+
+        // Step 3: update the moving average of mean and variance (optional)
+
+        if(updateMovingAverage && block_local_id == 0 && thread_k_cluster_id == 0)
+        {
+            StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+                running_mean_thread_buf;
+            StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+                running_var_thread_buf;
+
+            auto running_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultRunningMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto running_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultRunningVariance, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto threadwise_mean_var_load_m =
+                ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                                 AccDataType,
+                                                 MeanVarGridDesc_M,
+                                                 decltype(thread_buffer_desc_m),
+                                                 ThreadBufferLengths_M,
+                                                 Sequence<0>,
+                                                 0,
+                                                 MeanVarSrcDstVectorSize,
+                                                 1,
+                                                 true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize));
+
+            threadwise_mean_var_load_m.Run(mean_var_grid_desc_m,
+                                           running_mean_global_buf,
+                                           thread_buffer_desc_m,
+                                           make_tuple(I0),
+                                           running_mean_thread_buf);
+
+            threadwise_mean_var_load_m.Run(mean_var_grid_desc_m,
+                                           running_var_global_buf,
+                                           thread_buffer_desc_m,
+                                           make_tuple(I0),
+                                           running_var_thread_buf);
+
+            AccDataType oneMinusAverageFactor = type_convert<AccDataType>(1.0) - averageFactor;
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                running_mean_thread_buf(I) = running_mean_thread_buf[I] * oneMinusAverageFactor +
+                                             welford_mean_thread_buf[I] * averageFactor;
+                running_var_thread_buf(I) = running_var_thread_buf[I] * oneMinusAverageFactor +
+                                            welford_var_thread_buf[I] * averageFactor;
+            });
+
+            auto threadwise_mean_var_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   MeanVarDataType,
+                                                   decltype(thread_buffer_desc_m),
+                                                   MeanVarGridDesc_M,
+                                                   PassThroughOp,
+                                                   ThreadBufferLengths_M,
+                                                   Sequence<0>,
+                                                   0,
+                                                   MeanVarSrcDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_mean_var_store.Run(thread_buffer_desc_m,
+                                          make_tuple(I0),
+                                          running_mean_thread_buf,
+                                          mean_var_grid_desc_m,
+                                          running_mean_global_buf);
+
+            threadwise_mean_var_store.Run(thread_buffer_desc_m,
+                                          make_tuple(I0),
+                                          running_var_thread_buf,
+                                          mean_var_grid_desc_m,
+                                          running_var_global_buf);
+        };
+
+        // Step 4: save mean and inv-variance (optional)
+
+        if(saveMeanInvVariance && block_local_id == 0 && thread_k_cluster_id == 0)
+        {
+            auto result_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultSaveMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto result_inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultSaveInvVariance, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            // calculate inv-variance as 1/sqrt(epsilon+variance)
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                welford_var_thread_buf(I) =
+                    type_convert<AccDataType>(1.0f) / sqrt(epsilon + welford_var_thread_buf[I]);
+            });
+
+            auto threadwise_mean_inv_var_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   MeanVarDataType,
+                                                   decltype(thread_buffer_desc_m),
+                                                   MeanVarGridDesc_M,
+                                                   PassThroughOp,
+                                                   ThreadBufferLengths_M,
+                                                   Sequence<0>,
+                                                   0,
+                                                   MeanVarSrcDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              welford_mean_thread_buf,
+                                              mean_var_grid_desc_m,
+                                              result_mean_global_buf);
+
+            threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              welford_var_thread_buf,
+                                              mean_var_grid_desc_m,
+                                              result_inv_var_global_buf);
+        };
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp
new file mode 100644
index 0000000..42b7e17
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp
@@ -0,0 +1,556 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseWelfordSecondHalfReduceFirstHalf_,
+          typename XDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename MeanVarGridDesc_M,
+          typename MeanVarCountGridDesc_M_K,
+          typename DscaleDbiasGridDesc_M_G>
+__global__ void kernel_welford_second_half_reduce_first_half(
+    const XYGridDesc_M_K x_grid_desc_m_k,
+    const XYGridDesc_M_K dy_grid_desc_m_k,
+    const MeanVarGridDesc_M mean_var_grid_desc_m,
+    const MeanVarCountGridDesc_M_K mean_var_count_grid_desc_m_k,
+    const DscaleDbiasGridDesc_M_G dscale_dbias_grid_desc_m_g,
+    index_t blkgroup_size,
+    index_t num_xy_k_block_tile_iteration,
+    index_t num_mean_var_count_k_block_tile_iteration,
+    AccDataType epsilon,
+    bool haveSavedMeanInvVar,
+    const MeanVarDataType* const __restrict__ p_savedMean,
+    const MeanVarDataType* const __restrict__ p_savedInvVar,
+    const MeanVarDataType* const __restrict__ p_in_welford_mean,
+    const MeanVarDataType* const __restrict__ p_in_welford_variance,
+    const int32_t* const __restrict__ p_in_welford_count,
+    const DyElementwiseOp dy_elementwise_op,
+    MeanVarDataType* const __restrict__ p_out_welford_mean,
+    MeanVarDataType* const __restrict__ p_out_welford_inv_variance,
+    const XDataType* const __restrict__ p_x,
+    const DyDataType* const __restrict__ p_dy,
+    DscaleDbiasDataType* const __restrict__ p_reduce_dscale,
+    DscaleDbiasDataType* const __restrict__ p_reduce_dbias)
+{
+    GridwiseWelfordSecondHalfReduceFirstHalf_::Run(x_grid_desc_m_k,
+                                                   dy_grid_desc_m_k,
+                                                   mean_var_grid_desc_m,
+                                                   mean_var_count_grid_desc_m_k,
+                                                   dscale_dbias_grid_desc_m_g,
+                                                   blkgroup_size,
+                                                   num_xy_k_block_tile_iteration,
+                                                   num_mean_var_count_k_block_tile_iteration,
+                                                   epsilon,
+                                                   haveSavedMeanInvVar,
+                                                   p_savedMean,
+                                                   p_savedInvVar,
+                                                   p_in_welford_mean,
+                                                   p_in_welford_variance,
+                                                   p_in_welford_count,
+                                                   dy_elementwise_op,
+                                                   p_out_welford_mean,
+                                                   p_out_welford_inv_variance,
+                                                   p_x,
+                                                   p_dy,
+                                                   p_reduce_dscale,
+                                                   p_reduce_dbias);
+};
+
+template <typename XDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename MeanVarGridDesc_M,
+          typename MeanVarCountGridDesc_M_K,
+          typename DscaleDbiasGridDesc_M_G,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XDyVectorDim,
+          index_t XSrcVectorSize,
+          index_t DySrcVectorSize,
+          index_t MeanVarSrcVectorSize>
+struct GridwiseWelfordSecondHalfReduceFirstHalf
+{
+    static_assert((XDyVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0 &&
+                   MThreadSliceSize % DySrcVectorSize == 0) ||
+                      (XDyVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0 &&
+                       KThreadSliceSize % DySrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XDyVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceSrcDesc_M_1 = decltype(
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelfordMerge<AccDataType, ThreadReduceSrcDesc_M_1, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    using BlockwiseReduce = PartitionedBlockwiseReduction<AccDataType,
+                                                          BlockSize,
+                                                          ThreadClusterLengths_M_K,
+                                                          ThreadClusterArrangeOrder,
+                                                          ck::reduce::Add,
+                                                          false>;
+
+    using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                 ThreadReduceSrcDesc_M_K,
+                                                 ThreadReduceDstDesc_M,
+                                                 ck::reduce::Add,
+                                                 false>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    // clang-format off
+    // Two of the steps of Multiblock BatchNorm Backward
+    // Step 1: Second half of Welford method to calculate mean and variance, as well as getting inv-variance = 1/sqrt(epsilon+variance) 
+    // Step 2: First half of Reduction: dbias = sum(dy), dscale = sum(dy * (x-mean) * inv-variance)
+    // clang-format on
+    __device__ static void Run(const XYGridDesc_M_K& x_grid_desc_m_k,
+                               const XYGridDesc_M_K& dy_grid_desc_m_k,
+                               const MeanVarGridDesc_M& mean_var_grid_desc_m,
+                               const MeanVarCountGridDesc_M_K& mean_var_count_grid_desc_m_k,
+                               const DscaleDbiasGridDesc_M_G& dscale_dbias_grid_desc_m_g,
+                               index_t blkgroup_size,
+                               index_t num_xy_k_block_tile_iteration,
+                               index_t num_mean_var_count_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               bool haveSavedMeanInvVar,
+                               const MeanVarDataType* const __restrict__ p_savedMean,
+                               const MeanVarDataType* const __restrict__ p_savedInvVar,
+                               const MeanVarDataType* const __restrict__ p_in_welford_mean,
+                               const MeanVarDataType* const __restrict__ p_in_welford_variance,
+                               const int32_t* const __restrict__ p_in_welford_count,
+                               const DyElementwiseOp dy_elementwise_op,
+                               MeanVarDataType* const __restrict__ p_out_welford_mean,
+                               MeanVarDataType* const __restrict__ p_out_welford_inv_variance,
+                               const XDataType* const __restrict__ p_x,
+                               const DyDataType* const __restrict__ p_dy,
+                               DscaleDbiasDataType* const __restrict__ p_reduce_dscale,
+                               DscaleDbiasDataType* const __restrict__ p_reduce_dbias)
+    {
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * 1, true>
+            in_welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * 1, true>
+            in_welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize * 1, true>
+            in_welford_count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
+            welford_count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>& mean_thread_buf =
+            welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>&
+            inv_var_thread_buf = welford_var_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            dy_thread_buf;
+
+        // buffer of values of dy * (x-mean) * inv-variance, used as input of Blockwise reduction
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            tmp1_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            reduce_dscale_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            reduce_dbias_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / blkgroup_size;
+        const index_t block_local_id  = block_global_id % blkgroup_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M           = Sequence<MThreadSliceSize>;
+        using ThreadBufferLengths_M_1         = Sequence<MThreadSliceSize, 1>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m_1 = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<1>{}));
+
+        // clang-format off
+        // Step 1: load existing mean and inv-variance, or do final welford reduction on mean and variance as well as get inv-variance = 1/sqrt(epsilon+variance)
+        // clang-format on
+
+        if(haveSavedMeanInvVar)
+        {
+            const auto mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_savedMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            const auto inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_savedInvVar, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto threadwise_mean_inv_var_load =
+                ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                                 AccDataType,
+                                                 MeanVarGridDesc_M,
+                                                 decltype(thread_buffer_desc_m),
+                                                 ThreadBufferLengths_M,
+                                                 Sequence<0>,
+                                                 0,
+                                                 MeanVarSrcVectorSize,
+                                                 1,
+                                                 true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize));
+
+            threadwise_mean_inv_var_load.Run(mean_var_grid_desc_m,
+                                             mean_global_buf,
+                                             thread_buffer_desc_m,
+                                             make_tuple(I0),
+                                             mean_thread_buf);
+
+            threadwise_mean_inv_var_load.Run(mean_var_grid_desc_m,
+                                             inv_var_global_buf,
+                                             thread_buffer_desc_m,
+                                             make_tuple(I0),
+                                             inv_var_thread_buf);
+        }
+        else
+        {
+            const auto welford_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_in_welford_mean, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
+
+            const auto welford_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_in_welford_variance, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
+
+            const auto welford_count_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_in_welford_count, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
+
+            auto threadwise_mean_var_load_m_k =
+                ThreadwiseTensorSliceTransfer_v2<AccDataType,
+                                                 AccDataType,
+                                                 MeanVarCountGridDesc_M_K,
+                                                 decltype(thread_buffer_desc_m_1),
+                                                 ThreadBufferLengths_M_1,
+                                                 Sequence<0, 1>,
+                                                 1,
+                                                 1,
+                                                 1,
+                                                 true>(
+                    mean_var_count_grid_desc_m_k,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize,
+                                     thread_k_cluster_id * 1));
+
+            auto threadwise_count_load_m_k =
+                ThreadwiseTensorSliceTransfer_v2<int32_t,
+                                                 int32_t,
+                                                 MeanVarCountGridDesc_M_K,
+                                                 decltype(thread_buffer_desc_m_1),
+                                                 ThreadBufferLengths_M_1,
+                                                 Sequence<0, 1>,
+                                                 1,
+                                                 1,
+                                                 1,
+                                                 true>(
+                    mean_var_count_grid_desc_m_k,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize,
+                                     thread_k_cluster_id * 1));
+
+            constexpr auto mean_var_count_thread_copy_step_m_k =
+                make_multi_index(0, KThreadClusterSize * 1);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                welford_mean_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+                welford_var_thread_buf(I)   = type_convert<AccDataType>(0.0f);
+                welford_count_thread_buf(I) = 0;
+            });
+
+            for(index_t reducedTiles = 0; reducedTiles < num_mean_var_count_k_block_tile_iteration;
+                ++reducedTiles)
+            {
+                threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                                 welford_mean_global_buf,
+                                                 thread_buffer_desc_m_1,
+                                                 make_tuple(I0, I0),
+                                                 in_welford_mean_thread_buf);
+
+                threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                                 welford_var_global_buf,
+                                                 thread_buffer_desc_m_1,
+                                                 make_tuple(I0, I0),
+                                                 in_welford_var_thread_buf);
+
+                threadwise_count_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                              welford_count_global_buf,
+                                              thread_buffer_desc_m_1,
+                                              make_tuple(I0, I0),
+                                              in_welford_count_thread_buf);
+
+                ThreadwiseWelford::Run(in_welford_mean_thread_buf,
+                                       in_welford_var_thread_buf,
+                                       in_welford_count_thread_buf,
+                                       welford_mean_thread_buf,
+                                       welford_var_thread_buf,
+                                       welford_count_thread_buf);
+
+                threadwise_mean_var_load_m_k.MoveSrcSliceWindow(
+                    mean_var_count_grid_desc_m_k, mean_var_count_thread_copy_step_m_k);
+                threadwise_count_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
+                                                             mean_var_count_thread_copy_step_m_k);
+            }
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                BlockwiseWelford::Run(welford_mean_thread_buf(I),
+                                      welford_var_thread_buf(I),
+                                      welford_count_thread_buf(I));
+            });
+
+            // calculate inv-variance as 1/sqrt(epsilon+variance), stored in place of variance
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                welford_var_thread_buf(I) =
+                    type_convert<AccDataType>(1.0) / sqrt(welford_var_thread_buf[I] + epsilon);
+            });
+
+            if(block_local_id == 0 && thread_k_cluster_id == 0)
+            {
+
+                auto threadwise_mean_inv_var_store =
+                    ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                       MeanVarDataType,
+                                                       decltype(thread_buffer_desc_m),
+                                                       MeanVarGridDesc_M,
+                                                       PassThroughOp,
+                                                       ThreadBufferLengths_M,
+                                                       Sequence<0>,
+                                                       0,
+                                                       1,
+                                                       InMemoryDataOperationEnum::Set,
+                                                       1,
+                                                       true>(
+                        mean_var_grid_desc_m,
+                        make_multi_index(blkgroup_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize),
+                        PassThroughOp{});
+
+                auto mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_out_welford_mean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+                auto inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_out_welford_inv_variance, mean_var_grid_desc_m.GetElementSpaceSize());
+
+                threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                                  make_tuple(I0),
+                                                  mean_thread_buf,
+                                                  mean_var_grid_desc_m,
+                                                  mean_global_buf);
+
+                threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                                  make_tuple(I0),
+                                                  inv_var_thread_buf,
+                                                  mean_var_grid_desc_m,
+                                                  inv_var_global_buf);
+            };
+        };
+
+        const index_t workSizePerBlock = K_BlockTileSize * num_xy_k_block_tile_iteration;
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XYGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XDyVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             workSizePerBlock * block_local_id +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dy_load = ThreadwiseTensorSliceTransfer_v2<DyDataType,
+                                                                   AccDataType,
+                                                                   XYGridDesc_M_K,
+                                                                   decltype(thread_buffer_desc_m_k),
+                                                                   ThreadBufferLengths_M_K,
+                                                                   ThreadBufferDimAccessOrder,
+                                                                   XDyVectorDim,
+                                                                   DySrcVectorSize,
+                                                                   1,
+                                                                   true>(
+            dy_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             workSizePerBlock * block_local_id +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        const auto x_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto dy_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dy, dy_grid_desc_m_k.GetElementSpaceSize());
+
+        constexpr auto xy_thread_copy_step_m_k = make_multi_index(0, K_BlockTileSize);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            reduce_dscale_thread_buf(I) = type_convert<AccDataType>(0);
+            reduce_dbias_thread_buf(I)  = type_convert<AccDataType>(0);
+        });
+
+        // clang-format off
+        // Step 2: first-half of reduction: dbias = sum(dy), dscale = sum(dy * (x-mean) * inv-variance)
+        // clang-format on
+
+        for(index_t reducedTiles = 0; reducedTiles < num_xy_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_dy_load.Run(dy_grid_desc_m_k,
+                                   dy_global_buf,
+                                   thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   dy_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    dy_elementwise_op(dy_thread_buf(Number<offset>{}),
+                                      dy_thread_buf[Number<offset>{}]);
+
+                    AccDataType norm_x = (x_thread_buf[Number<offset>{}] - mean_thread_buf[iM]) *
+                                         inv_var_thread_buf[iM];
+
+                    tmp1_thread_buf(Number<offset>{}) = norm_x * dy_thread_buf[Number<offset>{}];
+                });
+            });
+
+            ThreadwiseReduce::Reduce(tmp1_thread_buf, reduce_dscale_thread_buf);
+            ThreadwiseReduce::Reduce(dy_thread_buf, reduce_dbias_thread_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, xy_thread_copy_step_m_k);
+            threadwise_dy_load.MoveSrcSliceWindow(dy_grid_desc_m_k, xy_thread_copy_step_m_k);
+        };
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            BlockwiseReduce::Reduce(reduce_work_buf, reduce_dscale_thread_buf(I));
+            block_sync_lds();
+            BlockwiseReduce::Reduce(reduce_work_buf, reduce_dbias_thread_buf(I));
+        });
+
+        auto threadwise_dscale_dbias_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               DscaleDbiasDataType,
+                                               decltype(thread_buffer_desc_m_1),
+                                               DscaleDbiasGridDesc_M_G,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_1,
+                                               Sequence<0, 1>,
+                                               1,
+                                               1,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                dscale_dbias_grid_desc_m_g,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 block_local_id),
+                PassThroughOp{});
+
+        auto reduce_dscale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_reduce_dscale, dscale_dbias_grid_desc_m_g.GetElementSpaceSize());
+
+        auto reduce_dbias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_reduce_dbias, dscale_dbias_grid_desc_m_g.GetElementSpaceSize());
+
+        if(thread_k_cluster_id == 0)
+        {
+            threadwise_dscale_dbias_store.Run(thread_buffer_desc_m_1,
+                                              make_tuple(I0, I0),
+                                              reduce_dscale_thread_buf,
+                                              dscale_dbias_grid_desc_m_g,
+                                              reduce_dscale_global_buf);
+
+            threadwise_dscale_dbias_store.Run(thread_buffer_desc_m_1,
+                                              make_tuple(I0, I0),
+                                              reduce_dbias_thread_buf,
+                                              dscale_dbias_grid_desc_m_g,
+                                              reduce_dbias_global_buf);
+        };
+    };
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
new file mode 100644
index 0000000..460b684
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
@@ -0,0 +1,546 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/number.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+
+namespace ck {
+
+// Rows of column-vectors
+template <index_t MPerBlock,
+          index_t NPerBlock,
+          typename CGridDesc_M_N,
+          bool DeviceCTileIndexCheck = false>
+struct BlockToCTileMap_M00_N0_M01
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    __host__ __device__ BlockToCTileMap_M00_N0_M01() = default;
+
+    __host__ __device__ BlockToCTileMap_M00_N0_M01(const CGridDesc_M_N& c_grid_desc_m_n,
+                                                   index_t M01 = 1)
+        : M01_(M01), underlying_map_(GetBlockToCTileMap(c_grid_desc_m_n, M01))
+    {
+    }
+
+    __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const auto M00 = math::integer_divide_ceil(M0, M01_);
+
+        const index_t grid_size = M00 * M01_ * N0;
+
+        return grid_size;
+    }
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        return underlying_map_.CalculateBottomIndex(idx_top);
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                             const CTileDim& c_tile_dim) const
+    {
+        if constexpr(DeviceCTileIndexCheck)
+            return DefaultValidCTileIndex(c_tile_idx, c_tile_dim);
+        else
+            return true;
+    }
+
+    __host__ bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        if constexpr(DeviceCTileIndexCheck)
+            return true; // validity check moved to kernel
+
+        const index_t M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        if(M0 % M01_ == 0)
+        {
+            return true;
+        }
+        else
+        {
+            return false;
+        }
+    }
+
+    private:
+    __host__ __device__ static constexpr auto
+    GetBlockToCTileMap(const CGridDesc_M_N& c_grid_desc_m_n, index_t M01)
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const auto M00 = math::integer_divide_ceil(M0, M01);
+
+        const auto m00_n0_m01_to_m0_n0_block_cluster_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_insert_transform(1),
+                       make_unmerge_transform(make_tuple(M00, M01)),
+                       make_pass_through_transform(make_tuple(N0))),
+            make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3>{}, Sequence<2>{}));
+
+        const auto cblockid_to_m00_n0_m01_block_cluster_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(1, M00, N0, M01))),
+            make_tuple(Sequence<0, 1, 2, 3>{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto cblockid_to_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(m00_n0_m01_to_m0_n0_block_cluster_adaptor,
+                                  cblockid_to_m00_n0_m01_block_cluster_adaptor);
+
+        return cblockid_to_m0_n0_block_cluster_adaptor;
+    }
+
+    index_t M01_;
+    using UnderlyingMap = decltype(GetBlockToCTileMap(CGridDesc_M_N{}, 1));
+    UnderlyingMap underlying_map_;
+};
+
+// Rows of column-vectors
+// This C-tile map dynamically adjusts M01 when C-tile index is out of range
+template <index_t MPerBlock, index_t NPerBlock, typename CGridDesc_M_N>
+struct BlockToCTileMap_M00_N0_M01Adapt
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    __host__ __device__ BlockToCTileMap_M00_N0_M01Adapt() = default;
+
+    __host__ __device__ BlockToCTileMap_M00_N0_M01Adapt(const CGridDesc_M_N& c_grid_desc_m_n,
+                                                        index_t M01 = 8)
+        : M01_(M01), c_grid_desc_m_n_(c_grid_desc_m_n)
+    {
+    }
+
+    __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const index_t grid_size = M0 * N0;
+
+        return grid_size;
+    }
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        auto block_1d_id = idx_top[I0];
+
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I1), NPerBlock);
+
+        block_1d_id = block_1d_id % (M0 * N0); // swallow batch index
+
+        index_t idx_N0 = block_1d_id % N0;
+        index_t idx_M0 = block_1d_id / N0;
+
+        const auto M01_adapt = (idx_M0 < M0 - M0 % M01_) ? M01_ : M0 % M01_;
+
+        index_t idx_M00          = idx_M0 / M01_;
+        index_t idx_M01          = idx_M0 % M01_;
+        index_t idx_N0_M01_local = idx_N0 + idx_M01 * N0;
+
+        return make_tuple(idx_N0_M01_local % M01_adapt + idx_M00 * M01_,
+                          idx_N0_M01_local / M01_adapt);
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
+                                             const CTileDim& /* c_tile_dim */) const
+    {
+        return true; // always valid provided that user gets grid size from CalculateGridSize()
+    }
+
+    __host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const { return true; }
+
+    private:
+    index_t M01_;
+    CGridDesc_M_N c_grid_desc_m_n_;
+};
+
+// 2D slices of column-vectors in 3D space
+// This C-tile map dynamically adjusts M01 when C-tile index is out of range
+template <index_t MPerBlock, index_t NPerBlock, typename CGridDesc_M_N>
+struct BlockToCTileMap_KSplit_M00_N0_M01Adapt
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    __host__ __device__ BlockToCTileMap_KSplit_M00_N0_M01Adapt() = default;
+
+    __host__ __device__ BlockToCTileMap_KSplit_M00_N0_M01Adapt(const CGridDesc_M_N& c_grid_desc_m_n,
+                                                               index_t M01    = 8,
+                                                               index_t KSplit = 1)
+        : M01_(M01), KSplit_(KSplit), c_grid_desc_m_n_(c_grid_desc_m_n)
+    {
+    }
+
+    __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const index_t grid_size = M0 * N0 * KSplit_;
+
+        return grid_size;
+    }
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        auto block_1d_id = idx_top[I0];
+
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I1), NPerBlock);
+
+        block_1d_id = block_1d_id % (M0 * N0 * KSplit_); // hide groups
+
+        const index_t idx_ksplit = block_1d_id / (M0 * N0);
+        block_1d_id              = block_1d_id % (M0 * N0);
+
+        index_t idx_N0 = block_1d_id % N0;
+        index_t idx_M0 = block_1d_id / N0;
+
+        const auto M01_adapt = (idx_M0 < M0 - M0 % M01_) ? M01_ : M0 % M01_;
+
+        index_t idx_M00          = idx_M0 / M01_;
+        index_t idx_M01          = idx_M0 % M01_;
+        index_t idx_N0_M01_local = idx_N0 + idx_M01 * N0;
+
+        return make_tuple(idx_ksplit,
+                          idx_N0_M01_local % M01_adapt + idx_M00 * M01_,
+                          idx_N0_M01_local / M01_adapt);
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
+                                             const CTileDim& /* c_tile_dim */) const
+    {
+        return true; // always valid provided that user gets grid size from CalculateGridSize()
+    }
+
+    __host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const { return true; }
+
+    private:
+    index_t M01_;
+    index_t KSplit_;
+    CGridDesc_M_N c_grid_desc_m_n_;
+};
+
+// Blocks of row-vectors
+template <index_t MPerBlock,
+          index_t NPerBlock,
+          typename CGridDesc_M_N,
+          bool DeviceCTileIndexCheck = false>
+struct BlockToCTileMap_M00_N00_M01_N01
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    __host__ __device__ BlockToCTileMap_M00_N00_M01_N01() = default;
+
+    __host__ __device__ BlockToCTileMap_M00_N00_M01_N01(const CGridDesc_M_N& c_grid_desc_m_n,
+                                                        index_t M01 = 1,
+                                                        index_t N01 = 1)
+        : M01_(M01), N01_(N01), underlying_map_(GetBlockToCTileMap(c_grid_desc_m_n, M01, N01))
+    {
+    }
+
+    __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const auto M00 = math::integer_divide_ceil(M0, M01_);
+        const auto N00 = math::integer_divide_ceil(N0, N01_);
+
+        const index_t grid_size = M00 * M01_ * N00 * N01_;
+
+        return grid_size;
+    }
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        return underlying_map_.CalculateBottomIndex(idx_top);
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                             const CTileDim& c_tile_dim) const
+    {
+        if constexpr(DeviceCTileIndexCheck)
+            return DefaultValidCTileIndex(c_tile_idx, c_tile_dim);
+        else
+            return true;
+    }
+
+    __host__ bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        if constexpr(DeviceCTileIndexCheck)
+            return true; // validity check moved to kernel
+
+        const index_t M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const index_t N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+        if(M0 % M01_ == 0 && N0 % N01_ == 0)
+        {
+            return true;
+        }
+        else
+        {
+            return false;
+        }
+    }
+
+    private:
+    __host__ __device__ static constexpr auto
+    GetBlockToCTileMap(const CGridDesc_M_N& c_grid_desc_m_n, index_t M01, index_t N01)
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const auto M00 = math::integer_divide_ceil(M0, M01);
+        const auto N00 = math::integer_divide_ceil(N0, N01);
+
+        const auto m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_insert_transform(1), // swallow the carry from lower dimensions
+                           make_unmerge_transform(make_tuple(M00, M01)),
+                           make_unmerge_transform(make_tuple(N00, N01))),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 3>{}, Sequence<2, 4>{}));
+
+        const auto cblockid_to_m00_m01_n00_n01_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(1, M00, N00, M01, N01))),
+                make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto cblockid_to_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
+                                  cblockid_to_m00_m01_n00_n01_block_cluster_adaptor);
+
+        return cblockid_to_m0_n0_block_cluster_adaptor;
+    }
+
+    index_t M01_, N01_;
+    using UnderlyingMap = decltype(GetBlockToCTileMap(CGridDesc_M_N{}, 1, 1));
+    UnderlyingMap underlying_map_;
+};
+
+// 2D slices of row-vectors in 3D space
+template <index_t MPerBlock,
+          index_t NPerBlock,
+          typename CGridDesc_M_N,
+          bool DeviceCTileIndexCheck = false>
+struct BlockToCTileMap_KSplit_M00_N00_M01_N01
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    __host__ BlockToCTileMap_KSplit_M00_N00_M01_N01() = default;
+
+    __host__ BlockToCTileMap_KSplit_M00_N00_M01_N01(const CGridDesc_M_N& c_grid_desc_m_n,
+                                                    index_t M01    = 1,
+                                                    index_t N01    = 1,
+                                                    index_t KSplit = 1)
+        : c_grid_desc_m_n_(c_grid_desc_m_n),
+          M01_(M01),
+          N01_(N01),
+          KSplit_(KSplit),
+          underlying_map_(GetBlockToCTileMap(c_grid_desc_m_n, M01, N01, KSplit))
+    {
+    }
+
+    __host__ __device__ constexpr index_t
+    CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const auto M00 = math::integer_divide_ceil(M0, M01_);
+        const auto N00 = math::integer_divide_ceil(N0, N01_);
+
+        const index_t grid_size = M00 * M01_ * N00 * N01_ * KSplit_;
+
+        return grid_size;
+    }
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        static_assert(TopIdx::Size() == 1);
+
+        return underlying_map_.CalculateBottomIndex(
+            make_multi_index(idx_top[I0] % CalculateGridSize()));
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                             const CTileDim& c_tile_dim) const
+    {
+        if constexpr(DeviceCTileIndexCheck)
+            return DefaultValidCTileIndex(c_tile_idx, c_tile_dim);
+        else
+            return true;
+    }
+
+    __host__ bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        if constexpr(DeviceCTileIndexCheck)
+            return true; // validity check moved to kernel
+
+        const index_t M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const index_t N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+        if(M0 % M01_ == 0 && N0 % N01_ == 0)
+        {
+            return true;
+        }
+        else
+        {
+            return false;
+        }
+    }
+
+    private:
+    __device__ constexpr index_t CalculateGridSize() const
+    {
+        return CalculateGridSize(c_grid_desc_m_n_);
+    }
+
+    __host__ static constexpr auto GetBlockToCTileMap(const CGridDesc_M_N& c_grid_desc_m_n,
+                                                      index_t M01,
+                                                      index_t N01,
+                                                      index_t KSplit)
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const auto M00 = math::integer_divide_ceil(M0, M01);
+        const auto N00 = math::integer_divide_ceil(N0, N01);
+
+        const auto ksplit_m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_pass_through_transform(KSplit),
+                           make_unmerge_transform(make_tuple(M00, M01)),
+                           make_unmerge_transform(make_tuple(N00, N01))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 3>{}, Sequence<2, 4>{}));
+
+        const auto c_blockid_to_ksplit_m00_m01_n00_n01_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(KSplit, M00, N00, M01, N01))),
+                make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto c_blockid_to_ksplit_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(ksplit_m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
+                                  c_blockid_to_ksplit_m00_m01_n00_n01_block_cluster_adaptor);
+
+        return c_blockid_to_ksplit_m0_n0_block_cluster_adaptor;
+    }
+
+    CGridDesc_M_N c_grid_desc_m_n_;
+    index_t M01_, N01_, KSplit_;
+    using UnderlyingMap = decltype(GetBlockToCTileMap(CGridDesc_M_N{}, 1, 1, 1));
+    UnderlyingMap underlying_map_;
+};
+
+template <typename CTileIdx, typename CTileDim>
+__host__ __device__ bool DefaultValidCTileIndex(const CTileIdx& c_tile_idx,
+                                                const CTileDim& c_tile_dim)
+{
+    bool is_valid = false;
+
+    const index_t m_block = c_tile_dim[Number<0>{}];
+    const index_t n_block = c_tile_dim[Number<1>{}];
+
+    if constexpr(CTileIdx::Size() == 2)
+    {
+        const index_t m_block_idx = c_tile_idx[Number<0>{}];
+        const index_t n_block_idx = c_tile_idx[Number<1>{}];
+        if(0 <= m_block_idx && m_block_idx < m_block && 0 <= n_block_idx && n_block_idx < n_block)
+        {
+            is_valid = true;
+        }
+    }
+    else if constexpr(CTileIdx::Size() == 3)
+    {
+        const index_t ksplit_idx  = c_tile_idx[Number<0>{}];
+        const index_t m_block_idx = c_tile_idx[Number<1>{}];
+        const index_t n_block_idx = c_tile_idx[Number<2>{}];
+        if(0 <= m_block_idx && m_block_idx < m_block && 0 <= n_block_idx && n_block_idx < n_block)
+        {
+            is_valid = true;
+        }
+        ignore = ksplit_idx;
+    }
+
+    return is_valid;
+}
+
+// This wrapper class is for grouped gemm where it subtracts blockIdx by a value so that the
+// workgroups assigned to a given gemm problem have top index offsetted to range [0,
+// grid_size_per_gemm]
+template <typename UnderlyingBlockToCTileMap>
+struct OffsettedBlockToCTileMap
+{
+    using underlying_type = UnderlyingBlockToCTileMap;
+
+    OffsettedBlockToCTileMap(UnderlyingBlockToCTileMap block_to_ctile_map, index_t block_start)
+    {
+        block_to_ctile_map_ = block_to_ctile_map;
+        block_start_        = block_start;
+    }
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        return block_to_ctile_map_.CalculateBottomIndex(
+            make_multi_index(idx_top[Number<0>{}] - block_start_));
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                             const CTileDim& c_tile_dim) const
+    {
+        return block_to_ctile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        return block_to_ctile_map_.CheckValidity(c_grid_desc_m_n);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        return block_to_ctile_map_.CalculateGridSize(c_grid_desc_m_n);
+    }
+
+    UnderlyingBlockToCTileMap block_to_ctile_map_;
+    index_t block_start_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp
new file mode 100644
index 0000000..bdebe38
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp
@@ -0,0 +1,321 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseMultipleReduction,
+          index_t NumReduction,
+          typename InDataType,
+          typename OutDataTypePointerTuple,
+          typename AccDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M_Tuple,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple>
+__global__ void
+kernel_multiple_reduce_multiblock(const InGridDesc_M_K in_grid_desc_m_k,
+                                  const OutGridDesc_M_Tuple out_grid_desc_m_tuple,
+                                  const InElementwiseOperationTuple in_elementwise_op_tuple,
+                                  const AccElementwiseOperationTuple acc_elementwise_op_tuple,
+                                  index_t block_group_size,
+                                  index_t num_k_block_tile_iteration,
+                                  Array<AccDataType, NumReduction> alpha_values,
+                                  const InDataType* const __restrict__ p_in_value_global,
+                                  Array<AccDataType, NumReduction> beta_values,
+                                  OutDataTypePointerTuple p_out_value_global_tuple)
+{
+    GridwiseMultipleReduction::Run(in_grid_desc_m_k,
+                                   out_grid_desc_m_tuple,
+                                   in_elementwise_op_tuple,
+                                   acc_elementwise_op_tuple,
+                                   block_group_size,
+                                   num_k_block_tile_iteration,
+                                   alpha_values,
+                                   p_in_value_global,
+                                   beta_values,
+                                   p_out_value_global_tuple);
+};
+
+template <index_t NumReduction,
+          typename InDataType,
+          typename OutDataTypePointerTuple,
+          typename AccDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M_Tuple,
+          typename ReduceOperation,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          typename OutDstVectorSizeSeq>
+struct GridwiseMultipleReduction_mk_to_m_multiblock
+{
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(NumReduction == OutDataTypePointerTuple::Size() &&
+                      NumReduction == OutGridDesc_M_Tuple::Size() &&
+                      NumReduction == OutDstVectorSizeSeq::Size() &&
+                      NumReduction == InElementwiseOperationTuple::Size() &&
+                      NumReduction == AccElementwiseOperationTuple::Size(),
+                  "All tuple should have the same size as the number of Reductions!");
+
+    static constexpr bool reorder_thread_cluster = (InSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using BlockwiseReduce = PartitionedBlockwiseReduction<AccDataType,
+                                                          BlockSize,
+                                                          ThreadClusterLengths_M_K,
+                                                          ThreadClusterArrangeOrder,
+                                                          ReduceOperation,
+                                                          PropagateNan>;
+
+    using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                 ThreadReduceSrcDesc_M_K,
+                                                 ThreadReduceDstDesc_M,
+                                                 ReduceOperation,
+                                                 PropagateNan>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    using Accumulation = detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+    __device__ static void Run(const InGridDesc_M_K& in_grid_desc_m_k,
+                               const OutGridDesc_M_Tuple& out_grid_desc_m_tuple,
+                               const InElementwiseOperationTuple& in_elementwise_op_tuple,
+                               const AccElementwiseOperationTuple& acc_elementwise_op_tuple,
+                               index_t block_group_size,
+                               index_t num_k_block_tile_iteration,
+                               Array<AccDataType, NumReduction> alpha_values,
+                               const InDataType* const __restrict__ p_in_value_global,
+                               Array<AccDataType, NumReduction> beta_values,
+                               OutDataTypePointerTuple p_out_value_global_tuple)
+    {
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+        // LDS,  reused by all reductions
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
+        auto out_global_val_buf_tuple = generate_tuple(
+            [&](auto iR) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_out_value_global_tuple[iR], out_grid_desc_m_tuple[iR].GetElementSpaceSize());
+            },
+            Number<NumReduction>{});
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_buf;
+
+        auto in_thread_buf_tuple = generate_tuple(
+            [&](auto iR) {
+                (void)iR;
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * KThreadSliceSize,
+                                    true>{};
+            },
+            Number<NumReduction>{});
+
+        auto accu_value_buf_tuple = generate_tuple(
+            [&](auto iR) {
+                (void)iR;
+                return StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>{};
+            },
+            Number<NumReduction>{});
+
+        static_for<0, NumReduction, 1>{}([&](auto iR) {
+            static_for<0, MThreadSliceSize, 1>{}(
+                [&](auto J) { accu_value_buf_tuple(iR)(J) = identityVal; });
+        });
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / block_group_size;
+        const index_t block_local_id  = block_global_id % block_group_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        const index_t reduceSizePerBlock = K_BlockTileSize * num_k_block_tile_iteration;
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        auto threadwise_src_load = ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                                                    AccDataType,
+                                                                    InGridDesc_M_K,
+                                                                    decltype(thread_buffer_desc),
+                                                                    ThreadBufferLengths,
+                                                                    ThreadBufferDimAccessOrder,
+                                                                    InSrcVectorDim,
+                                                                    InSrcVectorSize,
+                                                                    1,
+                                                                    false>(
+            in_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             block_local_id * reduceSizePerBlock +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        constexpr auto in_thread_copy_step = make_multi_index(0, K_BlockTileSize);
+
+        index_t reducedTiles = 0;
+        do
+        {
+            threadwise_src_load.Run(in_grid_desc_m_k,
+                                    in_global_val_buf,
+                                    thread_buffer_desc,
+                                    make_tuple(I0, I0),
+                                    in_thread_buf);
+
+            static_for<0, NumReduction, 1>{}([&](auto iR) {
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    // do element-wise pre-reduction operation
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                        in_elementwise_op_tuple[iR](in_thread_buf_tuple(iR)(Number<offset>{}),
+                                                    in_thread_buf(Number<offset>{}));
+                    });
+                });
+
+                ThreadwiseReduce::Reduce(in_thread_buf_tuple(iR), accu_value_buf_tuple(iR));
+            });
+
+            threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+            reducedTiles++;
+        } while(reducedTiles < num_k_block_tile_iteration);
+
+        constexpr auto reduced_data_desc = ThreadReduceDstDesc_M{};
+
+        static_for<0, NumReduction, 1>{}([&](auto iR) {
+            using OutDataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[iR])>;
+            using OutDataType        = remove_cvref_t<remove_pointer_t<OutDataTypePointer>>;
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                BlockwiseReduce::Reduce(reduce_work_buf, accu_value_buf_tuple(iR)(I));
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if(thread_k_cluster_id == 0)
+                {
+                    acc_elementwise_op_tuple[iR](accu_value_buf_tuple(iR)(I),
+                                                 accu_value_buf_tuple(iR)(I));
+
+                    accu_value_buf_tuple(iR)(I) *= alpha_values[iR];
+                }
+            });
+
+            if(thread_k_cluster_id == 0)
+            {
+                if(block_group_size == 0 && !float_equal_zero{}(beta_values[iR]))
+                {
+                    StaticBuffer<AddressSpaceEnum::Vgpr, OutDataType, MThreadSliceSize, true>
+                        priorDstValueBuf;
+
+                    auto threadwise_dst_load =
+                        ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                         OutDataType,
+                                                         decltype(out_grid_desc_m_tuple[iR]),
+                                                         decltype(reduced_data_desc),
+                                                         Sequence<MThreadSliceSize>,
+                                                         Sequence<0>,
+                                                         0,
+                                                         OutDstVectorSizeSeq::At(iR),
+                                                         1,
+                                                         false>(
+                            out_grid_desc_m_tuple[iR],
+                            make_multi_index(blkgroup_id * M_BlockTileSize +
+                                             thread_m_cluster_id * MThreadSliceSize));
+
+                    threadwise_dst_load.Run(out_grid_desc_m_tuple[iR],
+                                            out_global_val_buf_tuple(iR),
+                                            reduced_data_desc,
+                                            make_tuple(I0),
+                                            priorDstValueBuf);
+
+                    static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                        accu_value_buf_tuple(iR)(I) +=
+                            type_convert<AccDataType>(priorDstValueBuf[I]) * beta_values[iR];
+                    });
+                };
+
+                auto threadwise_dst_store =
+                    ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                       OutDataType,
+                                                       decltype(reduced_data_desc),
+                                                       decltype(out_grid_desc_m_tuple[iR]),
+                                                       PassThroughOp,
+                                                       Sequence<MThreadSliceSize>,
+                                                       Sequence<0>,
+                                                       0,
+                                                       OutDstVectorSizeSeq::At(iR),
+                                                       OutMemoryDataOperation,
+                                                       1,
+                                                       true>(
+                        out_grid_desc_m_tuple[iR],
+                        make_multi_index(blkgroup_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize),
+                        PassThroughOp{});
+
+                threadwise_dst_store.Run(reduced_data_desc,
+                                         make_tuple(I0),
+                                         accu_value_buf_tuple[iR],
+                                         out_grid_desc_m_tuple[iR],
+                                         out_global_val_buf_tuple(iR));
+            };
+        });
+    };
+}; // namespace ck
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp
new file mode 100644
index 0000000..1313ec9
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp
@@ -0,0 +1,264 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseMultipleReduction,
+          index_t NumReduction,
+          typename InDataType,
+          typename OutDataTypePointerTuple,
+          typename AccDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M_Tuple,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple>
+__global__ void
+kernel_multiple_reduce_threadwise(const InGridDesc_M_K in_grid_desc_m_k,
+                                  const OutGridDesc_M_Tuple out_grid_desc_m_tuple,
+                                  const InElementwiseOperationTuple in_elementwise_op_tuple,
+                                  const AccElementwiseOperationTuple acc_elementwise_op_tuple,
+                                  Array<AccDataType, NumReduction> alpha_values,
+                                  const InDataType* const __restrict__ p_in_value_global,
+                                  Array<AccDataType, NumReduction> beta_values,
+                                  OutDataTypePointerTuple p_out_value_global_tuple)
+{
+    GridwiseMultipleReduction::Run(in_grid_desc_m_k,
+                                   out_grid_desc_m_tuple,
+                                   in_elementwise_op_tuple,
+                                   acc_elementwise_op_tuple,
+                                   alpha_values,
+                                   p_in_value_global,
+                                   beta_values,
+                                   p_out_value_global_tuple);
+};
+
+template <index_t NumReduction,
+          typename InDataType,
+          typename OutDataTypePointerTuple,
+          typename AccDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M_Tuple,
+          typename ReduceOperation,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          typename OutDstVectorSizeSeq>
+struct GridwiseMultipleReduction_mk_to_m_threadwise
+{
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(NumReduction == OutDataTypePointerTuple::Size() &&
+                      NumReduction == OutGridDesc_M_Tuple::Size() &&
+                      NumReduction == OutDstVectorSizeSeq::Size() &&
+                      NumReduction == InElementwiseOperationTuple::Size() &&
+                      NumReduction == AccElementwiseOperationTuple::Size(),
+                  "All tuple should have the same size as the number of Reductions!");
+
+    static constexpr bool reorder_thread_cluster = (InSrcVectorDim == 0);
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                 ThreadReduceSrcDesc_M_K,
+                                                 ThreadReduceDstDesc_M,
+                                                 ReduceOperation,
+                                                 PropagateNan>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+
+    using Accumulation = detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+    __device__ static void Run(const InGridDesc_M_K& in_grid_desc_m_k,
+                               const OutGridDesc_M_Tuple& out_grid_desc_m_tuple,
+                               const InElementwiseOperationTuple& in_elementwise_op_tuple,
+                               const AccElementwiseOperationTuple& acc_elementwise_op_tuple,
+                               Array<AccDataType, NumReduction> alpha_values,
+                               const InDataType* const __restrict__ p_in_value_global,
+                               Array<AccDataType, NumReduction> beta_values,
+                               OutDataTypePointerTuple p_out_value_global_tuple)
+    {
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
+        auto out_global_val_buf_tuple = generate_tuple(
+            [&](auto iR) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_out_value_global_tuple[iR], out_grid_desc_m_tuple[iR].GetElementSpaceSize());
+            },
+            Number<NumReduction>{});
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_buf;
+
+        auto in_thread_buf_tuple = generate_tuple(
+            [&](auto iR) {
+                (void)iR;
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * KThreadSliceSize,
+                                    true>{};
+            },
+            Number<NumReduction>{});
+
+        auto accu_value_buf_tuple = generate_tuple(
+            [&](auto iR) {
+                (void)iR;
+                return StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>{};
+            },
+            Number<NumReduction>{});
+
+        static_for<0, NumReduction, 1>{}([&](auto iR) {
+            static_for<0, MThreadSliceSize, 1>{}(
+                [&](auto J) { accu_value_buf_tuple(iR)(J) = identityVal; });
+        });
+
+        const index_t thread_global_1d_id = get_thread_global_1d_id();
+
+        const auto toReduceLength = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        auto threadwise_src_load = ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                                                    AccDataType,
+                                                                    InGridDesc_M_K,
+                                                                    decltype(thread_buffer_desc),
+                                                                    ThreadBufferLengths,
+                                                                    ThreadBufferDimAccessOrder,
+                                                                    InSrcVectorDim,
+                                                                    InSrcVectorSize,
+                                                                    1,
+                                                                    false>(
+            in_grid_desc_m_k, make_multi_index(thread_global_1d_id * MThreadSliceSize, 0));
+
+        constexpr auto in_thread_copy_step = make_multi_index(0, KThreadSliceSize);
+
+        index_t reducedLength = 0;
+        do
+        {
+            threadwise_src_load.Run(in_grid_desc_m_k,
+                                    in_global_val_buf,
+                                    thread_buffer_desc,
+                                    make_tuple(I0, I0),
+                                    in_thread_buf);
+
+            static_for<0, NumReduction, 1>{}([&](auto iR) {
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    // do element-wise pre-reduction operation
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                        in_elementwise_op_tuple[iR](in_thread_buf_tuple(iR)(Number<offset>{}),
+                                                    in_thread_buf(Number<offset>{}));
+                    });
+                });
+
+                ThreadwiseReduce::Reduce(in_thread_buf_tuple(iR), accu_value_buf_tuple(iR));
+            });
+
+            threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+            reducedLength += KThreadSliceSize;
+        } while(reducedLength < toReduceLength);
+
+        constexpr auto reduced_data_desc = ThreadReduceDstDesc_M{};
+
+        static_for<0, NumReduction, 1>{}([&](auto iR) {
+            using OutDataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[iR])>;
+            using OutDataType        = remove_cvref_t<remove_pointer_t<OutDataTypePointer>>;
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                acc_elementwise_op_tuple[iR](accu_value_buf_tuple(iR)(I),
+                                             accu_value_buf_tuple(iR)(I));
+
+                accu_value_buf_tuple(iR)(I) *= alpha_values[iR];
+            });
+
+            if(!float_equal_zero{}(beta_values[iR]))
+            {
+                StaticBuffer<AddressSpaceEnum::Vgpr, OutDataType, MThreadSliceSize, true>
+                    priorDstValueBuf;
+
+                auto threadwise_dst_load =
+                    ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                     OutDataType,
+                                                     decltype(out_grid_desc_m_tuple[iR]),
+                                                     decltype(reduced_data_desc),
+                                                     Sequence<MThreadSliceSize>,
+                                                     Sequence<0>,
+                                                     0,
+                                                     OutDstVectorSizeSeq::At(iR),
+                                                     1,
+                                                     false>(
+                        out_grid_desc_m_tuple[iR],
+                        make_multi_index(thread_global_1d_id * MThreadSliceSize));
+
+                threadwise_dst_load.Run(out_grid_desc_m_tuple[iR],
+                                        out_global_val_buf_tuple(iR),
+                                        reduced_data_desc,
+                                        make_tuple(I0),
+                                        priorDstValueBuf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                    accu_value_buf_tuple(iR)(I) +=
+                        type_convert<AccDataType>(priorDstValueBuf[I]) * beta_values[iR];
+                });
+            };
+
+            auto threadwise_dst_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   OutDataType,
+                                                   decltype(reduced_data_desc),
+                                                   decltype(out_grid_desc_m_tuple[iR]),
+                                                   PassThroughOp,
+                                                   Sequence<MThreadSliceSize>,
+                                                   Sequence<0>,
+                                                   0,
+                                                   OutDstVectorSizeSeq::At(iR),
+                                                   OutMemoryDataOperation,
+                                                   1,
+                                                   true>(
+                    out_grid_desc_m_tuple[iR],
+                    make_multi_index(thread_global_1d_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_dst_store.Run(reduced_data_desc,
+                                     make_tuple(I0),
+                                     accu_value_buf_tuple[iR],
+                                     out_grid_desc_m_tuple[iR],
+                                     out_global_val_buf_tuple(iR));
+        });
+    };
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp
new file mode 100644
index 0000000..6836a66
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp
@@ -0,0 +1,613 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseReduction,
+          bool OutputIndex,
+          bool HaveIndexInput,
+          typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation>
+__global__ void kernel_reduce_multiblock(const InGridDesc_M_K in_grid_desc_m_k,
+                                         const OutGridDesc_M out_grid_desc_m,
+                                         const InElementwiseOperation in_elementwise_op,
+                                         const AccElementwiseOperation acc_elementwise_op,
+                                         index_t block_group_size,
+                                         index_t num_k_block_tile_iteration,
+                                         AccDataType alpha,
+                                         const InDataType* const __restrict__ p_in_value_global,
+                                         const IndexDataType* const __restrict__ p_in_index_global,
+                                         AccDataType beta,
+                                         OutDataType* const __restrict__ p_out_value_global,
+                                         IndexDataType* const __restrict__ p_out_index_global)
+{
+    if constexpr(!OutputIndex)
+    {
+        (void)p_in_index_global;
+        (void)p_out_index_global;
+
+        GridwiseReduction::Run(in_grid_desc_m_k,
+                               out_grid_desc_m,
+                               in_elementwise_op,
+                               acc_elementwise_op,
+                               block_group_size,
+                               num_k_block_tile_iteration,
+                               alpha,
+                               p_in_value_global,
+                               beta,
+                               p_out_value_global);
+    }
+    else
+    {
+        GridwiseReduction::template RunWithIndex<HaveIndexInput>(in_grid_desc_m_k,
+                                                                 out_grid_desc_m,
+                                                                 in_elementwise_op,
+                                                                 acc_elementwise_op,
+                                                                 num_k_block_tile_iteration,
+                                                                 alpha,
+                                                                 p_in_value_global,
+                                                                 p_in_index_global,
+                                                                 beta,
+                                                                 p_out_value_global,
+                                                                 p_out_index_global);
+    };
+};
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize>
+struct GridwiseReduction_mk_to_m_multiblock
+{
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)) &&
+                      (MThreadSliceSize % OutDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (InSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using BlockwiseReduce = PartitionedBlockwiseReduction<AccDataType,
+                                                          BlockSize,
+                                                          ThreadClusterLengths_M_K,
+                                                          ThreadClusterArrangeOrder,
+                                                          ReduceOperation,
+                                                          PropagateNan>;
+
+    using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                 ThreadReduceSrcDesc_M_K,
+                                                 ThreadReduceDstDesc_M,
+                                                 ReduceOperation,
+                                                 PropagateNan>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    using Accumulation = detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+    __device__ static void Run(const InGridDesc_M_K& in_grid_desc_m_k,
+                               const OutGridDesc_M& out_grid_desc_m,
+                               const InElementwiseOperation& in_elementwise_op,
+                               const AccElementwiseOperation& acc_elementwise_op,
+                               index_t block_group_size,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType alpha,
+                               const InDataType* const __restrict__ p_in_value_global,
+                               AccDataType beta,
+                               OutDataType* const __restrict__ p_out_value_global)
+    {
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+        // LDS
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
+        auto out_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_value_global, out_grid_desc_m.GetElementSpaceSize());
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> accu_value_buf;
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) { accu_value_buf(I) = identityVal; });
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / block_group_size;
+        const index_t block_local_id  = block_global_id % block_group_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        const index_t reduceSizePerBlock = K_BlockTileSize * num_k_block_tile_iteration;
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        auto threadwise_src_load = ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                                                    AccDataType,
+                                                                    InGridDesc_M_K,
+                                                                    decltype(thread_buffer_desc),
+                                                                    ThreadBufferLengths,
+                                                                    ThreadBufferDimAccessOrder,
+                                                                    InSrcVectorDim,
+                                                                    InSrcVectorSize,
+                                                                    1,
+                                                                    false>(
+            in_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             block_local_id * reduceSizePerBlock +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        constexpr auto in_thread_copy_step = make_multi_index(0, K_BlockTileSize);
+
+        index_t reducedTiles = 0;
+        do
+        {
+            threadwise_src_load.Run(in_grid_desc_m_k,
+                                    in_global_val_buf,
+                                    thread_buffer_desc,
+                                    make_tuple(I0, I0),
+                                    in_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                // do element-wise pre-reduction operation
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset = thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                    in_elementwise_op(in_thread_buf(Number<offset>{}),
+                                      in_thread_buf(Number<offset>{}));
+                });
+            });
+
+            ThreadwiseReduce::Reduce(in_thread_buf, accu_value_buf);
+
+            threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+            reducedTiles++;
+        } while(reducedTiles < num_k_block_tile_iteration);
+
+        constexpr auto reduced_data_desc = ThreadReduceDstDesc_M{};
+
+        static_for<0, MThreadSliceSize, 1>{}(
+            [&](auto I) { BlockwiseReduce::Reduce(reduce_work_buf, accu_value_buf(I)); });
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if(thread_k_cluster_id == 0)
+            {
+                acc_elementwise_op(accu_value_buf(I), accu_value_buf(I));
+
+                accu_value_buf(I) *= alpha;
+            }
+        });
+
+        if(thread_k_cluster_id == 0)
+        {
+            if(block_group_size == 0 && !float_equal_zero{}(beta))
+            {
+                StaticBuffer<AddressSpaceEnum::Vgpr, OutDataType, MThreadSliceSize, true>
+                    priorDstValueBuf;
+
+                auto threadwise_dst_load =
+                    ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                     OutDataType,
+                                                     OutGridDesc_M,
+                                                     decltype(reduced_data_desc),
+                                                     Sequence<MThreadSliceSize>,
+                                                     Sequence<0>,
+                                                     0,
+                                                     OutDstVectorSize,
+                                                     1,
+                                                     false>(
+                        out_grid_desc_m,
+                        make_multi_index(blkgroup_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize));
+
+                threadwise_dst_load.Run(out_grid_desc_m,
+                                        out_global_val_buf,
+                                        reduced_data_desc,
+                                        make_tuple(I0),
+                                        priorDstValueBuf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                    accu_value_buf(I) += type_convert<AccDataType>(priorDstValueBuf[I]) * beta;
+                });
+            };
+
+            auto threadwise_dst_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   OutDataType,
+                                                   decltype(reduced_data_desc),
+                                                   OutGridDesc_M,
+                                                   PassThroughOp,
+                                                   Sequence<MThreadSliceSize>,
+                                                   Sequence<0>,
+                                                   0,
+                                                   OutDstVectorSize,
+                                                   OutMemoryDataOperation,
+                                                   1,
+                                                   true>(
+                    out_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_dst_store.Run(reduced_data_desc,
+                                     make_tuple(I0),
+                                     accu_value_buf,
+                                     out_grid_desc_m,
+                                     out_global_val_buf);
+        }
+    };
+
+    template <bool HaveIndexInput>
+    __device__ static void RunWithIndex(const InGridDesc_M_K& in_grid_desc_m_k,
+                                        const OutGridDesc_M& out_grid_desc_m,
+                                        const InElementwiseOperation in_elementwise_op,
+                                        const AccElementwiseOperation acc_elementwise_op,
+                                        index_t num_k_block_tile_iteration,
+                                        AccDataType alpha,
+                                        const InDataType* const __restrict__ p_in_value_global,
+                                        const IndexDataType* const __restrict__ p_in_index_global,
+                                        AccDataType beta,
+                                        OutDataType* const __restrict__ p_out_value_global,
+                                        IndexDataType* const __restrict__ p_out_index_global)
+    {
+        using BlockwiseReduceWithIndex =
+            PartitionedBlockwiseReductionWithIndex<AccDataType,
+                                                   IndexDataType,
+                                                   BlockSize,
+                                                   Sequence<MThreadClusterSize, KThreadClusterSize>,
+                                                   ThreadClusterArrangeOrder,
+                                                   ReduceOperation,
+                                                   PropagateNan>;
+
+        using AccumulationWithIndex = detail::AccumulateWithIndexAndNanCheck<PropagateNan,
+                                                                             ReduceOperation,
+                                                                             AccDataType,
+                                                                             IndexDataType>;
+
+        (void)in_elementwise_op;
+
+        // LDS
+        __shared__ AccDataType p_reduce_work_val_buffer[BlockSize];
+        __shared__ IndexDataType p_reduce_work_idx_buffer[BlockSize];
+
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
+        const auto in_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_index_global, in_grid_desc_m_k.GetElementSpaceSize());
+        auto out_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_value_global, out_grid_desc_m.GetElementSpaceSize());
+        auto out_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_index_global, out_grid_desc_m.GetElementSpaceSize());
+
+        auto reduce_work_val_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_val_buffer, BlockSize);
+        auto reduce_work_idx_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_idx_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_val_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     IndexDataType,
+                     MThreadSliceSize * KThreadSliceSize,
+                     true>
+            in_thread_idx_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> accu_value_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, IndexDataType, MThreadSliceSize, true> accu_index_buf;
+
+        const index_t thread_local_id    = get_thread_local_1d_id();
+        const index_t block_global_1d_id = get_block_1d_id();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        auto threadwise_src_val_load =
+            ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                             AccDataType,
+                                             InGridDesc_M_K,
+                                             decltype(thread_buffer_desc),
+                                             ThreadBufferLengths,
+                                             ThreadBufferDimAccessOrder,
+                                             InSrcVectorDim,
+                                             InSrcVectorSize,
+                                             1,
+                                             false>(
+                in_grid_desc_m_k,
+                make_multi_index(block_global_1d_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            accu_value_buf(I) = identityVal;
+            accu_index_buf(I) = 0;
+        });
+
+        constexpr auto in_thread_copy_step = make_multi_index(0, K_BlockTileSize);
+
+        index_t reducedTiles = 0;
+
+        if constexpr(HaveIndexInput)
+        {
+            auto threadwise_src_idx_load =
+                ThreadwiseTensorSliceTransfer_v2<IndexDataType,
+                                                 IndexDataType,
+                                                 InGridDesc_M_K,
+                                                 decltype(thread_buffer_desc),
+                                                 ThreadBufferLengths,
+                                                 ThreadBufferDimAccessOrder,
+                                                 InSrcVectorDim,
+                                                 InSrcVectorSize,
+                                                 1,
+                                                 false>(
+                    in_grid_desc_m_k,
+                    make_multi_index(block_global_1d_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize,
+                                     thread_k_cluster_id * KThreadSliceSize));
+
+            do
+            {
+                // load the thread slice
+                threadwise_src_val_load.Run(in_grid_desc_m_k,
+                                            in_global_val_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_val_buf);
+                threadwise_src_idx_load.Run(in_grid_desc_m_k,
+                                            in_global_idx_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_idx_buf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    AccDataType tmpValue   = identityVal;
+                    IndexDataType tmpIndex = 0;
+
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+
+                        AccumulationWithIndex::Calculate(tmpValue,
+                                                         in_thread_val_buf[Number<offset>{}],
+                                                         tmpIndex,
+                                                         in_thread_idx_buf[Number<offset>{}]);
+                    });
+
+                    BlockwiseReduceWithIndex::Reduce(
+                        reduce_work_val_buf, reduce_work_idx_buf, tmpValue, tmpIndex);
+
+                    AccumulationWithIndex::Calculate(
+                        accu_value_buf(iM), tmpValue, accu_index_buf(iM), tmpIndex);
+                });
+
+                threadwise_src_val_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+                threadwise_src_idx_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+                reducedTiles++;
+            } while(reducedTiles < num_k_block_tile_iteration);
+        }
+        else
+        {
+            index_t indexOffset = 0;
+
+            do
+            {
+                // load the thread slice
+                threadwise_src_val_load.Run(in_grid_desc_m_k,
+                                            in_global_val_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_val_buf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+
+                        // initialize the indices for the per-thread to-reduce values
+                        in_thread_idx_buf(Number<offset>{}) =
+                            indexOffset + thread_k_cluster_id * KThreadSliceSize + iK();
+
+                        // do element-wise pre-reduction operation
+                        in_elementwise_op(in_thread_val_buf(Number<offset>{}),
+                                          in_thread_val_buf(Number<offset>{}));
+                    });
+
+                    AccDataType tmpValue   = identityVal;
+                    IndexDataType tmpIndex = 0;
+
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+
+                        AccumulationWithIndex::Calculate(tmpValue,
+                                                         in_thread_val_buf[Number<offset>{}],
+                                                         tmpIndex,
+                                                         in_thread_idx_buf[Number<offset>{}]);
+                    });
+
+                    BlockwiseReduceWithIndex::Reduce(
+                        reduce_work_val_buf, reduce_work_idx_buf, tmpValue, tmpIndex);
+
+                    AccumulationWithIndex::Calculate(
+                        accu_value_buf(iM), tmpValue, accu_index_buf(iM), tmpIndex);
+                });
+
+                threadwise_src_val_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+                indexOffset += K_BlockTileSize;
+                reducedTiles++;
+            } while(reducedTiles < num_k_block_tile_iteration);
+        };
+
+        constexpr auto reduced_data_desc = ThreadReduceDstDesc_M{};
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if(thread_k_cluster_id == 0)
+            {
+                // for indiced operation, acc_elementwise_op shoud do nothing
+                acc_elementwise_op(accu_value_buf(I), accu_value_buf(I));
+
+                accu_value_buf(I) *= alpha;
+            }
+        });
+
+        if(thread_k_cluster_id == 0)
+        {
+            if(!float_equal_zero{}(beta))
+            {
+                StaticBuffer<AddressSpaceEnum::Vgpr, OutDataType, MThreadSliceSize, true>
+                    priorDstValueBuf;
+
+                auto threadwise_dst_load =
+                    ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                     OutDataType,
+                                                     OutGridDesc_M,
+                                                     decltype(reduced_data_desc),
+                                                     Sequence<MThreadSliceSize>,
+                                                     Sequence<0>,
+                                                     0,
+                                                     OutDstVectorSize,
+                                                     1,
+                                                     true>(
+                        out_grid_desc_m,
+                        make_multi_index(block_global_1d_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize));
+
+                threadwise_dst_load.Run(out_grid_desc_m,
+                                        out_global_val_buf,
+                                        reduced_data_desc,
+                                        make_tuple(I0),
+                                        priorDstValueBuf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                    accu_value_buf(I) += type_convert<AccDataType>(priorDstValueBuf[I]) * beta;
+                });
+            };
+
+            auto threadwise_dst_val_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   OutDataType,
+                                                   decltype(reduced_data_desc),
+                                                   OutGridDesc_M,
+                                                   PassThroughOp,
+                                                   Sequence<MThreadSliceSize>,
+                                                   Sequence<0>,
+                                                   0,
+                                                   OutDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    out_grid_desc_m,
+                    make_multi_index(block_global_1d_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            auto threadwise_dst_idx_store =
+                ThreadwiseTensorSliceTransfer_v1r3<IndexDataType,
+                                                   IndexDataType,
+                                                   decltype(reduced_data_desc),
+                                                   OutGridDesc_M,
+                                                   PassThroughOp,
+                                                   Sequence<MThreadSliceSize>,
+                                                   Sequence<0>,
+                                                   0,
+                                                   OutDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    out_grid_desc_m,
+                    make_multi_index(block_global_1d_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_dst_val_store.Run(reduced_data_desc,
+                                         make_tuple(I0),
+                                         accu_value_buf,
+                                         out_grid_desc_m,
+                                         out_global_val_buf);
+            threadwise_dst_idx_store.Run(reduced_data_desc,
+                                         make_tuple(I0),
+                                         accu_index_buf,
+                                         out_grid_desc_m,
+                                         out_global_idx_buf);
+        }
+    };
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp
new file mode 100644
index 0000000..6c5bd29
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp
@@ -0,0 +1,474 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseReduction,
+          bool OutputIndex,
+          bool HaveIndexInput,
+          typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation>
+__global__ void kernel_reduce_threadwise(const InGridDesc_M_K in_grid_desc_m_k,
+                                         const OutGridDesc_M out_grid_desc_m,
+                                         const InElementwiseOperation in_elementwise_op,
+                                         const AccElementwiseOperation acc_elementwise_op,
+                                         AccDataType alpha,
+                                         const InDataType* const __restrict__ p_in_value_global,
+                                         const IndexDataType* const __restrict__ p_in_index_global,
+                                         AccDataType beta,
+                                         OutDataType* const __restrict__ p_out_value_global,
+                                         IndexDataType* const __restrict__ p_out_index_global)
+{
+    if constexpr(!OutputIndex)
+    {
+        GridwiseReduction::Run(in_grid_desc_m_k,
+                               out_grid_desc_m,
+                               in_elementwise_op,
+                               acc_elementwise_op,
+                               alpha,
+                               p_in_value_global,
+                               beta,
+                               p_out_value_global);
+    }
+    else
+    {
+        GridwiseReduction::template RunWithIndex<HaveIndexInput>(in_grid_desc_m_k,
+                                                                 out_grid_desc_m,
+                                                                 in_elementwise_op,
+                                                                 acc_elementwise_op,
+                                                                 alpha,
+                                                                 p_in_value_global,
+                                                                 p_in_index_global,
+                                                                 beta,
+                                                                 p_out_value_global,
+                                                                 p_out_index_global);
+    };
+};
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize>
+struct GridwiseReduction_mk_to_m_threadwise
+{
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)) &&
+                      (MThreadSliceSize % OutDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<InSrcVectorDim == 0, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ static void Run(const InGridDesc_M_K& in_grid_desc_m_k,
+                               const OutGridDesc_M& out_grid_desc_m,
+                               const InElementwiseOperation& in_elementwise_op,
+                               const AccElementwiseOperation& acc_elementwise_op,
+                               AccDataType alpha,
+                               const InDataType* const __restrict__ p_in_value_global,
+                               AccDataType beta,
+                               OutDataType* const __restrict__ p_out_value_global)
+    {
+        using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                     ThreadReduceSrcDesc_M_K,
+                                                     ThreadReduceDstDesc_M,
+                                                     ReduceOperation,
+                                                     PropagateNan>;
+
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
+        auto dst_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_value_global, out_grid_desc_m.GetElementSpaceSize());
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> accu_value_buf;
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) { accu_value_buf(I) = identityVal; });
+
+        const auto toReduceLength = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        index_t thread_global_1d_id = get_block_1d_id() * BlockSize + get_thread_local_1d_id();
+
+        auto threadwise_src_val_load =
+            ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                             AccDataType,
+                                             InGridDesc_M_K,
+                                             decltype(thread_buffer_desc),
+                                             ThreadBufferLengths,
+                                             ThreadBufferDimAccessOrder,
+                                             InSrcVectorDim,
+                                             InSrcVectorSize,
+                                             1,
+                                             false>(
+                in_grid_desc_m_k, make_multi_index(thread_global_1d_id * MThreadSliceSize, 0));
+
+        constexpr auto in_thread_copy_step = make_multi_index(0, KThreadSliceSize);
+
+        index_t reducedLength = 0;
+        do
+        {
+            threadwise_src_val_load.Run(in_grid_desc_m_k,
+                                        in_global_val_buf,
+                                        thread_buffer_desc,
+                                        make_tuple(I0, I0),
+                                        in_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                // do element-wise pre-reduction operation
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset = thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                    in_elementwise_op(in_thread_buf(Number<offset>{}),
+                                      in_thread_buf(Number<offset>{}));
+                });
+            });
+
+            ThreadwiseReduce::Reduce(in_thread_buf, accu_value_buf);
+
+            threadwise_src_val_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+            reducedLength += KThreadSliceSize;
+        } while(reducedLength < toReduceLength);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            acc_elementwise_op(accu_value_buf(I), accu_value_buf(I));
+
+            accu_value_buf(I) *= alpha;
+        });
+
+        constexpr auto reduced_data_desc = ThreadReduceDstDesc_M{};
+
+        if(!float_equal_zero{}(beta))
+        {
+            auto threadwise_dst_load = ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                                        OutDataType,
+                                                                        OutGridDesc_M,
+                                                                        decltype(reduced_data_desc),
+                                                                        Sequence<MThreadSliceSize>,
+                                                                        Sequence<0>,
+                                                                        0,
+                                                                        1,
+                                                                        1,
+                                                                        true>(
+                out_grid_desc_m, make_multi_index(thread_global_1d_id * MThreadSliceSize));
+
+            StaticBuffer<AddressSpaceEnum::Vgpr, OutDataType, MThreadSliceSize, true>
+                priorDstValue_buf;
+
+            threadwise_dst_load.Run(out_grid_desc_m,
+                                    dst_global_buf,
+                                    reduced_data_desc,
+                                    make_tuple(I0),
+                                    priorDstValue_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                accu_value_buf(I) += type_convert<AccDataType>(priorDstValue_buf[I]) * beta;
+            });
+        };
+
+        auto threadwise_dst_store = ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                                       OutDataType,
+                                                                       decltype(reduced_data_desc),
+                                                                       OutGridDesc_M,
+                                                                       PassThroughOp,
+                                                                       Sequence<MThreadSliceSize>,
+                                                                       Sequence<0>,
+                                                                       0,
+                                                                       OutDstVectorSize,
+                                                                       OutMemoryDataOperation,
+                                                                       1,
+                                                                       false>(
+            out_grid_desc_m,
+            make_multi_index(thread_global_1d_id * MThreadSliceSize),
+            PassThroughOp{});
+
+        threadwise_dst_store.Run(
+            reduced_data_desc, make_tuple(I0), accu_value_buf, out_grid_desc_m, dst_global_buf);
+    };
+
+    template <bool HaveIndexInput>
+    __device__ static void RunWithIndex(const InGridDesc_M_K& in_grid_desc_m_k,
+                                        const OutGridDesc_M& out_grid_desc_m,
+                                        const InElementwiseOperation& in_elementwise_op,
+                                        const AccElementwiseOperation& acc_elementwise_op,
+                                        AccDataType alpha,
+                                        const InDataType* const __restrict__ p_in_value_global,
+                                        const IndexDataType* const __restrict__ p_in_index_global,
+                                        AccDataType beta,
+                                        OutDataType* const __restrict__ p_out_value_global,
+                                        IndexDataType* const __restrict__ p_out_index_global)
+    {
+        using ThreadwiseReduceWithIndex = ThreadwiseReductionWithIndex<AccDataType,
+                                                                       IndexDataType,
+                                                                       ThreadReduceSrcDesc_M_K,
+                                                                       ThreadReduceDstDesc_M,
+                                                                       ReduceOperation,
+                                                                       PropagateNan>;
+
+        (void)acc_elementwise_op;
+
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
+        const auto in_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_index_global, in_grid_desc_m_k.GetElementSpaceSize());
+
+        auto out_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_value_global, out_grid_desc_m.GetElementSpaceSize());
+        auto out_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_index_global, out_grid_desc_m.GetElementSpaceSize());
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_val_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     IndexDataType,
+                     MThreadSliceSize * KThreadSliceSize,
+                     true>
+            in_thread_idx_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> accu_value_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, IndexDataType, MThreadSliceSize, true> accu_index_buf;
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            accu_value_buf(I) = identityVal;
+            accu_index_buf(I) = 0;
+        });
+
+        const auto toReduceLength = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        index_t thread_global_1d_id = get_block_1d_id() * BlockSize + get_thread_local_1d_id();
+
+        auto threadwise_src_val_load =
+            ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                             AccDataType,
+                                             InGridDesc_M_K,
+                                             decltype(thread_buffer_desc),
+                                             ThreadBufferLengths,
+                                             ThreadBufferDimAccessOrder,
+                                             InSrcVectorDim,
+                                             InSrcVectorSize,
+                                             1,
+                                             false>(
+                in_grid_desc_m_k, make_multi_index(thread_global_1d_id * MThreadSliceSize, 0));
+
+        constexpr auto in_thread_copy_step = make_multi_index(0, KThreadSliceSize);
+
+        index_t indexStart    = 0;
+        index_t reducedLength = 0;
+        if constexpr(HaveIndexInput)
+        {
+            auto threadwise_src_idx_load =
+                ThreadwiseTensorSliceTransfer_v2<IndexDataType,
+                                                 IndexDataType,
+                                                 InGridDesc_M_K,
+                                                 decltype(thread_buffer_desc),
+                                                 ThreadBufferLengths,
+                                                 ThreadBufferDimAccessOrder,
+                                                 InSrcVectorDim,
+                                                 InSrcVectorSize,
+                                                 1,
+                                                 false>(
+                    in_grid_desc_m_k, make_multi_index(thread_global_1d_id * MThreadSliceSize, 0));
+
+            do
+            {
+                threadwise_src_val_load.Run(in_grid_desc_m_k,
+                                            in_global_val_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_val_buf);
+
+                threadwise_src_idx_load.Run(in_grid_desc_m_k,
+                                            in_global_idx_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_idx_buf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    // do element-wise pre-reduction operation
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+
+                        in_elementwise_op(in_thread_val_buf(Number<offset>{}),
+                                          in_thread_val_buf(Number<offset>{}));
+                    });
+                });
+
+                ThreadwiseReduceWithIndex::Reduce(
+                    in_thread_val_buf, in_thread_idx_buf, accu_value_buf, accu_index_buf);
+
+                threadwise_src_val_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+                threadwise_src_idx_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+                indexStart += KThreadSliceSize;
+                reducedLength += KThreadSliceSize;
+            } while(reducedLength < toReduceLength);
+        }
+        else
+        {
+            do
+            {
+                threadwise_src_val_load.Run(in_grid_desc_m_k,
+                                            in_global_val_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_val_buf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    // do element-wise pre-reduction operation
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+
+                        in_thread_idx_buf(Number<offset>{}) = indexStart + iK();
+
+                        in_elementwise_op(in_thread_val_buf(Number<offset>{}),
+                                          in_thread_val_buf(Number<offset>{}));
+                    });
+                });
+
+                ThreadwiseReduceWithIndex::Reduce(
+                    in_thread_val_buf, in_thread_idx_buf, accu_value_buf, accu_index_buf);
+
+                threadwise_src_val_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+                indexStart += KThreadSliceSize;
+                reducedLength += KThreadSliceSize;
+            } while(reducedLength < toReduceLength);
+        };
+
+        // for indiced operation, acc_elementwise_op shoud do nothing
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            acc_elementwise_op(accu_value_buf(I), accu_value_buf(I));
+
+            accu_value_buf(I) *= alpha;
+        });
+
+        constexpr auto reduced_data_desc = ThreadReduceDstDesc_M{};
+
+        if(!float_equal_zero{}(beta))
+        {
+            auto threadwise_dst_load = ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                                        OutDataType,
+                                                                        OutGridDesc_M,
+                                                                        decltype(reduced_data_desc),
+                                                                        Sequence<MThreadSliceSize>,
+                                                                        Sequence<0>,
+                                                                        0,
+                                                                        1,
+                                                                        1,
+                                                                        false>(
+                out_grid_desc_m, make_multi_index(thread_global_1d_id * MThreadSliceSize));
+
+            StaticBuffer<AddressSpaceEnum::Vgpr, OutDataType, MThreadSliceSize, true>
+                priorDstValue_buf;
+
+            threadwise_dst_load.Run(out_grid_desc_m,
+                                    out_global_val_buf,
+                                    reduced_data_desc,
+                                    make_tuple(I0),
+                                    priorDstValue_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                accu_value_buf(I) += type_convert<AccDataType>(priorDstValue_buf[I]) * beta;
+            });
+        };
+
+        auto threadwise_dst_val_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               OutDataType,
+                                               decltype(reduced_data_desc),
+                                               OutGridDesc_M,
+                                               PassThroughOp,
+                                               Sequence<MThreadSliceSize>,
+                                               Sequence<0>,
+                                               0,
+                                               OutDstVectorSize,
+                                               OutMemoryDataOperation,
+                                               1,
+                                               false>(
+                out_grid_desc_m,
+                make_multi_index(thread_global_1d_id * MThreadSliceSize),
+                PassThroughOp{});
+
+        auto threadwise_dst_idx_store =
+            ThreadwiseTensorSliceTransfer_v1r3<IndexDataType,
+                                               IndexDataType,
+                                               decltype(reduced_data_desc),
+                                               OutGridDesc_M,
+                                               PassThroughOp,
+                                               Sequence<MThreadSliceSize>,
+                                               Sequence<0>,
+                                               0,
+                                               OutDstVectorSize,
+                                               OutMemoryDataOperation,
+                                               1,
+                                               false>(
+                out_grid_desc_m,
+                make_multi_index(thread_global_1d_id * MThreadSliceSize),
+                PassThroughOp{});
+
+        threadwise_dst_val_store.Run(
+            reduced_data_desc, make_tuple(I0), accu_value_buf, out_grid_desc_m, out_global_val_buf);
+
+        threadwise_dst_idx_store.Run(
+            reduced_data_desc, make_tuple(I0), accu_index_buf, out_grid_desc_m, out_global_idx_buf);
+    };
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000..fccb127
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp
@@ -0,0 +1,931 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t B1K1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1ThreadTransferSrcResetCoordinateAfterRun,
+          index_t B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseBatchedGemmGemm_Xdl_CShuffle
+{
+    static_assert(LoopSched == LoopScheduler::Default,
+                  "Non-default loop scheduler is currently not supported");
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    // Gemm0
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+    // Gemm1
+    static constexpr auto B1K0 = Number<Gemm1KPerBlock / B1K1Value>{};
+    static constexpr auto B1K1 = Number<B1K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, MWaves, MPerXdl>(
+            ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, 1, 1>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t Gemm1NWaves = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, Gemm1NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B1 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(B1K0, Number<Gemm1NPerBlock>{}, B1K1),
+            make_tuple(Number<Gemm1NPerBlock + B1BlockLdsExtraN>{} * B1K1, B1K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        const index_t gemm0_bytes_end = (SharedMemTrait::a_block_space_size_aligned +
+                                         SharedMemTrait::b_block_space_size_aligned) *
+                                        sizeof(FloatAB);
+        const index_t gemm1_bytes_end =
+            (SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
+            sizeof(FloatAB);
+        const index_t c_block_bytes_end =
+            SharedMemTrait::c_block_space_size * sizeof(FloatCShuffle);
+
+        return math::max(gemm0_bytes_end, gemm1_bytes_end, c_block_bytes_end);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+        const auto Gemm1N = b1_grid_desc_bk0_n_bk1.GetLength(I1);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && Gemm1N == c_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0 &&
+             Gemm1N % Gemm1NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gemm0 gridwise gemm pipeline
+        const auto num_gemm0_k_loop = K / KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm0_k_loop))
+        {
+            return false;
+        }
+
+        // check gemm1 gridwise gemm pipeline
+        if(!(NPerBlock % Gemm1KPerBlock == 0))
+        {
+            return false;
+        }
+
+        const auto num_gemm1_k_inner_loop = NPerBlock / Gemm1KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm1_k_inner_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / Gemm1NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<Gemm1NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, Gemm1NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    struct SharedMemTrait
+    {
+        // LDS allocation for A and B: be careful of alignment
+        static constexpr auto a_block_desc_ak0_m_ak1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        static constexpr auto b_block_desc_bk0_n_bk1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        static constexpr auto b1_block_desc_bk0_n_bk1 =
+            GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        static constexpr auto max_lds_align = math::lcm(math::lcm(AK1, BK1), B1K1);
+
+        static constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        static constexpr auto a_block_space_offset  = 0;
+        static constexpr auto b_block_space_offset  = a_block_space_size_aligned.value;
+        static constexpr auto b1_block_space_offset = 0;
+
+        // LDS allocation for C shuffle in LDS
+        static constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+        static constexpr auto c_block_space_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+    };
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               const FloatAB* __restrict__ p_b1_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const AccElementwiseOperation& acc_element_op,
+                               const B1ElementwiseOperation& b1_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                               const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        //
+        // set up Gemm0
+        //
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0), // will loop over GemmN dimension
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // Fused Gemm+Gemm pipeline
+        // for n in N0:
+        //   for k in K0:
+        //     acc[m][n] += A[m][k] * B0[k][n]
+        //   acc1[m][o] += acc[m][n] * B1[n][o]
+
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
+            decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            NPerBlock,
+            KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            true>{}; // TransposeC
+
+        auto acc_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::a_block_space_offset,
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b_block_space_offset,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+        const auto a_block_reset_copy_step =
+            make_multi_index(-a_grid_desc_ak0_m_ak1.GetLength(I0), 0, 0);
+        const auto b_block_reset_copy_step =
+            make_multi_index(-b_grid_desc_bk0_n_bk1.GetLength(I0), NPerBlock, 0);
+
+        // gridwise GEMM pipeline
+        // Only supports LoopScheduler::Default
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_Selector<PipelineVer,
+                                                                          NumGemmKPrefetchStage,
+                                                                          LoopScheduler::Default>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        //
+        // set up Gemm1
+        //
+
+        // Acc matrix threadwise copy: AccVGPR to VGPR and downcast to XDL input data type
+        constexpr auto acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        constexpr auto m0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+        constexpr auto n0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        constexpr auto m1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+        constexpr auto n1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        constexpr auto m2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+        constexpr auto n2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        constexpr auto n3 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        constexpr auto n4 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        constexpr auto b1_block_slice_copy_step = make_multi_index(Gemm1KPerBlock / B1K1, 0, 0);
+
+        // acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to acc_thread_desc_k0_m_k1
+        // n0_n1_n2_n3 -> k0
+        // m0_m1_m2 -> m
+        // n4 -> k1
+        // NOTE: had to use merge_v3 or will spit out compilation errors
+        constexpr auto acc_thread_desc_k0_m_k1 = transform_tensor_descriptor(
+            acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
+                       make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
+                       make_pass_through_transform(n4)),
+            make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // A1 matrix in AccVGPR
+        // N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
+        constexpr auto AccN3 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I6);
+
+        constexpr auto A1ThreadSlice_K0_M_K1 =
+            make_tuple(Number<Gemm1KPerBlock / n4 / AccN3>{}, Number<m0 * m1 * m2>{}, Number<n4>{});
+
+        constexpr auto A1ThreadSliceK0        = A1ThreadSlice_K0_M_K1[I0];
+        constexpr auto A1ThreadSliceM         = A1ThreadSlice_K0_M_K1[I1];
+        constexpr auto A1ThreadSliceK1        = A1ThreadSlice_K0_M_K1[I2];
+        constexpr auto a1_thread_desc_k0_m_k1 = make_naive_tensor_descriptor(
+            A1ThreadSlice_K0_M_K1,
+            make_tuple(A1ThreadSliceM * A1ThreadSliceK1, A1ThreadSliceK1, I1));
+
+        // B1 matrix in LDS memory, dst of blockwise copy
+        constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A1 matrix blockwise copy
+        auto a1_blockwise_copy = ThreadwiseTensorSliceTransfer_StaticToStatic<
+            FloatGemmAcc,
+            FloatAB,
+            decltype(acc_thread_desc_k0_m_k1),
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(acc_element_op),
+            Sequence<A1ThreadSliceK0, A1ThreadSliceM, A1ThreadSliceK1>,
+            Sequence<1, 0, 2>,
+            2,
+            n4>{acc_element_op};
+
+        // B1 matrix blockwise copy
+        auto b1_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<B1K0, Gemm1NPerBlock, B1K1>,
+                                                B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                B1BlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b1_grid_desc_bk0_n_bk1),
+                                                decltype(b1_block_desc_bk0_n_bk1),
+                                                B1BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                B1BlockTransferSrcVectorDim,
+                                                2,
+                                                B1BlockTransferSrcScalarPerVector,
+                                                B1BlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                B1ThreadTransferSrcResetCoordinateAfterRun,
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b1_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b1_element_op,
+                b1_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a1_thread_desc_k0_m_k1.GetElementSpaceSize());
+
+        // reuse LDS space for gemm0's b_block_buf
+        auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b1_block_space_offset,
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        // selected_mfma.group_size or B1K1 <= Gemm1KPack <= selected_mfma.group_size
+        // selected_mfma.k_per_blk <= Gemm1KPack
+        //
+        // Following similar rationale behind Gemm0KPack, let Gemm1KPack be the lowest common
+        // multiples of A1K1 (predetermined by selected_mfma.group_size) and B1K1. But in this case
+        // Gemm1KPack can't be higher than A1K1 itself because A1 matrix is distributed in VGPRs
+        // with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
+        // cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
+        // therefore we may just as well assign Gemm1KPack = group_size
+        constexpr index_t Gemm1KPack =
+            MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
+
+        auto gemm1_blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(b1_block_desc_bk0_n_bk1),
+            decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a1_thread_desc_k0_m_k1)),
+            decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b1_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            Gemm1NPerBlock,
+            Gemm1KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            Gemm1NXdlPerWave,
+            Gemm1KPack,
+            false,      // TransposeC
+            Gemm1KPack, // AMmaKStride
+            Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
+            // BMmaKStride
+            make_tuple(0, 0, 0, 0)}; // A_origin
+
+        auto c_thread_buf = gemm1_blockwise_gemm.GetCThreadBuffer();
+
+        const index_t num_gemm1_k_block_outer_loop =
+            b_grid_desc_bk0_n_bk1.GetLength(I1) / NPerBlock;
+        constexpr index_t num_gemm1_k_block_inner_loop = NPerBlock / Gemm1KPerBlock;
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // gemm1 K loop
+        index_t gemm1_k_block_outer_index = 0;
+        do
+        {
+            // gemm0
+            gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                                   a_block_desc_ak0_m_ak1,
+                                                                   a_blockwise_copy,
+                                                                   a_grid_buf,
+                                                                   a_block_buf,
+                                                                   a_block_slice_copy_step,
+                                                                   b_grid_desc_bk0_n_bk1,
+                                                                   b_block_desc_bk0_n_bk1,
+                                                                   b_blockwise_copy,
+                                                                   b_grid_buf,
+                                                                   b_block_buf,
+                                                                   b_block_slice_copy_step,
+                                                                   blockwise_gemm,
+                                                                   acc_thread_buf,
+                                                                   num_k_block_main_loop);
+            // gemm1
+            {
+                // TODO: explore using dynamic buffer for a1 thread buffer
+                // For a1_blockwise_copy, the goal is to satisfy pipeline requirements RunRead(),
+                // RunWrite(), and MoveSliceWindow(). But it is impossible to implement given that
+                // the A1 source buffer is static buffer holding the output of first GEMM and
+                // requires constexpr offset by design. Therefore, we pass tensor coordinate offset
+                // explicitly in Run() below.
+
+                // preload data into LDS
+                b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                     b1_block_slice_copy_step);
+
+                block_sync_lds(); // wait for gemm0 LDS read
+
+                b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+
+                // main body
+                if constexpr(num_gemm1_k_block_inner_loop > 1)
+                {
+                    static_for<0, num_gemm1_k_block_inner_loop - 1, 1>{}([&](auto i) {
+                        a1_blockwise_copy.Run(acc_thread_desc_k0_m_k1,
+                                              make_tuple(Number<i * A1ThreadSliceK0>{}, I0, I0),
+                                              acc_thread_buf,
+                                              a1_thread_desc_k0_m_k1,
+                                              make_tuple(I0, I0, I0),
+                                              a1_thread_buf);
+
+                        b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                        block_sync_lds();
+
+                        gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, c_thread_buf);
+
+                        block_sync_lds();
+
+                        b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                             b1_block_slice_copy_step);
+
+                        b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+                    });
+                }
+                // tail
+                {
+                    a1_blockwise_copy.Run(
+                        acc_thread_desc_k0_m_k1,
+                        make_tuple(
+                            Number<(num_gemm1_k_block_inner_loop - 1) * A1ThreadSliceK0>{}, I0, I0),
+                        acc_thread_buf,
+                        a1_thread_desc_k0_m_k1,
+                        make_tuple(I0, I0, I0),
+                        a1_thread_buf);
+
+                    block_sync_lds();
+
+                    gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, c_thread_buf);
+                }
+            } // end gemm1
+
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_ak0_m_ak1,
+                                                a_block_reset_copy_step); // rewind K
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_bk0_n_bk1,
+                                                b_block_reset_copy_step); // rewind K and step N
+
+            block_sync_lds(); // wait for gemm1 LDS read
+        } while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              Gemm1NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                gemm1_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                gemm1_blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, Gemm1NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, Gemm1NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000..b9f4a30
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp
@@ -0,0 +1,1268 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename A0B0B1DataType, // FIXME: don't assume A0/B0/B1 have same datatype
+          typename Acc0DataType,
+          typename D0sDataType,
+          typename Acc1DataType,
+          typename C1ShuffleDataType,
+          typename D1sDataType,
+          typename E1DataType,
+          typename A0ElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename CDE0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CDE1ElementwiseOperation,
+          InMemoryDataOperationEnum E1GlobalMemoryDataOperation,
+          typename A0GridDesc_M_K,
+          typename B0GridDesc_N_K,
+          typename D0sGridDesc_M_N,
+          typename B1GridDesc_N_K,
+          typename D1sGridDesc_M_N,
+          typename E1GridDesc_M_N,
+          index_t NumGemm0KPrefetchStage,
+          index_t BlockSize,
+          index_t Gemm0MPerBlock,
+          index_t Gemm0NPerBlock,
+          index_t Gemm0KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t A0K1Value,
+          index_t B0K1Value,
+          index_t B1K1Value,
+          index_t Gemm0MPerXdl,
+          index_t Gemm0NPerXdl,
+          index_t Gemm0MXdlPerWave,
+          index_t Gemm0NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename A0BlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename A0BlockTransferThreadClusterArrangeOrder,
+          typename A0BlockTransferSrcAccessOrder,
+          index_t A0BlockTransferSrcVectorDim,
+          index_t A0BlockTransferSrcScalarPerVector,
+          index_t A0BlockTransferDstScalarPerVector_AK1,
+          bool A0ThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t A0BlockLdsExtraM,
+          typename B0BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B0BlockTransferThreadClusterArrangeOrder,
+          typename B0BlockTransferSrcAccessOrder,
+          index_t B0BlockTransferSrcVectorDim,
+          index_t B0BlockTransferSrcScalarPerVector,
+          index_t B0BlockTransferDstScalarPerVector_BK1,
+          bool B0ThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t B0BlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1ThreadTransferSrcResetCoordinateAfterRun,
+          index_t B1BlockLdsExtraN,
+          index_t C1ShuffleGemm0MXdlPerWavePerShuffle,
+          index_t C1ShuffleGemm0NXdlPerWavePerShuffle,
+          typename CDE1ShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDE1ShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched>
+struct GridwiseBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
+{
+    static_assert(LoopSched == LoopScheduler::Default,
+                  "Non-default loop scheduler is currently not supported");
+
+    static constexpr index_t NumD0Tensor = D0sDataType::Size();
+    static constexpr index_t NumD1Tensor = D1sDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    static constexpr auto WaveSize = 64;
+    // K1 should be Number<...>
+    // Gemm0
+    static constexpr auto A0K1 = Number<A0K1Value>{};
+    static constexpr auto B0K1 = Number<B0K1Value>{};
+
+    static constexpr auto A0K0PerBlock = Number<Gemm0KPerBlock / A0K1Value>{};
+    static constexpr auto B0K0PerBlock = Number<Gemm0KPerBlock / B0K1Value>{};
+
+    static constexpr auto Gemm0MWaves = Gemm0MPerBlock / (Gemm0MPerXdl * Gemm0MXdlPerWave);
+    static constexpr auto Gemm0NWaves = Gemm0NPerBlock / (Gemm0NPerXdl * Gemm0NXdlPerWave);
+    // Gemm1
+    static constexpr auto B1K1         = Number<B1K1Value>{};
+    static constexpr auto B1K0PerBlock = Number<Gemm1KPerBlock / B1K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemm0KPrefetchStage>;
+
+    // ck::Tuple<const D0DataType1*, const D0DataType2*, ...>
+    static constexpr auto MakeD0sGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+
+                return static_cast<const D0DataType*>(nullptr);
+            },
+            Number<NumD0Tensor>{});
+    }
+
+    // ck::Tuple<const D1DataType1*, const D1DataType2*, ...>
+    static constexpr auto MakeD1sGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using D1DataType = remove_cvref_t<tuple_element_t<i.value, D1sDataType>>;
+
+                return static_cast<const D1DataType*>(nullptr);
+            },
+            Number<NumD1Tensor>{});
+    }
+
+    __device__ static auto GetGemm0WaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(Gemm0MWaves, Gemm0NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto GetGemm0WaveMNIdx(const index_t thread_id)
+    {
+        constexpr auto wave_threadid_to_mn_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(WaveSize / Gemm0NPerXdl, Gemm0NPerXdl))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        return wave_threadid_to_mn_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    template <typename A0BlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const A0BlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = Gemm0MPerBlock / (Gemm0MXdlPerWave * Gemm0MPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm0MXdlPerWave, MWaves, Gemm0MPerXdl>(
+            A0BlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t NWaves = Gemm0NPerBlock / (Gemm0NXdlPerWave * Gemm0NPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm0NXdlPerWave, NWaves, Gemm0NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename A0BlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const A0BlockDesc_AK0_M_AK1&)
+    {
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm0MXdlPerWave, 1, 1>(
+            A0BlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t Gemm1NWaves = Gemm1NPerBlock / (Gemm1NXdlPerWave * Gemm0NPerXdl);
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, Gemm1NWaves, Gemm0NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    __host__ __device__ static constexpr auto GetA0BlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A0 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(A0K0PerBlock, Number<Gemm0MPerBlock>{}, A0K1),
+            make_tuple(Number<Gemm0MPerBlock + A0BlockLdsExtraM>{} * A0K1, A0K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetB0BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B0 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(B0K0PerBlock, Number<Gemm0NPerBlock>{}, B0K1),
+            make_tuple(Number<Gemm0NPerBlock + B0BlockLdsExtraN>{} * B0K1, B0K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B1 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(B1K0PerBlock, Number<Gemm1NPerBlock>{}, B1K1),
+            make_tuple(Number<Gemm1NPerBlock + B1BlockLdsExtraN>{} * B1K1, B1K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetC1ShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = Gemm0MPerBlock / (Gemm0MXdlPerWave * Gemm0MPerXdl);
+        constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * Gemm0NPerXdl);
+
+        constexpr auto c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<C1ShuffleGemm0MXdlPerWavePerShuffle * MWave * Gemm0MPerXdl>{},
+                           I1,
+                           Number<C1ShuffleGemm0NXdlPerWavePerShuffle * NWave * Gemm0NPerXdl>{}));
+
+        return c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        const index_t gemm0_bytes_end = (SharedMemTrait::a0_block_space_size_aligned +
+                                         SharedMemTrait::b0_block_space_size_aligned) *
+                                        sizeof(A0B0B1DataType);
+        const index_t gemm1_bytes_end =
+            (SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
+            sizeof(A0B0B1DataType);
+        const index_t c1_block_bytes_end =
+            SharedMemTrait::c1_block_space_size * sizeof(C1ShuffleDataType);
+
+        return math::max(gemm0_bytes_end, gemm1_bytes_end, c1_block_bytes_end);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2E1TileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const A0GridDesc_M_K& a0_grid_desc_m_k,
+                  const B0GridDesc_N_K& b0_grid_desc_n_k,
+                  const B1GridDesc_N_K& b1_grid_desc_n_k,
+                  const E1GridDesc_M_N& e1_grid_desc_m_n,
+                  const Block2E1TileMap& block_2_e1tile_map)
+    {
+        static_assert((Gemm0MPerBlock % (Gemm0MPerXdl * Gemm0MXdlPerWave) == 0) &&
+                          (Gemm0NPerBlock % (Gemm0NXdlPerWave * Gemm0NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M      = a0_grid_desc_m_k.GetLength(I0);
+        const auto N      = b0_grid_desc_n_k.GetLength(I0);
+        const auto K      = a0_grid_desc_m_k.GetLength(I1);
+        const auto Gemm1N = b1_grid_desc_n_k.GetLength(I0);
+
+        if(!(M == e1_grid_desc_m_n.GetLength(I0) && Gemm1N == e1_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        if(!(M % Gemm0MPerBlock == 0 && N % Gemm0NPerBlock == 0 && K % Gemm0KPerBlock == 0 &&
+             Gemm1N % Gemm1NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gemm0 gridwise gemm pipeline
+        const auto num_gemm0_k_loop = K / Gemm0KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm0_k_loop))
+        {
+            return false;
+        }
+
+        // check gemm1 gridwise gemm pipeline
+        if(!(Gemm0NPerBlock % Gemm1KPerBlock == 0))
+        {
+            return false;
+        }
+
+        const auto num_gemm1_k_inner_loop = Gemm0NPerBlock / Gemm1KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm1_k_inner_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_e1tile_map.CheckValidity(e1_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / Gemm0KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    // A0 desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultA0GridDescriptor_AK0_M_AK1(const A0GridDesc_M_K& a0_grid_desc_m_k)
+    {
+        const auto M = a0_grid_desc_m_k.GetLength(I0);
+        const auto K = a0_grid_desc_m_k.GetLength(I1);
+
+        const auto A0K0 = K / A0K1;
+
+        return transform_tensor_descriptor(
+            a0_grid_desc_m_k,
+            make_tuple(make_unmerge_transform(make_tuple(A0K0, A0K1)),
+                       make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B0 desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultB0GridDescriptor_BK0_N_BK1(const B0GridDesc_N_K& b0_grid_desc_n_k)
+    {
+        const auto N = b0_grid_desc_n_k.GetLength(I0);
+        const auto K = b0_grid_desc_n_k.GetLength(I1);
+
+        const auto B0K0 = K / B0K1;
+
+        return transform_tensor_descriptor(
+            b0_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(B0K0, B0K1)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // D0 desc for source in blockwise copy
+    template <typename D0GridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0GridDesc_M_N& d0_grid_desc_m_n)
+    {
+        const auto M = d0_grid_desc_m_n.GetLength(I0);
+        const auto N = d0_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto mfma =
+            MfmaSelector<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl>::selected_mfma;
+        constexpr auto N3 = mfma.num_groups_per_blk;
+        constexpr auto N5 = mfma.group_size;
+        return transform_tensor_descriptor(
+            d0_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(
+                           M / Gemm0MPerBlock, Gemm0MXdlPerWave, Gemm0MWaves, Gemm0MPerXdl)),
+                       make_unmerge_transform(make_tuple(N / Gemm0NPerBlock,
+                                                         Gemm0NXdlPerWave,
+                                                         Gemm0NWaves,
+                                                         N3,
+                                                         WaveSize / Gemm0NPerXdl,
+                                                         N5))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
+    }
+
+    // B1 desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultB1GridDescriptor_BK0_N_BK1(const B1GridDesc_N_K& b1_grid_desc_n_k)
+    {
+        const auto N = b1_grid_desc_n_k.GetLength(I0);
+        const auto K = b1_grid_desc_n_k.GetLength(I1);
+
+        const auto B1K0 = K / B1K1;
+
+        return transform_tensor_descriptor(
+            b1_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // C1 desc for destination in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeE1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const E1GridDesc_M_N& e1_grid_desc_m_n)
+    {
+        const auto M = e1_grid_desc_m_n.GetLength(I0);
+        const auto N = e1_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / Gemm0MPerBlock;
+        const auto NBlock = N / Gemm1NPerBlock;
+
+        const auto e1_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e1_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<Gemm0MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<Gemm1NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e1_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+    // D0s desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0sGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ds_grid_desc_m_n[i]);
+            },
+            Number<NumD0Tensor>{});
+    }
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto
+    MakeD1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const DsGridDescriptor_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeE1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumD1Tensor>{});
+    }
+
+    // return block_id to C1 matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2E1TileMap(const E1GridDesc_M_N& e1_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<Gemm0MPerBlock, Gemm1NPerBlock, E1GridDesc_M_N>(
+            e1_grid_desc_m_n);
+    }
+
+    using E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeE1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(E1GridDesc_M_N{}))>;
+
+    using D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype(
+        MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(D0sGridDesc_M_N{}))>;
+
+    using D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeD1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(D1sGridDesc_M_N{}))>;
+
+    using DefaultBlock2E1TileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2E1TileMap(E1GridDesc_M_N{}))>;
+
+    struct SharedMemTrait
+    {
+        // LDS allocation for A0 and B0: be careful of alignment
+        static constexpr auto a0_block_desc_ak0_m_ak1 =
+            GetA0BlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        static constexpr auto b0_block_desc_bk0_n_bk1 =
+            GetB0BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        static constexpr auto b1_block_desc_bk0_n_bk1 =
+            GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        static constexpr auto max_lds_align = math::lcm(math::lcm(A0K1, B0K1), B1K1);
+
+        static constexpr auto a0_block_space_size_aligned = math::integer_least_multiple(
+            a0_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b0_block_space_size_aligned = math::integer_least_multiple(
+            b0_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        static constexpr auto a0_block_space_offset = 0;
+        static constexpr auto b0_block_space_offset = a0_block_space_size_aligned.value;
+        static constexpr auto b1_block_space_offset = 0;
+
+        // LDS allocation for C1 shuffle in LDS
+        static constexpr auto c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetC1ShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+        static constexpr auto c1_block_space_size =
+            c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+    };
+
+    using D0sGridPointer = decltype(MakeD0sGridPointer());
+    using D1sGridPointer = decltype(MakeD1sGridPointer());
+
+    template <bool HasMainKBlockLoop,
+              typename A0GridDesc_AK0_M_AK1,
+              typename B0GridDesc_BK0_N_BK1,
+              typename B1GridDesc_BK0_N_BK1,
+              typename Block2E1TileMap>
+    __device__ static void Run(const A0B0B1DataType* __restrict__ p_a0_grid,
+                               const A0B0B1DataType* __restrict__ p_b0_grid,
+                               D0sGridPointer p_d0s_grid,
+                               const A0B0B1DataType* __restrict__ p_b1_grid,
+                               D1sGridPointer p_d1s_grid,
+                               E1DataType* __restrict__ p_e1_grid,
+                               void* __restrict__ p_shared,
+                               const A0ElementwiseOperation& a0_element_op,
+                               const B0ElementwiseOperation& b0_element_op,
+                               const CDE0ElementwiseOperation& cde0_element_op,
+                               const B1ElementwiseOperation& b1_element_op,
+                               const CDE1ElementwiseOperation& cde1_element_op,
+                               const A0GridDesc_AK0_M_AK1& a0_grid_desc_ak0_m_ak1,
+                               const B0GridDesc_BK0_N_BK1& b0_grid_desc_bk0_n_bk1,
+                               const D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5&
+                                   d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                               const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                               const D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   d1s_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e1_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2E1TileMap& block_2_e1tile_map)
+    {
+        const auto a0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a0_grid, a0_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b0_grid, b0_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto e1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e1_grid, e1_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        const auto d0s_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_d0s_grid[i],
+                    d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i].GetElementSpaceSize());
+            },
+            Number<NumD0Tensor>{});
+        const auto d1s_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_d1s_grid[i],
+                    d1s_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumD1Tensor>{});
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_e1tile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_e1tile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(e1_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e1_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * Gemm0MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
+
+        // A0 matrix in LDS memory, dst of blockwise copy
+        constexpr auto a0_block_desc_ak0_m_ak1 = GetA0BlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B0 matrix in LDS memory, dst of blockwise copy
+        constexpr auto b0_block_desc_bk0_n_bk1 = GetB0BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        //
+        // set up Gemm0
+        //
+
+        // A0 matrix blockwise copy
+        auto a0_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                A0ElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<A0K0PerBlock, Gemm0MPerBlock, A0K1>,
+                                                A0BlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                A0BlockTransferThreadClusterArrangeOrder,
+                                                A0B0B1DataType,
+                                                A0B0B1DataType,
+                                                decltype(a0_grid_desc_ak0_m_ak1),
+                                                decltype(a0_block_desc_ak0_m_ak1),
+                                                A0BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                A0BlockTransferSrcVectorDim,
+                                                2,
+                                                A0BlockTransferSrcScalarPerVector,
+                                                A0BlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemm0KPrefetchStage>(
+                a0_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a0_element_op,
+                a0_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // B0 matrix blockwise copy
+        auto b0_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                B0ElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<B0K0PerBlock, Gemm0NPerBlock, B0K1>,
+                                                B0BlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                B0BlockTransferThreadClusterArrangeOrder,
+                                                A0B0B1DataType,
+                                                A0B0B1DataType,
+                                                decltype(b0_grid_desc_bk0_n_bk1),
+                                                decltype(b0_block_desc_bk0_n_bk1),
+                                                B0BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                B0BlockTransferSrcVectorDim,
+                                                2,
+                                                B0BlockTransferSrcScalarPerVector,
+                                                B0BlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemm0KPrefetchStage>(
+                b0_grid_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0), // will loop over GemmN dimension
+                b0_element_op,
+                b0_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // Fused Gemm+Gemm pipeline
+        // for n in N0:
+        //   for k in K0:
+        //     acc[m][n] += A[m][k] * B0[k][n]
+        //   acc1[m][o] += acc[m][n] * B1[n][o]
+
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(A0K1, B0K1),
+            MfmaSelector<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm0 = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            A0B0B1DataType,
+            Acc0DataType,
+            decltype(a0_block_desc_ak0_m_ak1),
+            decltype(b0_block_desc_bk0_n_bk1),
+            decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a0_block_desc_ak0_m_ak1)),
+            decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b0_block_desc_bk0_n_bk1)),
+            Gemm0MPerBlock,
+            Gemm0NPerBlock,
+            Gemm0KPerBlock,
+            Gemm0MPerXdl,
+            Gemm0NPerXdl,
+            Gemm0MXdlPerWave,
+            Gemm0NXdlPerWave,
+            KPack,
+            true>{}; // TransposeC
+
+        auto acc0_thread_buf = blockwise_gemm0.GetCThreadBuffer();
+
+        // LDS allocation for A0 and B0: be careful of alignment
+        auto a0_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<A0B0B1DataType*>(p_shared) + SharedMemTrait::a0_block_space_offset,
+            a0_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b0_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<A0B0B1DataType*>(p_shared) + SharedMemTrait::b0_block_space_offset,
+            b0_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a0_block_slice_copy_step = make_multi_index(Gemm0KPerBlock / A0K1, 0, 0);
+        constexpr auto b0_block_slice_copy_step = make_multi_index(Gemm0KPerBlock / B0K1, 0, 0);
+        const auto a0_block_reset_copy_step =
+            make_multi_index(-a0_grid_desc_ak0_m_ak1.GetLength(I0), 0, 0);
+        const auto b0_block_reset_copy_step =
+            make_multi_index(-b0_grid_desc_bk0_n_bk1.GetLength(I0), Gemm0NPerBlock, 0);
+
+        // gridwise GEMM pipeline
+        // Only supports LoopScheduler::Default
+        const auto gridwise_gemm0_pipeline =
+            GridwiseGemmPipeline_v1_Selector<NumGemm0KPrefetchStage, LoopScheduler::Default>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a0_grid_desc_ak0_m_ak1.GetLength(I0) * a0_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            Gemm0KPerBlock);
+
+        //
+        // set up Gemm1
+        //
+
+        // Acc0 matrix threadwise copy: AccVGPR to VGPR and downcast to XDL input data type
+        constexpr auto acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm0.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        constexpr auto m0 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+        constexpr auto n0 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        constexpr auto m1 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+        constexpr auto n1 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        constexpr auto m2 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+        constexpr auto n2 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        constexpr auto n3 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        constexpr auto n4 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        constexpr auto b1_block_slice_copy_step = make_multi_index(Gemm1KPerBlock / B1K1, 0, 0);
+
+        // d0 matrix threadwise copy
+        constexpr auto d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,   // MBlockId
+                                                           I1,   // NBlockID
+                                                           I1,   // MRepeat
+                                                           I1,   // NRepeat
+                                                           I1,   // MWaveId
+                                                           I1,   // NWaveId
+                                                           I1,   // MPerXdl
+                                                           I1,   // NGroupNum
+                                                           I1,   // NInputNum
+                                                           n4)); // registerNum
+
+        auto d0s_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<
+                    AddressSpaceEnum::Vgpr,
+                    A0B0B1DataType,
+                    d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize(),
+                    true>{};
+            },
+            Number<NumD0Tensor>{});
+
+        const auto wave_id     = GetGemm0WaveIdx();
+        const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
+
+        constexpr auto acc0_thread_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<Gemm0MXdlPerWave>{}, Number<Gemm0NXdlPerWave>{}, n2, n4));
+
+        auto d0s_threadwise_copy = generate_tuple(
+            [&](auto i) {
+                return ThreadwiseTensorSliceTransfer_v2<
+                    A0B0B1DataType,
+                    A0B0B1DataType,
+                    decltype(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i]),
+                    decltype(d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
+                    Sequence<I1, I1, I1, I1, I1, I1, I1, I1, I1, n4>,
+                    Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
+                    9,
+                    n4,
+                    1,
+                    false>(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                           make_multi_index(block_work_idx[I0], // MBlockId
+                                            0,                  // NBlockId
+                                            0,                  // mrepeat
+                                            0,                  // nrepeat
+                                            wave_id[I0],        // MWaveId
+                                            wave_id[I1],        // NWaveId
+                                            wave_m_n_id[I1],    // MPerXdl
+                                            0,                  // group
+                                            wave_m_n_id[I0],    // NInputIndex
+                                            0));                // register number
+            },
+            Number<NumD0Tensor>{});
+        // acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to acc0_thread_desc_k0_m_k1
+        // n0_n1_n2_n3 -> k0
+        // m0_m1_m2 -> m
+        // n4 -> k1
+        // NOTE: had to use merge_v3 or will spit out compilation errors
+        constexpr auto acc0_thread_desc_k0_m_k1 = transform_tensor_descriptor(
+            acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
+                       make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
+                       make_pass_through_transform(n4)),
+            make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // A1 matrix in AccVGPR
+        // N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
+        constexpr auto Acc0N3 =
+            blockwise_gemm0.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I6);
+
+        constexpr auto A1ThreadSlice_K0_M_K1 = make_tuple(
+            Number<Gemm1KPerBlock / n4 / Acc0N3>{}, Number<m0 * m1 * m2>{}, Number<n4>{});
+
+        constexpr auto A1ThreadSliceK0        = A1ThreadSlice_K0_M_K1[I0];
+        constexpr auto A1ThreadSliceM         = A1ThreadSlice_K0_M_K1[I1];
+        constexpr auto A1ThreadSliceK1        = A1ThreadSlice_K0_M_K1[I2];
+        constexpr auto a1_thread_desc_k0_m_k1 = make_naive_tensor_descriptor(
+            A1ThreadSlice_K0_M_K1,
+            make_tuple(A1ThreadSliceM * A1ThreadSliceK1, A1ThreadSliceK1, I1));
+
+        // B1 matrix in LDS memory, dst of blockwise copy
+        constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A1 matrix blockwise copy
+        auto a1_blockwise_copy = ThreadwiseTensorSliceTransfer_StaticToStatic<
+            Acc0DataType,
+            A0B0B1DataType,
+            decltype(acc0_thread_desc_k0_m_k1),
+            decltype(a1_thread_desc_k0_m_k1),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<A1ThreadSliceK0, A1ThreadSliceM, A1ThreadSliceK1>,
+            Sequence<1, 0, 2>,
+            2,
+            n4>{tensor_operation::element_wise::PassThrough{}};
+
+        // B1 matrix blockwise copy
+        auto b1_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                B0ElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<B1K0PerBlock, Gemm1NPerBlock, B1K1>,
+                                                B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                B1BlockTransferThreadClusterArrangeOrder,
+                                                A0B0B1DataType,
+                                                A0B0B1DataType,
+                                                decltype(b1_grid_desc_bk0_n_bk1),
+                                                decltype(b1_block_desc_bk0_n_bk1),
+                                                B1BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                B1BlockTransferSrcVectorDim,
+                                                2,
+                                                B1BlockTransferSrcScalarPerVector,
+                                                B1BlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                B1ThreadTransferSrcResetCoordinateAfterRun,
+                                                true, // DstResetCoord
+                                                1>(b1_grid_desc_bk0_n_bk1,
+                                                   make_multi_index(0, n_block_data_idx_on_grid, 0),
+                                                   b1_element_op,
+                                                   b1_block_desc_bk0_n_bk1,
+                                                   make_multi_index(0, 0, 0),
+                                                   tensor_operation::element_wise::PassThrough{});
+
+        auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, A0B0B1DataType>(
+            a1_thread_desc_k0_m_k1.GetElementSpaceSize());
+
+        // reuse LDS space for gemm0's b0_block_buf
+        auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<A0B0B1DataType*>(p_shared) + SharedMemTrait::b1_block_space_offset,
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr index_t Gemm1KPack = math::max(
+            math::lcm(
+                MfmaSelector<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl>::selected_mfma.group_size,
+                B1K1),
+            MfmaSelector<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm1 = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            A0B0B1DataType,
+            Acc1DataType,
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(b1_block_desc_bk0_n_bk1),
+            decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a1_thread_desc_k0_m_k1)),
+            decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b1_block_desc_bk0_n_bk1)),
+            Gemm0MPerBlock,
+            Gemm1NPerBlock,
+            Gemm1KPerBlock,
+            Gemm0MPerXdl,
+            Gemm0NPerXdl,
+            Gemm0MXdlPerWave,
+            Gemm1NXdlPerWave,
+            Gemm1KPack,
+            false,      // TransposeC
+            Gemm1KPack, // AMmaKStride
+            Gemm1KPack * XdlopsGemm<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl, Gemm1KPack, false>{}
+                             .K0PerXdlops>{                         // BMmaKStride
+                                           make_tuple(0, 0, 0, 0)}; // A_origin
+
+        auto c1_thread_buf = blockwise_gemm1.GetCThreadBuffer();
+
+        const index_t num_gemm1_k_block_outer_loop =
+            b0_grid_desc_bk0_n_bk1.GetLength(I1) / Gemm0NPerBlock;
+        constexpr index_t num_gemm1_k_block_inner_loop = Gemm0NPerBlock / Gemm1KPerBlock;
+
+        // Initialize C1
+        c1_thread_buf.Clear();
+
+        // gemm1 K loop
+        index_t gemm1_k_block_outer_index = 0;
+        do
+        {
+            // gemm0
+            gridwise_gemm0_pipeline.template Run<HasMainKBlockLoop>(a0_grid_desc_ak0_m_ak1,
+                                                                    a0_block_desc_ak0_m_ak1,
+                                                                    a0_blockwise_copy,
+                                                                    a0_grid_buf,
+                                                                    a0_block_buf,
+                                                                    a0_block_slice_copy_step,
+                                                                    b0_grid_desc_bk0_n_bk1,
+                                                                    b0_block_desc_bk0_n_bk1,
+                                                                    b0_blockwise_copy,
+                                                                    b0_grid_buf,
+                                                                    b0_block_buf,
+                                                                    b0_block_slice_copy_step,
+                                                                    blockwise_gemm0,
+                                                                    acc0_thread_buf,
+                                                                    num_k_block_main_loop);
+            // bias+gelu
+            {
+                static_for<0, Gemm0MXdlPerWave, 1>{}([&](auto mr) {
+                    static_for<0, Gemm0NXdlPerWave, 1>{}([&](auto nr) {
+                        static_for<0, n2, 1>{}([&](auto groupid) {
+                            static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                                d0s_threadwise_copy(i).Run(
+                                    d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                                    d0s_grid_buf[i],
+                                    d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                    make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                    d0s_thread_buf(i));
+                            });
+
+                            static_for<0, n4, 1>{}([&](auto i) {
+                                constexpr index_t c_offset = acc0_thread_desc.CalculateOffset(
+                                    make_tuple(mr, nr, groupid, i));
+
+                                // get reference to src data
+                                const auto src_data_refs = generate_tie(
+                                    // return type should be lvalue
+                                    [&](auto iSrc) -> const auto& {
+                                        return d0s_thread_buf[iSrc][i];
+                                    },
+                                    Number<NumD0Tensor>{});
+
+                                // get reference to dst data
+                                auto dst_data_refs = generate_tie(
+                                    // return type should be lvalue
+                                    [&](auto) -> auto& {
+                                        return acc0_thread_buf(Number<c_offset>{});
+                                    },
+                                    Number<2>{});
+
+                                unpack2(cde0_element_op, dst_data_refs, src_data_refs);
+                            });
+                            static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                                d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                                    d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                                    make_multi_index(0, 0, 0, 0, 0, 0, 0, 1, 0, 0));
+                            });
+                        });
+                        static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                            d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                                d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                                make_multi_index(0, 0, 0, 1, 0, 0, 0, -n2.value, 0, 0));
+                        });
+                    });
+                    static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                        d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                            d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                            make_multi_index(0, 0, 1, -Gemm0NXdlPerWave, 0, 0, 0, 0, 0, 0));
+                    });
+                });
+                static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                    d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                        d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                        make_multi_index(0, 1, -Gemm0MXdlPerWave, 0, 0, 0, 0, 0, 0, 0));
+                });
+            }
+            // gemm1
+            {
+                // TODO: explore using dynamic buffer for a1 thread buffer
+                // For a1_blockwise_copy, the goal is to satisfy pipeline requirements RunRead(),
+                // RunWrite(), and MoveSliceWindow(). But it is impossible to implement given that
+                // the A1 source buffer is static buffer holding the output of first GEMM and
+                // requires constexpr offset by design. Therefore, we pass tensor coordinate offset
+                // explicitly in Run() below.
+
+                // preload data into LDS
+                b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                     b1_block_slice_copy_step);
+
+                block_sync_lds(); // wait for gemm0 LDS read
+
+                b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+
+                // main body
+                if constexpr(num_gemm1_k_block_inner_loop > 1)
+                {
+                    static_for<0, num_gemm1_k_block_inner_loop - 1, 1>{}([&](auto i) {
+                        a1_blockwise_copy.Run(acc0_thread_desc_k0_m_k1,
+                                              make_tuple(Number<i * A1ThreadSliceK0>{}, I0, I0),
+                                              acc0_thread_buf,
+                                              a1_thread_desc_k0_m_k1,
+                                              make_tuple(I0, I0, I0),
+                                              a1_thread_buf);
+
+                        b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                        block_sync_lds();
+
+                        blockwise_gemm1.Run(a1_thread_buf, b1_block_buf, c1_thread_buf);
+
+                        block_sync_lds();
+
+                        b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                             b1_block_slice_copy_step);
+
+                        b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+                    });
+                }
+                // tail
+                {
+                    a1_blockwise_copy.Run(
+                        acc0_thread_desc_k0_m_k1,
+                        make_tuple(
+                            Number<(num_gemm1_k_block_inner_loop - 1) * A1ThreadSliceK0>{}, I0, I0),
+                        acc0_thread_buf,
+                        a1_thread_desc_k0_m_k1,
+                        make_tuple(I0, I0, I0),
+                        a1_thread_buf);
+
+                    block_sync_lds();
+
+                    blockwise_gemm1.Run(a1_thread_buf, b1_block_buf, c1_thread_buf);
+                }
+            } // end gemm1
+
+            a0_blockwise_copy.MoveSrcSliceWindow(a0_grid_desc_ak0_m_ak1,
+                                                 a0_block_reset_copy_step); // rewind K
+            b0_blockwise_copy.MoveSrcSliceWindow(b0_grid_desc_bk0_n_bk1,
+                                                 b0_block_reset_copy_step); // rewind K and step N
+
+            block_sync_lds(); // wait for gemm1 LDS read
+        } while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
+
+        // shuffle C1 and write out
+        {
+            static_assert(Gemm0MXdlPerWave % C1ShuffleGemm0MXdlPerWavePerShuffle == 0 &&
+                              Gemm1NXdlPerWave % C1ShuffleGemm0NXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = Gemm0MPerBlock / (Gemm0MXdlPerWave * Gemm0MPerXdl);
+            constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * Gemm0NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c1_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm1.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm1.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetC1ShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c1_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<C1ShuffleDataType*>(p_shared),
+                c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<C1ShuffleGemm0MXdlPerWavePerShuffle>{}, // M0 (Gemm0MXdlPerWave) per
+                                                                       // shuffle
+                        M1,                                            // M1 = MWave
+                        M2, // M2 * M3 * M4 = Gemm0MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<C1ShuffleGemm0NXdlPerWavePerShuffle>{}, // N0 (Gemm0NXdlPerWave) per
+                                                                       // shuffle
+                        N1,                                            // N1 = NWave
+                        N2))),                                         // N2 = Gemm0NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM C1 matrix starting index
+            const auto c1_thread_mtx_on_block =
+                blockwise_gemm1.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c1_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c1_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c1_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<Acc1DataType,
+                                                   C1ShuffleDataType,
+                                                   decltype(c1_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<C1ShuffleGemm0MXdlPerWavePerShuffle,
+                                                            C1ShuffleGemm0NXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    tensor_operation::element_wise::PassThrough{}};
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c1_d1s_desc_refs = concat_tuple_of_reference(
+                tie(c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return d1s_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumD1Tensor>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c1_d1s_buf_refs = concat_tuple_of_reference(
+                tie(c1_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return d1s_grid_buf[i]; },
+                    Number<NumD1Tensor>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c1_d1s_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumD1Tensor>{}));
+
+            // shuffle: blockwise copy C from LDS to global
+            auto cde1_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(C1ShuffleDataType{}), D1sDataType{})),
+                Tuple<E1DataType>,
+                decltype(c1_d1s_desc_refs),
+                decltype(tie(e1_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDE1ElementwiseOperation,
+                Sequence<static_cast<index_t>(E1GlobalMemoryDataOperation)>, // FIXME: make Sequence
+                                                                             // support arbitray
+                                                                             // type
+                Sequence<1,
+                         C1ShuffleGemm0MXdlPerWavePerShuffle * MWave * Gemm0MPerXdl,
+                         1,
+                         C1ShuffleGemm0NXdlPerWavePerShuffle * NWave *
+                             Gemm0NPerXdl>, // BlockSliceLengths,
+                CDE1ShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                3,                    // index_t VectorDim,
+                CDE1ShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumD1Tensor,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c1_d1s_desc_refs,
+                 idx_c1_d1s_block_begin,
+                 tie(e1_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
+                 cde1_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c1_vgpr =
+                SpaceFillingCurve<Sequence<Gemm0MXdlPerWave, Gemm1NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<C1ShuffleGemm0MXdlPerWavePerShuffle,
+                                           C1ShuffleGemm0NXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_e1_global = SpaceFillingCurve<
+                Sequence<1, Gemm0MPerBlock, 1, Gemm1NPerBlock>,
+                Sequence<0, 2, 1, 3>,
+                Sequence<1,
+                         C1ShuffleGemm0MXdlPerWavePerShuffle * MWave * Gemm0MPerXdl,
+                         1,
+                         C1ShuffleGemm0NXdlPerWavePerShuffle * NWave * Gemm0NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c1_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_e1_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c1_thread_copy_vgpr_to_lds.Run(c1_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                               sfc_c1_vgpr.GetIndexTupleOfNumber(access_id),
+                                               c1_thread_buf,
+                                               c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                               c1_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde1_shuffle_block_copy_lds_to_global.Run(
+                    c1_d1s_desc_refs,
+                    c1_d1s_buf_refs,
+                    tie(e1_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e1_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto e1_global_step = sfc_e1_global.GetForwardStep(access_id);
+
+                    // move on D1s
+                    static_for<0, NumD1Tensor, 1>{}([&](auto i) {
+                        cde1_shuffle_block_copy_lds_to_global.MoveSrcSliceWindow(
+                            c1_d1s_desc_refs, i + I1, e1_global_step);
+                    });
+
+                    // move on C
+                    cde1_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        tie(e1_grid_desc_mblock_mperblock_nblock_nperblock), I0, e1_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000..7f4001a
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
@@ -0,0 +1,1131 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_softmax.hpp"
+
+namespace ck {
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t B1K1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1ThreadTransferSrcResetCoordinateAfterRun,
+          index_t B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          bool PadN,
+          bool MaskOutUpperTriangle,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
+{
+    static_assert(LoopSched == LoopScheduler::Default,
+                  "Non-default loop scheduler is currently not supported");
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    // Gemm0
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    static constexpr auto Gemm0MWaves = MPerBlock / (MPerXdl * MXdlPerWave);
+    static constexpr auto Gemm0NWaves = NPerBlock / (NPerXdl * NXdlPerWave);
+
+    // Gemm1
+    static constexpr auto B1K0 = Number<Gemm1KPerBlock / B1K1Value>{};
+    static constexpr auto B1K1 = Number<B1K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, MWaves, MPerXdl>(
+            ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, 1, 1>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t Gemm1NWaves = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, Gemm1NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B1 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(B1K0, Number<Gemm1NPerBlock>{}, B1K1),
+            make_tuple(Number<Gemm1NPerBlock + B1BlockLdsExtraN>{} * B1K1, B1K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        const index_t gemm0_bytes_end = (SharedMemTrait::a_block_space_size_aligned +
+                                         SharedMemTrait::b_block_space_size_aligned) *
+                                        sizeof(FloatAB);
+        const index_t gemm1_bytes_end =
+            (SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
+            sizeof(FloatAB);
+        const index_t softmax_bytes_end = (SharedMemTrait::reduction_space_offset +
+                                           SharedMemTrait::reduction_space_size_aligned) *
+                                          sizeof(FloatGemmAcc);
+        const index_t c_block_bytes_end =
+            SharedMemTrait::c_block_space_size * sizeof(FloatCShuffle);
+
+        return math::max(gemm0_bytes_end, gemm1_bytes_end, softmax_bytes_end, c_block_bytes_end);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+        const auto Gemm1N = b1_grid_desc_bk0_n_bk1.GetLength(I1);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && Gemm1N == c_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0 &&
+             Gemm1N % Gemm1NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gemm0 gridwise gemm pipeline
+        const auto num_gemm0_k_loop = K / KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm0_k_loop))
+        {
+            return false;
+        }
+
+        // check gemm1 gridwise gemm pipeline
+        if(!(NPerBlock % Gemm1KPerBlock == 0))
+        {
+            return false;
+        }
+
+        const auto num_gemm1_k_inner_loop = NPerBlock / Gemm1KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm1_k_inner_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / Gemm1NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<Gemm1NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, Gemm1NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    struct SharedMemTrait
+    {
+        // LDS allocation for A and B: be careful of alignment
+        static constexpr auto a_block_desc_ak0_m_ak1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        static constexpr auto b_block_desc_bk0_n_bk1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        static constexpr auto b1_block_desc_bk0_n_bk1 =
+            GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        static constexpr auto max_lds_align = math::lcm(math::lcm(AK1, BK1), B1K1);
+
+        static constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        static constexpr auto a_block_space_offset  = 0;
+        static constexpr auto b_block_space_offset  = a_block_space_size_aligned.value;
+        static constexpr auto b1_block_space_offset = 0;
+
+        // LDS allocation for reduction
+        static constexpr index_t reduction_space_size_aligned =
+            math::integer_least_multiple(BlockSize, max_lds_align);
+
+        static constexpr auto reduction_space_offset = 0;
+
+        // LDS allocation for C shuffle in LDS
+        static constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+        static constexpr auto c_block_space_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+    };
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap, typename C0MatrixMask>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               const FloatAB* __restrict__ p_b1_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const AccElementwiseOperation& acc_element_op,
+                               const B1ElementwiseOperation& b1_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                               const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2CTileMap& block_2_ctile_map,
+                               const C0MatrixMask& c0_matrix_mask)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/gemm1_n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t gemm1_n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        //
+        // set up Gemm0
+        //
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0), // will loop over GemmN dimension
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // Fused Gemm+Gemm pipeline
+        // for n in N0:
+        //   for k in K0:
+        //     acc[m][n] += A[m][k] * B0[k][n]
+        //   acc1[m][o] += acc[m][n] * B1[n][o]
+
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
+            decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            NPerBlock,
+            KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            true>{}; // TransposeC
+
+        auto acc_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::a_block_space_offset,
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b_block_space_offset,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+        const auto a_block_reset_copy_step =
+            make_multi_index(-a_grid_desc_ak0_m_ak1.GetLength(I0), 0, 0);
+        const auto b_block_reset_copy_step =
+            make_multi_index(-b_grid_desc_bk0_n_bk1.GetLength(I0), NPerBlock, 0);
+
+        // gridwise GEMM pipeline
+        // Only supports LoopScheduler::Default
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_Selector<PipelineVer,
+                                                                          NumGemmKPrefetchStage,
+                                                                          LoopScheduler::Default>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        //
+        // set up Gemm1
+        //
+
+        // Acc matrix threadwise copy: AccVGPR to VGPR and downcast to XDL input data type
+        constexpr auto acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        constexpr auto m0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+        constexpr auto n0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        constexpr auto m1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+        constexpr auto n1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        constexpr auto m2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+        constexpr auto n2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        constexpr auto n3 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        constexpr auto n4 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        constexpr auto b1_block_slice_copy_step = make_multi_index(Gemm1KPerBlock / B1K1, 0, 0);
+
+        // acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to acc_thread_desc_k0_m_k1
+        // n0_n1_n2_n3 -> k0
+        // m0_m1_m2 -> m
+        // n4 -> k1
+        // NOTE: had to use merge_v3 or will spit out compilation errors
+        constexpr auto acc_thread_desc_k0_m_k1 = transform_tensor_descriptor(
+            acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
+                       make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
+                       make_pass_through_transform(n4)),
+            make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // A1 matrix in AccVGPR
+        // N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
+        constexpr auto AccN3 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I6);
+
+        constexpr auto A1ThreadSlice_K0_M_K1 =
+            make_tuple(Number<Gemm1KPerBlock / n4 / AccN3>{}, Number<m0 * m1 * m2>{}, Number<n4>{});
+
+        constexpr auto A1ThreadSliceK0        = A1ThreadSlice_K0_M_K1[I0];
+        constexpr auto A1ThreadSliceM         = A1ThreadSlice_K0_M_K1[I1];
+        constexpr auto A1ThreadSliceK1        = A1ThreadSlice_K0_M_K1[I2];
+        constexpr auto a1_thread_desc_k0_m_k1 = make_naive_tensor_descriptor(
+            A1ThreadSlice_K0_M_K1,
+            make_tuple(A1ThreadSliceM * A1ThreadSliceK1, A1ThreadSliceK1, I1));
+
+        // B1 matrix in LDS memory, dst of blockwise copy
+        constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A1 matrix blockwise copy
+        auto a1_blockwise_copy = ThreadwiseTensorSliceTransfer_StaticToStatic<
+            FloatGemmAcc,
+            FloatAB,
+            decltype(acc_thread_desc_k0_m_k1),
+            decltype(a1_thread_desc_k0_m_k1),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<A1ThreadSliceK0, A1ThreadSliceM, A1ThreadSliceK1>,
+            Sequence<1, 0, 2>,
+            2,
+            n4>{tensor_operation::element_wise::PassThrough{}};
+
+        // B1 matrix blockwise copy
+        auto b1_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<B1K0, Gemm1NPerBlock, B1K1>,
+                                                B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                B1BlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b1_grid_desc_bk0_n_bk1),
+                                                decltype(b1_block_desc_bk0_n_bk1),
+                                                B1BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                B1BlockTransferSrcVectorDim,
+                                                2,
+                                                B1BlockTransferSrcScalarPerVector,
+                                                B1BlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                B1ThreadTransferSrcResetCoordinateAfterRun,
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b1_grid_desc_bk0_n_bk1,
+                make_multi_index(0, gemm1_n_block_data_idx_on_grid, 0),
+                b1_element_op,
+                b1_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a1_thread_desc_k0_m_k1.GetElementSpaceSize());
+
+        // reuse LDS space for gemm0's b_block_buf
+        auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b1_block_space_offset,
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        // selected_mfma.group_size or B1K1 <= Gemm1KPack <= selected_mfma.group_size
+        // selected_mfma.k_per_blk <= Gemm1KPack
+        //
+        // Following similar rationale behind Gemm0KPack, let Gemm1KPack be the lowest common
+        // multiples of A1K1 (predetermined by selected_mfma.group_size) and B1K1. But in this case
+        // Gemm1KPack can't be higher than A1K1 itself because A1 matrix is distributed in VGPRs
+        // with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
+        // cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
+        // therefore we may just as well assign Gemm1KPack = group_size
+        constexpr index_t Gemm1KPack =
+            MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
+
+        auto gemm1_blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(b1_block_desc_bk0_n_bk1),
+            decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a1_thread_desc_k0_m_k1)),
+            decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b1_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            Gemm1NPerBlock,
+            Gemm1KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            Gemm1NXdlPerWave,
+            Gemm1KPack,
+            true,       // TransposeC
+            Gemm1KPack, // AMmaKStride
+            Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
+            // BMmaKStride
+            make_tuple(0, 0, 0, 0)}; // A_origin
+
+        auto acc1_thread_buf = gemm1_blockwise_gemm.GetCThreadBuffer();
+
+        //
+        // Blockwise softmax
+        //
+        auto workspace_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatGemmAcc*>(p_shared) + SharedMemTrait::reduction_space_offset,
+            SharedMemTrait::reduction_space_size_aligned);
+
+        // get acc0 8D thread cluster
+        constexpr auto thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths() /
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+        constexpr auto tm0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I0);
+        constexpr auto tn0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I1);
+        constexpr auto tm1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I2);
+        constexpr auto tn1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I3);
+        constexpr auto tm2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I4);
+        constexpr auto tn2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I5);
+        constexpr auto tn3 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I6);
+        constexpr auto tn4 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I7);
+
+        // get acc0 thread map
+        constexpr auto m0_n_m1_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(tm0 * tm1, tm2)),
+                       make_pass_through_transform(I1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        constexpr auto threadid_to_m0_n_m1_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(
+                make_merge_transform(make_tuple(tm0 * tm1, tn0 * tn1 * tn2 * tn3 * tn4, tm2))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+        const auto threadid_to_m_n_thread_cluster_adaptor =
+            chain_tensor_adaptors(m0_n_m1_to_m_n_adaptor, threadid_to_m0_n_m1_adaptor);
+
+        // get acc0 2D thread cluster & 2D thread slice
+        constexpr auto thread_cluster_desc_m_n = make_naive_tensor_descriptor_packed(
+            make_tuple(tm0 * tm1 * tm2, tn0 * tn1 * tn2 * tn3 * tn4));
+        constexpr auto thread_slice_desc_m_n =
+            make_naive_tensor_descriptor_packed(make_tuple(m0 * m1 * m2, n0 * n1 * n2 * n3 * n4));
+
+        auto blockwise_softmax = BlockwiseSoftmax<BlockSize,
+                                                  FloatGemmAcc,
+                                                  decltype(threadid_to_m_n_thread_cluster_adaptor),
+                                                  decltype(thread_cluster_desc_m_n),
+                                                  decltype(thread_slice_desc_m_n)>{};
+
+        const index_t num_gemm1_k_block_outer_loop =
+            b_grid_desc_bk0_n_bk1.GetLength(I1) / NPerBlock;
+        constexpr index_t num_gemm1_k_block_inner_loop = NPerBlock / Gemm1KPerBlock;
+
+        // Initialize C
+        StaticBuffer<AddressSpaceEnum::Vgpr, FloatGemmAcc, acc1_thread_buf.Size(), true>
+            c_thread_buf;
+        c_thread_buf.Clear();
+
+        // Initialize running sum and max of exponentiating row vectors
+        using SoftmaxBuf = typename decltype(blockwise_softmax)::BufferType;
+        SoftmaxBuf running_sum, running_sum_new, running_max, running_max_new;
+        running_sum     = 0;
+        running_sum_new = 0;
+        running_max     = NumericLimits<FloatGemmAcc>::Lowest();
+        running_max_new = NumericLimits<FloatGemmAcc>::Lowest();
+
+        // gemm1 K loop
+        index_t gemm1_k_block_outer_index = 0;
+        do
+        {
+            auto n_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(gemm1_k_block_outer_index * NPerBlock);
+            if(c0_matrix_mask.IsTileSkippable(
+                   m_block_data_idx_on_grid, n_block_data_idx_on_grid, MPerBlock, NPerBlock))
+            {
+                continue;
+            }
+            // gemm0
+            gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                                   a_block_desc_ak0_m_ak1,
+                                                                   a_blockwise_copy,
+                                                                   a_grid_buf,
+                                                                   a_block_buf,
+                                                                   a_block_slice_copy_step,
+                                                                   b_grid_desc_bk0_n_bk1,
+                                                                   b_block_desc_bk0_n_bk1,
+                                                                   b_blockwise_copy,
+                                                                   b_grid_buf,
+                                                                   b_block_buf,
+                                                                   b_block_slice_copy_step,
+                                                                   blockwise_gemm,
+                                                                   acc_thread_buf,
+                                                                   num_k_block_main_loop);
+
+            // do MNK padding or upper triangular masking
+            if constexpr(MaskOutUpperTriangle || PadN)
+            {
+                // 8d thread_desc in thread scope
+                constexpr auto c_thread_lengths =
+                    blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                // 8d block_desc in block scope
+                constexpr auto c_block_lengths =
+                    blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                constexpr auto M0 = c_block_lengths[I0];
+                constexpr auto N0 = c_block_lengths[I1];
+                constexpr auto M1 = c_block_lengths[I2];
+                constexpr auto N1 = c_block_lengths[I3];
+                constexpr auto M2 = c_block_lengths[I4];
+                constexpr auto N2 = c_block_lengths[I5];
+                constexpr auto N3 = c_block_lengths[I6];
+                constexpr auto N4 = c_block_lengths[I7];
+
+                // works like multi-dimension static_for (static_ford), but provides both the linear
+                // index as well as n-d index
+                using Acc0TileIterator = SpaceFillingCurve<
+                    decltype(c_thread_lengths),
+                    typename arithmetic_sequence_gen<0, c_thread_lengths.Size(), 1>::type,
+                    typename uniform_sequence_gen<c_thread_lengths.Size(), 1>::type,
+                    false>; // SnakeCurved
+
+                auto acc0_thread_origin = blockwise_gemm.CalculateCThreadOriginDataIndex8D(
+                    Number<0>{}, Number<0>{}, Number<0>{}, Number<0>{});
+
+                constexpr auto block_idx_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+                    make_tuple(make_unmerge_transform(make_tuple(M0, M1, M2)),
+                               make_unmerge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+                static_for<0, Acc0TileIterator::GetNumOfAccess(), 1>{}([&](auto i) {
+                    auto acc0_thread_idx = Acc0TileIterator::GetIndex(i) + acc0_thread_origin;
+                    auto m_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I0];
+                    auto n_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I1];
+                    auto m_global = m_local + m_block_data_idx_on_grid;
+                    auto n_global = n_local + n_block_data_idx_on_grid;
+                    if(c0_matrix_mask.IsMaskedElement(m_global, n_global))
+                    {
+                        acc_thread_buf(i) = -ck::NumericLimits<float>::Infinity();
+                    }
+                    else
+                    {
+                        acc_element_op(acc_thread_buf(i), acc_thread_buf[i]);
+                    }
+                });
+            }
+            else
+            {
+                static_for<0, acc_thread_buf.Size(), 1>{}(
+                    [&](auto i) { acc_element_op(acc_thread_buf(i), acc_thread_buf[i]); });
+            }
+
+            block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
+
+            // softmax
+            SoftmaxBuf& max = blockwise_softmax.max_value_buf;
+            SoftmaxBuf& sum = blockwise_softmax.sum_value_buf;
+
+            blockwise_softmax.Run(acc_thread_buf, workspace_buf);
+
+            // TODO: may convert to log domain
+            running_max_new = mathext::max(max, running_max);
+            running_sum_new = mathext::exp(running_max - running_max_new) * running_sum +
+                              mathext::exp(max - running_max_new) * sum;
+
+            // gemm1
+            {
+                // TODO: explore using dynamic buffer for a1 thread buffer
+                // For a1_blockwise_copy, the goal is to satisfy pipeline requirements RunRead(),
+                // RunWrite(), and MoveSliceWindow(). But it is impossible to implement given that
+                // the A1 source buffer is static buffer holding the output of first GEMM and
+                // requires constexpr offset by design. Therefore, we pass tensor coordinate offset
+                // explicitly in Run() below.
+
+                // Initialize acc1
+                acc1_thread_buf.Clear();
+
+                // preload data into LDS
+                b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                     b1_block_slice_copy_step);
+
+                block_sync_lds(); // wait for reduction LDS read
+
+                b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+
+                // main body
+                if constexpr(num_gemm1_k_block_inner_loop > 1)
+                {
+                    static_for<0, num_gemm1_k_block_inner_loop - 1, 1>{}([&](auto i) {
+                        a1_blockwise_copy.Run(acc_thread_desc_k0_m_k1,
+                                              make_tuple(Number<i * A1ThreadSliceK0>{}, I0, I0),
+                                              acc_thread_buf,
+                                              a1_thread_desc_k0_m_k1,
+                                              make_tuple(I0, I0, I0),
+                                              a1_thread_buf);
+
+                        b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                        block_sync_lds();
+
+                        gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, acc1_thread_buf);
+
+                        block_sync_lds();
+
+                        b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                             b1_block_slice_copy_step);
+
+                        b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+                    });
+                }
+                // tail
+                {
+                    a1_blockwise_copy.Run(
+                        acc_thread_desc_k0_m_k1,
+                        make_tuple(
+                            Number<(num_gemm1_k_block_inner_loop - 1) * A1ThreadSliceK0>{}, I0, I0),
+                        acc_thread_buf,
+                        a1_thread_desc_k0_m_k1,
+                        make_tuple(I0, I0, I0),
+                        a1_thread_buf);
+
+                    block_sync_lds();
+
+                    gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, acc1_thread_buf);
+                }
+            } // end gemm1
+
+            // workaround compiler issue; see ck/ck.hpp
+            if constexpr(CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE == 1 &&
+                         is_same_v<FloatAB, bhalf_t> && MPerBlock == 256 && NPerBlock == 128 &&
+                         Gemm1NPerBlock == 128)
+            {
+                __builtin_amdgcn_s_barrier();
+            }
+
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+            constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+            constexpr auto cn0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+            constexpr auto cm1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+            constexpr auto cn1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+            constexpr auto cm2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+            constexpr auto cn2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+            constexpr auto cn3 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+            constexpr auto cn4 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+            constexpr auto c_thread_slice_desc_m_n = make_naive_tensor_descriptor_packed(
+                make_tuple(cm0 * cm1 * cm2, cn0 * cn1 * cn2 * cn3 * cn4));
+            constexpr auto c_thread_buf_slice_m = c_thread_slice_desc_m_n.GetLength(I0);
+            constexpr auto c_thread_buf_slice_n = c_thread_slice_desc_m_n.GetLength(I1);
+
+            static_for<0, c_thread_buf_slice_m, 1>{}([&](auto iM) {
+                static_for<0, c_thread_buf_slice_n, 1>{}([&](auto iN) {
+                    auto I = Number<c_thread_slice_desc_m_n.CalculateOffset(make_tuple(iM, iN))>{};
+                    FloatGemmAcc acc1 = acc1_thread_buf[I]; // P*V
+                    FloatGemmAcc c    = c_thread_buf[I];    // O
+                    FloatGemmAcc c_new =
+                        (running_sum[iM] * math::exp(running_max[iM] - running_max_new[iM]) * c +
+                         math::exp(max[iM] - running_max_new[iM]) * acc1) /
+                        running_sum_new[iM]; // Formula by Dao et al.,
+                                             // https://arxiv.org/pdf/2205.14135v2.pdf section 3.1
+
+                    c_thread_buf(I) = c_new; // O_new
+                });
+            });
+
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_ak0_m_ak1,
+                                                a_block_reset_copy_step); // rewind K
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_bk0_n_bk1,
+                                                b_block_reset_copy_step); // rewind K and step N
+
+            // update before next j iteration
+            running_max = running_max_new;
+            running_sum = running_sum_new;
+
+            block_sync_lds(); // wait for gemm1 LDS read
+        } while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              Gemm1NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp =
+                gemm1_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5);
+            constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6);
+            constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2)),                                    // M2 = MPerXdl
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2,                                      // N2 * N3 * N4 = NPerXdl
+                        N3,
+                        N4))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                gemm1_blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            I1,
+                                                            N2,
+                                                            I1,
+                                                            N4>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     n_thread_data_on_block_idx[I2],
+                                     n_thread_data_on_block_idx[I3],
+                                     n_thread_data_on_block_idx[I4]),
+                    tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, Gemm1NXdlPerWave, 1, 1, 1, N2, 1, N4>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           1,
+                                           N2,
+                                           1,
+                                           N4>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, Gemm1NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp
new file mode 100644
index 0000000..ede6a96
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp
@@ -0,0 +1,554 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseBatchrNormBackwardWithBlockwiseWelford_,
+          typename XDataType,
+          typename DyDataType,
+          typename DxDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M,
+          typename GetReduceCountPerThreadFunctor>
+__global__ void kernel_batchnorm_backward_with_blockwise_welford(
+    const XYGridDesc_M_K x_grid_desc_m_k,
+    const XYGridDesc_M_K dy_grid_desc_m_k,
+    const XYGridDesc_M_K dx_grid_desc_m_k,
+    const ScaleBiasGridDesc_M scale_grid_desc_m,
+    const ScaleBiasGridDesc_M dscale_dbias_grid_desc_m,
+    const MeanVarGridDesc_M mean_var_grid_desc_m,
+    const GetReduceCountPerThreadFunctor get_reduce_count_per_thread,
+    long_index_t reduce_size,
+    index_t num_k_block_tile_iteration,
+    AccDataType epsilon,
+    const XDataType* const __restrict__ p_x,
+    const DyDataType* const __restrict__ p_dy,
+    const ScaleDataType* const __restrict__ p_scale,
+    bool haveSavedMeanInvVar,
+    const MeanVarDataType* const __restrict__ p_savedMean,
+    const MeanVarDataType* const __restrict__ p_savedInvVar,
+    const DyElementwiseOp dy_elementwise_op,
+    DxDataType* const __restrict__ p_dx,
+    DscaleDbiasDataType* const __restrict__ p_dscale,
+    DscaleDbiasDataType* const __restrict__ p_dbias)
+{
+    GridwiseBatchrNormBackwardWithBlockwiseWelford_::Run(x_grid_desc_m_k,
+                                                         dy_grid_desc_m_k,
+                                                         dx_grid_desc_m_k,
+                                                         scale_grid_desc_m,
+                                                         dscale_dbias_grid_desc_m,
+                                                         mean_var_grid_desc_m,
+                                                         get_reduce_count_per_thread,
+                                                         reduce_size,
+                                                         num_k_block_tile_iteration,
+                                                         epsilon,
+                                                         p_x,
+                                                         p_dy,
+                                                         p_scale,
+                                                         haveSavedMeanInvVar,
+                                                         p_savedMean,
+                                                         p_savedInvVar,
+                                                         dy_elementwise_op,
+                                                         p_dx,
+                                                         p_dscale,
+                                                         p_dbias);
+};
+
+template <typename XDataType,
+          typename DyDataType,
+          typename DxDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M,
+          typename GetReduceCountPerThreadFunctor,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XDyDxVectorDim,
+          index_t XSrcVectorSize,
+          index_t DySrcVectorSize,
+          index_t DxDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t DscaleDbiasDstVectorSize,
+          index_t MeanVarSrcVectorSize>
+struct GridwiseBatchNormBackwardWithBlockwiseWelford
+{
+    static_assert((XDyDxVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0 &&
+                   MThreadSliceSize % DySrcVectorSize == 0 &&
+                   MThreadSliceSize % DxDstVectorSize == 0) ||
+                      (XDyDxVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0 &&
+                       KThreadSliceSize % DySrcVectorSize == 0 &&
+                       KThreadSliceSize % DxDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XDyDxVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    using BlockwiseReduce = PartitionedBlockwiseReduction<AccDataType,
+                                                          BlockSize,
+                                                          ThreadClusterLengths_M_K,
+                                                          ThreadClusterArrangeOrder,
+                                                          ck::reduce::Add,
+                                                          false>;
+
+    using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                 ThreadReduceSrcDesc_M_K,
+                                                 ThreadReduceDstDesc_M,
+                                                 ck::reduce::Add,
+                                                 false>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    // clang-format off
+    // Blockwise BatchNorm Backward
+    // Input: x, dy, scale, savedMean and savedInvVar (optional), reduce_size
+    // Output: dx, dscale, dbias
+    // Step 1: calculating mean and inv-variance using welford method (if savedMean/savedInvVar not available), where inv-variance = 1/sqrt(epsilon+variance)
+    // Step 2: reduction: dbias = sum(dy),  dscale = sum(dy *(x-mean) * inv-variance)
+    // Step 3: calculating dx = 1/reduce_size * inv-variance * scale * (reduce_size * dy - dbias - dscale * (x - mean) * inv-variance)) elementwise-ly
+    // clang-format on
+    __device__ static void Run(const XYGridDesc_M_K x_grid_desc_m_k,
+                               const XYGridDesc_M_K dy_grid_desc_m_k,
+                               const XYGridDesc_M_K dx_grid_desc_m_k,
+                               const ScaleBiasGridDesc_M scale_grid_desc_m,
+                               const ScaleBiasGridDesc_M dscale_dbias_grid_desc_m,
+                               const MeanVarGridDesc_M mean_var_grid_desc_m,
+                               const GetReduceCountPerThreadFunctor get_reduce_count_per_thread,
+                               long_index_t reduce_size,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               const XDataType* const __restrict__ p_x,
+                               const DyDataType* const __restrict__ p_dy,
+                               const ScaleDataType* const __restrict__ p_scale,
+                               bool haveSavedMeanInvVar,
+                               const MeanVarDataType* const __restrict__ p_savedMean,
+                               const MeanVarDataType* const __restrict__ p_savedInvVar,
+                               const DyElementwiseOp dy_elementwise_op,
+                               DxDataType* const __restrict__ p_dx,
+                               DscaleDbiasDataType* const __restrict__ p_dscale,
+                               DscaleDbiasDataType* const __restrict__ p_dbias)
+    {
+        using ck::math::sqrt;
+
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            dy_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            dx_thread_buf;
+
+        // buffer of values of dy * (x-mean) * invVariance, used as input of Blockwise reduction
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            tmp1_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> scale_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>&
+            inv_var_thread_buf = var_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> dscale_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> dbias_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M           = Sequence<MThreadSliceSize>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XYGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XDyDxVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dy_load = ThreadwiseTensorSliceTransfer_v2<DyDataType,
+                                                                   AccDataType,
+                                                                   XYGridDesc_M_K,
+                                                                   decltype(thread_buffer_desc_m_k),
+                                                                   ThreadBufferLengths_M_K,
+                                                                   ThreadBufferDimAccessOrder,
+                                                                   XDyDxVectorDim,
+                                                                   XSrcVectorSize,
+                                                                   1,
+                                                                   true>(
+            dy_grid_desc_m_k,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dx_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               DxDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               XYGridDesc_M_K,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               XDyDxVectorDim,
+                                               DxDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                dx_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * KThreadSliceSize),
+                PassThroughOp{});
+
+        auto threadwise_scale_load =
+            ThreadwiseTensorSliceTransfer_v2<ScaleDataType,
+                                             AccDataType,
+                                             ScaleBiasGridDesc_M,
+                                             decltype(thread_buffer_desc_m),
+                                             ThreadBufferLengths_M,
+                                             Sequence<0>,
+                                             0,
+                                             ScaleSrcVectorSize,
+                                             1,
+                                             true>(
+                scale_grid_desc_m,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize));
+
+        auto threadwise_dscale_dbias_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               DscaleDbiasDataType,
+                                               decltype(thread_buffer_desc_m),
+                                               ScaleBiasGridDesc_M,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M,
+                                               Sequence<0>,
+                                               0,
+                                               DscaleDbiasDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                dscale_dbias_grid_desc_m,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize),
+                PassThroughOp{});
+
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileSize);
+        constexpr auto thread_copy_bwd_step_m_k = make_multi_index(0, -K_BlockTileSize);
+
+        const auto x_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto dy_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dy, dy_grid_desc_m_k.GetElementSpaceSize());
+
+        auto dx_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dx, dx_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto scale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_scale, scale_grid_desc_m.GetElementSpaceSize());
+
+        auto dscale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dscale, dscale_dbias_grid_desc_m.GetElementSpaceSize());
+
+        auto dbias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dbias, dscale_dbias_grid_desc_m.GetElementSpaceSize());
+
+        // clang-format off
+        // Step 1: calculating mean and inv-variance using welford method (if savedMean/savedInvVar not available), where inv-variance = 1/sqrt(epsilon+variance)
+        // clang-format on
+
+        if(haveSavedMeanInvVar)
+        {
+            const auto mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_savedMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            const auto inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_savedInvVar, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto threadwise_mean_inv_var_load =
+                ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                                 AccDataType,
+                                                 MeanVarGridDesc_M,
+                                                 decltype(thread_buffer_desc_m),
+                                                 ThreadBufferLengths_M,
+                                                 Sequence<0>,
+                                                 0,
+                                                 MeanVarSrcVectorSize,
+                                                 1,
+                                                 true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize));
+
+            threadwise_mean_inv_var_load.Run(mean_var_grid_desc_m,
+                                             mean_global_buf,
+                                             thread_buffer_desc_m,
+                                             make_tuple(I0),
+                                             mean_thread_buf);
+
+            threadwise_mean_inv_var_load.Run(mean_var_grid_desc_m,
+                                             inv_var_global_buf,
+                                             thread_buffer_desc_m,
+                                             make_tuple(I0),
+                                             inv_var_thread_buf);
+        }
+        else
+        {
+            auto threadwise_welford       = ThreadwiseWelford();
+            threadwise_welford.max_count_ = get_reduce_count_per_thread(thread_k_cluster_id);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
+                var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+            });
+
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+            {
+
+                threadwise_x_load.Run(x_grid_desc_m_k,
+                                      x_global_buf,
+                                      thread_buffer_desc_m_k,
+                                      make_tuple(I0, I0),
+                                      x_thread_buf);
+
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                threadwise_welford.Run(x_thread_buf, mean_thread_buf, var_thread_buf);
+            }
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                int count = threadwise_welford.cur_count_;
+                BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
+            });
+
+            // calculate inv-variance as 1/sqrt(epsilon+variance)
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                inv_var_thread_buf(I) =
+                    type_convert<AccDataType>(1.0) / sqrt(var_thread_buf[I] + epsilon);
+            });
+
+            threadwise_x_load.SetSrcSliceOrigin(
+                x_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * KThreadSliceSize));
+        };
+
+        // clang-format off
+        // Step 2: reduction: dbias = sum(dy),  dscale = sum(dy *(x-mean) * inv-variance)
+        // clang-format on
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            dscale_thread_buf(I) = type_convert<AccDataType>(0);
+            dbias_thread_buf(I)  = type_convert<AccDataType>(0);
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_dy_load.Run(dx_grid_desc_m_k,
+                                   dy_global_buf,
+                                   thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   dy_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    dy_elementwise_op(dy_thread_buf(Number<offset>{}),
+                                      dy_thread_buf[Number<offset>{}]);
+
+                    AccDataType norm_x = (x_thread_buf[Number<offset>{}] - mean_thread_buf[iM]) *
+                                         inv_var_thread_buf[iM];
+
+                    tmp1_thread_buf(Number<offset>{}) = norm_x * dy_thread_buf[Number<offset>{}];
+                });
+            });
+
+            ThreadwiseReduce::Reduce(tmp1_thread_buf, dscale_thread_buf);
+            ThreadwiseReduce::Reduce(dy_thread_buf, dbias_thread_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+            threadwise_dy_load.MoveSrcSliceWindow(dy_grid_desc_m_k, thread_copy_fwd_step_m_k);
+        };
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+            BlockwiseReduce::Reduce(reduce_work_buf, dscale_thread_buf(I));
+            block_sync_lds();
+            BlockwiseReduce::Reduce(reduce_work_buf, dbias_thread_buf(I));
+        });
+
+        if(thread_k_cluster_id == 0)
+        {
+            threadwise_dscale_dbias_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              dscale_thread_buf,
+                                              dscale_dbias_grid_desc_m,
+                                              dscale_global_buf);
+
+            threadwise_dscale_dbias_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              dbias_thread_buf,
+                                              dscale_dbias_grid_desc_m,
+                                              dbias_global_buf);
+        };
+
+        // clang-format off
+        // Step 3: calculating dx = 1/reduce_size * inv-variance * scale * (reduce_size * dy - dbias - dscale * (x - mean) * inv-variance)) elementwise-ly
+        // clang-format on
+
+        threadwise_scale_load.Run(scale_grid_desc_m,
+                                  scale_global_buf,
+                                  thread_buffer_desc_m,
+                                  make_tuple(I0),
+                                  scale_thread_buf);
+
+        auto thread_copy_tail_m_k = (num_k_block_tile_iteration - 1) * thread_copy_fwd_step_m_k;
+
+        threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        threadwise_dy_load.MoveSrcSliceWindow(dy_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        threadwise_dx_store.MoveDstSliceWindow(dx_grid_desc_m_k, thread_copy_tail_m_k);
+
+        AccDataType inv_reduce_size =
+            type_convert<AccDataType>(1.0) / type_convert<AccDataType>(reduce_size);
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_dy_load.Run(dy_grid_desc_m_k,
+                                   dy_global_buf,
+                                   thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   dy_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                AccDataType multiplier =
+                    inv_reduce_size * inv_var_thread_buf[iM] * scale_thread_buf[iM];
+
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    dy_elementwise_op(dy_thread_buf(Number<offset>{}),
+                                      dy_thread_buf[Number<offset>{}]);
+
+                    AccDataType norm_x = (x_thread_buf[Number<offset>{}] - mean_thread_buf[iM]) *
+                                         inv_var_thread_buf[iM];
+
+                    AccDataType tmpVal = norm_x * dscale_thread_buf[iM];
+
+                    dx_thread_buf(Number<offset>{}) =
+                        multiplier *
+                        (type_convert<AccDataType>(reduce_size) * dy_thread_buf[Number<offset>{}] -
+                         dbias_thread_buf[iM] - tmpVal);
+                });
+            });
+
+            threadwise_dx_store.Run(thread_buffer_desc_m_k,
+                                    make_tuple(I0, I0),
+                                    dx_thread_buf,
+                                    dx_grid_desc_m_k,
+                                    dx_global_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_dy_load.MoveSrcSliceWindow(dy_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_dx_store.MoveDstSliceWindow(dx_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp
new file mode 100644
index 0000000..33c45a0
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp
@@ -0,0 +1,483 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseBatchrNormForwardWithBlockwiseWelford_,
+          typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M,
+          typename GetReduceCountPerThreadFunctor>
+__global__ void kernel_batchnorm_forward_with_blockwise_welford(
+    const XYGridDesc_M_K x_grid_desc_m_k,
+    const XYGridDesc_M_K y_grid_desc_m_k,
+    const ScaleBiasGridDesc_M scale_grid_desc_m,
+    const ScaleBiasGridDesc_M bias_grid_desc_m,
+    const MeanVarGridDesc_M mean_var_grid_desc_m,
+    const GetReduceCountPerThreadFunctor get_reduce_count_per_thread,
+    index_t num_k_block_tile_iteration,
+    AccDataType epsilon,
+    const XDataType* const __restrict__ p_x,
+    const ScaleDataType* const __restrict__ p_scale,
+    const BiasDataType* const __restrict__ p_bias,
+    const YElementwiseOp y_elementwise_op,
+    YDataType* const __restrict__ p_y,
+    bool updateMovingAverage,
+    AccDataType averageFactor,
+    MeanVarDataType* const __restrict__ resultRunningMean,
+    MeanVarDataType* const __restrict__ resultRunningVariance,
+    bool saveMeanInvVariance,
+    MeanVarDataType* const __restrict__ resultSaveMean,
+    MeanVarDataType* const __restrict__ resultSaveInvVariance)
+{
+    GridwiseBatchrNormForwardWithBlockwiseWelford_::Run(x_grid_desc_m_k,
+                                                        y_grid_desc_m_k,
+                                                        scale_grid_desc_m,
+                                                        bias_grid_desc_m,
+                                                        mean_var_grid_desc_m,
+                                                        get_reduce_count_per_thread,
+                                                        num_k_block_tile_iteration,
+                                                        epsilon,
+                                                        p_x,
+                                                        p_scale,
+                                                        p_bias,
+                                                        y_elementwise_op,
+                                                        p_y,
+                                                        updateMovingAverage,
+                                                        averageFactor,
+                                                        resultRunningMean,
+                                                        resultRunningVariance,
+                                                        saveMeanInvVariance,
+                                                        resultSaveMean,
+                                                        resultSaveInvVariance);
+};
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M,
+          typename GetReduceCountPerThreadFunctor,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcYDstVectorDim,
+          index_t XSrcVectorSize,
+          index_t YDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t BiasSrcVectorSize,
+          index_t MeanVarSrcDstVectorSize>
+struct GridwiseBatchNormForwardWithBlockwiseWelford
+{
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XSrcYDstVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    __device__ static void Run(const XYGridDesc_M_K& x_grid_desc_m_k,
+                               const XYGridDesc_M_K& y_grid_desc_m_k,
+                               const ScaleBiasGridDesc_M& scale_grid_desc_m,
+                               const ScaleBiasGridDesc_M& bias_grid_desc_m,
+                               const MeanVarGridDesc_M& mean_var_grid_desc_m,
+                               const GetReduceCountPerThreadFunctor& get_reduce_count_per_thread,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               const XDataType* const __restrict__ p_x,
+                               const ScaleDataType* const __restrict__ p_scale,
+                               const BiasDataType* const __restrict__ p_bias,
+                               const YElementwiseOp y_elementwise_op,
+                               YDataType* const __restrict__ p_y,
+                               bool updateMovingAverage,
+                               AccDataType averageFactor,
+                               MeanVarDataType* const __restrict__ resultRunningMean,
+                               MeanVarDataType* const __restrict__ resultRunningVariance,
+                               bool saveMeanInvVariance,
+                               MeanVarDataType* const __restrict__ resultSaveMean,
+                               MeanVarDataType* const __restrict__ resultSaveInvVariance)
+    {
+        using ck::math::sqrt;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> scale_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> bias_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            y_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M           = Sequence<MThreadSliceSize>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XYGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcYDstVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               XYGridDesc_M_K,
+                                               YElementwiseOp,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               XSrcYDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                y_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * KThreadSliceSize),
+                y_elementwise_op);
+
+        auto threadwise_scale_load =
+            ThreadwiseTensorSliceTransfer_v2<ScaleDataType,
+                                             AccDataType,
+                                             ScaleBiasGridDesc_M,
+                                             decltype(thread_buffer_desc_m),
+                                             ThreadBufferLengths_M,
+                                             Sequence<0>,
+                                             0,
+                                             ScaleSrcVectorSize,
+                                             1,
+                                             true>(
+                scale_grid_desc_m,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize));
+
+        auto threadwise_bias_load = ThreadwiseTensorSliceTransfer_v2<BiasDataType,
+                                                                     AccDataType,
+                                                                     ScaleBiasGridDesc_M,
+                                                                     decltype(thread_buffer_desc_m),
+                                                                     ThreadBufferLengths_M,
+                                                                     Sequence<0>,
+                                                                     0,
+                                                                     BiasSrcVectorSize,
+                                                                     1,
+                                                                     true>(
+            bias_grid_desc_m,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                             thread_m_cluster_id * MThreadSliceSize));
+
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileSize);
+        constexpr auto thread_copy_bwd_step_m_k = make_multi_index(0, -K_BlockTileSize);
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto scale_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_scale, scale_grid_desc_m.GetElementSpaceSize());
+
+        const auto bias_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_bias, bias_grid_desc_m.GetElementSpaceSize());
+
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y, y_grid_desc_m_k.GetElementSpaceSize());
+
+        // Step 1:  do welford reduction to get mean and variance
+
+        auto threadwise_welford       = ThreadwiseWelford();
+        threadwise_welford.max_count_ = get_reduce_count_per_thread(thread_k_cluster_id);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
+            var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_val_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+            threadwise_welford.Run(x_thread_buf, mean_thread_buf, var_thread_buf);
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            int count = threadwise_welford.cur_count_;
+            BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
+        });
+
+        // Step 2: do normalization and output y
+
+        threadwise_scale_load.Run(scale_grid_desc_m,
+                                  scale_global_val_buf,
+                                  thread_buffer_desc_m,
+                                  make_tuple(I0),
+                                  scale_thread_buf);
+
+        threadwise_bias_load.Run(bias_grid_desc_m,
+                                 bias_global_val_buf,
+                                 thread_buffer_desc_m,
+                                 make_tuple(I0),
+                                 bias_thread_buf);
+
+        auto thread_copy_tail_m_k = (num_k_block_tile_iteration - 1) * thread_copy_fwd_step_m_k;
+
+        threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_val_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                AccDataType multiplier =
+                    scale_thread_buf[Number<iM>{}] / sqrt(var_thread_buf[iM] + epsilon);
+
+                AccDataType fused_mean_bias =
+                    bias_thread_buf[Number<iM>{}] - mean_thread_buf[iM] * multiplier;
+
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    // normalize
+                    y_thread_buf(Number<offset>{}) =
+                        x_thread_buf[Number<offset>{}] * multiplier + fused_mean_bias;
+                });
+            });
+
+            threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   y_thread_buf,
+                                   y_grid_desc_m_k,
+                                   y_global_val_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        }
+
+        // Step 3: update the moving average of mean and variance (optional)
+
+        if(updateMovingAverage && thread_k_cluster_id == 0)
+        {
+            StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+                running_mean_thread_buf;
+            StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+                running_var_thread_buf;
+
+            auto running_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultRunningMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto running_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultRunningVariance, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto threadwise_mean_var_load =
+                ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                                 AccDataType,
+                                                 MeanVarGridDesc_M,
+                                                 decltype(thread_buffer_desc_m),
+                                                 ThreadBufferLengths_M,
+                                                 Sequence<0>,
+                                                 0,
+                                                 MeanVarSrcDstVectorSize,
+                                                 1,
+                                                 true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize));
+
+            threadwise_mean_var_load.Run(mean_var_grid_desc_m,
+                                         running_mean_global_buf,
+                                         thread_buffer_desc_m,
+                                         make_tuple(I0),
+                                         running_mean_thread_buf);
+
+            threadwise_mean_var_load.Run(mean_var_grid_desc_m,
+                                         running_var_global_buf,
+                                         thread_buffer_desc_m,
+                                         make_tuple(I0),
+                                         running_var_thread_buf);
+
+            AccDataType oneMinusAverageFactor = type_convert<AccDataType>(1.0) - averageFactor;
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                running_mean_thread_buf(I) = running_mean_thread_buf[I] * oneMinusAverageFactor +
+                                             mean_thread_buf[I] * averageFactor;
+                running_var_thread_buf(I) = running_var_thread_buf[I] * oneMinusAverageFactor +
+                                            var_thread_buf[I] * averageFactor;
+            });
+
+            auto threadwise_mean_var_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   MeanVarDataType,
+                                                   decltype(thread_buffer_desc_m),
+                                                   MeanVarGridDesc_M,
+                                                   PassThroughOp,
+                                                   ThreadBufferLengths_M,
+                                                   Sequence<0>,
+                                                   0,
+                                                   MeanVarSrcDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_mean_var_store.Run(thread_buffer_desc_m,
+                                          make_tuple(I0),
+                                          running_mean_thread_buf,
+                                          mean_var_grid_desc_m,
+                                          running_mean_global_buf);
+
+            threadwise_mean_var_store.Run(thread_buffer_desc_m,
+                                          make_tuple(I0),
+                                          running_var_thread_buf,
+                                          mean_var_grid_desc_m,
+                                          running_var_global_buf);
+        };
+
+        // Step 4: save mean and inv-variance (optional)
+
+        if(saveMeanInvVariance && thread_k_cluster_id == 0)
+        {
+            auto result_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultSaveMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto result_inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultSaveInvVariance, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            // calculate inv-variance as 1/sqrt(epsilon+variance), stored in place of variance
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                var_thread_buf(I) =
+                    type_convert<AccDataType>(1.0f) / sqrt(epsilon + var_thread_buf[I]);
+            });
+
+            auto threadwise_mean_inv_var_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   MeanVarDataType,
+                                                   decltype(thread_buffer_desc_m),
+                                                   MeanVarGridDesc_M,
+                                                   PassThroughOp,
+                                                   ThreadBufferLengths_M,
+                                                   Sequence<0>,
+                                                   0,
+                                                   MeanVarSrcDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              mean_thread_buf,
+                                              mean_var_grid_desc_m,
+                                              result_mean_global_buf);
+
+            threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              var_thread_buf,
+                                              mean_var_grid_desc_m,
+                                              result_inv_var_global_buf);
+        };
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_contraction_dlops_v1r2.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_contraction_dlops_v1r2.hpp
new file mode 100644
index 0000000..2369f51
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_contraction_dlops_v1r2.hpp
@@ -0,0 +1,662 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_GRIDWISE_CONTRACTION_DLOPS_V1R2_HPP
+#define CK_GRIDWISE_CONTRACTION_DLOPS_V1R2_HPP
+
+#include "common_header.hpp"
+#include "multi_index_transform_helper.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "blockwise_gemm_dlops_v2r3.hpp"
+#include "blockwise_tensor_slice_transfer_v2.hpp"
+#include "threadwise_tensor_slice_transfer.hpp"
+#include "threadwise_tensor_slice_set.hpp"
+
+namespace ck {
+
+template <typename GridwiseContraction,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_GK0_GM0_GM10_GM11_GK1,
+          typename BGridDesc_GK0_GN0_GN10_GN11_GK1,
+          typename CGridDesc_GM10_BM0_BM1_GN10_BN0_BN1,
+          typename CGridBlockCluster_BlockId_To_GM10_GN10,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_contraction_dlops_v1r2(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AGridDesc_GK0_GM0_GM10_GM11_GK1 a_grid_desc_gk0_gm0_gm10_gm11_gk1,
+            const BGridDesc_GK0_GN0_GN10_GN11_GK1 b_grid_desc_gk0_gn0_gn10_gn11_gk1,
+            const CGridDesc_GM10_BM0_BM1_GN10_BN0_BN1 c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1,
+            const CGridBlockCluster_BlockId_To_GM10_GN10 c_grid_block_cluster_blockid_to_gm10_gn10)
+{
+    constexpr index_t shared_block_size =
+        GridwiseContraction::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseContraction::Run(p_a_grid,
+                             p_b_grid,
+                             p_c_grid,
+                             p_shared_block,
+                             a_grid_desc_gk0_gm0_gm10_gm11_gk1,
+                             b_grid_desc_gk0_gn0_gn10_gn11_gk1,
+                             c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1,
+                             c_grid_block_cluster_blockid_to_gm10_gn10,
+                             integral_constant<bool, HasMainKBlockLoop>{},
+                             integral_constant<bool, HasDoubleTailKBlockLoop>{});
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_GK0_GM0_GM1_GK1,
+          typename BGridDesc_GK0_GN0_GN1_GK1,
+          typename CGridDesc_GM0_GM1_GN0_GN1,
+          index_t GM1PerBlockGM11,
+          index_t GN1PerBlockGN11,
+          index_t GK0PerBlock,
+          index_t BM1PerThreadBM11,
+          index_t BN1PerThreadBN11,
+          index_t BK0PerThread,
+          typename BM10BN10ThreadClusterBM10Xs,
+          typename BM10BN10ThreadClusterBN10Xs,
+          typename ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1,
+          typename ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1,
+          typename ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1,
+          typename BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1,
+          typename BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          typename AGridStepHacks,
+          typename BGridStepHacks,
+          typename CGridStepHacks,
+          typename AGridMoveSliceWindowStepHacks,
+          typename BGridMoveSliceWindowStepHacks>
+struct GridwiseContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0_GM1_GN0_GN1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    // GM0 and GN0 need to known at compile-time
+    static constexpr auto GM0 = CGridDesc_GM0_GM1_GN0_GN1{}.GetLength(I0);
+    static constexpr auto GN0 = CGridDesc_GM0_GM1_GN0_GN1{}.GetLength(I2);
+    static constexpr auto GK1 = AGridDesc_GK0_GM0_GM1_GK1{}.GetLength(I3);
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // lds max alignment
+        // TODO: part of them should be moved into blockwise-gemm
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = GK1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_block_desc_gk0_gm0_gm10_gm11_gk1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<GK0PerBlock>{}, GM0, I1, Number<GM1PerBlockGM11>{}, GK1),
+            max_lds_align);
+
+        // B matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto b_block_desc_gk0_gn0_gn10_gn11_gk1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<GK0PerBlock>{}, GN0, I1, Number<GN1PerBlockGN11>{}, GK1),
+            max_lds_align);
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
+            a_block_desc_gk0_gm0_gm10_gm11_gk1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
+            b_block_desc_gk0_gn0_gn10_gn11_gk1.GetElementSpaceSize(), max_lds_align);
+
+        return 2 * (a_block_aligned_space_size + b_block_aligned_space_size) * sizeof(FloatAB);
+    }
+
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_GK0_GM0_GM1_GK1& a_grid_desc_gk0_gm0_gm1_gk1,
+                  const BGridDesc_GK0_GN0_GN1_GK1& b_grid_desc_gk0_gn0_gn1_gk1,
+                  const CGridDesc_GM0_GM1_GN0_GN1& c_grid_desc_gm0_gm1_gn0_gn1)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(GM0)>>::value &&
+                          is_known_at_compile_time<remove_cv_t<decltype(GN0)>>::value,
+                      "wrong! GM0 and GN0 need to be known at compile-time");
+
+        const auto GM1 = a_grid_desc_gk0_gm0_gm1_gk1.GetLength(I2);
+        const auto GN1 = b_grid_desc_gk0_gn0_gn1_gk1.GetLength(I2);
+        const auto GK0 = a_grid_desc_gk0_gm0_gm1_gk1.GetLength(I0);
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+
+        return (
+            (GM0 == c_grid_desc_gm0_gm1_gn0_gn1.GetLength(I0) &&
+             GM1 == c_grid_desc_gm0_gm1_gn0_gn1.GetLength(I1) &&
+             GN0 == c_grid_desc_gm0_gm1_gn0_gn1.GetLength(I2) &&
+             GN1 == c_grid_desc_gm0_gm1_gn0_gn1.GetLength(I3) &&
+             GM0 == a_grid_desc_gk0_gm0_gm1_gk1.GetLength(I1) &&
+             GM1 == a_grid_desc_gk0_gm0_gm1_gk1.GetLength(I2) &&
+             GN0 == b_grid_desc_gk0_gn0_gn1_gk1.GetLength(I1) &&
+             GN1 == b_grid_desc_gk0_gn0_gn1_gk1.GetLength(I2) &&
+             GK0 == b_grid_desc_gk0_gn0_gn1_gk1.GetLength(I0) &&
+             GK1 == b_grid_desc_gk0_gn0_gn1_gk1.GetLength(I3)) &&
+            (GM1 % GM1PerBlockGM11 == 0 && GN1 % GN1PerBlockGN11 == 0 && GK0 % GK0PerBlock == 0));
+    }
+
+    __host__ __device__ static constexpr index_t
+    CalculateGridSize(const CGridDesc_GM0_GM1_GN0_GN1& c_grid_desc_gm0_gm1_gn0_gn1)
+    {
+        const auto GM1 = c_grid_desc_gm0_gm1_gn0_gn1.GetLength(I1);
+        const auto GN1 = c_grid_desc_gm0_gm1_gn0_gn1.GetLength(I3);
+
+        constexpr index_t GM11 = GM1PerBlockGM11;
+        constexpr index_t GN11 = GN1PerBlockGN11;
+
+        const index_t GM10 = GM1 / GM11;
+        const index_t GN10 = GN1 / GN11;
+
+        const index_t grid_size = GM10 * GN10;
+
+        return grid_size;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t GK0)
+    {
+        const bool has_main_k_block_loop = (GK0 + GK0PerBlock) / (2 * GK0PerBlock) > 1;
+
+        return has_main_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasDoubleTailKBlockLoop(index_t GK0)
+    {
+        const bool has_double_tail_k_block_loop = (GK0 / GK0PerBlock) % 2 == 0;
+
+        return has_double_tail_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto MakeAGridDescriptor_GK0_GM0_GM10_GM11_GK1(
+        const AGridDesc_GK0_GM0_GM1_GK1& a_grid_desc_gk0_gm0_gm1_gk1)
+    {
+        const auto GK0 = a_grid_desc_gk0_gm0_gm1_gk1.GetLength(I0);
+        const auto GM1 = a_grid_desc_gk0_gm0_gm1_gk1.GetLength(I2);
+
+        const auto GM11 = Number<GM1PerBlockGM11>{};
+        const auto GM10 = GM1 / GM11;
+
+        const auto a_grid_desc_gk0_gm0_gm10_gm11_gk1 = transform_tensor_descriptor(
+            a_grid_desc_gk0_gm0_gm1_gk1,
+            make_tuple(make_pass_through_transform(GK0),
+                       make_pass_through_transform(GM0),
+                       make_unmerge_transform(make_tuple(GM10, GM11)),
+                       make_pass_through_transform(GK1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}));
+
+        return a_grid_desc_gk0_gm0_gm10_gm11_gk1;
+    }
+
+    __host__ __device__ static constexpr auto MakeBGridDescriptor_GK0_GN0_GN10_GN11_GK1(
+        const BGridDesc_GK0_GN0_GN1_GK1& b_grid_desc_gk0_gn0_gn1_gk1)
+    {
+        const auto GK0 = b_grid_desc_gk0_gn0_gn1_gk1.GetLength(I0);
+        const auto GN1 = b_grid_desc_gk0_gn0_gn1_gk1.GetLength(I2);
+
+        const auto GN11 = Number<GN1PerBlockGN11>{};
+        const auto GN10 = GN1 / GN11;
+
+        const auto b_grid_desc_gk0_gn0_gn10_gn11_gk1 = transform_tensor_descriptor(
+            b_grid_desc_gk0_gn0_gn1_gk1,
+            make_tuple(make_pass_through_transform(GK0),
+                       make_pass_through_transform(GN0),
+                       make_unmerge_transform(make_tuple(GN10, GN11)),
+                       make_pass_through_transform(GK1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}));
+
+        return b_grid_desc_gk0_gn0_gn10_gn11_gk1;
+    }
+
+    __host__ __device__ static constexpr auto MakeCGridDescriptor_GM10_BM0_BM1_GN10_BN0_BN1(
+        const CGridDesc_GM0_GM1_GN0_GN1& c_grid_desc_gm0_gm1_gn0_gn1)
+    {
+        const auto GM1 = c_grid_desc_gm0_gm1_gn0_gn1.GetLength(I1);
+        const auto GN1 = c_grid_desc_gm0_gm1_gn0_gn1.GetLength(I3);
+
+        constexpr auto GM11 = Number<GM1PerBlockGM11>{};
+        constexpr auto GN11 = Number<GN1PerBlockGN11>{};
+
+        const auto GM10 = GM1 / GM11;
+        const auto GN10 = GN1 / GN11;
+
+        constexpr auto BM = GM0 * GM11;
+        constexpr auto BN = GN0 * GN11;
+
+        constexpr auto BM1 =
+            Number<container_reduce(BM10BN10ThreadClusterBM10Xs{}, math::multiplies{}, I1) *
+                   BM1PerThreadBM11>{};
+        constexpr auto BN1 =
+            Number<container_reduce(BM10BN10ThreadClusterBN10Xs{}, math::multiplies{}, I1) *
+                   BN1PerThreadBN11>{};
+
+        constexpr auto BM0 = BM / BM1;
+        constexpr auto BN0 = BN / BN1;
+
+        const auto c_gm0_gm10_gm11_gn0_gn10_gn11_grid_desc = transform_tensor_descriptor(
+            c_grid_desc_gm0_gm1_gn0_gn1,
+            make_tuple(make_pass_through_transform(GM0),
+                       make_unmerge_transform(make_tuple(GM10, GM11)),
+                       make_pass_through_transform(GN0),
+                       make_unmerge_transform(make_tuple(GN10, GN11))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}, Sequence<4, 5>{}));
+
+        const auto c_gm10_bm_gn10_bn_grid_desc = transform_tensor_descriptor(
+            c_gm0_gm10_gm11_gn0_gn10_gn11_grid_desc,
+            make_tuple(make_pass_through_transform(GM10),
+                       make_merge_transform(make_tuple(GM0, GM11)),
+                       make_pass_through_transform(GN10),
+                       make_merge_transform(make_tuple(GN0, GN11))),
+            make_tuple(Sequence<1>{}, Sequence<0, 2>{}, Sequence<4>{}, Sequence<3, 5>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        const auto c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1 = transform_tensor_descriptor(
+            c_gm10_bm_gn10_bn_grid_desc,
+            make_tuple(make_pass_through_transform(GM10),
+                       make_unmerge_transform(make_tuple(BM0, BM1)),
+                       make_pass_through_transform(GN10),
+                       make_unmerge_transform(make_tuple(BN0, BN1))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}, Sequence<4, 5>{}));
+
+        return c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1;
+    }
+
+    __host__ __device__ static constexpr auto MakeCGridBlockCluster_BlockId_To_GM10_GN10(
+        const CGridDesc_GM0_GM1_GN0_GN1& c_grid_desc_gm0_gm1_gn0_gn1)
+    {
+        const auto GM1 = c_grid_desc_gm0_gm1_gn0_gn1.GetLength(I1);
+        const auto GN1 = c_grid_desc_gm0_gm1_gn0_gn1.GetLength(I3);
+
+        constexpr auto GM11 = Number<GM1PerBlockGM11>{};
+        constexpr auto GN11 = Number<GN1PerBlockGN11>{};
+
+        const auto GM10 = GM1 / GM11;
+        const auto GN10 = GN1 / GN11;
+
+        const auto c_grid_block_cluster_blockid_to_gm10_gn10 = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(GM10, GN10))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        return c_grid_block_cluster_blockid_to_gm10_gn10;
+    }
+
+    using AGridDesc_GK0_GM0_GM10_GM11_GK1 =
+        decltype(MakeAGridDescriptor_GK0_GM0_GM10_GM11_GK1(AGridDesc_GK0_GM0_GM1_GK1{}));
+    using BGridDesc_GK0_GN0_GN10_GN11_GK1 =
+        decltype(MakeBGridDescriptor_GK0_GN0_GN10_GN11_GK1(BGridDesc_GK0_GN0_GN1_GK1{}));
+    using CGridDesc_GM10_BM0_BM1_GN10_BN0_BN1 =
+        decltype(MakeCGridDescriptor_GM10_BM0_BM1_GN10_BN0_BN1(CGridDesc_GM0_GM1_GN0_GN1{}));
+    using CGridBlockCluster_BlockId_To_GM10_GN10 =
+        decltype(MakeCGridBlockCluster_BlockId_To_GM10_GN10(CGridDesc_GM0_GM1_GN0_GN1{}));
+
+    template <bool HasMainKBlockLoop, bool HasDoubleTailKBlockLoop>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        FloatAB* __restrict__ p_shared_block,
+        const AGridDesc_GK0_GM0_GM10_GM11_GK1& a_grid_desc_gk0_gm0_gm10_gm11_gk1,
+        const BGridDesc_GK0_GN0_GN10_GN11_GK1& b_grid_desc_gk0_gn0_gn10_gn11_gk1,
+        const CGridDesc_GM10_BM0_BM1_GN10_BN0_BN1& c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1,
+        const CGridBlockCluster_BlockId_To_GM10_GN10& c_grid_block_cluster_blockid_to_gm10_gn10,
+        integral_constant<bool, HasMainKBlockLoop>,
+        integral_constant<bool, HasDoubleTailKBlockLoop>)
+    {
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_gk0_gm0_gm10_gm11_gk1.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_gk0_gn0_gn10_gn11_gk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1.GetElementSpaceSize());
+
+        const auto GK0 = a_grid_desc_gk0_gm0_gm10_gm11_gk1.GetLength(I0);
+
+        // divide block work by [GM10, GN10]
+        const auto c_gm10_gn10_block_cluster_idx =
+            c_grid_block_cluster_blockid_to_gm10_gn10.CalculateBottomIndex(
+                make_multi_index(get_block_1d_id()));
+
+        // HACK: this force index data into SGPR
+        const index_t igm10 = __builtin_amdgcn_readfirstlane(c_gm10_gn10_block_cluster_idx[I0]);
+        const index_t ign10 = __builtin_amdgcn_readfirstlane(c_gm10_gn10_block_cluster_idx[I1]);
+
+        // lds max alignment
+        // TODO: part of them should be moved into blockwise-gemm
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = GK1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_block_desc_gk0_gm0_gm10_gm11_gk1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<GK0PerBlock>{}, GM0, I1, Number<GM1PerBlockGM11>{}, GK1),
+            max_lds_align);
+
+        // B matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto b_block_desc_gk0_gn0_gn10_gn11_gk1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<GK0PerBlock>{}, GN0, I1, Number<GN1PerBlockGN11>{}, GK1),
+            max_lds_align);
+
+        // A matrix in LDS memory for blockwise GEMM
+        //   be careful of LDS alignment
+        constexpr auto a_block_desc_gk0_bm_gk1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<GK0PerBlock>{}, GM0 * Number<GM1PerBlockGM11>{}, GK1), max_lds_align);
+
+        // B matrix in LDS memory for blockwise GEMM
+        //   be careful of LDS alignment
+        constexpr auto b_block_desc_gk0_bn_gk1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<GK0PerBlock>{}, GN0 * Number<GN1PerBlockGN11>{}, GK1), max_lds_align);
+
+        static_assert(a_block_desc_gk0_gm0_gm10_gm11_gk1.GetElementSpaceSize() ==
+                              a_block_desc_gk0_bm_gk1.GetElementSpaceSize() &&
+                          b_block_desc_gk0_gn0_gn10_gn11_gk1.GetElementSpaceSize() ==
+                              b_block_desc_gk0_bn_gk1.GetElementSpaceSize(),
+                      "wrong!");
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
+            BlockSize,
+            InMemoryDataOperationEnum::Set,
+            Sequence<GK0PerBlock, GM0, 1, GM1PerBlockGM11, GK1.value>,
+            ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1,
+            ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1,
+            ABlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            decltype(a_grid_desc_gk0_gm0_gm10_gm11_gk1),
+            decltype(a_block_desc_gk0_gm0_gm10_gm11_gk1),
+            ABlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2, 3, 4>,
+            ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1, // SrcVectorTensorLengths
+            ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1, // DstVectorTensorLengths
+            ABlockTransferSrcVectorTensorContiguousDimOrder, // SrcVectorTensorContiguousDimOrder
+            Sequence<0, 1, 2, 3, 4>,                         // DstVectorTensorContiguousDimOrder
+            false,
+            true>(a_grid_desc_gk0_gm0_gm10_gm11_gk1,
+                  make_multi_index(0, 0, igm10, 0, 0),
+                  a_block_desc_gk0_gm0_gm10_gm11_gk1,
+                  make_multi_index(0, 0, 0, 0, 0));
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
+            BlockSize,
+            InMemoryDataOperationEnum::Set,
+            Sequence<GK0PerBlock, GN0, 1, GN1PerBlockGN11, GK1.value>,
+            BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1,
+            BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1,
+            BBlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            decltype(b_grid_desc_gk0_gn0_gn10_gn11_gk1),
+            decltype(b_block_desc_gk0_gn0_gn10_gn11_gk1),
+            BBlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2, 3, 4>,
+            BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1, // SrcVectorTensorLengths
+            BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1, // DstVectorTensorLengths
+            BBlockTransferSrcVectorTensorContiguousDimOrder, // SrcVectorTensorContiguousDimOrder
+            Sequence<0, 1, 2, 3, 4>,                         // DstVectorTensorContiguousDimOrder
+            false,
+            true>(b_grid_desc_gk0_gn0_gn10_gn11_gk1,
+                  make_multi_index(0, 0, ign10, 0, 0),
+                  b_block_desc_gk0_gn0_gn10_gn11_gk1,
+                  make_multi_index(0, 0, 0, 0, 0));
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[GK0PerBlock, GM1PerBlockGM11] is in LDS
+        //     b_mtx[KPerBlocl, GN1PerBlockGN11] is in LDS
+        //     c_mtx[GM1PerBlockGM11, GN1PerBlockGN11] is distributed among threads, and saved in
+        //       register
+        const auto blockwise_gemm =
+            BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2<
+                BlockSize,
+                FloatAB,
+                FloatAB,
+                FloatAcc,
+                decltype(a_block_desc_gk0_bm_gk1),
+                decltype(b_block_desc_gk0_bn_gk1),
+                BM1PerThreadBM11,
+                BN1PerThreadBN11,
+                BK0PerThread,
+                BM10BN10ThreadClusterBM10Xs,
+                BM10BN10ThreadClusterBN10Xs,
+                BM1PerThreadBM11,
+                BN1PerThreadBN11>{};
+
+        constexpr auto c_thread_tensor_lengths_bm0_bm1_bn0_bn1 =
+            decltype(blockwise_gemm)::GetCThreadTensorLengths_BM0_BM1_BN0_BN1();
+
+        constexpr auto c_thread_desc_bm0_bm1_bn0_bn1 = make_naive_tensor_descriptor_packed(
+            sequence_to_tuple_of_number(c_thread_tensor_lengths_bm0_bm1_bn0_bn1));
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
+            a_block_desc_gk0_gm0_gm10_gm11_gk1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
+            b_block_desc_gk0_gn0_gn10_gn11_gk1.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block_double = p_shared_block;
+        FloatAB* p_b_block_double = p_shared_block + 2 * a_block_aligned_space_size;
+
+        // register allocation for output
+        auto c_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAcc>(
+            c_thread_desc_bm0_bm1_bn0_bn1.GetElementSpaceSize());
+
+        ThreadwiseTensorSliceSet_v1<FloatAcc,
+                                    decltype(c_thread_desc_bm0_bm1_bn0_bn1),
+                                    decltype(c_thread_tensor_lengths_bm0_bm1_bn0_bn1)>{}
+            .Run(c_thread_desc_bm0_bm1_bn0_bn1,
+                 make_tuple(I0, I0, I0, I0),
+                 c_thread_buf,
+                 FloatAcc{0});
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(GK0PerBlock, 0, 0, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(GK0PerBlock, 0, 0, 0, 0);
+
+        auto a_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double, a_block_desc_gk0_gm0_gm10_gm11_gk1.GetElementSpaceSize());
+        auto b_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double, b_block_desc_gk0_gn0_gn10_gn11_gk1.GetElementSpaceSize());
+
+        auto a_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double + a_block_aligned_space_size,
+            a_block_desc_gk0_gm0_gm10_gm11_gk1.GetElementSpaceSize());
+        auto b_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double + b_block_aligned_space_size,
+            b_block_desc_gk0_gn0_gn10_gn11_gk1.GetElementSpaceSize());
+
+        // LDS double buffer: preload data into LDS
+        {
+            a_blockwise_copy.RunRead(
+                a_grid_desc_gk0_gm0_gm10_gm11_gk1, a_global_buf, AGridStepHacks{});
+            b_blockwise_copy.RunRead(
+                b_grid_desc_gk0_gn0_gn10_gn11_gk1, b_global_buf, BGridStepHacks{});
+
+            a_blockwise_copy.RunWrite(a_block_desc_gk0_gm0_gm10_gm11_gk1, a_block_even_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_gk0_gn0_gn10_gn11_gk1, b_block_even_buf);
+        }
+
+        if constexpr(HasMainKBlockLoop)
+        {
+            index_t gk0_block_on_grid = 0;
+
+            // LDS double buffer: main body
+            // use Do-While loop instead of For loop to simplify control flow
+            do
+            {
+                // even iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_gk0_gm0_gm10_gm11_gk1,
+                                                    a_block_slice_copy_step,
+                                                    AGridMoveSliceWindowStepHacks{});
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_gk0_gn0_gn10_gn11_gk1,
+                                                    b_block_slice_copy_step,
+                                                    BGridMoveSliceWindowStepHacks{});
+
+                __syncthreads();
+
+                // LDS doubel buffer: load next data from device mem
+                a_blockwise_copy.RunRead(
+                    a_grid_desc_gk0_gm0_gm10_gm11_gk1, a_global_buf, AGridStepHacks{});
+                b_blockwise_copy.RunRead(
+                    b_grid_desc_gk0_gn0_gn10_gn11_gk1, b_global_buf, BGridStepHacks{});
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(c_thread_desc_bm0_bm1_bn0_bn1,
+                                   a_block_even_buf,
+                                   b_block_even_buf,
+                                   c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_block_desc_gk0_gm0_gm10_gm11_gk1, a_block_odd_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_gk0_gn0_gn10_gn11_gk1, b_block_odd_buf);
+
+                // odd iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_gk0_gm0_gm10_gm11_gk1,
+                                                    a_block_slice_copy_step,
+                                                    AGridMoveSliceWindowStepHacks{});
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_gk0_gn0_gn10_gn11_gk1,
+                                                    b_block_slice_copy_step,
+                                                    BGridMoveSliceWindowStepHacks{});
+
+                __syncthreads();
+
+                // LDS doubel buffer: load next data from device mem
+                a_blockwise_copy.RunRead(
+                    a_grid_desc_gk0_gm0_gm10_gm11_gk1, a_global_buf, AGridStepHacks{});
+                b_blockwise_copy.RunRead(
+                    b_grid_desc_gk0_gn0_gn10_gn11_gk1, b_global_buf, BGridStepHacks{});
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(
+                    c_thread_desc_bm0_bm1_bn0_bn1, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_block_desc_gk0_gm0_gm10_gm11_gk1, a_block_even_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_gk0_gn0_gn10_gn11_gk1, b_block_even_buf);
+
+                gk0_block_on_grid += 2 * GK0PerBlock;
+            } while(gk0_block_on_grid < GK0 - 2 * GK0PerBlock);
+        }
+
+        // LDS double buffer: tail
+        if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
+        {
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_gk0_gm0_gm10_gm11_gk1,
+                                                a_block_slice_copy_step,
+                                                AGridMoveSliceWindowStepHacks{});
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_gk0_gn0_gn10_gn11_gk1,
+                                                b_block_slice_copy_step,
+                                                BGridMoveSliceWindowStepHacks{});
+
+            __syncthreads();
+
+            // LDS double buffer: load last data from device mem
+            a_blockwise_copy.RunRead(
+                a_grid_desc_gk0_gm0_gm10_gm11_gk1, a_global_buf, AGridStepHacks{});
+            b_blockwise_copy.RunRead(
+                b_grid_desc_gk0_gn0_gn10_gn11_gk1, b_global_buf, BGridStepHacks{});
+
+            // LDS double buffer: GEMM on 2nd-last data
+            blockwise_gemm.Run(
+                c_thread_desc_bm0_bm1_bn0_bn1, a_block_even_buf, b_block_even_buf, c_thread_buf);
+
+            // LDS double buffer: store last data to LDS
+            a_blockwise_copy.RunWrite(a_block_desc_gk0_gm0_gm10_gm11_gk1, a_block_odd_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_gk0_gn0_gn10_gn11_gk1, b_block_odd_buf);
+
+            __syncthreads();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_thread_desc_bm0_bm1_bn0_bn1, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+        }
+        else // if has 1 iteration left
+        {
+            __syncthreads();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_thread_desc_bm0_bm1_bn0_bn1, a_block_even_buf, b_block_even_buf, c_thread_buf);
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_gm10_bm0_bm1_gn10_bn0_bn1 =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<c_thread_tensor_lengths_bm0_bm1_bn0_bn1[I0]>{},
+                               Number<c_thread_tensor_lengths_bm0_bm1_bn0_bn1[I1]>{},
+                               I1,
+                               Number<c_thread_tensor_lengths_bm0_bm1_bn0_bn1[I2]>{},
+                               Number<c_thread_tensor_lengths_bm0_bm1_bn0_bn1[I3]>{}));
+
+            const auto c_thread_origin_on_block_bm0_bm1_bn0_bn1 =
+                blockwise_gemm.CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
+                    get_thread_local_1d_id());
+
+            ThreadwiseTensorSliceTransfer_v1r3<
+                FloatAcc,
+                FloatC,
+                decltype(c_thread_desc_gm10_bm0_bm1_gn10_bn0_bn1),
+                decltype(c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1),
+                Sequence<1,
+                         c_thread_tensor_lengths_bm0_bm1_bn0_bn1[I0],
+                         c_thread_tensor_lengths_bm0_bm1_bn0_bn1[I1],
+                         1,
+                         c_thread_tensor_lengths_bm0_bm1_bn0_bn1[I2],
+                         c_thread_tensor_lengths_bm0_bm1_bn0_bn1[I3]>,
+                CThreadTransferSrcDstAccessOrder,
+                CThreadTransferSrcDstVectorDim,
+                CThreadTransferDstScalarPerVector,
+                CGlobalMemoryDataOperation,
+                1,
+                false>{c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1,
+                       make_multi_index(igm10,
+                                        c_thread_origin_on_block_bm0_bm1_bn0_bn1[I0],
+                                        c_thread_origin_on_block_bm0_bm1_bn0_bn1[I1],
+                                        ign10,
+                                        c_thread_origin_on_block_bm0_bm1_bn0_bn1[I2],
+                                        c_thread_origin_on_block_bm0_bm1_bn0_bn1[I3])}
+                .Run(c_thread_desc_gm10_bm0_bm1_gn10_bn0_bn1,
+                     make_tuple(I0, I0, I0, I0, I0, I0),
+                     c_thread_buf,
+                     c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1,
+                     c_grid_buf,
+                     CGridStepHacks{});
+        }
+    }
+};
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp
new file mode 100644
index 0000000..8b82b65
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp
@@ -0,0 +1,195 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseElementwise1dFunctor,
+          typename InGrid1dDescTuple,
+          typename OutGrid1dDescTuple,
+          typename InDataTypePointerTuple,
+          typename OutDataTypePointerTuple,
+          typename ElementwiseOperation>
+__global__ void kernel_elementwise_1d(const InGrid1dDescTuple in_grid_1d_desc_tuple,
+                                      const OutGrid1dDescTuple out_grid_1d_desc_tuple,
+                                      const InDataTypePointerTuple p_in_global_tuple,
+                                      const OutDataTypePointerTuple p_out_global_tuple,
+                                      const ElementwiseOperation elementwise_op)
+{
+    GridwiseElementwise1dFunctor::Run(in_grid_1d_desc_tuple,
+                                      out_grid_1d_desc_tuple,
+                                      p_in_global_tuple,
+                                      p_out_global_tuple,
+                                      elementwise_op);
+}
+
+template <typename InGrid1dDescTuple,
+          typename OutGrid1dDescTuple,
+          typename InDataTypePointerTuple,
+          typename OutDataTypePointerTuple,
+          typename ElementwiseOperation,
+          index_t MPerThread,
+          typename InScalarPerVectorSeq,
+          typename OutScalarPerVectorSeq>
+struct GridwiseElementwise_1D
+{
+    static constexpr index_t NumInput  = InDataTypePointerTuple::Size();
+    static constexpr index_t NumOutput = OutDataTypePointerTuple::Size();
+
+    static_assert(NumInput == InScalarPerVectorSeq::Size() &&
+                      NumOutput == OutScalarPerVectorSeq::Size() &&
+                      NumInput == InGrid1dDescTuple::Size() &&
+                      NumOutput == OutGrid1dDescTuple::Size(),
+                  "Tuple size is inconsistent with the number of in/out!");
+
+    static constexpr auto I0 = Number<0>{};
+
+    static constexpr auto thread_buffer_desc_m =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MPerThread>{}));
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    __device__ static void Run(const InGrid1dDescTuple in_grid_1d_desc_tuple,
+                               const OutGrid1dDescTuple out_grid_1d_desc_tuple,
+                               const InDataTypePointerTuple p_in_global_tuple,
+                               const OutDataTypePointerTuple p_out_global_tuple,
+                               const ElementwiseOperation elementwise_op)
+    {
+        const index_t thread_global_id = get_thread_global_1d_id();
+
+        auto in_thread_buf_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(InDataTypePointerTuple{}[I])>;
+                using DataType        = remove_cv_t<remove_pointer_t<DataTypePointer>>;
+
+                return StaticBuffer<AddressSpaceEnum::Vgpr, DataType, MPerThread, true>{};
+            },
+            Number<NumInput>{});
+
+        auto out_thread_buf_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[I])>;
+                using DataType        = remove_pointer_t<DataTypePointer>;
+
+                return StaticBuffer<AddressSpaceEnum::Vgpr, DataType, MPerThread, true>{};
+            },
+            Number<NumOutput>{});
+
+        auto in_global_buf_tuple = generate_tuple(
+            [&](auto I) {
+                static_assert(in_grid_1d_desc_tuple[I].GetNumOfDimension() == 1);
+
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_in_global_tuple[I], in_grid_1d_desc_tuple[I].GetElementSpaceSize());
+            },
+            Number<NumInput>{});
+
+        auto out_global_buf_tuple = generate_tuple(
+            [&](auto I) {
+                static_assert(out_grid_1d_desc_tuple[I].GetNumOfDimension() == 1);
+
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_out_global_tuple[I], out_grid_1d_desc_tuple[I].GetElementSpaceSize());
+            },
+            Number<NumOutput>{});
+
+        const auto thread_global_offset = make_multi_index(thread_global_id * MPerThread);
+
+        const index_t blockSize    = get_block_size();
+        const index_t blockPerGrid = get_grid_size();
+        const auto M               = in_grid_1d_desc_tuple[I0].GetLength(I0);
+        const index_t loop_step    = blockPerGrid * blockSize * MPerThread;
+        const auto loop_step_index = make_multi_index(loop_step);
+
+        auto in_global_load_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(InDataTypePointerTuple{}[I])>;
+                using DataType        = remove_cv_t<remove_pointer_t<DataTypePointer>>;
+
+                return ThreadwiseTensorSliceTransfer_v2<DataType,
+                                                        DataType,
+                                                        decltype(in_grid_1d_desc_tuple[I]),
+                                                        decltype(thread_buffer_desc_m),
+                                                        Sequence<MPerThread>, // SliceLengths
+                                                        Sequence<0>,          // DimAccessOrder
+                                                        0,                    // SrcVectorDim
+                                                        InScalarPerVectorSeq::At(
+                                                            I), // ScalarPerVector
+                                                        1,      // SrcScalarStrideInVector
+                                                        false>{in_grid_1d_desc_tuple[I],
+                                                               thread_global_offset};
+            },
+            Number<NumInput>{});
+
+        auto out_global_store_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[I])>;
+                using DataType        = remove_pointer_t<DataTypePointer>;
+
+                return ThreadwiseTensorSliceTransfer_v1r3<DataType,
+                                                          DataType,
+                                                          decltype(thread_buffer_desc_m),
+                                                          decltype(out_grid_1d_desc_tuple[I]),
+                                                          PassThroughOp,
+                                                          Sequence<MPerThread>, // SliceLengths
+                                                          Sequence<0>,          // DimAccessOrder
+                                                          0,                    // SrcVectorDim
+                                                          OutScalarPerVectorSeq::At(I),
+                                                          InMemoryDataOperationEnum::Set,
+                                                          1,
+                                                          false>(
+                    out_grid_1d_desc_tuple[I], thread_global_offset, PassThroughOp{});
+            },
+            Number<NumOutput>{});
+
+        index_t num_iter = M / (loop_step);
+        do
+        {
+            static_for<0, NumInput, 1>{}([&](auto I) {
+                in_global_load_tuple(I).Run(in_grid_1d_desc_tuple[I],
+                                            in_global_buf_tuple[I],
+                                            thread_buffer_desc_m,
+                                            make_tuple(I0),
+                                            in_thread_buf_tuple(I));
+
+                in_global_load_tuple(I).MoveSrcSliceWindow(in_grid_1d_desc_tuple[I],
+                                                           loop_step_index);
+            });
+
+            static_for<0, MPerThread, 1>{}([&](auto iM) {
+                // get reference to in data
+                const auto in_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto I) -> const auto& { return in_thread_buf_tuple(I)(iM); },
+                    Number<NumInput>{});
+
+                // get reference to dst data
+                auto out_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto I) -> auto& { return out_thread_buf_tuple(I)(iM); },
+                    Number<NumOutput>{});
+
+                unpack2(elementwise_op, out_data_refs, in_data_refs);
+            });
+
+            static_for<0, NumOutput, 1>{}([&](auto I) {
+                out_global_store_tuple(I).Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              out_thread_buf_tuple[I],
+                                              out_grid_1d_desc_tuple[I],
+                                              out_global_buf_tuple(I));
+
+                out_global_store_tuple(I).MoveDstSliceWindow(out_grid_1d_desc_tuple[I],
+                                                             loop_step_index);
+            });
+        } while(--num_iter);
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp
new file mode 100644
index 0000000..05257d1
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp
@@ -0,0 +1,230 @@
+// SPDX-License-Identifier: MIT
+// // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+//
+#pragma once
+
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseElementwise2dFunctor,
+          typename InGrid2dDescTuple,
+          typename OutGrid2dDescTuple,
+          typename InDataTypePointerTuple,
+          typename OutDataTypePointerTuple,
+          typename ElementwiseOperation>
+__global__ void kernel_elementwise_2d(const InGrid2dDescTuple in_grid_2d_desc_tuple,
+                                      const OutGrid2dDescTuple out_grid_2d_desc_tuple,
+                                      const InDataTypePointerTuple p_in_global_tuple,
+                                      const OutDataTypePointerTuple p_out_global_tuple,
+                                      const ElementwiseOperation elementwise_op,
+                                      const index_t num_threads_m,
+                                      const index_t num_threads_n)
+{
+    GridwiseElementwise2dFunctor::Run(in_grid_2d_desc_tuple,
+                                      out_grid_2d_desc_tuple,
+                                      p_in_global_tuple,
+                                      p_out_global_tuple,
+                                      elementwise_op,
+                                      num_threads_m,
+                                      num_threads_n);
+}
+
+template <typename InGrid2dDescTuple,
+          typename OutGrid2dDescTuple,
+          typename InDataTypePointerTuple,
+          typename OutDataTypePointerTuple,
+          typename ElementwiseOperation,
+          index_t MPerThread,
+          index_t NPerThread,
+          typename InScalarPerVectorSeq,
+          typename OutScalarPerVectorSeq>
+struct GridwiseElementwise_2D
+{
+    static constexpr index_t NumInput  = InDataTypePointerTuple::Size();
+    static constexpr index_t NumOutput = OutDataTypePointerTuple::Size();
+
+    static_assert(NumInput == InScalarPerVectorSeq::Size() &&
+                      NumOutput == OutScalarPerVectorSeq::Size() &&
+                      NumInput == InGrid2dDescTuple::Size() &&
+                      NumOutput == OutGrid2dDescTuple::Size(),
+                  "Tuple size is inconsistent with the number of in/out!");
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr auto thread_buffer_desc_mn =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MPerThread>{}, Number<NPerThread>{}));
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    __device__ static void Run(const InGrid2dDescTuple in_grid_2d_desc_tuple,
+                               const OutGrid2dDescTuple out_grid_2d_desc_tuple,
+                               const InDataTypePointerTuple p_in_global_tuple,
+                               const OutDataTypePointerTuple p_out_global_tuple,
+                               const ElementwiseOperation elementwise_op,
+                               const index_t num_threads_m,
+                               const index_t num_threads_n)
+    {
+        auto in_thread_buf_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(InDataTypePointerTuple{}[I])>;
+                using DataType        = remove_cv_t<remove_pointer_t<DataTypePointer>>;
+
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    DataType,
+                                    MPerThread * NPerThread,
+                                    true>{};
+            },
+            Number<NumInput>{});
+
+        auto out_thread_buf_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[I])>;
+                using DataType        = remove_pointer_t<DataTypePointer>;
+
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    DataType,
+                                    MPerThread * NPerThread,
+                                    true>{};
+            },
+            Number<NumOutput>{});
+
+        auto in_global_buf_tuple = generate_tuple(
+            [&](auto I) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_in_global_tuple[I], in_grid_2d_desc_tuple[I].GetElementSpaceSize());
+            },
+            Number<NumInput>{});
+
+        auto out_global_buf_tuple = generate_tuple(
+            [&](auto I) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_out_global_tuple[I], out_grid_2d_desc_tuple[I].GetElementSpaceSize());
+            },
+            Number<NumOutput>{});
+
+        const auto M = in_grid_2d_desc_tuple[I0].GetLength(I0);
+        const auto N = in_grid_2d_desc_tuple[I0].GetLength(I1);
+
+        const index_t loop_step_m = num_threads_m * MPerThread;
+        const index_t loop_step_n = num_threads_n * NPerThread;
+
+        const index_t thread_1d_id = get_thread_global_1d_id();
+        index_t tid_m              = thread_1d_id / num_threads_n;
+        index_t tid_n              = thread_1d_id % num_threads_n;
+
+        const auto thread_global_offset = make_multi_index(tid_m * MPerThread, tid_n * NPerThread);
+
+        auto in_global_load_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(InDataTypePointerTuple{}[I])>;
+                using DataType        = remove_cv_t<remove_pointer_t<DataTypePointer>>;
+
+                return ThreadwiseTensorSliceTransfer_v2<
+                    DataType,
+                    DataType,
+                    decltype(in_grid_2d_desc_tuple[I]),
+                    decltype(thread_buffer_desc_mn),
+                    Sequence<MPerThread, NPerThread>, // SliceLengths
+                    Sequence<0, 1>,                   // DimAccessOrder
+                    0,                                // SrcVectorDim
+                    InScalarPerVectorSeq::At(I),      // ScalarPerVector
+                    1,                                // SrcScalarStrideInVector
+                    true>{in_grid_2d_desc_tuple[I], thread_global_offset};
+            },
+            Number<NumInput>{});
+
+        auto out_global_store_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[I])>;
+                using DataType        = remove_pointer_t<DataTypePointer>;
+
+                return ThreadwiseTensorSliceTransfer_v1r3<
+                    DataType,
+                    DataType,
+                    decltype(thread_buffer_desc_mn),
+                    decltype(out_grid_2d_desc_tuple[I]),
+                    PassThroughOp,
+                    Sequence<MPerThread, NPerThread>, // SliceLengths
+                    Sequence<0, 1>,                   // DimAccessOrder
+                    1,                                // SrcVectorDim
+                    1,                                // OutScalarPerVectorSeq::At(I),
+                    InMemoryDataOperationEnum::Set,
+                    1,
+                    true>(out_grid_2d_desc_tuple[I], thread_global_offset, PassThroughOp{});
+            },
+            Number<NumOutput>{});
+
+        index_t num_iter_m = M / (loop_step_m);
+        do
+        {
+            index_t num_iter_n = N / (loop_step_n);
+            do
+            {
+                static_for<0, NumInput, 1>{}([&](auto I) {
+                    in_global_load_tuple(I).Run(in_grid_2d_desc_tuple[I],
+                                                in_global_buf_tuple[I],
+                                                thread_buffer_desc_mn,
+                                                make_tuple(I0, I0),
+                                                in_thread_buf_tuple(I));
+
+                    in_global_load_tuple(I).MoveSrcSliceWindow(in_grid_2d_desc_tuple[I],
+                                                               make_multi_index(0, loop_step_n));
+                });
+
+                static_for<0, MPerThread, 1>{}([&](auto iM) {
+                    static_for<0, NPerThread, 1>{}([&](auto iN) {
+                        constexpr auto offset =
+                            thread_buffer_desc_mn.CalculateOffset(make_tuple(iM, iN));
+                        // get reference to in data
+                        const auto in_data_refs = generate_tie(
+                            // return type should be lvalue
+                            [&](auto I) -> const auto& {
+                                return in_thread_buf_tuple(I)(Number<offset>{});
+                            },
+                            Number<NumInput>{});
+
+                        // get referenec to dst data
+                        auto out_data_refs = generate_tie(
+                            // return type should be lvalue
+                            [&](auto I) -> auto& {
+                                return out_thread_buf_tuple(I)(Number<offset>{});
+                            },
+                            Number<NumOutput>{});
+                        unpack2(elementwise_op, out_data_refs, in_data_refs);
+                    });
+                });
+
+                static_for<0, NumOutput, 1>{}([&](auto I) {
+                    out_global_store_tuple(I).Run(thread_buffer_desc_mn,
+                                                  make_tuple(I0, I0),
+                                                  out_thread_buf_tuple[I],
+                                                  out_grid_2d_desc_tuple[I],
+                                                  out_global_buf_tuple(I));
+
+                    out_global_store_tuple(I).MoveDstSliceWindow(out_grid_2d_desc_tuple[I],
+                                                                 make_multi_index(0, loop_step_n));
+                });
+
+            } while(--num_iter_n);
+
+            static_for<0, NumInput, 1>{}([&](auto I) {
+                in_global_load_tuple(I).MoveSrcSliceWindow(
+                    in_grid_2d_desc_tuple[I],
+                    make_multi_index(loop_step_m, -(N / loop_step_n) * loop_step_n));
+            });
+
+            static_for<0, NumOutput, 1>{}([&](auto I) {
+                out_global_store_tuple(I).MoveDstSliceWindow(
+                    out_grid_2d_desc_tuple[I],
+                    make_multi_index(loop_step_m, -(N / loop_step_n) * loop_step_n));
+            });
+        } while(--num_iter_m);
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp
new file mode 100644
index 0000000..989f7f1
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp
@@ -0,0 +1,500 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// X = Elementwise(input1, input2, input3, ...)
+// Y = Normalization(X, beta, gamma)
+template <typename InDataTypePointerTuple,
+          typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename XElementwiseOperation,
+          typename YElementwiseOperation,
+          typename InGrid2dDescTuple,
+          typename GridDesc_M_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorDim,
+          index_t YDstVectorSize,
+          bool SweepOnce>
+struct GridwiseElementwiseLayernormWelfordVariance_mk_to_mk
+{
+    static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((YDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr index_t NumInput = InDataTypePointerTuple::Size();
+
+    static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr index_t M_BlockTileSize     = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize     = KThreadClusterSize * KThreadSliceSize;
+    static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
+
+    static constexpr auto XThreadBufferNumber     = Number<KThreadSliceSize / XSrcVectorSize>{};
+    static constexpr auto GammaThreadBufferNumber = Number<KThreadSliceSize / GammaSrcVectorSize>{};
+    static constexpr auto BetaThreadBufferNumber  = Number<KThreadSliceSize / BetaSrcVectorSize>{};
+    static constexpr auto YThreadBufferNumber     = Number<KThreadSliceSize / YDstVectorSize>{};
+
+    __device__ static int GetKPerThread(const GridDesc_M_K& x_grid_desc_m_k,
+                                        int thread_k_cluster_id)
+    {
+        int kPerBlock = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
+        int kPerThread =
+            kPerBlock < K_BlockTileSize ? 0 : KThreadSliceSize * (kPerBlock / K_BlockTileSize);
+        int kPerBlockTail = kPerBlock - kPerThread * KThreadClusterSize;
+
+        if(kPerBlockTail > 0)
+        {
+            static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
+                int thread_max_len =
+                    (thread_k_cluster_id + 1) * XSrcVectorSize + K_BlockTileStepSize * i;
+                int delta = thread_max_len - kPerBlockTail;
+                delta     = math::clamp(thread_max_len - kPerBlockTail, 0, XSrcVectorSize);
+                kPerThread += XSrcVectorSize - delta;
+            });
+        }
+
+        return kPerThread;
+    }
+
+    __device__ static void Run(const InGrid2dDescTuple in_grid_2d_desc_tuple,
+                               const GridDesc_M_K& x_grid_desc_m_k,
+                               const GridDesc_M_K& gamma_grid_desc_m_k,
+                               const GridDesc_M_K& beta_grid_desc_m_k,
+                               const GridDesc_M_K& y_grid_desc_m_k,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               const InDataTypePointerTuple p_in_global_tuple,
+                               XDataType* const __restrict__ p_x_lds,
+                               const GammaDataType* const __restrict__ p_gamma_global,
+                               const BetaDataType* const __restrict__ p_beta_global,
+                               YDataType* const __restrict__ p_y_global,
+                               const XElementwiseOperation x_elementwise_op,
+                               const YElementwiseOperation y_elementwise_op)
+    {
+        if constexpr(SweepOnce)
+        {
+            num_k_block_tile_iteration = 1;
+        }
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t grid_size       = get_grid_size();
+
+        auto in_global_buf_tuple = generate_tuple(
+            [&](auto I) {
+                static_assert(in_grid_2d_desc_tuple[I].GetNumOfDimension() ==
+                              2); // matrix dimension
+
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_in_global_tuple[I], in_grid_2d_desc_tuple[I].GetElementSpaceSize());
+            },
+            Number<NumInput>{});
+
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
+
+        auto x_lds_val_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_x_lds, x_grid_desc_m_k.GetElementSpaceSize() / grid_size);
+
+        auto in_thread_buf_tuple = generate_tuple(
+            [&](auto) {
+                return generate_tuple(
+                    [&](auto) {
+                        return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                            AccDataType,
+                                            MThreadSliceSize * XSrcVectorSize,
+                                            true>{};
+                    },
+                    Number<NumInput>{});
+            },
+            Number<XThreadBufferNumber>{});
+
+        auto x_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * XSrcVectorSize,
+                                    true>{};
+            },
+            Number<XThreadBufferNumber>{});
+
+        auto gamma_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * GammaSrcVectorSize,
+                                    true>{};
+            },
+            Number<GammaThreadBufferNumber>{});
+
+        auto beta_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * BetaSrcVectorSize,
+                                    true>{};
+            },
+            Number<BetaThreadBufferNumber>{});
+
+        auto y_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * YDstVectorSize,
+                                    true>{};
+            },
+            Number<YThreadBufferNumber>{});
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, XSrcVectorSize>;
+
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
+
+        auto in_global_load_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(InDataTypePointerTuple{}[I])>;
+                using DataType        = remove_cv_t<remove_pointer_t<DataTypePointer>>;
+
+                return ThreadwiseTensorSliceTransfer_v2<DataType,
+                                                        AccDataType,
+                                                        decltype(in_grid_2d_desc_tuple[I]),
+                                                        decltype(thread_buffer_desc_m_k),
+                                                        ThreadBufferLengths_M_K,
+                                                        ThreadBufferDimAccessOrder,
+                                                        XSrcVectorDim,
+                                                        XSrcVectorSize,
+                                                        1,
+                                                        false>{
+                    in_grid_2d_desc_tuple[I],
+                    make_multi_index(block_global_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize,
+                                     thread_k_cluster_id * XSrcVectorSize)};
+            },
+            Number<NumInput>{});
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  GridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * XSrcVectorSize));
+
+        auto threadwise_gamma_load =
+            ThreadwiseTensorSliceTransfer_v2<GammaDataType,
+                                             AccDataType,
+                                             GridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             GammaSrcVectorDim,
+                                             GammaSrcVectorSize,
+                                             1,
+                                             true>(
+                gamma_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * GammaSrcVectorSize));
+
+        auto threadwise_beta_load =
+            ThreadwiseTensorSliceTransfer_v2<BetaDataType,
+                                             AccDataType,
+                                             GridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             BetaSrcVectorDim,
+                                             BetaSrcVectorSize,
+                                             1,
+                                             true>(
+                beta_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * BetaSrcVectorSize));
+
+        using PassThrough = tensor_operation::element_wise::PassThrough;
+        PassThrough pass_through_op;
+        auto threadwise_x_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               XDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               GridDesc_M_K,
+                                               PassThrough,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               XSrcVectorDim,
+                                               XSrcVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                x_grid_desc_m_k,
+                make_multi_index(thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * XSrcVectorSize),
+                pass_through_op);
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               GridDesc_M_K,
+                                               YElementwiseOperation,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               YDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                y_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * YDstVectorSize),
+                y_elementwise_op);
+
+        // Copy x from Cache
+        // one pass: fwd, second pass: bwd
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
+        constexpr auto thread_copy_bwd_step_m_k =
+            make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
+
+        const auto gamma_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_gamma_global, gamma_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto beta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_beta_global, beta_grid_desc_m_k.GetElementSpaceSize());
+
+        auto threadwise_welford       = ThreadwiseWelford();
+        threadwise_welford.max_count_ = GetKPerThread(x_grid_desc_m_k, thread_k_cluster_id);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
+            var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
+                static_for<0, NumInput, 1>{}([&](auto I) { // input load loop
+                    in_global_load_tuple(I).Run(in_grid_2d_desc_tuple[I],
+                                                in_global_buf_tuple[I],
+                                                thread_buffer_desc_m_k,
+                                                make_tuple(I0, I0),
+                                                in_thread_buf_tuple(iK0)(I));
+
+                    in_global_load_tuple(I).MoveSrcSliceWindow(in_grid_2d_desc_tuple[I],
+                                                               thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) { // input add loop
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // get reference to in data
+                        const auto in_data_refs = generate_tie(
+                            // return type should be lvalue
+                            [&](auto I) -> const auto& {
+                                return in_thread_buf_tuple(iK0)(I)(Number<offset_m_k>{});
+                            },
+                            Number<NumInput>{});
+
+                        // get reference to dst data
+                        auto out_data_refs = generate_tie(
+                            // return type should be lvalue
+                            [&](auto) -> auto& { return x_thread_buf(iK0)(Number<offset_m_k>{}); },
+                            I1);
+
+                        unpack2(x_elementwise_op, out_data_refs, in_data_refs);
+                    });
+                });
+                threadwise_welford.Run(x_thread_buf[iK0], mean_thread_buf, var_thread_buf);
+
+                if constexpr(!SweepOnce)
+                {
+                    threadwise_x_store.Run(thread_buffer_desc_m_k,
+                                           make_tuple(I0, I0),
+                                           x_thread_buf(iK0),
+                                           x_grid_desc_m_k,
+                                           x_lds_val_buf);
+                    threadwise_x_store.MoveDstSliceWindow(x_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+                }
+            });
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            int count = threadwise_welford.cur_count_;
+            BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
+        });
+
+        auto thread_copy_tail_m_k =
+            (num_k_block_tile_iteration - 1) * XThreadBufferNumber * thread_copy_fwd_step_m_k;
+
+        if constexpr(!SweepOnce)
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_tail_m_k);
+        threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
+        threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
+        threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            if constexpr(!SweepOnce)
+            {
+                static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_lds_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                });
+            }
+
+            static_for<0, GammaThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                          gamma_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          gamma_thread_buf(i));
+                threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                         thread_copy_fwd_step_m_k);
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                auto divisor = 1 / __builtin_amdgcn_sqrtf(var_thread_buf(iM) + epsilon);
+                static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // normalize
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                            divisor;
+
+                        // gamma
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                            gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
+
+            static_for<0, BetaThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                         beta_global_val_buf,
+                                         thread_buffer_desc_m_k,
+                                         make_tuple(I0, I0),
+                                         beta_thread_buf(i));
+                threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                        thread_copy_fwd_step_m_k);
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // beta
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                            beta_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
+
+            static_for<0, YThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                       make_tuple(I0, I0),
+                                       y_thread_buf(i),
+                                       y_grid_desc_m_k,
+                                       y_global_val_buf);
+                threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_fwd_step_m_k);
+            });
+
+            if constexpr(!SweepOnce)
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
+            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                     2 * thread_copy_bwd_step_m_k);
+            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                    2 * thread_copy_bwd_step_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000..16ba232
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
@@ -0,0 +1,997 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename FloatC0,
+          typename FloatC1,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename C1ElementwiseOperation,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename ReduceGridDescriptor_MBlock_MPerBlock,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_bias_add_reduce_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const FloatC0* __restrict__ p_bias_grid,
+            const FloatC1* __restrict__ p_d0_grid,
+            ReducePtrsGlobal p_reduces_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const C1ElementwiseOperation c1_element_op,
+            const ReduceInElementwiseOperations reduce_in_element_ops,
+            const ReduceAccElementwiseOperations reduce_out_element_ops,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c0_grid_desc_mblock_mperblock_nblock_nperblock,
+            const C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c1_grid_desc_mblock_mperblock_nblock_nperblock,
+            const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_bias_grid,
+                                                  p_d0_grid,
+                                                  p_reduces_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  c1_element_op,
+                                                  reduce_in_element_ops,
+                                                  reduce_out_element_ops,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  c1_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  reduce_grid_desc_mblock_mperblock,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_bias_grid;
+    ignore = p_d0_grid;
+    ignore = p_reduces_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = c1_element_op;
+    ignore = reduce_in_element_ops;
+    ignore = reduce_out_element_ops;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = c0_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = c1_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = reduce_grid_desc_mblock_mperblock;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename FloatC0,
+          typename FloatC1,
+          typename FloatReduceAcc,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename C1ElementwiseOperation,
+          typename ReduceOperations,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename ReduceGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          typename C0GridDesc_M_N,
+          typename C1GridDesc_M_N,
+          typename ReduceGridDesc_M,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+          index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmBiasAddReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        // static_assert(is_known_at_compile_time<remove_cv_t<decltype(AK1)>>::value &&
+        //               is_known_at_compile_time<remove_cv_t<decltype(BK1)>>::value,
+        //               "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    template <typename CGridDesc_M_N_>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N_& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeReduceGridDescriptor_MBlock_MPerBlock(const ReduceGridDesc_M& d_grid_desc_m)
+    {
+        const auto M      = d_grid_desc_m.GetLength(I0);
+        const auto MBlock = M / MPerBlock;
+
+        const auto reduce_grid_desc_mblock_mperblock = transform_tensor_descriptor(
+            d_grid_desc_m,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{}))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1>{}));
+
+        return reduce_grid_desc_mblock_mperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+
+    using C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(C0GridDesc_M_N{}))>;
+
+    using C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(C1GridDesc_M_N{}))>;
+
+    using ReduceGridDescriptor_MBlock_MPerBlock =
+        remove_cvref_t<decltype(MakeReduceGridDescriptor_MBlock_MPerBlock(ReduceGridDesc_M{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        const FloatC0* __restrict__ p_bias_grid,
+        const FloatC1* __restrict__ p_d0_grid,
+        ReducePtrsGlobal p_reduces_grid,
+        void* __restrict__ p_shared,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const C1ElementwiseOperation& c1_element_op,
+        const ReduceInElementwiseOperations& reduce_in_element_ops,
+        const ReduceAccElementwiseOperations& reduce_out_element_ops,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            c_grid_desc_mblock_mperblock_nblock_nperblock,
+        const C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            c0_grid_desc_mblock_mperblock_nblock_nperblock,
+        const C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            c1_grid_desc_mblock_mperblock_nblock_nperblock,
+        const ReduceGridDescriptor_MBlock_MPerBlock& reduce_grid_desc_mblock_mperblock,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        auto c0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_bias_grid, c0_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        auto c1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_d0_grid, c1_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C + reduction + write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            // TODO: this should be implemented as a blockwise reduction
+            // LDS c_reduce_block_desc_mperblock_nperblock
+            constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
+
+            static_assert(CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) *
+                                  CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                              BlockSize,
+                          "wrong!");
+
+            static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) ==
+                                  0 &&
+                              (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                                  0,
+                          "wrong!");
+
+            constexpr index_t mreduce_per_thread =
+                (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0);
+
+            constexpr index_t nreduce_per_thread =
+                (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1);
+
+            constexpr auto c_reduce_thread_lengths_mperblock_nperblock =
+                Sequence<mreduce_per_thread, nreduce_per_thread>{};
+
+            // VGPR c_reduce_thread_desc_mperblock_nperblock
+            constexpr auto c_reduce_thread_desc_mperblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(Number<mreduce_per_thread>{}, Number<nreduce_per_thread>{}));
+
+            // VGPR reduce_thread_desc_mperblock
+            constexpr auto reduce_thread_desc_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(Number<mreduce_per_thread>{}));
+
+            // VGPR reduce_thread_desc_mblock_mperblock
+            constexpr auto reduce_thread_desc_mblock_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(I1, Number<mreduce_per_thread>{}));
+
+            auto c_reduce_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                c_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
+
+            // reduce: threadwise copy from LDS to VGPR
+            constexpr auto c_reduce_thread_cluster_desc = make_cluster_descriptor(
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock{}, Sequence<1, 0>{});
+
+            const auto c_reduce_thread_cluster_idx =
+                c_reduce_thread_cluster_desc.CalculateBottomIndex(
+                    make_multi_index(get_thread_local_1d_id()));
+
+            const auto c_reduce_thread_data_idx_begin =
+                c_reduce_thread_cluster_idx * c_reduce_thread_lengths_mperblock_nperblock;
+
+            auto c_reduce_thread_copy_lds_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatCShuffle,
+                FloatReduceAcc,
+                decltype(c_reduce_block_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_lengths_mperblock_nperblock),
+                Sequence<0, 1>,
+                1,
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>{c_reduce_block_desc_mperblock_nperblock, c_reduce_thread_data_idx_begin};
+
+            auto reduce_tuple_thread_copy_vgpr_to_global = generate_tuple(
+                [&](auto I) {
+                    auto p_reduce_grid         = p_reduces_grid[I];
+                    auto reduce_acc_element_op = reduce_out_element_ops[I];
+
+                    return ThreadwiseTensorSliceTransfer_v1r3<
+                        FloatReduceAcc,
+                        remove_pointer_t<decltype(p_reduce_grid)>,
+                        decltype(reduce_thread_desc_mblock_mperblock),
+                        decltype(reduce_grid_desc_mblock_mperblock),
+                        decltype(reduce_acc_element_op),
+                        Sequence<1, mreduce_per_thread>,
+                        Sequence<0, 1>,
+                        1,
+                        CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+                        ReduceGlobalMemoryDataOperation::At(I),
+                        1,
+                        false>{reduce_grid_desc_mblock_mperblock,
+                               make_multi_index(block_work_idx[I0],                  // mblock
+                                                c_reduce_thread_data_idx_begin[I0]), // mperblock
+                               reduce_acc_element_op};
+                },
+                Number<p_reduces_grid.Size()>{});
+
+            // c0 and c1
+            constexpr auto c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mreduce_per_thread>{}, I1, Number<nreduce_per_thread>{}));
+
+            constexpr auto c1_reduce_thread_desc_mblock_mperblock_nblock_nperblock =
+                c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock;
+
+            auto c01_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            auto c0_thread_copy_global_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatC0,
+                FloatReduceAcc,
+                decltype(c0_grid_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>(
+                c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]));
+
+            auto c1_thread_copy_global_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatC1,
+                FloatReduceAcc,
+                decltype(c1_grid_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c1_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>(
+                c1_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]));
+
+            constexpr auto c_reduce_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mreduce_per_thread>{}, I1, Number<nreduce_per_thread>{}));
+
+            auto c_reduce_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+                FloatReduceAcc,
+                FloatC,
+                decltype(c_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                tensor_operation::element_wise::PassThrough,
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>, // SliceLengths
+                Sequence<0, 1, 2, 3>,                                     // DimAccessOrder
+                3,                                                        // DstVectorDim
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]),
+                tensor_operation::element_wise::PassThrough{}};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+                {
+                    c_reduce_thread_copy_lds_to_vgpr.Run(c_reduce_block_desc_mperblock_nperblock,
+                                                         c_shuffle_block_buf,
+                                                         c_reduce_thread_desc_mperblock_nperblock,
+                                                         make_tuple(I0, I0),
+                                                         c_reduce_thread_buf);
+
+                    c0_thread_copy_global_to_vgpr.Run(
+                        c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c0_grid_buf,
+                        c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c01_thread_buf);
+
+                    // c = activation(c + bias)
+                    static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                        [&](auto i) {
+                            FloatReduceAcc out;
+                            c_element_op(out, c_reduce_thread_buf(i) + c01_thread_buf(i));
+                            c_reduce_thread_buf(i) = out;
+                        });
+
+                    c1_thread_copy_global_to_vgpr.Run(
+                        c1_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c1_grid_buf,
+                        c1_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c01_thread_buf);
+
+                    // c = c + c1_functior(c1)
+                    static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                        [&](auto i) {
+                            c1_element_op(c01_thread_buf(i), c01_thread_buf(i));
+                            c_reduce_thread_buf(i) += c01_thread_buf(i);
+                        });
+
+                    c_reduce_thread_copy_vgpr_to_global.Run(
+                        c_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c_reduce_thread_buf,
+                        c_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c_grid_buf);
+
+                    static_for<0, p_reduces_grid.Size(), 1>{}([&](auto In) {
+                        auto& p_reduce_grid = p_reduces_grid[In];
+
+                        auto reduce_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                            p_reduce_grid, reduce_grid_desc_mblock_mperblock.GetElementSpaceSize());
+
+                        auto reduce_thread_buf =
+                            make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                                reduce_thread_desc_mperblock.GetElementSpaceSize());
+
+                        auto& reduce_in_element_op = reduce_in_element_ops[In];
+
+                        auto& reduce_thread_copy_vgpr_to_global =
+                            reduce_tuple_thread_copy_vgpr_to_global(In);
+
+                        using ReduceOperation = remove_cvref_t<decltype(ReduceOperations{}[In])>;
+                        using ThreadwiseReduce =
+                            ThreadwiseReduction<FloatReduceAcc,
+                                                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                                                decltype(reduce_thread_desc_mperblock),
+                                                ReduceOperation,
+                                                false>;
+
+                        // Global write Gemm shuffle + reduction
+                        const auto reduce_identityVal =
+                            ReduceOperation::template GetIdentityValue<FloatReduceAcc>();
+
+                        static_for<0, mreduce_per_thread, 1>{}(
+                            [&](auto I) { reduce_thread_buf(I) = reduce_identityVal; });
+
+                        // reduce in VGPR
+                        static_for<0, mreduce_per_thread, 1>{}([&](auto im) {
+                            static_for<0, nreduce_per_thread, 1>{}([&](auto in) {
+                                constexpr auto offset =
+                                    Number<c_reduce_thread_desc_mperblock_nperblock.CalculateOffset(
+                                        make_tuple(im, in))>{};
+
+                                reduce_in_element_op(c_reduce_thread_buf(offset),
+                                                     c_reduce_thread_buf(offset));
+                            });
+                        });
+
+                        ThreadwiseReduce::Reduce(c_reduce_thread_buf, reduce_thread_buf);
+
+                        // copy from VGPR to Global
+                        reduce_thread_copy_vgpr_to_global.Run(reduce_thread_desc_mblock_mperblock,
+                                                              make_tuple(I0, I0),
+                                                              reduce_thread_buf,
+                                                              reduce_grid_desc_mblock_mperblock,
+                                                              reduce_grid_buf);
+
+                        if constexpr(access_id < num_access - 1)
+                        {
+                            constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+                            reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                                reduce_grid_desc_mblock_mperblock,
+                                make_tuple(c_global_step[I0], c_global_step[I1]));
+                        }
+                    });
+                }
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+
+                    // move on C0
+                    c0_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                        c0_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+
+                    // move on C1
+                    c1_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                        c1_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        } // Reduction
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp
new file mode 100644
index 0000000..25bb714
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp
@@ -0,0 +1,685 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename DsDataType,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M_N,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t K1Value,
+          index_t M1PerThreadM111,
+          index_t N1PerThreadN111,
+          index_t KPerThread,
+          typename M11N11ThreadClusterM110Xs,
+          typename M11N11ThreadClusterN110Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct GridwiseGemmDlMultipleD_km_kn_mn
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = K1;
+
+        // TODO: check alignment
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k_m = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k_n = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size =
+            math::integer_least_multiple(a_block_desc_k_m.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size =
+            math::integer_least_multiple(b_block_desc_k_n.GetElementSpaceSize(), max_lds_align);
+
+        return 2 * (a_block_aligned_space_size + b_block_aligned_space_size) * sizeof(FloatAB);
+    }
+
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+
+        bool ret=(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+                K0 == b_grid_desc_k0_n_k1.GetLength(I0) &&
+                K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+                K1 == b_grid_desc_k0_n_k1.GetLength(I2)) &&
+               (M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0);
+        if(!ret){
+            std::cout<<"M="<<M<<" N="<<N<<" K0="<<K0<<" c_grid_desc_m_n[0]="<<c_grid_desc_m_n.GetLength(I0)<<" c_grid_desc_m_n[1]="<<c_grid_desc_m_n.GetLength(I1)
+                <<" b_grid_desc_k0_n_k1[0]="<<b_grid_desc_k0_n_k1.GetLength(I0)<<" b_grid_desc_k0_n_k1[2]="<<b_grid_desc_k0_n_k1.GetLength(I2)
+                <<" a_grid_desc_k0_m_k1[2]="<<a_grid_desc_k0_m_k1.GetLength(I2)
+                <<" K1="<<K1<<" MPerBlock="<<MPerBlock<<" NPerBlock="<<NPerBlock<<" K0PerBlock="<<K0PerBlock<<std::endl;
+        }
+        return ret;
+    }
+
+    __host__ __device__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
+    {
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+
+        return grid_size;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K0)
+    {
+        const bool has_main_k_block_loop = (K0 + K0PerBlock) / (2 * K0PerBlock) > 1;
+
+        return has_main_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasDoubleTailKBlockLoop(index_t K0)
+    {
+        const bool has_double_tail_k_block_loop = (K0 / K0PerBlock) % 2 == 0;
+
+        return has_double_tail_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeAGridDescriptor_K0_M0_M1_K1(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1)
+    {
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+
+        const auto M1 = Number<MPerBlock>{};
+        const auto M0 = M / M1;
+
+        const auto a_grid_desc_k0_m0_m1_k1 =
+            transform_tensor_descriptor(a_grid_desc_k0_m_k1,
+                                        make_tuple(make_pass_through_transform(K0),
+                                                   make_unmerge_transform(make_tuple(M0, M1)),
+                                                   make_pass_through_transform(K1)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return a_grid_desc_k0_m0_m1_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBGridDescriptor_K0_N0_N1_K1(const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1)
+    {
+        const auto K0 = b_grid_desc_k0_n_k1.GetLength(I0);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+
+        const auto N1 = Number<NPerBlock>{};
+        const auto N0 = N / N1;
+
+        const auto b_grid_desc_k0_n0_n1_k1 =
+            transform_tensor_descriptor(b_grid_desc_k0_n_k1,
+                                        make_tuple(make_pass_through_transform(K0),
+                                                   make_unmerge_transform(make_tuple(N0, N1)),
+                                                   make_pass_through_transform(K1)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return b_grid_desc_k0_n0_n1_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        constexpr auto M11 =
+            Number<container_reduce(M11N11ThreadClusterM110Xs{}, math::multiplies{}, I1) *
+                   M1PerThreadM111>{};
+        constexpr auto N11 =
+            Number<container_reduce(M11N11ThreadClusterN110Xs{}, math::multiplies{}, I1) *
+                   N1PerThreadN111>{};
+
+        constexpr auto M10 = M1 / M11;
+        constexpr auto N10 = N1 / N11;
+
+        const auto c_grid_desc_m0_m10_m11_n0_n10_n11 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M0, M10, M11)),
+                       make_unmerge_transform(make_tuple(N0, N10, N11))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return c_grid_desc_m0_m10_m11_n0_n10_n11;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(const DsGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) { return MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(ds_grid_desc_m_n[i]); },
+            Number<NumDTensor>{});
+    }
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N00_M01_N01<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using AGridDesc_K0_M0_M1_K1 = decltype(MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
+    using BGridDesc_K0_N0_N1_K1 = decltype(MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+    using Block2CTileMap = decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}));
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <typename DsGridDesc_M0_M10_M11_N0_N10_N11,
+              bool HasMainKBlockLoop,
+              bool HasDoubleTailKBlockLoop>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        DsGridPointer p_ds_grid,
+        FloatC* __restrict__ p_c_grid,
+        FloatAB* __restrict__ p_shared_block,
+        const AElementwiseOperation&,
+        const BElementwiseOperation&,
+        const CDEElementwiseOperation& cde_element_op,
+        const AGridDesc_K0_M0_M1_K1& a_grid_desc_k0_m0_m1_k1,
+        const BGridDesc_K0_N0_N1_K1& b_grid_desc_k0_n0_n1_k1,
+        const DsGridDesc_M0_M10_M11_N0_N10_N11& ds_grid_desc_m0_m10_m11_n0_n10_n11,
+        const CGridDesc_M0_M10_M11_N0_N10_N11& c_grid_desc_m0_m10_m11_n0_n10_n11,
+        const Block2CTileMap& block_2_ctile_map,
+        integral_constant<bool, HasMainKBlockLoop>,
+        integral_constant<bool, HasDoubleTailKBlockLoop>)
+    {
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n0_n1_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_m0_m10_m11_n0_n10_n11.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto c_m0_n0_block_cluster_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force index data into SGPR
+        const index_t im0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I0]);
+        const index_t in0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I1]);
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               make_tuple(im0, in0),
+               make_tuple(c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I0),
+                          c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I3))))
+        {
+            return;
+        }
+
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = K1;
+
+        // TODO: check alignment
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_block_desc_k0_m0_m1_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, I1, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto b_block_desc_k0_n0_n1_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, I1, Number<NPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // A matrix in LDS memory, for blockwise GEMM
+        constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, for blockwise GEMM
+        constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+
+        static_assert(a_block_desc_k0_m0_m1_k1.GetElementSpaceSize() ==
+                          a_k0_m_k1_block_desc.GetElementSpaceSize() &&
+                      b_block_desc_k0_n0_n1_k1.GetElementSpaceSize() ==
+                          b_k0_n_k1_block_desc.GetElementSpaceSize() &&
+                      "wrong!");
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
+            BlockSize,
+            InMemoryDataOperationEnum::Set,
+            Sequence<K0PerBlock, 1, MPerBlock, K1.value>,
+            ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+            ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+            ABlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            remove_reference_t<decltype(a_grid_desc_k0_m0_m1_k1)>,
+            decltype(a_block_desc_k0_m0_m1_k1),
+            ABlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2, 3>,
+            ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1, // SrcVectorTensorLengths
+            ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1, // DstVectorTensorLengths
+            ABlockTransferSrcVectorTensorContiguousDimOrder,  // SrcVectorTensorContiguousDimOrder
+            Sequence<0, 1, 2, 3>,                             // DstVectorTensorContiguousDimOrder
+            false,
+            true>(a_grid_desc_k0_m0_m1_k1,
+                  make_multi_index(0, im0, 0, 0),
+                  a_block_desc_k0_m0_m1_k1,
+                  make_multi_index(0, 0, 0, 0));
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
+            BlockSize,
+            InMemoryDataOperationEnum::Set,
+            Sequence<K0PerBlock, 1, NPerBlock, K1.value>,
+            BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+            BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+            BBlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            remove_reference_t<decltype(b_grid_desc_k0_n0_n1_k1)>,
+            decltype(b_block_desc_k0_n0_n1_k1),
+            BBlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2, 3>,
+            BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1, // SrcVectorTensorLengths
+            BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1, // DstVectorTensorLengths
+            BBlockTransferSrcVectorTensorContiguousDimOrder,  // SrcVectorTensorContiguousDimOrder
+            Sequence<0, 1, 2, 3>,                             // DstVectorTensorContiguousDimOrder
+            false,
+            true>(b_grid_desc_k0_n0_n1_k1,
+                  make_multi_index(0, in0, 0, 0),
+                  b_block_desc_k0_n0_n1_k1,
+                  make_multi_index(0, 0, 0, 0));
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[KPerBlocl, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        const auto blockwise_gemm =
+            BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2<
+                BlockSize,
+                FloatAB,
+                FloatAB,
+                FloatAcc,
+                decltype(a_k0_m_k1_block_desc),
+                decltype(b_k0_n_k1_block_desc),
+                M1PerThreadM111,
+                N1PerThreadN111,
+                KPerThread,
+                M11N11ThreadClusterM110Xs,
+                M11N11ThreadClusterN110Xs,
+                M1PerThreadM111,
+                N1PerThreadN111>{};
+
+        constexpr auto c_m10_m11_n10_n11_thread_tensor_lengths =
+            decltype(blockwise_gemm)::GetCThreadTensorLengths_BM0_BM1_BN0_BN1();
+
+        constexpr auto c_thread_desc_m10_m11_n10_n11 = make_naive_tensor_descriptor_packed(
+            sequence_to_tuple_of_number(c_m10_m11_n10_n11_thread_tensor_lengths));
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
+            a_block_desc_k0_m0_m1_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
+            b_block_desc_k0_n0_n1_k1.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block_double = p_shared_block;
+        FloatAB* p_b_block_double = p_shared_block + 2 * a_block_aligned_space_size;
+
+        // register allocation for output
+        auto c_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAcc>(
+            c_thread_desc_m10_m11_n10_n11.GetElementSpaceSize());
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0, 0);
+
+        auto a_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double, a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        auto b_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double, b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
+
+        auto a_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double + a_block_aligned_space_size,
+            a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        auto b_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double + b_block_aligned_space_size,
+            b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
+
+        // LDS double buffer: preload data into LDS
+        {
+            a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+            b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+            a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_even_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_even_buf);
+        }
+
+        if constexpr(HasMainKBlockLoop)
+        {
+            const auto K0 = a_grid_desc_k0_m0_m1_k1.GetLength(I0);
+
+            index_t k_block_data_begin = 0;
+
+            // LDS double buffer: main body
+            // use Do-While loop instead of For loop to simplify control flow
+            do
+            {
+                // even iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1,
+                                                    a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1,
+                                                    b_block_slice_copy_step);
+
+                // LDS doubel buffer: load next data from device mem
+                a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+                b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+                block_sync_lds();
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(c_thread_desc_m10_m11_n10_n11,
+                                   a_block_even_buf,
+                                   b_block_even_buf,
+                                   c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_odd_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_odd_buf);
+
+                // odd iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1,
+                                                    a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1,
+                                                    b_block_slice_copy_step);
+
+                // LDS doubel buffer: load next data from device mem
+                a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+                b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+                block_sync_lds();
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(
+                    c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_even_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_even_buf);
+
+                k_block_data_begin += 2 * K0PerBlock;
+            } while(k_block_data_begin < K0 - 2 * K0PerBlock);
+        }
+
+        // LDS double buffer: tail
+        if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
+        {
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1, a_block_slice_copy_step);
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1, b_block_slice_copy_step);
+
+            block_sync_lds();
+
+            // LDS double buffer: load last data from device mem
+            a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+            b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+            // LDS double buffer: GEMM on 2nd-last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
+
+            // LDS double buffer: store last data to LDS
+            a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_odd_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_odd_buf);
+
+            block_sync_lds();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+        }
+        else // if has 1 iteration left
+        {
+            __syncthreads();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_m10_m11_n0_n10_n11 =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I0]>{},
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I1]>{},
+                               I1,
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I2]>{},
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I3]>{}));
+
+            const auto c_m10_m11_n10_n11_thread_origin_idx_on_block =
+                blockwise_gemm.CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
+                    get_thread_local_1d_id());
+
+            const auto ds_grid_buf = generate_tuple(
+                [&](auto i) {
+                    return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                        p_ds_grid[i], ds_grid_desc_m0_m10_m11_n0_n10_n11[i].GetElementSpaceSize());
+                },
+                Number<NumDTensor>{});
+
+            auto ds_thread_buf = generate_tuple(
+                [&](auto i) {
+                    using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                    return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                        DDataType,
+                                        c_m10_m11_n10_n11_thread_tensor_lengths[I3],
+                                        true>{};
+                },
+                Number<NumDTensor>{});
+
+            auto ds_threadwise_copy = generate_tuple(
+                [&](auto i) {
+                    using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                    return ThreadwiseTensorSliceTransfer_v2<
+                        DDataType,
+                        DDataType,
+                        decltype(ds_grid_desc_m0_m10_m11_n0_n10_n11[i]),
+                        decltype(c_thread_desc_m0_m10_m11_n0_n10_n11),
+                        Sequence<I1,
+                                 I1,
+                                 I1,
+                                 I1,
+                                 I1,
+                                 Number<c_m10_m11_n10_n11_thread_tensor_lengths[I3]>{}>,
+                        CThreadTransferSrcDstAccessOrder,
+                        CThreadTransferSrcDstVectorDim,
+                        CThreadTransferDstScalarPerVector,
+                        1,
+                        false>(ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
+                               make_multi_index(im0,
+                                                c_m10_m11_n10_n11_thread_origin_idx_on_block[I0],
+                                                c_m10_m11_n10_n11_thread_origin_idx_on_block[I1],
+                                                in0,
+                                                c_m10_m11_n10_n11_thread_origin_idx_on_block[I2],
+                                                c_m10_m11_n10_n11_thread_origin_idx_on_block[I3]));
+                },
+                Number<NumDTensor>{});
+
+            static_for<0, c_m10_m11_n10_n11_thread_tensor_lengths[I0], 1>{}([&](auto m10) {
+                static_for<0, c_m10_m11_n10_n11_thread_tensor_lengths[I1], 1>{}([&](auto m11) {
+                    static_for<0, c_m10_m11_n10_n11_thread_tensor_lengths[I2], 1>{}([&](auto n10) {
+                        // load d matrix data
+                        static_for<0, NumDTensor, 1>{}([&](auto i) {
+                            ds_threadwise_copy(i).Run(ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
+                                                      ds_grid_buf[i],
+                                                      c_thread_desc_m0_m10_m11_n0_n10_n11,
+                                                      make_tuple(I0, I0, I0, I0, I0, I0),
+                                                      ds_thread_buf(i));
+                        });
+                        // cal element op
+                        static_for<0, c_m10_m11_n10_n11_thread_tensor_lengths[I3], 1>{}(
+                            [&](auto i) {
+                                // get reference to src data
+                                const auto src_data_refs = generate_tie(
+                                    // return type should be lvalue
+                                    [&](auto iSrc) -> const auto& {
+                                        return ds_thread_buf[iSrc][i];
+                                    },
+                                    Number<NumDTensor>{});
+
+                                // get reference to dst data
+                                constexpr index_t c_offset =
+                                    c_thread_desc_m0_m10_m11_n0_n10_n11.CalculateOffset(
+                                        make_tuple(0, m10, m11, 0, n10, i));
+                                auto dst_data_refs = generate_tie(
+                                    // return type should be lvalue
+                                    [&](auto) -> auto& { return c_thread_buf(Number<c_offset>{}); },
+                                    Number<2>{});
+
+                                unpack2(cde_element_op, dst_data_refs, src_data_refs);
+                            });
+
+                        static_for<0, NumDTensor, 1>{}([&](auto i) {
+                            ds_threadwise_copy(i).MoveSrcSliceWindow(
+                                ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
+                                make_multi_index(0, 0, 0, 0, 1, 0));
+                        });
+                    });
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        ds_threadwise_copy(i).MoveSrcSliceWindow(
+                            ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
+                            make_multi_index(
+                                0, 0, 1, 0, -c_m10_m11_n10_n11_thread_tensor_lengths[I2], 0));
+                    });
+                });
+                static_for<0, NumDTensor, 1>{}([&](auto i) {
+                    ds_threadwise_copy(i).MoveSrcSliceWindow(
+                        ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
+                        make_multi_index(
+                            0, 1, -c_m10_m11_n10_n11_thread_tensor_lengths[I1], 0, 0, 0));
+                });
+            });
+
+            ThreadwiseTensorSliceTransfer_v1r3<
+                FloatAcc,
+                FloatC,
+                decltype(c_thread_desc_m0_m10_m11_n0_n10_n11),
+                decltype(c_grid_desc_m0_m10_m11_n0_n10_n11),
+                ck::tensor_operation::element_wise::PassThrough,
+                Sequence<1,
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I0],
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I1],
+                         1,
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I2],
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I3]>,
+                CThreadTransferSrcDstAccessOrder,
+                CThreadTransferSrcDstVectorDim,
+                CThreadTransferDstScalarPerVector,
+                CGlobalMemoryDataOperation,
+                1,
+                true>{c_grid_desc_m0_m10_m11_n0_n10_n11,
+                      make_multi_index(im0,
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I0],
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I1],
+                                       in0,
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I2],
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I3]),
+                      ck::tensor_operation::element_wise::PassThrough{}}
+                .Run(c_thread_desc_m0_m10_m11_n0_n10_n11,
+                     make_tuple(I0, I0, I0, I0, I0, I0),
+                     c_thread_buf,
+                     c_grid_desc_m0_m10_m11_n0_n10_n11,
+                     c_grid_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp
new file mode 100644
index 0000000..c839cde
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp
@@ -0,0 +1,577 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M0_M1_K1,
+          typename BGridDesc_K0_N0_N1_K1,
+          typename CGridDesc_M0_M10_M11_N0_N10_N11,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_dl_v1r3(const FloatAB* __restrict__ p_a_grid,
+                            const FloatAB* __restrict__ p_b_grid,
+                            FloatC* __restrict__ p_c_grid,
+                            const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1,
+                            const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1,
+                            const CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11,
+                            const Block2CTileMap block_2_ctile_map)
+{
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::Run(p_a_grid,
+                      p_b_grid,
+                      p_c_grid,
+                      p_shared_block,
+                      a_grid_desc_k0_m0_m1_k1,
+                      b_grid_desc_k0_n0_n1_k1,
+                      c_grid_desc_m0_m10_m11_n0_n10_n11,
+                      block_2_ctile_map,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M_N,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t K1Value,
+          index_t M1PerThreadM111,
+          index_t N1PerThreadN111,
+          index_t KPerThread,
+          typename M11N11ThreadClusterM110Xs,
+          typename M11N11ThreadClusterN110Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct GridwiseGemmDl_km_kn_mn_v1r3
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = K1;
+
+        // TODO: check alignment
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k_m = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k_n = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size =
+            math::integer_least_multiple(a_block_desc_k_m.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size =
+            math::integer_least_multiple(b_block_desc_k_n.GetElementSpaceSize(), max_lds_align);
+
+        return 2 * (a_block_aligned_space_size + b_block_aligned_space_size) * sizeof(FloatAB);
+    }
+
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+
+        return (M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+                K0 == b_grid_desc_k0_n_k1.GetLength(I0) &&
+                K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+                K1 == b_grid_desc_k0_n_k1.GetLength(I2)) &&
+               (M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0);
+    }
+
+    __host__ __device__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
+    {
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+
+        return grid_size;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K0)
+    {
+        const bool has_main_k_block_loop = (K0 + K0PerBlock) / (2 * K0PerBlock) > 1;
+
+        return has_main_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasDoubleTailKBlockLoop(index_t K0)
+    {
+        const bool has_double_tail_k_block_loop = (K0 / K0PerBlock) % 2 == 0;
+
+        return has_double_tail_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeAGridDescriptor_K0_M0_M1_K1(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1)
+    {
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+
+        const auto M1 = Number<MPerBlock>{};
+        const auto M0 = M / M1;
+
+        const auto a_grid_desc_k0_m0_m1_k1 =
+            transform_tensor_descriptor(a_grid_desc_k0_m_k1,
+                                        make_tuple(make_pass_through_transform(K0),
+                                                   make_unmerge_transform(make_tuple(M0, M1)),
+                                                   make_pass_through_transform(K1)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return a_grid_desc_k0_m0_m1_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBGridDescriptor_K0_N0_N1_K1(const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1)
+    {
+        const auto K0 = b_grid_desc_k0_n_k1.GetLength(I0);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+
+        const auto N1 = Number<NPerBlock>{};
+        const auto N0 = N / N1;
+
+        const auto b_grid_desc_k0_n0_n1_k1 =
+            transform_tensor_descriptor(b_grid_desc_k0_n_k1,
+                                        make_tuple(make_pass_through_transform(K0),
+                                                   make_unmerge_transform(make_tuple(N0, N1)),
+                                                   make_pass_through_transform(K1)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return b_grid_desc_k0_n0_n1_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        constexpr auto M11 =
+            Number<container_reduce(M11N11ThreadClusterM110Xs{}, math::multiplies{}, I1) *
+                   M1PerThreadM111>{};
+        constexpr auto N11 =
+            Number<container_reduce(M11N11ThreadClusterN110Xs{}, math::multiplies{}, I1) *
+                   N1PerThreadN111>{};
+
+        constexpr auto M10 = M1 / M11;
+        constexpr auto N10 = N1 / N11;
+
+        const auto c_grid_desc_m0_m10_m11_n0_n10_n11 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M0, M10, M11)),
+                       make_unmerge_transform(make_tuple(N0, N10, N11))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return c_grid_desc_m0_m10_m11_n0_n10_n11;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N00_M01_N01<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using AGridDesc_K0_M0_M1_K1 = decltype(MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
+    using BGridDesc_K0_N0_N1_K1 = decltype(MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+    using Block2CTileMap = decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}));
+
+    template <bool HasMainKBlockLoop, bool HasDoubleTailKBlockLoop>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        FloatAB* __restrict__ p_shared_block,
+        const AGridDesc_K0_M0_M1_K1& a_grid_desc_k0_m0_m1_k1,
+        const BGridDesc_K0_N0_N1_K1& b_grid_desc_k0_n0_n1_k1,
+        const CGridDesc_M0_M10_M11_N0_N10_N11& c_grid_desc_m0_m10_m11_n0_n10_n11,
+        const Block2CTileMap& block_2_ctile_map,
+        integral_constant<bool, HasMainKBlockLoop>,
+        integral_constant<bool, HasDoubleTailKBlockLoop>)
+    {
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n0_n1_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_m0_m10_m11_n0_n10_n11.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto c_m0_n0_block_cluster_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force index data into SGPR
+        const index_t im0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I0]);
+        const index_t in0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I1]);
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               make_tuple(im0, in0),
+               make_tuple(c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I0),
+                          c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I3))))
+        {
+            return;
+        }
+
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = K1;
+
+        // TODO: check alignment
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_block_desc_k0_m0_m1_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, I1, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto b_block_desc_k0_n0_n1_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, I1, Number<NPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // A matrix in LDS memory, for blockwise GEMM
+        constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, for blockwise GEMM
+        constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+
+        static_assert(a_block_desc_k0_m0_m1_k1.GetElementSpaceSize() ==
+                          a_k0_m_k1_block_desc.GetElementSpaceSize() &&
+                      b_block_desc_k0_n0_n1_k1.GetElementSpaceSize() ==
+                          b_k0_n_k1_block_desc.GetElementSpaceSize() &&
+                      "wrong!");
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
+            BlockSize,
+            InMemoryDataOperationEnum::Set,
+            Sequence<K0PerBlock, 1, MPerBlock, K1.value>,
+            ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+            ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+            ABlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            remove_reference_t<decltype(a_grid_desc_k0_m0_m1_k1)>,
+            decltype(a_block_desc_k0_m0_m1_k1),
+            ABlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2, 3>,
+            ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1, // SrcVectorTensorLengths
+            ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1, // DstVectorTensorLengths
+            ABlockTransferSrcVectorTensorContiguousDimOrder,  // SrcVectorTensorContiguousDimOrder
+            Sequence<0, 1, 2, 3>,                             // DstVectorTensorContiguousDimOrder
+            false,
+            true>(a_grid_desc_k0_m0_m1_k1,
+                  make_multi_index(0, im0, 0, 0),
+                  a_block_desc_k0_m0_m1_k1,
+                  make_multi_index(0, 0, 0, 0));
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
+            BlockSize,
+            InMemoryDataOperationEnum::Set,
+            Sequence<K0PerBlock, 1, NPerBlock, K1.value>,
+            BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+            BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+            BBlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            remove_reference_t<decltype(b_grid_desc_k0_n0_n1_k1)>,
+            decltype(b_block_desc_k0_n0_n1_k1),
+            BBlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2, 3>,
+            BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1, // SrcVectorTensorLengths
+            BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1, // DstVectorTensorLengths
+            BBlockTransferSrcVectorTensorContiguousDimOrder,  // SrcVectorTensorContiguousDimOrder
+            Sequence<0, 1, 2, 3>,                             // DstVectorTensorContiguousDimOrder
+            false,
+            true>(b_grid_desc_k0_n0_n1_k1,
+                  make_multi_index(0, in0, 0, 0),
+                  b_block_desc_k0_n0_n1_k1,
+                  make_multi_index(0, 0, 0, 0));
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[KPerBlocl, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        const auto blockwise_gemm =
+            BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2<
+                BlockSize,
+                FloatAB,
+                FloatAB,
+                FloatAcc,
+                decltype(a_k0_m_k1_block_desc),
+                decltype(b_k0_n_k1_block_desc),
+                M1PerThreadM111,
+                N1PerThreadN111,
+                KPerThread,
+                M11N11ThreadClusterM110Xs,
+                M11N11ThreadClusterN110Xs,
+                M1PerThreadM111,
+                N1PerThreadN111>{};
+
+        constexpr auto c_m10_m11_n10_n11_thread_tensor_lengths =
+            decltype(blockwise_gemm)::GetCThreadTensorLengths_BM0_BM1_BN0_BN1();
+
+        constexpr auto c_thread_desc_m10_m11_n10_n11 = make_naive_tensor_descriptor_packed(
+            sequence_to_tuple_of_number(c_m10_m11_n10_n11_thread_tensor_lengths));
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
+            a_block_desc_k0_m0_m1_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
+            b_block_desc_k0_n0_n1_k1.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block_double = p_shared_block;
+        FloatAB* p_b_block_double = p_shared_block + 2 * a_block_aligned_space_size;
+
+        // register allocation for output
+        auto c_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAcc>(
+            c_thread_desc_m10_m11_n10_n11.GetElementSpaceSize());
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0, 0);
+
+        auto a_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double, a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        auto b_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double, b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
+
+        auto a_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double + a_block_aligned_space_size,
+            a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        auto b_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double + b_block_aligned_space_size,
+            b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
+
+        // LDS double buffer: preload data into LDS
+        {
+            a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+            b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+            a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_even_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_even_buf);
+        }
+
+        if constexpr(HasMainKBlockLoop)
+        {
+            const auto K0 = a_grid_desc_k0_m0_m1_k1.GetLength(I0);
+
+            index_t k_block_data_begin = 0;
+
+            // LDS double buffer: main body
+            // use Do-While loop instead of For loop to simplify control flow
+            do
+            {
+                // even iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1,
+                                                    a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1,
+                                                    b_block_slice_copy_step);
+
+                // LDS doubel buffer: load next data from device mem
+                a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+                b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+                block_sync_lds();
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(c_thread_desc_m10_m11_n10_n11,
+                                   a_block_even_buf,
+                                   b_block_even_buf,
+                                   c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_odd_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_odd_buf);
+
+                // odd iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1,
+                                                    a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1,
+                                                    b_block_slice_copy_step);
+
+                // LDS doubel buffer: load next data from device mem
+                a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+                b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+                block_sync_lds();
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(
+                    c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_even_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_even_buf);
+
+                k_block_data_begin += 2 * K0PerBlock;
+            } while(k_block_data_begin < K0 - 2 * K0PerBlock);
+        }
+
+        // LDS double buffer: tail
+        if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
+        {
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1, a_block_slice_copy_step);
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1, b_block_slice_copy_step);
+
+            block_sync_lds();
+
+            // LDS double buffer: load last data from device mem
+            a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+            b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+            // LDS double buffer: GEMM on 2nd-last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
+
+            // LDS double buffer: store last data to LDS
+            a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_odd_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_odd_buf);
+
+            block_sync_lds();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+        }
+        else // if has 1 iteration left
+        {
+            __syncthreads();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_m10_m11_n0_n10_n11 =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I0]>{},
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I1]>{},
+                               I1,
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I2]>{},
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I3]>{}));
+
+            const auto c_m10_m11_n10_n11_thread_origin_idx_on_block =
+                blockwise_gemm.CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
+                    get_thread_local_1d_id());
+
+            ThreadwiseTensorSliceTransfer_v1r3<
+                FloatAcc,
+                FloatC,
+                decltype(c_thread_desc_m0_m10_m11_n0_n10_n11),
+                decltype(c_grid_desc_m0_m10_m11_n0_n10_n11),
+                ck::tensor_operation::element_wise::PassThrough,
+                Sequence<1,
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I0],
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I1],
+                         1,
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I2],
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I3]>,
+                CThreadTransferSrcDstAccessOrder,
+                CThreadTransferSrcDstVectorDim,
+                CThreadTransferDstScalarPerVector,
+                CGlobalMemoryDataOperation,
+                1,
+                true>{c_grid_desc_m0_m10_m11_n0_n10_n11,
+                      make_multi_index(im0,
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I0],
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I1],
+                                       in0,
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I2],
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I3]),
+                      ck::tensor_operation::element_wise::PassThrough{}}
+                .Run(c_thread_desc_m0_m10_m11_n0_n10_n11,
+                     make_tuple(I0, I0, I0, I0, I0, I0),
+                     c_thread_buf,
+                     c_grid_desc_m0_m10_m11_n0_n10_n11,
+                     c_grid_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dlops_v1r2.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dlops_v1r2.hpp
new file mode 100644
index 0000000..84e033e
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dlops_v1r2.hpp
@@ -0,0 +1,608 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_GRIDWISE_GEMM_DLOPS_V1R2_HPP
+#define CK_GRIDWISE_GEMM_DLOPS_V1R2_HPP
+
+#include "common_header.hpp"
+#include "multi_index_transform_helper.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "blockwise_gemm_dlops_v2r2.hpp"
+#include "blockwise_tensor_slice_transfer.hpp"
+#include "threadwise_tensor_slice_transfer.hpp"
+#include "threadwise_tensor_slice_set.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AKM0M1GridDesc,
+          typename BKN0N1GridDesc,
+          typename CM0M10M11N0N10N11GridDesc,
+          typename CBlockIdToM0N0BlockClusterAdaptor,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_dlops_v1r2(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AKM0M1GridDesc a_k_m0_m1_grid_desc,
+            const BKN0N1GridDesc b_k_n0_n1_grid_desc,
+            const CM0M10M11N0N10N11GridDesc c_m0_m10_m11_n0_n10_n11_grid_desc,
+            const CBlockIdToM0N0BlockClusterAdaptor cblockid_to_m0_n0_block_cluster_adaptor)
+{
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::Run(p_a_grid,
+                      p_b_grid,
+                      p_c_grid,
+                      p_shared_block,
+                      a_k_m0_m1_grid_desc,
+                      b_k_n0_n1_grid_desc,
+                      c_m0_m10_m11_n0_n10_n11_grid_desc,
+                      cblockid_to_m0_n0_block_cluster_adaptor,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AKMGridDesc,
+          typename BKNGridDesc,
+          typename CMNGridDesc,
+          index_t MPerBlockM1,
+          index_t NPerBlockN1,
+          index_t KPerBlock,
+          index_t M1PerThreadM111,
+          index_t N1PerThreadN111,
+          index_t KPerThread,
+          index_t M11N11ThreadClusterM1100,
+          index_t M11N11ThreadClusterN1100,
+          index_t M11N11ThreadClusterM1101,
+          index_t M11N11ThreadClusterN1101,
+          typename ABlockTransferThreadSliceLengths_K_M0_M1,
+          typename ABlockTransferThreadClusterLengths_K_M0_M1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_M1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          typename BBlockTransferThreadSliceLengths_K_N0_N1,
+          typename BBlockTransferThreadClusterLengths_K_N0_N1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_N1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          typename AGridStepHacks,
+          typename BGridStepHacks,
+          typename CGridStepHacks,
+          typename AGridMoveSliceWindowStepHacks,
+          typename BGridMoveSliceWindowStepHacks>
+struct GridwiseGemmDlops_km_kn_mn_v1r2
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = math::lcm(Number<ABlockTransferDstScalarPerVector_M1>{},
+                                                 Number<BBlockTransferDstScalarPerVector_N1>{},
+                                                 Number<M1PerThreadM111>{},
+                                                 Number<N1PerThreadN111>{});
+
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_k_m_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<KPerBlock>{}, Number<MPerBlockM1>{}), max_lds_align);
+
+        // B matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto b_k_n_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<KPerBlock>{}, Number<NPerBlockN1>{}), max_lds_align);
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size =
+            math::integer_least_multiple(a_k_m_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size =
+            math::integer_least_multiple(b_k_n_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        return 2 * (a_block_aligned_space_size + b_block_aligned_space_size) * sizeof(FloatAB);
+    }
+
+    __host__ __device__ static constexpr bool CheckValidity(const AKMGridDesc& a_k_m_grid_desc,
+                                                            const BKNGridDesc& b_k_n_grid_desc,
+                                                            const CMNGridDesc& c_m_n_grid_desc)
+    {
+        const auto M = a_k_m_grid_desc.GetLength(I1);
+        const auto N = b_k_n_grid_desc.GetLength(I1);
+        const auto K = a_k_m_grid_desc.GetLength(I0);
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+
+        return (M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
+                K == b_k_n_grid_desc.GetLength(I0)) &&
+               (M % MPerBlockM1 == 0 && N % NPerBlockN1 == 0 && K % KPerBlock == 0);
+    }
+
+    __host__ __device__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
+    {
+        const index_t grid_size = (M / MPerBlockM1) * (N / NPerBlockN1);
+
+        return grid_size;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const bool has_main_k_block_loop = (K + KPerBlock) / (2 * KPerBlock) > 1;
+
+        return has_main_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasDoubleTailKBlockLoop(index_t K)
+    {
+        const bool has_double_tail_k_block_loop = (K / KPerBlock) % 2 == 0;
+
+        return has_double_tail_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeAKM0M1GridDescriptor(const AKMGridDesc& a_k_m_grid_desc)
+    {
+        const auto K = a_k_m_grid_desc.GetLength(I0);
+        const auto M = a_k_m_grid_desc.GetLength(I1);
+
+        const auto M1 = Number<MPerBlockM1>{};
+        const auto M0 = M / M1;
+
+        const auto a_k_m0_m1_grid_desc = transform_tensor_descriptor(
+            a_k_m_grid_desc,
+            make_tuple(make_pass_through_transform(K), make_unmerge_transform(make_tuple(M0, M1))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}));
+
+        return a_k_m0_m1_grid_desc;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBKN0N1GridDescriptor(const BKNGridDesc& b_k_n_grid_desc)
+    {
+        const auto K = b_k_n_grid_desc.GetLength(I0);
+        const auto N = b_k_n_grid_desc.GetLength(I1);
+
+        const auto N1 = Number<NPerBlockN1>{};
+        const auto N0 = N / N1;
+
+        const auto b_k_n0_n1_grid_desc = transform_tensor_descriptor(
+            b_k_n_grid_desc,
+            make_tuple(make_pass_through_transform(K), make_unmerge_transform(make_tuple(N0, N1))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}));
+
+        return b_k_n0_n1_grid_desc;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCM0M10M11N0N10N11GridDescriptor(const CMNGridDesc& c_m_n_grid_desc)
+    {
+        const auto M = c_m_n_grid_desc.GetLength(I0);
+        const auto N = c_m_n_grid_desc.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlockM1>{};
+        constexpr auto N1 = Number<NPerBlockN1>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        constexpr auto M11 =
+            Number<M11N11ThreadClusterM1100 * M11N11ThreadClusterM1101 * M1PerThreadM111>{};
+        constexpr auto N11 =
+            Number<M11N11ThreadClusterN1100 * M11N11ThreadClusterN1101 * N1PerThreadN111>{};
+
+        constexpr auto M10 = M1 / M11;
+        constexpr auto N10 = N1 / N11;
+
+        const auto c_m0_m10_m11_n0_n10_n11_grid_desc = transform_tensor_descriptor(
+            c_m_n_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(M0, M10, M11)),
+                       make_unmerge_transform(make_tuple(N0, N10, N11))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return c_m0_m10_m11_n0_n10_n11_grid_desc;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCBlockIdToM0N0BlockClusterAdaptor(const CMNGridDesc& c_m_n_grid_desc)
+    {
+        const auto M = c_m_n_grid_desc.GetLength(I0);
+        const auto N = c_m_n_grid_desc.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlockM1>{};
+        constexpr auto N1 = Number<NPerBlockN1>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        const auto cblockid_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(make_tuple(make_merge_transform(make_tuple(M0, N0))),
+                                             make_tuple(Sequence<0, 1>{}),
+                                             make_tuple(Sequence<0>{}));
+
+        return cblockid_to_m0_n0_block_cluster_adaptor;
+    }
+
+    using AKM0M1GridDesc            = decltype(MakeAKM0M1GridDescriptor(AKMGridDesc{}));
+    using BKN0N1GridDesc            = decltype(MakeBKN0N1GridDescriptor(BKNGridDesc{}));
+    using CM0M10M11N0N10N11GridDesc = decltype(MakeCM0M10M11N0N10N11GridDescriptor(CMNGridDesc{}));
+    using CBlockIdToM0N0BlockClusterAdaptor =
+        decltype(MakeCBlockIdToM0N0BlockClusterAdaptor(CMNGridDesc{}));
+
+    template <bool HasMainKBlockLoop, bool HasDoubleTailKBlockLoop>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        FloatAB* __restrict__ p_shared_block,
+        const AKM0M1GridDesc& a_k_m0_m1_grid_desc,
+        const BKN0N1GridDesc& b_k_n0_n1_grid_desc,
+        const CM0M10M11N0N10N11GridDesc& c_m0_m10_m11_n0_n10_n11_grid_desc,
+        const CBlockIdToM0N0BlockClusterAdaptor& cblockid_to_m0_n0_block_cluster_adaptor,
+        integral_constant<bool, HasMainKBlockLoop>,
+        integral_constant<bool, HasDoubleTailKBlockLoop>)
+    {
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_k_m0_m1_grid_desc.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_k_n0_n1_grid_desc.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_m0_m10_m11_n0_n10_n11_grid_desc.GetElementSpaceSize());
+
+        const auto K = a_k_m0_m1_grid_desc.GetLength(I0);
+
+        // divide block work by [M, N]
+        const auto c_m0_n0_block_cluster_idx =
+            cblockid_to_m0_n0_block_cluster_adaptor.CalculateBottomIndex(
+                make_multi_index(get_block_1d_id()));
+
+        // HACK: this force index data into SGPR
+        const index_t im0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I0]);
+        const index_t in0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I1]);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(Number<ABlockTransferDstScalarPerVector_M1>{},
+                                                 Number<BBlockTransferDstScalarPerVector_N1>{},
+                                                 Number<M1PerThreadM111>{},
+                                                 Number<N1PerThreadN111>{});
+
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_k_m_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<KPerBlock>{}, Number<MPerBlockM1>{}), max_lds_align);
+
+        // B matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto b_k_n_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<KPerBlock>{}, Number<NPerBlockN1>{}), max_lds_align);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_k_m0_m1_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<KPerBlock>{}, I1, Number<MPerBlockM1>{}), max_lds_align);
+
+        // B matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto b_k_n0_n1_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<KPerBlock>{}, I1, Number<NPerBlockN1>{}), max_lds_align);
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            BlockwiseTensorSliceTransfer_v4<BlockSize,
+                                            InMemoryDataOperationEnum::Set,
+                                            Sequence<KPerBlock, 1, MPerBlockM1>,
+                                            ABlockTransferThreadSliceLengths_K_M0_M1,
+                                            ABlockTransferThreadClusterLengths_K_M0_M1,
+                                            ABlockTransferThreadClusterArrangeOrder,
+                                            FloatAB,
+                                            FloatAB,
+                                            decltype(a_k_m0_m1_grid_desc),
+                                            decltype(a_k_m0_m1_block_desc),
+                                            ABlockTransferSrcAccessOrder,
+                                            Sequence<0, 1, 2>,
+                                            ABlockTransferSrcVectorDim,
+                                            2,
+                                            ABlockTransferSrcScalarPerVector,
+                                            ABlockTransferDstScalarPerVector_M1,
+                                            1,
+                                            1,
+                                            AThreadTransferSrcResetCoordinateAfterRun,
+                                            true>(a_k_m0_m1_grid_desc,
+                                                  make_multi_index(0, im0, 0),
+                                                  a_k_m0_m1_block_desc,
+                                                  make_multi_index(0, 0, 0));
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            BlockwiseTensorSliceTransfer_v4<BlockSize,
+                                            InMemoryDataOperationEnum::Set,
+                                            Sequence<KPerBlock, 1, NPerBlockN1>,
+                                            BBlockTransferThreadSliceLengths_K_N0_N1,
+                                            BBlockTransferThreadClusterLengths_K_N0_N1,
+                                            BBlockTransferThreadClusterArrangeOrder,
+                                            FloatAB,
+                                            FloatAB,
+                                            decltype(b_k_n0_n1_grid_desc),
+                                            decltype(b_k_n0_n1_block_desc),
+                                            BBlockTransferSrcAccessOrder,
+                                            Sequence<0, 1, 2>,
+                                            BBlockTransferSrcVectorDim,
+                                            2,
+                                            BBlockTransferSrcScalarPerVector,
+                                            BBlockTransferDstScalarPerVector_N1,
+                                            1,
+                                            1,
+                                            BThreadTransferSrcResetCoordinateAfterRun,
+                                            true>(b_k_n0_n1_grid_desc,
+                                                  make_multi_index(0, in0, 0),
+                                                  b_k_n0_n1_block_desc,
+                                                  make_multi_index(0, 0, 0));
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[KPerBlock, MPerBlockM1] is in LDS
+        //     b_mtx[KPerBlocl, NPerBlockN1] is in LDS
+        //     c_mtx[MPerBlockM1, NPerBlockN1] is distributed among threads, and saved in
+        //       register
+        const auto blockwise_gemm =
+            BlockwiseGemmDlops_km_kn_m0m1n0n1_v2r2_pipeline_2x2<BlockSize,
+                                                                FloatAB,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_k_m_block_desc),
+                                                                decltype(b_k_n_block_desc),
+                                                                M1PerThreadM111,
+                                                                N1PerThreadN111,
+                                                                KPerThread,
+                                                                M11N11ThreadClusterM1100,
+                                                                M11N11ThreadClusterN1100,
+                                                                M11N11ThreadClusterM1101,
+                                                                M11N11ThreadClusterN1101,
+                                                                M1PerThreadM111,
+                                                                N1PerThreadN111>{};
+        constexpr auto c_m10_m11_n10_n11_thread_tensor_lengths =
+            decltype(blockwise_gemm)::GetCM0M1N0N1ThreadTensorLengths();
+
+        constexpr auto c_m10_m11_n10_n11_thread_desc = make_naive_tensor_descriptor_packed(
+            sequence_to_tuple_of_number(c_m10_m11_n10_n11_thread_tensor_lengths));
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size =
+            math::integer_least_multiple(a_k_m0_m1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size =
+            math::integer_least_multiple(b_k_n0_n1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block_double = p_shared_block;
+        FloatAB* p_b_block_double = p_shared_block + 2 * a_block_aligned_space_size;
+
+        // register allocation for output
+        auto c_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAcc>(
+            c_m10_m11_n10_n11_thread_desc.GetElementSpaceSize());
+
+        ThreadwiseTensorSliceSet_v1<FloatAcc,
+                                    decltype(c_m10_m11_n10_n11_thread_desc),
+                                    decltype(c_m10_m11_n10_n11_thread_tensor_lengths)>{}
+            .Run(c_m10_m11_n10_n11_thread_desc,
+                 make_tuple(I0, I0, I0, I0),
+                 c_thread_buf,
+                 FloatAcc{0});
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock, 0, 0);
+
+        // hack to control index calculation when iterating over A and B matrix for threadwise copy
+        constexpr auto a_k_m0_m1_global_step_hacks = AGridStepHacks{};
+        constexpr auto b_k_n0_n1_global_step_hacks = BGridStepHacks{};
+
+        // hack to control index calculation when move slice window for A and B matrix for
+        // threadwise copy
+        constexpr auto a_k_m0_m1_global_move_slice_window_step_hack =
+            AGridMoveSliceWindowStepHacks{};
+        constexpr auto b_k_n0_n1_global_move_slice_window_step_hack =
+            BGridMoveSliceWindowStepHacks{};
+
+        auto a_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double, a_k_m0_m1_block_desc.GetElementSpaceSize());
+        auto b_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double, b_k_n0_n1_block_desc.GetElementSpaceSize());
+
+        auto a_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double + a_block_aligned_space_size,
+            a_k_m0_m1_block_desc.GetElementSpaceSize());
+        auto b_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double + b_block_aligned_space_size,
+            b_k_n0_n1_block_desc.GetElementSpaceSize());
+
+        // LDS double buffer: preload data into LDS
+        {
+            a_blockwise_copy.RunRead(
+                a_k_m0_m1_grid_desc, a_global_buf, a_k_m0_m1_global_step_hacks);
+            b_blockwise_copy.RunRead(
+                b_k_n0_n1_grid_desc, b_global_buf, b_k_n0_n1_global_step_hacks);
+
+            a_blockwise_copy.RunWrite(a_k_m0_m1_block_desc, a_block_even_buf);
+            b_blockwise_copy.RunWrite(b_k_n0_n1_block_desc, b_block_even_buf);
+        }
+
+        if constexpr(HasMainKBlockLoop)
+        {
+            index_t k_block_data_begin = 0;
+
+            // LDS double buffer: main body
+            // use Do-While loop instead of For loop to simplify control flow
+            do
+            {
+                // even iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_k_m0_m1_grid_desc,
+                                                    a_block_slice_copy_step,
+                                                    a_k_m0_m1_global_move_slice_window_step_hack);
+                b_blockwise_copy.MoveSrcSliceWindow(b_k_n0_n1_grid_desc,
+                                                    b_block_slice_copy_step,
+                                                    b_k_n0_n1_global_move_slice_window_step_hack);
+
+                __syncthreads();
+
+                // LDS doubel buffer: load next data from device mem
+                a_blockwise_copy.RunRead(
+                    a_k_m0_m1_grid_desc, a_global_buf, a_k_m0_m1_global_step_hacks);
+                b_blockwise_copy.RunRead(
+                    b_k_n0_n1_grid_desc, b_global_buf, b_k_n0_n1_global_step_hacks);
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(c_m10_m11_n10_n11_thread_desc,
+                                   a_block_even_buf,
+                                   b_block_even_buf,
+                                   c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_k_m0_m1_block_desc, a_block_odd_buf);
+                b_blockwise_copy.RunWrite(b_k_n0_n1_block_desc, b_block_odd_buf);
+
+                // odd iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_k_m0_m1_grid_desc,
+                                                    a_block_slice_copy_step,
+                                                    a_k_m0_m1_global_move_slice_window_step_hack);
+                b_blockwise_copy.MoveSrcSliceWindow(b_k_n0_n1_grid_desc,
+                                                    b_block_slice_copy_step,
+                                                    b_k_n0_n1_global_move_slice_window_step_hack);
+
+                __syncthreads();
+
+                // LDS doubel buffer: load next data from device mem
+                a_blockwise_copy.RunRead(
+                    a_k_m0_m1_grid_desc, a_global_buf, a_k_m0_m1_global_step_hacks);
+                b_blockwise_copy.RunRead(
+                    b_k_n0_n1_grid_desc, b_global_buf, b_k_n0_n1_global_step_hacks);
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(
+                    c_m10_m11_n10_n11_thread_desc, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_k_m0_m1_block_desc, a_block_even_buf);
+                b_blockwise_copy.RunWrite(b_k_n0_n1_block_desc, b_block_even_buf);
+
+                k_block_data_begin += 2 * KPerBlock;
+            } while(k_block_data_begin < K - 2 * KPerBlock);
+        }
+
+        // LDS double buffer: tail
+        if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
+        {
+            a_blockwise_copy.MoveSrcSliceWindow(a_k_m0_m1_grid_desc,
+                                                a_block_slice_copy_step,
+                                                a_k_m0_m1_global_move_slice_window_step_hack);
+            b_blockwise_copy.MoveSrcSliceWindow(b_k_n0_n1_grid_desc,
+                                                b_block_slice_copy_step,
+                                                b_k_n0_n1_global_move_slice_window_step_hack);
+
+            __syncthreads();
+
+            // LDS double buffer: load last data from device mem
+            a_blockwise_copy.RunRead(
+                a_k_m0_m1_grid_desc, a_global_buf, a_k_m0_m1_global_step_hacks);
+            b_blockwise_copy.RunRead(
+                b_k_n0_n1_grid_desc, b_global_buf, b_k_n0_n1_global_step_hacks);
+
+            // LDS double buffer: GEMM on 2nd-last data
+            blockwise_gemm.Run(
+                c_m10_m11_n10_n11_thread_desc, a_block_even_buf, b_block_even_buf, c_thread_buf);
+
+            // LDS double buffer: store last data to LDS
+            a_blockwise_copy.RunWrite(a_k_m0_m1_block_desc, a_block_odd_buf);
+            b_blockwise_copy.RunWrite(b_k_n0_n1_block_desc, b_block_odd_buf);
+
+            __syncthreads();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_m10_m11_n10_n11_thread_desc, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+        }
+        else // if has 1 iteration left
+        {
+            __syncthreads();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_m10_m11_n10_n11_thread_desc, a_block_even_buf, b_block_even_buf, c_thread_buf);
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_m0_m10_m11_n0_n10_n11_thread_desc =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I0]>{},
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I1]>{},
+                               I1,
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I2]>{},
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I3]>{}));
+
+            const auto c_m10_m11_n10_n11_thread_origin_idx_on_block =
+                blockwise_gemm.CalculateCM0M1N0N1ThreadOriginOnBlock(get_thread_local_1d_id());
+
+            ThreadwiseTensorSliceTransfer_v1r3<
+                FloatAcc,
+                FloatC,
+                decltype(c_m0_m10_m11_n0_n10_n11_thread_desc),
+                decltype(c_m0_m10_m11_n0_n10_n11_grid_desc),
+                Sequence<1,
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I0],
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I1],
+                         1,
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I2],
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I3]>,
+                CThreadTransferSrcDstAccessOrder,
+                CThreadTransferSrcDstVectorDim,
+                CThreadTransferDstScalarPerVector,
+                CGlobalMemoryDataOperation,
+                1,
+                true>{c_m0_m10_m11_n0_n10_n11_grid_desc,
+                      make_multi_index(im0,
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I0],
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I1],
+                                       in0,
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I2],
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I3])}
+                .Run(c_m0_m10_m11_n0_n10_n11_thread_desc,
+                     make_tuple(I0, I0, I0, I0, I0, I0),
+                     c_thread_buf,
+                     c_m0_m10_m11_n0_n10_n11_grid_desc,
+                     c_grid_buf,
+                     CGridStepHacks{});
+        }
+    }
+};
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dlops_v2.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dlops_v2.hpp
new file mode 100644
index 0000000..b1dfb0c
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dlops_v2.hpp
@@ -0,0 +1,461 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_GRIDWISE_GEMM_V2_HPP
+#define CK_GRIDWISE_GEMM_V2_HPP
+
+#include "common_header.hpp"
+#include "multi_index_transform_helper.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "blockwise_tensor_slice_transfer.hpp"
+#include "threadwise_tensor_slice_transfer.hpp"
+#include "blockwise_gemm_dlops_v3.hpp"
+
+namespace ck {
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGlobalDesc,
+          typename BGlobalDesc,
+          typename CGlobalDesc,
+          index_t KPerBlock,
+          index_t HoPerBlock,
+          index_t WoPerBlock,
+          index_t EPerBlock,
+          index_t KPerThread,
+          index_t HoPerThread,
+          index_t WoPerThread,
+          index_t EPerThread,
+          typename ABlockTransferThreadSliceLengths_E_K,
+          typename ABlockTransferThreadClusterLengths_E_K,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          typename AGlobalStepHacks,
+          typename BGlobalStepHacks,
+          typename CGlobalStepHacks,
+          typename AGlobalMoveSliceWindowStepHacks,
+          typename BGlobalMoveSliceWindowStepHacks>
+struct GridwiseGemmDlops_km_kn_mn_v3
+{
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto E = EPerBlock * 3 * 3;
+
+        constexpr auto max_lds_align =
+            math::lcm(Number<ABlockTransferDstScalarPerVector_K>{}, Number<KPerBlock>{});
+
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_e_k_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<E>{}, Number<KPerBlock>{}), max_lds_align);
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_e_k_desc.GetElementSpaceSize(), max_lds_align);
+
+        return a_block_space_size * sizeof(FloatAB);
+    }
+
+    template <bool HasMainKBlockLoop, bool HasDoubleTailKBlockLoop>
+    __device__ void Run(const AGlobalDesc& a_e_k_global_desc,
+                        const FloatAB* __restrict__ p_a_global,
+                        const BGlobalDesc& b_e_n_ho_wo_global_desc,
+                        const FloatAB* __restrict__ p_b_global,
+                        const CGlobalDesc& c_k_n_ho_wo_global_desc,
+                        FloatC* __restrict__ p_c_global,
+                        FloatAB* __restrict__ p_shared_block,
+                        integral_constant<bool, HasMainKBlockLoop>,
+                        integral_constant<bool, HasDoubleTailKBlockLoop>) const
+    {
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+        constexpr auto I2 = Number<2>{};
+        constexpr auto I3 = Number<3>{};
+
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_global, a_e_k_global_desc.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_global, b_e_n_ho_wo_global_desc.GetElementSpaceSize());
+        auto c_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_global, c_k_n_ho_wo_global_desc.GetElementSpaceSize());
+
+        constexpr auto E = EPerBlock * 3 * 3;
+
+        // const auto E = a_e_k_global_desc.GetLength(I0);
+        const auto K = a_e_k_global_desc.GetLength(I1);
+
+        const auto N  = b_e_n_ho_wo_global_desc.GetLength(I1);
+        const auto Ho = b_e_n_ho_wo_global_desc.GetLength(I2);
+        const auto Wo = b_e_n_ho_wo_global_desc.GetLength(I3);
+
+// divide block work by [M, N]
+#if 0
+        const auto ho_block_work_num  = Ho / Number<HoPerBlock>{};
+        const auto wo_block_work_num  = Wo / Number<WoPerBlock>{};
+        const auto hwo_block_work_num = ho_block_work_num * wo_block_work_num;
+
+        const index_t k_block_work_id   = get_block_1d_id() / hwo_block_work_num;
+        const index_t hwo_block_work_id = get_block_1d_id() - k_block_work_id * hwo_block_work_num;
+
+        const index_t ho_block_work_id = hwo_block_work_id / wo_block_work_num;
+        const index_t wo_block_work_id = hwo_block_work_id - ho_block_work_id * wo_block_work_num;
+#else
+        // Hack: this force result into SGPR
+        const index_t ho_block_work_num  = __builtin_amdgcn_readfirstlane(Ho / HoPerBlock);
+        const index_t wo_block_work_num  = __builtin_amdgcn_readfirstlane(Wo / WoPerBlock);
+        const index_t hwo_block_work_num = ho_block_work_num * wo_block_work_num;
+
+        const index_t k_block_work_id =
+            __builtin_amdgcn_readfirstlane(get_block_1d_id() / hwo_block_work_num);
+        const index_t hwo_block_work_id = get_block_1d_id() - k_block_work_id * hwo_block_work_num;
+
+        const index_t ho_block_work_id =
+            __builtin_amdgcn_readfirstlane(hwo_block_work_id / wo_block_work_num);
+        const index_t wo_block_work_id = hwo_block_work_id - ho_block_work_id * wo_block_work_num;
+#endif
+
+        // lds max alignment
+        constexpr auto max_lds_align =
+            math::lcm(Number<ABlockTransferDstScalarPerVector_K>{}, Number<KPerBlock>{});
+
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_e_k_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<EPerBlock>{}, Number<KPerBlock>{}), max_lds_align);
+
+        constexpr auto a_e_k_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<E>{}, Number<KPerBlock>{}), max_lds_align);
+
+        // B matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto b_e_n_ho_wo_block_desc = make_naive_tensor_descriptor_packed(make_tuple(
+            Number<EPerBlock>{}, Number<1>{}, Number<HoPerBlock>{}, Number<WoPerBlock>{}));
+
+        // c_thread_mtx definition: this is a mess
+        // TODO:: more elegent way of defining c_thread_mtx
+        constexpr auto c_k_n_ho_wo_thread_desc = make_naive_tensor_descriptor_packed(make_tuple(
+            Number<KPerThread>{}, Number<1>{}, Number<HoPerThread>{}, Number<WoPerThread>{}));
+
+        auto blockwise_gemm =
+            BlockwiseGemmDlops_km_kn_m0m1n0n1_v3<BlockSize,
+                                                 FloatAB,
+                                                 FloatAB,
+                                                 FloatAcc,
+                                                 decltype(a_e_k_block_desc),
+                                                 decltype(b_e_n_ho_wo_block_desc),
+                                                 decltype(c_k_n_ho_wo_thread_desc),
+                                                 KPerThread,
+                                                 HoPerThread,
+                                                 WoPerThread,
+                                                 EPerThread,
+                                                 ABlockTransferSrcScalarPerVector,
+                                                 ABlockTransferDstScalarPerVector_K>{};
+
+        auto c_thread_mtx_index = blockwise_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
+
+        const auto k_thread_id  = c_thread_mtx_index.k;
+        const auto ho_thread_id = c_thread_mtx_index.h;
+        const auto wo_thread_id = c_thread_mtx_index.w;
+
+        const index_t k_block_data_on_global  = k_block_work_id * KPerBlock;
+        const index_t ho_block_data_on_global = ho_block_work_id * HoPerBlock;
+        const index_t wo_block_data_on_global = wo_block_work_id * WoPerBlock;
+
+        const index_t ho_thread_data_on_global =
+            ho_block_data_on_global + ho_thread_id * HoPerThread;
+        const index_t wo_thread_data_on_global =
+            wo_block_data_on_global + wo_thread_id * WoPerThread;
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            BlockwiseTensorSliceTransfer_v4<BlockSize,
+                                            InMemoryDataOperationEnum::Set,
+                                            Sequence<E, KPerBlock>,
+                                            ABlockTransferThreadSliceLengths_E_K,
+                                            ABlockTransferThreadClusterLengths_E_K,
+                                            ABlockTransferThreadClusterArrangeOrder,
+                                            FloatAB,
+                                            FloatAB,
+                                            decltype(a_e_k_global_desc),
+                                            decltype(a_e_k_desc),
+                                            ABlockTransferSrcAccessOrder,
+                                            Sequence<0, 1>,
+                                            ABlockTransferSrcVectorDim,
+                                            1,
+                                            ABlockTransferSrcScalarPerVector,
+                                            ABlockTransferDstScalarPerVector_K,
+                                            1,
+                                            1,
+                                            AThreadTransferSrcResetCoordinateAfterRun,
+                                            true>(a_e_k_global_desc,
+                                                  make_multi_index(0, k_block_data_on_global),
+                                                  a_e_k_desc,
+                                                  make_multi_index(0, 0));
+
+        constexpr auto b_e_n_ho_wo_thread_desc = make_naive_tensor_descriptor_packed(make_tuple(
+            Number<EPerBlock>{}, Number<1>{}, Number<HoPerThread>{}, Number<WoPerThread>{}));
+
+        auto b_threadwise_transfer =
+            ThreadwiseTensorSliceTransfer_v2<FloatAB,
+                                             FloatAB,
+                                             decltype(b_e_n_ho_wo_global_desc),
+                                             decltype(b_e_n_ho_wo_thread_desc),
+                                             Sequence<EPerBlock, 1, HoPerThread, WoPerThread>,
+                                             BBlockTransferSrcAccessOrder,
+                                             BBlockTransferSrcVectorDim,
+                                             BBlockTransferSrcScalarPerVector,
+                                             1,
+                                             true>(
+                b_e_n_ho_wo_global_desc,
+                make_multi_index(0, 0, ho_thread_data_on_global, wo_thread_data_on_global));
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_shared_block, a_e_k_desc.GetElementSpaceSize());
+
+        // register allocation for output
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     FloatAcc,
+                     c_k_n_ho_wo_thread_desc.GetElementSpaceSize(),
+                     true>
+            c_thread_buf;
+
+        // initialize output thread tensor
+        ThreadwiseTensorSliceSet_v1<FloatAcc,
+                                    decltype(c_k_n_ho_wo_thread_desc),
+                                    Sequence<KPerThread, 1, HoPerThread, WoPerThread>>{}
+            .Run(c_k_n_ho_wo_thread_desc, make_tuple(I0, I0, I0, I0), c_thread_buf, FloatAcc{0});
+
+        constexpr auto b_thread_slice_copy_step = make_multi_index(EPerBlock, 0, 0, 0);
+
+        // hack to control index calculation when iterating over A and B matrix for threadwise copy
+        constexpr auto a_e_k_global_step_hacks       = AGlobalStepHacks{};
+        constexpr auto b_e_n_ho_wo_global_step_hacks = BGlobalStepHacks{};
+
+        // hack to control index calculation when move slice window for A and B matrix for
+        // threadwise copy
+        constexpr auto a_e_k_global_move_slice_window_step_hack = AGlobalMoveSliceWindowStepHacks{};
+        constexpr auto b_e_n_ho_wo_global_move_slice_window_step_hack =
+            BGlobalMoveSliceWindowStepHacks{};
+
+        // double regsiter buffer for b
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     FloatAB,
+                     b_e_n_ho_wo_thread_desc.GetElementSpaceSize(),
+                     true>
+            b_thread_even_buf, b_thread_odd_buf;
+
+        // LDS double buffer: preload data
+        {
+            a_blockwise_copy.RunRead(a_e_k_global_desc, a_global_buf, a_e_k_global_step_hacks);
+
+            b_threadwise_transfer.Run(b_e_n_ho_wo_global_desc,
+                                      b_global_buf,
+                                      b_e_n_ho_wo_thread_desc,
+                                      make_tuple(I0, I0, I0, I0),
+                                      b_thread_even_buf,
+                                      b_e_n_ho_wo_global_step_hacks);
+
+            a_blockwise_copy.RunWrite(a_e_k_desc, a_block_buf);
+        }
+
+        __syncthreads();
+
+        if constexpr(HasMainKBlockLoop)
+        {
+            index_t e_block_data_begin = 0;
+
+            // LDS double buffer: main body
+            // use Do-While loop instead of For loop to simplify control flow
+            do
+            {
+                // even iteration
+                b_threadwise_transfer.MoveSrcSliceWindow(b_e_n_ho_wo_global_desc,
+                                                         b_thread_slice_copy_step);
+
+                b_threadwise_transfer.Run(b_e_n_ho_wo_global_desc,
+                                          b_global_buf,
+                                          b_e_n_ho_wo_thread_desc,
+                                          make_tuple(I0, I0, I0, I0),
+                                          b_thread_odd_buf,
+                                          b_e_n_ho_wo_global_step_hacks);
+
+                // LDS double buffer: GEMM on current data
+                // TODO: @Zhang Jing: blockwise gemm should be able to move slice window
+                blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
+
+                blockwise_gemm.MoveASliceWindow(a_e_k_block_desc, make_tuple(EPerBlock, 0));
+
+                b_threadwise_transfer.MoveSrcSliceWindow(b_e_n_ho_wo_global_desc,
+                                                         b_thread_slice_copy_step);
+
+                b_threadwise_transfer.Run(b_e_n_ho_wo_global_desc,
+                                          b_global_buf,
+                                          b_e_n_ho_wo_thread_desc,
+                                          make_tuple(I0, I0, I0, I0),
+                                          b_thread_even_buf,
+                                          b_e_n_ho_wo_global_step_hacks);
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(a_block_buf, b_thread_odd_buf, c_thread_buf);
+
+                blockwise_gemm.MoveASliceWindow(a_e_k_block_desc, make_tuple(EPerBlock, 0));
+
+                e_block_data_begin += 2 * EPerBlock;
+
+            } while(e_block_data_begin < E - 2 * EPerBlock);
+        }
+
+        // LDS double buffer: tail
+        if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
+        {
+            b_threadwise_transfer.MoveSrcSliceWindow(b_e_n_ho_wo_global_desc,
+                                                     b_thread_slice_copy_step);
+
+            b_threadwise_transfer.Run(b_e_n_ho_wo_global_desc,
+                                      b_global_buf,
+                                      b_e_n_ho_wo_thread_desc,
+                                      make_tuple(I0, I0, I0, I0),
+                                      b_thread_odd_buf,
+                                      b_e_n_ho_wo_global_step_hacks);
+
+            // LDS double buffer: GEMM on 2nd-last data
+            blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
+
+            blockwise_gemm.MoveASliceWindow(a_e_k_block_desc, make_tuple(EPerBlock, 0));
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(a_block_buf, b_thread_odd_buf, c_thread_buf);
+        }
+        else // if has 1 iteration left
+        {
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
+        }
+
+        // output: register to global memory
+        {
+            // hack to control index calculation when iterating over c_k_n_ho_wo_global tensor
+            constexpr auto c_k_n_ho_wo_global_tensor_step_hacks = CGlobalStepHacks{};
+
+            const index_t k_thread_data_on_global =
+                k_block_data_on_global + k_thread_id * KPerThread;
+
+            ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                               FloatC,
+                                               decltype(c_k_n_ho_wo_thread_desc),
+                                               decltype(c_k_n_ho_wo_global_desc),
+                                               Sequence<KPerThread, 1, HoPerThread, WoPerThread>,
+                                               CThreadTransferSrcDstAccessOrder,
+                                               CThreadTransferSrcDstVectorDim,
+                                               CThreadTransferDstScalarPerVector,
+                                               CGlobalMemoryDataOperation,
+                                               1,
+                                               true>(
+                c_k_n_ho_wo_global_desc,
+                make_multi_index(
+                    k_thread_data_on_global, 0, ho_thread_data_on_global, wo_thread_data_on_global))
+                .Run(c_k_n_ho_wo_thread_desc,
+                     make_tuple(I0, I0, I0, I0),
+                     c_thread_buf,
+                     c_k_n_ho_wo_global_desc,
+                     c_global_buf,
+                     c_k_n_ho_wo_global_tensor_step_hacks);
+        }
+    }
+
+    // pass tensor descriptor by reference
+    template <bool HasMainKBlockLoop, bool HasDoubleTailKBlockLoop>
+    __device__ void Run(const AGlobalDesc& a_e_k_global_desc,
+                        const FloatAB* __restrict__ p_a_global,
+                        const BGlobalDesc& b_e_n_ho_wo_global_desc,
+                        const FloatAB* __restrict__ p_b_global,
+                        const CGlobalDesc& c_k_n_ho_wo_global_desc,
+                        FloatC* __restrict__ p_c_global,
+                        integral_constant<bool, HasMainKBlockLoop>,
+                        integral_constant<bool, HasDoubleTailKBlockLoop>) const
+    {
+        constexpr index_t shared_block_size = GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+        __shared__ FloatAB p_shared_block[shared_block_size];
+
+        Run(a_e_k_global_desc,
+            p_a_global,
+            b_e_n_ho_wo_global_desc,
+            p_b_global,
+            c_k_n_ho_wo_global_desc,
+            p_c_global,
+            p_shared_block,
+            integral_constant<bool, HasMainKBlockLoop>{},
+            integral_constant<bool, HasDoubleTailKBlockLoop>{});
+    }
+
+    // pass tensor descriptors by their pointers
+    template <bool HasMainKBlockLoop, bool HasDoubleTailKBlockLoop>
+    __device__ void Run(const AGlobalDesc* p_a_e_k_global_desc,
+                        const FloatAB* __restrict__ p_a_global,
+                        const BGlobalDesc* p_b_e_n_ho_wo_global_desc,
+                        const FloatAB* __restrict__ p_b_global,
+                        const CGlobalDesc* p_c_k_n_ho_wo_global_desc,
+                        FloatC* __restrict__ p_c_global,
+                        integral_constant<bool, HasMainKBlockLoop>,
+                        integral_constant<bool, HasDoubleTailKBlockLoop>) const
+    {
+        const auto a_e_k_global_desc       = *p_a_e_k_global_desc;
+        const auto b_e_n_ho_wo_global_desc = *p_b_e_n_ho_wo_global_desc;
+        const auto c_k_n_ho_wo_global_desc = *p_c_k_n_ho_wo_global_desc;
+
+        Run(a_e_k_global_desc,
+            p_a_global,
+            b_e_n_ho_wo_global_desc,
+            p_b_global,
+            c_k_n_ho_wo_global_desc,
+            p_c_global,
+            integral_constant<bool, HasMainKBlockLoop>{},
+            integral_constant<bool, HasDoubleTailKBlockLoop>{});
+    }
+
+    // pass tensor descriptors by void*
+    template <bool HasMainKBlockLoop, bool HasDoubleTailKBlockLoop>
+    __device__ void Run(const void* p_a_e_k_global_desc,
+                        const FloatAB* __restrict__ p_a_global,
+                        const void* p_b_e_n_ho_wo_global_desc,
+                        const FloatAB* __restrict__ p_b_global,
+                        const void* p_c_k_n_ho_wo_global_desc,
+                        FloatC* __restrict__ p_c_global,
+                        integral_constant<bool, HasMainKBlockLoop>,
+                        integral_constant<bool, HasDoubleTailKBlockLoop>) const
+    {
+        const auto a_e_k_global_desc = *reinterpret_cast<const AGlobalDesc*>(p_a_e_k_global_desc);
+        const auto b_e_n_ho_wo_global_desc =
+            *reinterpret_cast<const BGlobalDesc*>(p_b_e_n_ho_wo_global_desc);
+        const auto c_k_n_ho_wo_global_desc =
+            *reinterpret_cast<const CGlobalDesc*>(p_c_k_n_ho_wo_global_desc);
+
+        Run(a_e_k_global_desc,
+            p_a_global,
+            b_e_n_ho_wo_global_desc,
+            p_b_global,
+            c_k_n_ho_wo_global_desc,
+            p_c_global,
+            integral_constant<bool, HasMainKBlockLoop>{},
+            integral_constant<bool, HasDoubleTailKBlockLoop>{});
+    }
+};
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dlops_v3.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dlops_v3.hpp
new file mode 100644
index 0000000..ace8443
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dlops_v3.hpp
@@ -0,0 +1,1597 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_GRIDWISE_GEMM_V3_HPP
+#define CK_GRIDWISE_GEMM_V3_HPP
+
+#include "common_header.hpp"
+#include "multi_index_transform_helper.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "blockwise_tensor_slice_transfer.hpp"
+#include "threadwise_tensor_slice_transfer.hpp"
+#include "threadwise_tensor_slice_set.hpp"
+#include "blockwise_gemm_dlops_v3.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_E0_E1_K0_K1_E2,
+          typename BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2,
+          typename CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2,
+          typename CBlockIdToBlockClusterAdaptor_K_N_H_W,
+          bool HasMainE0BlockLoop,
+          ActivTypeEnum ActivType>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_dlops_v3(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            const FloatC* __restrict__ p_bias_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AGridDesc_E0_E1_K0_K1_E2 a_e0_e1_k0_k1_e2_grid_desc,
+            const BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2 b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+            const CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2 c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
+            const CBlockIdToBlockClusterAdaptor_K_N_H_W cblockid_to_k_n_h_w_block_cluster_adaptor)
+{
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::ConvBiasActiv(p_a_grid,
+                                p_b_grid,
+                                p_bias_grid,
+                                p_c_grid,
+                                p_shared_block,
+                                a_e0_e1_k0_k1_e2_grid_desc,
+                                b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                                c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
+                                cblockid_to_k_n_h_w_block_cluster_adaptor,
+                                integral_constant<bool, HasMainE0BlockLoop>{},
+                                integral_constant<ActivTypeEnum, ActivType>{});
+}
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_E0_E1_K0_K1_E2,
+          typename BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2,
+          typename CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2,
+          typename DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx,
+          typename CBlockIdToBlockClusterAdaptor_K_N_H_W,
+          bool HasMainE0BlockLoop,
+          ActivTypeEnum ActivType>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_dlops_v3_resize_add(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            const FloatC* __restrict__ p_bias_grid,
+            FloatC* __restrict__ p_d_grid,
+            const AGridDesc_E0_E1_K0_K1_E2 a_e0_e1_k0_k1_e2_grid_desc,
+            const BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2 b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+            const CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2 c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
+            const DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
+            const CBlockIdToBlockClusterAdaptor_K_N_H_W cblockid_to_k_n_h_w_block_cluster_adaptor)
+{
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::ConvBiasActivResizeAdd(p_a_grid,
+                                         p_b_grid,
+                                         p_bias_grid,
+                                         p_d_grid,
+                                         p_shared_block,
+                                         a_e0_e1_k0_k1_e2_grid_desc,
+                                         b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                                         c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
+                                         d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
+                                         cblockid_to_k_n_h_w_block_cluster_adaptor,
+                                         integral_constant<bool, HasMainE0BlockLoop>{},
+                                         integral_constant<ActivTypeEnum, ActivType>{});
+}
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_E0_E1_K0_K1_E2,
+          typename BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2,
+          typename CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2,
+          typename DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx,
+          typename CBlockIdToBlockClusterAdaptor_K_N_H_W,
+          bool HasMainE0BlockLoop,
+          ActivTypeEnum ActivType>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_dlops_v3_maxpool(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            const FloatC* __restrict__ p_bias_grid,
+            FloatC* __restrict__ p_c_grid,
+            FloatC* __restrict__ p_d_grid,
+            const AGridDesc_E0_E1_K0_K1_E2 a_e0_e1_k0_k1_e2_grid_desc,
+            const BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2 b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+            const CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2 c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
+            const DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
+            const CBlockIdToBlockClusterAdaptor_K_N_H_W cblockid_to_k_n_h_w_block_cluster_adaptor)
+{
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::ConvBiasActivMaxpool(p_a_grid,
+                                       p_b_grid,
+                                       p_bias_grid,
+                                       p_c_grid,
+                                       p_d_grid,
+                                       p_shared_block,
+                                       a_e0_e1_k0_k1_e2_grid_desc,
+                                       b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                                       c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
+                                       d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
+                                       cblockid_to_k_n_h_w_block_cluster_adaptor,
+                                       integral_constant<bool, HasMainE0BlockLoop>{},
+                                       integral_constant<ActivTypeEnum, ActivType>{});
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_E0_E1_K_E2,
+          typename BGridDesc_E0_E1_N_Ho_Wo_E2,
+          typename CGridDesc_K_N_Ho_Wo,
+          typename DGridDesc_K_N_Hx_Wx,
+          index_t E1_,
+          index_t E2_,
+          index_t K2_,
+          index_t KPerBlock,
+          index_t HoPerBlock,
+          index_t WoPerBlock,
+          index_t E1PerBlock,
+          index_t KPerThread,
+          index_t HoPerThread,
+          index_t WoPerThread,
+          index_t EPerThread,
+          typename ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2,
+          typename ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_E2,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          typename AGlobalStepHacks,
+          typename BGlobalStepHacks,
+          typename CGlobalStepHacks,
+          typename DGlobalStepHacks,
+          typename AGlobalMoveSliceWindowStepHacks,
+          typename BGlobalMoveSliceWindowStepHacks>
+struct GridwiseGemmDlops_km_kn_mn_v3
+{
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto E1 = Number<E1_>{};
+    static constexpr auto E2 = Number<E2_>{};
+    static constexpr auto K2 = Number<K2_>{};
+
+    static constexpr auto NPerBlock = I1;
+
+    static constexpr FloatAcc alpha = 0.3;
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = Number<ABlockTransferDstScalarPerVector_E2>{};
+
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_e0_e1_k1_e2_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(I1, Number<E1>{}, Number<KPerBlock>{}, Number<E2>{}), max_lds_align);
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size = math::integer_least_multiple(
+            a_e0_e1_k1_e2_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        return a_block_space_size * sizeof(FloatAB);
+    }
+
+    __host__ __device__ static constexpr index_t
+    CalculateGridSize(const CGridDesc_K_N_Ho_Wo& c_k_n_ho_wo_grid_desc)
+    {
+        const auto K  = c_k_n_ho_wo_grid_desc.GetLength(I0);
+        const auto N  = c_k_n_ho_wo_grid_desc.GetLength(I1);
+        const auto Ho = c_k_n_ho_wo_grid_desc.GetLength(I2);
+        const auto Wo = c_k_n_ho_wo_grid_desc.GetLength(I3);
+
+        const auto K0 = K / KPerBlock;
+        const auto N0 = N / NPerBlock;
+        const auto H0 = Ho / HoPerBlock;
+        const auto W0 = Wo / WoPerBlock;
+
+        const index_t grid_size = K0 * N0 * H0 * W0;
+
+        return grid_size;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainE0BlockLoop(const index_t E0)
+    {
+        const bool has_main_e0_block_loop = E0 > 1;
+
+        return has_main_e0_block_loop;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainE1BlockLoop()
+    {
+        const bool has_main_e1_block_loop = ((E1 + E1PerBlock) / (2 * E1PerBlock)) > 1;
+
+        return has_main_e1_block_loop;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasDoubleTailE1BlockLoop()
+    {
+        const bool has_double_tail_e1_block_loop = (E1 / E1PerBlock) % 2 == 0;
+
+        return has_double_tail_e1_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeAE0E1K0K1E2GridDescriptor(const AGridDesc_E0_E1_K_E2& a_e0_e1_k_e2_grid_desc)
+    {
+        const auto E0 = a_e0_e1_k_e2_grid_desc.GetLength(I0);
+        const auto K  = a_e0_e1_k_e2_grid_desc.GetLength(I2);
+
+        const auto K1 = Number<KPerBlock>{};
+        const auto K0 = K / K1;
+
+        const auto a_e0_e1_k0_k1_e2_grid_desc = transform_tensor_descriptor(
+            a_e0_e1_k_e2_grid_desc,
+            make_tuple(make_pass_through_transform(E0),
+                       make_pass_through_transform(E1),
+                       make_unmerge_transform(make_tuple(K0, K1)),
+                       make_pass_through_transform(E2)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}));
+
+        return a_e0_e1_k0_k1_e2_grid_desc;
+    }
+
+    __host__ __device__ static constexpr auto MakeBE0E1NH0H1H2W0W1W2E2GridDescriptor(
+        const BGridDesc_E0_E1_N_Ho_Wo_E2& b_e0_e1_n_ho_wo_e2_grid_desc)
+    {
+        const auto E0 = b_e0_e1_n_ho_wo_e2_grid_desc.GetLength(I0);
+        // const auto E1 = b_e0_e1_n_ho_wo_e2_grid_desc.GetLength(I1);
+        const auto N  = b_e0_e1_n_ho_wo_e2_grid_desc.GetLength(I2);
+        const auto Ho = b_e0_e1_n_ho_wo_e2_grid_desc.GetLength(I3);
+        const auto Wo = b_e0_e1_n_ho_wo_e2_grid_desc.GetLength(I4);
+        // const auto E2 = b_e0_e1_n_ho_wo_e2_grid_desc.GetLength(I5);
+
+        const auto H2 = Number<HoPerThread>{};
+        const auto H1 = Number<HoPerBlock / HoPerThread>{};
+        const auto H0 = Ho / (H1 * H2);
+
+        const auto W2 = Number<WoPerThread>{};
+        const auto W1 = Number<WoPerBlock / WoPerThread>{};
+        const auto W0 = Wo / (W1 * W2);
+
+        const auto b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc =
+            transform_tensor_descriptor(b_e0_e1_n_ho_wo_e2_grid_desc,
+                                        make_tuple(make_pass_through_transform(E0),
+                                                   make_pass_through_transform(E1),
+                                                   make_pass_through_transform(N),
+                                                   make_unmerge_transform(make_tuple(H0, H1, H2)),
+                                                   make_unmerge_transform(make_tuple(W0, W1, W2)),
+                                                   make_pass_through_transform(E2)),
+                                        make_tuple(Sequence<0>{},
+                                                   Sequence<1>{},
+                                                   Sequence<2>{},
+                                                   Sequence<3>{},
+                                                   Sequence<4>{},
+                                                   Sequence<5>{}),
+                                        make_tuple(Sequence<0>{},
+                                                   Sequence<1>{},
+                                                   Sequence<2>{},
+                                                   Sequence<3, 4, 5>{},
+                                                   Sequence<6, 7, 8>{},
+                                                   Sequence<9>{}));
+
+        return b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCK0K1NH0H1H2W0W1W2GridDescriptor(const CGridDesc_K_N_Ho_Wo& c_k_n_ho_wo_grid_desc)
+    {
+        const auto K  = c_k_n_ho_wo_grid_desc.GetLength(I0);
+        const auto N  = c_k_n_ho_wo_grid_desc.GetLength(I1);
+        const auto Ho = c_k_n_ho_wo_grid_desc.GetLength(I2);
+        const auto Wo = c_k_n_ho_wo_grid_desc.GetLength(I3);
+
+        const auto K1 = Number<KPerBlock>{};
+        const auto K0 = K / K1;
+
+        const auto H2 = Number<HoPerThread>{};
+        const auto H1 = Number<HoPerBlock / HoPerThread>{};
+        const auto H0 = Ho / (H1 * H2);
+
+        const auto W2 = Number<WoPerThread>{};
+        const auto W1 = Number<WoPerBlock / WoPerThread>{};
+        const auto W0 = Wo / (W1 * W2);
+
+        const auto c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc = transform_tensor_descriptor(
+            c_k_n_ho_wo_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                       make_pass_through_transform(N),
+                       make_unmerge_transform(make_tuple(H0, H1, H2)),
+                       make_unmerge_transform(make_tuple(W0, W1, W2))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}, Sequence<6, 7, 8>{}));
+
+        return c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeDK0K1NH0H1HxW0W1WxGridDescriptorMaxPool(const DGridDesc_K_N_Hx_Wx& d_k_n_hx_wx_grid_desc)
+    {
+        const auto K  = d_k_n_hx_wx_grid_desc.GetLength(I0);
+        const auto N  = d_k_n_hx_wx_grid_desc.GetLength(I1);
+        const auto Hx = d_k_n_hx_wx_grid_desc.GetLength(I2);
+        const auto Wx = d_k_n_hx_wx_grid_desc.GetLength(I3);
+
+        const auto K1 = Number<KPerBlock>{};
+        const auto K0 = K / K1;
+
+#if CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR
+        const auto H2 = Number<HoPerThread / 2>{};
+        const auto H1 = Number<HoPerBlock / HoPerThread>{};
+        const auto H0 = Number<Hx / (H1 * H2)>{};
+
+        const auto W2 = Number<WoPerThread / 2>{};
+        const auto W1 = Number<WoPerBlock / WoPerThread>{};
+        const auto W0 = Number<Wx / (W1 * W2)>{};
+#else
+        const auto H2 = HoPerThread / 2;
+        const auto H1 = HoPerBlock / HoPerThread;
+        const auto H0 = Hx / (H1 * H2);
+
+        const auto W2 = WoPerThread / 2;
+        const auto W1 = WoPerBlock / WoPerThread;
+        const auto W0 = Wx / (W1 * W2);
+#endif
+
+        const auto d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc = transform_tensor_descriptor(
+            d_k_n_hx_wx_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                       make_pass_through_transform(N),
+                       make_unmerge_transform(make_tuple(H0, H1, H2)),
+                       make_unmerge_transform(make_tuple(W0, W1, W2))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}, Sequence<6, 7, 8>{}));
+
+        return d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeDK0K1NH0H1HxW0W1WxGridDescriptorResizeAdd(const DGridDesc_K_N_Hx_Wx& d_k_n_hx_wx_grid_desc)
+    {
+        const auto K  = d_k_n_hx_wx_grid_desc.GetLength(I0);
+        const auto N  = d_k_n_hx_wx_grid_desc.GetLength(I1);
+        const auto Hx = d_k_n_hx_wx_grid_desc.GetLength(I2);
+        const auto Wx = d_k_n_hx_wx_grid_desc.GetLength(I3);
+
+        const auto K1 = Number<KPerBlock>{};
+        const auto K0 = K / K1;
+
+        const auto H2 = Number<HoPerThread * 2>{};
+        const auto H1 = Number<HoPerBlock / HoPerThread>{};
+
+        const auto W2 = Number<WoPerThread * 2>{};
+        const auto W1 = Number<WoPerBlock / WoPerThread>{};
+
+#if CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR
+        const auto H0 = Number<Hx / (H1 * H2)>{};
+        const auto W0 = Number<Wx / (W1 * W2)>{};
+#else
+        const auto H0 = Hx / (H1 * H2);
+        const auto W0 = Wx / (W1 * W2);
+#endif
+
+        const auto d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc = transform_tensor_descriptor(
+            d_k_n_hx_wx_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                       make_pass_through_transform(N),
+                       make_unmerge_transform(make_tuple(H0, H1, H2)),
+                       make_unmerge_transform(make_tuple(W0, W1, W2))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}, Sequence<6, 7, 8>{}));
+
+        return d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCBlockIdToKNHoWoBlockClusterAdaptor(const CGridDesc_K_N_Ho_Wo& c_k_n_ho_wo_grid_desc)
+    {
+        const auto K  = c_k_n_ho_wo_grid_desc.GetLength(I0);
+        const auto N  = c_k_n_ho_wo_grid_desc.GetLength(I1);
+        const auto Ho = c_k_n_ho_wo_grid_desc.GetLength(I2);
+        const auto Wo = c_k_n_ho_wo_grid_desc.GetLength(I3);
+
+#if CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR
+        const auto K0 = Number<K / KPerBlock>{};
+        const auto N0 = Number<N / NPerBlock>{};
+        const auto H0 = Number<Ho / HoPerBlock>{};
+        const auto W0 = Number<Wo / WoPerBlock>{};
+#else
+        const auto K0 = K / KPerBlock;
+        const auto N0 = N / NPerBlock;
+        const auto H0 = Ho / HoPerBlock;
+        const auto W0 = Wo / WoPerBlock;
+#endif
+
+        const auto cblockid_to_k_n_ho_wo_block_cluster_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(K0, N0, H0, W0))),
+            make_tuple(Sequence<0, 1, 2, 3>{}),
+            make_tuple(Sequence<0>{}));
+
+        return cblockid_to_k_n_ho_wo_block_cluster_adaptor;
+    }
+
+    // using AGridDesc_E0_E1_K0_K1_E2 =
+    // decltype(MakeAE0E1K0K1E2GridDescriptor(AGridDesc_E0_E1_K_E2{}));
+    // using BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2 =
+    // decltype(MakeBE0E1NH0H1H2W0W1W2E2GridDescriptor(BGridDesc_E0_E1_N_Ho_Wo_E2{}));
+    // using CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2 =
+    // decltype(MakeCK0K1NH0H1H2W0W1W2GridDescriptor(CGridDesc_K_N_Ho_Wo{}));
+    // using DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx =
+    // decltype(MakeDK0K1NH0H1HxW0W1WxGridDescriptor(DGridDesc_K_N_Hx_Wx{}));
+
+    using CBlockIdToBlockClusterAdaptor_K_N_H_W =
+        decltype(MakeCBlockIdToKNHoWoBlockClusterAdaptor(CGridDesc_K_N_Ho_Wo{}));
+
+    template <typename CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2>
+    __host__ __device__ static constexpr auto MakeBiasK0K1GridDescriptor(
+        const CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2& c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc)
+    {
+        const auto K0 = c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc.GetLength(I0);
+        const auto K1 = c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc.GetLength(I1);
+
+        return make_naive_tensor_descriptor_packed(make_tuple(K0, K1));
+    }
+
+    __host__ __device__ static constexpr auto MakeCK1NH2W2ThreadDescriptor()
+    {
+        constexpr auto c_k1_n_h2_w2_thread_gemm_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<KPerThread>{}, I1, Number<HoPerThread>{}, Number<WoPerThread>{}));
+        return c_k1_n_h2_w2_thread_gemm_desc;
+    }
+
+    // using CThreadDesc_K1_N_H2_W2 = decltype(MakeCK1NH2W2ThreadDescriptor());
+
+    __host__ __device__ static constexpr auto GetBlockWiseGemm()
+    {
+        constexpr auto max_lds_align = Number<ABlockTransferDstScalarPerVector_E2>{};
+
+        constexpr auto a_e1_k1_e2_block_gemm_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<E1PerBlock>{}, Number<KPerBlock>{}, Number<E2>{}), max_lds_align);
+
+        constexpr auto b_e1_n_h_w_e2_block_gemm_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<E1PerBlock>{},
+                                                           I1,
+                                                           Number<HoPerBlock>{},
+                                                           Number<WoPerBlock>{},
+                                                           Number<E2>{}));
+
+        constexpr auto c_k1_n_h2_w2_thread_gemm_desc = MakeCK1NH2W2ThreadDescriptor();
+
+        auto blockwise_gemm =
+            BlockwiseGemmDlops_km_kn_m0m1n0n1_v3<BlockSize,
+                                                 FloatAB,
+                                                 FloatAB,
+                                                 FloatAcc,
+                                                 decltype(a_e1_k1_e2_block_gemm_desc),
+                                                 decltype(b_e1_n_h_w_e2_block_gemm_desc),
+                                                 decltype(c_k1_n_h2_w2_thread_gemm_desc),
+                                                 EPerThread,
+                                                 K2>{};
+
+        return blockwise_gemm;
+    }
+
+    __device__ static constexpr auto GetCThreadIndex()
+    {
+        auto blockwise_gemm = GetBlockWiseGemm();
+        auto c_thread_mtx_index =
+            blockwise_gemm.GetBeginOfCThreadDesc_K_N_Ho_Wo(get_thread_local_1d_id());
+
+        return c_thread_mtx_index;
+    };
+
+    __device__ static constexpr auto GetCBlockIndex(
+        const CBlockIdToBlockClusterAdaptor_K_N_H_W& cblockid_to_k_n_h_w_block_cluster_adaptor)
+    {
+        const auto c_k_n_h_w_block_cluster_idx =
+            cblockid_to_k_n_h_w_block_cluster_adaptor.CalculateBottomIndex(
+                make_multi_index(get_block_1d_id()));
+        return c_k_n_h_w_block_cluster_idx;
+    }
+
+    template <typename BiasGlobalBuff,
+              typename CThreadBuff,
+              typename CBlockIndex,
+              typename CThreadIndex,
+              typename BiasGridDesc_K0_K1,
+              typename CThreadDesc_K1_N_H2_W2>
+    __device__ static void BiasOp(BiasGlobalBuff& bias_global_buf,
+                                  CThreadBuff& c_thread_buf,
+                                  const CBlockIndex& c_block_idx,
+                                  const CThreadIndex& c_thread_idx,
+                                  const BiasGridDesc_K0_K1& bias_k0_k1_grid_desc,
+                                  const CThreadDesc_K1_N_H2_W2&)
+
+    {
+        const index_t k_block_work_id = __builtin_amdgcn_readfirstlane(c_block_idx[I0]);
+
+        const auto k_thread_id = c_thread_idx[I0];
+
+        constexpr auto c_k1_n_h2_w2_thread_gemm_desc = CThreadDesc_K1_N_H2_W2{};
+
+        constexpr auto bias_k0_k1_thread_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(I1, Number<KPerThread>{}));
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     FloatC,
+                     bias_k0_k1_thread_desc.GetElementSpaceSize(),
+                     true>
+            bias_thread_buf;
+
+        const index_t k_thread_data_on_global = k_thread_id * KPerThread;
+
+        auto bias_threadwise_transfer =
+            ThreadwiseTensorSliceTransfer_v2<FloatC,
+                                             FloatC,
+                                             decltype(bias_k0_k1_grid_desc),
+                                             decltype(bias_k0_k1_thread_desc),
+                                             Sequence<I1, Number<KPerThread>{}>,
+                                             Sequence<0, 1>,
+                                             1,
+                                             CThreadTransferDstScalarPerVector,
+                                             false,
+                                             true>(
+                bias_k0_k1_grid_desc, make_multi_index(k_block_work_id, k_thread_data_on_global));
+
+        constexpr auto bias_k0_k1_global_tensor_step_hacks = make_tuple(
+            make_tuple(Sequence<0>{}, Sequence<0>{}), make_tuple(Sequence<0>{}, Sequence<0>{}));
+
+        bias_threadwise_transfer.Run(bias_k0_k1_grid_desc,
+                                     bias_global_buf,
+                                     bias_k0_k1_thread_desc,
+                                     make_tuple(I0, I0),
+                                     bias_thread_buf,
+                                     bias_k0_k1_global_tensor_step_hacks);
+
+        static_for<0, KPerThread, 1>{}([&](auto ki) {
+            static_for<0, HoPerThread, 1>{}([&](auto hi) {
+                static_for<0, WoPerThread, 1>{}([&](auto wi) {
+                    constexpr index_t c_offset =
+                        c_k1_n_h2_w2_thread_gemm_desc.CalculateOffset(make_tuple(ki, 0, hi, wi));
+                    c_thread_buf(Number<c_offset>{}) =
+                        c_thread_buf[Number<c_offset>{}] + bias_thread_buf[ki];
+                });
+            });
+        });
+    }
+
+    template <typename CThreadBuff, typename CThreadDesc_K1_N_H2_W2, ActivTypeEnum activ_type_>
+    __device__ static void Activation(CThreadBuff& c_thread_buf,
+                                      const CThreadDesc_K1_N_H2_W2&,
+                                      integral_constant<ActivTypeEnum, activ_type_>)
+    {
+        constexpr auto c_k1_n_h2_w2_thread_gemm_desc = CThreadDesc_K1_N_H2_W2{};
+
+        static_for<0, c_k1_n_h2_w2_thread_gemm_desc.GetElementSpaceSize(), 1>{}([&](auto i) {
+            if constexpr(activ_type_ == 1)
+            {
+                c_thread_buf(i) = c_thread_buf[i] >= 0 ? c_thread_buf[i] : alpha * c_thread_buf[i];
+            }
+            else if constexpr(activ_type_ == 2)
+            {
+                FloatAcc x = 1.0 + exp(-c_thread_buf[i]);
+
+                asm volatile("\n \
+                        v_rcp_f32 %0, %1 \n"
+                             : "=v"(x)
+                             : "0"(x));
+
+                c_thread_buf(i) = x;
+            }
+        });
+    }
+
+    template <typename CThreadBuff,
+              typename CGlobalBuff,
+              typename CBlockIndex,
+              typename CThreadIndex,
+              typename CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2>
+    __device__ static void
+    WriteOut(const CThreadBuff& c_thread_buf,
+             CGlobalBuff& c_global_buf,
+             const CBlockIndex& c_block_idx,
+             const CThreadIndex& c_thread_idx,
+             const CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2& c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc)
+    {
+        const index_t k_block_work_id  = __builtin_amdgcn_readfirstlane(c_block_idx[I0]);
+        const index_t n_block_work_id  = __builtin_amdgcn_readfirstlane(c_block_idx[I1]);
+        const index_t ho_block_work_id = __builtin_amdgcn_readfirstlane(c_block_idx[I2]);
+        const index_t wo_block_work_id = __builtin_amdgcn_readfirstlane(c_block_idx[I3]);
+
+        const auto k_thread_id  = c_thread_idx[I0];
+        const auto ho_thread_id = c_thread_idx[I2];
+        const auto wo_thread_id = c_thread_idx[I3];
+
+        // hack to control index calculation when iterating over c_k_n_h0_h1_h2_w0_w1_w2_global
+        // tensor
+        constexpr auto c_k_n_h0_h1_h2_w0_w1_w2_global_tensor_step_hacks = CGlobalStepHacks{};
+
+        constexpr auto c_k0_k1_n_h0_h1_h2_w0_w1_w2_thread_copy_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<KPerThread>{},
+                                                           I1,
+                                                           I1,
+                                                           I1,
+                                                           Number<HoPerThread>{},
+                                                           I1,
+                                                           I1,
+                                                           Number<WoPerThread>{}));
+
+        const index_t k_thread_data_on_global = k_thread_id * KPerThread;
+
+        ThreadwiseTensorSliceTransfer_v1r3<
+            FloatAcc,
+            FloatC,
+            decltype(c_k0_k1_n_h0_h1_h2_w0_w1_w2_thread_copy_desc),
+            decltype(c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc),
+            Sequence<I1, KPerThread, I1, I1, I1, HoPerThread, I1, I1, WoPerThread>,
+            CThreadTransferSrcDstAccessOrder,
+            CThreadTransferSrcDstVectorDim,
+            CThreadTransferDstScalarPerVector,
+            CGlobalMemoryDataOperation,
+            1,
+            true>(c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
+                  make_multi_index(k_block_work_id,
+                                   k_thread_data_on_global,
+                                   n_block_work_id,
+                                   ho_block_work_id,
+                                   ho_thread_id,
+                                   0,
+                                   wo_block_work_id,
+                                   wo_thread_id,
+                                   0))
+            .Run(c_k0_k1_n_h0_h1_h2_w0_w1_w2_thread_copy_desc,
+                 make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                 c_thread_buf,
+                 c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
+                 c_global_buf,
+                 c_k_n_h0_h1_h2_w0_w1_w2_global_tensor_step_hacks);
+    }
+
+    template <typename CThreadBuff,
+              typename DGlobalBuff,
+              typename CBlockIndex,
+              typename CThreadIndex,
+              typename CThreadDesc_K1_N_H2_W2,
+              typename DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx>
+    __device__ static void
+    MaxPool(const CThreadBuff& c_thread_buf,
+            DGlobalBuff& d_global_buf,
+            const CBlockIndex& c_block_idx,
+            const CThreadIndex& c_thread_idx,
+            const CThreadDesc_K1_N_H2_W2&,
+            const DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx& d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc)
+    {
+
+        const index_t k_block_work_id  = __builtin_amdgcn_readfirstlane(c_block_idx[I0]);
+        const index_t n_block_work_id  = __builtin_amdgcn_readfirstlane(c_block_idx[I1]);
+        const index_t ho_block_work_id = __builtin_amdgcn_readfirstlane(c_block_idx[I2]);
+        const index_t wo_block_work_id = __builtin_amdgcn_readfirstlane(c_block_idx[I3]);
+
+        const auto k_thread_id  = c_thread_idx[I0];
+        const auto ho_thread_id = c_thread_idx[I2];
+        const auto wo_thread_id = c_thread_idx[I3];
+
+        constexpr auto c_k1_n_h2_w2_thread_gemm_desc = CThreadDesc_K1_N_H2_W2{};
+
+        static_assert(HoPerThread % 2 == 0 && WoPerThread % 2 == 0, "");
+
+        constexpr auto HoPerThread_2 = HoPerThread / 2;
+        constexpr auto WoPerThread_2 = WoPerThread / 2;
+
+        constexpr auto d_k0_k1_n_h0_h1_hx_w0_w1_wx_thread_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<KPerThread>{},
+                                                           I1,
+                                                           I1,
+                                                           I1,
+                                                           Number<HoPerThread_2>{},
+                                                           I1,
+                                                           I1,
+                                                           Number<WoPerThread_2>{}));
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     FloatC,
+                     d_k0_k1_n_h0_h1_hx_w0_w1_wx_thread_desc.GetElementSpaceSize(),
+                     true>
+            d_thread_buf;
+
+        static_for<0, KPerThread, 1>{}([&](auto ki) {
+            static_for<0, HoPerThread_2, 1>{}([&](auto hi) {
+                static_for<0, WoPerThread_2, 1>{}([&](auto wi) {
+                    constexpr index_t d_offset =
+                        d_k0_k1_n_h0_h1_hx_w0_w1_wx_thread_desc.CalculateOffset(
+                            make_tuple(0, ki, 0, 0, 0, hi, 0, 0, wi));
+
+                    constexpr index_t c_offset_0 = c_k1_n_h2_w2_thread_gemm_desc.CalculateOffset(
+                        make_tuple(ki, 0, hi * 2, wi * 2));
+                    constexpr index_t c_offset_1 = c_k1_n_h2_w2_thread_gemm_desc.CalculateOffset(
+                        make_tuple(ki, 0, hi * 2, wi * 2 + 1));
+                    constexpr index_t c_offset_2 = c_k1_n_h2_w2_thread_gemm_desc.CalculateOffset(
+                        make_tuple(ki, 0, hi * 2 + 1, wi * 2));
+                    constexpr index_t c_offset_3 = c_k1_n_h2_w2_thread_gemm_desc.CalculateOffset(
+                        make_tuple(ki, 0, hi * 2 + 1, wi * 2 + 1));
+
+                    d_thread_buf(Number<d_offset>{}) = c_thread_buf[Number<c_offset_0>{}];
+                    d_thread_buf(Number<d_offset>{}) =
+                        fmaxf(c_thread_buf[Number<c_offset_1>{}], d_thread_buf(Number<d_offset>{}));
+                    d_thread_buf(Number<d_offset>{}) =
+                        fmaxf(c_thread_buf[Number<c_offset_2>{}], d_thread_buf(Number<d_offset>{}));
+                    d_thread_buf(Number<d_offset>{}) =
+                        fmax(c_thread_buf[Number<c_offset_3>{}], d_thread_buf(Number<d_offset>{}));
+                });
+            });
+        });
+
+        const index_t k_thread_data_on_global = k_thread_id * KPerThread;
+
+        constexpr auto d_k_n_h0_h1_hx_w0_w1_wx_global_tensor_step_hacks = DGlobalStepHacks{};
+
+        ThreadwiseTensorSliceTransfer_v1r3<
+            FloatC,
+            FloatC,
+            decltype(d_k0_k1_n_h0_h1_hx_w0_w1_wx_thread_desc),
+            decltype(d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc),
+            Sequence<I1, KPerThread, I1, I1, I1, HoPerThread_2, I1, I1, WoPerThread_2>,
+            CThreadTransferSrcDstAccessOrder,
+            CThreadTransferSrcDstVectorDim,
+            CThreadTransferDstScalarPerVector,
+            InMemoryDataOperationEnum::Set,
+            1,
+            true>(d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
+                  make_multi_index(k_block_work_id,
+                                   k_thread_data_on_global,
+                                   n_block_work_id,
+                                   ho_block_work_id,
+                                   ho_thread_id,
+                                   0,
+                                   wo_block_work_id,
+                                   wo_thread_id,
+                                   0))
+            .Run(d_k0_k1_n_h0_h1_hx_w0_w1_wx_thread_desc,
+                 make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                 d_thread_buf,
+                 d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
+                 d_global_buf,
+                 d_k_n_h0_h1_hx_w0_w1_wx_global_tensor_step_hacks);
+    }
+
+    template <typename CThreadBuff,
+              typename DGlobalBuff,
+              typename CBlockIndex,
+              typename CThreadIndex,
+              typename CThreadDesc_K1_N_H2_W2,
+              typename DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx>
+    __device__ static void
+    ResizeAdd(const CThreadBuff& c_thread_buf,
+              DGlobalBuff& d_global_buf,
+              const CBlockIndex& c_block_idx,
+              const CThreadIndex& c_thread_idx,
+              const CThreadDesc_K1_N_H2_W2&,
+              const DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx& d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc)
+    {
+
+        const index_t k_block_work_id  = __builtin_amdgcn_readfirstlane(c_block_idx[I0]);
+        const index_t n_block_work_id  = __builtin_amdgcn_readfirstlane(c_block_idx[I1]);
+        const index_t ho_block_work_id = __builtin_amdgcn_readfirstlane(c_block_idx[I2]);
+        const index_t wo_block_work_id = __builtin_amdgcn_readfirstlane(c_block_idx[I3]);
+
+        const auto k_thread_id  = c_thread_idx[I0];
+        const auto ho_thread_id = c_thread_idx[I2];
+        const auto wo_thread_id = c_thread_idx[I3];
+
+        constexpr auto c_k1_n_h2_w2_thread_gemm_desc = CThreadDesc_K1_N_H2_W2{};
+
+        constexpr auto HoPerThreadx2 = HoPerThread * 2;
+        constexpr auto WoPerThreadx2 = WoPerThread * 2;
+
+        constexpr auto d_k0_k1_n_h0_h1_hx_w0_w1_wx_thread_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<KPerThread>{},
+                                                           I1,
+                                                           I1,
+                                                           I1,
+                                                           Number<HoPerThreadx2>{},
+                                                           I1,
+                                                           I1,
+                                                           Number<WoPerThreadx2>{}));
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     FloatC,
+                     d_k0_k1_n_h0_h1_hx_w0_w1_wx_thread_desc.GetElementSpaceSize(),
+                     true>
+            d_thread_buf;
+
+        static_for<0, KPerThread, 1>{}([&](auto k_i) {
+            static_for<0, HoPerThreadx2, 1>{}([&](auto h_i) {
+                static_for<0, WoPerThreadx2, 1>{}([&](auto w_i) {
+                    d_thread_buf(Number<d_k0_k1_n_h0_h1_hx_w0_w1_wx_thread_desc.CalculateOffset(
+                                     make_tuple(0, k_i, 0, 0, 0, h_i, 0, 0, w_i))>{}) =
+                        c_thread_buf[Number<c_k1_n_h2_w2_thread_gemm_desc.CalculateOffset(
+                            make_tuple(k_i, 0, h_i / 2, w_i / 2))>{}];
+                });
+            });
+        });
+
+        // hack to control index calculation when iterating over d_k_n_ho_wo_global tensor
+        constexpr auto d_k_n_h0_h1_hx_w0_w1_wx_global_tensor_step_hacks = DGlobalStepHacks{};
+
+        const index_t k_thread_data_on_global = k_thread_id * KPerThread;
+
+        ThreadwiseTensorSliceTransfer_v1r3<
+            FloatC,
+            FloatC,
+            decltype(d_k0_k1_n_h0_h1_hx_w0_w1_wx_thread_desc),
+            decltype(d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc),
+            Sequence<I1, KPerThread, I1, I1, I1, HoPerThreadx2, I1, I1, WoPerThreadx2>,
+            CThreadTransferSrcDstAccessOrder,
+            CThreadTransferSrcDstVectorDim,
+            CThreadTransferDstScalarPerVector,
+            InMemoryDataOperationEnum::Add,
+            1,
+            true>(d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
+                  make_multi_index(k_block_work_id,
+                                   k_thread_data_on_global,
+                                   n_block_work_id,
+                                   ho_block_work_id,
+                                   ho_thread_id,
+                                   0,
+                                   wo_block_work_id,
+                                   wo_thread_id,
+                                   0))
+            .Run(d_k0_k1_n_h0_h1_hx_w0_w1_wx_thread_desc,
+                 make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                 d_thread_buf,
+                 d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
+                 d_global_buf,
+                 d_k_n_h0_h1_hx_w0_w1_wx_global_tensor_step_hacks);
+    }
+
+    template <typename AGlobalBuff,
+              typename BGlobalBuff,
+              typename CThreadBuff,
+              typename CBlockIndex,
+              typename CThreadIndex,
+              typename AGridDesc_E0_E1_K0_K1_E2,
+              typename BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2,
+              typename CThreadDesc_K1_N_H2_W2,
+              bool HasMainE0BlockLoop>
+    __device__ static void
+    GemmOp(const AGlobalBuff& a_global_buf,
+           const BGlobalBuff& b_global_buf,
+           CThreadBuff& c_thread_buf,
+           FloatAB* __restrict__ p_shared_block,
+           const CBlockIndex& c_block_idx,
+           const CThreadIndex& c_thread_idx,
+           const AGridDesc_E0_E1_K0_K1_E2& a_e0_e1_k0_k1_e2_grid_desc,
+           const BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2& b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+           const CThreadDesc_K1_N_H2_W2&,
+           integral_constant<bool, HasMainE0BlockLoop>)
+    {
+        constexpr auto HasMainE1BlockLoop       = CalculateHasMainE1BlockLoop();
+        constexpr auto HasDoubleTailE1BlockLoop = CalculateHasDoubleTailE1BlockLoop();
+
+        // const auto c_k_n_h_w_block_cluster_idx =
+        // GetCBlockIndex(cblockid_to_k_n_h_w_block_cluster_adaptor);
+        // cblockid_to_k_n_h_w_block_cluster_adaptor.CalculateBottomIndex(
+        // make_multi_index(get_block_1d_id()));
+
+        const index_t k_block_work_id  = __builtin_amdgcn_readfirstlane(c_block_idx[I0]);
+        const index_t n_block_work_id  = __builtin_amdgcn_readfirstlane(c_block_idx[I1]);
+        const index_t ho_block_work_id = __builtin_amdgcn_readfirstlane(c_block_idx[I2]);
+        const index_t wo_block_work_id = __builtin_amdgcn_readfirstlane(c_block_idx[I3]);
+
+        constexpr auto max_lds_align = Number<ABlockTransferDstScalarPerVector_E2>{};
+
+        constexpr auto a_e1_k1_e2_block_gemm_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<E1PerBlock>{}, Number<KPerBlock>{}, Number<E2>{}), max_lds_align);
+
+        constexpr auto b_e1_n_h_w_e2_block_gemm_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<E1PerBlock>{},
+                                                           I1,
+                                                           Number<HoPerBlock>{},
+                                                           Number<WoPerBlock>{},
+                                                           Number<E2>{}));
+
+        constexpr auto c_k1_n_h2_w2_thread_gemm_desc = CThreadDesc_K1_N_H2_W2{};
+
+        auto blockwise_gemm =
+            BlockwiseGemmDlops_km_kn_m0m1n0n1_v3<BlockSize,
+                                                 FloatAB,
+                                                 FloatAB,
+                                                 FloatAcc,
+                                                 decltype(a_e1_k1_e2_block_gemm_desc),
+                                                 decltype(b_e1_n_h_w_e2_block_gemm_desc),
+                                                 decltype(c_k1_n_h2_w2_thread_gemm_desc),
+                                                 EPerThread,
+                                                 K2>{};
+        // blockwise_gemm.GetBeginOfCThreadDesc_K_N_Ho_Wo(get_thread_local_1d_id());
+
+        const auto ho_thread_id = c_thread_idx[I2];
+        const auto wo_thread_id = c_thread_idx[I3];
+
+        constexpr auto a_e0_e1_k0_k1_e2_block_copy_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<I1>{}, Number<E1>{}, I1, Number<KPerBlock>{}, Number<E2>{}),
+            max_lds_align);
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            BlockwiseTensorSliceTransfer_v4<BlockSize,
+                                            InMemoryDataOperationEnum::Set,
+                                            Sequence<I1, E1, I1, KPerBlock, E2>,
+                                            ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2,
+                                            ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2,
+                                            ABlockTransferThreadClusterArrangeOrder,
+                                            FloatAB,
+                                            FloatAB,
+                                            decltype(a_e0_e1_k0_k1_e2_grid_desc),
+                                            decltype(a_e0_e1_k0_k1_e2_block_copy_desc),
+                                            ABlockTransferSrcAccessOrder,
+                                            Sequence<0, 1, 2, 3, 4>,
+                                            ABlockTransferSrcVectorDim,
+                                            4,
+                                            ABlockTransferSrcScalarPerVector,
+                                            ABlockTransferDstScalarPerVector_E2,
+                                            1,
+                                            1,
+                                            AThreadTransferSrcResetCoordinateAfterRun,
+                                            false>(a_e0_e1_k0_k1_e2_grid_desc,
+                                                   make_multi_index(0, 0, k_block_work_id, 0, 0),
+                                                   a_e0_e1_k0_k1_e2_block_copy_desc,
+                                                   make_multi_index(0, 0, 0, 0, 0));
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(I1, 0, 0, 0, 0);
+
+        constexpr auto b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_thread_copy_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<E1PerBlock>{},
+                                                           I1,
+                                                           I1,
+                                                           I1,
+                                                           Number<HoPerThread>{},
+                                                           I1,
+                                                           I1,
+                                                           Number<WoPerThread>{},
+                                                           Number<E2>{}));
+
+        auto b_threadwise_transfer = ThreadwiseTensorSliceTransfer_v2<
+            FloatAB,
+            FloatAB,
+            decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc),
+            decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_thread_copy_desc),
+            Sequence<I1, E1PerBlock, I1, I1, I1, HoPerThread, I1, I1, WoPerThread, E2>,
+            BBlockTransferSrcAccessOrder,
+            BBlockTransferSrcVectorDim,
+            BBlockTransferSrcScalarPerVector,
+            BThreadTransferSrcResetCoordinateAfterRun,
+            true>(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                  make_multi_index(0,
+                                   0,
+                                   n_block_work_id,
+                                   ho_block_work_id,
+                                   ho_thread_id,
+                                   0,
+                                   wo_block_work_id,
+                                   wo_thread_id,
+                                   0,
+                                   0));
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_shared_block, a_e0_e1_k0_k1_e2_block_copy_desc.GetElementSpaceSize());
+
+        //// register allocation for output
+        // StaticBuffer<AddressSpaceEnum::Vgpr,
+        // FloatAcc,
+        // c_k1_n_h2_w2_thread_gemm_desc.GetElementSpaceSize(),
+        // true>
+        // c_thread_buf;
+
+        // initialize output thread tensor
+        ThreadwiseTensorSliceSet_v1<FloatAcc,
+                                    decltype(c_k1_n_h2_w2_thread_gemm_desc),
+                                    Sequence<KPerThread, I1, HoPerThread, WoPerThread>>{}
+            .Run(c_k1_n_h2_w2_thread_gemm_desc,
+                 make_tuple(I0, I0, I0, I0),
+                 c_thread_buf,
+                 FloatAcc{0});
+
+        constexpr auto b_thread_slice_copy_step =
+            make_multi_index(0, E1PerBlock, 0, 0, 0, 0, 0, 0, 0, 0);
+
+        // hack to control index calculation when iterating over A and B matrix for threadwise copy
+        constexpr auto a_e0_e1_k_e2_global_step_hacks                   = AGlobalStepHacks{};
+        constexpr auto b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks = BGlobalStepHacks{};
+
+        // double regsiter buffer for b
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     FloatAB,
+                     b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_thread_copy_desc.GetElementSpaceSize(),
+                     true>
+            b_thread_even_buf, b_thread_odd_buf;
+
+        if constexpr(HasMainE0BlockLoop)
+        {
+            const auto E0 = b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc.GetLength(I0);
+
+            index_t e0_block_data_begin = 0;
+
+            do
+            {
+                // LDS double buffer: preload data
+                {
+                    a_blockwise_copy.RunRead(
+                        a_e0_e1_k0_k1_e2_grid_desc, a_global_buf, a_e0_e1_k_e2_global_step_hacks);
+
+                    b_threadwise_transfer.Run(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                                              b_global_buf,
+                                              b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_thread_copy_desc,
+                                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                              b_thread_even_buf,
+                                              b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks);
+
+                    a_blockwise_copy.RunWrite(a_e0_e1_k0_k1_e2_block_copy_desc, a_block_buf);
+                }
+
+                __syncthreads();
+
+                if constexpr(HasMainE1BlockLoop)
+                {
+                    index_t e1_block_data_begin = 0;
+
+                    // LDS double buffer: main body
+                    // use Do-While loop instead of For loop to simplify control flow
+                    do
+                    {
+                        // even iteration
+                        b_threadwise_transfer.MoveSrcSliceWindow(
+                            b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                            b_thread_slice_copy_step,
+                            BGlobalMoveSliceWindowStepHacks{});
+
+                        b_threadwise_transfer.Run(
+                            b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                            b_global_buf,
+                            b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_thread_copy_desc,
+                            make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                            b_thread_odd_buf,
+                            b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks);
+
+                        // LDS double buffer: GEMM on current data
+                        blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
+
+                        blockwise_gemm.MoveABlockSliceWindow(make_tuple(E1PerBlock, 0, 0));
+
+                        b_threadwise_transfer.MoveSrcSliceWindow(
+                            b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                            b_thread_slice_copy_step,
+                            BGlobalMoveSliceWindowStepHacks{});
+
+                        b_threadwise_transfer.Run(
+                            b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                            b_global_buf,
+                            b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_thread_copy_desc,
+                            make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                            b_thread_even_buf,
+                            b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks);
+
+                        // LDS double buffer: GEMM on current data
+                        blockwise_gemm.Run(a_block_buf, b_thread_odd_buf, c_thread_buf);
+
+                        blockwise_gemm.MoveABlockSliceWindow(make_tuple(E1PerBlock, 0, 0));
+
+                        e1_block_data_begin += 2 * E1PerBlock;
+
+                    } while(e1_block_data_begin < E1 - 2 * E1PerBlock);
+                }
+
+                // LDS double buffer: tail
+                if constexpr(HasDoubleTailE1BlockLoop) // if has 2 iteration left
+                {
+                    b_threadwise_transfer.MoveSrcSliceWindow(
+                        b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                        b_thread_slice_copy_step,
+                        BGlobalMoveSliceWindowStepHacks{});
+
+                    b_threadwise_transfer.Run(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                                              b_global_buf,
+                                              b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_thread_copy_desc,
+                                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                              b_thread_odd_buf,
+                                              b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks);
+
+                    // LDS double buffer: GEMM on 2nd-last data
+                    blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
+
+                    blockwise_gemm.MoveABlockSliceWindow(make_tuple(E1PerBlock, 0, 0));
+
+                    // LDS double buffer: GEMM on last data
+                    blockwise_gemm.Run(a_block_buf, b_thread_odd_buf, c_thread_buf);
+                }
+                else // if has 1 iteration left
+                {
+                    // LDS double buffer: GEMM on last data
+                    blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
+                }
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_e0_e1_k0_k1_e2_grid_desc,
+                                                    a_block_slice_copy_step,
+                                                    AGlobalMoveSliceWindowStepHacks{});
+
+                blockwise_gemm.MoveABlockSliceWindow(make_tuple(-(E1 - E1PerBlock), 0, 0));
+
+                b_threadwise_transfer.MoveSrcSliceWindow(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                                                         b_thread_slice_copy_step,
+                                                         BGlobalMoveSliceWindowStepHacks{});
+
+                e0_block_data_begin += 1;
+
+            } while(e0_block_data_begin < E0);
+        }
+        else
+        {
+            // LDS double buffer: preload data
+            {
+                a_blockwise_copy.RunRead(
+                    a_e0_e1_k0_k1_e2_grid_desc, a_global_buf, a_e0_e1_k_e2_global_step_hacks);
+
+                b_threadwise_transfer.Run(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                                          b_global_buf,
+                                          b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_thread_copy_desc,
+                                          make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                          b_thread_even_buf,
+                                          b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks);
+
+                a_blockwise_copy.RunWrite(a_e0_e1_k0_k1_e2_block_copy_desc, a_block_buf);
+            }
+
+            __syncthreads();
+
+            if constexpr(HasMainE1BlockLoop)
+            {
+                index_t e1_block_data_begin = 0;
+
+                // LDS double buffer: main body
+                // use Do-While loop instead of For loop to simplify control flow
+                do
+                {
+                    // even iteration
+                    b_threadwise_transfer.MoveSrcSliceWindow(
+                        b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                        b_thread_slice_copy_step,
+                        BGlobalMoveSliceWindowStepHacks{});
+
+                    b_threadwise_transfer.Run(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                                              b_global_buf,
+                                              b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_thread_copy_desc,
+                                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                              b_thread_odd_buf,
+                                              b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks);
+
+                    // LDS double buffer: GEMM on current data
+                    blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
+
+                    blockwise_gemm.MoveABlockSliceWindow(make_tuple(E1PerBlock, 0, 0));
+
+                    b_threadwise_transfer.MoveSrcSliceWindow(
+                        b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                        b_thread_slice_copy_step,
+                        BGlobalMoveSliceWindowStepHacks{});
+
+                    b_threadwise_transfer.Run(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                                              b_global_buf,
+                                              b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_thread_copy_desc,
+                                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                              b_thread_even_buf,
+                                              b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks);
+
+                    // LDS double buffer: GEMM on current data
+                    blockwise_gemm.Run(a_block_buf, b_thread_odd_buf, c_thread_buf);
+
+                    blockwise_gemm.MoveABlockSliceWindow(make_tuple(E1PerBlock, 0, 0));
+
+                    e1_block_data_begin += 2 * E1PerBlock;
+
+                } while(e1_block_data_begin < E1 - 2 * E1PerBlock);
+            }
+
+            // LDS double buffer: tail
+            if constexpr(HasDoubleTailE1BlockLoop) // if has 2 iteration left
+            {
+                b_threadwise_transfer.MoveSrcSliceWindow(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                                                         b_thread_slice_copy_step,
+                                                         BGlobalMoveSliceWindowStepHacks{});
+
+                b_threadwise_transfer.Run(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+                                          b_global_buf,
+                                          b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_thread_copy_desc,
+                                          make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                          b_thread_odd_buf,
+                                          b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks);
+
+                // LDS double buffer: GEMM on 2nd-last data
+                blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
+
+                blockwise_gemm.MoveABlockSliceWindow(make_tuple(E1PerBlock, 0, 0));
+
+                // LDS double buffer: GEMM on last data
+                blockwise_gemm.Run(a_block_buf, b_thread_odd_buf, c_thread_buf);
+            }
+            else // if has 1 iteration left
+            {
+                // LDS double buffer: GEMM on last data
+                blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
+            }
+        }
+    }
+
+    template <typename AGridDesc_E0_E1_K0_K1_E2,
+              typename BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2,
+              typename CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2,
+              typename DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx,
+              typename CBlockIdToBlockClusterAdaptor_K_N_H_W,
+              bool HasMainE0BlockLoop>
+    __device__ static void
+    Conv(const FloatAB* __restrict__ p_a_global,
+         const FloatAB* __restrict__ p_b_global,
+         const FloatC* __restrict__ p_bias_global,
+         FloatC* __restrict__ p_c_global,
+         FloatC* __restrict__ p_d_global,
+         FloatAB* __restrict__ p_shared_block,
+         const AGridDesc_E0_E1_K0_K1_E2& a_e0_e1_k0_k1_e2_grid_desc,
+         const BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2& b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+         const CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2& c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
+         const DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx& d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
+         const CBlockIdToBlockClusterAdaptor_K_N_H_W& cblockid_to_k_n_h_w_block_cluster_adaptor,
+         integral_constant<bool, HasMainE0BlockLoop>)
+    {
+        const auto bias_k0_k1_grid_desc =
+            MakeBiasK0K1GridDescriptor(c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc);
+
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_global, a_e0_e1_k0_k1_e2_grid_desc.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_global, b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc.GetElementSpaceSize());
+        auto c_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_global, c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc.GetElementSpaceSize());
+        auto d_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_d_global, d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc.GetElementSpaceSize());
+        auto bias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_bias_global, bias_k0_k1_grid_desc.GetElementSpaceSize());
+
+        constexpr auto c_k1_n_h2_w2_thread_gemm_desc = MakeCK1NH2W2ThreadDescriptor();
+
+        // register allocation for output
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     FloatAcc,
+                     c_k1_n_h2_w2_thread_gemm_desc.GetElementSpaceSize(),
+                     true>
+            c_thread_buf;
+
+        const auto c_k_n_h_w_block_cluster_idx =
+            GetCBlockIndex(cblockid_to_k_n_h_w_block_cluster_adaptor);
+
+        const auto c_thread_mtx_index = GetCThreadIndex();
+
+        // GemmOp
+        GemmOp(a_global_buf,
+               b_global_buf,
+               c_thread_buf,
+               p_shared_block,
+               c_k_n_h_w_block_cluster_idx,
+               c_thread_mtx_index,
+               a_e0_e1_k0_k1_e2_grid_desc,
+               b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+               c_k1_n_h2_w2_thread_gemm_desc,
+               integral_constant<bool, HasMainE0BlockLoop>{});
+
+        // Output
+        WriteOut(c_thread_buf,
+                 c_global_buf,
+                 c_k_n_h_w_block_cluster_idx,
+                 c_thread_mtx_index,
+                 c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc);
+    }
+
+    template <typename AGridDesc_E0_E1_K0_K1_E2,
+              typename BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2,
+              typename CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2,
+              typename CBlockIdToBlockClusterAdaptor_K_N_H_W,
+              bool HasMainE0BlockLoop,
+              ActivTypeEnum ActivType>
+    __device__ static void ConvBiasActiv(
+        const FloatAB* __restrict__ p_a_global,
+        const FloatAB* __restrict__ p_b_global,
+        const FloatC* __restrict__ p_bias_global,
+        FloatC* __restrict__ p_c_global,
+        FloatAB* __restrict__ p_shared_block,
+        const AGridDesc_E0_E1_K0_K1_E2& a_e0_e1_k0_k1_e2_grid_desc,
+        const BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2& b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+        const CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2& c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
+        const CBlockIdToBlockClusterAdaptor_K_N_H_W& cblockid_to_k_n_h_w_block_cluster_adaptor,
+        integral_constant<bool, HasMainE0BlockLoop>,
+        integral_constant<ActivTypeEnum, ActivType>)
+    {
+        static constexpr auto activ_type = integral_constant<ActivTypeEnum, ActivType>{};
+
+        const auto bias_k0_k1_grid_desc =
+            MakeBiasK0K1GridDescriptor(c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc);
+
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_global, a_e0_e1_k0_k1_e2_grid_desc.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_global, b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc.GetElementSpaceSize());
+        auto c_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_global, c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc.GetElementSpaceSize());
+        auto bias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_bias_global, bias_k0_k1_grid_desc.GetElementSpaceSize());
+
+        constexpr auto c_k1_n_h2_w2_thread_gemm_desc = MakeCK1NH2W2ThreadDescriptor();
+
+        // register allocation for output
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     FloatAcc,
+                     c_k1_n_h2_w2_thread_gemm_desc.GetElementSpaceSize(),
+                     true>
+            c_thread_buf;
+
+        const auto c_k_n_h_w_block_cluster_idx =
+            GetCBlockIndex(cblockid_to_k_n_h_w_block_cluster_adaptor);
+
+        const auto c_thread_mtx_index = GetCThreadIndex();
+
+        // GemmOp
+        GemmOp(a_global_buf,
+               b_global_buf,
+               c_thread_buf,
+               p_shared_block,
+               c_k_n_h_w_block_cluster_idx,
+               c_thread_mtx_index,
+               a_e0_e1_k0_k1_e2_grid_desc,
+               b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+               c_k1_n_h2_w2_thread_gemm_desc,
+               integral_constant<bool, HasMainE0BlockLoop>{});
+
+        // Bias
+        BiasOp(bias_global_buf,
+               c_thread_buf,
+               c_k_n_h_w_block_cluster_idx,
+               c_thread_mtx_index,
+               bias_k0_k1_grid_desc,
+               c_k1_n_h2_w2_thread_gemm_desc);
+
+        // Activ
+        Activation(c_thread_buf, c_k1_n_h2_w2_thread_gemm_desc, activ_type);
+
+        // Output
+        WriteOut(c_thread_buf,
+                 c_global_buf,
+                 c_k_n_h_w_block_cluster_idx,
+                 c_thread_mtx_index,
+                 c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc);
+    }
+
+    template <typename AGridDesc_E0_E1_K0_K1_E2,
+              typename BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2,
+              typename CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2,
+              typename DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx,
+              typename CBlockIdToBlockClusterAdaptor_K_N_H_W,
+              bool HasMainE0BlockLoop,
+              ActivTypeEnum ActivType>
+    __device__ static void ConvBiasActivMaxpool(
+        const FloatAB* __restrict__ p_a_global,
+        const FloatAB* __restrict__ p_b_global,
+        const FloatC* __restrict__ p_bias_global,
+        FloatC* __restrict__ p_c_global,
+        FloatC* __restrict__ p_d_global,
+        FloatAB* __restrict__ p_shared_block,
+        const AGridDesc_E0_E1_K0_K1_E2& a_e0_e1_k0_k1_e2_grid_desc,
+        const BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2& b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+        const CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2& c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
+        const DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx& d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
+        const CBlockIdToBlockClusterAdaptor_K_N_H_W& cblockid_to_k_n_h_w_block_cluster_adaptor,
+        integral_constant<bool, HasMainE0BlockLoop>,
+        integral_constant<ActivTypeEnum, ActivType>)
+    {
+        static constexpr auto activ_type = integral_constant<ActivTypeEnum, ActivType>{};
+
+        const auto bias_k0_k1_grid_desc =
+            MakeBiasK0K1GridDescriptor(c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc);
+
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_global, a_e0_e1_k0_k1_e2_grid_desc.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_global, b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc.GetElementSpaceSize());
+        auto c_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_global, c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc.GetElementSpaceSize());
+        auto d_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_d_global, d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc.GetElementSpaceSize());
+        auto bias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_bias_global, bias_k0_k1_grid_desc.GetElementSpaceSize());
+
+        constexpr auto c_k1_n_h2_w2_thread_gemm_desc = MakeCK1NH2W2ThreadDescriptor();
+
+        // register allocation for output
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     FloatAcc,
+                     c_k1_n_h2_w2_thread_gemm_desc.GetElementSpaceSize(),
+                     true>
+            c_thread_buf;
+
+        const auto c_k_n_h_w_block_cluster_idx =
+            GetCBlockIndex(cblockid_to_k_n_h_w_block_cluster_adaptor);
+
+        const auto c_thread_mtx_index = GetCThreadIndex();
+
+        // GemmOp
+        GemmOp(a_global_buf,
+               b_global_buf,
+               c_thread_buf,
+               p_shared_block,
+               c_k_n_h_w_block_cluster_idx,
+               c_thread_mtx_index,
+               a_e0_e1_k0_k1_e2_grid_desc,
+               b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+               c_k1_n_h2_w2_thread_gemm_desc,
+               integral_constant<bool, HasMainE0BlockLoop>{});
+
+        // Bias
+        BiasOp(bias_global_buf,
+               c_thread_buf,
+               c_k_n_h_w_block_cluster_idx,
+               c_thread_mtx_index,
+               bias_k0_k1_grid_desc,
+               c_k1_n_h2_w2_thread_gemm_desc);
+
+        // Activ
+        Activation(c_thread_buf, c_k1_n_h2_w2_thread_gemm_desc, activ_type);
+
+        // Output
+        WriteOut(c_thread_buf,
+                 c_global_buf,
+                 c_k_n_h_w_block_cluster_idx,
+                 c_thread_mtx_index,
+                 c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc);
+
+        // MaxPool
+        MaxPool(c_thread_buf,
+                d_global_buf,
+                c_k_n_h_w_block_cluster_idx,
+                c_thread_mtx_index,
+                c_k1_n_h2_w2_thread_gemm_desc,
+                d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc);
+    }
+
+    template <typename AGridDesc_E0_E1_K0_K1_E2,
+              typename BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2,
+              typename CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2,
+              typename DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx,
+              typename CBlockIdToBlockClusterAdaptor_K_N_H_W,
+              bool HasMainE0BlockLoop,
+              ActivTypeEnum ActivType>
+    __device__ static void ConvBiasActivResizeAdd(
+        const FloatAB* __restrict__ p_a_global,
+        const FloatAB* __restrict__ p_b_global,
+        const FloatC* __restrict__ p_bias_global,
+        FloatC* __restrict__ p_d_global,
+        FloatAB* __restrict__ p_shared_block,
+        const AGridDesc_E0_E1_K0_K1_E2& a_e0_e1_k0_k1_e2_grid_desc,
+        const BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2& b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+        const CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2& c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
+        const DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx& d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
+        const CBlockIdToBlockClusterAdaptor_K_N_H_W& cblockid_to_k_n_h_w_block_cluster_adaptor,
+        integral_constant<bool, HasMainE0BlockLoop>,
+        integral_constant<ActivTypeEnum, ActivType>)
+    {
+        static constexpr auto activ_type = integral_constant<ActivTypeEnum, ActivType>{};
+
+        const auto bias_k0_k1_grid_desc =
+            MakeBiasK0K1GridDescriptor(c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc);
+
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_global, a_e0_e1_k0_k1_e2_grid_desc.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_global, b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc.GetElementSpaceSize());
+        auto d_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_d_global, d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc.GetElementSpaceSize());
+        auto bias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_bias_global, bias_k0_k1_grid_desc.GetElementSpaceSize());
+
+        constexpr auto c_k1_n_h2_w2_thread_gemm_desc = MakeCK1NH2W2ThreadDescriptor();
+
+        // register allocation for output
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     FloatAcc,
+                     c_k1_n_h2_w2_thread_gemm_desc.GetElementSpaceSize(),
+                     true>
+            c_thread_buf;
+
+        const auto c_k_n_h_w_block_cluster_idx =
+            GetCBlockIndex(cblockid_to_k_n_h_w_block_cluster_adaptor);
+
+        const auto c_thread_mtx_index = GetCThreadIndex();
+
+        // GemmOp
+        GemmOp(a_global_buf,
+               b_global_buf,
+               c_thread_buf,
+               p_shared_block,
+               c_k_n_h_w_block_cluster_idx,
+               c_thread_mtx_index,
+               a_e0_e1_k0_k1_e2_grid_desc,
+               b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
+               c_k1_n_h2_w2_thread_gemm_desc,
+               integral_constant<bool, HasMainE0BlockLoop>{});
+
+        // Bias
+        BiasOp(bias_global_buf,
+               c_thread_buf,
+               c_k_n_h_w_block_cluster_idx,
+               c_thread_mtx_index,
+               bias_k0_k1_grid_desc,
+               c_k1_n_h2_w2_thread_gemm_desc);
+
+        // Activ
+        Activation(c_thread_buf, c_k1_n_h2_w2_thread_gemm_desc, activ_type);
+
+        // Resize_Add
+        ResizeAdd(c_thread_buf,
+                  d_global_buf,
+                  c_k_n_h_w_block_cluster_idx,
+                  c_thread_mtx_index,
+                  c_k1_n_h2_w2_thread_gemm_desc,
+                  d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc);
+    }
+};
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
new file mode 100644
index 0000000..578665e
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
@@ -0,0 +1,944 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename DsDataType,
+          typename FloatE,
+          typename FloatReduceAcc,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation,
+          typename ThreadReduceOperations,
+          InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+          typename RsGlobalMemoryDataOperation,
+          typename AGridDesc_M_K,
+          typename BGridDesc_N_K,
+          typename EGridDesc_M_N,
+          typename RGridDesc_M,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
+          index_t CDEReduceThreadTransferScalarPerVector_NPerBlock,
+          index_t RThreadTransferDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleDMultipleR_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    static constexpr index_t NumRTensor = RsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const T0DataType*, const T1DataType*, ...>
+    template <typename Ts, bool isConst = true>
+    static constexpr auto MakeTsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using T = remove_cvref_t<tuple_element_t<i.value, Ts>>;
+                if constexpr(isConst)
+                    return static_cast<const T*>(nullptr);
+                else
+                    return static_cast<T*>(nullptr);
+            },
+            Number<Ts::Size()>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // A desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2ETileMap>
+    __host__ __device__ static constexpr bool CheckValidity(const AGridDesc_M_K& a_grid_desc_m_k,
+                                                            const BGridDesc_N_K& b_grid_desc_n_k,
+                                                            const EGridDesc_M_N& e_grid_desc_m_n,
+                                                            const RGridDesc_M& r_grid_desc_m,
+                                                            const Block2ETileMap& block_2_etile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        static_assert(AGridDesc_M_K::GetNumOfDimension() == 2);
+        static_assert(BGridDesc_N_K::GetNumOfDimension() == 2);
+        static_assert(EGridDesc_M_N::GetNumOfDimension() == 2);
+
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        if(M != r_grid_desc_m.GetLength(I0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeRGridDescriptor_MBlock_MPerBlock(const RGridDesc_M& r_grid_desc_m)
+    {
+        const auto M      = r_grid_desc_m.GetLength(I0);
+        const auto MBlock = M / MPerBlock;
+
+        const auto r_grid_desc_mblock_mperblock = transform_tensor_descriptor(
+            r_grid_desc_m,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{}))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1>{}));
+
+        return r_grid_desc_mblock_mperblock;
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    using DefaultAGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using DefaultBGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+
+    // Support 2 dimension in the future. Not only M
+    using RGridDescriptor_MBlock_MPerBlock =
+        remove_cvref_t<decltype(MakeRGridDescriptor_MBlock_MPerBlock(RGridDesc_M{}))>;
+
+    using DefaultBlock2ETileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    using DsGridPointer = decltype(MakeTsGridPointer<DsDataType, true>());
+    using RsGridPointer = decltype(MakeTsGridPointer<RsDataType, false>());
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AK0_M_AK1,
+              typename BGridDesc_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        DsGridPointer p_ds_grid,
+        FloatE* __restrict__ p_e_grid,
+        RsGridPointer p_rs_grid,
+        void* __restrict__ p_shared,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const QsElementwiseOperation& qs_element_op,
+        const RsElementwiseOperation& rs_element_op,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const StaticallyIndexedArray<EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                                     NumDTensor>&
+            ds_grid_desc_mblock_mperblock_nblock_nperblock, // FIXME: Ds desc may be of different
+        const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            e_grid_desc_mblock_mperblock_nblock_nperblock,
+        const StaticallyIndexedArray<RGridDescriptor_MBlock_MPerBlock,
+                                     NumRTensor>&
+            rs_grid_desc_mblock_mperblock, // FIXME: Rs desc may be of different
+        const Block2ETileMap& block_2_etile_map)
+    {
+        // FIXME - Share code with other gemm kernel
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        auto rs_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_rs_grid(i), rs_grid_desc_mblock_mperblock[i].GetElementSpaceSize());
+            },
+            Number<NumRTensor>{});
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_etile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C + Ds + reduction + write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_der_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            // TODO: this should be implemented as a blockwise reduction
+            // LDS c_reduce_block_desc_mperblock_nperblock
+            constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
+
+            static_assert(CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I0) *
+                                  CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                              BlockSize,
+                          "wrong!");
+
+            static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
+                                      CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I0) ==
+                                  0 &&
+                              (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
+                                      CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                                  0,
+                          "wrong!");
+
+            constexpr index_t mreduce_per_thread =
+                (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
+                CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I0);
+
+            constexpr index_t nreduce_per_thread =
+                (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
+                CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I1);
+
+            constexpr auto c_reduce_thread_lengths_mperblock_nperblock =
+                Sequence<mreduce_per_thread, nreduce_per_thread>{};
+
+            // VGPR cde_reduce_thread_desc_mperblock_nperblock
+            constexpr auto cde_reduce_thread_desc_mperblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(Number<mreduce_per_thread>{}, Number<nreduce_per_thread>{}));
+
+            constexpr auto r_thread_desc_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(Number<mreduce_per_thread>{}));
+
+            constexpr auto r_thread_desc_mblock_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(I1, Number<mreduce_per_thread>{}));
+
+            auto e_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                cde_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
+
+            // reduce: threadwise copy from LDS to VGPR
+            constexpr auto c_reduce_thread_cluster_desc = make_cluster_descriptor(
+                CDRThreadTransferClusterLengths_MPerBlock_NPerBlock{}, Sequence<1, 0>{});
+
+            const auto c_reduce_thread_cluster_idx =
+                c_reduce_thread_cluster_desc.CalculateBottomIndex(
+                    make_multi_index(get_thread_local_1d_id()));
+
+            const auto c_reduce_thread_data_idx_begin =
+                c_reduce_thread_cluster_idx * c_reduce_thread_lengths_mperblock_nperblock;
+
+            // To apply D0, D1, ... and reduction.
+            // Copy c shuffle from LDS back to VGPR
+            auto c_reduce_thread_copy_lds_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatCShuffle,
+                FloatReduceAcc,
+                decltype(c_reduce_block_desc_mperblock_nperblock),
+                decltype(cde_reduce_thread_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_lengths_mperblock_nperblock),
+                Sequence<0, 1>,
+                1,
+                CDEReduceThreadTransferScalarPerVector_NPerBlock,
+                1,
+                true>{c_reduce_block_desc_mperblock_nperblock, c_reduce_thread_data_idx_begin};
+
+            // Copy result of reduction back from VGPR to global
+            auto reduce_tuple_thread_copy_vgpr_to_global = generate_tuple(
+                [&](auto I) {
+                    auto p_r_grid                     = p_rs_grid[I];
+                    auto r_element_op                 = rs_element_op[I];
+                    auto r_grid_desc_mblock_mperblock = rs_grid_desc_mblock_mperblock[I];
+
+                    return ThreadwiseTensorSliceTransfer_v1r3<
+                        FloatReduceAcc,
+                        remove_pointer_t<decltype(p_r_grid)>,
+                        decltype(r_thread_desc_mblock_mperblock),
+                        decltype(r_grid_desc_mblock_mperblock),
+                        decltype(r_element_op),
+                        Sequence<1, mreduce_per_thread>,
+                        Sequence<0, 1>,
+                        1,
+                        RThreadTransferDstScalarPerVector_MPerBlock,
+                        RsGlobalMemoryDataOperation::At(I),
+                        1,
+                        false>{r_grid_desc_mblock_mperblock,
+                               make_multi_index(block_work_idx[I0],                  // mblock
+                                                c_reduce_thread_data_idx_begin[I0]), // mperblock
+                               r_element_op};
+                },
+                Number<NumRTensor>{});
+
+            // D0, D1, ..., Dn
+            constexpr auto cde_reduce_thread_desc_I1_mperblock_I1_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mreduce_per_thread>{}, I1, Number<nreduce_per_thread>{}));
+
+            // FIXME: Decrease usage of VGPR
+            // Apply pointwise lambda function from multi-source (Global and LDS) into VGPR
+            auto ds_thread_buf = generate_tuple(
+                [&](auto) {
+                    return make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                        cde_reduce_thread_desc_I1_mperblock_I1_nperblock.GetElementSpaceSize());
+                },
+                Number<NumDTensor>{});
+
+            // Copy D0, D1, ..., Dn from global to VGPR
+            auto ds_thread_copy_global_to_vgpr = generate_tuple(
+                [&](auto I) {
+                    using DDataType = remove_cvref_t<tuple_element_t<I.value, DsDataType>>;
+                    return ThreadwiseTensorSliceTransfer_v2<
+                        DDataType,
+                        FloatReduceAcc,
+                        decltype(ds_grid_desc_mblock_mperblock_nblock_nperblock[I]),
+                        decltype(cde_reduce_thread_desc_I1_mperblock_I1_nperblock),
+                        Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
+                        Sequence<0, 1, 2, 3>,
+                        3,
+                        CDEReduceThreadTransferScalarPerVector_NPerBlock,
+                        1,
+                        true>(ds_grid_desc_mblock_mperblock_nblock_nperblock[I],
+                              make_multi_index(
+                                  I0,
+                                  m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                  I0,
+                                  n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]));
+                },
+                Number<NumDTensor>{});
+
+            auto e_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+                FloatReduceAcc,
+                FloatE,
+                decltype(cde_reduce_thread_desc_I1_mperblock_I1_nperblock),
+                decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                tensor_operation::element_wise::PassThrough,
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>, // SliceLengths
+                Sequence<0, 1, 2, 3>,                                     // DimAccessOrder
+                3,                                                        // DstVectorDim
+                CDEReduceThreadTransferScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]),
+                tensor_operation::element_wise::PassThrough{}};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_der_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each thread shuffle data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // Get shuffle data from LDS to VGPR
+                c_reduce_thread_copy_lds_to_vgpr.Run(c_reduce_block_desc_mperblock_nperblock,
+                                                     c_shuffle_block_buf,
+                                                     cde_reduce_thread_desc_mperblock_nperblock,
+                                                     make_tuple(I0, I0),
+                                                     e_thread_buf);
+
+                // Global read D0, D1, ...
+                static_for<0, NumDTensor, 1>{}([&](auto Id) {
+                    auto& d_thread_copy_global_to_vgpr = ds_thread_copy_global_to_vgpr(Id);
+                    d_thread_copy_global_to_vgpr.Run(
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock[Id],
+                        ds_grid_buf[Id],
+                        cde_reduce_thread_desc_I1_mperblock_I1_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        ds_thread_buf(Id));
+
+                    if constexpr(access_id < num_access - 1)
+                    {
+                        // move on D0, D1, ...
+                        constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
+                        d_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                            ds_grid_desc_mblock_mperblock_nblock_nperblock[Id], de_global_step);
+                    }
+                });
+
+                // cde_element_op(e, c, d0, d1, ...);
+                static_for<0, cde_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                    [&](auto i) {
+                        const auto c_ds_src_data_refs = concat_tuple_of_reference(
+                            tie(e_thread_buf[i]),
+                            generate_tie(
+                                [&](auto Id) -> const auto& { return ds_thread_buf[Id][i]; },
+                                Number<NumDTensor>{}));
+                        auto e_dst_data_refs = tie(e_thread_buf(i));
+                        unpack2(cde_element_op, e_dst_data_refs, c_ds_src_data_refs);
+                    });
+
+                // Global write E
+                e_thread_copy_vgpr_to_global.Run(cde_reduce_thread_desc_I1_mperblock_I1_nperblock,
+                                                 make_tuple(I0, I0, I0, I0),
+                                                 e_thread_buf,
+                                                 e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                 e_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    // move on E
+                    constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
+                    e_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                        e_grid_desc_mblock_mperblock_nblock_nperblock, de_global_step);
+                }
+
+                // reduction
+                static_for<0, NumRTensor, 1>{}([&](auto Ir) {
+                    auto r_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                        r_thread_desc_mperblock.GetElementSpaceSize());
+
+                    auto& reduce_thread_copy_vgpr_to_global =
+                        reduce_tuple_thread_copy_vgpr_to_global(Ir);
+
+                    using ThreadReduceOperation =
+                        remove_cvref_t<decltype(ThreadReduceOperations{}[Ir])>;
+
+                    using ThreadwiseReduce =
+                        ThreadwiseReduction<FloatReduceAcc,
+                                            decltype(cde_reduce_thread_desc_mperblock_nperblock),
+                                            decltype(r_thread_desc_mperblock),
+                                            ThreadReduceOperation,
+                                            false>;
+
+                    // threadwise reduction
+                    const auto reduce_identityVal =
+                        ThreadReduceOperation::template GetIdentityValue<FloatReduceAcc>();
+                    static_for<0, mreduce_per_thread, 1>{}(
+                        [&](auto I) { r_thread_buf(I) = reduce_identityVal; });
+                    static_for<0, mreduce_per_thread, 1>{}([&](auto im) {
+                        static_for<0, nreduce_per_thread, 1>{}([&](auto in) {
+                            constexpr auto offset =
+                                Number<cde_reduce_thread_desc_mperblock_nperblock.CalculateOffset(
+                                    make_tuple(im, in))>{};
+
+                            qs_element_op[Ir](e_thread_buf(offset), e_thread_buf(offset));
+                        });
+                    });
+                    ThreadwiseReduce::Reduce(e_thread_buf, r_thread_buf);
+
+                    // gridwise reduction
+                    reduce_thread_copy_vgpr_to_global.Run(r_thread_desc_mblock_mperblock,
+                                                          make_tuple(I0, I0),
+                                                          r_thread_buf,
+                                                          rs_grid_desc_mblock_mperblock[Ir],
+                                                          rs_grid_buf(Ir));
+
+                    if constexpr(access_id < num_access - 1)
+                    {
+                        // move on R0, R1, ...
+                        constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
+                        reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                            rs_grid_desc_mblock_mperblock[Ir],
+                            make_tuple(de_global_step[I0], de_global_step[I1]));
+                    }
+                });
+            }); // copy c, d, e + reduction
+
+        } // shuffle C + Ds + reduction + write out
+    }     // Run
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000..da0b0ce
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,753 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ABDataType, // FIXME: don't assume A/B have same datatype
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleD_xdl_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(ABDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    // A desc for source in blockwise copy
+    template <typename AGridDesc_M_K>
+    __host__ __device__ static constexpr auto
+    MakeDefaultAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    template <typename BGridDesc_N_K>
+    __host__ __device__ static constexpr auto
+    MakeDefaultBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename AGridDesc_M_K,
+              typename BGridDesc_N_K,
+              typename DsGridDesc_M_N,
+              typename EGridDesc_M_N,
+              typename Block2ETileMap>
+    __host__ __device__ static constexpr bool CheckValidity(const AGridDesc_M_K& a_grid_desc_m_k,
+                                                            const BGridDesc_N_K& b_grid_desc_n_k,
+                                                            const DsGridDesc_M_N& ds_grid_desc_m_n,
+                                                            const EGridDesc_M_N& e_grid_desc_m_n,
+                                                            const Block2ETileMap& block_2_etile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        // check consistency of desc
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
+                              N == ds_grid_desc_m_n[i].GetLength(I1));
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // check block-to-E-tile
+        if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_m_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             b_grid_desc_n_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AK0_M_AK1,
+              typename BGridDesc_BK0_N_BK1,
+              typename DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename Block2ETileMap>
+    __device__ static void Run(const ABDataType* __restrict__ p_a_grid,
+                               const ABDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_etile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ABDataType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumDTensor>{}));
+
+            // blockwise copy C/D/E between LDS and global
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
+                                                                            // support arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                3,                    // index_t VectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
+                 cde_element_op};
+
+            // space filling curve for threadwise C in VGPR before shuffle
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
new file mode 100644
index 0000000..98331d8
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
@@ -0,0 +1,43 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp"
+
+namespace ck {
+
+enum struct PipelineVersion
+{
+    v1,
+    v2,
+};
+
+template <PipelineVersion PipelineVer,
+          index_t NumPrefetch     = 1,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+constexpr auto GridwiseGemmPipeline_Selector()
+{
+    if constexpr(PipelineVer == PipelineVersion::v1)
+    {
+        if constexpr(LoopSched == LoopScheduler::Default)
+        {
+            return GridwiseGemmPipeline_v1<NumPrefetch>{};
+        }
+        else if constexpr(LoopSched == LoopScheduler::Interwave)
+        {
+            return GridwiseGemmPipelineInterwave_v1<NumPrefetch>{};
+        }
+    }
+    else if constexpr(PipelineVer == PipelineVersion::v2)
+    {
+        return GridwiseGemmPipeline_v2{};
+    }
+    else
+    {
+        std::cerr << "GridwiseGemmPipeline configuration is not available" << std::endl;
+    }
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp
new file mode 100644
index 0000000..e909755
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp
@@ -0,0 +1,369 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+
+namespace ck {
+
+template <index_t NumPrefetch>
+struct GridwiseGemmPipeline_v1;
+
+// 1-stage prefetch
+template <>
+struct GridwiseGemmPipeline_v1<1>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    __host__ __device__ static constexpr bool IsSupported(index_t /* num_loop */) { return true; }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return num_loop > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    __device__ static void Run(const AGridDesc& a_grid_desc,
+                               const ABlockDesc& a_block_desc,
+                               ABlockTransfer& a_blockwise_copy,
+                               const AGridBuffer& a_grid_buf,
+                               ABlockBuffer& a_block_buf,
+                               const ABlockTransferStep& a_block_copy_step,
+                               const BGridDesc& b_grid_desc,
+                               const BBlockDesc& b_block_desc,
+                               BBlockTransfer& b_blockwise_copy,
+                               const BGridBuffer& b_grid_buf,
+                               BBlockBuffer& b_block_buf,
+                               const BBlockTransferStep& b_block_copy_step,
+                               const BlockwiseGemm& blockwise_gemm,
+                               CThreadBuffer& c_thread_buf,
+                               index_t num_loop)
+    {
+        // preload data into LDS
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+
+                block_sync_lds();
+
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                block_sync_lds();
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                ++i;
+            } while(i < (num_loop - 1));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+// 2-stage prefetch
+template <>
+struct GridwiseGemmPipeline_v1<2>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    __host__ __device__ static constexpr bool IsSupported(index_t num_loop)
+    {
+        // TODO: improve applicability
+        return num_loop % 2 == 0;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return (num_loop / 2) > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    static __device__ void Run(const AGridDesc& a_grid_desc,
+                               const ABlockDesc& a_block_desc,
+                               ABlockTransfer& a_blockwise_copy,
+                               const AGridBuffer& a_grid_buf,
+                               ABlockBuffer& a_block_buf,
+                               const ABlockTransferStep& a_block_copy_step,
+                               const BGridDesc& b_grid_desc,
+                               const BBlockDesc& b_block_desc,
+                               BBlockTransfer& b_blockwise_copy,
+                               const BGridBuffer& b_grid_buf,
+                               BBlockBuffer& b_block_buf,
+                               const BBlockTransferStep& b_block_copy_step,
+                               const BlockwiseGemm& blockwise_gemm,
+                               CThreadBuffer& c_thread_buf,
+                               index_t num_loop)
+    {
+        // preload data into LDS
+        {
+            // Read 0
+            a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+            b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
+
+            // Move
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+            // Read 1
+            a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
+            b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1);
+        }
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                // Move
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                // Write i
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I0);
+
+                // Read i+2
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
+
+                // Sync
+                block_sync_lds();
+
+                // Gemm i
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                // Sync
+                block_sync_lds();
+
+                // Move
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                // Write i+1
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I1);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I1);
+
+                // Read i+3
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1);
+
+                // Sync
+                block_sync_lds();
+
+                // Gemm i+1
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                // Sync
+                block_sync_lds();
+
+                i += 2;
+            } while(i < (num_loop - 2));
+        }
+
+        // tail
+        {
+            // Write num_loop - 2
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I0);
+
+            // Sync
+            block_sync_lds();
+
+            // Gemm num_loop - 2
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+            // Sync
+            block_sync_lds();
+
+            // Write num_loop - 1
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I1);
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I1);
+
+            // Sync
+            block_sync_lds();
+
+            // Gemm num_loop - 1
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+template <index_t NumPrefetch>
+struct GridwiseGemmPipelineInterwave_v1;
+
+template <>
+struct GridwiseGemmPipelineInterwave_v1<1>
+{
+    __host__ __device__ static constexpr bool IsSupported(index_t /* num_loop */) { return true; }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return num_loop > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    static __device__ void Run(const AGridDesc& a_grid_desc,
+                               const ABlockDesc& a_block_desc,
+                               ABlockTransfer& a_blockwise_copy,
+                               const AGridBuffer& a_grid_buf,
+                               ABlockBuffer& a_block_buf,
+                               const ABlockTransferStep& a_block_copy_step,
+                               const BGridDesc& b_grid_desc,
+                               const BBlockDesc& b_block_desc,
+                               BBlockTransfer& b_blockwise_copy,
+                               const BGridBuffer& b_grid_buf,
+                               BBlockBuffer& b_block_buf,
+                               const BBlockTransferStep& b_block_copy_step,
+                               const BlockwiseGemm& blockwise_gemm,
+                               CThreadBuffer& c_thread_buf,
+                               index_t num_loop)
+    {
+        // preload data into LDS
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+
+                block_sync_lds();
+
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                // block_sync_lds(); // moved into blockwise_gemm
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                ++i;
+            } while(i < (num_loop - 1));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+// Note: 2 stage prefetch not optimized for inter-wave loop scheduler
+template <>
+struct GridwiseGemmPipelineInterwave_v1<2> : public GridwiseGemmPipeline_v1<2>
+{
+};
+
+// TODO: deprecate as GridwiseGemmPipeline_Selector covers the functionality
+template <index_t NumPrefetch, LoopScheduler LoopSched>
+constexpr auto GridwiseGemmPipeline_v1_Selector()
+{
+    if constexpr(LoopSched == LoopScheduler::Default)
+    {
+        return GridwiseGemmPipeline_v1<NumPrefetch>{};
+    }
+    else if constexpr(LoopSched == LoopScheduler::Interwave)
+    {
+        return GridwiseGemmPipelineInterwave_v1<NumPrefetch>{};
+    }
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp
new file mode 100644
index 0000000..3281b91
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp
@@ -0,0 +1,128 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+namespace ck {
+
+struct GridwiseGemmPipeline_v2
+{
+    __host__ __device__ static constexpr bool IsSupported(index_t num_loop)
+    {
+        // TODO: improve applicability
+        return num_loop % 2 == 0;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return (num_loop / 2) > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    __device__ static void Run(const AGridDesc& a_grid_desc,
+                               const ABlockDesc& a_block_desc,
+                               ABlockTransfer& a_blockwise_copy,
+                               const AGridBuffer& a_grid_buf,
+                               ABlockBuffer& a_block_buf,
+                               const ABlockTransferStep& a_block_copy_step,
+                               const BGridDesc& b_grid_desc,
+                               const BBlockDesc& b_block_desc,
+                               BBlockTransfer& b_blockwise_copy,
+                               const BGridBuffer& b_grid_buf,
+                               BBlockBuffer& b_block_buf,
+                               const BBlockTransferStep& b_block_copy_step,
+                               const BlockwiseGemm& blockwise_gemm,
+                               CThreadBuffer& c_thread_buf,
+                               index_t num_loop)
+    {
+        // global read 0
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        // move to 1
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // LDS write 0
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        // global Read 1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+
+        // LDS write 0
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+        // global Read 1
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                block_sync_lds();
+
+                // GEMM i
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                block_sync_lds();
+
+                // move to i + 2
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                // LDS write i + 1
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                // global read i + 2
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+
+                // LDS write i + 1
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+                // global read i + 2
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                ++i;
+            } while(i < (num_loop - 2));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            // GEMM num_loop - 2
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+            block_sync_lds();
+
+            // LDS write num_loop - 1
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+            block_sync_lds();
+
+            // GEMM num_loop - 1
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000..2fe5506
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
@@ -0,0 +1,879 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename ReduceGridDescriptor_MBlock_MPerBlock,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_reduce_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            ReducePtrsGlobal p_reduces_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const ReduceInElementwiseOperations reduce_in_element_ops,
+            const ReduceAccElementwiseOperations reduce_out_element_ops,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_reduces_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  reduce_in_element_ops,
+                                                  reduce_out_element_ops,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  reduce_grid_desc_mblock_mperblock,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_reduces_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = reduce_in_element_ops;
+    ignore = reduce_out_element_ops;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = reduce_grid_desc_mblock_mperblock;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename FloatReduceAcc,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename ReduceOperations,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename ReduceGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          typename ReduceGridDesc_M,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+          index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        // static_assert(is_known_at_compile_time<remove_cv_t<decltype(AK1)>>::value &&
+        //               is_known_at_compile_time<remove_cv_t<decltype(BK1)>>::value,
+        //               "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeReduceGridDescriptor_MBlock_MPerBlock(const ReduceGridDesc_M& d_grid_desc_m)
+    {
+        const auto M      = d_grid_desc_m.GetLength(I0);
+        const auto MBlock = M / MPerBlock;
+
+        const auto reduce_grid_desc_mblock_mperblock = transform_tensor_descriptor(
+            d_grid_desc_m,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{}))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1>{}));
+
+        return reduce_grid_desc_mblock_mperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+
+    using ReduceGridDescriptor_MBlock_MPerBlock =
+        remove_cvref_t<decltype(MakeReduceGridDescriptor_MBlock_MPerBlock(ReduceGridDesc_M{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        ReducePtrsGlobal p_reduces_grid,
+        void* __restrict__ p_shared,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const ReduceInElementwiseOperations& reduce_in_element_ops,
+        const ReduceAccElementwiseOperations& reduce_out_element_ops,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            c_grid_desc_mblock_mperblock_nblock_nperblock,
+        const ReduceGridDescriptor_MBlock_MPerBlock& reduce_grid_desc_mblock_mperblock,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C + reduction + write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            // TODO: this should be implemented as a blockwise reduction
+            // LDS c_reduce_block_desc_mperblock_nperblock
+            constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
+
+            static_assert(CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) *
+                                  CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                              BlockSize,
+                          "wrong!");
+
+            static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) ==
+                                  0 &&
+                              (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                                  0,
+                          "wrong!");
+
+            constexpr index_t mreduce_per_thread =
+                (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0);
+
+            constexpr index_t nreduce_per_thread =
+                (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1);
+
+            constexpr auto c_reduce_thread_lengths_mperblock_nperblock =
+                Sequence<mreduce_per_thread, nreduce_per_thread>{};
+
+            // VGPR c_reduce_thread_desc_mperblock_nperblock
+            constexpr auto c_reduce_thread_desc_mperblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(Number<mreduce_per_thread>{}, Number<nreduce_per_thread>{}));
+
+            // VGPR reduce_thread_desc_mperblock
+            constexpr auto reduce_thread_desc_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(Number<mreduce_per_thread>{}));
+
+            // VGPR reduce_thread_desc_mblock_mperblock
+            constexpr auto reduce_thread_desc_mblock_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(I1, Number<mreduce_per_thread>{}));
+
+            auto c_reduce_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                c_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
+
+            // reduce: threadwise copy from LDS to VGPR
+            constexpr auto c_reduce_thread_cluster_desc = make_cluster_descriptor(
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock{}, Sequence<1, 0>{});
+
+            const auto c_reduce_thread_cluster_idx =
+                c_reduce_thread_cluster_desc.CalculateBottomIndex(
+                    make_multi_index(get_thread_local_1d_id()));
+
+            const auto c_reduce_thread_data_idx_begin =
+                c_reduce_thread_cluster_idx * c_reduce_thread_lengths_mperblock_nperblock;
+
+            auto c_reduce_thread_copy_lds_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatCShuffle,
+                FloatReduceAcc,
+                decltype(c_reduce_block_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_lengths_mperblock_nperblock),
+                Sequence<0, 1>,
+                1,
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>{c_reduce_block_desc_mperblock_nperblock, c_reduce_thread_data_idx_begin};
+
+            auto reduce_tuple_thread_copy_vgpr_to_global = generate_tuple(
+                [&](auto I) {
+                    auto p_reduce_grid         = p_reduces_grid[I];
+                    auto reduce_acc_element_op = reduce_out_element_ops[I];
+
+                    return ThreadwiseTensorSliceTransfer_v1r3<
+                        FloatReduceAcc,
+                        remove_pointer_t<decltype(p_reduce_grid)>,
+                        decltype(reduce_thread_desc_mblock_mperblock),
+                        decltype(reduce_grid_desc_mblock_mperblock),
+                        decltype(reduce_acc_element_op),
+                        Sequence<1, mreduce_per_thread>,
+                        Sequence<0, 1>,
+                        1,
+                        CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+                        ReduceGlobalMemoryDataOperation::At(I),
+                        1,
+                        false>{reduce_grid_desc_mblock_mperblock,
+                               make_multi_index(block_work_idx[I0],                  // mblock
+                                                c_reduce_thread_data_idx_begin[I0]), // mperblock
+                               reduce_acc_element_op};
+                },
+                Number<p_reduces_grid.Size()>{});
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                // TODO - extract following into reduction_blockwise
+                {
+                    c_reduce_thread_copy_lds_to_vgpr.Run(c_reduce_block_desc_mperblock_nperblock,
+                                                         c_shuffle_block_buf,
+                                                         c_reduce_thread_desc_mperblock_nperblock,
+                                                         make_tuple(I0, I0),
+                                                         c_reduce_thread_buf);
+
+                    static_for<0, p_reduces_grid.Size(), 1>{}([&](auto In) {
+                        auto& p_reduce_grid = p_reduces_grid[In];
+
+                        auto reduce_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                            p_reduce_grid, reduce_grid_desc_mblock_mperblock.GetElementSpaceSize());
+
+                        auto reduce_thread_buf =
+                            make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                                reduce_thread_desc_mperblock.GetElementSpaceSize());
+
+                        auto& reduce_in_element_op = reduce_in_element_ops[In];
+
+                        auto& reduce_thread_copy_vgpr_to_global =
+                            reduce_tuple_thread_copy_vgpr_to_global(In);
+
+                        using ReduceOperation = remove_cvref_t<decltype(ReduceOperations{}[In])>;
+                        using ThreadwiseReduce =
+                            ThreadwiseReduction<FloatReduceAcc,
+                                                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                                                decltype(reduce_thread_desc_mperblock),
+                                                ReduceOperation,
+                                                false>;
+
+                        // Global write Gemm shuffle + reduction
+                        const auto reduce_identityVal =
+                            ReduceOperation::template GetIdentityValue<FloatReduceAcc>();
+
+                        static_for<0, mreduce_per_thread, 1>{}(
+                            [&](auto I) { reduce_thread_buf(I) = reduce_identityVal; });
+
+                        // reduce in VGPR
+                        static_for<0, mreduce_per_thread, 1>{}([&](auto im) {
+                            static_for<0, nreduce_per_thread, 1>{}([&](auto in) {
+                                constexpr auto offset =
+                                    Number<c_reduce_thread_desc_mperblock_nperblock.CalculateOffset(
+                                        make_tuple(im, in))>{};
+
+                                reduce_in_element_op(c_reduce_thread_buf(offset),
+                                                     c_reduce_thread_buf(offset));
+                            });
+                        });
+
+                        ThreadwiseReduce::Reduce(c_reduce_thread_buf, reduce_thread_buf);
+
+                        // copy from VGPR to Global
+                        reduce_thread_copy_vgpr_to_global.Run(reduce_thread_desc_mblock_mperblock,
+                                                              make_tuple(I0, I0),
+                                                              reduce_thread_buf,
+                                                              reduce_grid_desc_mblock_mperblock,
+                                                              reduce_grid_buf);
+
+                        if constexpr(access_id < num_access - 1)
+                        {
+                            constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+                            reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                                reduce_grid_desc_mblock_mperblock,
+                                make_tuple(c_global_step[I0], c_global_step[I1]));
+                        }
+                    });
+                }
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+
+            // Reduction
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000..aa89bff
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,1263 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ABDataType, // FIXME: don't assume A/B have same datatype
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+          typename AGridDesc_M_K,
+          typename BGridDesc_N_K,
+          typename DsGridDesc_M_N,
+          typename EGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched>
+struct GridwiseGemmSplitKMultipleD_xdl_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemmKPrefetchStage>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, src of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AKB_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(AK0PerBlock * Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       AK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, src of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BKB_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(BK0PerBlock * Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       BK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(ABDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    // A desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k,
+                                             const int split_k)
+    {
+        const auto MRaw = a_grid_desc_m_k.GetLength(I0);
+        const auto KRaw = a_grid_desc_m_k.GetLength(I1);
+
+        const index_t AK0 =
+            (math::integer_divide_ceil(KRaw, KPerBlock * split_k) * KPerBlock) / AK1;
+        const index_t K = split_k * AK0 * AK1;
+        const auto KPad = K - KRaw;
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+        return transform_tensor_descriptor(
+            a_grid_desc_m_kpad,
+            make_tuple(make_unmerge_transform(make_tuple(split_k, AK0, AK1)),
+                       make_pass_through_transform(MRaw)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+    }
+
+    // B desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k,
+                                             const int split_k)
+    {
+        const auto NRaw = b_grid_desc_n_k.GetLength(I0);
+        const auto KRaw = b_grid_desc_n_k.GetLength(I1);
+
+        const index_t BK0 =
+            (math::integer_divide_ceil(KRaw, KPerBlock * split_k) * KPerBlock) / BK1;
+        const index_t K = split_k * BK0 * BK1;
+        const auto KPad = K - KRaw;
+
+        const auto b_grid_desc_n_kpad = transform_tensor_descriptor(
+            b_grid_desc_n_k,
+            make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return transform_tensor_descriptor(
+            b_grid_desc_n_kpad,
+            make_tuple(make_unmerge_transform(make_tuple(split_k, BK0, BK1)),
+                       make_pass_through_transform(NRaw)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const EGridDescriptor_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const DsGridDescriptor_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n, const int split_k)
+    {
+        return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n, 8, split_k);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename AGridDesc_AKB_AK0_M_AK1,
+              typename BGridDesc_BKB_BK0_N_BK1,
+              typename Block2ETileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AKB_AK0_M_AK1& a_grid_desc_akb_ak0_m_ak1,
+                  const BGridDesc_BKB_BK0_N_BK1& b_grid_desc_bkb_bk0_n_bk1,
+                  const DsGridDesc_M_N& ds_grid_desc_m_n,
+                  const EGridDesc_M_N& e_grid_desc_m_n,
+                  const Block2ETileMap& block_2_etile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_akb_ak0_m_ak1.GetLength(I2);
+        const auto N = b_grid_desc_bkb_bk0_n_bk1.GetLength(I2);
+        const auto K =
+            a_grid_desc_akb_ak0_m_ak1.GetLength(I1) * a_grid_desc_akb_ak0_m_ak1.GetLength(I3);
+
+        if(K != b_grid_desc_bkb_bk0_n_bk1.GetLength(I1) * b_grid_desc_bkb_bk0_n_bk1.GetLength(I3))
+        {
+            return false;
+        }
+        if(a_grid_desc_akb_ak0_m_ak1.GetLength(I0) != b_grid_desc_bkb_bk0_n_bk1.GetLength(I0))
+        {
+            return false;
+        }
+
+        // check consistency of desc
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
+                              N == ds_grid_desc_m_n[i].GetLength(I1));
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // check block-to-E-tile
+        if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_akb_ak0_m_ak1.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             b_grid_desc_bkb_bk0_n_bk1.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DefaultAGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(AGridDesc_M_K{}, 1))>;
+    using DefaultBGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(BGridDesc_N_K{}, 1))>;
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+    using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+
+    using DefaultBlock2ETileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}, 1))>;
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AKB_AK0_M_AK1,
+              typename BGridDesc_BKB_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void Run(const ABDataType* __restrict__ p_a_grid,
+                               const ABDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const AGridDesc_AKB_AK0_M_AK1& a_grid_desc_akb_ak0_m_ak1,
+                               const BGridDesc_BKB_BK0_N_BK1& b_grid_desc_bkb_bk0_n_bk1,
+                               const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(block_work_idx[Number<0>{}] == 0)
+        {
+            Run0<HasMainKBlockLoop>(p_a_grid,
+                                    p_b_grid,
+                                    p_ds_grid,
+                                    p_e_grid,
+                                    p_shared,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op,
+                                    a_grid_desc_akb_ak0_m_ak1,
+                                    b_grid_desc_bkb_bk0_n_bk1,
+                                    ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    block_2_etile_map);
+        }
+        else
+        {
+            Run1<HasMainKBlockLoop>(p_a_grid,
+                                    p_b_grid,
+                                    p_e_grid,
+                                    p_shared,
+                                    a_element_op,
+                                    b_element_op,
+                                    a_grid_desc_akb_ak0_m_ak1,
+                                    b_grid_desc_bkb_bk0_n_bk1,
+                                    ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    block_2_etile_map);
+        }
+    }
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AKB_AK0_M_AK1,
+              typename BGridDesc_BKB_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void Run0(const ABDataType* __restrict__ p_a_grid,
+                                const ABDataType* __restrict__ p_b_grid,
+                                DsGridPointer p_ds_grid,
+                                EDataType* __restrict__ p_e_grid,
+                                void* __restrict__ p_shared,
+                                const AElementwiseOperation& a_element_op,
+                                const BElementwiseOperation& b_element_op,
+                                const CDEElementwiseOperation& cde_element_op,
+                                const AGridDesc_AKB_AK0_M_AK1& a_grid_desc_akb_ak0_m_ak1,
+                                const BGridDesc_BKB_BK0_N_BK1& b_grid_desc_bkb_bk0_n_bk1,
+                                const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                    ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                    e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_akb_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bkb_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_etile_map.ValidCTileIndex(
+               make_tuple(block_work_idx[I1], block_work_idx[I2]),
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t k_batch_id = block_work_idx[I0];
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto a_block_desc_akb_ak0_m_ak1 =
+            GetABlockDescriptor_AKB_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        constexpr auto b_block_desc_bkb_bk0_n_bk1 =
+            GetBBlockDescriptor_BKB_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<I1, AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(a_grid_desc_akb_ak0_m_ak1),
+                                                decltype(a_block_desc_akb_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_akb_ak0_m_ak1,
+                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_akb_ak0_m_ak1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<I1, BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(b_grid_desc_bkb_bk0_n_bk1),
+                                                decltype(b_block_desc_bkb_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bkb_bk0_n_bk1,
+                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bkb_bk0_n_bk1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ABDataType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_akb_ak0_m_ak1.GetLength(I1) * a_grid_desc_akb_ak0_m_ak1.GetLength(I3)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_akb_ak0_m_ak1,
+                                                               a_block_desc_akb_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bkb_bk0_n_bk1,
+                                                               b_block_desc_bkb_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+            {
+                // tuple of reference to C/Ds tensor descriptors
+                const auto c_ds_desc_refs = concat_tuple_of_reference(
+                    tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                    generate_tie(
+                        [&](auto i) -> const auto& // return type should be reference
+                        { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                        Number<NumDTensor>{}));
+
+                // tuple of reference to C/Ds tensor descriptors
+                const auto c_ds_buf_refs = concat_tuple_of_reference(
+                    tie(c_shuffle_block_buf),
+                    generate_tie(
+                        [&](auto i) -> const auto& // return type should be reference
+                        { return ds_grid_buf[i]; },
+                        Number<NumDTensor>{}));
+
+                // tuple of starting index of C/Ds blockwise copy
+                const auto idx_c_ds_block_begin = container_concat(
+                    make_tuple(make_multi_index(0, 0, 0, 0)),
+                    generate_tuple(
+                        [&](auto) {
+                            return make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0);
+                        },
+                        Number<NumDTensor>{}));
+
+                // blockwise copy C/D/E between LDS and global
+                auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                    ThisThreadBlock,
+                    decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                    Tuple<EDataType>,
+                    decltype(c_ds_desc_refs),
+                    decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                    CDEElementwiseOperation,
+                    Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make
+                                                                                // Sequence support
+                                                                                // arbitray type
+                    Sequence<1,
+                             CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                             1,
+                             CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                    CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                    3,                    // index_t VectorDim,
+                    CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                    sequence_merge_t<Sequence<true>,
+                                     uniform_sequence_gen_t<
+                                         NumDTensor,
+                                         false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                    Sequence<false>>              // ThreadTransferDstResetCoordinateAfterRunFlags
+                    {c_ds_desc_refs,
+                     idx_c_ds_block_begin,
+                     tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                     make_tuple(make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0)),
+                     cde_element_op};
+
+                // space filling curve for threadwise C in VGPR before shuffle
+                constexpr auto sfc_c_vgpr =
+                    SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                      Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                      Sequence<CShuffleMXdlPerWavePerShuffle,
+                                               CShuffleNXdlPerWavePerShuffle,
+                                               1,
+                                               1,
+                                               M2,
+                                               1,
+                                               M4,
+                                               1>>{};
+
+                // space filling curve for shuffled blockwise C/D/E
+                constexpr auto sfc_cde_block =
+                    SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                      Sequence<0, 2, 1, 3>,
+                                      Sequence<1,
+                                               CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                               1,
+                                               CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+                constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+                static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+                static_for<0, num_access, 1>{}([&](auto access_id) {
+                    // make sure it's safe to write to LDS
+                    block_sync_lds();
+
+                    // each thread write its data from VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                                  c_thread_buf,
+                                                  c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  c_shuffle_block_buf);
+
+                    // make sure it's safe to read from LDS
+                    block_sync_lds();
+
+                    // each block copy its data from LDS to global
+                    cde_block_copy_lds_and_global.Run(
+                        c_ds_desc_refs,
+                        c_ds_buf_refs,
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        tie(e_grid_buf));
+
+                    if constexpr(access_id < num_access - 1)
+                    {
+                        constexpr auto cde_lds_and_global_step =
+                            sfc_cde_block.GetForwardStep(access_id);
+
+                        // move on Ds
+                        static_for<0, NumDTensor, 1>{}([&](auto i) {
+                            cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                                c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                        });
+
+                        // move on E
+                        cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                            tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                            I0,
+                            cde_lds_and_global_step);
+                    }
+                });
+            }
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AKB_AK0_M_AK1,
+              typename BGridDesc_BKB_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void Run1(const ABDataType* __restrict__ p_a_grid,
+                                const ABDataType* __restrict__ p_b_grid,
+                                EDataType* __restrict__ p_e_grid,
+                                void* __restrict__ p_shared,
+                                const AElementwiseOperation& a_element_op,
+                                const BElementwiseOperation& b_element_op,
+                                const AGridDesc_AKB_AK0_M_AK1& a_grid_desc_akb_ak0_m_ak1,
+                                const BGridDesc_BKB_BK0_N_BK1& b_grid_desc_bkb_bk0_n_bk1,
+                                const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&,
+                                const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                    e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_akb_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bkb_bk0_n_bk1.GetElementSpaceSize());
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_etile_map.ValidCTileIndex(
+               make_tuple(block_work_idx[I1], block_work_idx[I2]),
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t k_batch_id = block_work_idx[I0];
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto a_block_desc_akb_ak0_m_ak1 =
+            GetABlockDescriptor_AKB_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        constexpr auto b_block_desc_bkb_bk0_n_bk1 =
+            GetBBlockDescriptor_BKB_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<I1, AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(a_grid_desc_akb_ak0_m_ak1),
+                                                decltype(a_block_desc_akb_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_akb_ak0_m_ak1,
+                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_akb_ak0_m_ak1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<I1, BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(b_grid_desc_bkb_bk0_n_bk1),
+                                                decltype(b_block_desc_bkb_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bkb_bk0_n_bk1,
+                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bkb_bk0_n_bk1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ABDataType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_akb_ak0_m_ak1.GetLength(I1) * a_grid_desc_akb_ak0_m_ak1.GetLength(I3)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_akb_ak0_m_ak1,
+                                                               a_block_desc_akb_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bkb_bk0_n_bk1,
+                                                               b_block_desc_bkb_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+            {
+                // shuffle: blockwise copy C from LDS to global
+                auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                    ThisThreadBlock,                                 // ThreadGroup
+                    ck::tensor_operation::element_wise::PassThrough, // ElementwiseOperation,
+                    EGlobalMemoryDataOperation,                      // DstInMemOp,
+                    Sequence<1,
+                             CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                             1,
+                             CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                    CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    CShuffleDataType,     // typename SrcData,
+                    EDataType,            // typename DstData,
+                    decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                    decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    Sequence<0, 1, 2, 3>,                             // typename DimAccessOrder,
+                    3,                                                // index_t VectorDim,
+                    CDEShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                    true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                    false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                    {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(0, 0, 0, 0),
+                     e_grid_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0),
+                     ck::tensor_operation::element_wise::PassThrough{}};
+
+                // space filling curve for threadwise C in VGPR
+                constexpr auto sfc_c_vgpr =
+                    SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                      Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                      Sequence<CShuffleMXdlPerWavePerShuffle,
+                                               CShuffleNXdlPerWavePerShuffle,
+                                               1,
+                                               1,
+                                               M2,
+                                               1,
+                                               M4,
+                                               1>>{};
+
+                // space filling curve for shuffled blockwise C in global mem
+                constexpr auto sfc_c_global =
+                    SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                      Sequence<0, 2, 1, 3>,
+                                      Sequence<1,
+                                               CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                               1,
+                                               CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+                constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+                static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+                static_for<0, num_access, 1>{}([&](auto access_id) {
+                    // make sure it's safe to write to LDS
+                    block_sync_lds();
+
+                    // each thread write its data from VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                                  c_thread_buf,
+                                                  c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  c_shuffle_block_buf);
+
+                    // make sure it's safe to read from LDS
+                    block_sync_lds();
+
+                    // each block copy its data from LDS to global
+                    c_shuffle_block_copy_lds_to_global.Run(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                        c_shuffle_block_buf,
+                        e_grid_desc_mblock_mperblock_nblock_nperblock,
+                        e_grid_buf);
+
+                    if constexpr(access_id < num_access - 1)
+                    {
+                        constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                        // move on C
+                        c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                            e_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                    }
+                });
+            }
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000..ecc528a
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
@@ -0,0 +1,653 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdl_cshuffle_v1(const FloatAB* __restrict__ p_a_grid,
+                                    const FloatAB* __restrict__ p_b_grid,
+                                    FloatC* __restrict__ p_c_grid,
+                                    const AElementwiseOperation a_element_op,
+                                    const BElementwiseOperation b_element_op,
+                                    const CElementwiseOperation c_element_op,
+                                    const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+                                    const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+                                    const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                                        c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    // FIXME: pass GridwiseGemmPipe as a template arguement into GridwiseGemm
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        static_assert(std::is_default_constructible_v<GridwiseGemmPipe>);
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipe{};
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
new file mode 100644
index 0000000..94e181c
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
@@ -0,0 +1,1068 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+
+namespace ck {
+
+// D = Layernorm(acc_element_op(A * B + broadcast(bias)) + add) * broadcast(gamma) + broadcast(beta)
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename FloatC0,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename CElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename C0GridDescriptor_NBlock_NPerBlock,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_layernorm_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,               // MxN
+            const FloatC0* __restrict__ p_c0_bias_grid,  // 1xN
+            const FloatC0* __restrict__ p_c0_add_grid,   // MxN
+            const FloatC0* __restrict__ p_c0_gamma_grid, // 1xN
+            const FloatC0* __restrict__ p_c0_beta_grid,  // 1xN
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const AccElementwiseOperation acc_element_op,
+            const CElementwiseOperation c_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const C0GridDescriptor_NBlock_NPerBlock c0_grid_desc_nblock_nperblock,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    // TODO ANT: separate into MMA + Epilogue
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_c0_bias_grid,
+                                                  p_c0_add_grid,
+                                                  p_c0_gamma_grid,
+                                                  p_c0_beta_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  acc_element_op,
+                                                  c_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  c0_grid_desc_nblock_nperblock,
+                                                  block_2_ctile_map);
+
+    // TODO ANT: Run layernorm epilogue here
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_c0_bias_grid;
+    ignore = p_c0_add_grid;
+    ignore = p_c0_gamma_grid;
+    ignore = p_c0_beta_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = acc_element_op;
+    ignore = c_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = c0_grid_desc_nblock_nperblock;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// The GEMM + Layernorm implementation is a specialized kernel which allows fusing both layers
+// together given the condition GEMM extents N of MNK is spanned by a single workgroup. For example,
+// a kernel configured with NPerBlock = 128 allows to operate on all GEMM sizes if N <= 128
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename FloatC0,
+          typename FloatReduceAcc, // Data type after shuffle
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename CElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          typename C0GridDesc_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmLayernorm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        // Align 16 bytes (maximum LDS read/write width)
+        constexpr auto c_block_size_aligned =
+            math::integer_least_multiple(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize() *
+                    sizeof(FloatCShuffle),
+                16) /
+            sizeof(FloatCShuffle);
+
+        // LDS allocation for reduction workspace
+        constexpr index_t c_lds_workspace_size = BlockSize;
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size_aligned * sizeof(FloatCShuffle) +
+                             c_lds_workspace_size * sizeof(FloatReduceAcc));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        // in order to reduce N dim without elaborate sync across CUs in single kernel, one
+        // workgroup must span the entire N extent
+        if(math::integer_divide_ceil(N, NPerBlock) > 1)
+        {
+            return false;
+        }
+
+        // static check: all waves in the workgroups combined must cover whole N extent in order
+        // to have efficient N-dim reduction
+        static_assert(CShuffleNXdlPerWavePerShuffle == NXdlPerWave,
+                      "condition not met for efficient layernorm");
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // for bias, beta, gamma
+    __host__ __device__ static constexpr auto
+    MakeC0GridDescriptor_NBlock_NPerBlock(const C0GridDesc_N& c0_grid_desc_n)
+    {
+        const auto N      = c0_grid_desc_n.GetLength(I0);
+        const auto NBlock = N / NPerBlock;
+
+        const auto c0_grid_desc_nblock_nperblock = transform_tensor_descriptor(
+            c0_grid_desc_n,
+            make_tuple(make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1>{}));
+
+        return c0_grid_desc_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+
+    using C0GridDescriptor_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeC0GridDescriptor_NBlock_NPerBlock(C0GridDesc_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        const FloatC0* __restrict__ p_c0_bias_grid,  // 1xN
+        const FloatC0* __restrict__ p_c0_add_grid,   // MxN
+        const FloatC0* __restrict__ p_c0_gamma_grid, // 1xN
+        const FloatC0* __restrict__ p_c0_beta_grid,  // 1xN
+        void* __restrict__ p_shared,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const AccElementwiseOperation& acc_element_op,
+        const CElementwiseOperation& c_element_op,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            c_grid_desc_mblock_mperblock_nblock_nperblock,
+        const C0GridDescriptor_NBlock_NPerBlock& c0_grid_desc_nblock_nperblock,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        auto c0_bias_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_bias_grid, c0_grid_desc_nblock_nperblock.GetElementSpaceSize());
+        // Note: c0_add is of same layout as c so we don't declare new c0_add_desc here
+        auto c0_add_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_add_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        auto c0_gamma_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_gamma_grid, c0_grid_desc_nblock_nperblock.GetElementSpaceSize());
+        auto c0_beta_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_beta_grid, c0_grid_desc_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            const auto NBlock = c0_grid_desc_nblock_nperblock.GetLength(I0);
+
+            // for broadcasting bias, beta, gamma
+            const auto c0_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+                c0_grid_desc_nblock_nperblock,
+                make_tuple(make_insert_transform(I1),
+                           make_insert_transform(I1),
+                           make_pass_through_transform(NBlock),
+                           make_pass_through_transform(NPerBlock)),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            // LDS c_reduce_block_desc_mperblock_nperblock
+            constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
+
+            static_assert(CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) *
+                                  CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                              BlockSize,
+                          "wrong!");
+
+            static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) ==
+                                  0 &&
+                              (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                                  0,
+                          "wrong!");
+
+            constexpr index_t mreduce_per_thread =
+                (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0);
+
+            constexpr index_t nreduce_per_thread =
+                (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1);
+
+            constexpr auto c_reduce_thread_lengths_mperblock_nperblock =
+                Sequence<mreduce_per_thread, nreduce_per_thread>{};
+
+            // pytorch default
+            // https://pytorch.org/docs/stable/generated/torch.nn.LayerNorm.html
+            static constexpr FloatReduceAcc epsilon = 1e-5;
+
+            // VGPR c_reduce_thread_desc_mperblock_nperblock
+            constexpr auto c_reduce_thread_desc_mperblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(Number<mreduce_per_thread>{}, Number<nreduce_per_thread>{}));
+
+            constexpr auto c_reduce_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mreduce_per_thread>{}, I1, Number<nreduce_per_thread>{}));
+
+            // VGPR d_reduce_thread_desc_mperblock
+            constexpr auto d_reduce_thread_desc_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(Number<mreduce_per_thread>{}));
+
+            // TODO: this should be implemented as a blockwise reduction
+            auto c_reduce_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                c_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
+
+            auto c0_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatC0>(
+                c_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
+
+            // Align 16 bytes (maximum LDS read/write width)
+            constexpr auto c_block_size_aligned =
+                math::integer_least_multiple(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize() *
+                        sizeof(FloatCShuffle),
+                    16) /
+                sizeof(FloatCShuffle);
+
+            auto d_reduce_work_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                reinterpret_cast<FloatReduceAcc*>(static_cast<FloatCShuffle*>(p_shared) +
+                                                  c_block_size_aligned),
+                BlockSize);
+
+            // Sum thread workspace
+            auto d0_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                d_reduce_thread_desc_mperblock.GetElementSpaceSize());
+
+            // Squared sum thread workspace
+            auto d1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                d_reduce_thread_desc_mperblock.GetElementSpaceSize());
+
+            // reduce: threadwise copy from LDS to VGPR
+            constexpr auto c_reduce_thread_cluster_desc = make_cluster_descriptor(
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock{}, Sequence<1, 0>{});
+
+            const auto c_reduce_thread_cluster_idx =
+                c_reduce_thread_cluster_desc.CalculateBottomIndex(
+                    make_multi_index(get_thread_local_1d_id()));
+
+            const auto c_reduce_thread_data_idx_begin =
+                c_reduce_thread_cluster_idx * c_reduce_thread_lengths_mperblock_nperblock;
+
+            auto c_reduce_thread_copy_lds_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatCShuffle,
+                FloatReduceAcc,
+                decltype(c_reduce_block_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_lengths_mperblock_nperblock),
+                Sequence<0, 1>,
+                1,
+                CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>{c_reduce_block_desc_mperblock_nperblock, c_reduce_thread_data_idx_begin};
+
+            auto c_reduce_thread_copy_vgpr_to_lds = ThreadwiseTensorSliceTransfer_v1r3<
+                FloatReduceAcc,
+                FloatCShuffle,
+                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                decltype(c_reduce_block_desc_mperblock_nperblock),
+                tensor_operation::element_wise::PassThrough,
+                decltype(c_reduce_thread_lengths_mperblock_nperblock),
+                Sequence<0, 1>,
+                1,
+                CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{c_reduce_block_desc_mperblock_nperblock,
+                      c_reduce_thread_data_idx_begin,
+                      tensor_operation::element_wise::PassThrough{}};
+
+            auto c0_thread_copy_global_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatC0,
+                FloatC0,
+                decltype(c0_grid_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>(c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(block_work_idx[I0],
+                                       c_reduce_thread_data_idx_begin[I0],
+                                       block_work_idx[I1],
+                                       c_reduce_thread_data_idx_begin[I1]));
+
+            // Note: c0_add is of same layout as c so we don't declare new c0_add_desc here
+            auto c0_add_thread_copy_global_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatC0,
+                FloatC0,
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>(c_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(block_work_idx[I0],
+                                       c_reduce_thread_data_idx_begin[I0],
+                                       block_work_idx[I1],
+                                       c_reduce_thread_data_idx_begin[I1]));
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                block_sync_lds();
+
+                // load from LDS and global, add bias
+                c_reduce_thread_copy_lds_to_vgpr.Run(c_reduce_block_desc_mperblock_nperblock,
+                                                     c_shuffle_block_buf,
+                                                     c_reduce_thread_desc_mperblock_nperblock,
+                                                     make_tuple(I0, I0),
+                                                     c_reduce_thread_buf);
+
+                c0_thread_copy_global_to_vgpr.Run(
+                    c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c0_bias_grid_buf,
+                    c_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                    make_tuple(I0, I0, I0, I0),
+                    c0_thread_buf);
+
+                static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                    [&](auto i) {
+                        FloatReduceAcc out;
+                        acc_element_op(out,
+                                       c_reduce_thread_buf(i) +
+                                           static_cast<FloatReduceAcc>(c0_thread_buf(i)));
+                        c_reduce_thread_buf(i) = out; // acc_element_op(acc + bias)
+                    });
+
+                c0_add_thread_copy_global_to_vgpr.Run(
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c0_add_grid_buf,
+                    c_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                    make_tuple(I0, I0, I0, I0),
+                    c0_thread_buf);
+
+                static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                    [&](auto i) {
+                        c_reduce_thread_buf(i) +=
+                            static_cast<FloatReduceAcc>(c0_thread_buf(i)); // add
+                    });
+
+                // layernorm
+                {
+                    using ThreadwiseReduceD0 =
+                        ThreadwiseReduction<FloatReduceAcc,
+                                            decltype(c_reduce_thread_desc_mperblock_nperblock),
+                                            decltype(d_reduce_thread_desc_mperblock),
+                                            reduce::Add,
+                                            false>;
+                    using ThreadwiseReduceD1 =
+                        ThreadwiseReduction<FloatReduceAcc,
+                                            decltype(c_reduce_thread_desc_mperblock_nperblock),
+                                            decltype(d_reduce_thread_desc_mperblock),
+                                            reduce::SquaredAdd,
+                                            false>;
+
+                    const auto d0_zeroVal =
+                        ThreadwiseReduceD0::Op::template GetIdentityValue<FloatReduceAcc>();
+                    const auto d1_zeroVal =
+                        ThreadwiseReduceD1::Op::template GetIdentityValue<FloatReduceAcc>();
+                    static_for<0, mreduce_per_thread, 1>{}(
+                        [&](auto i) { d0_thread_buf(i) = d0_zeroVal; });
+                    static_for<0, mreduce_per_thread, 1>{}(
+                        [&](auto i) { d1_thread_buf(i) = d1_zeroVal; });
+
+                    // reduce sum in VGPR
+                    ThreadwiseReduceD0::Reduce(c_reduce_thread_buf, d0_thread_buf);
+
+                    // reduce squared sum in VGPR
+                    ThreadwiseReduceD1::Reduce(c_reduce_thread_buf, d1_thread_buf);
+
+                    // reduce within workgroup
+                    using BlockwiseReduce = PartitionedBlockwiseReduction<
+                        FloatReduceAcc,
+                        BlockSize,
+                        CReduceThreadClusterLengths_MPerBlock_NPerBlock, // ThreadClusterLengths_M_K
+                        Sequence<1, 0>, // ThreadClusterArrangeOrder
+                        reduce::Add,
+                        false>;
+
+                    static_for<0, mreduce_per_thread, 1>{}([&](auto i) {
+                        block_sync_lds();
+                        BlockwiseReduce::Reduce(d_reduce_work_buf,
+                                                d0_thread_buf(i)); // blockwise reduced sum
+                        block_sync_lds();
+                        BlockwiseReduce::Reduce(d_reduce_work_buf,
+                                                d1_thread_buf(i)); // blockwise reduced squared sum
+                    });
+
+                    // normalize
+                    const index_t NRaw =
+                        c_grid_desc_mblock_mperblock_nblock_nperblock.GetTransforms()[I0]
+                            .GetUpperLengths()[I1]; // TODO: proper handle
+
+                    static_for<0, mreduce_per_thread, 1>{}([&](auto im) {
+                        static_for<0, nreduce_per_thread, 1>{}([&](auto in) {
+                            constexpr auto dst_offset =
+                                Number<c_reduce_thread_desc_mperblock_nperblock.CalculateOffset(
+                                    make_tuple(im, in))>{};
+
+                            constexpr auto src_offset =
+                                Number<d_reduce_thread_desc_mperblock.CalculateOffset(
+                                    make_tuple(im))>{};
+
+                            FloatReduceAcc avg_sum         = d0_thread_buf(src_offset) / NRaw;
+                            FloatReduceAcc avg_squared_sum = d1_thread_buf(src_offset) / NRaw;
+
+                            FloatReduceAcc numerator = c_reduce_thread_buf(dst_offset) - avg_sum;
+                            FloatReduceAcc divisor = epsilon + avg_squared_sum - avg_sum * avg_sum;
+                            FloatReduceAcc divisor_sqrt;
+                            tensor_operation::element_wise::UnarySqrt{}(divisor_sqrt, divisor);
+
+                            c_reduce_thread_buf(dst_offset) = numerator / divisor_sqrt;
+                        });
+                    });
+
+                    // scaling
+                    c0_thread_copy_global_to_vgpr.Run(
+                        c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c0_gamma_grid_buf,
+                        c_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c0_thread_buf);
+
+                    static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                        [&](auto i) {
+                            c_reduce_thread_buf(i) *=
+                                static_cast<FloatReduceAcc>(c0_thread_buf(i)); // * gamma
+                        });
+
+                    c0_thread_copy_global_to_vgpr.Run(
+                        c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c0_beta_grid_buf,
+                        c_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c0_thread_buf);
+
+                    static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                        [&](auto i) {
+                            c_reduce_thread_buf(i) +=
+                                static_cast<FloatReduceAcc>(c0_thread_buf(i)); // + beta
+                        });
+
+                    block_sync_lds();
+
+                    c_reduce_thread_copy_vgpr_to_lds.Run(c_reduce_thread_desc_mperblock_nperblock,
+                                                         make_tuple(I0, I0),
+                                                         c_reduce_thread_buf,
+                                                         c_reduce_block_desc_mperblock_nperblock,
+                                                         c_shuffle_block_buf);
+
+                } // end layernorm
+
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+
+                    // move on C0
+                    c0_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                        c0_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+
+                    // move on C0_add
+                    c0_add_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
new file mode 100644
index 0000000..126887c
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
@@ -0,0 +1,983 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// Implementation of "Merge" transformation primitive that uses division and mod. It is supposed to
+// be used for low_lengths that are known at compile time and are power of 2, otherwise performance
+// will be very bad
+template <typename LowLengths>
+struct Merge_v4_no_carry
+{
+    static constexpr index_t NDimLow = LowLengths::Size();
+
+    using LowerIndex = MultiIndex<NDimLow>;
+    using UpperIndex = MultiIndex<1>;
+
+    using LowLengthsScan =
+        decltype(container_reverse_exclusive_scan(LowLengths{}, math::multiplies{}, Number<1>{}));
+
+    using UpLengths =
+        decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies{}, Number<1>{})));
+
+    LowLengths low_lengths_;
+    LowLengthsScan low_lengths_scan_;
+    UpLengths up_lengths_;
+
+    __host__ __device__ constexpr Merge_v4_no_carry() = default;
+
+    __host__ __device__ constexpr Merge_v4_no_carry(const LowLengths& low_lengths)
+        : low_lengths_{low_lengths},
+          low_lengths_scan_{
+              container_reverse_exclusive_scan(low_lengths, math::multiplies{}, Number<1>{})},
+          up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies{}, Number<1>{}))}
+    {
+        static_assert(LowerIndex::Size() == NDimLow, "wrong!");
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return NDimLow; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        index_t tmp = idx_up[Number<0>{}];
+
+        // division and mod
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            idx_low(i) = tmp / this->low_lengths_scan_[i];
+            tmp %= this->low_lengths_scan_[i];
+        });
+
+        idx_low(Number<NDimLow - 1>{}) = tmp;
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff& idx_up_diff,
+                                              LowIdx& idx_low,
+                                              const UpIdx& idx_up_new,
+                                              Number<Hack>) const
+    {
+        static_assert(LowIdxDiff::Size() == NDimLow && UpIdxDiff::Size() == 1 &&
+                          LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0   = Number<0>{};
+        constexpr auto INm1 = Number<NDimLow - 1>{};
+
+        index_t tmp = idx_up_new[I0];
+
+        idx_low(INm1)      = tmp;
+        idx_diff_low(INm1) = idx_up_diff[I0];
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return false; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<LowLengths>::value &&
+               is_known_at_compile_time<LowLengthsScan>::value &&
+               is_known_at_compile_time<UpLengths>::value;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("Merge_v3_direct_division_mod_wrw, ");
+        printf("low_lengths_ ");
+        print_multi_index(low_lengths_);
+        printf("low_lengths_scan_ ");
+        print_multi_index(low_lengths_scan_);
+        printf("up_lengths_ ");
+        print_multi_index(up_lengths_);
+        printf("}");
+    }
+};
+
+template <typename LowLengths>
+__host__ __device__ constexpr auto make_merge_transform_v4_no_carry(const LowLengths& low_lengths)
+{
+    return Merge_v4_no_carry<LowLengths>{low_lengths};
+}
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_B_K0_M_K1,
+          typename BGridDesc_B_K0_N_K1,
+          typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename CBlockClusterAdaptor,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_bwd_weight(const FloatAB* __restrict__ p_a_grid,
+                                      const FloatAB* __restrict__ p_b_grid,
+                                      FloatC* __restrict__ p_c_grid,
+                                      const AGridDesc_B_K0_M_K1 a_b_k0_m_k1_grid_desc,
+                                      const BGridDesc_B_K0_N_K1 b_b_k0_n_k1_grid_desc,
+                                      const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                                          c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                      const AElementwiseOperation a_element_op,
+                                      const BElementwiseOperation b_element_op,
+                                      const CElementwiseOperation c_element_op,
+                                      const CBlockClusterAdaptor c_block_cluster_adaptor)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_shared_block,
+                                                  a_b_k0_m_k1_grid_desc,
+                                                  b_b_k0_n_k1_grid_desc,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  c_block_cluster_adaptor);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_b_k0_m_k1_grid_desc;
+    ignore = b_b_k0_n_k1_grid_desc;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = c_block_cluster_adaptor;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_B_K0_M_K1,
+          typename BGridDesc_B_K0_N_K1,
+          typename CMNGridDesc,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MRepeat,
+          index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          index_t ABlockLdsM1PerBlock,
+          index_t ABlockLdsM0PerBlock,
+          index_t ABlockLdsM1Padding,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          bool BBlockLdsExtraN,
+          index_t BBlockLdsN1PerBlock,
+          index_t BBlockLdsN0PerBlock,
+          index_t BBlockLdsN1Padding,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          bool ABlockLdsExtraM1Wrw      = false,
+          bool BBlockLdsExtraN1Wrw      = false,
+          index_t NumGemmKPrefetchStage = 1,
+          PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    // M0/M1/M1Padding
+    static constexpr auto M1PerBlock = Number<ABlockLdsM1PerBlock>{};
+    static constexpr auto M0PerBlock = Number<ABlockLdsM0PerBlock>{};
+    static constexpr auto M1Padding  = Number<ABlockLdsM1Padding>{};
+
+    // N0/N1/N1Padding
+    static constexpr auto N1PerBlock = Number<BBlockLdsN1PerBlock>{};
+    static constexpr auto N0PerBlock = Number<BBlockLdsN0PerBlock>{};
+    static constexpr auto N1Padding  = Number<BBlockLdsN1Padding>{};
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                if constexpr(ABlockLdsExtraM1Wrw)
+                {
+                    constexpr auto a_block_desc_k0_m0_m1_k1 = make_naive_tensor_descriptor(
+                        make_tuple(
+                            Number<K0PerBlock>{}, Number<M0PerBlock>{}, Number<M1PerBlock>{}, K1),
+                        make_tuple(Number<M0PerBlock>{} * (Number<M1PerBlock>{} * K1 + M1Padding),
+                                   Number<M1PerBlock>{} * K1 + M1Padding,
+                                   K1,
+                                   I1));
+
+                    constexpr auto a_block_desc_k0_m_k1_tmp = transform_tensor_descriptor(
+                        a_block_desc_k0_m0_m1_k1,
+                        make_tuple(make_pass_through_transform(Number<K0PerBlock>{}),
+                                   make_merge_transform_v3_division_mod(
+                                       make_tuple(Number<M0PerBlock>{}, Number<M1PerBlock>{})),
+                                   make_pass_through_transform(K1)),
+                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+                    return a_block_desc_k0_m_k1_tmp;
+                }
+                else
+                {
+                    return make_naive_tensor_descriptor(
+                        make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                        make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+                }
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_Batch_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_b_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                if constexpr(ABlockLdsExtraM1Wrw)
+                {
+                    constexpr auto a_block_desc_b_k0_m0_m1_k1 = make_naive_tensor_descriptor(
+                        make_tuple(Number<1>{},
+                                   Number<K0PerBlock>{},
+                                   Number<M0PerBlock>{},
+                                   Number<M1PerBlock>{},
+                                   K1),
+                        make_tuple(Number<K0PerBlock>{} * Number<M0PerBlock>{} *
+                                       (Number<M1PerBlock>{} * K1 + M1Padding),
+                                   Number<M0PerBlock>{} * (Number<M1PerBlock>{} * K1 + M1Padding),
+                                   Number<M1PerBlock>{} * K1 + M1Padding,
+                                   K1,
+                                   I1));
+
+                    constexpr auto a_block_desc_b_k0_m_k1_tmp = transform_tensor_descriptor(
+                        a_block_desc_b_k0_m0_m1_k1,
+                        make_tuple(make_pass_through_transform(Number<1>{}),
+                                   make_pass_through_transform(Number<K0PerBlock>{}),
+                                   make_merge_transform_v4_no_carry(
+                                       make_tuple(Number<M0PerBlock>{}, Number<M1PerBlock>{})),
+                                   make_pass_through_transform(K1)),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+                    return a_block_desc_b_k0_m_k1_tmp;
+                }
+                else
+                {
+                    return make_naive_tensor_descriptor(
+                        make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                        make_tuple(Number<K0PerBlock>{} * Number<MPerBlock + 1>{} * K1,
+                                   Number<MPerBlock + 1>{} * K1,
+                                   K1,
+                                   I1));
+                }
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+
+        return a_block_desc_b_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                if constexpr(BBlockLdsExtraN1Wrw)
+                {
+                    constexpr auto b_block_desc_k0_n0_n1_k1 = make_naive_tensor_descriptor(
+                        make_tuple(
+                            Number<K0PerBlock>{}, Number<N0PerBlock>{}, Number<N1PerBlock>{}, K1),
+                        make_tuple(Number<N0PerBlock>{} * (Number<N1PerBlock>{} * K1 + N1Padding),
+                                   Number<N1PerBlock>{} * K1 + N1Padding,
+                                   K1,
+                                   I1));
+
+                    constexpr auto b_block_desc_k0_n_k1_tmp = transform_tensor_descriptor(
+                        b_block_desc_k0_n0_n1_k1,
+                        make_tuple(make_pass_through_transform(Number<K0PerBlock>{}),
+                                   make_merge_transform_v3_division_mod(
+                                       make_tuple(Number<N0PerBlock>{}, Number<N1PerBlock>{})),
+                                   make_pass_through_transform(K1)),
+                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+                    return b_block_desc_k0_n_k1_tmp;
+                }
+                else
+                {
+                    return make_naive_tensor_descriptor(
+                        make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                        make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+                }
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_k0_n_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_Batch_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_b_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                if constexpr(BBlockLdsExtraN1Wrw)
+                {
+                    constexpr auto b_block_desc_b_k0_n0_n1_k1 = make_naive_tensor_descriptor(
+                        make_tuple(Number<1>{},
+                                   Number<K0PerBlock>{},
+                                   Number<N0PerBlock>{},
+                                   Number<N1PerBlock>{},
+                                   K1),
+                        make_tuple(Number<K0PerBlock>{} * Number<N0PerBlock>{} *
+                                       (Number<N1PerBlock>{} * K1 + N1Padding),
+                                   Number<N0PerBlock>{} * (Number<N1PerBlock>{} * K1 + N1Padding),
+                                   Number<N1PerBlock>{} * K1 + N1Padding,
+                                   K1,
+                                   I1));
+
+                    constexpr auto b_block_desc_b_k0_n_k1_tmp = transform_tensor_descriptor(
+                        b_block_desc_b_k0_n0_n1_k1,
+                        make_tuple(make_pass_through_transform(Number<1>{}),
+                                   make_pass_through_transform(Number<K0PerBlock>{}),
+                                   make_merge_transform_v4_no_carry(
+                                       make_tuple(Number<N0PerBlock>{}, Number<N1PerBlock>{})),
+                                   make_pass_through_transform(K1)),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+                    return b_block_desc_b_k0_n_k1_tmp;
+                }
+                else
+                {
+                    return make_naive_tensor_descriptor(
+                        make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                        make_tuple(Number<K0PerBlock>{} * Number<NPerBlock + 1>{} * K1,
+                                   Number<NPerBlock + 1>{} * K1,
+                                   K1,
+                                   I1));
+                }
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+
+        return b_block_desc_b_k0_n_k1;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_b_k0_m_k1_block_desc = GetABlockDescriptor_Batch_K0PerBlock_MPerBlock_K1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_b_k0_n_k1_block_desc = GetBBlockDescriptor_Batch_K0PerBlock_NPerBlock_K1();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size = math::integer_least_multiple(
+            a_b_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size = math::integer_least_multiple(
+            b_b_k0_n_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto c_block_size =
+            GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock().GetElementSpaceSize();
+
+        return math::max((a_block_space_size + b_block_space_size) * sizeof(FloatAB),
+                         c_block_size * sizeof(FloatC));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
+                  const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
+                  const CMNGridDesc& c_m_n_grid_desc,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MRepeat) == 0) &&
+                          (NPerBlock % (NRepeat * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M      = a_b_k0_m_k1_grid_desc.GetLength(I2);
+        const auto N      = b_b_k0_n_k1_grid_desc.GetLength(I2);
+        const auto K0     = a_b_k0_m_k1_grid_desc.GetLength(I1);
+        const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!(M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
+             K0 == b_b_k0_n_k1_grid_desc.GetLength(I1) &&
+             K1 == a_b_k0_m_k1_grid_desc.GetLength(I3) &&
+             K1 == b_b_k0_n_k1_grid_desc.GetLength(I3) &&
+             KBatch == b_b_k0_n_k1_grid_desc.GetLength(I0)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        if(!block_2_ctile_map.CheckValidity(c_m_n_grid_desc))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        // const bool has_main_k0_block_loop = K0 > K0PerBlock;
+        const index_t num_loop = K0 / K0PerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+
+        // return has_main_k0_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CMNGridDesc& c_m_n_grid_desc)
+    {
+        const auto M = c_m_n_grid_desc.GetLength(I0);
+        const auto N = c_m_n_grid_desc.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        return transform_tensor_descriptor(
+            c_m_n_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
+        const CMNGridDesc& c_m_n_grid_desc, index_t M01, index_t N01, index_t KBatch)
+    {
+        return BlockToCTileMap_KSplit_M00_N00_M01_N01<MPerBlock, NPerBlock, CMNGridDesc>(
+            c_m_n_grid_desc, M01, N01, KBatch);
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1,
+                       Number<CShuffleMRepeatPerShuffle * MWave * MPerXDL>{},
+                       I1,
+                       Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
+    }
+
+    using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        decltype(MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CMNGridDesc{}));
+    using CBlockClusterAdaptor = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1));
+
+    template <bool HasMainKBlockLoop>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               FloatAB* __restrict__ p_shared_block,
+                               const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
+                               const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
+                               const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const CBlockClusterAdaptor& c_block_cluster_adaptor)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t k_batch_id = block_work_idx[I0];
+
+        if(!c_block_cluster_adaptor.ValidCTileIndex(
+               make_tuple(block_work_idx[I1], block_work_idx[I2]),
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto a_b_k0_m_k1_block_desc = GetABlockDescriptor_Batch_K0PerBlock_MPerBlock_K1();
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        constexpr auto b_b_k0_n_k1_block_desc = GetBBlockDescriptor_Batch_K0PerBlock_NPerBlock_K1();
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, K0PerBlock, MPerBlock, K1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_b_k0_m_k1_grid_desc),
+                                                decltype(a_b_k0_m_k1_block_desc),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                a_b_k0_m_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_b_k0_m_k1_block_desc,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_b_k0_n_k1_grid_desc),
+                                                decltype(b_b_k0_n_k1_block_desc),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                b_b_k0_n_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_b_k0_n_k1_block_desc,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+
+        constexpr index_t KPack =
+            math::max(K1, MfmaSelector<FloatAB, MPerXDL, NPerXDL>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_k0_m_k1_block_desc),
+                                                                decltype(b_k0_n_k1_block_desc),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                KPack>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block = p_shared_block;
+        FloatAB* p_b_block = p_shared_block + a_block_space_size;
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
+
+        // gridwise GEMM pipeline
+        const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
+
+        GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_b_k0_m_k1_grid_desc,
+                                                          a_b_k0_m_k1_block_desc,
+                                                          a_blockwise_copy,
+                                                          a_grid_buf,
+                                                          a_block_buf,
+                                                          a_block_slice_copy_step,
+                                                          b_b_k0_n_k1_grid_desc,
+                                                          b_b_k0_n_k1_block_desc,
+                                                          b_blockwise_copy,
+                                                          b_grid_buf,
+                                                          b_block_buf,
+                                                          b_block_slice_copy_step,
+                                                          blockwise_gemm,
+                                                          c_thread_buf,
+                                                          K0BlockMainLoop);
+
+        // output: register to global memory
+        {
+            constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+            constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
+            constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
+            constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
+            constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
+            constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
+            constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
+            constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
+            constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
+
+            constexpr auto c_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            void* p_shared = static_cast<void*>(p_shared_block);
+
+            auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatC*>(p_shared),
+                c_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            static_assert(M1 == MWave, "");
+            static_assert(N1 == NWave, "");
+            static_assert(M2 * M3 * M4 == MPerXDL, "");
+            static_assert(N2 == NPerXDL, "");
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0), // freeze mblock
+                    make_unmerge_transform(make_tuple(CShuffleMRepeatPerShuffle,
+                                                      M1,
+                                                      M2,
+                                                      M3,
+                                                      M4)), // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                    make_freeze_transform(I0),              // freeze nblock
+                    make_unmerge_transform(make_tuple(CShuffleNRepeatPerShuffle,
+                                                      N1,
+                                                      N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMRepeatPerShuffle,
+                                                            CShuffleNRepeatPerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // LDS to global
+            auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // index_t BlockSize,
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMRepeatPerShuffle * MWave * MPerXDL,
+                         1,
+                         CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
+                CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatC,               // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
+                3,                                          // index_t VectorDim,
+                CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun
+                {c_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0),
+                 c_element_op};
+
+            constexpr auto mxdlperwave_forward_step =
+                make_multi_index(0, CShuffleMRepeatPerShuffle * MWave * MPerXDL, 0, 0);
+            constexpr auto nxdlperwave_forward_step =
+                make_multi_index(0, 0, 0, CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+            constexpr auto nxdlperwave_backward_step =
+                make_multi_index(0, 0, 0, -CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+
+            static_for<0, MRepeat, CShuffleMRepeatPerShuffle>{}([&](auto mxdlperwave_iter) {
+                constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                static_for<0, NRepeat, CShuffleNRepeatPerShuffle>{}([&](auto nxdlperwave_iter) {
+                    constexpr bool nxdlperwave_forward_sweep =
+                        (mxdlperwave % (2 * CShuffleMRepeatPerShuffle) == 0);
+
+                    constexpr index_t nxdlperwave_value =
+                        nxdlperwave_forward_sweep
+                            ? nxdlperwave_iter
+                            : (NRepeat - nxdlperwave_iter - CShuffleNRepeatPerShuffle);
+
+                    constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                    // make sure it's safe to do ds_write
+                    block_sync_lds();
+
+                    // VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(
+                        c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
+                        make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                        c_thread_buf,
+                        c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        c_block_buf);
+
+                    // make sure it's safe to do ds_read
+                    block_sync_lds();
+
+                    // LDS to global
+                    c_block_copy_lds_to_global.Run(c_block_desc_mblock_mperblock_nblock_nperblock,
+                                                   c_block_buf,
+                                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                   c_grid_buf);
+
+                    // move on nxdlperwave dimension
+                    if constexpr(nxdlperwave_forward_sweep &&
+                                 (nxdlperwave < NRepeat - CShuffleNRepeatPerShuffle))
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mperblock_nblock_nperblock,
+                            nxdlperwave_forward_step);
+                    }
+                    else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mperblock_nblock_nperblock,
+                            nxdlperwave_backward_step);
+                    }
+                });
+
+                // move on mxdlperwave dimension
+                if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle)
+                {
+                    c_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, mxdlperwave_forward_step);
+                }
+            });
+        }
+    }
+}; // namespace ck
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
new file mode 100644
index 0000000..2aad712
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
@@ -0,0 +1,678 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3,
+          typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainK0BlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_skip_b_lds_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3 b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+            const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid,
+                                                   p_b_grid,
+                                                   p_c_grid,
+                                                   p_shared,
+                                                   a_grid_desc_k0_m_k1,
+                                                   b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                                   c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                   a_element_op,
+                                                   b_element_op,
+                                                   c_element_op,
+                                                   block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3;
+    ignore = c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M_N,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          index_t BBlockTransferSrcScalarPerVector,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockBufferSize,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_skip_b_lds_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    static constexpr index_t WaveSize = 64;
+    static constexpr index_t MWaves   = MPerBlock / (MXdlPerWave * MPerXDL);
+    static constexpr index_t NWaves   = NPerBlock / (NXdlPerWave * NPerXDL);
+
+    static constexpr auto xdlops_gemm    = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, K1>{};
+    static constexpr index_t K0PerThread = K0PerBlock / xdlops_gemm.K0PerXdlops;
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock * BBlockBufferSize>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock * BBlockBufferSize>{}, Number<MPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size_aligned) * sizeof(FloatAB);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  index_t M01,
+                  index_t N01)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // 2-stage prefetch currently only support even number of K0 loop
+        // TODO: add support for odd number of K0 loop
+        if(!((K0 / K0PerBlock) % BBlockBufferSize == 0))
+        {
+            return false;
+        }
+
+        // check M01, N01
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        if(!(M0 % M01 == 0 && N0 % N01 == 0))
+            return false;
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr index_t
+    CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+
+        return grid_size;
+    }
+
+    // TODO move this function into GEMM-pipeline class
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        const bool has_main_k0_block_loop = (K0 / (BBlockBufferSize * K0PerBlock)) > 1;
+
+        return has_main_k0_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBGridDescriptor_K0_K1_K2_N0_N1_N2_N3_K3(const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1)
+    {
+        const auto K0 = b_grid_desc_k0_n_k1.GetLength(I0);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+
+        const auto b_griddesc_k0_nblockid_nrepeat_waves_nperxdlops_k1 = transform_tensor_descriptor(
+            b_grid_desc_k0_n_k1,
+            make_tuple(make_unmerge_transform(
+                           make_tuple(K0 / K0PerBlock, xdlops_gemm.K0PerXdlops, K0PerThread)),
+                       make_unmerge_transform(make_tuple(
+                           N / (NXdlPerWave * NWaves * NPerXDL), NXdlPerWave, NWaves, NPerXDL)),
+                       make_pass_through_transform(K1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5, 6>{}, Sequence<7>{}));
+        return b_griddesc_k0_nblockid_nrepeat_waves_nperxdlops_k1;
+    }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto GetWaveKNIdx(const index_t thread_id)
+    {
+        constexpr auto wave_threadid_to_nk_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(xdlops_gemm.K0PerXdlops, NPerXDL))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        return wave_threadid_to_nk_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix threadwise copy
+        constexpr auto b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           I1,
+                                                           Number<K0PerThread>{}, // K0PerThread
+                                                           I1,                    // NBlockId
+                                                           Number<NXdlPerWave>{}, // repeat
+                                                           I1,                    // waves
+                                                           I1,                    // NPerXdlops
+                                                           Number<K1>{}));
+
+        using BlockwiseGemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1r1<
+            BlockSize,
+            FloatAB,
+            FloatAcc,
+            decltype(a_block_desc_k0_m_k1),
+            decltype(b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3),
+            MPerBlock,
+            NPerBlock,
+            K0PerBlock,
+            MPerXDL,
+            NPerXDL,
+            MXdlPerWave,
+            NXdlPerWave,
+            K1>;
+
+        return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n);
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n, index_t M01, index_t N01)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        const auto M00 = M0 / M01;
+        const auto N00 = N0 / N01;
+
+        const auto m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_unmerge_transform(make_tuple(M00, M01)),
+                           make_unmerge_transform(make_tuple(N00, N01))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}));
+
+        const auto cblockid_to_m00_m01_n00_n01_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(M00, N00, M01, N01))),
+                make_tuple(Sequence<0, 1, 2, 3>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto cblockid_to_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
+                                  cblockid_to_m00_m01_n00_n01_block_cluster_adaptor);
+
+        return cblockid_to_m0_n0_block_cluster_adaptor;
+    }
+
+    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
+        decltype(MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
+    using DefaultBlock2CTileMap = decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
+    using BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3 =
+        decltype(MakeBGridDescriptor_K0_K1_K2_N0_N1_N2_N3_K3(BGridDesc_K0_N_K1{}));
+
+    template <bool HasMainK0BlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        void* __restrict__ p_shared,
+        const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+        const BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3 b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+        const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2& c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetElementSpaceSize());
+
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy = ThreadGroupTensorSliceTransfer_v4r1<
+            ThisThreadBlock,
+            AElementwiseOperation,
+            ck::tensor_operation::element_wise::PassThrough,
+            InMemoryDataOperationEnum::Set,
+            Sequence<K0PerBlock * BBlockBufferSize, MPerBlock, K1>,
+            ABlockTransferThreadClusterLengths_K0_M_K1,
+            ABlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            decltype(a_grid_desc_k0_m_k1),
+            decltype(a_block_desc_k0_m_k1),
+            ABlockTransferSrcAccessOrder,
+            Sequence<1, 0, 2>,
+            ABlockTransferSrcVectorDim,
+            2,
+            ABlockTransferSrcScalarPerVector,
+            ABlockTransferDstScalarPerVector_K1,
+            1,
+            1,
+            AThreadTransferSrcResetCoordinateAfterRun,
+            true,
+            1>(a_grid_desc_k0_m_k1,
+               make_multi_index(0, m_block_data_idx_on_grid, 0),
+               a_element_op,
+               a_block_desc_k0_m_k1,
+               make_multi_index(0, 0, 0),
+               ck::tensor_operation::element_wise::PassThrough{});
+
+        ignore = b_element_op;
+        // B matrix threadwise copy
+        constexpr auto b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           I1,
+                                                           Number<K0PerThread>{}, // K0PerThread
+                                                           I1,                    // NBlockId
+                                                           Number<NXdlPerWave>{}, // repeat
+                                                           I1,                    // waves
+                                                           I1,                    // NPerXdlops
+                                                           Number<K1>{}));
+
+        auto b_thread_buf = generate_tuple(
+            [&](auto i) {
+                ignore = i;
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    FloatAB,
+                                    b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3.GetElementSpaceSize(),
+                                    true>{};
+            },
+            Number<BBlockBufferSize>{});
+
+        const auto wave_id     = GetWaveIdx();
+        const auto wave_k_n_id = GetWaveKNIdx(wave_id[I2]);
+
+#if 0
+        const index_t block_id  = get_block_1d_id();
+        const index_t thread_id = get_thread_local_1d_id();
+        printf("block id: %d  m blockid: %d n block id: %d ,thread id: %d, wave id :{%d %d %d} "
+               "kn id: {%d %d}\n",
+               block_id,
+               block_work_idx[I0],
+               block_work_idx[I1],
+               thread_id,
+               wave_id[I0],
+               wave_id[I1],
+               wave_id[I2],
+               wave_k_n_id[I0],
+               wave_k_n_id[I1]);
+        printf("mfma thread k per xdlops: %d K0PerThread: %d HasMainK0BlockLoop: %d K0: %d  \t", 
+                xdlops_gemm.K0PerXdlops, K0PerThread, HasMainK0BlockLoop, b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3.GetLength(I0));
+#endif
+
+        auto b_threadwise_copy =
+            ThreadwiseTensorSliceTransfer_v2<FloatAB,
+                                             FloatAB,
+                                             decltype(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3),
+                                             decltype(b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3),
+                                             Sequence<I1,
+                                                      I1,
+                                                      Number<K0PerThread>{},
+                                                      I1,
+                                                      Number<NXdlPerWave>{},
+                                                      I1,
+                                                      I1,
+                                                      Number<K1>{}>,
+                                             Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                             7,
+                                             BBlockTransferSrcScalarPerVector,
+                                             BThreadTransferSrcResetCoordinateAfterRun,
+                                             true>(
+                b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                make_multi_index(
+                    0, wave_k_n_id[I0], 0, block_work_idx[I1], 0, wave_id[I1], wave_k_n_id[I1], 0));
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1r1<
+            BlockSize,
+            FloatAB,
+            FloatAcc,
+            decltype(a_block_desc_k0_m_k1),
+            decltype(b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3),
+            MPerBlock,
+            NPerBlock,
+            K0PerBlock,
+            MPerXDL,
+            NPerXDL,
+            MXdlPerWave,
+            NXdlPerWave,
+            K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
+
+        // gridwise GEMM pipeline
+        constexpr auto a_block_slice_copy_step =
+            make_multi_index(K0PerBlock * BBlockBufferSize, 0, 0);
+        constexpr auto b_thread_slice_copy_step = make_multi_index(1, 0, 0, 0, 0, 0, 0, 0);
+        // preload data to regiester and LDS
+        {
+            // Read
+            a_blockwise_copy.RunRead(a_grid_desc_k0_m_k1, a_grid_buf);
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m_k1, a_block_slice_copy_step);
+
+            static_for<0, BBlockBufferSize, 1>{}([&](auto ii) {
+                b_threadwise_copy.Run(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                      b_grid_buf,
+                                      b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                      make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                                      b_thread_buf(Number<ii>{}));
+                b_threadwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                                     b_thread_slice_copy_step);
+            });
+
+            // Initialize C
+            c_thread_buf.Clear();
+            // a data write to lds
+            a_blockwise_copy.RunWrite(a_block_desc_k0_m_k1, a_block_buf);
+            // main body
+            if constexpr(HasMainK0BlockLoop)
+            {
+                index_t K0BlockMainLoop =
+                    __builtin_amdgcn_readfirstlane(K0 / (BBlockBufferSize * K0PerBlock));
+                index_t i = 0;
+                do
+                {
+                    a_blockwise_copy.RunRead(a_grid_desc_k0_m_k1, a_grid_buf);
+                    blockwise_gemm.ResetABlockStartWindow();
+                    block_sync_lds();
+
+                    static_for<0, BBlockBufferSize, 1>{}([&](auto ii) {
+                        blockwise_gemm.Run(a_block_buf, b_thread_buf(Number<ii>{}), c_thread_buf);
+                        blockwise_gemm.MoveABlockSliceWindow();
+                        s_nop();
+
+                        b_threadwise_copy.Run(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                              b_grid_buf,
+                                              b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                                              b_thread_buf(Number<ii>{}));
+                        b_threadwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                                             b_thread_slice_copy_step);
+                    });
+
+                    block_sync_lds();
+                    a_blockwise_copy.RunWrite(a_block_desc_k0_m_k1, a_block_buf);
+                    // move a and b window
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m_k1,
+                                                        a_block_slice_copy_step);
+
+                    i += 1;
+                } while(i < (K0BlockMainLoop - 1));
+            }
+
+            // tail
+            {
+                block_sync_lds();
+
+                blockwise_gemm.ResetABlockStartWindow();
+
+                static_for<0, BBlockBufferSize, 1>{}([&](auto ii) {
+                    blockwise_gemm.Run(a_block_buf, b_thread_buf(Number<ii>{}), c_thread_buf);
+                    blockwise_gemm.MoveABlockSliceWindow();
+                });
+            }
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I7);
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_grid =
+                m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
+
+            const index_t n_thread_data_on_grid =
+                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_grid_idx =
+                m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_grid));
+
+            const auto n_thread_data_on_grid_to_n0_n1_n2_adaptor = make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_grid_idx =
+                n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_grid));
+
+            auto c_thread_copy =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   CElementwiseOperation,
+                                                   Sequence<M0, N0, I1, I1, M2, I1, M4, I1>,
+                                                   CThreadTransferSrcDstAccessOrder,
+                                                   CThreadTransferSrcDstVectorDim,
+                                                   CThreadTransferDstScalarPerVector,
+                                                   CGlobalMemoryDataOperation,
+                                                   1,
+                                                   true>{
+                    c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(m_thread_data_on_grid_idx[I0],
+                                     n_thread_data_on_grid_idx[I0],
+                                     m_thread_data_on_grid_idx[I1],
+                                     n_thread_data_on_grid_idx[I1],
+                                     m_thread_data_on_grid_idx[I2],
+                                     m_thread_data_on_grid_idx[I3],
+                                     m_thread_data_on_grid_idx[I4],
+                                     n_thread_data_on_grid_idx[I2]),
+                    c_element_op};
+
+            c_thread_copy.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                              c_thread_buf,
+                              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              c_grid_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
new file mode 100644
index 0000000..d1149c0
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
@@ -0,0 +1,557 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v2r3(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_shared,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M_N,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          bool BBlockLdsExtraN,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          index_t NumGemmKPrefetchStage = 1,
+          LoopScheduler LoopSched       = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_k0_n_k1;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned =
+            math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size_aligned + b_block_space_size_aligned) * sizeof(FloatAB);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / (K0PerBlock * K1);
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        using BlockwiseGemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MXdlPerWave,
+                                                                NXdlPerWave,
+                                                                K1>;
+
+        return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n);
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
+        const CGridDesc_M_N& c_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
+        decltype(MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
+    using DefaultBlock2CTileMap = decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        void* __restrict__ p_shared,
+        const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+        const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+        const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2& c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetElementSpaceSize());
+
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I0),
+                          c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I1))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, MPerBlock, K1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_k0_m_k1),
+                                                decltype(a_block_desc_k0_m_k1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_k0_m_k1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_k0_m_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_k0_n_k1),
+                                                decltype(b_block_desc_k0_n_k1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_k0_n_k1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_k0_n_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatAB,
+            FloatAcc,
+            decltype(a_block_desc_k0_m_k1),
+            decltype(b_block_desc_k0_n_k1),
+            MPerXDL,
+            NPerXDL,
+            MXdlPerWave,
+            NXdlPerWave,
+            K1,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        // gridwise GEMM pipeline
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
+
+        GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
+                                                          a_block_desc_k0_m_k1,
+                                                          a_blockwise_copy,
+                                                          a_grid_buf,
+                                                          a_block_buf,
+                                                          a_block_slice_copy_step,
+                                                          b_grid_desc_k0_n_k1,
+                                                          b_block_desc_k0_n_k1,
+                                                          b_blockwise_copy,
+                                                          b_grid_buf,
+                                                          b_block_buf,
+                                                          b_block_slice_copy_step,
+                                                          blockwise_gemm,
+                                                          c_thread_buf,
+                                                          num_k_block_main_loop);
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I7);
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_grid =
+                m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
+
+            const index_t n_thread_data_on_grid =
+                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_grid_idx =
+                m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_grid));
+
+            const auto n_thread_data_on_grid_to_n0_n1_n2_adaptor = make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_grid_idx =
+                n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_grid));
+
+            auto c_thread_copy =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   CElementwiseOperation,
+                                                   Sequence<M0, N0, I1, I1, M2, I1, M4, I1>,
+                                                   CThreadTransferSrcDstAccessOrder,
+                                                   CThreadTransferSrcDstVectorDim,
+                                                   CThreadTransferDstScalarPerVector,
+                                                   CGlobalMemoryDataOperation,
+                                                   1,
+                                                   true>{
+                    c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(m_thread_data_on_grid_idx[I0],
+                                     n_thread_data_on_grid_idx[I0],
+                                     m_thread_data_on_grid_idx[I1],
+                                     n_thread_data_on_grid_idx[I1],
+                                     m_thread_data_on_grid_idx[I2],
+                                     m_thread_data_on_grid_idx[I3],
+                                     m_thread_data_on_grid_idx[I4],
+                                     n_thread_data_on_grid_idx[I2]),
+                    c_element_op};
+
+            c_thread_copy.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                              c_thread_buf,
+                              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              c_grid_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
new file mode 100644
index 0000000..949d564
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
@@ -0,0 +1,616 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename ABK0MK1GridDesc,
+          typename BBK0NK1GridDesc,
+          typename CM0N0M1N1M2M3M4N2GridDesc,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename CBlockClusterAdaptor,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v2r4(const FloatAB* __restrict__ p_a_grid,
+                                const FloatAB* __restrict__ p_b_grid,
+                                FloatC* __restrict__ p_c_grid,
+                                const ABK0MK1GridDesc a_b_k0_m_k1_grid_desc,
+                                const BBK0NK1GridDesc b_b_k0_n_k1_grid_desc,
+                                const CM0N0M1N1M2M3M4N2GridDesc c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                                const AElementwiseOperation a_element_op,
+                                const BElementwiseOperation b_element_op,
+                                const CElementwiseOperation c_element_op,
+                                const CBlockClusterAdaptor c_block_cluster_adaptor)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_shared_block,
+                                                  a_b_k0_m_k1_grid_desc,
+                                                  b_b_k0_n_k1_grid_desc,
+                                                  c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  c_block_cluster_adaptor);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_b_k0_m_k1_grid_desc;
+    ignore = b_b_k0_n_k1_grid_desc;
+    ignore = c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = c_block_cluster_adaptor;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename ABK0MK1GridDesc,
+          typename BBK0NK1GridDesc,
+          typename CMNGridDesc,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MRepeat,
+          index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          bool BBlockLdsExtraN,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size =
+            math::integer_least_multiple(b_k0_n_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size + b_block_space_size) * sizeof(FloatAB);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const ABK0MK1GridDesc& a_b_k0_m_k1_grid_desc,
+                  const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
+                  const CMNGridDesc& c_m_n_grid_desc,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MRepeat) == 0) &&
+                          (NPerBlock % (NRepeat * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M      = a_b_k0_m_k1_grid_desc.GetLength(I2);
+        const auto N      = b_b_k0_n_k1_grid_desc.GetLength(I2);
+        const auto K0     = a_b_k0_m_k1_grid_desc.GetLength(I1);
+        const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
+
+        if(!(M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
+             K0 == b_b_k0_n_k1_grid_desc.GetLength(I1) &&
+             K1 == a_b_k0_m_k1_grid_desc.GetLength(I3) &&
+             K1 == b_b_k0_n_k1_grid_desc.GetLength(I3) &&
+             KBatch == b_b_k0_n_k1_grid_desc.GetLength(I0)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        if(!block_2_ctile_map.CheckValidity(c_m_n_grid_desc))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        const bool has_main_k0_block_loop = K0 > K0PerBlock;
+
+        return has_main_k0_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCM0N0M1N1M2M3M4N2GridDescriptor(const CMNGridDesc& c_m_n_grid_desc)
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        using BlockwiseGemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_k0_m_k1_block_desc),
+                                                                decltype(b_k0_n_k1_block_desc),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                K1>;
+
+        return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_m_n_grid_desc);
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
+        const CMNGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
+    {
+        return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CMNGridDesc>(
+            c_m_n_grid_desc, 8, KBatch);
+    }
+
+    using CM0N0M1N1M2M3M4N2GridDesc = decltype(MakeCM0N0M1N1M2M3M4N2GridDescriptor(CMNGridDesc{}));
+    using CBlockClusterAdaptor      = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1));
+
+    template <bool HasMainKBlockLoop>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               FloatAB* __restrict__ p_shared_block,
+                               const ABK0MK1GridDesc& a_b_k0_m_k1_grid_desc,
+                               const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
+                               const CM0N0M1N1M2M3M4N2GridDesc& c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const CBlockClusterAdaptor& c_block_cluster_adaptor)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetElementSpaceSize());
+
+        const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!c_block_cluster_adaptor.ValidCTileIndex(
+               make_tuple(block_work_idx[I1], block_work_idx[I2]),
+               make_tuple(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetLength(I0),
+                          c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetLength(I1))))
+        {
+            return;
+        }
+
+        const index_t k_batch_id = block_work_idx[I0];
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        constexpr auto a_b_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<K0PerBlock>{} * Number<MPerBlock + 1>{} * K1,
+                               Number<MPerBlock + 1>{} * K1,
+                               K1,
+                               I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        constexpr auto b_b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<K0PerBlock>{} * Number<NPerBlock + 1>{} * K1,
+                               Number<NPerBlock + 1>{} * K1,
+                               K1,
+                               I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, K0PerBlock, MPerBlock, K1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_b_k0_m_k1_grid_desc),
+                                                decltype(a_b_k0_m_k1_block_desc),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                a_b_k0_m_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_b_k0_m_k1_block_desc,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_b_k0_n_k1_grid_desc),
+                                                decltype(b_b_k0_n_k1_block_desc),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                b_b_k0_n_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_b_k0_n_k1_block_desc,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_k0_m_k1_block_desc),
+                                                                decltype(b_k0_n_k1_block_desc),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block = p_shared_block;
+        FloatAB* p_b_block = p_shared_block + a_block_space_size;
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
+
+        // preload data into LDS
+        {
+            a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
+            b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
+
+            a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
+            b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
+        }
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainKBlockLoop)
+        {
+            index_t k0_block_data_begin = 0;
+
+            do
+            {
+                a_blockwise_copy.MoveSrcSliceWindow(a_b_k0_m_k1_grid_desc, a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_b_k0_n_k1_grid_desc, b_block_slice_copy_step);
+
+                a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
+
+                block_sync_lds();
+
+                b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
+
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                block_sync_lds();
+
+                a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
+
+                k0_block_data_begin += K0PerBlock;
+            } while(k0_block_data_begin < (K0 - K0PerBlock));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
+            constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
+            constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
+            constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
+            constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
+            constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
+            constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
+            constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(
+                    Number<M0>{}, Number<N0>{}, I1, I1, Number<M2>{}, I1, Number<M4>{}, I1));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_grid =
+                m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
+
+            const index_t n_thread_data_on_grid =
+                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_grid_idx =
+                m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_grid));
+
+            const auto n_thread_data_on_grid_to_n0_n1_n2_adaptor = make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_grid_idx =
+                n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_grid));
+
+            auto c_thread_copy =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
+                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc),
+                                                   CElementwiseOperation,
+                                                   Sequence<M0, N0, I1, I1, M2, I1, M4, I1>,
+                                                   CThreadTransferSrcDstAccessOrder,
+                                                   CThreadTransferSrcDstVectorDim,
+                                                   CThreadTransferDstScalarPerVector,
+                                                   CGlobalMemoryDataOperation,
+                                                   1,
+                                                   true>{
+
+                    c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                    make_multi_index(m_thread_data_on_grid_idx[I0],
+                                     n_thread_data_on_grid_idx[I0],
+                                     m_thread_data_on_grid_idx[I1],
+                                     n_thread_data_on_grid_idx[I1],
+                                     m_thread_data_on_grid_idx[I2],
+                                     m_thread_data_on_grid_idx[I3],
+                                     m_thread_data_on_grid_idx[I4],
+                                     n_thread_data_on_grid_idx[I2]),
+                    c_element_op};
+
+            c_thread_copy.Run(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
+                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                              c_thread_buf,
+                              c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                              c_grid_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
new file mode 100644
index 0000000..190194f
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
@@ -0,0 +1,721 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_B_K0_M_K1,
+          typename BGridDesc_B_K0_N_K1,
+          typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename CBlockClusterAdaptor,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v2r4r2(const FloatAB* __restrict__ p_a_grid,
+                                  const FloatAB* __restrict__ p_b_grid,
+                                  FloatC* __restrict__ p_c_grid,
+                                  const AGridDesc_B_K0_M_K1 a_b_k0_m_k1_grid_desc,
+                                  const BGridDesc_B_K0_N_K1 b_b_k0_n_k1_grid_desc,
+                                  const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                                      c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                  const AElementwiseOperation a_element_op,
+                                  const BElementwiseOperation b_element_op,
+                                  const CElementwiseOperation c_element_op,
+                                  const CBlockClusterAdaptor c_block_cluster_adaptor)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  static_cast<void*>(p_shared_block),
+                                                  a_b_k0_m_k1_grid_desc,
+                                                  b_b_k0_n_k1_grid_desc,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  c_block_cluster_adaptor);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_b_k0_m_k1_grid_desc;
+    ignore = b_b_k0_n_k1_grid_desc;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = c_block_cluster_adaptor;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_B_K0_M_K1,
+          typename BGridDesc_B_K0_N_K1,
+          typename CMNGridDesc,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MRepeat,
+          index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>
+struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size =
+            math::integer_least_multiple(b_k0_n_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto c_block_size =
+            GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock().GetElementSpaceSize();
+
+        return math::max((a_block_space_size + b_block_space_size) * sizeof(FloatAB),
+                         c_block_size * sizeof(FloatC));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
+                  const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
+                  const CMNGridDesc& c_m_n_grid_desc,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MRepeat) == 0) &&
+                          (NPerBlock % (NRepeat * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M      = a_b_k0_m_k1_grid_desc.GetLength(I2);
+        const auto N      = b_b_k0_n_k1_grid_desc.GetLength(I2);
+        const auto K0     = a_b_k0_m_k1_grid_desc.GetLength(I1);
+        const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
+
+        if(!(M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
+             K0 == b_b_k0_n_k1_grid_desc.GetLength(I1) &&
+             K1 == a_b_k0_m_k1_grid_desc.GetLength(I3) &&
+             K1 == b_b_k0_n_k1_grid_desc.GetLength(I3) &&
+             KBatch == b_b_k0_n_k1_grid_desc.GetLength(I0)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        if(!block_2_ctile_map.CheckValidity(c_m_n_grid_desc))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        const bool has_main_k0_block_loop = K0 > K0PerBlock;
+
+        return has_main_k0_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CMNGridDesc& c_m_n_grid_desc)
+    {
+        const auto M = c_m_n_grid_desc.GetLength(I0);
+        const auto N = c_m_n_grid_desc.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        return transform_tensor_descriptor(
+            c_m_n_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
+        const CMNGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
+    {
+        return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CMNGridDesc>(
+            c_m_n_grid_desc, 8, KBatch);
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1,
+                       Number<CShuffleMRepeatPerShuffle * MWave * MPerXDL>{},
+                       I1,
+                       Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
+    }
+
+    using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        decltype(MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CMNGridDesc{}));
+    using CBlockClusterAdaptor = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1));
+
+    template <bool HasMainKBlockLoop>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared_block,
+                               const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
+                               const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
+                               const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const CBlockClusterAdaptor& c_block_cluster_adaptor)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!c_block_cluster_adaptor.ValidCTileIndex(
+               make_tuple(block_work_idx[I1], block_work_idx[I2]),
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        const index_t k_batch_id = block_work_idx[I0];
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        constexpr auto a_b_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<K0PerBlock>{} * Number<MPerBlock + 1>{} * K1,
+                               Number<MPerBlock + 1>{} * K1,
+                               K1,
+                               I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        constexpr auto b_b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<K0PerBlock>{} * Number<NPerBlock + 1>{} * K1,
+                               Number<NPerBlock + 1>{} * K1,
+                               K1,
+                               I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, K0PerBlock, MPerBlock, K1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_b_k0_m_k1_grid_desc),
+                                                decltype(a_b_k0_m_k1_block_desc),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                a_b_k0_m_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_b_k0_m_k1_block_desc,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_b_k0_n_k1_grid_desc),
+                                                decltype(b_b_k0_n_k1_block_desc),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                b_b_k0_n_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_b_k0_n_k1_block_desc,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_k0_m_k1_block_desc),
+                                                                decltype(b_k0_n_k1_block_desc),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block = static_cast<FloatAB*>(p_shared_block);
+        FloatAB* p_b_block = static_cast<FloatAB*>(p_shared_block) + a_block_space_size;
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
+
+        // preload data into LDS
+        {
+            a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
+            b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
+
+            a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
+            b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
+        }
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainKBlockLoop)
+        {
+            index_t k0_block_data_begin = 0;
+
+            do
+            {
+                a_blockwise_copy.MoveSrcSliceWindow(a_b_k0_m_k1_grid_desc, a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_b_k0_n_k1_grid_desc, b_block_slice_copy_step);
+
+                a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
+
+                block_sync_lds();
+
+                b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
+
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                block_sync_lds();
+
+                a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
+
+                k0_block_data_begin += K0PerBlock;
+            } while(k0_block_data_begin < (K0 - K0PerBlock));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+
+        // output: register to global memory
+        {
+            constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+            constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
+            constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
+            constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
+            constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
+            constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
+            constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
+            constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
+            constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
+
+            constexpr auto c_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatC*>(p_shared_block),
+                c_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0), // freeze mblock
+                    make_unmerge_transform(make_tuple(CShuffleMRepeatPerShuffle,
+                                                      M1,
+                                                      M2,
+                                                      M3,
+                                                      M4)), // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                    make_freeze_transform(I0),              // freeze nblock
+                    make_unmerge_transform(make_tuple(CShuffleNRepeatPerShuffle,
+                                                      N1,
+                                                      N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMRepeatPerShuffle,
+                                                            CShuffleNRepeatPerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // LDS to global
+            auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // index_t BlockSize,
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMRepeatPerShuffle * MWave * MPerXDL,
+                         1,
+                         CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
+                CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatC,               // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
+                3,                                          // index_t VectorDim,
+                CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun
+                {c_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0),
+                 c_element_op};
+
+            constexpr auto mxdlperwave_forward_step =
+                make_multi_index(0, CShuffleMRepeatPerShuffle * MWave * MPerXDL, 0, 0);
+            constexpr auto nxdlperwave_forward_step =
+                make_multi_index(0, 0, 0, CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+            constexpr auto nxdlperwave_backward_step =
+                make_multi_index(0, 0, 0, -CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+
+            static_for<0, MRepeat, CShuffleMRepeatPerShuffle>{}([&](auto mxdlperwave_iter) {
+                constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                static_for<0, NRepeat, CShuffleNRepeatPerShuffle>{}([&](auto nxdlperwave_iter) {
+                    constexpr bool nxdlperwave_forward_sweep =
+                        (mxdlperwave % (2 * CShuffleMRepeatPerShuffle) == 0);
+
+                    constexpr index_t nxdlperwave_value =
+                        nxdlperwave_forward_sweep
+                            ? nxdlperwave_iter
+                            : (NRepeat - nxdlperwave_iter - CShuffleNRepeatPerShuffle);
+
+                    constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                    // make sure it's safe to do ds_write
+                    block_sync_lds();
+
+                    // VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(
+                        c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
+                        make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                        c_thread_buf,
+                        c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        c_block_buf);
+
+                    // make sure it's safe to do ds_read
+                    block_sync_lds();
+
+                    // LDS to global
+                    c_block_copy_lds_to_global.Run(c_block_desc_mblock_mperblock_nblock_nperblock,
+                                                   c_block_buf,
+                                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                   c_grid_buf);
+
+                    // move on nxdlperwave dimension
+                    if constexpr(nxdlperwave_forward_sweep &&
+                                 (nxdlperwave < NRepeat - CShuffleNRepeatPerShuffle))
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mperblock_nblock_nperblock,
+                            nxdlperwave_forward_step);
+                    }
+                    else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mperblock_nblock_nperblock,
+                            nxdlperwave_backward_step);
+                    }
+                });
+
+                // move on mxdlperwave dimension
+                if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle)
+                {
+                    c_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, mxdlperwave_forward_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
new file mode 100644
index 0000000..ffb2926
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
@@ -0,0 +1,723 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainK0BlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v3r1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainK0BlockLoop>(
+        p_a_grid,
+        p_b_grid,
+        p_c_grid,
+        p_shared,
+        a_grid_desc_ak0_m_ak1,
+        b_grid_desc_bk0_n_bk1,
+        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        a_element_op,
+        b_element_op,
+        c_element_op,
+        block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <
+    index_t BlockSize,
+    typename FloatAB,
+    typename FloatAcc,
+    typename FloatCShuffle,
+    typename FloatC,
+    InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+    typename AGridDesc_AK0_M_AK1,
+    typename BGridDesc_BK0_N_BK1,
+    typename CGridDesc_M_N,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t KPerBlock,
+    index_t AK1Value,
+    index_t BK1Value,
+    index_t MPerXdl,
+    index_t NPerXdl,
+    index_t MXdlPerWave,
+    index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    index_t ABlockTransferSrcVectorDim,
+    index_t ABlockTransferSrcScalarPerVector,
+    index_t ABlockTransferDstScalarPerVector_K1,
+    bool AThreadTransferSrcResetCoordinateAfterRun,
+    bool ABlockLdsExtraM,
+    typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    index_t BBlockTransferSrcVectorDim,
+    index_t BBlockTransferSrcScalarPerVector,
+    index_t BBlockTransferDstScalarPerVector_K1,
+    bool BThreadTransferSrcResetCoordinateAfterRun,
+    bool BBlockLdsExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+    index_t CBlockTransferScalarPerVector_NWaveNPerXdl,
+    index_t NumGemmKPrefetchStage = 1,
+    PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        constexpr auto max_lds_align = AK1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(AK0, Number<MPerBlock>{}, AK1),
+                    make_tuple(Number<MPerBlock + 1>{} * AK1, AK1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(AK0, Number<MPerBlock>{}, AK1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_ak0_m_ak1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        constexpr auto max_lds_align = BK1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(BK0, Number<NPerBlock>{}, BK1),
+                    make_tuple(Number<NPerBlock + 1>{} * BK1, BK1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(BK0, Number<NPerBlock>{}, BK1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_bk0_n_bk1;
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto
+            c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<CShuffleMXdlPerWavePerShuffle>{},
+                               Number<MWave * MPerXdl>{},
+                               I1,
+                               Number<CShuffleNXdlPerWavePerShuffle>{},
+                               Number<NWave * NPerXdl>{}));
+
+        return c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_ak0_m_ak1.GetElementSpaceSize(), AK1);
+
+        constexpr auto b_block_space_size_aligned =
+            math::integer_least_multiple(b_block_desc_bk0_n_bk1.GetElementSpaceSize(), BK1);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+            GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
+
+        constexpr auto c_block_size =
+            c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        // static_assert(is_known_at_compile_time<remove_cv_t<decltype(AK1)>>::value &&
+        //               is_known_at_compile_time<remove_cv_t<decltype(BK1)>>::value,
+        //               "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        const auto c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+            transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_unmerge_transform(make_tuple(
+                               MBlock, Number<MXdlPerWave>{}, Number<MWave * MPerXdl>{})),
+                           make_unmerge_transform(make_tuple(
+                               NBlock, Number<NXdlPerWave>{}, Number<NWave * NPerXdl>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
+        const CGridDesc_M_N& c_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+    using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
+        remove_cvref_t<decltype(
+            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                CGridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
+
+    template <bool HasMainK0BlockLoop, typename Block2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        void* __restrict__ p_shared,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
+            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid,
+            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(
+                   c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                       .GetLength(I0),
+                   c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                       .GetLength(I3))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t k_pack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_ak0_m_ak1),
+                                                                decltype(b_block_desc_bk0_n_bk1),
+                                                                MPerXdl,
+                                                                NPerXdl,
+                                                                MXdlPerWave,
+                                                                NXdlPerWave,
+                                                                k_pack>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        GridwiseGemmPipe::template Run<HasMainK0BlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                           a_block_desc_ak0_m_ak1,
+                                                           a_blockwise_copy,
+                                                           a_grid_buf,
+                                                           a_block_buf,
+                                                           a_block_slice_copy_step,
+                                                           b_grid_desc_bk0_n_bk1,
+                                                           b_block_desc_bk0_n_bk1,
+                                                           b_blockwise_copy,
+                                                           b_grid_buf,
+                                                           b_block_buf,
+                                                           b_block_slice_copy_step,
+                                                           blockwise_gemm,
+                                                           c_thread_buf,
+                                                           num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+                GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                    .GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                make_tuple(
+                    make_freeze_transform(I0), // freeze mblock
+                    make_pass_through_transform(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}), // M0 (MXdlPerWave) per shuffle
+                    make_unmerge_transform(
+                        make_tuple(M1, M2, M3, M4)), // M1 = MWave, M2 * M3 * M4 = MPerXdl
+                    make_freeze_transform(I0),       // freeze nblock
+                    make_pass_through_transform(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}), // N0 (NXdlPerWave) per shuffle
+                    make_unmerge_transform(
+                        make_tuple(N1, N2))), // M1 = MWave, M2 * M3 * M4 = MPerXdl
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<>{},
+                           Sequence<0>{},
+                           Sequence<2, 4, 5, 6>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // LDS to global
+            auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle,
+                         MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle,
+                         NWave * NPerXdl>, // BlockSliceLengths,
+                CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+                Sequence<0, 1, 2, 3, 4, 5>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,              // typename SrcData,
+                FloatC,                     // typename DstData,
+                decltype(
+                    c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(
+                    c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                Sequence<0, 1, 2, 3, 4, 5>,                 // typename DimAccessOrder,
+                5,                                          // index_t VectorDim,
+                CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(0, 0, 0, 0, 0, 0),
+                 c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
+                 c_element_op};
+
+            constexpr auto mxdlperwave_forward_step =
+                make_multi_index(0, CShuffleMXdlPerWavePerShuffle, 0, 0, 0, 0);
+            constexpr auto nxdlperwave_forward_step =
+                make_multi_index(0, 0, 0, 0, CShuffleNXdlPerWavePerShuffle, 0);
+            constexpr auto nxdlperwave_backward_step =
+                make_multi_index(0, 0, 0, 0, -CShuffleNXdlPerWavePerShuffle, 0);
+
+            static_for<0, MXdlPerWave, CShuffleMXdlPerWavePerShuffle>{}([&](auto mxdlperwave_iter) {
+                constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                static_for<0,
+                           NXdlPerWave,
+                           CShuffleNXdlPerWavePerShuffle>{}([&](auto nxdlperwave_iter) {
+                    constexpr bool nxdlperwave_forward_sweep =
+                        (mxdlperwave % (2 * CShuffleMXdlPerWavePerShuffle) == 0);
+
+                    constexpr index_t nxdlperwave_value =
+                        nxdlperwave_forward_sweep
+                            ? nxdlperwave_iter
+                            : (NXdlPerWave - nxdlperwave_iter - CShuffleNXdlPerWavePerShuffle);
+
+                    constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                    // make sure it's safe to do ds_write
+                    block_sync_lds();
+
+                    // VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(
+                        c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                        c_thread_buf,
+                        c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        c_shuffle_block_buf);
+
+                    // make sure it's safe to do ds_read
+                    block_sync_lds();
+
+                    // LDS to global
+                    c_block_copy_lds_to_global.Run(
+                        c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c_shuffle_block_buf,
+                        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c_grid_buf);
+
+                    // move on nxdlperwave dimension
+                    if constexpr(nxdlperwave_forward_sweep &&
+                                 (nxdlperwave < NXdlPerWave - CShuffleNXdlPerWavePerShuffle))
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_forward_step);
+                    }
+                    else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_backward_step);
+                    }
+                });
+
+                // move on mxdlperwave dimension
+                if constexpr(mxdlperwave < MXdlPerWave - CShuffleMXdlPerWavePerShuffle)
+                {
+                    c_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        mxdlperwave_forward_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
new file mode 100644
index 0000000..7e6dbb3
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
@@ -0,0 +1,762 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+          typename C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v3r2(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const FloatC* __restrict__ p_c0_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+            const C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        p_a_grid,
+        p_b_grid,
+        p_c_grid,
+        p_c0_grid,
+        p_shared,
+        a_grid_desc_k0_m_k1,
+        b_grid_desc_k0_n_k1,
+        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        a_element_op,
+        b_element_op,
+        c_element_op,
+        block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_c0_grid;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    ignore = c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <
+    index_t BlockSize,
+    typename FloatAB,
+    typename FloatAcc,
+    typename FloatC,
+    InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+    typename AGridDesc_K0_M_K1,
+    typename BGridDesc_K0_N_K1,
+    typename CGridDesc_M_N,
+    typename C0GridDesc_M_N,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t K0PerBlock,
+    index_t MPerXdl,
+    index_t NPerXdl,
+    index_t K1Value,
+    index_t MXdlPerWave,
+    index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_K0_M_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    index_t ABlockTransferSrcVectorDim,
+    index_t ABlockTransferSrcScalarPerVector,
+    index_t ABlockTransferDstScalarPerVector_K1,
+    bool AThreadTransferSrcResetCoordinateAfterRun,
+    bool ABlockLdsExtraM,
+    typename BBlockTransferThreadClusterLengths_K0_N_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    index_t BBlockTransferSrcVectorDim,
+    index_t BBlockTransferSrcScalarPerVector,
+    index_t BBlockTransferDstScalarPerVector_K1,
+    bool BThreadTransferSrcResetCoordinateAfterRun,
+    bool BBlockLdsExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+    index_t CBlockTransferScalarPerVector_NWaveNPerXdl,
+    index_t NumGemmKPrefetchStage = 1,
+    PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_k0_n_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto
+            c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<CShuffleMXdlPerWavePerShuffle>{},
+                               Number<MWave * MPerXdl>{},
+                               I1,
+                               Number<CShuffleNXdlPerWavePerShuffle>{},
+                               Number<NWave * NPerXdl>{}));
+
+        return c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned =
+            math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+            GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
+
+        constexpr auto c_block_size =
+            c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatC));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / (K0PerBlock * K1);
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    template <typename CGridDesc_M_N_>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        const CGridDesc_M_N_& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        const auto c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+            transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_unmerge_transform(make_tuple(
+                               MBlock, Number<MXdlPerWave>{}, Number<MWave * MPerXdl>{})),
+                           make_unmerge_transform(make_tuple(
+                               NBlock, Number<NXdlPerWave>{}, Number<NWave * NPerXdl>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
+        const CGridDesc_M_N& c_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
+        remove_cvref_t<decltype(
+            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                CGridDesc_M_N{}))>;
+
+    using C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
+        remove_cvref_t<decltype(
+            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                C0GridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        const FloatC* __restrict__ p_c0_grid,
+        void* __restrict__ p_shared,
+        const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+        const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+        const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
+            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        const C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
+            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid,
+            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize());
+        auto c0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_grid,
+            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize());
+
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(
+                   c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                       .GetLength(I0),
+                   c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                       .GetLength(I3))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, MPerBlock, K1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_k0_m_k1),
+                                                decltype(a_block_desc_k0_m_k1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_k0_m_k1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_k0_m_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_k0_n_k1),
+                                                decltype(b_block_desc_k0_n_k1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_k0_n_k1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_k0_n_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXdl,
+                                                                NPerXdl,
+                                                                MXdlPerWave,
+                                                                NXdlPerWave,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        // gridwise GEMM pipeline
+        const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
+
+        GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
+                                                          a_block_desc_k0_m_k1,
+                                                          a_blockwise_copy,
+                                                          a_grid_buf,
+                                                          a_block_buf,
+                                                          a_block_slice_copy_step,
+                                                          b_grid_desc_k0_n_k1,
+                                                          b_block_desc_k0_n_k1,
+                                                          b_blockwise_copy,
+                                                          b_grid_buf,
+                                                          b_block_buf,
+                                                          b_block_slice_copy_step,
+                                                          blockwise_gemm,
+                                                          c_thread_buf,
+                                                          K0BlockMainLoop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+                GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
+
+            auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatC*>(p_shared),
+                c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                    .GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                make_tuple(
+                    make_freeze_transform(I0), // freeze mblock
+                    make_pass_through_transform(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}), // M0 (MXdlPerWave) per shuffle
+                    make_unmerge_transform(
+                        make_tuple(M1, M2, M3, M4)), // M1 = MWave, M2 * M3 * M4 = MPerXdl
+                    make_freeze_transform(I0),       // freeze nblock
+                    make_pass_through_transform(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}), // N0 (NXdlPerWave) per shuffle
+                    make_unmerge_transform(
+                        make_tuple(N1, N2))), // M1 = MWave, M2 * M3 * M4 = MPerXdl
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<>{},
+                           Sequence<0>{},
+                           Sequence<2, 4, 5, 6>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<3, 7>{})
+
+            );
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r2<
+                ThisThreadBlock,            // index_t BlockSize,
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle,
+                         MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle,
+                         NWave * NPerXdl>, // BlockSliceLengths,
+                CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+                Sequence<0, 1, 2, 3, 4, 5>, // typename ThreadClusterArrangeOrder,
+                FloatC,                     // typename Src0Data,
+                FloatC,                     // typename Src1Data,
+                FloatC,                     // typename DstData,
+                decltype(
+                    c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(
+                    c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(
+                    c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                Sequence<0, 1, 2, 3, 4, 5>,                 // typename DimAccessOrder,
+                5,                                          // index_t VectorDim,
+                CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrc0ResetCoordinateAfterRun,
+                false, // bool ThreadTransferSrc1ResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(0, 0, 0, 0, 0, 0),
+                 c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
+                 c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
+                 c_element_op};
+
+            constexpr auto mxdlperwave_forward_step =
+                make_multi_index(0, CShuffleMXdlPerWavePerShuffle, 0, 0, 0, 0);
+            constexpr auto nxdlperwave_forward_step =
+                make_multi_index(0, 0, 0, 0, CShuffleNXdlPerWavePerShuffle, 0);
+            constexpr auto nxdlperwave_backward_step =
+                make_multi_index(0, 0, 0, 0, -CShuffleNXdlPerWavePerShuffle, 0);
+
+            static_for<0, MXdlPerWave, CShuffleMXdlPerWavePerShuffle>{}([&](auto mxdlperwave_iter) {
+                constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                static_for<0,
+                           NXdlPerWave,
+                           CShuffleNXdlPerWavePerShuffle>{}([&](auto nxdlperwave_iter) {
+                    constexpr bool nxdlperwave_forward_sweep =
+                        (mxdlperwave % (2 * CShuffleMXdlPerWavePerShuffle) == 0);
+
+                    constexpr index_t nxdlperwave_value =
+                        nxdlperwave_forward_sweep
+                            ? nxdlperwave_iter
+                            : (NXdlPerWave - nxdlperwave_iter - CShuffleNXdlPerWavePerShuffle);
+
+                    constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                    // make sure it's safe to do ds_write
+                    block_sync_lds();
+
+                    // VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(
+                        c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                        c_thread_buf,
+                        c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        c_block_buf);
+
+                    // make sure it's safe to do ds_read
+                    block_sync_lds();
+
+                    // LDS to global
+                    c_block_copy_lds_to_global.Run(
+                        c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c_block_buf,
+                        c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c0_grid_buf,
+                        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c_grid_buf);
+
+                    // move on nxdlperwave dimension
+                    if constexpr(nxdlperwave_forward_sweep &&
+                                 (nxdlperwave < NXdlPerWave - CShuffleNXdlPerWavePerShuffle))
+                    {
+                        c_block_copy_lds_to_global.MoveSrc1SliceWindow(
+                            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_forward_step);
+
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_forward_step);
+                    }
+                    else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                    {
+                        c_block_copy_lds_to_global.MoveSrc1SliceWindow(
+                            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_backward_step);
+
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_backward_step);
+                    }
+                });
+
+                // move on mxdlperwave dimension
+                if constexpr(mxdlperwave < MXdlPerWave - CShuffleMXdlPerWavePerShuffle)
+                {
+                    c_block_copy_lds_to_global.MoveSrc1SliceWindow(
+                        c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        mxdlperwave_forward_step);
+
+                    c_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        mxdlperwave_forward_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
new file mode 100644
index 0000000..fb1e34b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
@@ -0,0 +1,801 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+          typename C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+          typename C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v3r3(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const FloatC* __restrict__ p_c0_grid,
+            const FloatC* __restrict__ p_c1_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+            const C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+            const C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        p_a_grid,
+        p_b_grid,
+        p_c_grid,
+        p_c0_grid,
+        p_c1_grid,
+        p_shared,
+        a_grid_desc_k0_m_k1,
+        b_grid_desc_k0_n_k1,
+        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        a_element_op,
+        b_element_op,
+        c_element_op,
+        block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_c0_grid;
+    ignore = p_c1_grid;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    ignore = c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    ignore = c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <
+    index_t BlockSize,
+    typename FloatAB,
+    typename FloatAcc,
+    typename FloatC,
+    InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+    typename AGridDesc_K0_M_K1,
+    typename BGridDesc_K0_N_K1,
+    typename CGridDesc_M_N,
+    typename C0GridDesc_M_N,
+    typename C1GridDesc_M_N,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t K0PerBlock,
+    index_t MPerXdl,
+    index_t NPerXdl,
+    index_t K1Value,
+    index_t MXdlPerWave,
+    index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_K0_M_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    index_t ABlockTransferSrcVectorDim,
+    index_t ABlockTransferSrcScalarPerVector,
+    index_t ABlockTransferDstScalarPerVector_K1,
+    bool AThreadTransferSrcResetCoordinateAfterRun,
+    bool ABlockLdsExtraM,
+    typename BBlockTransferThreadClusterLengths_K0_N_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    index_t BBlockTransferSrcVectorDim,
+    index_t BBlockTransferSrcScalarPerVector,
+    index_t BBlockTransferDstScalarPerVector_K1,
+    bool BThreadTransferSrcResetCoordinateAfterRun,
+    bool BBlockLdsExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+    index_t CBlockTransferScalarPerVector_NWaveNPerXdl,
+    index_t NumGemmKPrefetchStage = 1,
+    PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_k0_n_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto
+            c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<CShuffleMXdlPerWavePerShuffle>{},
+                               Number<MWave * MPerXdl>{},
+                               I1,
+                               Number<CShuffleNXdlPerWavePerShuffle>{},
+                               Number<NWave * NPerXdl>{}));
+
+        return c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned =
+            math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+            GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
+
+        constexpr auto c_block_size =
+            c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatC));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / (K0PerBlock * K1);
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    template <typename CGridDesc_M_N_>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        const CGridDesc_M_N_& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        const auto c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+            transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_unmerge_transform(make_tuple(
+                               MBlock, Number<MXdlPerWave>{}, Number<MWave * MPerXdl>{})),
+                           make_unmerge_transform(make_tuple(
+                               NBlock, Number<NXdlPerWave>{}, Number<NWave * NPerXdl>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
+        const CGridDesc_M_N& c_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+    using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
+        remove_cvref_t<decltype(
+            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                CGridDesc_M_N{}))>;
+
+    using C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
+        remove_cvref_t<decltype(
+            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                C0GridDesc_M_N{}))>;
+
+    using C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
+        remove_cvref_t<decltype(
+            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                C1GridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        const FloatC* __restrict__ p_c0_grid,
+        const FloatC* __restrict__ p_c1_grid,
+        void* __restrict__ p_shared,
+        const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+        const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+        const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
+            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        const C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
+            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        const C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
+            c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid,
+            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize());
+        auto c0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_grid,
+            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize());
+        auto c1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c1_grid,
+            c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize());
+
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(
+                   c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                       .GetLength(I0),
+                   c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                       .GetLength(I3))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, MPerBlock, K1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_k0_m_k1),
+                                                decltype(a_block_desc_k0_m_k1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                a_grid_desc_k0_m_k1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_k0_m_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_k0_n_k1),
+                                                decltype(b_block_desc_k0_n_k1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                b_grid_desc_k0_n_k1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_k0_n_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXdl,
+                                                                NPerXdl,
+                                                                MXdlPerWave,
+                                                                NXdlPerWave,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        // gridwise GEMM pipeline
+        const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
+
+        GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
+                                                          a_block_desc_k0_m_k1,
+                                                          a_blockwise_copy,
+                                                          a_grid_buf,
+                                                          a_block_buf,
+                                                          a_block_slice_copy_step,
+                                                          b_grid_desc_k0_n_k1,
+                                                          b_block_desc_k0_n_k1,
+                                                          b_blockwise_copy,
+                                                          b_grid_buf,
+                                                          b_block_buf,
+                                                          b_block_slice_copy_step,
+                                                          blockwise_gemm,
+                                                          c_thread_buf,
+                                                          K0BlockMainLoop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+                GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
+
+            auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatC*>(p_shared),
+                c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                    .GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                make_tuple(make_freeze_transform(I0), // freeze mblock
+                           make_pass_through_transform(
+                               Number<CShuffleMXdlPerWavePerShuffle>{}), // M0 (MXdlPerWave) per
+                                                                         // shuffle
+                           make_unmerge_transform(
+                               make_tuple(M1, M2, M3, M4)), // M1 = MWave, M2 * M3 * M4 = MPerXdl
+                           make_freeze_transform(I0),       // freeze nblock
+                           make_pass_through_transform(
+                               Number<CShuffleNXdlPerWavePerShuffle>{}), // N0 (NXdlPerWave) per
+                                                                         // shuffle
+                           make_unmerge_transform(
+                               make_tuple(N1, N2))), // M1 = MWave, M2 * M3 * M4 = MPerXdl
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<>{},
+                           Sequence<0>{},
+                           Sequence<2, 4, 5, 6>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<3, 7>{})
+
+            );
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r3<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle,
+                         MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle,
+                         NWave * NPerXdl>, // BlockSliceLengths,
+                CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+                Sequence<0, 1, 2, 3, 4, 5>, // typename ThreadClusterArrangeOrder,
+                FloatC,                     // typename Src0Data,
+                FloatC,                     // typename Src1Data,
+                FloatC,                     // typename Src2Data,
+                FloatC,                     // typename DstData,
+                decltype(
+                    c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(
+                    c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(
+                    c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(
+                    c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                Sequence<0, 1, 2, 3, 4, 5>,                 // typename DimAccessOrder,
+                5,                                          // index_t VectorDim,
+                CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrc0ResetCoordinateAfterRun,
+                false, // bool ThreadTransferSrc1ResetCoordinateAfterRun,
+                false, // bool ThreadTransferSrc2ResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(0, 0, 0, 0, 0, 0),
+                 c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
+                 c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
+                 c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
+                 c_element_op};
+
+            constexpr auto mxdlperwave_forward_step =
+                make_multi_index(0, CShuffleMXdlPerWavePerShuffle, 0, 0, 0, 0);
+            constexpr auto nxdlperwave_forward_step =
+                make_multi_index(0, 0, 0, 0, CShuffleNXdlPerWavePerShuffle, 0);
+            constexpr auto nxdlperwave_backward_step =
+                make_multi_index(0, 0, 0, 0, -CShuffleNXdlPerWavePerShuffle, 0);
+
+            static_for<0, MXdlPerWave, CShuffleMXdlPerWavePerShuffle>{}([&](auto mxdlperwave_iter) {
+                constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                static_for<0,
+                           NXdlPerWave,
+                           CShuffleNXdlPerWavePerShuffle>{}([&](auto nxdlperwave_iter) {
+                    constexpr bool nxdlperwave_forward_sweep =
+                        (mxdlperwave % (2 * CShuffleMXdlPerWavePerShuffle) == 0);
+
+                    constexpr index_t nxdlperwave_value =
+                        nxdlperwave_forward_sweep
+                            ? nxdlperwave_iter
+                            : (NXdlPerWave - nxdlperwave_iter - CShuffleNXdlPerWavePerShuffle);
+
+                    constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                    // make sure it's safe to do ds_write
+                    block_sync_lds();
+
+                    // VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(
+                        c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                        c_thread_buf,
+                        c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        c_block_buf);
+
+                    // make sure it's safe to do ds_read
+                    block_sync_lds();
+
+                    // LDS to global
+                    c_block_copy_lds_to_global.Run(
+                        c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c_block_buf,
+                        c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c0_grid_buf,
+                        c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c1_grid_buf,
+                        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c_grid_buf);
+
+                    // move on nxdlperwave dimension
+                    if constexpr(nxdlperwave_forward_sweep &&
+                                 (nxdlperwave < NXdlPerWave - CShuffleNXdlPerWavePerShuffle))
+                    {
+                        c_block_copy_lds_to_global.MoveSrc1SliceWindow(
+                            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_forward_step);
+
+                        c_block_copy_lds_to_global.MoveSrc2SliceWindow(
+                            c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_forward_step);
+
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_forward_step);
+                    }
+                    else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                    {
+                        c_block_copy_lds_to_global.MoveSrc1SliceWindow(
+                            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_backward_step);
+
+                        c_block_copy_lds_to_global.MoveSrc2SliceWindow(
+                            c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_backward_step);
+
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_backward_step);
+                    }
+                });
+
+                // move on mxdlperwave dimension
+                if constexpr(mxdlperwave < MXdlPerWave - CShuffleMXdlPerWavePerShuffle)
+                {
+                    c_block_copy_lds_to_global.MoveSrc1SliceWindow(
+                        c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        mxdlperwave_forward_step);
+
+                    c_block_copy_lds_to_global.MoveSrc2SliceWindow(
+                        c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        mxdlperwave_forward_step);
+
+                    c_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        mxdlperwave_forward_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_normalization_naive_variance.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_normalization_naive_variance.hpp
new file mode 100644
index 0000000..89efea4
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_normalization_naive_variance.hpp
@@ -0,0 +1,360 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// Y = Normalization(X, Beta, Gamma)
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename AccElementwiseOperation,
+          typename GridDesc_M_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorDim,
+          index_t YDstVectorSize,
+          bool SweepOnce>
+struct GridwiseNormalizationNaiveVariance_mk_to_mk
+{
+    static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((YDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using BlockwiseSumReduce = PartitionedBlockwiseReduction<AccDataType,
+                                                             BlockSize,
+                                                             ThreadClusterLengths_M_K,
+                                                             ThreadClusterArrangeOrder,
+                                                             reduce::Add,
+                                                             true>;
+
+    using ThreadwiseSumReduce = ThreadwiseReduction<AccDataType,
+                                                    ThreadReduceSrcDesc_M_K,
+                                                    ThreadReduceDstDesc_M,
+                                                    reduce::Add,
+                                                    true>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    __device__ static void Run(const GridDesc_M_K& x_grid_desc_m_k,
+                               const GridDesc_M_K& gamma_grid_desc_m_k,
+                               const GridDesc_M_K& beta_grid_desc_m_k,
+                               const GridDesc_M_K& y_grid_desc_m_k,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               const XDataType* const __restrict__ p_x_global,
+                               const GammaDataType* const __restrict__ p_gamma_global,
+                               const BetaDataType* const __restrict__ p_beta_global,
+                               YDataType* const __restrict__ p_y_global,
+                               const AccElementwiseOperation acc_elementwise_op)
+    {
+        if constexpr(SweepOnce)
+        {
+            num_k_block_tile_iteration = 1;
+        }
+
+        // LDS
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            gamma_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     AccDataType,
+                     MThreadSliceSize * KThreadSliceSize,
+                     true>& beta_thread_buf = gamma_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            y_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     AccDataType,
+                     MThreadSliceSize * KThreadSliceSize,
+                     true>& x_square_thread_buf = y_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            mean_square_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>& var_thread_buf =
+            mean_square_thread_buf;
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I)        = reduce::Add::template GetIdentityValue<AccDataType>();
+            mean_square_thread_buf(I) = reduce::Add::template GetIdentityValue<AccDataType>();
+        });
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  GridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_gamma_load =
+            ThreadwiseTensorSliceTransfer_v2<GammaDataType,
+                                             AccDataType,
+                                             GridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             GammaSrcVectorDim,
+                                             GammaSrcVectorSize,
+                                             1,
+                                             true>(
+                gamma_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_beta_load =
+            ThreadwiseTensorSliceTransfer_v2<BetaDataType,
+                                             AccDataType,
+                                             GridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             BetaSrcVectorDim,
+                                             BetaSrcVectorSize,
+                                             1,
+                                             true>(
+                beta_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               GridDesc_M_K,
+                                               AccElementwiseOperation,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               YDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                y_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * KThreadSliceSize),
+                acc_elementwise_op);
+
+        // Copy x from Cache
+        // one pass: fwd, second pass: bwd
+        constexpr auto thread_copy_fwd_step_m_k =
+            make_multi_index(0, SweepOnce ? 0 : K_BlockTileSize);
+        constexpr auto thread_copy_bwd_step_m_k =
+            make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x_global, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto gamma_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_gamma_global, gamma_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto beta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_beta_global, beta_grid_desc_m_k.GetElementSpaceSize());
+
+        // E(x), E[x^2], var(x)
+        // FIXME: Should not hack the transform from deviceOP
+        int reduce_length = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
+
+        index_t reducedTiles = 0;
+        do
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_val_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset_m_k =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                    x_square_thread_buf(Number<offset_m_k>{}) =
+                        x_thread_buf(Number<offset_m_k>{}) * x_thread_buf(Number<offset_m_k>{});
+                });
+            });
+
+            ThreadwiseSumReduce::Reduce(x_thread_buf, mean_thread_buf);
+            ThreadwiseSumReduce::Reduce(x_square_thread_buf, mean_square_thread_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+
+            ++reducedTiles;
+        } while(reducedTiles < num_k_block_tile_iteration);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            BlockwiseSumReduce::Reduce(reduce_work_buf, mean_thread_buf(I));
+            mean_thread_buf(I) = mean_thread_buf(I) / reduce_length;
+
+            block_sync_lds();
+
+            BlockwiseSumReduce::Reduce(reduce_work_buf, mean_square_thread_buf(I));
+            mean_square_thread_buf(I) = mean_square_thread_buf(I) / reduce_length;
+
+            // var(x) = E[x^2] - E[x]^2
+            var_thread_buf(I) =
+                mean_square_thread_buf(I) - (mean_thread_buf(I) * mean_thread_buf(I));
+        });
+
+        // y = (x - E[x]) / sqrt(var[x] + epsilon)
+        auto thread_copy_tail_m_k = (num_k_block_tile_iteration - 1) * thread_copy_fwd_step_m_k;
+
+        threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
+        threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
+        threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+
+        reducedTiles = 0;
+        do
+        {
+            if constexpr(!SweepOnce)
+            {
+                threadwise_x_load.Run(x_grid_desc_m_k,
+                                      x_global_val_buf,
+                                      thread_buffer_desc_m_k,
+                                      make_tuple(I0, I0),
+                                      x_thread_buf);
+            }
+
+            threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                      gamma_global_val_buf,
+                                      thread_buffer_desc_m_k,
+                                      make_tuple(I0, I0),
+                                      gamma_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset_m_k =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    // normalize
+                    y_thread_buf(Number<offset_m_k>{}) =
+                        (x_thread_buf(Number<offset_m_k>{}) - mean_thread_buf(iM)) /
+                        sqrt(var_thread_buf(iM) + epsilon);
+
+                    // gamma
+                    y_thread_buf(Number<offset_m_k>{}) =
+                        y_thread_buf(Number<offset_m_k>{}) * gamma_thread_buf(Number<offset_m_k>{});
+                });
+            });
+
+            threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                     beta_global_val_buf,
+                                     thread_buffer_desc_m_k,
+                                     make_tuple(I0, I0),
+                                     beta_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset_m_k =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    // beta
+                    y_thread_buf(Number<offset_m_k>{}) =
+                        y_thread_buf(Number<offset_m_k>{}) + beta_thread_buf(Number<offset_m_k>{});
+                });
+            });
+
+            threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   y_thread_buf,
+                                   y_grid_desc_m_k,
+                                   y_global_val_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_bwd_step_m_k);
+
+            ++reducedTiles;
+        } while(reducedTiles < num_k_block_tile_iteration);
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp
new file mode 100644
index 0000000..7aefd3c
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp
@@ -0,0 +1,384 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// Y = Normalization(X, Beta, Gamma)
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename AccElementwiseOperation,
+          typename GridDesc_M_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorDim,
+          index_t YDstVectorSize,
+          bool SweepOnce>
+struct GridwiseNormalizationWelfordVariance_mk_to_mk
+{
+    static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((YDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr index_t M_BlockTileSize     = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize     = KThreadClusterSize * KThreadSliceSize;
+    static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
+
+    static constexpr auto XThreadBufferNumber     = Number<KThreadSliceSize / XSrcVectorSize>{};
+    static constexpr auto GammaThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
+    static constexpr auto BetaThreadBufferNumber  = Number<KThreadSliceSize / XSrcVectorSize>{};
+    static constexpr auto YThreadBufferNumber     = Number<KThreadSliceSize / XSrcVectorSize>{};
+
+    __device__ static int GetKPerThread(const GridDesc_M_K& x_grid_desc_m_k,
+                                        int thread_k_cluster_id)
+    {
+        // FIXME: Should not hack the transform from deviceOP
+        int kPerBlock = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
+        int kPerThread =
+            kPerBlock < K_BlockTileSize ? 0 : KThreadSliceSize * (kPerBlock / K_BlockTileSize);
+        int kPerBlockTail = kPerBlock - kPerThread * KThreadClusterSize;
+
+        if(kPerBlockTail > 0)
+        {
+            static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
+                int thread_max_len =
+                    (thread_k_cluster_id + 1) * XSrcVectorSize + K_BlockTileStepSize * i;
+                int delta = thread_max_len - kPerBlockTail;
+                delta     = math::clamp(thread_max_len - kPerBlockTail, 0, XSrcVectorSize);
+                kPerThread += XSrcVectorSize - delta;
+            });
+        }
+
+        return kPerThread;
+    }
+
+    __device__ static void Run(const GridDesc_M_K& x_grid_desc_m_k,
+                               const GridDesc_M_K& gamma_grid_desc_m_k,
+                               const GridDesc_M_K& beta_grid_desc_m_k,
+                               const GridDesc_M_K& y_grid_desc_m_k,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               const XDataType* const __restrict__ p_x_global,
+                               const GammaDataType* const __restrict__ p_gamma_global,
+                               const BetaDataType* const __restrict__ p_beta_global,
+                               YDataType* const __restrict__ p_y_global,
+                               const AccElementwiseOperation acc_elementwise_op)
+    {
+        if constexpr(SweepOnce)
+        {
+            num_k_block_tile_iteration = 1;
+        }
+
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
+
+        auto x_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * XSrcVectorSize,
+                                    true>{};
+            },
+            Number<XThreadBufferNumber>{});
+
+        auto gamma_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * GammaSrcVectorSize,
+                                    true>{};
+            },
+            Number<GammaThreadBufferNumber>{});
+
+        auto beta_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * BetaSrcVectorSize,
+                                    true>{};
+            },
+            Number<BetaThreadBufferNumber>{});
+
+        auto y_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * YDstVectorSize,
+                                    true>{};
+            },
+            Number<YThreadBufferNumber>{});
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, XSrcVectorSize>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  GridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * XSrcVectorSize));
+
+        auto threadwise_gamma_load =
+            ThreadwiseTensorSliceTransfer_v2<GammaDataType,
+                                             AccDataType,
+                                             GridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             GammaSrcVectorDim,
+                                             GammaSrcVectorSize,
+                                             1,
+                                             true>(
+                gamma_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * GammaSrcVectorSize));
+
+        auto threadwise_beta_load =
+            ThreadwiseTensorSliceTransfer_v2<BetaDataType,
+                                             AccDataType,
+                                             GridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             BetaSrcVectorDim,
+                                             BetaSrcVectorSize,
+                                             1,
+                                             true>(
+                beta_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * BetaSrcVectorSize));
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               GridDesc_M_K,
+                                               AccElementwiseOperation,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               YDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                y_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * YDstVectorSize),
+                acc_elementwise_op);
+
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
+        constexpr auto thread_copy_bwd_step_m_k =
+            make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x_global, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto gamma_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_gamma_global, gamma_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto beta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_beta_global, beta_grid_desc_m_k.GetElementSpaceSize());
+
+        auto threadwise_welford       = ThreadwiseWelford();
+        threadwise_welford.max_count_ = GetKPerThread(x_grid_desc_m_k, thread_k_cluster_id);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
+            var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_x_load.Run(x_grid_desc_m_k,
+                                      x_global_val_buf,
+                                      thread_buffer_desc_m_k,
+                                      make_tuple(I0, I0),
+                                      x_thread_buf(i));
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
+            });
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            int count = threadwise_welford.cur_count_;
+            BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
+        });
+
+        auto thread_copy_tail_m_k =
+            (num_k_block_tile_iteration - 1) * XThreadBufferNumber * thread_copy_fwd_step_m_k;
+
+        threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
+        threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
+        threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            if constexpr(!SweepOnce)
+            {
+                static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                });
+            }
+
+            static_for<0, GammaThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                          gamma_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          gamma_thread_buf(i));
+
+                threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                         thread_copy_fwd_step_m_k);
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                auto divisor = 1 / __builtin_amdgcn_sqrtf(var_thread_buf(iM) + epsilon);
+                static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // normalize
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                            divisor;
+
+                        // gamma
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                            gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
+
+            static_for<0, BetaThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                         beta_global_val_buf,
+                                         thread_buffer_desc_m_k,
+                                         make_tuple(I0, I0),
+                                         beta_thread_buf(i));
+                threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                        thread_copy_fwd_step_m_k);
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // beta
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                            beta_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
+
+            static_for<0, YThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                       make_tuple(I0, I0),
+                                       y_thread_buf(i),
+                                       y_grid_desc_m_k,
+                                       y_global_val_buf);
+                threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_fwd_step_m_k);
+            });
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
+            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                     2 * thread_copy_bwd_step_m_k);
+            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                    2 * thread_copy_bwd_step_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp
new file mode 100644
index 0000000..de1ae91
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp
@@ -0,0 +1,339 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <numeric>
+#include <iterator>
+
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwisePermute,
+          typename InGridDesc,
+          typename OutGridDesc,
+          typename InDataType,
+          typename OutDataType,
+          typename ElementwiseOperation,
+          typename Block2TileMap>
+__global__ void kernel_nd_permute(const InGridDesc in_grid_desc,
+                                  const OutGridDesc out_grid_desc,
+                                  const InDataType* p_in_global,
+                                  OutDataType* p_out_global,
+                                  const ElementwiseOperation elementwise_op,
+                                  const Block2TileMap block_2_tile_map)
+{
+    __shared__ char p_shared[GridwisePermute::GetSharedMemoryNumberOfByte()];
+
+    GridwisePermute::Run(in_grid_desc,
+                         out_grid_desc,
+                         p_in_global,
+                         p_out_global,
+                         p_shared,
+                         elementwise_op,
+                         block_2_tile_map);
+}
+
+template <typename InGridDesc,
+          typename OutGridDesc,
+          typename InDataType,
+          typename OutDataType,
+          typename ElementwiseOperation,
+          index_t BlockSize,
+          index_t NPerBlock,
+          index_t HPerBlock,
+          index_t WPerBlock,
+          index_t InBlockLdsExtraW,
+          typename InBlockTransferThreadClusterLengths,
+          typename InBlockTransferThreadClusterArrangeOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector>
+struct GridwisePermute
+{
+    static_assert(InGridDesc::GetNumOfDimension() == OutGridDesc::GetNumOfDimension());
+    static_assert(3 <= InGridDesc::GetNumOfDimension());
+    static_assert((InGridDesc::GetNumOfDimension() - 2) <= SrcVectorDim &&
+                  SrcVectorDim < InGridDesc::GetNumOfDimension());
+    static_assert((OutGridDesc::GetNumOfDimension() - 2) <= DstVectorDim &&
+                  DstVectorDim < OutGridDesc::GetNumOfDimension());
+    static_assert(SrcVectorDim != DstVectorDim);
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    struct Block2TileMap
+    {
+        static constexpr index_t NumDim = InGridDesc::GetNumOfDimension();
+        static_assert(3 <= NumDim);
+
+        static constexpr auto I0 = Number<0>{};
+
+        Block2TileMap()                     = delete;
+        Block2TileMap(const Block2TileMap&) = default;
+        Block2TileMap(Block2TileMap&&)      = delete;
+
+        ~Block2TileMap() = default;
+
+        Block2TileMap& operator=(const Block2TileMap&) = delete;
+        Block2TileMap& operator=(Block2TileMap&&) = delete;
+
+        explicit Block2TileMap(const InGridDesc& desc) : desc_(desc) {}
+
+        __host__ constexpr index_t CalculateGridSize(const InGridDesc& desc) const
+        {
+            const auto N0 =
+                math::integer_divide_ceil(desc.GetLength(Number<NumDim - 3>{}), NPerBlock);
+            const auto H0 =
+                math::integer_divide_ceil(desc.GetLength(Number<NumDim - 2>{}), HPerBlock);
+            const auto W0 =
+                math::integer_divide_ceil(desc.GetLength(Number<NumDim - 1>{}), WPerBlock);
+
+            const index_t grid_size = N0 * H0 * W0;
+
+            return grid_size;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            static_assert(TopIdx::Size() == 1);
+
+            auto block_1d_id = idx_top[I0];
+
+            const auto N0 =
+                math::integer_divide_ceil(desc_.GetLength(Number<NumDim - 3>{}), NPerBlock);
+            const auto H0 =
+                math::integer_divide_ceil(desc_.GetLength(Number<NumDim - 2>{}), HPerBlock);
+            const auto W0 =
+                math::integer_divide_ceil(desc_.GetLength(Number<NumDim - 1>{}), WPerBlock);
+
+            block_1d_id = block_1d_id % (N0 * H0 * W0);
+
+            index_t idx_N0 = block_1d_id / (H0 * W0);
+            index_t idx_H0 = (block_1d_id % (H0 * W0)) / W0;
+            index_t idx_W0 = block_1d_id % W0;
+
+            return make_tuple(idx_N0, idx_H0, idx_W0);
+        }
+
+        private:
+        const InGridDesc desc_;
+    };
+
+    using DefaultBlock2TileMap = Block2TileMap;
+
+    // use an [NPerBlock, HPerBlock, WPerBlock] tensor as element-copy relay
+    __host__ __device__ static constexpr auto GetInBlockDesc_NPerBlock_HPerBlock_WPerBlock()
+    {
+        return make_naive_tensor_descriptor(
+            make_tuple(Number<NPerBlock>{}, Number<HPerBlock>{}, Number<WPerBlock>{}),
+            make_tuple(Number<HPerBlock*(WPerBlock + InBlockLdsExtraW)>{},
+                       Number<WPerBlock + InBlockLdsExtraW>{},
+                       I1));
+    }
+
+    // for N-dimension descriptor, reserve its last 2 dimensions, then merge its leading dimensions
+    // into single one. finally, form a 3D descriptor: [d(0), d(1), ..., d(N - 2), d(N - 1)] ->
+    // [(d(0) x d(1) x ...), d(N - 2), d(N - 1)]
+    template <typename GridDesc>
+    __host__ __device__ static constexpr auto GetMergedDesc(const GridDesc& desc)
+    {
+        constexpr index_t NumDim = GridDesc::GetNumOfDimension();
+        static_assert(3 <= NumDim);
+
+        const auto merged_desc = transform_tensor_descriptor(
+            desc,
+            make_tuple(make_merge_transform(generate_tuple(
+                           [&](auto I) { return desc.GetLength(I); }, Number<NumDim - 2>{})),
+                       make_pass_through_transform(desc.GetLength(Number<NumDim - 2>{})),
+                       make_pass_through_transform(desc.GetLength(Number<NumDim - 1>{}))),
+            make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<NumDim - 2>{}),
+                       Sequence<NumDim - 2>{},
+                       Sequence<NumDim - 1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+        return merged_desc;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto in_block_desc_nperblock_hperblock_wperblock =
+            GetInBlockDesc_NPerBlock_HPerBlock_WPerBlock();
+
+        return in_block_desc_nperblock_hperblock_wperblock.GetElementSpaceSize() *
+               sizeof(InDataType);
+    }
+
+    __host__ __device__ static constexpr auto MakeDefaultBlock2TileMap(const InGridDesc& desc)
+    {
+        return DefaultBlock2TileMap{desc};
+    }
+
+    __host__ __device__ static constexpr bool CheckValidity(const InGridDesc& in_grid_desc,
+                                                            const OutGridDesc& out_grid_desc)
+    {
+        constexpr index_t NumDim = InGridDesc::GetNumOfDimension();
+
+        // check if we only swap last 2 dimensions
+        bool valid = true;
+        static_for<0, NumDim - 2, 1>{}([&](auto I) {
+            if(valid && in_grid_desc.GetLength(I) != out_grid_desc.GetLength(I))
+            {
+                valid = false;
+            }
+        });
+
+        return valid &&
+               (in_grid_desc.GetLength(Number<NumDim - 1>{}) ==
+                out_grid_desc.GetLength(Number<NumDim - 2>{})) &&
+               (in_grid_desc.GetLength(Number<NumDim - 2>{}) ==
+                out_grid_desc.GetLength(Number<NumDim - 1>{}));
+    }
+
+    template <typename Block2TileMap>
+    __device__ static void Run(const InGridDesc in_grid_desc,
+                               const OutGridDesc out_grid_desc,
+                               const InDataType* p_in_global,
+                               OutDataType* p_out_global,
+                               void* __restrict__ p_shared,
+                               const ElementwiseOperation elementwise_op,
+                               const Block2TileMap& block_2_tile_map)
+    {
+        auto in_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_global, in_grid_desc.GetElementSpaceSize());
+
+        auto out_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_global, out_grid_desc.GetElementSpaceSize());
+
+        // each workgroup handles an [NPerBlock, HPerBlock, WPerBLock] slice-transpose problem
+        const auto block_work_idx =
+            block_2_tile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * NPerBlock);
+
+        const index_t h_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * HPerBlock);
+
+        const index_t w_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * WPerBlock);
+
+        // create [NPerBlock, HPerBlock, WPerBLock] shaped LDS buffer
+        constexpr auto in_block_desc_nperblock_hperblock_wperblock =
+            GetInBlockDesc_NPerBlock_HPerBlock_WPerBlock();
+
+        auto in_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<InDataType*>(p_shared),
+            in_block_desc_nperblock_hperblock_wperblock.GetElementSpaceSize());
+
+        using BlockSliceLengths          = Sequence<NPerBlock, HPerBlock, WPerBlock>;
+        using InBlockTransferAccessOrder = Sequence<0, 1, 2>;
+
+        constexpr index_t SrcVectorDimAfterMerge =
+            SrcVectorDim - (InGridDesc::GetNumOfDimension() - 3);
+        constexpr index_t DstVectorDimAfterMerge = SrcVectorDimAfterMerge;
+
+        using ck::tensor_operation::element_wise::PassThrough;
+
+        // merge input descriptor into [(in_grid_desc.GetLength(0) x in_grid_desc.GetLength(1) x
+        // ...), in_grid_desc.GetLength(NumDim - 2), in_grid_desc.GetLength(NumDim - 1)]
+        const auto in_grid_desc_n_h_w = GetMergedDesc(in_grid_desc);
+
+        // a workgroup copies an [NPerBlock, HPerBlock, WPerBlock] slice from global memory to LDS
+        auto in_global_load = ThreadGroupTensorSliceTransfer_v4r1<
+            ThisThreadBlock,
+            ElementwiseOperation,
+            PassThrough,
+            InMemoryDataOperationEnum::Set,
+            BlockSliceLengths,
+            InBlockTransferThreadClusterLengths,
+            InBlockTransferThreadClusterArrangeOrder,
+            InDataType,
+            InDataType,
+            decltype(in_grid_desc_n_h_w),
+            decltype(in_block_desc_nperblock_hperblock_wperblock),
+            InBlockTransferAccessOrder,
+            InBlockTransferAccessOrder,
+            SrcVectorDimAfterMerge,
+            2,
+            SrcScalarPerVector,
+            1,
+            1,
+            1,
+            true,
+            true>(in_grid_desc_n_h_w,
+                  make_multi_index(
+                      n_block_data_idx_on_grid, h_block_data_idx_on_grid, w_block_data_idx_on_grid),
+                  PassThrough{},
+                  in_block_desc_nperblock_hperblock_wperblock,
+                  make_multi_index(0, 0, 0),
+                  PassThrough{});
+
+        // merge output descriptor into [(out_grid_desc.GetLength(0) x out_grid_desc.GetLength(1) x
+        // ...), out_grid_desc.GetLength(NumDim - 2), out_grid_desc.GetLength(NumDim - 1)]
+        const auto out_grid_desc_n_w_h = GetMergedDesc(out_grid_desc);
+
+        // create transposed view of output tensor
+        const auto out_grid_desc_n_h_w = transform_tensor_descriptor(
+            out_grid_desc_n_w_h,
+            make_tuple(make_pass_through_transform(out_grid_desc_n_w_h.GetLength(I0)),
+                       make_pass_through_transform(out_grid_desc_n_w_h.GetLength(I1)),
+                       make_pass_through_transform(out_grid_desc_n_w_h.GetLength(I2))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<2>{}, Sequence<1>{}));
+
+        // a workgroup copies an [NPerBlock, HPerBlock, WPerBlock] slice from LDS to global memory
+        auto out_global_store = ThreadGroupTensorSliceTransfer_v4r1<
+            ThisThreadBlock,
+            ElementwiseOperation,
+            PassThrough,
+            InMemoryDataOperationEnum::Set,
+            BlockSliceLengths,
+            InBlockTransferThreadClusterLengths,
+            InBlockTransferThreadClusterArrangeOrder,
+            InDataType,
+            OutDataType,
+            decltype(in_block_desc_nperblock_hperblock_wperblock),
+            decltype(out_grid_desc_n_h_w),
+            InBlockTransferAccessOrder,
+            InBlockTransferAccessOrder,
+            2,
+            DstVectorDimAfterMerge,
+            1,
+            DstScalarPerVector,
+            1,
+            1,
+            true,
+            true>(in_block_desc_nperblock_hperblock_wperblock,
+                  make_multi_index(0, 0, 0),
+                  PassThrough{},
+                  out_grid_desc_n_h_w,
+                  make_multi_index(
+                      n_block_data_idx_on_grid, h_block_data_idx_on_grid, w_block_data_idx_on_grid),
+                  elementwise_op);
+
+        in_global_load.Run(in_grid_desc_n_h_w,
+                           in_global_buf,
+                           in_block_desc_nperblock_hperblock_wperblock,
+                           in_block_buf,
+                           I0);
+
+        out_global_store.Run(in_block_desc_nperblock_hperblock_wperblock,
+                             in_block_buf,
+                             out_grid_desc_n_h_w,
+                             out_global_buf,
+                             I0);
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp
new file mode 100644
index 0000000..901e7ae
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp
@@ -0,0 +1,57 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+
+namespace ck {
+
+template <index_t BlockSize, typename DataType, typename Grid1dBufferDescType>
+__global__ void kernel_buffer_set_value(const Grid1dBufferDescType grid_1d_buffer_desc,
+                                        DataType* const __restrict__ p_global,
+                                        DataType value)
+
+{
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    constexpr auto I0 = Number<0>{};
+
+    const index_t thread_local_id = get_thread_local_1d_id();
+    const index_t block_global_id = get_block_1d_id();
+
+    const index_t thread_global_id = block_global_id * BlockSize + thread_local_id;
+
+    StaticBuffer<AddressSpaceEnum::Vgpr, DataType, 1, true> value_buf;
+
+    value_buf(I0) = value;
+
+    constexpr auto val_buff_desc = make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}));
+
+    auto global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+        p_global, grid_1d_buffer_desc.GetElementSpaceSize());
+
+    if(thread_global_id < grid_1d_buffer_desc.GetElementSize())
+    {
+        auto threadwise_store = ThreadwiseTensorSliceTransfer_v1r3<DataType,
+                                                                   DataType,
+                                                                   decltype(val_buff_desc),
+                                                                   Grid1dBufferDescType,
+                                                                   PassThroughOp,
+                                                                   Sequence<1>,
+                                                                   Sequence<0>,
+                                                                   0,
+                                                                   1,
+                                                                   InMemoryDataOperationEnum::Set,
+                                                                   1,
+                                                                   true>(
+            grid_1d_buffer_desc, make_multi_index(thread_global_id), PassThroughOp{});
+
+        threadwise_store.Run(
+            val_buff_desc, make_tuple(I0), value_buf, grid_1d_buffer_desc, global_buf);
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp
new file mode 100644
index 0000000..88c7b6a
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+
+namespace ck {
+
+template <typename Grid1dBufferDescTuple,
+          index_t NumBuffer,
+          index_t BlockSize,
+          typename DataTypePointerTuple,
+          typename DataTypeTuple>
+__global__ void
+kernel_multiple_buffer_set_value(const Grid1dBufferDescTuple grid_1d_buffer_desc_tuple,
+                                 DataTypePointerTuple p_global_tuple,
+                                 DataTypeTuple value_tuple)
+
+{
+    static_assert(NumBuffer == DataTypePointerTuple::Size() && NumBuffer == DataTypeTuple::Size(),
+                  "The tuple size should be same as NumBuffer!");
+
+    static_for<0, NumBuffer, 1>{}([&](auto iB) {
+        using DataTypePointer     = remove_cvref_t<decltype(DataTypePointerTuple{}[iB])>;
+        using DataTypeFromPointer = remove_pointer_t<DataTypePointer>;
+        using DataType            = remove_cvref_t<decltype(DataTypeTuple{}[iB])>;
+
+        static_assert(is_same<DataType, DataTypeFromPointer>::value,
+                      "Types in tuples does not match!");
+    });
+
+    constexpr auto I0 = Number<0>{};
+
+    const index_t thread_global_id = get_thread_global_1d_id();
+
+    auto value_buf_tuple = generate_tuple(
+        [&](auto iB) {
+            using DataType = remove_cvref_t<decltype(DataTypeTuple{}[iB])>;
+
+            return StaticBuffer<AddressSpaceEnum::Vgpr, DataType, 1, true>{};
+        },
+        Number<NumBuffer>{});
+
+    static_for<0, NumBuffer, 1>{}([&](auto iB) {
+        static_for<0, 1, 1>{}([&](auto J) { value_buf_tuple(iB)(J) = value_tuple[iB]; });
+    });
+
+    auto global_buf_tuple = generate_tuple(
+        [&](auto iB) {
+            return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_global_tuple(iB), grid_1d_buffer_desc_tuple[iB].GetElementSpaceSize());
+        },
+        Number<NumBuffer>{});
+
+    constexpr auto val_buff_desc = make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}));
+
+    static_for<0, NumBuffer, 1>{}([&](auto iB) {
+        using DataType      = remove_cvref_t<decltype(DataTypeTuple{}[iB])>;
+        using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+        auto threadwise_store =
+            ThreadwiseTensorSliceTransfer_v1r3<DataType,
+                                               DataType,
+                                               decltype(val_buff_desc),
+                                               decltype(Grid1dBufferDescTuple{}[iB]),
+                                               PassThroughOp,
+                                               Sequence<1>,
+                                               Sequence<0>,
+                                               0,
+                                               1,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                grid_1d_buffer_desc_tuple[iB], make_multi_index(thread_global_id), PassThroughOp{});
+
+        threadwise_store.Run(val_buff_desc,
+                             make_tuple(I0),
+                             value_buf_tuple(iB),
+                             grid_1d_buffer_desc_tuple[iB],
+                             global_buf_tuple(iB));
+    });
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp
new file mode 100644
index 0000000..0344e68
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp
@@ -0,0 +1,407 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseReduction,
+          typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename GridDesc_M_K>
+__global__ void kernel_softmax(const GridDesc_M_K in_grid_desc_m_k,
+                               const GridDesc_M_K out_grid_desc_m_k,
+                               index_t block_group_size,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType alpha,
+                               const InDataType* const __restrict__ p_in_value_global,
+                               AccDataType beta,
+                               OutDataType* const __restrict__ p_out_value_global)
+{
+    GridwiseReduction::Run(in_grid_desc_m_k,
+                           out_grid_desc_m_k,
+                           block_group_size,
+                           num_k_block_tile_iteration,
+                           alpha,
+                           p_in_value_global,
+                           beta,
+                           p_out_value_global);
+};
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename GridDesc_M_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize,
+          bool SweepOnce>
+struct GridwiseSoftmax_mk_to_mk
+{
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)) &&
+                      (KThreadSliceSize % OutDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (InSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    __device__ static void Run(const GridDesc_M_K& in_grid_desc_m_k,
+                               const GridDesc_M_K& out_grid_desc_m_k,
+                               index_t block_group_size,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType alpha,
+                               const InDataType* const __restrict__ p_in_value_global,
+                               AccDataType beta,
+                               OutDataType* const __restrict__ p_out_value_global)
+    {
+        if constexpr(SweepOnce)
+        {
+            num_k_block_tile_iteration = 1;
+        }
+
+        // LDS
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        auto out_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_value_global, out_grid_desc_m_k.GetElementSpaceSize());
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            out_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> max_value_buf;
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            max_value_buf(I) = reduce::Max::template GetIdentityValue<AccDataType>();
+        });
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> accu_value_buf;
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            accu_value_buf(I) = reduce::Add::template GetIdentityValue<AccDataType>();
+        });
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / block_group_size;
+        const index_t block_local_id  = block_global_id % block_group_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        const index_t reduceSizePerBlock = K_BlockTileSize * num_k_block_tile_iteration;
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        // Normally, 0 as invalid element value is adequate since 0 makes no contribution to
+        // accumulated result. However, in stable softmax, all values 0s or not are subtracted by
+        // another value_max. As numbers become non-zero, effectively it allows invalid values to
+        // slip through and contribute to the accumulated result.
+        //
+        // The trick here is leveraging the fact that many math functions (add, sub, exp, ...)
+        // propagate NaNs when operands have NaNs involved. By initialiing invalid element value
+        // with NaN, an invalid value doing math manipulations is still NaN, which in turn can still
+        // be identified as an invalid value. We can then discard the invalid values which
+        // originally failed the bound check during accumulation. This allows to ignore values that
+        // failed bound check even after multiple math manipulations.
+        //
+        // NOTE: reset coordinate after every step because the same threadwise copy will sweep
+        // through global memory 3 times back and forth
+        auto threadwise_src_load = ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                                                    AccDataType,
+                                                                    GridDesc_M_K,
+                                                                    decltype(thread_buffer_desc),
+                                                                    ThreadBufferLengths,
+                                                                    ThreadBufferDimAccessOrder,
+                                                                    InSrcVectorDim,
+                                                                    InSrcVectorSize,
+                                                                    1,
+                                                                    true /* ResetCoordAfterRun */,
+                                                                    true /* InvalidElementAsNaN */>(
+            in_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             block_local_id * reduceSizePerBlock +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dst_load = ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                                    AccDataType,
+                                                                    GridDesc_M_K,
+                                                                    decltype(thread_buffer_desc),
+                                                                    ThreadBufferLengths,
+                                                                    ThreadBufferDimAccessOrder,
+                                                                    InSrcVectorDim,
+                                                                    InSrcVectorSize,
+                                                                    1,
+                                                                    false>(
+            out_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             block_local_id * reduceSizePerBlock +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dst_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               OutDataType,
+                                               decltype(thread_buffer_desc),
+                                               GridDesc_M_K,
+                                               PassThroughOp,
+                                               ThreadBufferLengths,
+                                               ThreadBufferDimAccessOrder,
+                                               InSrcVectorDim,
+                                               OutDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                out_grid_desc_m_k,
+                make_multi_index(
+                    blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                    block_local_id * reduceSizePerBlock + thread_k_cluster_id * KThreadSliceSize),
+                PassThroughOp{});
+
+        constexpr auto in_thread_copy_fwd_step =
+            make_multi_index(0, SweepOnce ? 0 : K_BlockTileSize);
+        constexpr auto in_thread_copy_bwd_step =
+            make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
+
+        ///
+        /// max(x)
+        ///
+        using BlockwiseMaxReduce = PartitionedBlockwiseReduction<
+            AccDataType,
+            BlockSize,
+            ThreadClusterLengths_M_K,
+            ThreadClusterArrangeOrder,
+            reduce::Max,
+            false, // param ignored
+            detail::AccumulateWithNanIgnore<reduce::Max, AccDataType>>;
+
+        using ThreadwiseMaxReduce =
+            ThreadwiseReduction<AccDataType,
+                                ThreadReduceSrcDesc_M_K,
+                                ThreadReduceDstDesc_M,
+                                reduce::Max,
+                                false, // param ignored
+                                detail::AccumulateWithNanIgnore<reduce::Max, AccDataType>>;
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global, in_grid_desc_m_k.GetElementSpaceSize());
+
+        index_t reducedTiles = 0;
+        do
+        {
+            threadwise_src_load.Run(in_grid_desc_m_k,
+                                    in_global_val_buf,
+                                    thread_buffer_desc,
+                                    make_tuple(I0, I0),
+                                    in_thread_buf);
+
+            ThreadwiseMaxReduce::Reduce(in_thread_buf, max_value_buf);
+
+            threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_fwd_step);
+
+            reducedTiles++;
+        } while(reducedTiles < num_k_block_tile_iteration);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            BlockwiseMaxReduce::Reduce(reduce_work_buf, max_value_buf(I));
+            block_sync_lds();
+        });
+
+        threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_bwd_step);
+
+        ///
+        /// sum(exp(x - max(x)))
+        ///
+        using BlockwiseSumReduce = PartitionedBlockwiseReduction<
+            AccDataType,
+            BlockSize,
+            ThreadClusterLengths_M_K,
+            ThreadClusterArrangeOrder,
+            reduce::Add,
+            false, // ignored
+            detail::AccumulateWithNanIgnore<reduce::Add, AccDataType>>;
+
+        using ThreadwiseSumReduce =
+            ThreadwiseReduction<AccDataType,
+                                ThreadReduceSrcDesc_M_K,
+                                ThreadReduceDstDesc_M,
+                                reduce::Add,
+                                false, // ignored
+                                detail::AccumulateWithNanIgnore<reduce::Add, AccDataType>>;
+
+        reducedTiles = 0;
+        do
+        {
+            if constexpr(!SweepOnce)
+            {
+                threadwise_src_load.Run(in_grid_desc_m_k,
+                                        in_global_val_buf,
+                                        thread_buffer_desc,
+                                        make_tuple(I0, I0),
+                                        in_thread_buf);
+            }
+
+            // do element-wise pre-reduction operation
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset = thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                    out_thread_buf(Number<offset>{}) =
+                        math::exp(in_thread_buf(Number<offset>{}) - max_value_buf(iM));
+                });
+            });
+
+            ThreadwiseSumReduce::Reduce(out_thread_buf, accu_value_buf);
+
+            threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_bwd_step);
+
+            reducedTiles++;
+        } while(reducedTiles < num_k_block_tile_iteration);
+
+        block_sync_lds(); // wait for reading being complete before writing to LDS
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            BlockwiseSumReduce::Reduce(reduce_work_buf, accu_value_buf(I));
+            block_sync_lds();
+        });
+
+        threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_fwd_step);
+
+        ///
+        /// softmax
+        ///
+        reducedTiles = 0;
+        if(float_equal_zero{}(beta))
+        {
+            do
+            {
+                if constexpr(!SweepOnce)
+                {
+                    threadwise_src_load.Run(in_grid_desc_m_k,
+                                            in_global_val_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_buf);
+                }
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    // out = alpha * exp(x - max(x)) / sum(exp(x - max(x)))
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                        out_thread_buf(Number<offset>{}) =
+                            alpha * math::exp(in_thread_buf(Number<offset>{}) - max_value_buf(iM)) /
+                            accu_value_buf(iM);
+                    });
+                });
+
+                threadwise_dst_store.Run(thread_buffer_desc,
+                                         make_tuple(I0, I0),
+                                         out_thread_buf,
+                                         out_grid_desc_m_k,
+                                         out_global_val_buf);
+
+                threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_fwd_step);
+                threadwise_dst_store.MoveDstSliceWindow(out_grid_desc_m_k, in_thread_copy_fwd_step);
+
+                reducedTiles++;
+            } while(reducedTiles < num_k_block_tile_iteration);
+        }
+        else
+        {
+            StaticBuffer<AddressSpaceEnum::Vgpr,
+                         AccDataType,
+                         MThreadSliceSize * KThreadSliceSize,
+                         true>
+                in_prior_dst_buf;
+            do
+            {
+                if constexpr(!SweepOnce)
+                {
+                    threadwise_src_load.Run(in_grid_desc_m_k,
+                                            in_global_val_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_buf);
+                }
+                threadwise_dst_load.Run(out_grid_desc_m_k,
+                                        out_global_val_buf,
+                                        thread_buffer_desc,
+                                        make_tuple(I0, I0),
+                                        in_prior_dst_buf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    // out = alpha * exp(x - max(x)) / sum(exp(x - max(x))) + beta * prior_out
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                        out_thread_buf(Number<offset>{}) =
+                            alpha * math::exp(in_thread_buf(Number<offset>{}) - max_value_buf(iM)) /
+                                accu_value_buf(iM) +
+                            beta * in_prior_dst_buf(Number<offset>{});
+                    });
+                });
+
+                threadwise_dst_store.Run(thread_buffer_desc,
+                                         make_tuple(I0, I0),
+                                         out_thread_buf,
+                                         out_grid_desc_m_k,
+                                         out_global_val_buf);
+
+                threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_fwd_step);
+                threadwise_dst_store.MoveDstSliceWindow(out_grid_desc_m_k, in_thread_copy_fwd_step);
+                threadwise_dst_load.MoveSrcSliceWindow(out_grid_desc_m_k, in_thread_copy_fwd_step);
+
+                reducedTiles++;
+            } while(reducedTiles < num_k_block_tile_iteration);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_sparse_embedding3_forward_layernorm.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_sparse_embedding3_forward_layernorm.hpp
new file mode 100644
index 0000000..3de6aa0
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_sparse_embedding3_forward_layernorm.hpp
@@ -0,0 +1,344 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+
+namespace ck {
+
+template <typename GridwiseSparseEmbedding,
+          typename EmbType,
+          typename IndexType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename OutType,
+          typename OutGridDesc>
+#if CK_USE_LAUNCH_BOUNDS
+__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+    __global__ void kernel_sparse_embedding3_forward_layernorm(OutType* p_out,
+                                                               const EmbType* p_emb_a,
+                                                               const EmbType* p_emb_b,
+                                                               const EmbType* p_emb_c,
+                                                               const IndexType* p_index_a,
+                                                               const IndexType* p_index_b,
+                                                               const IndexType* p_index_c,
+                                                               const GammaDataType* p_gamma,
+                                                               const BetaDataType* p_beta,
+                                                               const OutGridDesc out_grid_desc,
+                                                               const AccDataType epsilon)
+{
+    GridwiseSparseEmbedding::Run(p_out,
+                                 p_emb_a,
+                                 p_emb_b,
+                                 p_emb_c,
+                                 p_index_a,
+                                 p_index_b,
+                                 p_index_c,
+                                 p_gamma,
+                                 p_beta,
+                                 out_grid_desc,
+                                 epsilon);
+}
+
+template <typename EmbType,
+          typename IndexType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename OutType,
+          typename OutGridDesc,
+          ck::index_t BlockSize,
+          ck::index_t DimClusterSize,
+          ck::index_t RowClusterSize,
+          ck::index_t DimPerBlock,   // Row x Dim, along Dim
+          ck::index_t RowPerBlock,   // Row x Dim, along Row
+          ck::index_t DimThreadSize, // this is actually not vector, but number of registers
+          ck::index_t RowVectorSize>
+struct GridwiseSparseEmbedding3ForwardLayernorm
+{
+    static constexpr auto I0          = Number<0>{};
+    static constexpr auto I1          = Number<1>{};
+    static constexpr auto I2          = Number<2>{};
+    static constexpr auto I3          = Number<3>{};
+    static constexpr index_t WaveSize = 64;
+
+    static_assert(BlockSize == RowClusterSize * DimClusterSize,
+                  "Invalid cluster distribution within block");
+    static_assert(RowClusterSize % WaveSize == 0, "need to be wavewise");
+
+    static_assert(DimPerBlock % (DimClusterSize * DimThreadSize) == 0, "");
+    static_assert(RowPerBlock % (RowClusterSize * RowVectorSize) == 0, "");
+
+    static constexpr auto DimSubBlocks = DimPerBlock / (DimClusterSize * DimThreadSize);
+    static constexpr auto RowSubBlocks = RowPerBlock / (RowClusterSize * RowVectorSize);
+
+    static_assert((DimPerBlock % DimSubBlocks == 0) && (RowPerBlock % RowSubBlocks == 0), "");
+    static constexpr auto DimPerSubBlock = DimPerBlock / DimSubBlocks;
+    static constexpr auto RowPerSubBlock = RowPerBlock / RowSubBlocks;
+
+    using ThreadwiseWolfordDesc2D = decltype(make_naive_tensor_descriptor_packed(make_tuple(
+        Number<DimSubBlocks * DimThreadSize>{}, Number<RowSubBlocks * RowVectorSize>{})));
+
+    using ThreadwiseWolfordDescReduce = decltype(
+        make_naive_tensor_descriptor_packed(make_tuple(Number<DimSubBlocks * DimThreadSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<AccDataType, ThreadwiseWolfordDesc2D, ThreadwiseWolfordDescReduce>;
+
+    using ThreadClusterLength = Sequence<DimClusterSize, RowClusterSize>;
+
+    using BlockwiseWelford =
+        BlockwiseWelford<AccDataType, BlockSize, ThreadClusterLength, Sequence<0, 1>>;
+
+    __device__ static void Run(OutType* p_out,
+                               const EmbType* p_emb_a,
+                               const EmbType* p_emb_b,
+                               const EmbType* p_emb_c,
+                               const IndexType* p_index_a,
+                               const IndexType* p_index_b,
+                               const IndexType* p_index_c,
+                               const GammaDataType* p_gamma,
+                               const BetaDataType* p_beta,
+                               const OutGridDesc,
+                               const AccDataType epsilon)
+    {
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+
+        // const auto index_length = out_grid_desc.GetLength(I0);
+        // const auto emb_dim      = out_grid_desc.GetLength(I1);
+
+        constexpr auto thread_cluster_desc =
+            make_cluster_descriptor(Sequence<DimClusterSize, RowClusterSize>{}, Sequence<0, 1>{});
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_dim_cluster_id = thread_cluster_idx[I0];
+        const auto thread_row_cluster_id = thread_cluster_idx[I1];
+
+        const auto wave_dim_id = __builtin_amdgcn_readfirstlane(thread_dim_cluster_id / WaveSize);
+
+        const auto index_start = block_global_id * DimPerBlock + wave_dim_id * DimThreadSize;
+
+        auto threadwise_welford       = ThreadwiseWelford();
+        threadwise_welford.max_count_ = RowSubBlocks * RowVectorSize;
+
+        constexpr auto thread_buf_size =
+            DimSubBlocks * DimThreadSize * RowSubBlocks * RowVectorSize;
+        constexpr auto thread_buf_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(DimSubBlocks, DimThreadSize, RowSubBlocks, RowVectorSize));
+        constexpr auto mean_var_buf_size = DimSubBlocks * DimThreadSize;
+        constexpr auto mean_var_buf_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(DimSubBlocks, DimThreadSize));
+        constexpr auto gamma_beta_buf_size = RowSubBlocks * RowVectorSize;
+        constexpr auto gamma_beta_buf_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(RowSubBlocks, RowVectorSize));
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, EmbType, thread_buf_size, true> in_thread_buf_a;
+        StaticBuffer<AddressSpaceEnum::Vgpr, EmbType, thread_buf_size, true> in_thread_buf_b;
+        StaticBuffer<AddressSpaceEnum::Vgpr, EmbType, thread_buf_size, true> in_thread_buf_c;
+
+        StaticBuffer<AddressSpaceEnum::Sgpr, IndexType, DimPerBlock, true> index_buf_a;
+        StaticBuffer<AddressSpaceEnum::Sgpr, IndexType, DimPerBlock, true> index_buf_b;
+        StaticBuffer<AddressSpaceEnum::Sgpr, IndexType, DimPerBlock, true> index_buf_c;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, thread_buf_size, true> acc_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, gamma_beta_buf_size, true>
+            gamma_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, gamma_beta_buf_size, true>
+            beta_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, mean_var_buf_size, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, mean_var_buf_size, true> var_thread_buf;
+
+        auto load_current_sub_row = [&](auto i_dim_sub_, auto i_row_sub_) {
+            vector_type_maker_t<EmbType, RowVectorSize> emb_vector_a;
+            vector_type_maker_t<EmbType, RowVectorSize> emb_vector_b;
+            vector_type_maker_t<EmbType, RowVectorSize> emb_vector_c;
+
+            using src_vector_t = typename decltype(emb_vector_a)::type;
+            static_for<0, DimThreadSize, 1>{}([&](auto i_dim_vec_) {
+                constexpr auto current_dim = i_dim_sub_ * DimPerSubBlock + i_dim_vec_;
+                IndexType index_a          = index_buf_a[Number<current_dim>{}];
+                IndexType index_b          = index_buf_b[Number<current_dim>{}];
+                IndexType index_c          = index_buf_c[Number<current_dim>{}];
+
+                auto thread_offset = (thread_row_cluster_id + i_row_sub_ * RowClusterSize) *
+                                     sizeof(EmbType) * RowVectorSize;
+
+                int32x4_t emb_res_a =
+                    make_wave_buffer_resource_with_default_range(p_emb_a + index_a * RowPerBlock);
+                int32x4_t emb_res_b =
+                    make_wave_buffer_resource_with_default_range(p_emb_b + index_b * RowPerBlock);
+                int32x4_t emb_res_c =
+                    make_wave_buffer_resource_with_default_range(p_emb_c + index_c * RowPerBlock);
+                emb_vector_a.template AsType<src_vector_t>()(I0) =
+                    amd_buffer_load_impl<EmbType, RowVectorSize>(emb_res_a, thread_offset, 0);
+                emb_vector_b.template AsType<src_vector_t>()(I0) =
+                    amd_buffer_load_impl<EmbType, RowVectorSize>(emb_res_b, thread_offset, 0);
+                emb_vector_c.template AsType<src_vector_t>()(I0) =
+                    amd_buffer_load_impl<EmbType, RowVectorSize>(emb_res_c, thread_offset, 0);
+
+                static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
+                    constexpr auto register_offset = thread_buf_desc.CalculateOffset(
+                        make_tuple(i_dim_sub_, i_dim_vec_, i_row_sub_, i_row_vec_));
+                    in_thread_buf_a(Number<register_offset>{}) =
+                        emb_vector_a.template AsType<EmbType>()[i_row_vec_];
+                    in_thread_buf_b(Number<register_offset>{}) =
+                        emb_vector_b.template AsType<EmbType>()[i_row_vec_];
+                    in_thread_buf_c(Number<register_offset>{}) =
+                        emb_vector_c.template AsType<EmbType>()[i_row_vec_];
+                });
+            });
+        };
+
+        auto accumulate_current_sub_row = [&](auto i_dim_sub_, auto i_row_sub_) {
+            static_for<0, DimThreadSize, 1>{}([&](auto i_dim_vec_) {
+                static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
+                    constexpr auto register_offset = thread_buf_desc.CalculateOffset(
+                        make_tuple(i_dim_sub_, i_dim_vec_, i_row_sub_, i_row_vec_));
+                    AccDataType va =
+                        ck::type_convert<AccDataType>(in_thread_buf_a(Number<register_offset>{}));
+                    AccDataType vb =
+                        ck::type_convert<AccDataType>(in_thread_buf_b(Number<register_offset>{}));
+                    AccDataType vc =
+                        ck::type_convert<AccDataType>(in_thread_buf_c(Number<register_offset>{}));
+
+                    acc_thread_buf(Number<register_offset>{}) += va + vb + vc;
+                });
+            });
+        };
+
+        auto threadwise_welford_sub_row = [&](auto i_dim_sub_, auto i_row_sub_) {
+            static_for<0, DimThreadSize, 1>{}([&](auto i_dim_vec_) {
+                static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
+                    constexpr auto register_offset = thread_buf_desc.CalculateOffset(
+                        make_tuple(i_dim_sub_, i_dim_vec_, i_row_sub_, i_row_vec_));
+                    constexpr auto mean_var_offset =
+                        mean_var_buf_desc.CalculateOffset(make_tuple(i_dim_sub_, i_dim_vec_));
+
+                    threadwise_welford.cur_count_++;
+                    threadwise_welford.Update(mean_thread_buf(Number<mean_var_offset>{}),
+                                              var_thread_buf(Number<mean_var_offset>{}),
+                                              acc_thread_buf(Number<register_offset>{}));
+                });
+            });
+        };
+
+        auto threadwise_normalize_store_out = [&](auto i_dim_sub_, auto i_row_sub_) {
+            int32x4_t out_res =
+                make_wave_buffer_resource_with_default_range(p_out + index_start * RowPerBlock);
+            static_for<0, DimThreadSize, 1>{}([&](auto i_dim_vec_) {
+                vector_type_maker_t<OutType, RowVectorSize> out_vector;
+                using dst_vector_t = typename decltype(out_vector)::type;
+
+                constexpr auto mean_var_offset =
+                    mean_var_buf_desc.CalculateOffset(make_tuple(i_dim_sub_, i_dim_vec_));
+
+                static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
+                    constexpr auto register_offset = thread_buf_desc.CalculateOffset(
+                        make_tuple(i_dim_sub_, i_dim_vec_, i_row_sub_, i_row_vec_));
+                    constexpr auto gamma_beta_offset =
+                        gamma_beta_buf_desc.CalculateOffset(make_tuple(i_row_sub_, i_row_vec_));
+
+                    auto acc_val = acc_thread_buf[Number<register_offset>{}];
+                    acc_val      = (acc_val - mean_thread_buf(Number<mean_var_offset>{})) /
+                              sqrt(var_thread_buf(Number<mean_var_offset>{}) + epsilon);
+                    acc_val = acc_val * gamma_thread_buf[Number<gamma_beta_offset>{}] +
+                              beta_thread_buf[Number<gamma_beta_offset>{}];
+
+                    out_vector.template AsType<OutType>()(Number<i_row_vec_>{}) =
+                        type_convert<OutType>(acc_val);
+                });
+
+                index_t thread_offset = (thread_row_cluster_id + i_row_sub_ * RowClusterSize) *
+                                        sizeof(OutType) * RowVectorSize;
+
+                amd_buffer_store_impl<OutType, RowVectorSize>(
+                    out_vector.template AsType<dst_vector_t>()[Number<0>{}],
+                    out_res,
+                    thread_offset,
+                    0);
+            });
+        };
+
+        // first load index
+        ck::static_for<0, DimPerBlock, 1>{}([&](auto i_idx_) {
+            // prefer use s_load
+            index_buf_a(i_idx_) = p_index_a[index_start + i_idx_.value];
+            index_buf_b(i_idx_) = p_index_b[index_start + i_idx_.value];
+            index_buf_c(i_idx_) = p_index_c[index_start + i_idx_.value];
+        });
+
+        // load gamma/beta
+        static_for<0, RowSubBlocks, 1>{}([&](auto i_row_sub_) {
+            vector_type_maker_t<GammaDataType, RowVectorSize> gamma_vector;
+            vector_type_maker_t<BetaDataType, RowVectorSize> beta_vector;
+
+            index_t thread_offset_gamma = (thread_row_cluster_id + i_row_sub_ * RowClusterSize) *
+                                          sizeof(GammaDataType) * RowVectorSize;
+            index_t thread_offset_beta = (thread_row_cluster_id + i_row_sub_ * RowClusterSize) *
+                                         sizeof(BetaDataType) * RowVectorSize;
+
+            int32x4_t gamma_res = make_wave_buffer_resource_with_default_range(p_gamma);
+            int32x4_t beta_res  = make_wave_buffer_resource_with_default_range(p_beta);
+
+            gamma_vector.template AsType<typename decltype(gamma_vector)::type>()(I0) =
+                amd_buffer_load_impl<GammaDataType, RowVectorSize>(
+                    gamma_res, thread_offset_gamma, 0);
+            beta_vector.template AsType<typename decltype(beta_vector)::type>()(I0) =
+                amd_buffer_load_impl<BetaDataType, RowVectorSize>(beta_res, thread_offset_beta, 0);
+
+            static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
+                constexpr auto offset =
+                    gamma_beta_buf_desc.CalculateOffset(make_tuple(i_row_sub_, i_row_vec_));
+                gamma_thread_buf(Number<offset>{}) = type_convert<AccDataType>(
+                    gamma_vector.template AsType<GammaDataType>()[Number<i_row_vec_>{}]);
+                beta_thread_buf(Number<offset>{}) = type_convert<AccDataType>(
+                    beta_vector.template AsType<BetaDataType>()[Number<i_row_vec_>{}]);
+            });
+        });
+
+        static_for<0, thread_buf_size, 1>{}(
+            [&](auto I) { acc_thread_buf(I) = type_convert<AccDataType>(0.0f); });
+
+        static_for<0, mean_var_buf_size, 1>{}([&](auto I) {
+            mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
+            var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+        });
+
+        static_for<0, DimSubBlocks, 1>{}([&](auto i_dim_sub) {
+            load_current_sub_row(i_dim_sub, Number<0>{});
+            static_for<0, RowSubBlocks - 1, 1>{}([&](auto i_row) {
+                load_current_sub_row(i_dim_sub, Number<1>{} + i_row);
+                accumulate_current_sub_row(i_dim_sub, i_row);
+                threadwise_welford_sub_row(i_dim_sub, i_row);
+            });
+            accumulate_current_sub_row(i_dim_sub, Number<RowSubBlocks - 1>{});
+            threadwise_welford_sub_row(i_dim_sub, Number<RowSubBlocks - 1>{});
+
+            // blockwise welford
+            static_for<0, mean_var_buf_size, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                BlockwiseWelford::Run(
+                    mean_thread_buf(I), var_thread_buf(I), threadwise_welford.cur_count_);
+            });
+
+            // store
+            static_for<0, RowSubBlocks, 1>{}(
+                [&](auto i_row) { threadwise_normalize_store_out(i_dim_sub, i_row); });
+        });
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp
new file mode 100644
index 0000000..188c62d
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp
@@ -0,0 +1,99 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_functions_accumulate.hpp"
+
+namespace ck {
+
+// Assume
+//  1) SrcDesc is known at compile-time
+//  2) DstDesc is known at compile-time
+//  3) SrcBuffer is static buffer
+//  4) DstBuffer is static buffer
+template <typename AccDataType,
+          typename SrcThreadDesc_M_K,
+          typename DstThreadDesc_M,
+          typename OpReduce,
+          bool PropagateNan,
+          typename Accumulation =
+              detail::AccumulateWithNanCheck<PropagateNan, OpReduce, AccDataType>>
+struct ThreadwiseReduction
+{
+    static constexpr auto src_thread_desc_m_k = SrcThreadDesc_M_K{};
+    static constexpr auto dst_thread_desc_m   = DstThreadDesc_M{};
+
+    static constexpr auto src_length_m = src_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto src_length_k = src_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto dst_length_m = dst_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(src_length_m == dst_length_m, "lengths of source and dst buffer must match!");
+
+    using Op = OpReduce;
+
+    template <typename SrcBufferType, typename DstBufferType>
+    __device__ static void Reduce(const SrcBufferType& src_buf, DstBufferType& dst_buf)
+    {
+        static_for<0, src_length_m, 1>{}([&](auto iM) {
+            constexpr index_t out_offset = dst_thread_desc_m.CalculateOffset(make_tuple(iM));
+
+            static_for<0, src_length_k, 1>{}([&](auto iK) {
+                constexpr auto offset = src_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                Accumulation::Calculate(dst_buf(Number<out_offset>{}), src_buf[Number<offset>{}]);
+            });
+        });
+    };
+};
+
+// Assume
+//  1) SrcDesc is known at compile-time
+//  2) DstDesc is known at compile-time
+//  3) SrcBuffer is static buffer
+//  4) DstBuffer is static buffer
+template <
+    typename AccDataType,
+    typename IndexDataType,
+    typename SrcThreadDesc_M_K,
+    typename DstThreadDesc_M,
+    typename OpReduce,
+    bool PropagateNan,
+    typename Accumulation =
+        detail::AccumulateWithIndexAndNanCheck<PropagateNan, OpReduce, AccDataType, IndexDataType>>
+struct ThreadwiseReductionWithIndex
+{
+    static constexpr auto src_thread_desc_m_k = SrcThreadDesc_M_K{};
+    static constexpr auto dst_thread_desc_m   = DstThreadDesc_M{};
+
+    static constexpr auto src_length_m = src_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto src_length_k = src_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto dst_length_m = dst_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(src_length_m == dst_length_m, "lengths of source and dst buffer must match!");
+
+    template <typename SrcValueBufferType,
+              typename SrcIndexBufferType,
+              typename DstValueBufferType,
+              typename DstIndexBufferType>
+    __device__ static void Reduce(const SrcValueBufferType& src_val_buf,
+                                  const SrcIndexBufferType& src_idx_buf,
+                                  DstValueBufferType& dst_val_buf,
+                                  DstIndexBufferType& dst_idx_buf)
+    {
+        static_for<0, src_length_m, 1>{}([&](auto iM) {
+            constexpr index_t out_offset = dst_thread_desc_m.CalculateOffset(make_tuple(iM));
+
+            static_for<0, src_length_k, 1>{}([&](auto iK) {
+                constexpr auto offset = src_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                Accumulation::Calculate(dst_val_buf(Number<out_offset>{}),
+                                        src_val_buf[Number<offset>{}],
+                                        dst_idx_buf(Number<out_offset>{}),
+                                        src_idx_buf[Number<offset>{}]);
+            });
+        });
+    };
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp
new file mode 100644
index 0000000..94cdfe0
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp
@@ -0,0 +1,224 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/math.hpp"
+
+namespace ck {
+
+// C[TM0, TM1, TN0, TN1] += A[TK, TM0, TM1] * B[TK, TN0, TN1]
+//   Tensor element can be vectorized data
+// Assume:
+//   1. AThreadDesc_TK0_TM0_TM1_TK1, BThreadDesc_TK0_TN0_TN1_TK1, CThreadDesc_TM0_TM1_TN0_TN1 are
+//   known at compile-time
+//   2. AOriginIdx, BOriginIdx, COriginIdx are known at compile-time
+template <typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename AThreadDesc_TK0_TM0_TM1_TK1,
+          typename BThreadDesc_TK0_TN0_TN1_TK1,
+          typename CThreadDesc_TM0_TM1_TN0_TN1,
+          typename TKLengths,
+          typename TMLengths,
+          typename TNLengths,
+          typename enable_if<AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                                 BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                                 CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseGemmDl_km0m1_kn0n1_m0m1n0n1
+{
+    __device__ constexpr ThreadwiseGemmDl_km0m1_kn0n1_m0m1n0n1()
+    {
+        static_assert(AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                          BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                          CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        // TODO: sanity-check: compare AThreadDesc_TK0_TM0_TM1_TK1, BThreadDesc_TK0_TN0_TN1_TK1,
+        // CThreadDesc_TM0_TM1_TN0_TN1 Size with KLenghts, TMLengths and TNLengths
+
+        // TODO remove this restriction
+        static_assert(TKLengths::Size() == 1 && TMLengths::Size() == 2 && TNLengths::Size() == 2,
+                      "wrong!");
+    }
+
+    template <typename ABuffer,
+              typename AOriginIdx,
+              typename BBuffer,
+              typename BOriginIdx,
+              typename CBuffer,
+              typename COriginIdx>
+    __device__ static void Run(const ABuffer& a_buf,
+                               AOriginIdx,
+                               const BBuffer& b_buf,
+                               BOriginIdx,
+                               CBuffer& c_buf,
+                               COriginIdx)
+    {
+        static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
+                      "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
+            is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
+            is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
+            "wrong! inconsistent type");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        constexpr auto TK  = TKLengths{}[I0];
+        constexpr auto TM0 = TMLengths{}[I0];
+        constexpr auto TM1 = TMLengths{}[I1];
+        constexpr auto TN0 = TNLengths{}[I0];
+        constexpr auto TN1 = TNLengths{}[I1];
+
+        constexpr auto a_origin_idx = to_multi_index(AOriginIdx{});
+        constexpr auto b_origin_idx = to_multi_index(BOriginIdx{});
+        constexpr auto c_origin_idx = to_multi_index(COriginIdx{});
+
+        static_for<0, TK, 1>{}([&](auto tk) {
+            static_for<0, TM0, 1>{}([&](auto tm0) {
+                static_for<0, TM1, 1>{}([&](auto tm1) {
+                    static_for<0, TN0, 1>{}([&](auto tn0) {
+                        static_for<0, TN1, 1>{}([&](auto tn1) {
+                            constexpr index_t a_offset =
+                                AThreadDesc_TK0_TM0_TM1_TK1{}.CalculateOffset(
+                                    a_origin_idx + make_multi_index(tk, tm0, tm1));
+                            constexpr index_t b_offset =
+                                BThreadDesc_TK0_TN0_TN1_TK1{}.CalculateOffset(
+                                    b_origin_idx + make_multi_index(tk, tn0, tn1));
+                            constexpr index_t c_offset =
+                                CThreadDesc_TM0_TM1_TN0_TN1{}.CalculateOffset(
+                                    c_origin_idx + make_multi_index(tm0, tm1, tn0, tn1));
+
+                            inner_product<FloatA, FloatB, FloatC>(a_buf[Number<a_offset>{}],
+                                                                  b_buf[Number<b_offset>{}],
+                                                                  c_buf(Number<c_offset>{}));
+                        });
+                    });
+                });
+            });
+        });
+    }
+};
+
+// C[TM0, TM1, TN0, TN1] += A[TK0, TM0, TM1, TK1] * B[TK0, TN0, TN1, TK1]
+//   Tensor element can be vectorized data
+// Assume:
+//   1. AThreadDesc_TK0_TM0_TM1_TK1, BThreadDesc_TK0_TN0_TN1_TK1, CThreadDesc_TM0_TM1_TN0_TN1 are
+//      known at compile-time
+//   2. AOriginIdx, BOriginIdx, COriginIdx are known at compile-time
+template <typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename AThreadDesc_TK0_TM0_TM1_TK1,
+          typename BThreadDesc_TK0_TN0_TN1_TK1,
+          typename CThreadDesc_TM0_TM1_TN0_TN1,
+          typename TKLengths,
+          typename TMLengths,
+          typename TNLengths,
+          typename enable_if<AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                                 BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                                 CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseContractionDl_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1
+{
+    __device__ constexpr ThreadwiseContractionDl_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1()
+    {
+        static_assert(AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                          BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                          CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        // TODO: sanity-check: compare AThreadDesc_TK0_TM0_TM1_TK1, BThreadDesc_TK0_TN0_TN1_TK1,
+        // CThreadDesc_TM0_TM1_TN0_TN1 Size with KLenghts, TMLengths and TNLengths
+
+        // TODO remove this restriction
+        static_assert(TKLengths::Size() == 2 && TMLengths::Size() == 2 && TNLengths::Size() == 2,
+                      "wrong!");
+    }
+
+    template <typename ABuffer,
+              typename AOriginIdx,
+              typename BBuffer,
+              typename BOriginIdx,
+              typename CBuffer,
+              typename COriginIdx>
+    __device__ static void Run(const ABuffer& a_buf,
+                               AOriginIdx,
+                               const BBuffer& b_buf,
+                               BOriginIdx,
+                               CBuffer& c_buf,
+                               COriginIdx)
+    {
+        static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
+                      "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
+            is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
+            is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
+            "wrong! inconsistent type");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        constexpr index_t TK0 = TKLengths{}[I0];
+        constexpr index_t TK1 = TKLengths{}[I1];
+        constexpr index_t TM0 = TMLengths{}[I0];
+        constexpr index_t TM1 = TMLengths{}[I1];
+        constexpr index_t TN0 = TNLengths{}[I0];
+        constexpr index_t TN1 = TNLengths{}[I1];
+
+        constexpr auto a_origin_idx = to_multi_index(AOriginIdx{});
+        constexpr auto b_origin_idx = to_multi_index(BOriginIdx{});
+        constexpr auto c_origin_idx = to_multi_index(COriginIdx{});
+
+        static_for<0, TK0, 1>{}([&](auto tk0) {
+            static_for<0, TM0, 1>{}([&](auto tm0) {
+                static_for<0, TM1, 1>{}([&](auto tm1) {
+                    static_for<0, TN0, 1>{}([&](auto tn0) {
+                        static_for<0, TN1, 1>{}([&](auto tn1) {
+                            vector_type<FloatA, TK1> a_vec;
+                            vector_type<FloatB, TK1> b_vec;
+
+                            static_for<0, TK1, 1>{}([&](auto tk1) {
+                                constexpr index_t a_offset =
+                                    AThreadDesc_TK0_TM0_TM1_TK1{}.CalculateOffset(
+                                        a_origin_idx + make_multi_index(tk0, tm0, tm1, tk1));
+
+                                constexpr index_t b_offset =
+                                    BThreadDesc_TK0_TN0_TN1_TK1{}.CalculateOffset(
+                                        b_origin_idx + make_multi_index(tk0, tn0, tn1, tk1));
+
+                                a_vec.template AsType<FloatA>()(tk1) = a_buf[Number<a_offset>{}];
+                                b_vec.template AsType<FloatB>()(tk1) = b_buf[Number<b_offset>{}];
+                            });
+
+                            using a_vector_t = typename vector_type<FloatA, TK1>::type;
+                            using b_vector_t = typename vector_type<FloatB, TK1>::type;
+
+                            constexpr index_t c_offset =
+                                CThreadDesc_TM0_TM1_TN0_TN1{}.CalculateOffset(
+                                    c_origin_idx + make_multi_index(tm0, tm1, tn0, tn1));
+
+                            inner_product<a_vector_t, b_vector_t, FloatC>(
+                                a_vec.template AsType<a_vector_t>()[I0],
+                                b_vec.template AsType<b_vector_t>()[I0],
+                                c_buf(Number<c_offset>{}));
+                        });
+                    });
+                });
+            });
+        });
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp
new file mode 100644
index 0000000..e045e3b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp
@@ -0,0 +1,168 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_THREADWISE_GEMM_DLOPS_V3_HPP
+#define CK_THREADWISE_GEMM_DLOPS_V3_HPP
+
+#include "common_header.hpp"
+#include "math.hpp"
+
+namespace ck {
+
+// C[M, N] += transpose(A[K, M]) * B[K, N]
+//   Element of matrix can be vectorized data
+// Assume:
+//   1. AThreadDesc_E1_K_E2, BThreadDesc_E1_N_Ho_Wo_E2, CThreadDesc_K_N_Ho_Wo are known at
+//   compile-time
+//   2. AOriginIdx, BOriginIdx, COriginIdx are known at compile-time
+template <typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename AThreadDesc_E1_K_E2,
+          typename BThreadDesc_E1_N_Ho_Wo_E2,
+          typename CThreadDesc_K_N_Ho_Wo,
+          typename enable_if<AThreadDesc_E1_K_E2::IsKnownAtCompileTime() &&
+                                 BThreadDesc_E1_N_Ho_Wo_E2::IsKnownAtCompileTime() &&
+                                 CThreadDesc_K_N_Ho_Wo::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseGemmDlops_km_kn_mn_v3
+{
+
+    template <typename ABuffer,
+              typename AOriginIdx,
+              typename BBuffer,
+              typename BOriginIdx,
+              typename CBuffer,
+              typename COriginIdx>
+    __device__ static void Run(const ABuffer& a_buf,
+                               AOriginIdx,
+                               const BBuffer& b_buf,
+                               BOriginIdx,
+                               CBuffer& c_buf,
+                               COriginIdx)
+    {
+
+        static_assert(AThreadDesc_E1_K_E2::IsKnownAtCompileTime() &&
+                          BThreadDesc_E1_N_Ho_Wo_E2::IsKnownAtCompileTime() &&
+                          CThreadDesc_K_N_Ho_Wo::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
+                      "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
+            is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
+            is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
+            "wrong! inconsistent type");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+        constexpr auto I2 = Number<2>{};
+        constexpr auto I3 = Number<3>{};
+
+        constexpr auto E1 = AThreadDesc_E1_K_E2{}.GetLength(I0);
+        constexpr auto K  = AThreadDesc_E1_K_E2{}.GetLength(I1);
+        constexpr auto E2 = AThreadDesc_E1_K_E2{}.GetLength(I2);
+
+        constexpr auto Ho = BThreadDesc_E1_N_Ho_Wo_E2{}.GetLength(I2);
+        constexpr auto Wo = BThreadDesc_E1_N_Ho_Wo_E2{}.GetLength(I3);
+
+        constexpr auto a_origin_idx = to_multi_index(AOriginIdx{});
+        constexpr auto b_origin_idx = to_multi_index(BOriginIdx{});
+        constexpr auto c_origin_idx = to_multi_index(COriginIdx{});
+
+        if constexpr((Ho % 2 == 0) && (Wo % 2 == 0))
+        {
+            constexpr auto SubHW = 2;
+
+            static_for<0, K, 1>{}([&](auto k) {
+                static_for<0, Ho, SubHW>{}([&](auto h) {
+                    static_for<0, Wo, SubHW>{}([&](auto w) {
+                        static_for<0, E1, 1>{}([&](auto e1) {
+                            static_for<0, E2, 1>{}([&](auto e2) {
+                                constexpr index_t a_offset = AThreadDesc_E1_K_E2{}.CalculateOffset(
+                                    a_origin_idx + make_tuple(e1, k, e2));
+
+                                constexpr index_t b0_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h, w, e2));
+
+                                constexpr index_t b1_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h, w + 1, e2));
+
+                                constexpr index_t b2_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h + 1, w, e2));
+
+                                constexpr index_t b3_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h + 1, w + 1, e2));
+
+                                constexpr index_t c0_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(c_origin_idx +
+                                                                            make_tuple(k, 0, h, w));
+
+                                constexpr index_t c1_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(
+                                        c_origin_idx + make_tuple(k, 0, h, w + 1));
+
+                                constexpr index_t c2_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(
+                                        c_origin_idx + make_tuple(k, 0, h + 1, w));
+
+                                constexpr index_t c3_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(
+                                        c_origin_idx + make_tuple(k, 0, h + 1, w + 1));
+
+                                amd_assembly_outer_product_1x4(a_buf[Number<a_offset>{}],
+                                                               b_buf[Number<b0_offset>{}],
+                                                               b_buf[Number<b1_offset>{}],
+                                                               b_buf[Number<b2_offset>{}],
+                                                               b_buf[Number<b3_offset>{}],
+                                                               c_buf(Number<c0_offset>{}),
+                                                               c_buf(Number<c1_offset>{}),
+                                                               c_buf(Number<c2_offset>{}),
+                                                               c_buf(Number<c3_offset>{}));
+                            });
+                        });
+                    });
+                });
+            });
+        }
+        else
+        {
+
+            static_for<0, K, 1>{}([&](auto k) {
+                static_for<0, Ho, 1>{}([&](auto h) {
+                    static_for<0, Wo, 1>{}([&](auto w) {
+                        static_for<0, E1, 1>{}([&](auto e1) {
+                            static_for<0, E2, 1>{}([&](auto e2) {
+                                constexpr index_t a_offset = AThreadDesc_E1_K_E2{}.CalculateOffset(
+                                    a_origin_idx + make_tuple(e1, k, e2));
+
+                                constexpr index_t b_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h, w, e2));
+
+                                constexpr index_t c_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(c_origin_idx +
+                                                                            make_tuple(k, 0, h, w));
+
+                                inner_product<FloatA, FloatB, FloatC>(a_buf[Number<a_offset>{}],
+                                                                      b_buf[Number<b_offset>{}],
+                                                                      c_buf(Number<c_offset>{}));
+                            });
+                        });
+                    });
+                });
+            });
+        }
+    }
+};
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp
new file mode 100644
index 0000000..0a1197a
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// Assume:
+//   1. Desc is known at compile-time
+//   2. Buffer is StaticBuffer
+//   3. OriginIdx is known at compile-time
+//   4. use #-step
+template <typename Data,
+          typename Desc,
+          typename SliceLengths,
+          typename enable_if<Desc::IsKnownAtCompileTime(), bool>::type = false>
+struct ThreadwiseTensorSliceSet_v1
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    template <typename OriginIdx, typename Buffer>
+    __device__ void Run(const Desc&, const OriginIdx&, Buffer& buf, const Data& initial_value) const
+    {
+        static_assert(Desc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_assert(Buffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<OriginIdx>>::value,
+                      "wrong! OriginIdx need to be known at compile-time");
+
+        // Desc is known at compile-time
+        constexpr auto desc = remove_cvref_t<Desc>{};
+
+        // OriginIdx is known at compile-time
+        constexpr auto origin_idx = to_multi_index(OriginIdx{});
+
+        static_ford<SliceLengths>{}([&](auto access_idx) {
+            constexpr auto coord = make_tensor_coordinate(desc, origin_idx + access_idx);
+
+            constexpr bool is_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(desc, coord);
+
+            constexpr index_t offset = coord.GetOffset();
+
+            if constexpr(is_valid)
+            {
+                buf(Number<offset>{}) = initial_value;
+            }
+        });
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
new file mode 100644
index 0000000..b0f453b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
@@ -0,0 +1,1301 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+namespace detail {
+// TODO: How to fix this? It uses an struct instead of lambda because lambda
+// doesn't have constructor
+template <index_t VectorDim, index_t ScalarPerVector>
+struct lambda_scalar_per_access
+{
+    __host__ __device__ constexpr auto operator()(index_t i) const
+    {
+        return (i == VectorDim) ? ScalarPerVector : 1;
+    }
+};
+
+template <index_t VectorDim>
+struct lambda_scalar_step_in_vector
+{
+    __host__ __device__ constexpr auto operator()(index_t i) const
+    {
+        return (i == VectorDim) ? 1 : 0;
+    }
+};
+} // namespace detail
+
+// Assume:
+//   1. src:
+//     1. SrcDesc is known at compile-time
+//     2. SrcBuffer is StaticBuffer
+//     3. SrcSliceOrginIdx is known at compile-time
+//   2. dst:
+//     1. DstDesc is not known at compile-time
+//     2. DstBuffer is DynamicBuffer
+//     3. DstSliceOrginIdx is not known at compile time
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t DstVectorDim,
+          index_t DstScalarPerVector,
+          InMemoryDataOperationEnum DstInMemOp,
+          index_t DstScalarStrideInVector,
+          bool DstResetCoordinateAfterRun,
+          typename enable_if<SrcDesc::IsKnownAtCompileTime(), bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_v1r3
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v1r3(const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin_idx,
+                                                            const ElementwiseOperation& element_op)
+        : dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin_idx)),
+          element_op_{element_op}
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc need to known at compile-time");
+        static_assert(SliceLengths::At(Number<DstVectorDim>{}) % DstScalarPerVector == 0,
+                      "wrong! Not divisible");
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcSliceOriginIdx, typename SrcBuffer, typename DstBuffer>
+    __device__ void Run(const SrcDesc&,
+                        const SrcSliceOriginIdx&,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc need to known at compile-time");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<SrcSliceOriginIdx>>::value,
+                      "wrong! SrcSliceOrigin need to known at compile-time");
+
+        static_assert(SrcBuffer::IsStaticBuffer(), "wrong! SrcBuffer need to be StaticBuffer");
+
+        // SrcDesc and src_slice_origin_idx are known at compile-time
+        constexpr auto src_desc             = remove_cvref_t<SrcDesc>{};
+        constexpr auto src_slice_origin_idx = to_multi_index(SrcSliceOriginIdx{});
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_scalar_step_in_vector =
+            generate_sequence(detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(dst_scalar_per_access)>>;
+
+        // TODO: Use SpaceFillingCurve::ScalarsPerAccess instread of DstScalarPerVector?
+        static_assert(DstScalarPerVector == SpaceFillingCurve::ScalarPerVector,
+                      "wrong!DstScalarPerVector != SpaceFillingCurve::ScalarPerVector");
+        typename vector_type_maker<DstData, DstScalarPerVector>::type dst_vector;
+        using dst_vector_t = typename vector_type_maker<DstData, DstScalarPerVector>::type::type;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            constexpr auto idx_md = SpaceFillingCurve::GetIndex(idx_1d);
+
+            // copy data from src_buf into dst_vector
+            // TODO: It's a hack here to use \p dst_scalar_step_in_vector. Use SpaceFillingCurve?
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t src_offset = src_desc.CalculateOffset(
+                    src_slice_origin_idx + idx_md + i * dst_scalar_step_in_vector);
+
+                SrcData v;
+
+                // apply element-wise operation
+                element_op_(v, src_buf[Number<src_offset>{}]);
+
+                // apply type convert
+                dst_vector.template AsType<DstData>()(i) = type_convert<DstData>(v);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
+
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(dst_scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+}; // namespace ThreadwiseTensorSliceTransfer_v1r3
+
+// Assume:
+//   1. src:
+//     1. SrcDesc is not known at compile-time
+//     2. SrcBuffer is DynamicBuffer
+//     3. src_slice_origin_idx is not known at compile-time
+//   2. dst:
+//     1. DstDesc is known at compile-time
+//     2. DstBuffer is StaticBuffer
+//     3. dst_slice_origin_idx is known at compile-time
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t SrcVectorDim,
+          index_t SrcScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          bool SrcResetCoordinateAfterRun,
+          bool InvalidElementAsNaN                                        = false,
+          typename enable_if<DstDesc::IsKnownAtCompileTime(), bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_v2
+{
+    static_assert((InvalidElementAsNaN && !std::is_integral<DstData>::value) ||
+                      (!InvalidElementAsNaN),
+                  "Filling invalid element as NaN is only for floating point types");
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v2(const SrcDesc& src_desc,
+                                                          const Index& src_slice_origin_idx)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin_idx))
+    {
+        static_assert(DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc need to known at compile-time");
+        static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
+                      "wrong! Not divisible");
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename DstBuffer, typename DstSliceOriginIdx>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc&,
+                        const DstSliceOriginIdx&,
+                        DstBuffer& dst_buf)
+    {
+        static_assert(DstDesc::IsKnownAtCompileTime(),
+                      "wrong! DstDesc need to known at compile-time");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<DstSliceOriginIdx>>::value,
+                      "wrong! DstSliceOrigin need to known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value &&
+            "wrong! inconsistent type");
+
+        // DstDesc and dst_slice_origin_idx are known at compile-time
+        constexpr auto dst_desc             = remove_cvref_t<DstDesc>{};
+        constexpr auto dst_slice_origin_idx = DstSliceOriginIdx{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_scalar_step_in_vector =
+            generate_sequence(detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(src_scalar_per_access)>>;
+
+        // loop over tensor and copy
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            typename vector_type_maker<SrcData, SrcScalarPerVector>::type src_vector;
+
+            using src_vector_t =
+                typename vector_type_maker<SrcData, SrcScalarPerVector>::type::type;
+            constexpr auto src_data_idx = SpaceFillingCurve::GetIndex(idx_1d);
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf into src_vector
+            src_vector.template AsType<src_vector_t>()(Number<0>{}) =
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid);
+
+            // copy data from src_vector into dst_buf
+            static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t dst_offset =
+                    dst_desc.CalculateOffset(to_multi_index(dst_slice_origin_idx) + src_data_idx +
+                                             i * src_scalar_step_in_vector);
+
+                if constexpr(InvalidElementAsNaN)
+                {
+                    dst_buf(Number<dst_offset>{}) =
+                        is_src_valid
+                            ? type_convert<DstData>(src_vector.template AsType<SrcData>()[i])
+                            : NumericLimits<DstData>::QuietNaN();
+                }
+                else
+                {
+                    dst_buf(Number<dst_offset>{}) =
+                        type_convert<DstData>(src_vector.template AsType<SrcData>()[i]);
+                }
+            });
+
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+
+                move_tensor_coordinate(
+                    src_desc, src_coord_, make_tensor_coordinate_step(src_desc, forward_step));
+            }
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetSrcCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(src_scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <typename SrcMoveSliceWindowStepHack>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDesc& src_desc,
+                       const Index& src_slice_origin_step_idx,
+                       const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(
+            src_desc, adjusted_step_idx, src_move_slice_window_step_hack);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    private:
+    SrcCoord src_coord_;
+}; // namespace ck
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+//   4. Use thread buffer
+template <typename SliceLengths,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          index_t DstScalarStrideInVector,
+          bool SrcResetCoordinateAfterRun, // control whether to move back src coordinate after each
+                                           // RunRead(),  will be fused with MoveSrcSliceWindow to
+                                           // save addr computation
+          bool DstResetCoordinateAfterRun> // control whether to move back dst coordinate after each
+                                           // RunWrite(),  will be fused with MoveDstSliceWindow to
+                                           // save addr computation
+struct ThreadwiseTensorSliceTransfer_v3
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+    using Index                   = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+    using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v3(const SrcDesc& src_desc,
+                                                          const Index& src_slice_origin,
+                                                          const DstDesc& dst_desc,
+                                                          const Index& dst_slice_origin)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin))
+    {
+        static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
+                      "wrong! Not divisible");
+        static_assert(SliceLengths::At(Number<DstVectorDim>{}) % DstScalarPerVector == 0,
+                      "wrong! Not divisible");
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename SrcStepHacks>
+    __device__ void
+    RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf, const SrcStepHacks& src_step_hacks)
+    {
+        static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
+            "wrong! SrcBuffer and SrcData data type are inconsistent");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_scalar_step_in_vector =
+            generate_sequence(detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // make forward steps
+        const auto src_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    src_desc, forward_step_idx, src_step_hacks[I0][i]);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto src_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    src_desc, backward_step_idx, src_step_hacks[I1][i]);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_src_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate src data index
+            constexpr auto src_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
+                                                      : ordered_src_access_lengths[i] - 1 -
+                                                            ordered_src_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                       src_scalar_per_access;
+            }();
+
+            vector_type_maker_t<SrcData, SrcScalarPerVector> src_tmp_vector;
+
+            using src_vector_t = typename decltype(src_tmp_vector)::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf to src_tmp_vector
+            src_tmp_vector.template AsType<src_vector_t>()(Number<0>{}) =
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid);
+
+            // copy data from src_tmp_vector to buffer_
+            static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t buffer_offset =
+                    buffer_desc_.CalculateOffset(src_data_idx + i * src_scalar_step_in_vector);
+
+                buffer_(Number<buffer_offset>{}) = src_tmp_vector.template AsType<SrcData>()[i];
+            });
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_src_access_idx[i] < ordered_src_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_src_access_idx[j] == ordered_src_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_forward_steps[src_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_backward_steps[src_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetSrcCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    template <typename DstBuffer, typename DstStepHacks>
+    __device__ void
+    RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf, const DstStepHacks& dst_step_hacks)
+    {
+        static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        // src scalar per access on each dim
+        // TODO: don't use this
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_scalar_step_in_vector =
+            generate_sequence(detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // make forward steps
+        const auto dst_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    dst_desc, forward_step_idx, dst_step_hacks[I0][i]);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto dst_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    dst_desc, backward_step_idx, dst_step_hacks[I1][i]);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_dst_access_lengths)>{}([&](auto ordered_dst_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_dst_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate dst data index
+            constexpr auto dst_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_idx[i]
+                                                      : ordered_dst_access_lengths[i] - 1 -
+                                                            ordered_dst_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                       dst_scalar_per_access;
+            }();
+
+            vector_type_maker_t<DstData, DstScalarPerVector> dst_tmp_vector;
+
+            // copy data from buffer_ to dst_tmp_vector
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t buffer_offset =
+                    buffer_desc_.CalculateOffset(dst_data_idx + i * dst_scalar_step_in_vector);
+
+                dst_tmp_vector.template AsType<DstData>()(i) =
+                    type_convert<DstData>(buffer_[Number<buffer_offset>{}]);
+            });
+
+            using dst_vector_t = typename decltype(dst_tmp_vector)::type;
+
+            // copy data from dst_tmp_vector to dst_buf
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            dst_buf.template Set<dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_tmp_vector.template AsType<dst_vector_t>()[Number<0>{}]);
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_dst_access_idx[i] < ordered_dst_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_dst_access_idx[j] == ordered_dst_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_forward_steps[dst_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_backward_steps[dst_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    template <typename SrcBuffer>
+    __device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf)
+    {
+        constexpr index_t ntransform_src = SrcDesc::GetNumOfTransform();
+
+        constexpr auto zeros = typename uniform_sequence_gen<ntransform_src, 0>::type{};
+
+        constexpr auto src_step_hacks =
+            make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
+                       generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
+
+        RunRead(src_desc, src_buf, src_step_hacks);
+    }
+
+    template <typename DstBuffer>
+    __device__ void RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf)
+    {
+        constexpr index_t ntransform_dst = DstDesc::GetNumOfTransform();
+
+        constexpr auto zeros = typename uniform_sequence_gen<ntransform_dst, 0>::type{};
+
+        constexpr auto dst_step_hacks =
+            make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
+                       generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
+
+        RunWrite(dst_desc, dst_buf, dst_step_hacks);
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        constexpr auto I0 = Number<0>{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_src_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate src data index after last iteration in RunRead(), if it has not being reset by
+        // RunRead()
+        constexpr auto src_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                   src_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_src_data_step = [&]() {
+            Index reset_src_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_src_data_step_(i) = -src_data_idx[i]; });
+
+            return reset_src_data_step_;
+        }();
+
+        return reset_src_data_step;
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        constexpr auto I0 = Number<0>{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_dst_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate dst data index after last iteration in RunWrite(), if it has not being reset by
+        // RunWrite()
+        constexpr auto dst_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                   dst_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_dst_data_step = [&]() {
+            Index reset_dst_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
+
+            return reset_dst_data_step_;
+        }();
+
+        return reset_dst_data_step;
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <typename SrcMoveSliceWindowStepHack>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDesc& src_desc,
+                       const Index& src_slice_origin_step_idx,
+                       const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(
+            src_desc, adjusted_step_idx, src_move_slice_window_step_hack);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by RunWrite(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    static constexpr auto buffer_desc_ =
+        make_naive_tensor_descriptor_packed(sequence_to_tuple_of_number(SliceLengths{}));
+
+    static constexpr auto buffer_size_ = buffer_desc_.GetElementSpaceSize();
+
+    StaticBuffer<AddressSpaceEnum::Vgpr, SrcData, buffer_size_, true> buffer_;
+
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+};
+
+// Assume:
+//   1. src:
+//     1. SrcDesc is known at compile-time
+//     2. SrcBuffer is DynamicBuffer
+//     3. src_ref_idx is known at run-time
+//     4. SrcRefToOriginDisplacement is known at compile-time
+//     5. use #-step
+//   2. dst:
+//     1. DstDesc is known at compile-time
+//     2. DstBuffer is StaticBuffer
+//     3. DstOriginIdx is known at compile-time
+//     4. use direct address calculation
+//   3. vector access on src
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t SrcVectorDim,
+          index_t SrcScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          typename enable_if<SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_v4
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v4(const Index& src_ref_idx)
+        : src_ref_coord_(make_tensor_coordinate(SrcDesc{}, src_ref_idx))
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
+                      "wrong! Not divisible");
+    }
+
+    template <typename SrcRefToOriginDisplacement,
+              typename DstOriginIdx,
+              typename SrcBuffer,
+              typename DstBuffer>
+    __device__ void Run(const SrcDesc&,
+                        const SrcRefToOriginDisplacement&,
+                        const SrcBuffer& src_buf,
+                        const DstDesc&,
+                        const DstOriginIdx&,
+                        DstBuffer& dst_buf) const
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value &&
+                is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<SrcRefToOriginDisplacement>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<DstOriginIdx>>::value,
+                      "wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known "
+                      "at compile-time");
+
+        // SrcDesc and DstDesc are known at compile-time
+        constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
+        constexpr auto dst_desc = remove_cvref_t<DstDesc>{};
+
+        // SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time
+        constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{});
+        constexpr auto dst_origin_idx             = to_multi_index(DstOriginIdx{});
+
+        // scalar per access of each dim
+        constexpr auto src_scalar_per_access = generate_sequence_v2(
+            [&](auto i) constexpr {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Number<SrcScalarPerVector>{};
+                }
+                else
+                {
+                    return Number<1>{};
+                }
+            },
+            Number<nDim>{});
+
+        // scalar step (if steping on SrcVectorDim) of each dim
+        constexpr auto src_scalar_step_in_vector = generate_sequence_v2(
+            [&](auto i) constexpr {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Number<1>{};
+                }
+                else
+                {
+                    return Number<0>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto dim_access_order = DimAccessOrder{};
+
+        constexpr auto ordered_access_lengths =
+            container_reorder_given_new2old(access_lengths, dim_access_order);
+
+        static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
+#if 0
+            // TODO: unable to compile
+            // position in slice window
+            constexpr auto data_to_origin_disp_idx =
+                container_reorder_given_old2new(ordered_access_idx, dim_access_order) *
+                src_scalar_per_access;
+#else
+            // position in slice window
+            constexpr auto data_to_origin_disp_idx =
+                ordered_access_idx.ReorderGivenOld2New(dim_access_order) * src_scalar_per_access;
+#endif
+            // src coordinate
+            constexpr auto src_ref_to_data_disp_idx =
+                src_ref_to_origin_disp_idx + data_to_origin_disp_idx;
+
+            constexpr auto src_ref_to_data_disp_coord_step =
+                make_tensor_coordinate_step(src_desc, src_ref_to_data_disp_idx);
+
+            auto src_data_coord = src_ref_coord_;
+
+            move_tensor_coordinate(src_desc, src_data_coord, src_ref_to_data_disp_coord_step);
+
+            vector_type_maker_t<SrcData, SrcScalarPerVector> src_tmp_vector;
+
+            using src_vector_t = typename decltype(src_tmp_vector)::type;
+
+            const bool is_src_valid = coordinate_has_valid_offset_assuming_visible_index_is_valid(
+                src_desc, src_data_coord);
+
+            // copy data from src_buf into src_tmp_vector
+            if constexpr(SrcBuffer::IsDynamicBuffer())
+            {
+                src_tmp_vector.template AsType<src_vector_t>()(Number<0>{}) =
+                    src_buf.template Get<src_vector_t>(src_data_coord.GetOffset(), is_src_valid);
+            }
+            else if constexpr(SrcBuffer::IsStaticBuffer())
+            {
+                static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                    constexpr index_t src_offset = src_desc.CalculateOffset(
+                        src_ref_to_origin_disp_idx + data_to_origin_disp_idx +
+                        i * src_scalar_step_in_vector);
+
+                    // apply type convert
+                    src_tmp_vector.template AsType<SrcData>()(i) = src_buf[Number<src_offset>{}];
+                });
+            }
+            // copy data from src_tmp_vector to dst_tmp_vector (data cast data from SrcData to
+            // DstData)
+            vector_type_maker_t<DstData, SrcScalarPerVector> dst_tmp_vector;
+
+            // TODO: if SrcData and DstData are vetor type, then static_cast may not compile
+            static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                dst_tmp_vector.template AsType<DstData>()(i) =
+                    type_convert<DstData>(src_tmp_vector.template AsType<SrcData>()[i]);
+            });
+
+            // copy data from dst_tmp_vector into dst_buf
+            static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t dst_offset = dst_desc.CalculateOffset(
+                    dst_origin_idx + data_to_origin_disp_idx + i * src_scalar_step_in_vector);
+
+                dst_buf(Number<dst_offset>{}) = dst_tmp_vector.template AsType<DstData>()[i];
+            });
+        });
+    }
+
+    template <typename SrcSliceMoveStepIdx>
+    __device__ void MoveSrcSliceWindow(const SrcDesc&,
+                                       const SrcSliceMoveStepIdx& src_slice_move_step_idx)
+    {
+        constexpr auto src_desc = SrcDesc{};
+
+        const auto src_slice_move_step_iter =
+            make_tensor_coordinate_step(src_desc, to_multi_index(src_slice_move_step_idx));
+
+        move_tensor_coordinate(SrcDesc{}, src_ref_coord_, src_slice_move_step_iter);
+    }
+    __device__ void SetSrcCoord(const Index& src_ref_idx)
+    {
+        src_ref_coord_ = make_tensor_coordinate(SrcDesc{}, src_ref_idx);
+    }
+
+    private:
+    SrcCoord src_ref_coord_;
+};
+
+// Do NOT involve any tensor coordinates with StaticBuffer
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t DstVectorDim,
+          index_t DstScalarPerVector,
+          typename enable_if<SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_StaticToStatic
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_StaticToStatic(
+        const ElementwiseOperation& element_op)
+        : element_op_{element_op}
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc need to known at compile-time");
+
+        static_assert(SliceLengths::At(Number<DstVectorDim>{}) % DstScalarPerVector == 0,
+                      "wrong! Not divisible");
+    }
+
+    template <typename SrcSliceOriginIdx,
+              typename DstSliceOriginIdx,
+              typename SrcBuffer,
+              typename DstBuffer>
+    __device__ void Run(const SrcDesc&,
+                        const SrcSliceOriginIdx&,
+                        const SrcBuffer& src_buf,
+                        const DstDesc&,
+                        const DstSliceOriginIdx&,
+                        DstBuffer& dst_buf)
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc need to known at compile-time");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<SrcSliceOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<DstSliceOriginIdx>>::value,
+                      "wrong! SliceOrigin need to known at compile-time");
+
+        static_assert(SrcBuffer::IsStaticBuffer() && DstBuffer::IsStaticBuffer(),
+                      "wrong! Buffer need to be StaticBuffer");
+
+        // SrcDesc and src_slice_origin_idx are known at compile-time
+        constexpr auto src_desc             = remove_cvref_t<SrcDesc>{};
+        constexpr auto dst_desc             = remove_cvref_t<DstDesc>{};
+        constexpr auto src_slice_origin_idx = to_multi_index(SrcSliceOriginIdx{});
+        constexpr auto dst_slice_origin_idx = to_multi_index(DstSliceOriginIdx{});
+
+        // scalar per access on each dim
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_scalar_step_in_vector =
+            generate_sequence(detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(dst_scalar_per_access)>>;
+
+        static_assert(DstScalarPerVector == SpaceFillingCurve::ScalarPerVector,
+                      "wrong!DstScalarPerVector != SpaceFillingCurve::ScalarPerVector");
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            constexpr auto idx_md = SpaceFillingCurve::GetIndex(idx_1d);
+
+            // copy data from src_buf into dst_vector
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t src_offset = src_desc.CalculateOffset(
+                    src_slice_origin_idx + idx_md + i * dst_scalar_step_in_vector);
+
+                constexpr index_t dst_offset = dst_desc.CalculateOffset(
+                    dst_slice_origin_idx + idx_md + i * dst_scalar_step_in_vector);
+
+                SrcData v;
+
+                // apply element-wise operation
+                element_op_(v, src_buf[Number<src_offset>{}]);
+
+                // apply type convert
+                dst_buf(Number<dst_offset>{}) = type_convert<DstData>(v);
+            });
+        });
+    }
+
+    ElementwiseOperation element_op_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
new file mode 100644
index 0000000..bb28c19
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
@@ -0,0 +1,794 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor/static_tensor.hpp"
+
+namespace ck {
+
+namespace detail {
+// TODO: How to fix this? It uses an struct instead of lambda because lambda
+// doesn't have constructor
+template <index_t SrcVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstVectorDim,
+          index_t DstScalarPerVector>
+struct lambda_scalar_per_access_for_src_and_dst
+{
+    __host__ __device__ constexpr auto operator()(index_t i) const
+    {
+        if(i == SrcVectorDim && i == DstVectorDim)
+        {
+            return math::lcm(SrcScalarPerVector, DstScalarPerVector);
+        }
+        else if(i == SrcVectorDim)
+        {
+            return SrcScalarPerVector;
+        }
+        else if(i == DstVectorDim)
+        {
+            return DstScalarPerVector;
+        }
+        else
+        {
+            return 1;
+        }
+    }
+};
+
+} // namespace detail
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+//   4. Use thread buffer
+template <typename SliceLengths,
+          typename SrcElementwiseOperation,
+          typename DstElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          index_t DstScalarStrideInVector,
+          bool SrcResetCoordinateAfterRun, // control whether to move back src coordinate after each
+                                           // RunRead(),  will be fused with MoveSrcSliceWindow to
+                                           // save addr computation
+          bool DstResetCoordinateAfterRun, // control whether to move back dst coordinate after each
+                                           // RunWrite(),  will be fused with MoveDstSliceWindow to
+                                           // save addr computation
+          index_t NumThreadScratch = 1>
+struct ThreadwiseTensorSliceTransfer_v3r1
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+    using Index                   = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+    using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v3r1(
+        const SrcDesc& src_desc,
+        const Index& src_slice_origin,
+        const SrcElementwiseOperation& src_element_op,
+        const DstDesc& dst_desc,
+        const Index& dst_slice_origin,
+        const DstElementwiseOperation& dst_element_op)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          src_element_op_(src_element_op),
+          dst_element_op_(dst_element_op)
+    {
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunRead(const SrcDesc& src_desc,
+                            const SrcBuffer& src_buf,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
+            "wrong! SrcBuffer and SrcData data type are inconsistent");
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // make forward steps
+        const auto src_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(src_desc, forward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto src_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(src_desc, backward_step_idx);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_src_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate src data index
+            constexpr auto src_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
+                                                      : ordered_src_access_lengths[i] - 1 -
+                                                            ordered_src_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                       src_scalar_per_access;
+            }();
+
+            constexpr auto src_data_idx_seq = generate_sequence_v2(
+                [&](auto i) { return Number<src_data_idx[i]>{}; }, Number<src_data_idx.Size()>{});
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            using src_vector_type = vector_type_maker_t<SrcData, SrcScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            // copy data from src_buf into src_vector_container
+            auto src_vector_container = src_vector_type{
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
+
+            // apply SrcElementwiseOperation on src_vector_container
+            static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                SrcData src_v;
+
+                src_element_op_(src_v, src_vector_container.template AsType<SrcData>()[i]);
+
+                src_vector_container.template AsType<SrcData>()(i) = src_v;
+            });
+
+            // copy data from src_vector_container into src_thread_scratch_
+            src_thread_scratch_tuple_(thread_scratch_id)
+                .template SetAsType<src_vector_t>(
+                    src_data_idx_seq, src_vector_container.template AsType<src_vector_t>()[I0]);
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_src_access_idx[i] < ordered_src_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_src_access_idx[j] == ordered_src_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move src coord
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_forward_steps[src_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_backward_steps[src_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetSrcCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    template <index_t ThreadScratchId>
+    __device__ void
+    TransferDataFromSrcThreadScratchToDstThreadScratch(Number<ThreadScratchId> thread_scratch_id)
+    {
+#if !CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE
+        static_ford<SliceLengths>{}([&](auto idx) {
+            // convert from SrcData to DstData here
+            dst_thread_scratch_(idx) =
+                type_convert<DstData>(src_thread_scratch_tuple_[thread_scratch_id][idx]);
+        });
+#else
+        // sub-dword transpose between src_thread_scratch_ and dst_thread_scratch_
+        // TODO make this logic more generic for more sub-dword datatype
+        if constexpr(SrcVectorDim != DstVectorDim &&
+                     ((is_same<half_t, remove_cvref_t<SrcData>>::value &&
+                       is_same<half_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
+                      (is_same<int8_t, remove_cvref_t<SrcData>>::value &&
+                       is_same<int8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
+        {
+            // each transpose does
+            // DstScalarPerVector # of src vectors in src_thread_scratch_
+            // SrcScalarPerVector # of dst vectors in dst_thread_scratch_
+            constexpr index_t num_src_vector = Number<DstScalarPerVector>{};
+            constexpr index_t num_dst_vector = Number<SrcScalarPerVector>{};
+
+            // Assume SrcVectorDim is not the same as DstVectorDim, so we do transpose
+            // TODO: make this logic generic for all scenario
+            static_assert(SrcVectorDim != DstVectorDim, "wrong");
+
+            constexpr auto src_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+            constexpr auto dst_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+            constexpr auto scalar_per_access = generate_sequence(
+                detail::lambda_scalar_per_access_for_src_and_dst<SrcVectorDim,
+                                                                 SrcScalarPerVector,
+                                                                 DstVectorDim,
+                                                                 DstScalarPerVector>{},
+                Number<nDim>{});
+
+            constexpr auto access_lengths = SliceLengths{} / scalar_per_access;
+
+            static_ford<decltype(access_lengths)>{}([&](auto access_idx) {
+                constexpr auto data_idx = access_idx * scalar_per_access;
+
+                constexpr auto data_idx_seq = generate_sequence_v2(
+                    [&](auto i) { return Number<data_idx[i]>{}; }, Number<nDim>{});
+
+                // TODO type_convert is not used yet!!!!!
+                using src_vector_t = vector_type_maker_t<SrcData, SrcScalarPerVector>;
+                using dst_vector_t = vector_type_maker_t<DstData, DstScalarPerVector>;
+
+                // get DstScalarPerVector # of read-only references to src vectors from
+                // src_thread_scratch_
+                const auto src_vector_refs = generate_tie(
+                    [&](auto i) -> const src_vector_t& {
+                        // i increment corresponds to movement in DstVectorDim
+                        return src_thread_scratch_tuple_[thread_scratch_id].GetVectorTypeReference(
+                            data_idx_seq + i * dst_scalar_step_in_vector);
+                    },
+                    Number<num_src_vector>{});
+
+                // get SrcScalarPerVector # of references to dst vectors from dst_thread_scratch_
+                auto dst_vector_refs = generate_tie(
+                    [&](auto i) -> dst_vector_t& {
+                        // i increment corresponds to movement in SrcVectorDim
+                        return dst_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * src_scalar_step_in_vector);
+                    },
+                    Number<num_dst_vector>{});
+
+                // do data transpose
+                // TODO type_convert is not used yet!!!!!
+                transpose_vectors<SrcData, DstScalarPerVector, SrcScalarPerVector>{}(
+                    src_vector_refs, dst_vector_refs);
+            });
+        }
+        else
+        {
+            static_ford<SliceLengths>{}([&](auto idx) {
+                // convert from SrcData to DstData here
+                dst_thread_scratch_(idx) =
+                    type_convert<DstData>(src_thread_scratch_tuple_[thread_scratch_id][idx]);
+            });
+        }
+#endif
+    }
+
+    template <typename DstBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunWrite(const DstDesc& dst_desc,
+                             DstBuffer& dst_buf,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // if there is transpose, it's done here
+        // TODO move this elsewhere
+        TransferDataFromSrcThreadScratchToDstThreadScratch(thread_scratch_id);
+
+        static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        // src scalar per access on each dim
+        // TODO: don't use this
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // make forward steps
+        const auto dst_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst_desc, forward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto dst_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst_desc, backward_step_idx);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_dst_access_lengths)>{}([&](auto ordered_dst_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_dst_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate dst data index
+            constexpr auto dst_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_idx[i]
+                                                      : ordered_dst_access_lengths[i] - 1 -
+                                                            ordered_dst_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                       dst_scalar_per_access;
+            }();
+
+            constexpr auto dst_data_idx_seq = generate_sequence_v2(
+                [&](auto i) { return Number<dst_data_idx[i]>{}; }, Number<dst_data_idx.Size()>{});
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            using dst_vector_type = vector_type_maker_t<DstData, DstScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            // copy data from dst_thread_scratch_ into dst_vector_container
+            auto dst_vector_container = dst_vector_type{
+                dst_thread_scratch_.template GetAsType<dst_vector_t>(dst_data_idx_seq)};
+
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                DstData dst_v;
+
+                // apply DstElementwiseOperation
+                dst_element_op_(dst_v, dst_vector_container.template AsType<DstData>()[i]);
+
+                dst_vector_container.template AsType<DstData>()(i) = dst_v;
+            });
+
+            // copy data from dst_vector_container to dst_buf
+            dst_buf.template Set<dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_dst_access_idx[i] < ordered_dst_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_dst_access_idx[j] == ordered_dst_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move dst coord
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_forward_steps[dst_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_backward_steps[dst_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_src_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate src data index after last iteration in RunRead(), if it has not being reset by
+        // RunRead()
+        constexpr auto src_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                   src_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_src_data_step = [&]() {
+            Index reset_src_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_src_data_step_(i) = -src_data_idx[i]; });
+
+            return reset_src_data_step_;
+        }();
+
+        return reset_src_data_step;
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_dst_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate dst data index after last iteration in RunWrite(), if it has not being reset by
+        // RunWrite()
+        constexpr auto dst_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                   dst_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_dst_data_step = [&]() {
+            Index reset_dst_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
+
+            return reset_dst_data_step_;
+        }();
+
+        return reset_dst_data_step;
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by RunWrite(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    __device__ static constexpr auto GetSrcThreadScratchDescriptor()
+    {
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(src_access_lengths), Number<SrcScalarPerVector>{});
+
+        // 1st stage of transforms
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(src_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(src_access_lengths_and_vector_length[i],
+                                   src_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(src_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    __device__ static constexpr auto GetDstThreadScratchDescriptor()
+    {
+        // 1st stage of transforms
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(dst_access_lengths), Number<DstScalarPerVector>{});
+
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(dst_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(dst_access_lengths_and_vector_length[i],
+                                   dst_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(dst_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    private:
+    static constexpr auto src_thread_scratch_desc_ = decltype(GetSrcThreadScratchDescriptor()){};
+    static constexpr auto dst_thread_scratch_desc_ = decltype(GetDstThreadScratchDescriptor()){};
+
+    using SrcThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                                             SrcData,
+                                                             SrcScalarPerVector,
+                                                             decltype(src_thread_scratch_desc_),
+                                                             true>;
+
+    using DstThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                                             DstData,
+                                                             DstScalarPerVector,
+                                                             decltype(dst_thread_scratch_desc_),
+                                                             true>;
+
+    StaticallyIndexedArray<SrcThreadScratch, NumThreadScratch> src_thread_scratch_tuple_;
+
+    DstThreadScratch dst_thread_scratch_;
+
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const SrcElementwiseOperation src_element_op_;
+    const DstElementwiseOperation dst_element_op_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r3.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r3.hpp
new file mode 100644
index 0000000..6a73466
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r3.hpp
@@ -0,0 +1,886 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_THREADWISE_TENSOR_SLICE_TRANSFER_V3R3_HPP
+#define CK_THREADWISE_TENSOR_SLICE_TRANSFER_V3R3_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "static_tensor.hpp"
+
+namespace ck {
+
+namespace detail {
+// TODO: How to fix this? It uses an struct instead of lambda because lambda
+// doesn't have constructor
+template <index_t SrcVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstVectorDim,
+          index_t DstScalarPerVector>
+struct lambda_scalar_per_access_for_src_and_dst
+{
+    __host__ __device__ constexpr auto operator()(index_t i) const
+    {
+        if(i == SrcVectorDim && i == DstVectorDim)
+        {
+            return math::lcm(SrcScalarPerVector, DstScalarPerVector);
+        }
+        else if(i == SrcVectorDim)
+        {
+            return SrcScalarPerVector;
+        }
+        else if(i == DstVectorDim)
+        {
+            return DstScalarPerVector;
+        }
+        else
+        {
+            return 1;
+        }
+    }
+};
+
+} // namespace detail
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+//   4. Use thread buffer
+template <typename SliceLengths,
+          typename SrcElementwiseOperation,
+          typename DstElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename Dst0Desc,
+          typename Dst1Desc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          index_t DstScalarStrideInVector,
+          bool SrcResetCoordinateAfterRun, // control whether to move back src coordinate after each
+                                           // RunRead(),  will be fused with MoveSrcSliceWindow to
+                                           // save addr computation
+          bool DstResetCoordinateAfterRun> // control whether to move back dst coordinate after each
+                                           // RunWrite(),  will be fused with MoveDstSliceWindow to
+                                           // save addr computation
+struct ThreadwiseTensorSliceTransfer_v3r3
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+    using Index                   = MultiIndex<nDim>;
+
+    using SrcCoord  = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord  = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+    using Dst0Coord = decltype(make_tensor_coordinate(Dst0Desc{}, Index{}));
+    using Dst1Coord = decltype(make_tensor_coordinate(Dst1Desc{}, Index{}));
+
+    using SrcCoordStep  = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+    using DstCoordStep  = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
+    using Dst0CoordStep = decltype(make_tensor_coordinate_step(Dst0Desc{}, Index{}));
+    using Dst1CoordStep = decltype(make_tensor_coordinate_step(Dst1Desc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v3r3(
+        const SrcDesc& src_desc,
+        const Index& src_slice_origin,
+        const SrcElementwiseOperation& src_element_op,
+        const DstDesc& dst_desc,
+        const Dst0Desc& dst0_desc,
+        const Dst1Desc& dst1_desc,
+        const Index& dst_slice_origin,
+        const DstElementwiseOperation& dst_element_op)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          dst0_coord_(make_tensor_coordinate(dst0_desc, dst_slice_origin)),
+          dst1_coord_(make_tensor_coordinate(dst1_desc, dst_slice_origin)),
+          src_element_op_(src_element_op),
+          dst_element_op_(dst_element_op)
+    {
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc,
+                                      const Dst0Desc& dst0_desc,
+                                      const Dst1Desc& dst1_desc,
+                                      const Index& dst_slice_origin_idx)
+    {
+        dst_coord_  = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+        dst0_coord_ = make_tensor_coordinate(dst0_desc, dst_slice_origin_idx);
+        dst1_coord_ = make_tensor_coordinate(dst1_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer>
+    __device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf)
+    {
+        static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
+            "wrong! SrcBuffer and SrcData data type are inconsistent");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // make forward steps
+        const auto src_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(src_desc, forward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto src_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(src_desc, backward_step_idx);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_src_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate src data index
+            constexpr auto src_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
+                                                      : ordered_src_access_lengths[i] - 1 -
+                                                            ordered_src_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                       src_scalar_per_access;
+            }();
+
+            constexpr auto src_data_idx_seq = generate_sequence_v2(
+                [&](auto i) { return Number<src_data_idx[i]>{}; }, Number<src_data_idx.Size()>{});
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            using src_vector_type = vector_type_maker_t<SrcData, SrcScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            // copy data from src_buf into src_vector_container
+            auto src_vector_container = src_vector_type{
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
+
+            // apply SrcElementwiseOperation on src_vector_container
+            static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                src_vector_container.template AsType<SrcData>()(i) =
+                    src_element_op_(src_vector_container.template AsType<SrcData>()[i]);
+            });
+
+            // copy data from src_vector_container into src_thread_scratch_
+            src_thread_scratch_.template SetAsType<src_vector_t>(
+                src_data_idx_seq, src_vector_container.template AsType<src_vector_t>()[I0]);
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_src_access_idx[i] < ordered_src_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_src_access_idx[j] == ordered_src_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move src coord
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_forward_steps[src_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_backward_steps[src_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetSrcCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    __device__ void TransferDataFromSrcThreadScratchToDstThreadScratch()
+    {
+#if !CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE
+        static_ford<SliceLengths>{}([&](auto idx) {
+            // convert from SrcData to DstData here
+            dst_thread_scratch_(idx) = type_convert<DstData>(src_thread_scratch_[idx]);
+        });
+#else
+        // sub-dword transpose between src_thread_scratch_ and dst_thread_scratch_
+        // TODO make this logic more generic for more sub-dword datatype
+        if constexpr(SrcVectorDim != DstVectorDim &&
+                     is_same<half_t, remove_cvref_t<SrcData>>::value &&
+                     is_same<half_t, remove_cvref_t<DstData>>::value &&
+                     SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0)
+        {
+            // each transpose does
+            // DstScalarPerVector # of src vectors in src_thread_scratch_
+            // SrcScalarPerVector # of dst vectors in dst_thread_scratch_
+            constexpr index_t num_src_vector = Number<DstScalarPerVector>{};
+            constexpr index_t num_dst_vector = Number<SrcScalarPerVector>{};
+
+            // Assume SrcVectorDim is not the same as DstVectorDim, so we do transpose
+            // TODO: make this logic generic for all scenario
+            static_assert(SrcVectorDim != DstVectorDim, "wrong");
+
+            constexpr auto src_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+            constexpr auto dst_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+            constexpr auto scalar_per_access = generate_sequence(
+                detail::lambda_scalar_per_access_for_src_and_dst<SrcVectorDim,
+                                                                 SrcScalarPerVector,
+                                                                 DstVectorDim,
+                                                                 DstScalarPerVector>{},
+                Number<nDim>{});
+
+            constexpr auto access_lengths = SliceLengths{} / scalar_per_access;
+
+            static_ford<decltype(access_lengths)>{}([&](auto access_idx) {
+                constexpr auto data_idx = access_idx * scalar_per_access;
+
+                constexpr auto data_idx_seq = generate_sequence_v2(
+                    [&](auto i) { return Number<data_idx[i]>{}; }, Number<nDim>{});
+
+                // TODO type_convert is not used yet!!!!!
+                using src_vector_t = vector_type_maker_t<SrcData, SrcScalarPerVector>;
+                using dst_vector_t = vector_type_maker_t<DstData, DstScalarPerVector>;
+
+                // get DstScalarPerVector # of read-only references to src vectors from
+                // src_thread_scratch_
+                const auto src_vector_refs = generate_tie(
+                    [&](auto i) -> const src_vector_t& {
+                        // i increment corresponds to movement in DstVectorDim
+                        return src_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * dst_scalar_step_in_vector);
+                    },
+                    Number<num_src_vector>{});
+
+                // get SrcScalarPerVector # of references to dst vectors from dst_thread_scratch_
+                auto dst_vector_refs = generate_tie(
+                    [&](auto i) -> dst_vector_t& {
+                        // i increment corresponds to movement in SrcVectorDim
+                        return dst_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * src_scalar_step_in_vector);
+                    },
+                    Number<num_dst_vector>{});
+
+                // do data transpose
+                // TODO type_convert is not used yet!!!!!
+                transpose_vectors<SrcData, DstScalarPerVector, SrcScalarPerVector>{}(
+                    src_vector_refs, dst_vector_refs);
+            });
+        }
+        else
+        {
+            static_ford<SliceLengths>{}([&](auto idx) {
+                // convert from SrcData to DstData here
+                dst_thread_scratch_(idx) = type_convert<DstData>(src_thread_scratch_[idx]);
+            });
+        }
+#endif
+    }
+
+    template <typename DstBuffer, typename Dst0Buffer, typename Dst1Buffer>
+    __device__ void RunWrite(const DstDesc& dst_desc,
+                             DstBuffer& dst_buf,
+                             const Dst0Desc& dst0_desc,
+                             const Dst0Buffer& dst0_buf,
+                             const Dst1Desc& dst1_desc,
+                             const Dst1Buffer& dst1_buf)
+    {
+        // if there is transpose, it's done here
+        // TODO move this elsewhere
+        TransferDataFromSrcThreadScratchToDstThreadScratch();
+
+        static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        // src scalar per access on each dim
+        // TODO: don't use this
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // make forward steps
+        const auto dst_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst_desc, forward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make forward steps: dst0
+        // WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
+        // DstScalarPerVector
+        // TODO: fix this
+        const auto dst0_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst0_desc, forward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make forward steps: dst1
+        // WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
+        // DstScalarPerVector
+        // TODO: fix this
+        const auto dst1_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst1_desc, forward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto dst_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst_desc, backward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make backward steps: dst0
+        // WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
+        // DstScalarPerVector
+        // TODO: fix this
+        const auto dst0_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst0_desc, backward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make backward steps: dst1
+        // WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
+        // DstScalarPerVector
+        // TODO: fix this
+        const auto dst1_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst1_desc, backward_step_idx);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_dst_access_lengths)>{}([&](auto ordered_dst_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_dst_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate dst data index
+            constexpr auto dst_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_idx[i]
+                                                      : ordered_dst_access_lengths[i] - 1 -
+                                                            ordered_dst_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                       dst_scalar_per_access;
+            }();
+
+            constexpr auto dst_data_idx_seq = generate_sequence_v2(
+                [&](auto i) { return Number<dst_data_idx[i]>{}; }, Number<dst_data_idx.Size()>{});
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            using dst_vector_type = vector_type_maker_t<DstData, DstScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            // copy data from dst_thread_scratch_ into dst_vector_container
+            auto dst_vector_container = dst_vector_type{
+                dst_thread_scratch_.template GetAsType<dst_vector_t>(dst_data_idx_seq)};
+
+            // apply DstElementwiseOperation on dst_vector_container
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                dst_vector_container.template AsType<DstData>()(i) =
+                    dst_element_op_(dst_vector_container.template AsType<DstData>()[i]);
+            });
+
+            // copy data from dst_vector_container to dst_buf
+            dst_buf.template Set<dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_dst_access_idx[i] < ordered_dst_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_dst_access_idx[j] == ordered_dst_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move dst coord
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_forward_steps[dst_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_backward_steps[dst_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        constexpr auto I0 = Number<0>{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            // TODO: BUG: should start at 1
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_src_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate src data index after last iteration in RunRead(), if it has not being reset by
+        // RunRead()
+        constexpr auto src_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                   src_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_src_data_step = [&]() {
+            Index reset_src_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_src_data_step_(i) = -src_data_idx[i]; });
+
+            return reset_src_data_step_;
+        }();
+
+        return reset_src_data_step;
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        constexpr auto I0 = Number<0>{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_dst_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate dst data index after last iteration in RunWrite(), if it has not being reset by
+        // RunWrite()
+        constexpr auto dst_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                   dst_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_dst_data_step = [&]() {
+            Index reset_dst_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
+
+            return reset_dst_data_step_;
+        }();
+
+        return reset_dst_data_step;
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Dst0Desc dst0_desc,
+                                       const Dst1Desc dst1_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by RunWrite(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+        move_tensor_coordinate(dst0_desc, dst0_coord_, adjusted_step);
+        move_tensor_coordinate(dst1_desc, dst1_coord_, adjusted_step);
+    }
+
+    __device__ static constexpr auto GetSrcThreadScratchDescriptor()
+    {
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(src_access_lengths), Number<SrcScalarPerVector>{});
+
+        // 1st stage of transforms
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(src_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(src_access_lengths_and_vector_length[i],
+                                   src_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(src_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    __device__ static constexpr auto GetDstThreadScratchDescriptor()
+    {
+        // 1st stage of transforms
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(dst_access_lengths), Number<DstScalarPerVector>{});
+
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(dst_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(dst_access_lengths_and_vector_length[i],
+                                   dst_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(dst_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    private:
+    static constexpr auto src_thread_scratch_desc_ = decltype(GetSrcThreadScratchDescriptor()){};
+    static constexpr auto dst_thread_scratch_desc_ = decltype(GetDstThreadScratchDescriptor()){};
+
+    StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                    SrcData,
+                                    SrcScalarPerVector,
+                                    decltype(src_thread_scratch_desc_),
+                                    true>
+        src_thread_scratch_;
+
+    StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                    DstData,
+                                    DstScalarPerVector,
+                                    decltype(dst_thread_scratch_desc_),
+                                    true>
+        dst_thread_scratch_;
+
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const SrcElementwiseOperation src_element_op_;
+    const DstElementwiseOperation dst_element_op_;
+};
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp
new file mode 100644
index 0000000..6e8a239
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp
@@ -0,0 +1,175 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+namespace ck {
+// Assume:
+//   1. src:
+//     1. SrcDesc is known at compile-time
+//     2. SrcBuffer is DynamicBuffer
+//     3. src_ref_idx is known at run-time
+//     4. SrcRefToOriginDisplacement is known at compile-time
+//     5. use #-step
+//   2. dst:
+//     1. DstDesc is known at compile-time
+//     2. DstBuffer is StaticBuffer
+//     3. DstOriginIdx is known at compile-time
+//     4. use direct address calculation
+//   3. vector access on src
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          typename SrcVectorTensorLengths,
+          typename SrcVectorTensorContiguousDimOrder,
+          typename enable_if<SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_v4r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v4r1(const Index& src_ref_idx)
+        : src_ref_coord_(make_tensor_coordinate(SrcDesc{}, src_ref_idx))
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_for<0, nDim, 1>{}([](auto i) {
+            static_assert(SliceLengths::At(i) % SrcVectorTensorLengths::At(i) == 0, "wrong!");
+        });
+    }
+
+    template <typename SrcRefToOriginDisplacement,
+              typename DstOriginIdx,
+              typename SrcBuffer,
+              typename DstBuffer>
+    __device__ void Run(const SrcDesc&,
+                        const SrcRefToOriginDisplacement&,
+                        const SrcBuffer& src_buf,
+                        const DstDesc&,
+                        const DstOriginIdx&,
+                        DstBuffer& dst_buf) const
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value &&
+                is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<SrcRefToOriginDisplacement>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<DstOriginIdx>>::value,
+                      "wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known "
+                      "at compile-time");
+
+        // SrcDesc and DstDesc are known at compile-time
+        constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
+        constexpr auto dst_desc = remove_cvref_t<DstDesc>{};
+
+        // SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time
+        constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{});
+        constexpr auto dst_origin_idx             = to_multi_index(DstOriginIdx{});
+
+        // tensor descriptor for src_vector
+        constexpr auto src_vector_tensor_lengths = SrcVectorTensorLengths{};
+
+        constexpr auto src_vector_tensor_strides = container_reorder_given_old2new(
+            container_reverse_exclusive_scan(
+                container_reorder_given_new2old(src_vector_tensor_lengths,
+                                                SrcVectorTensorContiguousDimOrder{}),
+                math::multiplies{},
+                I1),
+            SrcVectorTensorContiguousDimOrder{});
+
+        constexpr auto src_vector_desc =
+            make_naive_tensor_descriptor(sequence_to_tuple_of_number(src_vector_tensor_lengths),
+                                         sequence_to_tuple_of_number(src_vector_tensor_strides));
+
+        // access order and lengths
+        constexpr auto access_lengths = SliceLengths{} / src_vector_tensor_lengths;
+
+        constexpr auto dim_access_order = DimAccessOrder{};
+
+        constexpr auto ordered_access_lengths =
+            container_reorder_given_new2old(access_lengths, dim_access_order);
+
+        static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
+            // position in slice window
+            constexpr auto data_to_origin_disp_idx =
+                ordered_access_idx.ReorderGivenOld2New(dim_access_order) *
+                src_vector_tensor_lengths;
+
+            // src coordinate at starting point of src_vector
+            constexpr auto src_ref_to_data_disp_idx =
+                src_ref_to_origin_disp_idx + data_to_origin_disp_idx;
+
+            constexpr auto src_ref_to_data_disp_coord_step =
+                make_tensor_coordinate_step(src_desc, src_ref_to_data_disp_idx);
+
+            auto src_data_coord = src_ref_coord_;
+
+            move_tensor_coordinate(src_desc, src_data_coord, src_ref_to_data_disp_coord_step);
+
+            vector_type_maker_t<SrcData, src_vector_desc.GetElementSpaceSize()> src_vector;
+
+            using src_vector_t = typename decltype(src_vector)::type;
+
+            const bool is_src_valid = coordinate_has_valid_offset_assuming_visible_index_is_valid(
+                src_desc, src_data_coord);
+
+            // copy data from src_buf into src_vector
+            src_vector.template AsType<src_vector_t>()(I0) =
+                src_buf.template Get<src_vector_t>(src_data_coord.GetOffset(), is_src_valid);
+
+            // copy data from src_vector into dst_buf (also cast from SrcData to DstData)
+            static_ford<SrcVectorTensorLengths>{}([&](auto src_vector_idx_) {
+                constexpr auto src_vector_idx = to_multi_index(src_vector_idx_);
+
+                constexpr index_t src_vector_offset =
+                    src_vector_desc.CalculateOffset(src_vector_idx);
+
+                constexpr index_t dst_offset = dst_desc.CalculateOffset(
+                    dst_origin_idx + data_to_origin_disp_idx + src_vector_idx);
+
+                dst_buf(Number<dst_offset>{}) = type_convert<DstData>(
+                    src_vector.template AsType<DstData>()[Number<src_vector_offset>{}]);
+            });
+        });
+    }
+
+    template <typename SrcSliceMoveStepIdx>
+    __device__ void MoveSrcSliceWindow(const SrcDesc&,
+                                       const SrcSliceMoveStepIdx& src_slice_move_step_idx)
+    {
+        constexpr auto src_desc = SrcDesc{};
+
+        const auto src_slice_move_step_iter =
+            make_tensor_coordinate_step(src_desc, to_multi_index(src_slice_move_step_idx));
+
+        move_tensor_coordinate(SrcDesc{}, src_ref_coord_, src_slice_move_step_iter);
+    }
+
+    private:
+    SrcCoord src_ref_coord_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp
new file mode 100644
index 0000000..f13da34
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp
@@ -0,0 +1,614 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+//   4. Use thread buffer
+template <typename SliceLengths,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          typename SrcVectorTensorLengths,
+          typename DstVectorTensorLengths,
+          typename SrcVectorTensorContiguousDimOrder,
+          typename DstVectorTensorContiguousDimOrder,
+          bool SrcResetCoordinateAfterRun, // control whether to move back src coordinate after each
+                                           // RunRead(),  will be fused with MoveSrcSliceWindow to
+                                           // save addr computation
+          bool DstResetCoordinateAfterRun> // control whether to move back dst coordinate after each
+                                           // RunWrite(),  will be fused with MoveDstSliceWindow to
+                                           // save addr computation
+struct ThreadwiseTensorSliceTransfer_v5r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+    using Index                   = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+    using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v5r1(const SrcDesc& src_desc,
+                                                            const Index& src_slice_origin,
+                                                            const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin))
+    {
+        // TODO: fix this
+        static_assert(is_same<SrcData, DstData>::value,
+                      "wrong! current implementation assume SrcData and DstData are same type");
+
+        static_for<0, nDim, 1>{}([](auto i) {
+            static_assert(SliceLengths::At(i) % SrcVectorTensorLengths::At(i) == 0 &&
+                              SliceLengths::At(i) % DstVectorTensorLengths::At(i) == 0,
+                          "wrong!");
+        });
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename SrcStepHacks>
+    __device__ void
+    RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf, const SrcStepHacks& src_step_hacks)
+    {
+        static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
+            "wrong! SrcBuffer and SrcData data type are inconsistent");
+
+        // tensor descriptor for src_vector
+        constexpr auto src_vector_tensor_lengths = SrcVectorTensorLengths{};
+
+        constexpr auto src_vector_tensor_strides = container_reorder_given_old2new(
+            container_reverse_exclusive_scan(
+                container_reorder_given_new2old(src_vector_tensor_lengths,
+                                                SrcVectorTensorContiguousDimOrder{}),
+                math::multiplies{},
+                I1),
+            SrcVectorTensorContiguousDimOrder{});
+
+        constexpr auto src_vector_desc =
+            make_naive_tensor_descriptor(sequence_to_tuple_of_number(src_vector_tensor_lengths),
+                                         sequence_to_tuple_of_number(src_vector_tensor_strides));
+
+        // access order and lengths
+        constexpr auto src_access_lengths = SliceLengths{} / src_vector_tensor_lengths;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // make forward steps
+        const auto src_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? src_vector_tensor_lengths[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    src_desc, forward_step_idx, src_step_hacks[I0][i]);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto src_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -src_vector_tensor_lengths[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    src_desc, backward_step_idx, src_step_hacks[I1][i]);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_src_access_idx[I0];
+
+                    static_for<0, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate src data index
+            constexpr auto src_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
+                                                      : ordered_src_access_lengths[i] - 1 -
+                                                            ordered_src_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                       src_vector_tensor_lengths;
+            }();
+
+            vector_type_maker_t<SrcData, src_vector_desc.GetElementSpaceSize()> src_vector;
+
+            using src_vector_t = typename decltype(src_vector)::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf to src_vector
+            src_vector.template AsType<src_vector_t>()(I0) =
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid);
+
+            // copy data from src_vector to buffer_
+            static_ford<SrcVectorTensorLengths>{}([&](auto src_vector_idx_) {
+                constexpr auto src_vector_idx = to_multi_index(src_vector_idx_);
+
+                constexpr index_t src_vector_offset =
+                    src_vector_desc.CalculateOffset(src_vector_idx);
+
+                constexpr index_t buffer_offset =
+                    buffer_desc_.CalculateOffset(src_data_idx + src_vector_idx);
+
+                buffer_(Number<buffer_offset>{}) =
+                    src_vector.template AsType<SrcData>()[Number<src_vector_offset>{}];
+            });
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_src_access_idx[i] < ordered_src_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_src_access_idx[j] == ordered_src_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_forward_steps[src_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_backward_steps[src_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetSrcCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    template <typename DstBuffer, typename DstStepHacks>
+    __device__ void
+    RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf, const DstStepHacks& dst_step_hacks)
+    {
+        static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        // tensor descriptor for dst_vector
+        constexpr auto dst_vector_tensor_lengths = DstVectorTensorLengths{};
+
+        constexpr auto dst_vector_tensor_strides = container_reorder_given_old2new(
+            container_reverse_exclusive_scan(
+                container_reorder_given_new2old(dst_vector_tensor_lengths,
+                                                DstVectorTensorContiguousDimOrder{}),
+                math::multiplies{},
+                I1),
+            DstVectorTensorContiguousDimOrder{});
+
+        constexpr auto dst_vector_desc =
+            make_naive_tensor_descriptor(sequence_to_tuple_of_number(dst_vector_tensor_lengths),
+                                         sequence_to_tuple_of_number(dst_vector_tensor_strides));
+
+        // dst access order and lengths
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_vector_tensor_lengths;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // make forward steps
+        const auto dst_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_vector_tensor_lengths[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    dst_desc, forward_step_idx, dst_step_hacks[I0][i]);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto dst_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_vector_tensor_lengths[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    dst_desc, backward_step_idx, dst_step_hacks[I1][i]);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_dst_access_lengths)>{}([&](auto ordered_dst_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_dst_access_idx[I0];
+
+                    static_for<0, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate dst data index
+            constexpr auto dst_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_idx[i]
+                                                      : ordered_dst_access_lengths[i] - 1 -
+                                                            ordered_dst_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                       dst_vector_tensor_lengths;
+            }();
+
+            vector_type_maker_t<DstData, dst_vector_desc.GetElementSpaceSize()> dst_vector;
+
+            // copy data from buffer_ to dst_vector (also cast from SrcData to DstData)
+            static_ford<DstVectorTensorLengths>{}([&](auto dst_vector_idx_) {
+                constexpr auto dst_vector_idx = to_multi_index(dst_vector_idx_);
+
+                constexpr index_t buffer_offset =
+                    buffer_desc_.CalculateOffset(dst_data_idx + dst_vector_idx);
+
+                constexpr index_t dst_vector_offset =
+                    dst_vector_desc.CalculateOffset(dst_vector_idx);
+
+                dst_vector.template AsType<DstData>()(Number<dst_vector_offset>{}) =
+                    type_convert<DstData>(buffer_[Number<buffer_offset>{}]);
+            });
+
+            using dst_vector_t = typename decltype(dst_vector)::type;
+
+            // copy data from dst_vector to dst_buf
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            dst_buf.template Set<dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_dst_access_idx[i] < ordered_dst_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_dst_access_idx[j] == ordered_dst_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_forward_steps[dst_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_backward_steps[dst_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    template <typename SrcBuffer>
+    __device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf)
+    {
+        constexpr index_t ntransform_src = SrcDesc::GetNumOfTransform();
+
+        constexpr auto zeros = typename uniform_sequence_gen<ntransform_src, 0>::type{};
+
+        constexpr auto src_step_hacks =
+            make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
+                       generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
+
+        RunRead(src_desc, src_buf, src_step_hacks);
+    }
+
+    template <typename DstBuffer>
+    __device__ void RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf)
+    {
+        constexpr index_t ntransform_dst = DstDesc::GetNumOfTransform();
+
+        constexpr auto zeros = typename uniform_sequence_gen<ntransform_dst, 0>::type{};
+
+        constexpr auto dst_step_hacks =
+            make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
+                       generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
+
+        RunWrite(dst_desc, dst_buf, dst_step_hacks);
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        constexpr auto src_vector_tensor_lengths = SrcVectorTensorLengths{};
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_vector_tensor_lengths;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_src_access_lengths[I0] - 1;
+
+                static_for<0, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate src data index after last iteration in RunRead(), if it has not being reset by
+        // RunRead()
+        constexpr auto src_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                   src_vector_tensor_lengths;
+        }();
+
+        //
+        constexpr auto reset_src_data_step = [&]() {
+            Index reset_src_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_src_data_step_(i) = -src_data_idx[i]; });
+
+            return reset_src_data_step_;
+        }();
+
+        return reset_src_data_step;
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        constexpr auto dst_vector_tensor_lengths = DstVectorTensorLengths{};
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_vector_tensor_lengths;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_dst_access_lengths[I0] - 1;
+
+                static_for<0, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate dst data index after last iteration in RunWrite(), if it has not being reset by
+        // RunWrite()
+        constexpr auto dst_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                   dst_vector_tensor_lengths;
+        }();
+
+        //
+        constexpr auto reset_dst_data_step = [&]() {
+            Index reset_dst_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
+
+            return reset_dst_data_step_;
+        }();
+
+        return reset_dst_data_step;
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <typename SrcMoveSliceWindowStepHack>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDesc& src_desc,
+                       const Index& src_slice_origin_step_idx,
+                       const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(
+            src_desc, adjusted_step_idx, src_move_slice_window_step_hack);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by RunWrite(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    static constexpr auto buffer_desc_ =
+        make_naive_tensor_descriptor_packed(sequence_to_tuple_of_number(SliceLengths{}));
+
+    static constexpr auto buffer_size_ = buffer_desc_.GetElementSpaceSize();
+
+    StaticBuffer<AddressSpaceEnum::Vgpr, SrcData, buffer_size_, true> buffer_;
+
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp
new file mode 100644
index 0000000..9c91cd9
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp
@@ -0,0 +1,213 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          InMemoryDataOperationEnum DstInMemOp,
+          bool SrcResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r1
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r1(const SrcDesc& src_desc,
+                                                            const Index& src_slice_origin,
+                                                            const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin,
+                                                            const ElementwiseOperation& element_op)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename DstBuffer>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        // loop over space-filling curve
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src_vector_type = vector_type_maker_t<SrcData, ScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf into src_vector_container
+            auto src_vector_container = src_vector_type{
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                SrcData v;
+
+                // apply element-wise operation
+                element_op_(v, src_vector_container.template AsType<SrcData>()[i]);
+
+                // apply type convert
+                dst_vector_container.template AsType<DstData>()(i) = type_convert<DstData>(v);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(
+                    src_desc, src_coord_, make_tensor_coordinate_step(src_desc, forward_step));
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = SrcResetCoordinateAfterRun
+                                           ? src_slice_origin_step_idx
+                                           : src_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp
new file mode 100644
index 0000000..68bc272
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp
@@ -0,0 +1,260 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src0_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename Src0Data,
+          typename Src1Data,
+          typename DstData,
+          typename Src0Desc,
+          typename Src1Desc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          InMemoryDataOperationEnum DstInMemOp,
+          bool Src0ResetCoordinateAfterRun,
+          bool Src1ResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r2
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using Src0Coord = decltype(make_tensor_coordinate(Src0Desc{}, Index{}));
+    using Src1Coord = decltype(make_tensor_coordinate(Src1Desc{}, Index{}));
+    using DstCoord  = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r2(const Src0Desc& src0_desc,
+                                                            const Index& src0_slice_origin,
+                                                            const Src1Desc& src1_desc,
+                                                            const Index& src1_slice_origin,
+                                                            const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin,
+                                                            const ElementwiseOperation& element_op)
+        : src0_coord_(make_tensor_coordinate(src0_desc, src0_slice_origin)),
+          src1_coord_(make_tensor_coordinate(src1_desc, src1_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrc0SliceOrigin(const Src0Desc& src0_desc,
+                                       const Index& src0_slice_origin_idx)
+    {
+        src0_coord_ = make_tensor_coordinate(src0_desc, src0_slice_origin_idx);
+    }
+
+    __device__ void SetSrc1SliceOrigin(const Src1Desc& src1_desc,
+                                       const Index& src1_slice_origin_idx)
+    {
+        src1_coord_ = make_tensor_coordinate(src1_desc, src1_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename Src0Buffer, typename Src1Buffer, typename DstBuffer>
+    __device__ void Run(const Src0Desc& src0_desc,
+                        const Src0Buffer& src0_buf,
+                        const Src1Desc& src1_desc,
+                        const Src1Buffer& src1_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src0_vector_type = vector_type_maker_t<Src0Data, ScalarPerVector>;
+            using src0_vector_t    = typename src0_vector_type::type;
+
+            using src1_vector_type = vector_type_maker_t<Src1Data, ScalarPerVector>;
+            using src1_vector_t    = typename src1_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src0_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src0_desc, src0_coord_);
+
+            const bool is_src1_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src1_desc, src1_coord_);
+
+            // copy data from src0_buf into src0_vector_container
+            auto src0_vector_container = src0_vector_type{
+                src0_buf.template Get<src0_vector_t>(src0_coord_.GetOffset(), is_src0_valid)};
+
+            auto src1_vector_container = src1_vector_type{
+                src1_buf.template Get<src1_vector_t>(src1_coord_.GetOffset(), is_src1_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                element_op_(dst_vector_container.template AsType<DstData>()(i),
+                            src0_vector_container.template AsType<Src0Data>()[i],
+                            src1_vector_container.template AsType<Src1Data>()[i]);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(
+                    src0_desc, src0_coord_, make_tensor_coordinate_step(src0_desc, forward_step));
+                move_tensor_coordinate(
+                    src1_desc, src1_coord_, make_tensor_coordinate_step(src1_desc, forward_step));
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(Src0ResetCoordinateAfterRun)
+        {
+            const auto src0_reset_step =
+                make_tensor_coordinate_step(src0_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src0_desc, src0_coord_, src0_reset_step);
+        }
+
+        if constexpr(Src1ResetCoordinateAfterRun)
+        {
+            const auto src1_reset_step =
+                make_tensor_coordinate_step(src1_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src1_desc, src1_coord_, src1_reset_step);
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc0SliceWindow(const Src0Desc& src0_desc,
+                                        const Index& src0_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src0ResetCoordinateAfterRun
+                                           ? src0_slice_origin_step_idx
+                                           : src0_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src0_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src0_desc, src0_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc1SliceWindow(const Src1Desc& src1_desc,
+                                        const Index& src1_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src1ResetCoordinateAfterRun
+                                           ? src1_slice_origin_step_idx
+                                           : src1_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src1_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src1_desc, src1_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    Src0Coord src0_coord_;
+    Src1Coord src1_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp
new file mode 100644
index 0000000..0f5fb88
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp
@@ -0,0 +1,310 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src0_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename Src0Data,
+          typename Src1Data,
+          typename Src2Data,
+          typename DstData,
+          typename Src0Desc,
+          typename Src1Desc,
+          typename Src2Desc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          InMemoryDataOperationEnum DstInMemOp,
+          bool Src0ResetCoordinateAfterRun,
+          bool Src1ResetCoordinateAfterRun,
+          bool Src2ResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r3
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using Src0Coord = decltype(make_tensor_coordinate(Src0Desc{}, Index{}));
+    using Src1Coord = decltype(make_tensor_coordinate(Src1Desc{}, Index{}));
+    using Src2Coord = decltype(make_tensor_coordinate(Src2Desc{}, Index{}));
+    using DstCoord  = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r3(const Src0Desc& src0_desc,
+                                                            const Index& src0_slice_origin,
+                                                            const Src1Desc& src1_desc,
+                                                            const Index& src1_slice_origin,
+                                                            const Src2Desc& src2_desc,
+                                                            const Index& src2_slice_origin,
+                                                            const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin,
+                                                            const ElementwiseOperation& element_op)
+        : src0_coord_(make_tensor_coordinate(src0_desc, src0_slice_origin)),
+          src1_coord_(make_tensor_coordinate(src1_desc, src1_slice_origin)),
+          src2_coord_(make_tensor_coordinate(src2_desc, src2_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrc0SliceOrigin(const Src0Desc& src0_desc,
+                                       const Index& src0_slice_origin_idx)
+    {
+        src0_coord_ = make_tensor_coordinate(src0_desc, src0_slice_origin_idx);
+    }
+
+    __device__ void SetSrc1SliceOrigin(const Src1Desc& src1_desc,
+                                       const Index& src1_slice_origin_idx)
+    {
+        src1_coord_ = make_tensor_coordinate(src1_desc, src1_slice_origin_idx);
+    }
+
+    __device__ void SetSrc2SliceOrigin(const Src2Desc& src2_desc,
+                                       const Index& src2_slice_origin_idx)
+    {
+        src2_coord_ = make_tensor_coordinate(src2_desc, src2_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename Src0Buffer, typename Src1Buffer, typename Src2Buffer, typename DstBuffer>
+    __device__ void Run(const Src0Desc& src0_desc,
+                        const Src0Buffer& src0_buf,
+                        const Src1Desc& src1_desc,
+                        const Src1Buffer& src1_buf,
+                        const Src2Desc& src2_desc,
+                        const Src2Buffer& src2_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src0_vector_type = vector_type_maker_t<Src0Data, ScalarPerVector>;
+            using src0_vector_t    = typename src0_vector_type::type;
+
+            using src1_vector_type = vector_type_maker_t<Src1Data, ScalarPerVector>;
+            using src1_vector_t    = typename src1_vector_type::type;
+
+            using src2_vector_type = vector_type_maker_t<Src2Data, ScalarPerVector>;
+            using src2_vector_t    = typename src2_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src0_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src0_desc, src0_coord_);
+
+            const bool is_src1_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src1_desc, src1_coord_);
+
+            const bool is_src2_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src2_desc, src2_coord_);
+
+            // copy data from src0_buf into src0_vector_container
+            auto src0_vector_container = src0_vector_type{
+                src0_buf.template Get<src0_vector_t>(src0_coord_.GetOffset(), is_src0_valid)};
+
+            auto src1_vector_container = src1_vector_type{
+                src1_buf.template Get<src1_vector_t>(src1_coord_.GetOffset(), is_src1_valid)};
+
+            auto src2_vector_container = src2_vector_type{
+                src2_buf.template Get<src2_vector_t>(src2_coord_.GetOffset(), is_src2_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                element_op_(dst_vector_container.template AsType<DstData>()(i),
+                            src0_vector_container.template AsType<Src0Data>()[i],
+                            src1_vector_container.template AsType<Src1Data>()[i],
+                            src2_vector_container.template AsType<Src2Data>()[i]);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(
+                    src0_desc, src0_coord_, make_tensor_coordinate_step(src0_desc, forward_step));
+                move_tensor_coordinate(
+                    src1_desc, src1_coord_, make_tensor_coordinate_step(src1_desc, forward_step));
+                move_tensor_coordinate(
+                    src2_desc, src2_coord_, make_tensor_coordinate_step(src2_desc, forward_step));
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(Src0ResetCoordinateAfterRun)
+        {
+            const auto src0_reset_step =
+                make_tensor_coordinate_step(src0_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src0_desc, src0_coord_, src0_reset_step);
+        }
+
+        if constexpr(Src1ResetCoordinateAfterRun)
+        {
+            const auto src1_reset_step =
+                make_tensor_coordinate_step(src1_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src1_desc, src1_coord_, src1_reset_step);
+        }
+
+        if constexpr(Src2ResetCoordinateAfterRun)
+        {
+            const auto src2_reset_step =
+                make_tensor_coordinate_step(src2_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src2_desc, src2_coord_, src2_reset_step);
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc0SliceWindow(const Src0Desc& src0_desc,
+                                        const Index& src0_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src0ResetCoordinateAfterRun
+                                           ? src0_slice_origin_step_idx
+                                           : src0_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src0_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src0_desc, src0_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc1SliceWindow(const Src1Desc& src1_desc,
+                                        const Index& src1_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src1ResetCoordinateAfterRun
+                                           ? src1_slice_origin_step_idx
+                                           : src1_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src1_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src1_desc, src1_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc2SliceWindow(const Src2Desc& src2_desc,
+                                        const Index& src2_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src2ResetCoordinateAfterRun
+                                           ? src2_slice_origin_step_idx
+                                           : src2_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src2_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src2_desc, src2_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    Src0Coord src0_coord_;
+    Src1Coord src1_coord_;
+    Src2Coord src2_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp
new file mode 100644
index 0000000..2eb1b0e
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp
@@ -0,0 +1,298 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Thread-level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//   6. Does not need to know src_descs and dst_descs at compile-time
+//   7. Does not need to know src_slice_origins and dst_slice_origins at compile-time,
+//
+// Does following things to avoid scratch memory issue
+//   1. Use StaticallyIndexedArray or vector_type instead of C array for thread buffer
+//   2. Pass tensor descritpors by reference (or tuple of references)
+//   3. Does not keep reference to tensor descriptor
+//   4. Does not construct new tensor coordinate when call Run()
+template <typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          typename SrcResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          typename DstResetCoordinateAfterRunFlags> // Sequence<bool ...>
+struct ThreadwiseTensorSliceTransfer_v7
+{
+    static constexpr auto I0 = Number<0>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    static constexpr index_t nSrc = SrcDescs::Size();
+    static constexpr index_t nDst = DstDescs::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    // return a tuple of coordiantes for a tuple of tensor
+    template <typename Descs,
+              typename Indices,
+              enable_if_t<Descs::Size() == Indices::Size(), bool> = false>
+    static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices)
+    {
+        return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); },
+                              Number<Descs::Size()>{});
+    }
+
+    using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray<Index, nSrc>{}));
+    using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray<Index, nDst>{}));
+
+    // scalar per access on each dim
+    // FIXME: don't use lambda_scalar_per_access
+    static constexpr auto scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+    using SpaceFillingCurve =
+        SpaceFillingCurve<SliceLengths, DimAccessOrder, remove_cv_t<decltype(scalar_per_access)>>;
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v7(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_slice_origins,
+        const ElementwiseOperation& element_op)
+        : src_coords_(MakeCoordinates(src_descs, src_slice_origins)),
+          dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    template <typename Indices, enable_if_t<SrcDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetSrcSliceOrigins(const SrcDescs& src_descs,
+                                       const Indices& src_slice_origin_idxs)
+    {
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            src_coords_(i) = make_tensor_coordinate(src_descs[i], src_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename Indices, enable_if_t<DstDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetDstSliceOrigins(const DstDescs& dst_descs,
+                                       const Indices& dst_slice_origin_idxs)
+    {
+        static_for<0, nDst, 1>{}([&](auto i) {
+            dst_coords_(i) = make_tensor_coordinate(dst_descs[i], dst_slice_origin_idxs[i]);
+        });
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename SrcBuffers,
+              typename DstBuffers,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size() &&
+                              DstDescs::Size() == DstBuffers::Size(),
+                          bool> = false>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs)
+    {
+        auto generate_vectors = [&](auto data_types) {
+            constexpr index_t num = data_types.Size();
+
+            return generate_tuple(
+                [&](auto i) {
+                    using DataType = remove_cvref_t<decltype(data_types[i])>;
+
+                    return vector_type_maker_t<DataType, ScalarPerVector>{};
+                },
+                Number<num>{});
+        };
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto iAccess) {
+            auto src_vectors = generate_vectors(SrcDatas{});
+            auto dst_vectors = generate_vectors(DstDatas{});
+
+            // copy data from src_bufs into src_vectors
+            static_for<0, nSrc, 1>{}([&](auto i) {
+                using src_vector_t = typename remove_cvref_t<decltype(src_vectors[i])>::type;
+
+                const bool is_src_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(src_descs[i],
+                                                                                src_coords_[i]);
+
+                src_vectors(i).template AsType<src_vector_t>()(I0) =
+                    src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(),
+                                                           is_src_valid);
+            });
+
+            // apply pointwise function
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                // get reference to src data
+                const auto src_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iSrc) -> const auto& {
+                        using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
+
+                        return src_vectors[iSrc].template AsType<SrcData>()[i];
+                    },
+                    Number<nSrc>{});
+
+                // get reference to dst data
+                auto dst_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iDst) -> auto& {
+                        using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
+
+                        return dst_vectors(iDst).template AsType<DstData>()(i);
+                    },
+                    Number<nDst>{});
+
+                // apply pointwise function
+                // pointwise function signature:
+                // element_op_(dst_data_refs[I0],
+                //             dst_data_refs[I1],
+                //             ...,
+                //             src_data_refs[I0],
+                //             src_data_refs[I1],
+                //             ...)
+                unpack2(element_op_, dst_data_refs, src_data_refs);
+            });
+
+            // copy data from buf_vectors into dst_bufs
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using dst_vector_t = typename remove_cvref_t<decltype(dst_vectors[i])>::type;
+
+                const bool is_dst_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_descs[i],
+                                                                                dst_coords_[i]);
+
+                constexpr InMemoryDataOperationEnum DstInMemOp =
+                    static_cast<InMemoryDataOperationEnum>(DstInMemOps::At(i.value));
+
+                dst_bufs(i).template Update<DstInMemOp, dst_vector_t>(
+                    dst_coords_[i].GetOffset(),
+                    is_dst_valid,
+                    dst_vectors[i].template AsType<dst_vector_t>()[I0]);
+            });
+
+            // move coordinate
+            if constexpr(iAccess.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nSrc, 1>{}([&](auto i) {
+                    move_tensor_coordinate(src_descs[i],
+                                           src_coords_(i),
+                                           make_tensor_coordinate_step(src_descs[i], forward_step));
+                });
+
+                static_for<0, nDst, 1>{}([&](auto i) {
+                    move_tensor_coordinate(dst_descs[i],
+                                           dst_coords_(i),
+                                           make_tensor_coordinate_step(dst_descs[i], forward_step));
+                });
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            if constexpr(SrcResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto src_reset_step =
+                    make_tensor_coordinate_step(src_descs[i], GetCoordinateResetStep());
+
+                move_tensor_coordinate(src_descs[i], src_coords_(i), src_reset_step);
+            }
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            if constexpr(DstResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto dst_reset_step =
+                    make_tensor_coordinate_step(dst_descs[i], GetCoordinateResetStep());
+
+                move_tensor_coordinate(dst_descs[i], dst_coords_(i), dst_reset_step);
+            }
+        });
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t ISrc>
+    __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
+                                       Number<ISrc> iSrc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = SrcResetCoordinateAfterRunFlags::At(iSrc)
+                                           ? src_slice_origin_step_idx
+                                           : src_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_descs[iSrc], adjusted_step_idx);
+
+        move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t IDst>
+    __device__ void MoveDstSliceWindow(const DstDescs& dst_descs,
+                                       Number<IDst> iDst,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRunFlags::At(iDst)
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_descs[iDst], adjusted_step_idx);
+
+        move_tensor_coordinate(dst_descs[iDst], dst_coords_(iDst), adjusted_step);
+    }
+
+    private:
+    SrcCoords src_coords_;
+    DstCoords dst_coords_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp
new file mode 100644
index 0000000..12ba2c5
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp
@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/math_v2.hpp"
+
+namespace ck {
+
+// Assume
+//  1) XDesc is known at compile-time
+//  2) MeanVarDesc is known at compile-time
+//  3) XBuffer is static buffer
+//  4) MeanBuffer is static buffer
+//  5) VarBuffer is static buffer
+template <typename T, typename XThreadDesc_M_K, typename MeanVarThreadDesc_M>
+struct ThreadwiseWelford
+{
+    static constexpr auto x_thread_desc_m_k      = XThreadDesc_M_K{};
+    static constexpr auto mean_var_thread_desc_m = MeanVarThreadDesc_M{};
+
+    static constexpr auto thread_x_length_m        = x_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto thread_x_length_k        = x_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto thread_mean_var_length_m = mean_var_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(thread_x_length_m == thread_mean_var_length_m,
+                  "lengths of source and mean/var buffer must match!");
+
+    __device__ constexpr ThreadwiseWelford() : cur_count_(0), max_count_(0) {}
+
+    __device__ inline void Update(T& mean, T& var, T x)
+    {
+        using ck::math::isnan;
+
+        if(isnan(x))
+        {
+            mean = x;
+            var  = x;
+        }
+        else
+        {
+            T delta = x - mean;
+            mean += delta / cur_count_;
+            T delta2 = x - mean;
+            var += delta * delta2;
+        }
+    }
+
+    template <typename XBufferType, typename MeanBufferType, typename VarBufferType>
+    __device__ void
+    Run(const XBufferType& x_buf_m_k, MeanBufferType& mean_buf_m, VarBufferType& var_buf_m)
+    {
+        // FIXME - Better naming for var_buf_m
+
+        static_for<0, thread_x_length_k, 1>{}([&](auto iK) {
+            if(cur_count_ < max_count_)
+            {
+                ++cur_count_;
+
+                static_for<0, thread_x_length_m, 1>{}([&](auto iM) {
+                    constexpr index_t out_offset =
+                        mean_var_thread_desc_m.CalculateOffset(make_tuple(iM));
+
+                    constexpr auto in_offset =
+                        x_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                    Update(mean_buf_m(Number<out_offset>{}),
+                           var_buf_m(Number<out_offset>{}),
+                           x_buf_m_k[Number<in_offset>{}]);
+                });
+            }
+        });
+    };
+
+    int cur_count_;
+    int max_count_;
+};
+
+template <typename T,
+          typename SrcMeanVarCountThreadDesc_M_K,
+          typename DstMeanVarThreadDesc_M,
+          bool GetActualVariance = false>
+struct ThreadwiseWelfordMerge
+{
+    static constexpr auto src_thread_desc_m_k = SrcMeanVarCountThreadDesc_M_K{};
+    static constexpr auto dst_thread_desc_m   = DstMeanVarThreadDesc_M{};
+
+    static constexpr auto src_length_m = src_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto src_length_k = src_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto dst_length_m = dst_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(src_length_m == dst_length_m, "lengths of source and dst buffer must match!");
+
+    __device__ static void
+    Merge(T& mean_a, T& var_a, int32_t& count_a, T mean_b, T var_b, int32_t count_b)
+    {
+        int count            = count_a + count_b;
+        T count_b_over_count = count == 0 ? type_convert<T>(0) : type_convert<T>(count_b) / count;
+        T delta              = mean_b - mean_a;
+        mean_a += delta * count_b_over_count;
+        var_a += var_b + delta * delta * count_a * count_b_over_count;
+        count_a = count;
+    }
+
+    template <typename SrcMeanBufferType,
+              typename SrcVarBufferType,
+              typename SrcCountBufferType,
+              typename DstMeanBufferType,
+              typename DstVarBufferType,
+              typename DstCountBufferType>
+    __device__ static void Run(const SrcMeanBufferType& src_mean_buf,
+                               const SrcVarBufferType& src_var_buf,
+                               const SrcCountBufferType& src_count_buf,
+                               DstMeanBufferType& dst_mean_buf,
+                               DstVarBufferType& dst_var_buf,
+                               DstCountBufferType& dst_count_buf)
+    {
+        static_for<0, src_length_m, 1>{}([&](auto iM) {
+            static_for<0, src_length_k, 1>{}([&](auto iK) {
+                constexpr auto src_offset = src_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                Merge(dst_mean_buf(iM),
+                      dst_var_buf(iM),
+                      dst_count_buf(iM),
+                      src_mean_buf[Number<src_offset>{}],
+                      src_var_buf[Number<src_offset>{}],
+                      src_count_buf[Number<src_offset>{}]);
+            });
+
+            if constexpr(GetActualVariance)
+            {
+                dst_var_buf(iM) = dst_var_buf[iM] / dst_count_buf[iM];
+            };
+        });
+    };
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
new file mode 100644
index 0000000..4d53f0d
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
@@ -0,0 +1,851 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/utility/amd_xdlops.hpp"
+
+namespace ck {
+
+enum struct MfmaInstr
+{
+    mfma_f32_32x32x1xf32 = 0,
+    mfma_f32_16x16x1xf32,
+    mfma_f32_4x4x1xf32,
+    mfma_f32_32x32x2xf32,
+    mfma_f32_16x16x4xf32,
+    mfma_f32_32x32x4f16,
+    mfma_f32_16x16x4f16,
+    mfma_f32_4x4x4f16,
+    mfma_f32_32x32x8f16,
+    mfma_f32_16x16x16f16,
+    mfma_f32_32x32x8bf16_1k,
+    mfma_f32_16x16x16bf16_1k,
+    mfma_f32_32x32x4bf16,
+    mfma_f32_16x16x8bf16,
+    mfma_i32_32x32x8i8,
+    mfma_i32_16x16x16i8,
+    mfma_f64_16x16x4f64
+};
+
+template <MfmaInstr instr>
+struct mfma_type;
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x1xf32>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 2;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 1;
+    static constexpr bool is_k_reduction         = false;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x1f32<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x2xf32>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 1;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x2f32<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x4xf32>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 1;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x4f32<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x1xf32>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 4;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 1;
+    static constexpr bool is_k_reduction         = false;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x1f32<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+// treat 4x4x1 as a single-blk 4x64 mfma
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_4x4x1xf32>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 64;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 1;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 4;
+    static constexpr index_t n_per_blk           = 64;
+    static constexpr index_t k_per_blk           = 1;
+    static constexpr bool is_k_reduction         = false;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_4x4x1f32<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x4f16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 2;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = false;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x4f16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x8f16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x8f16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x16f16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x16f16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x4f16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 4;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = false;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x4f16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_4x4x4f16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 64;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 1;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 4;
+    static constexpr index_t n_per_blk           = 64;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = false;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_4x4x4f16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x8bf16_1k>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x8bf16_1k<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x16bf16_1k>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x16bf16_1k<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x4bf16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 2;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x4bf16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x8bf16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 2;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x8bf16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_i32_32x32x8i8>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_i32_32x32x8i8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_i32_16x16x16i8>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_i32_16x16x16i8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f64_16x16x4f64>
+{
+    static constexpr index_t group_size          = 1;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 4; // group_size * num_groups_per_blk;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4; // wave_size / num_threads_per_blk;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 1;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f64_16x16x4f64<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <typename base_type, index_t MPerXdlops, index_t NPerXdlops>
+struct MfmaSelector
+{
+    template <typename base_type_, index_t MPerXdlops_, index_t NPerXdlops_>
+    static constexpr auto GetMfma();
+
+    template <>
+    static constexpr auto GetMfma<double, 16, 16>()
+    {
+        return MfmaInstr::mfma_f64_16x16x4f64;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 64, 64>()
+    {
+        return MfmaInstr::mfma_f32_32x32x1xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 32, 64>()
+    {
+        return MfmaInstr::mfma_f32_32x32x1xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 16, 64>()
+    {
+        return MfmaInstr::mfma_f32_16x16x1xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 8, 64>()
+    {
+        return MfmaInstr::mfma_f32_4x4x1xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 4, 64>()
+    {
+        return MfmaInstr::mfma_f32_4x4x1xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 32, 32>()
+    {
+        return MfmaInstr::mfma_f32_32x32x2xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 16, 16>()
+    {
+        return MfmaInstr::mfma_f32_16x16x4xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 64, 64>()
+    {
+        return MfmaInstr::mfma_f32_32x32x4f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 32, 64>()
+    {
+        return MfmaInstr::mfma_f32_32x32x4f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 32, 32>()
+    {
+        return MfmaInstr::mfma_f32_32x32x8f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 16, 16>()
+    {
+        return MfmaInstr::mfma_f32_16x16x16f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 16, 64>()
+    {
+        return MfmaInstr::mfma_f32_16x16x4f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 8, 64>()
+    {
+        return MfmaInstr::mfma_f32_4x4x4f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 4, 64>()
+    {
+        return MfmaInstr::mfma_f32_4x4x4f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<bhalf_t, 32, 32>()
+    {
+#if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
+        return MfmaInstr::mfma_f32_32x32x8bf16_1k;
+#else
+        return MfmaInstr::mfma_f32_32x32x4bf16;
+#endif
+    }
+
+    template <>
+    static constexpr auto GetMfma<bhalf_t, 16, 16>()
+    {
+#if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
+        return MfmaInstr::mfma_f32_16x16x16bf16_1k;
+#else
+        return MfmaInstr::mfma_f32_16x16x8bf16;
+#endif
+    }
+
+    template <>
+    static constexpr auto GetMfma<int8_t, 32, 32>()
+    {
+        return MfmaInstr::mfma_i32_32x32x8i8;
+    }
+
+    template <>
+    static constexpr auto GetMfma<int8_t, 16, 16>()
+    {
+        return MfmaInstr::mfma_i32_16x16x16i8;
+    }
+
+    static constexpr auto selected_mfma = mfma_type<GetMfma<base_type, MPerXdlops, NPerXdlops>()>{};
+
+    __host__ __device__ constexpr MfmaSelector()
+    {
+        static_assert(selected_mfma.group_size * selected_mfma.num_groups_per_blk ==
+                          selected_mfma.num_regs_per_blk,
+                      "wrong! num_regs_per_blk");
+
+        static_assert(selected_mfma.num_threads_per_blk == selected_mfma.n_per_blk,
+                      "n_per_blk != num_threads_per_blk");
+
+        static_assert(selected_mfma.num_regs_per_blk * selected_mfma.num_input_blks ==
+                          selected_mfma.m_per_blk,
+                      "m_per_blk != num_input_blks * num_regs_per_blk");
+
+        static_assert(selected_mfma.num_output_blks == selected_mfma.num_input_blks ||
+                          selected_mfma.num_output_blks == 1,
+                      "incorrect num_output_blks");
+
+        static_assert(selected_mfma.num_regs_per_blk * selected_mfma.wave_size ==
+                          selected_mfma.m_per_blk * selected_mfma.n_per_blk,
+                      "num_regs_per_blk incorrect");
+
+        static_assert(selected_mfma.is_k_reduction ||
+                          (selected_mfma.num_input_blks == selected_mfma.num_output_blks),
+                      "is_k_reduction wrong!");
+    }
+
+    static constexpr bool IsABroadcast()
+    {
+        static_assert(NPerXdlops >= MPerXdlops, "only support ABroadcast");
+        return true;
+    }
+
+    static constexpr index_t GetKPerXdlops()
+    {
+        return (selected_mfma.is_k_reduction ? selected_mfma.num_input_blks : 1) *
+               selected_mfma.k_per_blk;
+    }
+
+    static constexpr index_t GetK1PerXdlops() { return selected_mfma.k_per_blk; }
+};
+
+template <typename base_type,
+          index_t MPerXdlops,
+          index_t NPerXdlops,
+          index_t KPack,
+          bool TransposeC = false>
+struct XdlopsGemm
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    using CIndex   = MultiIndex<2>;
+    using CIndex4D = MultiIndex<4>;
+
+    __device__ static constexpr index_t GetNumBlks() { return mfma_instr.num_output_blks; }
+
+    __device__ static constexpr index_t GetNumXdlops()
+    {
+        return MPerXdlops * NPerXdlops /
+               (mfma_instr.m_per_blk * mfma_instr.n_per_blk * mfma_instr.num_output_blks);
+    }
+
+    __host__ __device__ constexpr XdlopsGemm()
+    {
+        static_assert(NPerXdlops == 4 || NPerXdlops == 8 || NPerXdlops == 16 || NPerXdlops == 32 ||
+                          NPerXdlops == 64,
+                      "Only support GemmNPerXdlops == 4, 8, 16, 32 or 64 for xdlops");
+
+        static_assert(MPerXdlops == 4 || MPerXdlops == 8 || MPerXdlops == 16 || MPerXdlops == 32 ||
+                          MPerXdlops == 64,
+                      "Only support GemmMPerXdlops == 4, 8, 16, 32 or 64 for xdlops");
+
+        static_assert(KPack % mfma_instr.k_per_blk == 0, "KPack cannot be divided by k_per_blk");
+    }
+
+    // XDL output supporting C = A * B
+    // M2_N2 -> M2_M3_M4_N2
+    template <typename CDesc_M0_N0_M1_N1_M2_N2>
+    __host__ __device__ static constexpr auto
+    MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CDesc_M0_N0_M1_N1_M2_N2& c_desc_m0_n0_m1_n1_m2_n2)
+    {
+        const auto M0 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I0);
+        const auto N0 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I1);
+        const auto M1 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I2);
+        const auto N1 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I3);
+
+        return transform_tensor_descriptor(
+            c_desc_m0_n0_m1_n1_m2_n2,
+            make_tuple(make_pass_through_transform(M0),
+                       make_pass_through_transform(N0),
+                       make_pass_through_transform(M1),
+                       make_pass_through_transform(N1),
+                       make_unmerge_transform(make_tuple(Number<mfma_instr.num_groups_per_blk>{},
+                                                         Number<mfma_instr.num_input_blks>{},
+                                                         Number<mfma_instr.group_size>{})),
+                       make_pass_through_transform(Number<mfma_instr.num_threads_per_blk>{})),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5>{}),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4, 5, 6>{},
+                       Sequence<7>{}));
+    }
+
+    // transposed XDL output supporting C' = B' * A'
+    // M2_N2 -> M2_N2_N3_N4
+    template <typename CDesc_M0_N0_M1_N1_M2_N2>
+    __host__ __device__ static constexpr auto
+    MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(const CDesc_M0_N0_M1_N1_M2_N2& c_desc_m0_n0_m1_n1_m2_n2)
+    {
+        const auto M0 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I0);
+        const auto N0 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I1);
+        const auto M1 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I2);
+        const auto N1 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I3);
+
+        return transform_tensor_descriptor(
+            c_desc_m0_n0_m1_n1_m2_n2,
+            make_tuple(make_pass_through_transform(M0),
+                       make_pass_through_transform(N0),
+                       make_pass_through_transform(M1),
+                       make_pass_through_transform(N1),
+                       make_pass_through_transform(Number<mfma_instr.num_threads_per_blk>{}),
+                       make_unmerge_transform(make_tuple(Number<mfma_instr.num_groups_per_blk>{},
+                                                         Number<mfma_instr.num_input_blks>{},
+                                                         Number<mfma_instr.group_size>{}))),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5>{}),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5, 6, 7>{}));
+    }
+
+    template <typename CDesc_G_M0_N0_M1_N1_M2_N2>
+    __host__ __device__ static constexpr auto MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+        const CDesc_G_M0_N0_M1_N1_M2_N2& c_desc_g_m0_n0_m1_n1_m2_n2)
+    {
+        const auto G  = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I0);
+        const auto M0 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I1);
+        const auto N0 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I2);
+        const auto M1 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I3);
+        const auto N1 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I4);
+
+        return transform_tensor_descriptor(
+            c_desc_g_m0_n0_m1_n1_m2_n2,
+            make_tuple(make_pass_through_transform(G),
+                       make_pass_through_transform(M0),
+                       make_pass_through_transform(N0),
+                       make_pass_through_transform(M1),
+                       make_pass_through_transform(N1),
+                       make_unmerge_transform(make_tuple(mfma_instr.num_groups_per_blk,
+                                                         mfma_instr.num_input_blks,
+                                                         mfma_instr.group_size)),
+                       make_pass_through_transform(mfma_instr.num_threads_per_blk)),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5>{},
+                       Sequence<6>{}),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5, 6, 7>{},
+                       Sequence<8>{}));
+    }
+
+    __device__ static constexpr index_t GetRegSizePerXdlops()
+    {
+        return MPerXdlops * NPerXdlops / mfma_instr.wave_size;
+    }
+
+    __device__ static constexpr index_t GetWaveSize() { return mfma_instr.wave_size; }
+
+    template <class FloatA, class FloatB, class FloatC>
+    __device__ void Run(const FloatA& p_a_wave, const FloatB& p_b_wave, FloatC& p_c_thread) const
+    {
+        static_assert(is_same<base_type, double>::value || is_same<base_type, float>::value ||
+                          is_same<base_type, half_t>::value || is_same<base_type, bhalf_t>::value ||
+                          is_same<base_type, int8_t>::value,
+                      "base base_type must be double, float, half, bfloat16, and int8_t!");
+
+        static_for<0, KPack / mfma_instr.k_per_blk, 1>{}([&](auto k) {
+            if constexpr(!TransposeC)
+            {
+                mfma_instr.template run<MPerXdlops, NPerXdlops>(
+                    p_a_wave[k], p_b_wave[k], p_c_thread);
+            }
+            else
+            {
+                mfma_instr.template run<MPerXdlops, NPerXdlops>(
+                    p_b_wave[k], p_a_wave[k], p_c_thread);
+            }
+        });
+    }
+
+    __device__ static auto GetLaneId() { return get_thread_local_1d_id() % mfma_instr.wave_size; }
+
+    __device__ static auto GetBlkIdx()
+    {
+        const auto laneId = GetLaneId();
+
+        constexpr auto threadidx_to_blk_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(
+                make_tuple(1, mfma_instr.num_input_blks, mfma_instr.num_threads_per_blk))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto blk_idx =
+            threadidx_to_blk_idx_adaptor.CalculateBottomIndex(make_multi_index(laneId));
+
+        const auto blk_id = blk_idx[I1];
+        const auto blk_td = blk_idx[I2];
+
+        return make_tuple(blk_id, blk_td);
+    }
+
+    __host__ __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto laneId  = GetLaneId();
+        const auto blk_idx = GetBlkIdx();
+
+        const auto blk_id = blk_idx[I0];
+        const auto blk_td = blk_idx[I1];
+
+        if constexpr(mfma_instr.is_k_reduction)
+        {
+            return make_tuple(blk_id, blk_td);
+        }
+        else
+        {
+            return make_tuple(0, laneId);
+        }
+    }
+
+    __host__ __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto laneId  = GetLaneId();
+        const auto blk_idx = GetBlkIdx();
+
+        const auto blk_id = blk_idx[I0];
+        const auto blk_td = blk_idx[I1];
+
+        if constexpr(mfma_instr.is_k_reduction)
+        {
+            return make_tuple(blk_id, blk_td);
+        }
+        else
+        {
+            return make_tuple(0, laneId);
+        }
+    }
+
+    __device__ static CIndex GetBeginOfThreadBlk(index_t xdlops_i, index_t blk_i)
+    {
+        const auto blk_idx = GetBlkIdx();
+
+        const auto blk_id = blk_idx[I0];
+        const auto blk_td = blk_idx[I1];
+
+        index_t n_offset = blk_i * mfma_instr.n_per_blk + blk_td;
+        index_t m_offset = xdlops_i * mfma_instr.m_per_blk + blk_id * mfma_instr.group_size;
+
+        return TransposeC ? CIndex{n_offset, m_offset} : CIndex{m_offset, n_offset};
+    }
+
+    __device__ static CIndex4D GetBeginOfThreadBlk4D(index_t /* xdlops_i */, index_t /* blk_i */)
+    {
+        const auto blk_idx = GetBlkIdx();
+
+        const auto blk_id = blk_idx[I0];
+        const auto blk_td = blk_idx[I1];
+
+        return TransposeC ? CIndex4D{blk_td, I0, blk_id, I0} : CIndex4D{I0, blk_id, I0, blk_td};
+    }
+
+    static constexpr auto mfma = MfmaSelector<base_type, MPerXdlops, NPerXdlops>{};
+
+    static constexpr auto mfma_instr = mfma.selected_mfma;
+
+    static constexpr auto KPerXdlops  = mfma.GetKPerXdlops();
+    static constexpr auto K1PerXdlops = mfma.GetK1PerXdlops();
+    static constexpr auto K0PerXdlops = KPerXdlops / K1PerXdlops;
+
+    __host__ __device__ static constexpr auto GetCM0M1M2NThreadBlkLengths()
+    {
+        return make_tuple(
+            Number<mfma_instr.num_groups_per_blk>{}, I1, Number<mfma_instr.group_size>{}, I1);
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp
new file mode 100644
index 0000000..5fc11d9
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp
@@ -0,0 +1,288 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+
+// assume C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          device::TensorSpecialization TensorSpec>
+static auto MakeGridDescriptorPair(const std::vector<index_t>& gs_ms_ns_lengths_vec,
+                                   const std::vector<index_t>& gs_ms_ns_strides_vec)
+{
+    if(!(gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+         gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN))
+    {
+        throw std::runtime_error("wrong! dimension must match input lengths");
+    }
+
+    const auto to_tuple = [&](auto& vec, auto start, auto end) {
+        return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+    };
+
+    const auto gs_ms_ns_lengths =
+        to_tuple(gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+    const auto gs_ms_ns_strides =
+        to_tuple(gs_ms_ns_strides_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+    // dimension Ids for G0, G1, ...
+    constexpr auto gDimIds = typename arithmetic_sequence_gen<0, NumDimG, 1>::type{};
+
+    // dimension Ids for M0, M1, ...
+    constexpr auto mDimIds =
+        typename arithmetic_sequence_gen<NumDimG, NumDimG + NumDimM, 1>::type{};
+
+    // dimension Ids for N0, N1, ...
+    constexpr auto nDimIds =
+        typename arithmetic_sequence_gen<NumDimG + NumDimM, NumDimG + NumDimM + NumDimN, 1>::type{};
+
+    // lengths for G0, G1, ...
+    const auto gLengths = get_container_subset(gs_ms_ns_lengths, gDimIds);
+
+    // lengths for M0, M1, ...
+    const auto mLengths = get_container_subset(gs_ms_ns_lengths, mDimIds);
+
+    // lengths for N0, N1, ...
+    const auto nLengths = get_container_subset(gs_ms_ns_lengths, nDimIds);
+
+    if constexpr(TensorSpec == device::TensorSpecialization::Packed)
+    {
+        auto G = container_reduce(gLengths, math::multiplies{}, Number<1>{});
+        auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+        auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+        const auto grid_desc_g_mraw_nraw = make_naive_tensor_descriptor(
+            make_tuple(G, M, N),
+            make_tuple(gs_ms_ns_strides[Number<NumDimG - 1>{}],
+                       gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
+                       gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
+
+        const auto grid_desc_mraw_nraw = make_naive_tensor_descriptor(
+            make_tuple(M, N),
+            make_tuple(gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
+                       gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
+
+        return std::make_pair(grid_desc_g_mraw_nraw, grid_desc_mraw_nraw);
+    }
+    else
+    {
+        // naive tensor C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+        const auto grid_desc_gs_ms_ns =
+            make_naive_tensor_descriptor(gs_ms_ns_lengths, gs_ms_ns_strides);
+
+        // transformed tensor C[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
+        // N2 * ...]
+        // Note: This does not require padding as it only provides G offset calculation. Technically
+        // descriptor for only G is needed. Here we opt for backward compatibility purpose to return
+        // G_M_N
+        const auto grid_desc_g_mraw_nraw =
+            transform_tensor_descriptor(grid_desc_gs_ms_ns,
+                                        make_tuple(make_merge_transform(gLengths),
+                                                   make_merge_transform(mLengths),
+                                                   make_merge_transform(nLengths)),
+                                        make_tuple(gDimIds, mDimIds, nDimIds),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        const auto c_ms_ns_lengths = to_tuple(
+            gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto c_ms_ns_strides = to_tuple(
+            gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // transformed tensor C[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
+        // N2 * ...]
+        const auto grid_desc_ms_ns = make_naive_tensor_descriptor(c_ms_ns_lengths, c_ms_ns_strides);
+
+        const auto grid_desc_mraw_nraw = transform_tensor_descriptor(
+            grid_desc_ms_ns,
+            make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+            make_tuple(mDimIds - Number<NumDimG>{}, nDimIds - Number<NumDimG>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return std::make_pair(grid_desc_g_mraw_nraw, grid_desc_mraw_nraw);
+    }
+}
+
+template <typename NumDims_G_M_N_K_O, // Sequence<>
+          typename PerBlock_M_N_K_O,  // Sequence<>
+          device::GemmSpecialization GemmSpec,
+          device::TensorSpecialization ASpec,
+          device::TensorSpecialization B0Spec,
+          device::TensorSpecialization B1Spec,
+          device::TensorSpecialization CSpec>
+struct TransformBatchedContractionContractionToBatchedGemmGemm
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+
+    static constexpr index_t NumDimG = NumDims_G_M_N_K_O::At(I0);
+    static constexpr index_t NumDimM = NumDims_G_M_N_K_O::At(I1);
+    static constexpr index_t NumDimN = NumDims_G_M_N_K_O::At(I2);
+    static constexpr index_t NumDimK = NumDims_G_M_N_K_O::At(I3);
+    static constexpr index_t NumDimO = NumDims_G_M_N_K_O::At(I4);
+
+    static constexpr index_t MPerBlock = PerBlock_M_N_K_O::At(I0);
+    static constexpr index_t NPerBlock = PerBlock_M_N_K_O::At(I1);
+    static constexpr index_t KPerBlock = PerBlock_M_N_K_O::At(I2);
+    static constexpr index_t OPerBlock = PerBlock_M_N_K_O::At(I3);
+
+    static constexpr auto matrix_padder =
+        device::GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
+            MPerBlock, NPerBlock, KPerBlock, OPerBlock};
+
+    //
+    // A
+    //
+    static auto MakeAGridDescriptorPair(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        return MakeGridDescriptorPair<NumDimG, NumDimM, NumDimK, ASpec>(a_gs_ms_ks_lengths_vec,
+                                                                        a_gs_ms_ks_strides_vec);
+    }
+
+    // TODO: rename to G_MRaw_KRaw
+    static auto MakeAGridDescriptor_G_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                          const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        return MakeAGridDescriptorPair(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec).first;
+    }
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        return matrix_padder.PadADescriptor_M_K(
+            MakeAGridDescriptorPair(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec).second);
+    }
+
+    template <typename AGridDesc_M_K, typename Number>
+    __host__ __device__ static constexpr auto
+    MakeAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k, const Number& AK1)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    //
+    // B (alias of B0)
+    //
+    static auto MakeB0GridDescriptorPair(const std::vector<index_t>& b0_gs_ns_ks_lengths_vec,
+                                         const std::vector<index_t>& b0_gs_ns_ks_strides_vec)
+    {
+        return MakeGridDescriptorPair<NumDimG, NumDimN, NumDimK, B0Spec>(b0_gs_ns_ks_lengths_vec,
+                                                                         b0_gs_ns_ks_strides_vec);
+    }
+
+    // TODO: rename to G_MRaw_NRaw
+    static auto MakeB0GridDescriptor_G_N_K(const std::vector<index_t>& b0_gs_ns_ks_lengths_vec,
+                                           const std::vector<index_t>& b0_gs_ns_ks_strides_vec)
+    {
+        return MakeB0GridDescriptorPair(b0_gs_ns_ks_lengths_vec, b0_gs_ns_ks_strides_vec).first;
+    }
+    static auto MakeB0GridDescriptor_N_K(const std::vector<index_t>& b0_gs_ns_ks_lengths_vec,
+                                         const std::vector<index_t>& b0_gs_ns_ks_strides_vec)
+    {
+        // alias of matrix_padder.PadB0Descriptor_N_K
+        return matrix_padder.PadBDescriptor_N_K(
+            MakeB0GridDescriptorPair(b0_gs_ns_ks_lengths_vec, b0_gs_ns_ks_strides_vec).second);
+    }
+
+    template <typename BGridDesc_N_K, typename Number>
+    __host__ __device__ static constexpr auto
+    MakeB0GridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k, const Number& BK1)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    //
+    // B1
+    //
+    static auto MakeB1GridDescriptorPair(const std::vector<index_t>& b1_gs_os_ns_lengths_vec,
+                                         const std::vector<index_t>& b1_gs_os_ns_strides_vec)
+    {
+        return MakeGridDescriptorPair<NumDimG, NumDimO, NumDimN, B1Spec>(b1_gs_os_ns_lengths_vec,
+                                                                         b1_gs_os_ns_strides_vec);
+    }
+
+    // TODO: rename to G_NRaw_KRaw
+    static auto MakeB1GridDescriptor_G_N_K(const std::vector<index_t>& b1_gs_os_ns_lengths_vec,
+                                           const std::vector<index_t>& b1_gs_os_ns_strides_vec)
+    {
+        return MakeB1GridDescriptorPair(b1_gs_os_ns_lengths_vec, b1_gs_os_ns_strides_vec).first;
+    }
+    static auto MakeB1GridDescriptor_N_K(const std::vector<index_t>& b1_gs_os_ns_lengths_vec,
+                                         const std::vector<index_t>& b1_gs_os_ns_strides_vec)
+    {
+        // alias of matrix_padder.PadB1Descriptor_O_N
+        return matrix_padder.PadB1Descriptor_N_K(
+            MakeB1GridDescriptorPair(b1_gs_os_ns_lengths_vec, b1_gs_os_ns_strides_vec).second);
+    }
+
+    template <typename B1GridDesc_N_K, typename Number>
+    __host__ __device__ static constexpr auto
+    MakeB1GridDescriptor_BK0_N_BK1(const B1GridDesc_N_K& b1_grid_desc_n_k, const Number& B1K1)
+    {
+        const auto N = b1_grid_desc_n_k.GetLength(I0);
+        const auto K = b1_grid_desc_n_k.GetLength(I1);
+
+        const auto B1K0 = K / B1K1;
+
+        return transform_tensor_descriptor(
+            b1_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    //
+    // C
+    //
+    static auto MakeCGridDescriptorPair(const std::vector<index_t>& c_gs_ms_os_lengths_vec,
+                                        const std::vector<index_t>& c_gs_ms_os_strides_vec)
+    {
+        return MakeGridDescriptorPair<NumDimG, NumDimM, NumDimO, CSpec>(c_gs_ms_os_lengths_vec,
+                                                                        c_gs_ms_os_strides_vec);
+    }
+
+    // TODO: rename to G_MRaw_NRaw
+    static auto MakeCGridDescriptor_G_M_N(const std::vector<index_t>& c_gs_ms_os_lengths_vec,
+                                          const std::vector<index_t>& c_gs_ms_os_strides_vec)
+    {
+        return MakeCGridDescriptorPair(c_gs_ms_os_lengths_vec, c_gs_ms_os_strides_vec).first;
+    }
+    static auto MakeCGridDescriptor_M_N(const std::vector<index_t>& c_gs_ms_os_lengths_vec,
+                                        const std::vector<index_t>& c_gs_ms_os_strides_vec)
+    {
+        return matrix_padder.PadCDescriptor_M_N(
+            MakeCGridDescriptorPair(c_gs_ms_os_lengths_vec, c_gs_ms_os_strides_vec).second);
+    }
+};
+
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
new file mode 100644
index 0000000..13d0a28
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
@@ -0,0 +1,583 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+
+namespace ck {
+namespace tensor_operation {
+
+template <
+    index_t NDimSpatial,
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization ConvBwdDataSpecialization,
+    index_t AK1,
+    index_t BK1,
+    index_t GemmMPerBlock,
+    index_t GemmNPerBlock,
+    bool DoPadGemmM,
+    bool DoPadGemmN>
+struct TransformConvBwdDataToGemm_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    template <typename ALayout,
+              typename std::enable_if<NDimSpatial == 2 &&
+                                          is_same_v<ALayout, tensor_layout::convolution::GNHWK>,
+                                      bool>::type = false>
+    static auto MakeADescriptor_AK0_M_AK1(
+        const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& /* out_g_n_k_wos_strides */,
+        const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& /* wei_g_k_c_xs_strides */,
+        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& /* in_g_n_c_wis_strides */,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& /* input_right_pads */,
+        const std::array<index_t, NDimSpatial>& tildes)
+    {
+        index_t i_ytilde = tildes[0];
+        index_t i_xtilde = tildes[1];
+
+        const index_t N = in_g_n_c_wis_lengths[1];
+        const index_t K = wei_g_k_c_xs_lengths[1];
+
+        const index_t Hi = in_g_n_c_wis_lengths[3];
+        const index_t Wi = in_g_n_c_wis_lengths[4];
+
+        const index_t Ho = out_g_n_k_wos_lengths[3];
+        const index_t Wo = out_g_n_k_wos_lengths[4];
+
+        const index_t Y = wei_g_k_c_xs_lengths[3];
+        const index_t X = wei_g_k_c_xs_lengths[4];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t AK0 = K / AK1;
+
+        // assume packed
+        const auto out_n_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K));
+
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Ho * Wo),
+                           make_unmerge_transform(make_tuple(AK0, AK1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            const auto out_gemmak0_gemmm_gemmak1_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    out_gemmak0_gemmmraw_gemmak1_grid_desc,
+                    make_tuple(AK0, GemmMPerBlock, AK1),
+                    Sequence<false, DoPadGemmM, false>{});
+
+            return out_gemmak0_gemmm_gemmak1_grid_desc;
+        }
+        else
+        {
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_ho_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Ho, I0, I0),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                out_n_hop_wop_k_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(YDot, HTilde),
+                                         make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_slice_transform(YDot, I0, YDotSlice),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_slice_transform(XDot, I0, XDotSlice),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_unmerge_transform(make_tuple(AK0, AK1))),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5, 6>{}));
+
+            const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
+                out_n_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, AK0)),
+                           make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(AK1)),
+                make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto out_gemmak0_gemmm_gemmak1_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    out_gemmak0_gemmmraw_gemmak1_grid_desc,
+                    make_tuple(AK0, GemmMPerBlock, AK1),
+                    Sequence<false, DoPadGemmM, false>{});
+
+            return out_gemmak0_gemmm_gemmak1_grid_desc;
+        }
+    }
+
+    template <typename BLayout,
+              typename std::enable_if<NDimSpatial == 2 &&
+                                          is_same_v<BLayout, tensor_layout::convolution::GKYXC>,
+                                      bool>::type = false>
+    static auto MakeBDescriptor_BK0_N_BK1(
+        const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& /* out_g_n_k_wos_strides */,
+        const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& /* wei_g_k_c_xs_strides */,
+        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& /* in_g_n_c_wis_strides */,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& /* input_left_pads */,
+        const std::array<index_t, NDimSpatial>& /* input_right_pads */,
+        const std::array<index_t, NDimSpatial>& tildes)
+    {
+        index_t i_ytilde = tildes[0];
+        index_t i_xtilde = tildes[1];
+
+        const index_t N = in_g_n_c_wis_lengths[1];
+        const index_t K = wei_g_k_c_xs_lengths[1];
+        const index_t C = wei_g_k_c_xs_lengths[2];
+
+        const index_t Ho = out_g_n_k_wos_lengths[3];
+        const index_t Wo = out_g_n_k_wos_lengths[4];
+
+        const index_t Y = wei_g_k_c_xs_lengths[3];
+        const index_t X = wei_g_k_c_xs_lengths[4];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t BK0 = K / BK1;
+
+        // assume packed
+        const auto wei_k_y_x_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
+
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+            // B: weight tensor
+            const auto wei_gemmbk0_gemmnraw_gemmbk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+            make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, C), make_tuple(I0, I1));
+
+            const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    wei_gemmbk0_gemmnraw_gemmbk1_grid_desc,
+                    make_tuple(BK0, GemmNPerBlock, BK1),
+                    Sequence<false, DoPadGemmN, false>{});
+
+            return wei_gemmbk0_gemmn_gemmbk1_grid_desc;
+        }
+        else
+        {
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            // GemmK is different for each GEMM
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // B weight tensor
+            const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                wei_k_y_x_c_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_embed_transform(make_tuple(YDot, YTilde),
+                                                make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                           make_embed_transform(make_tuple(XDot, XTilde),
+                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto wei_bk0_bk1_ydotslice_xdotslice_c_grid_desc =
+                transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_slice_transform(YDot, I0, YDotSlice),
+                                                       make_slice_transform(XDot, I0, XDotSlice),
+                                                       make_freeze_transform(i_ytilde),
+                                                       make_freeze_transform(i_xtilde),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{},
+                                                       Sequence<1>{},
+                                                       Sequence<3>{},
+                                                       Sequence<2>{},
+                                                       Sequence<4>{},
+                                                       Sequence<5>{}),
+                                            make_tuple(Sequence<0, 1>{},
+                                                       Sequence<2>{},
+                                                       Sequence<3>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<4>{}));
+
+            const auto wei_gemmbk0_gemmnraw_gemmbk1_grid_desc = transform_tensor_descriptor(
+                wei_bk0_bk1_ydotslice_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, BK0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(BK1)),
+                make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    wei_gemmbk0_gemmnraw_gemmbk1_grid_desc,
+                    make_tuple(
+                        wei_gemmbk0_gemmnraw_gemmbk1_grid_desc.GetLength(I0), GemmNPerBlock, BK1),
+                    Sequence<false, DoPadGemmN, false>{});
+
+            return wei_gemmbk0_gemmn_gemmbk1_grid_desc;
+        }
+    }
+
+    template <typename CLayout,
+              typename std::enable_if<NDimSpatial == 2 &&
+                                          (is_same_v<CLayout, tensor_layout::convolution::GNHWC> ||
+                                           is_same_v<CLayout, tensor_layout::convolution::NHWGC> ||
+                                           is_same_v<CLayout, tensor_layout::convolution::G_NHW_C>),
+                                      bool>::type = false>
+    static auto
+    MakeCDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* out_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* wei_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads,
+                        const std::array<index_t, NDimSpatial>& tildes)
+    {
+        index_t i_ytilde = tildes[0];
+        index_t i_xtilde = tildes[1];
+
+        const index_t N = in_g_n_c_wis_lengths[1];
+        const index_t C = wei_g_k_c_xs_lengths[2];
+
+        const index_t Hi = in_g_n_c_wis_lengths[3];
+        const index_t Wi = in_g_n_c_wis_lengths[4];
+
+        const index_t Ho = out_g_n_k_wos_lengths[3];
+        const index_t Wo = out_g_n_k_wos_lengths[4];
+
+        const index_t Y = wei_g_k_c_xs_lengths[3];
+        const index_t X = wei_g_k_c_xs_lengths[4];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        // assume strided
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor(make_tuple(N, Hi, Wi, C),
+                                         make_tuple(in_g_n_c_wis_strides[1],
+                                                    in_g_n_c_wis_strides[3],
+                                                    in_g_n_c_wis_strides[4],
+                                                    in_g_n_c_wis_strides[2]));
+
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+            // C: input tensor
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmmraw_gemmnraw_grid_desc = transform_tensor_descriptor(
+                in_n_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = ck::tensor_operation::device::PadTensorDescriptor(
+                in_gemmmraw_gemmnraw_grid_desc,
+                make_tuple(GemmMPerBlock, GemmNPerBlock),
+                Sequence<DoPadGemmM, DoPadGemmN>{});
+
+            return in_gemmm_gemmn_grid_desc;
+        }
+        else
+        {
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // C: input tensor
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(YTilde, HTilde),
+                                                make_tuple(ConvDilationH, ConvStrideH)),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_ytilde),
+                           make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<>{},
+                           Sequence<2>{},
+                           Sequence<3>{}));
+
+            const auto in_gemmmraw_gemmnraw_grid_desc = transform_tensor_descriptor(
+                in_n_htildeslice_wtildeslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = ck::tensor_operation::device::PadTensorDescriptor(
+                in_gemmmraw_gemmnraw_grid_desc,
+                make_tuple(GemmMPerBlock, GemmNPerBlock),
+                Sequence<DoPadGemmM, DoPadGemmN>{});
+
+            return in_gemmm_gemmn_grid_desc;
+        }
+    }
+
+    // for input bias
+    template <typename CLayout,
+              typename std::enable_if<NDimSpatial == 2 &&
+                                          (is_same_v<CLayout, tensor_layout::convolution::GC> ||
+                                           is_same_v<CLayout, tensor_layout::convolution::G_C>),
+                                      bool>::type = false>
+    static auto
+    MakeCDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* out_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* wei_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* in_g_n_c_wis_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& /* input_right_pads */,
+                        const std::array<index_t, NDimSpatial>& /* tildes */)
+    {
+        const index_t N = in_g_n_c_wis_lengths[1];
+        const index_t C = wei_g_k_c_xs_lengths[2];
+
+        const index_t Hi = in_g_n_c_wis_lengths[3];
+        const index_t Wi = in_g_n_c_wis_lengths[4];
+
+        const index_t Ho = out_g_n_k_wos_lengths[3];
+        const index_t Wo = out_g_n_k_wos_lengths[4];
+
+        const index_t Y = wei_g_k_c_xs_lengths[3];
+        const index_t X = wei_g_k_c_xs_lengths[4];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+            const auto in_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, C), make_tuple(I0, I1));
+
+            return in_gemmm_gemmn_grid_desc;
+        }
+        else
+        {
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // bias tensor
+            const auto in_gemmmraw_gemmnraw_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N * HTildeSlice * WTildeSlice, C), make_tuple(I0, I1));
+
+            const auto in_gemmm_gemmn_grid_desc = ck::tensor_operation::device::PadTensorDescriptor(
+                in_gemmmraw_gemmnraw_grid_desc,
+                make_tuple(GemmMPerBlock, GemmNPerBlock),
+                Sequence<DoPadGemmM, DoPadGemmN>{});
+
+            return in_gemmm_gemmn_grid_desc;
+        }
+    }
+};
+
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp b/3rdparty/composable_kernel/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp
new file mode 100644
index 0000000..1b5e64b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp
@@ -0,0 +1,880 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/utility/numeric.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+
+template <index_t NDimSpatial, device::ConvolutionForwardSpecialization ConvForwardSpecialization>
+struct TransformConvFwdToGemm
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    template <typename ALayout,
+              typename std::enable_if<NDimSpatial == 1 &&
+                                          is_same_v<ALayout, tensor_layout::convolution::GNWC>,
+                                      bool>::type = false>
+    static auto
+    MakeADescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* a_g_n_c_wis_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Wi = a_g_n_c_wis_lengths[3];
+
+        const index_t Wo = c_g_n_k_wos_lengths[3];
+
+        const index_t ConvStrideW = conv_filter_strides[0];
+
+        if constexpr(ConvForwardSpecialization ==
+                     device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NWo =
+                N * ck::accumulate_n<index_t>(
+                        c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+            const auto in_gemmm_gemmk_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(NWo, C));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          device::ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            const auto in_n_wi_c_desc = make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+            const auto in_n_wo_c_desc = transform_tensor_descriptor(
+                in_n_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_gemmm_gemmk_desc = transform_tensor_descriptor(
+                in_n_wo_c_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Wo)), make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else
+        {
+            const index_t X             = b_g_k_c_xs_lengths[3];
+            const index_t ConvDilationW = conv_filter_dilations[0];
+            const index_t InLeftPadW    = input_left_pads[0];
+            const index_t InRightPadW   = input_right_pads[0];
+
+            const auto in_n_wi_c_desc = make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+            const auto in_n_wip_c_desc = transform_tensor_descriptor(
+                in_n_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_x_wo_c_desc = transform_tensor_descriptor(
+                in_n_wip_c_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmm_gemmk_desc =
+                transform_tensor_descriptor(in_n_x_wo_c_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Wo)),
+                                                       make_merge_transform(make_tuple(X, C))),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+    }
+
+    template <typename ALayout,
+              typename std::enable_if<NDimSpatial == 2 &&
+                                          is_same_v<ALayout, tensor_layout::convolution::GNHWC>,
+                                      bool>::type = false>
+    static auto
+    MakeADescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* a_g_n_c_wis_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Hi = a_g_n_c_wis_lengths[3];
+        const index_t Wi = a_g_n_c_wis_lengths[4];
+
+        const index_t Ho = c_g_n_k_wos_lengths[3];
+        const index_t Wo = c_g_n_k_wos_lengths[4];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        if constexpr(ConvForwardSpecialization ==
+                     device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NHoWo =
+                N * ck::accumulate_n<index_t>(
+                        c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+            const auto in_gemmm_gemmk_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(NHoWo, C));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          device::ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            const auto in_n_hi_wi_c_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_ho_wo_c_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmm_gemmk_desc =
+                transform_tensor_descriptor(in_n_ho_wo_c_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else
+        {
+            const index_t Y = b_g_k_c_xs_lengths[3];
+            const index_t X = b_g_k_c_xs_lengths[4];
+
+            const index_t ConvDilationH = conv_filter_dilations[0];
+            const index_t ConvDilationW = conv_filter_dilations[1];
+
+            const index_t InLeftPadH = input_left_pads[0];
+            const index_t InLeftPadW = input_left_pads[1];
+
+            const index_t InRightPadH = input_right_pads[0];
+            const index_t InRightPadW = input_right_pads[1];
+
+            const auto in_n_hi_wi_c_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmm_gemmk_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                                       make_merge_transform(make_tuple(Y, X, C))),
+                                            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+    }
+
+    template <typename ALayout,
+              typename std::enable_if<NDimSpatial == 3 &&
+                                          is_same_v<ALayout, tensor_layout::convolution::GNDHWC>,
+                                      bool>::type = false>
+    static auto
+    MakeADescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* a_g_n_c_wis_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Di = a_g_n_c_wis_lengths[3];
+        const index_t Hi = a_g_n_c_wis_lengths[4];
+        const index_t Wi = a_g_n_c_wis_lengths[5];
+
+        const index_t Do = c_g_n_k_wos_lengths[3];
+        const index_t Ho = c_g_n_k_wos_lengths[4];
+        const index_t Wo = c_g_n_k_wos_lengths[5];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        if constexpr(ConvForwardSpecialization ==
+                     device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NDoHoWo =
+                N * ck::accumulate_n<index_t>(
+                        c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+            const auto in_gemmm_gemmk_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(NDoHoWo, C));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          device::ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            const auto in_n_di_hi_wi_c_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+
+            const auto in_n_do_ho_wo_c_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Do), make_tuple(ConvStrideD)),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_gemmm_gemmk_desc = transform_tensor_descriptor(
+                in_n_do_ho_wo_c_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else
+        {
+            const index_t Z = b_g_k_c_xs_lengths[3];
+            const index_t Y = b_g_k_c_xs_lengths[4];
+            const index_t X = b_g_k_c_xs_lengths[5];
+
+            const index_t ConvDilationD = conv_filter_dilations[0];
+            const index_t ConvDilationH = conv_filter_dilations[1];
+            const index_t ConvDilationW = conv_filter_dilations[2];
+
+            const index_t InLeftPadD = input_left_pads[0];
+            const index_t InLeftPadH = input_left_pads[1];
+            const index_t InLeftPadW = input_left_pads[2];
+
+            const index_t InRightPadD = input_right_pads[0];
+            const index_t InRightPadH = input_right_pads[1];
+            const index_t InRightPadW = input_right_pads[2];
+
+            const auto in_n_di_hi_wi_c_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_z_do_y_ho_x_wo_c_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmm_gemmk_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_merge_transform(make_tuple(Z, Y, X, C))),
+                make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+    }
+
+    // TODO: implement ck::tensor_layout::convolution that describe packed/strided dimemsion as
+    // properties
+    template <typename ALayout,
+              typename std::enable_if<NDimSpatial == 1 &&
+                                          (is_same_v<ALayout, tensor_layout::convolution::G_NW_C> ||
+                                           is_same_v<ALayout, tensor_layout::convolution::NWGC>),
+                                      bool>::type = false>
+    static auto
+    MakeADescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Wi = a_g_n_c_wis_lengths[3];
+
+        const index_t Wo = c_g_n_k_wos_lengths[3];
+
+        const index_t ConvStrideW = conv_filter_strides[0];
+
+        if constexpr(ConvForwardSpecialization ==
+                     device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NHoWo =
+                N * ck::accumulate_n<index_t>(
+                        c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+            // This is different
+            const index_t WiStride = a_g_n_c_wis_strides[2 + NDimSpatial];
+            const auto CStride     = I1;
+
+            const auto in_gemmm_gemmk_desc =
+                make_naive_tensor_descriptor(make_tuple(NHoWo, C), make_tuple(WiStride, CStride));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          device::ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t WiStride = a_g_n_c_wis_strides[3];
+            const auto CStride     = I1;
+
+            const auto in_n_wi_c_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Wi, C), make_tuple(NStride, WiStride, CStride));
+
+            const auto in_n_wo_c_desc = transform_tensor_descriptor(
+                in_n_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_gemmm_gemmk_desc = transform_tensor_descriptor(
+                in_n_wo_c_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Wo)), make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else
+        {
+            const index_t X             = b_g_k_c_xs_lengths[3];
+            const index_t ConvDilationW = conv_filter_dilations[0];
+            const index_t InLeftPadW    = input_left_pads[0];
+            const index_t InRightPadW   = input_right_pads[0];
+
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t WiStride = a_g_n_c_wis_strides[3];
+            const auto CStride     = I1;
+
+            const auto in_n_wi_c_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Wi, C), make_tuple(NStride, WiStride, CStride));
+
+            const auto in_n_wip_c_desc = transform_tensor_descriptor(
+                in_n_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_x_wo_c_desc = transform_tensor_descriptor(
+                in_n_wip_c_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmm_gemmk_desc =
+                transform_tensor_descriptor(in_n_x_wo_c_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Wo)),
+                                                       make_merge_transform(make_tuple(X, C))),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+    }
+
+    template <typename ALayout,
+              typename std::enable_if<
+                  NDimSpatial == 2 && (is_same_v<ALayout, tensor_layout::convolution::G_NHW_C> ||
+                                       is_same_v<ALayout, tensor_layout::convolution::NHWGC>),
+                  bool>::type = false>
+    static auto
+    MakeADescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Hi = a_g_n_c_wis_lengths[3];
+        const index_t Wi = a_g_n_c_wis_lengths[4];
+
+        const index_t Ho = c_g_n_k_wos_lengths[3];
+        const index_t Wo = c_g_n_k_wos_lengths[4];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        if constexpr(ConvForwardSpecialization ==
+                     device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NHoWo =
+                N * ck::accumulate_n<index_t>(
+                        c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+            // This is different
+            const index_t WiStride = a_g_n_c_wis_strides[2 + NDimSpatial];
+            const auto CStride     = I1;
+
+            const auto in_gemmm_gemmk_desc =
+                make_naive_tensor_descriptor(make_tuple(NHoWo, C), make_tuple(WiStride, CStride));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          device::ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t HiStride = a_g_n_c_wis_strides[3];
+            const index_t WiStride = a_g_n_c_wis_strides[4];
+            const auto CStride     = I1;
+
+            const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Hi, Wi, C), make_tuple(NStride, HiStride, WiStride, CStride));
+
+            const auto in_n_ho_wo_c_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmm_gemmk_desc =
+                transform_tensor_descriptor(in_n_ho_wo_c_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else
+        {
+            const index_t Y = b_g_k_c_xs_lengths[3];
+            const index_t X = b_g_k_c_xs_lengths[4];
+
+            const index_t ConvDilationH = conv_filter_dilations[0];
+            const index_t ConvDilationW = conv_filter_dilations[1];
+
+            const index_t InLeftPadH = input_left_pads[0];
+            const index_t InLeftPadW = input_left_pads[1];
+
+            const index_t InRightPadH = input_right_pads[0];
+            const index_t InRightPadW = input_right_pads[1];
+
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t HiStride = a_g_n_c_wis_strides[3];
+            const index_t WiStride = a_g_n_c_wis_strides[4];
+            const auto CStride     = I1;
+
+            const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Hi, Wi, C), make_tuple(NStride, HiStride, WiStride, CStride));
+
+            const auto in_n_hip_wip_c_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmm_gemmk_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                                       make_merge_transform(make_tuple(Y, X, C))),
+                                            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+    }
+
+    template <typename ALayout,
+              typename std::enable_if<
+                  NDimSpatial == 3 && (is_same_v<ALayout, tensor_layout::convolution::G_NDHW_C> ||
+                                       is_same_v<ALayout, tensor_layout::convolution::NDHWGC>),
+                  bool>::type = false>
+    static auto
+    MakeADescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Di = a_g_n_c_wis_lengths[3];
+        const index_t Hi = a_g_n_c_wis_lengths[4];
+        const index_t Wi = a_g_n_c_wis_lengths[5];
+
+        const index_t Do = c_g_n_k_wos_lengths[3];
+        const index_t Ho = c_g_n_k_wos_lengths[4];
+        const index_t Wo = c_g_n_k_wos_lengths[5];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        if constexpr(ConvForwardSpecialization ==
+                     device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NDoHoWo =
+                N * ck::accumulate_n<index_t>(
+                        c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+            // This is different
+            const index_t WiStride = a_g_n_c_wis_strides[2 + NDimSpatial];
+            const auto CStride     = I1;
+
+            const auto in_gemmm_gemmk_desc =
+                make_naive_tensor_descriptor(make_tuple(NDoHoWo, C), make_tuple(WiStride, CStride));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          device::ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t DiStride = a_g_n_c_wis_strides[3];
+            const index_t HiStride = a_g_n_c_wis_strides[4];
+            const index_t WiStride = a_g_n_c_wis_strides[5];
+            const auto CStride     = I1;
+
+            const auto in_n_di_hi_wi_c_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Di, Hi, Wi, C),
+                make_tuple(NStride, DiStride, HiStride, WiStride, CStride));
+
+            const auto in_n_do_ho_wo_c_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Do), make_tuple(ConvStrideD)),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_gemmm_gemmk_desc = transform_tensor_descriptor(
+                in_n_do_ho_wo_c_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else
+        {
+            const index_t Z = b_g_k_c_xs_lengths[3];
+            const index_t Y = b_g_k_c_xs_lengths[4];
+            const index_t X = b_g_k_c_xs_lengths[5];
+
+            const index_t ConvDilationD = conv_filter_dilations[0];
+            const index_t ConvDilationH = conv_filter_dilations[1];
+            const index_t ConvDilationW = conv_filter_dilations[2];
+
+            const index_t InLeftPadD = input_left_pads[0];
+            const index_t InLeftPadH = input_left_pads[1];
+            const index_t InLeftPadW = input_left_pads[2];
+
+            const index_t InRightPadD = input_right_pads[0];
+            const index_t InRightPadH = input_right_pads[1];
+            const index_t InRightPadW = input_right_pads[2];
+
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t DiStride = a_g_n_c_wis_strides[3];
+            const index_t HiStride = a_g_n_c_wis_strides[4];
+            const index_t WiStride = a_g_n_c_wis_strides[5];
+            const auto CStride     = I1;
+
+            const auto in_n_di_hi_wi_c_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Di, Hi, Wi, C),
+                make_tuple(NStride, DiStride, HiStride, WiStride, CStride));
+
+            const auto in_n_hip_wip_c_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_z_do_y_ho_x_wo_c_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmm_gemmk_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_merge_transform(make_tuple(Z, Y, X, C))),
+                make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+    }
+
+    template <typename BLayout,
+              typename std::enable_if<is_same_v<BLayout, tensor_layout::convolution::GKXC> ||
+                                          is_same_v<BLayout, tensor_layout::convolution::GKYXC> ||
+                                          is_same_v<BLayout, tensor_layout::convolution::GKZYXC>,
+                                      bool>::type = false>
+    static auto
+    MakeBDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */)
+    {
+        const index_t K = b_g_k_c_xs_lengths[1];
+        const index_t C = b_g_k_c_xs_lengths[2];
+
+        const index_t YX = ck::accumulate_n<index_t>(
+            b_g_k_c_xs_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+        const auto wei_gemmn_gemmk_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, YX * C));
+
+        return wei_gemmn_gemmk_desc;
+    }
+
+    template <
+        typename BLayout,
+        typename std::enable_if<is_same_v<BLayout, tensor_layout::convolution::G_K_X_C> ||
+                                    is_same_v<BLayout, tensor_layout::convolution::G_K_YX_C> ||
+                                    is_same_v<BLayout, tensor_layout::convolution::G_K_ZYX_C> ||
+                                    is_same_v<BLayout, tensor_layout::convolution::KXGC> ||
+                                    is_same_v<BLayout, tensor_layout::convolution::KYXGC> ||
+                                    is_same_v<BLayout, tensor_layout::convolution::KZYXGC>,
+                                bool>::type = false>
+    static auto MakeBDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                    const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const index_t K = b_g_k_c_xs_lengths[1];
+        const index_t C = b_g_k_c_xs_lengths[2];
+
+        const index_t YX = ck::accumulate_n<index_t>(
+            b_g_k_c_xs_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+        const index_t KStride = b_g_k_c_xs_strides[1];
+        const index_t XStride = b_g_k_c_xs_strides[2 + NDimSpatial];
+        const auto CStride    = I1;
+
+        const auto wei_k_yx_c_desc = make_naive_tensor_descriptor(
+            make_tuple(K, YX, C), make_tuple(KStride, XStride, CStride));
+
+        const auto wei_gemmn_gemmk_desc = transform_tensor_descriptor(
+            wei_k_yx_c_desc,
+            make_tuple(make_pass_through_transform(K), make_merge_transform(make_tuple(YX, C))),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return wei_gemmn_gemmk_desc;
+    }
+
+    template <typename CLayout,
+              typename std::enable_if<is_same_v<CLayout, tensor_layout::convolution::GNWK> ||
+                                          is_same_v<CLayout, tensor_layout::convolution::GNHWK> ||
+                                          is_same_v<CLayout, tensor_layout::convolution::GNDHWK>,
+                                      bool>::type = false>
+    static auto
+    MakeCDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */)
+    {
+        const index_t N = c_g_n_k_wos_lengths[1];
+        const index_t K = c_g_n_k_wos_lengths[2];
+
+        const index_t NHoWo =
+            N * ck::accumulate_n<index_t>(
+                    c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+        const auto out_gemmm_gemmn_desc = make_naive_tensor_descriptor_packed(make_tuple(NHoWo, K));
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    template <
+        typename CLayout,
+        typename std::enable_if<is_same_v<CLayout, tensor_layout::convolution::G_NW_K> ||
+                                    is_same_v<CLayout, tensor_layout::convolution::G_NHW_K> ||
+                                    is_same_v<CLayout, tensor_layout::convolution::G_NDHW_K> ||
+                                    is_same_v<CLayout, tensor_layout::convolution::NWGK> ||
+                                    is_same_v<CLayout, tensor_layout::convolution::NHWGK> ||
+                                    is_same_v<CLayout, tensor_layout::convolution::NDHWGK>,
+                                bool>::type = false>
+    static auto MakeCDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                                    const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides)
+    {
+        const index_t N = c_g_n_k_wos_lengths[1];
+        const index_t K = c_g_n_k_wos_lengths[2];
+
+        const auto KStride     = I1;
+        const index_t WoStride = c_g_n_k_wos_strides[NDimSpatial + 2];
+
+        const index_t NHoWo =
+            N * ck::accumulate_n<index_t>(
+                    c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+        const auto out_gemmm_gemmn_desc =
+            make_naive_tensor_descriptor(make_tuple(NHoWo, K), make_tuple(WoStride, KStride));
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    // for output bias
+    template <typename CLayout,
+              typename std::enable_if<is_same_v<CLayout, tensor_layout::convolution::GK> ||
+                                          is_same_v<CLayout, tensor_layout::convolution::G_K>,
+                                      bool>::type = false>
+    static auto
+    MakeCDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */)
+    {
+        const index_t N = c_g_n_k_wos_lengths[1];
+        const index_t K = c_g_n_k_wos_lengths[2];
+
+        const index_t NHoWo =
+            N * ck::accumulate_n<index_t>(
+                    c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+        const auto out_gemmm_gemmn_desc =
+            make_naive_tensor_descriptor(make_tuple(NHoWo, K), make_tuple(I0, I1));
+
+        return out_gemmm_gemmn_desc;
+    }
+};
+
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/amd_address_space.hpp b/3rdparty/composable_kernel/include/ck/utility/amd_address_space.hpp
new file mode 100644
index 0000000..9f15259
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/amd_address_space.hpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "c_style_pointer_cast.hpp"
+
+// Address Space for AMDGCN
+// https://llvm.org/docs/AMDGPUUsage.html#address-space
+
+namespace ck {
+
+enum struct AddressSpaceEnum
+{
+    Generic,
+    Global,
+    Lds,
+    Sgpr,
+    Vgpr,
+};
+
+template <typename T>
+__device__ T* cast_pointer_to_generic_address_space(T CK_CONSTANT_ADDRESS_SPACE* p)
+{
+    // cast a pointer in "Constant" address space (4) to "Generic" address space (0)
+    // only c-style pointer cast seems be able to be compiled
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wold-style-cast"
+    return (T*)p; // NOLINT(old-style-cast)
+#pragma clang diagnostic pop
+}
+
+template <typename T>
+__host__ __device__ T CK_CONSTANT_ADDRESS_SPACE* cast_pointer_to_constant_address_space(T* p)
+{
+    // cast a pointer in "Generic" address space (0) to "Constant" address space (4)
+    // only c-style pointer cast seems be able to be compiled
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wold-style-cast"
+    return (T CK_CONSTANT_ADDRESS_SPACE*)p; // NOLINT(old-style-cast)
+#pragma clang diagnostic pop
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/amd_buffer_addressing.hpp b/3rdparty/composable_kernel/include/ck/utility/amd_buffer_addressing.hpp
new file mode 100644
index 0000000..7929535
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/amd_buffer_addressing.hpp
@@ -0,0 +1,1177 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "data_type.hpp"
+
+namespace ck {
+
+template <typename T>
+union BufferResource
+{
+    __device__ constexpr BufferResource() : content{} {}
+
+    // 128 bit SGPRs to supply buffer resource in buffer instructions
+    // https://rocm-documentation.readthedocs.io/en/latest/GCN_ISA_Manuals/testdocbook.html#vector-memory-buffer-instructions
+    int32x4_t content;
+    StaticallyIndexedArray<T*, 2> address;
+    StaticallyIndexedArray<int32_t, 4> range;
+    StaticallyIndexedArray<int32_t, 4> config;
+};
+
+template <typename T>
+__device__ int32x4_t make_wave_buffer_resource(T* p_wave, index_t element_space_size)
+{
+    BufferResource<T> wave_buffer_resource;
+
+    // wavewise base address (64 bit)
+    wave_buffer_resource.address(Number<0>{}) = const_cast<remove_cv_t<T>*>(p_wave);
+    // wavewise range (32 bit)
+    wave_buffer_resource.range(Number<2>{}) = element_space_size * sizeof(T);
+    // wavewise setting (32 bit)
+    wave_buffer_resource.config(Number<3>{}) = CK_BUFFER_RESOURCE_3RD_DWORD;
+
+    return wave_buffer_resource.content;
+}
+
+template <typename T>
+__device__ int32x4_t make_wave_buffer_resource_with_default_range(T* p_wave)
+{
+    BufferResource<T> wave_buffer_resource;
+
+    // wavewise base address (64 bit)
+    wave_buffer_resource.address(Number<0>{}) = const_cast<remove_cv_t<T>*>(p_wave);
+    // wavewise range (32 bit)
+    wave_buffer_resource.range(Number<2>{}) = 0xffffffff; // max possible range
+    // wavewise setting (32 bit)
+    wave_buffer_resource.config(Number<3>{}) = CK_BUFFER_RESOURCE_3RD_DWORD;
+
+    return wave_buffer_resource.content;
+}
+
+// buffer load i8
+__device__ int8_t
+llvm_amdgcn_raw_buffer_load_i8(int32x4_t srsrc,
+                               index_t voffset,
+                               index_t soffset,
+                               index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i8");
+
+__device__ int8x2_t
+llvm_amdgcn_raw_buffer_load_i8x2(int32x4_t srsrc,
+                                 index_t voffset,
+                                 index_t soffset,
+                                 index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i8");
+
+__device__ int8x4_t
+llvm_amdgcn_raw_buffer_load_i8x4(int32x4_t srsrc,
+                                 index_t voffset,
+                                 index_t soffset,
+                                 index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i8");
+
+// buffer load i16
+__device__ bhalf_t
+llvm_amdgcn_raw_buffer_load_i16(int32x4_t srsrc,
+                                index_t voffset,
+                                index_t soffset,
+                                index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i16");
+
+__device__ bhalf2_t
+llvm_amdgcn_raw_buffer_load_i16x2(int32x4_t srsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i16");
+
+__device__ bhalf4_t
+llvm_amdgcn_raw_buffer_load_i16x4(int32x4_t srsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i16");
+
+// buffer load i32
+__device__ int32_t
+llvm_amdgcn_raw_buffer_load_i32(int32x4_t srsrc,
+                                index_t voffset,
+                                index_t soffset,
+                                index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i32");
+
+__device__ int32x2_t
+llvm_amdgcn_raw_buffer_load_i32x2(int32x4_t srsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i32");
+
+__device__ int32x4_t
+llvm_amdgcn_raw_buffer_load_i32x4(int32x4_t srsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i32");
+
+// buffer load fp16
+__device__ half_t
+llvm_amdgcn_raw_buffer_load_fp16(int32x4_t srsrc,
+                                 index_t voffset,
+                                 index_t soffset,
+                                 index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.f16");
+
+__device__ half2_t
+llvm_amdgcn_raw_buffer_load_fp16x2(int32x4_t srsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2f16");
+
+__device__ half4_t
+llvm_amdgcn_raw_buffer_load_fp16x4(int32x4_t srsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4f16");
+
+// buffer load fp32
+__device__ float
+llvm_amdgcn_raw_buffer_load_fp32(int32x4_t srsrc,
+                                 index_t voffset,
+                                 index_t soffset,
+                                 index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.f32");
+
+__device__ float2_t
+llvm_amdgcn_raw_buffer_load_fp32x2(int32x4_t srsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2f32");
+
+__device__ float4_t
+llvm_amdgcn_raw_buffer_load_fp32x4(int32x4_t srsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4f32");
+
+// buffer store i8
+__device__ void
+llvm_amdgcn_raw_buffer_store_i8(int8_t vdata,
+                                int32x4_t rsrc,
+                                index_t voffset,
+                                index_t soffset,
+                                index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i8");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_i8x2(int8x2_t vdata,
+                                  int32x4_t rsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i8");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_i8x4(int8x4_t vdata,
+                                  int32x4_t rsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i8");
+
+// buffer store i16
+__device__ void
+llvm_amdgcn_raw_buffer_store_i16(bhalf_t vdata,
+                                 int32x4_t rsrc,
+                                 index_t voffset,
+                                 index_t soffset,
+                                 index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i16");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_i16x2(bhalf2_t vdata,
+                                   int32x4_t rsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i16");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_i16x4(bhalf4_t vdata,
+                                   int32x4_t rsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i16");
+
+// buffer store i32
+__device__ void
+llvm_amdgcn_raw_buffer_store_i32(int32_t vdata,
+                                 int32x4_t rsrc,
+                                 index_t voffset,
+                                 index_t soffset,
+                                 index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i32");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_i32x2(int32x2_t vdata,
+                                   int32x4_t rsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i32");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_i32x4(int32x4_t vdata,
+                                   int32x4_t rsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i32");
+
+// buffer store fp16
+__device__ void
+llvm_amdgcn_raw_buffer_store_fp16(half_t vdata,
+                                  int32x4_t rsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.f16");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_fp16x2(half2_t vdata,
+                                    int32x4_t rsrc,
+                                    index_t voffset,
+                                    index_t soffset,
+                                    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2f16");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_fp16x4(half4_t vdata,
+                                    int32x4_t rsrc,
+                                    index_t voffset,
+                                    index_t soffset,
+                                    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4f16");
+
+// buffer store fp32
+__device__ void
+llvm_amdgcn_raw_buffer_store_fp32(float vdata,
+                                  int32x4_t rsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.f32");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_fp32x2(float2_t vdata,
+                                    int32x4_t rsrc,
+                                    index_t voffset,
+                                    index_t soffset,
+                                    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2f32");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_fp32x4(float4_t vdata,
+                                    int32x4_t rsrc,
+                                    index_t voffset,
+                                    index_t soffset,
+                                    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4f32");
+
+// buffer atomic-add fp16
+__device__ half2_t llvm_amdgcn_raw_buffer_atomic_add_fp16x2(
+    half2_t vdata,
+    int32x4_t rsrc,
+    index_t voffset,
+    index_t soffset,
+    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.v2f16");
+
+// buffer atomic-add i32
+__device__ int32_t llvm_amdgcn_raw_buffer_atomic_add_i32(
+    int32_t vdata,
+    int32x4_t rsrc,
+    index_t voffset,
+    index_t soffset,
+    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.add.i32");
+
+// buffer atomic-add fp32
+__device__ float llvm_amdgcn_raw_buffer_atomic_add_fp32(
+    float vdata,
+    int32x4_t rsrc,
+    index_t voffset,
+    index_t soffset,
+    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.f32");
+
+// buffer atomic-add fp32
+__device__ double
+llvm_amdgcn_raw_buffer_atomic_max_fp64(double vdata,
+                                       int32x4_t rsrc, // dst_wave_buffer_resource
+                                       int voffset,    // dst_thread_addr_offset
+                                       int soffset,    // dst_wave_addr_offset
+                                       int glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fmax.f64");
+
+template <typename T, index_t N>
+__device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_wave_buffer_resource,
+                                                                 index_t src_thread_addr_offset,
+                                                                 index_t src_wave_addr_offset)
+{
+    static_assert(
+        (is_same<T, double>::value && (N == 1 || N == 2 || N == 4)) ||
+            (is_same<T, float>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, bhalf_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
+        "wrong! not implemented");
+
+    if constexpr(is_same<T, double>::value)
+    {
+        // use fp32 load to mimic fp64 load
+        if constexpr(N == 1)
+        {
+            const float2_t tmp = llvm_amdgcn_raw_buffer_load_fp32x2(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            return bit_cast<double>(tmp);
+        }
+        else if constexpr(N == 2)
+        {
+            const float4_t tmp = llvm_amdgcn_raw_buffer_load_fp32x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            return bit_cast<double2_t>(tmp);
+        }
+        else if constexpr(N == 4)
+        {
+            const float4_t f32_0 = llvm_amdgcn_raw_buffer_load_fp32x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            const float4_t f32_1 =
+                llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource,
+                                                   src_thread_addr_offset,
+                                                   src_wave_addr_offset + 4 * sizeof(float),
+                                                   0);
+            vector_type<double, 4> tmp;
+
+            tmp.AsType<double2_t>()(Number<0>{}) = bit_cast<double2_t>(f32_0);
+            tmp.AsType<double2_t>()(Number<1>{}) = bit_cast<double2_t>(f32_1);
+
+            return tmp.AsType<double4_t>()(Number<0>{});
+        }
+    }
+    else if constexpr(is_same<T, float>::value)
+    {
+        if constexpr(N == 1)
+        {
+            return llvm_amdgcn_raw_buffer_load_fp32(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 2)
+        {
+            return llvm_amdgcn_raw_buffer_load_fp32x2(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 4)
+        {
+            return llvm_amdgcn_raw_buffer_load_fp32x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 8)
+        {
+            vector_type<float, 8> tmp;
+
+            tmp.AsType<float4_t>()(Number<0>{}) = llvm_amdgcn_raw_buffer_load_fp32x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            tmp.AsType<float4_t>()(Number<1>{}) =
+                llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource,
+                                                   src_thread_addr_offset,
+                                                   src_wave_addr_offset + 4 * sizeof(float),
+                                                   0);
+
+            return tmp.AsType<float8_t>()(Number<0>{});
+        }
+    }
+    else if constexpr(is_same<T, half_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            return llvm_amdgcn_raw_buffer_load_fp16(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 2)
+        {
+            return llvm_amdgcn_raw_buffer_load_fp16x2(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 4)
+        {
+            return llvm_amdgcn_raw_buffer_load_fp16x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 8)
+        {
+            // use fp32 load to mimic fp16 load
+            float4_t tmp = llvm_amdgcn_raw_buffer_load_fp32x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            return bit_cast<half8_t>(tmp);
+        }
+    }
+    else if constexpr(is_same<T, bhalf_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            return llvm_amdgcn_raw_buffer_load_i16(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 2)
+        {
+            return llvm_amdgcn_raw_buffer_load_i16x2(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 4)
+        {
+            return llvm_amdgcn_raw_buffer_load_i16x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 8)
+        {
+            int32x4_t tmp = llvm_amdgcn_raw_buffer_load_i32x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            return bit_cast<bhalf8_t>(tmp);
+        }
+    }
+    else if constexpr(is_same<T, int32_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            return llvm_amdgcn_raw_buffer_load_i32(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 2)
+        {
+            return llvm_amdgcn_raw_buffer_load_i32x2(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 4)
+        {
+            return llvm_amdgcn_raw_buffer_load_i32x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 8)
+        {
+            vector_type<int32_t, 8> tmp;
+
+            tmp.AsType<int32x4_t>()(Number<0>{}) = llvm_amdgcn_raw_buffer_load_i32x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            tmp.AsType<int32x4_t>()(Number<1>{}) =
+                llvm_amdgcn_raw_buffer_load_i32x4(src_wave_buffer_resource,
+                                                  src_thread_addr_offset,
+                                                  src_wave_addr_offset + 4 * sizeof(int32_t),
+                                                  0);
+            return tmp.AsType<int32x8_t>()(Number<0>{});
+        }
+    }
+    else if constexpr(is_same<T, int8_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            return llvm_amdgcn_raw_buffer_load_i8(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+        }
+        else if constexpr(N == 2)
+        {
+#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
+            return llvm_amdgcn_raw_buffer_load_i8x2(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+#else
+            int16_t tmp = llvm_amdgcn_raw_buffer_load_i16(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            return bit_cast<int8x2_t>(tmp);
+#endif
+        }
+        else if constexpr(N == 4)
+        {
+#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
+            return llvm_amdgcn_raw_buffer_load_i8x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+#else
+            int32_t tmp = llvm_amdgcn_raw_buffer_load_i32(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            return bit_cast<int8x4_t>(tmp);
+#endif
+        }
+        else if constexpr(N == 8)
+        {
+#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
+            vector_type<int8_t, 8> tmp;
+
+            tmp.AsType<int8x4_t>()(Number<0>{}) = llvm_amdgcn_raw_buffer_load_i8x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            tmp.AsType<int8x4_t>()(Number<1>{}) =
+                llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource,
+                                                 src_thread_addr_offset,
+                                                 src_wave_addr_offset + 4 * sizeof(int8_t),
+                                                 0);
+
+            return tmp.AsType<int8x8_t>()(Number<0>{});
+#else
+            int32x2_t tmp = llvm_amdgcn_raw_buffer_load_i32x2(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            return bit_cast<int8x8_t>(tmp);
+#endif
+        }
+        else if constexpr(N == 16)
+        {
+#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
+            vector_type<int8_t, 16> tmp;
+
+            tmp.AsType<int8x4_t>()(Number<0>{}) = llvm_amdgcn_raw_buffer_load_i8x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            tmp.AsType<int8x4_t>()(Number<1>{}) =
+                llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource,
+                                                 src_thread_addr_offset,
+                                                 src_wave_addr_offset + 4 * sizeof(int8_t),
+                                                 0);
+
+            tmp.AsType<int8x4_t>()(Number<2>{}) =
+                llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource,
+                                                 src_thread_addr_offset,
+                                                 src_wave_addr_offset + 8 * sizeof(int8_t),
+                                                 0);
+
+            tmp.AsType<int8x4_t>()(Number<3>{}) =
+                llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource,
+                                                 src_thread_addr_offset,
+                                                 src_wave_addr_offset + 12 * sizeof(int8_t),
+                                                 0);
+
+            return tmp.AsType<int8x16_t>()(Number<0>{});
+#else
+            int32x4_t tmp = llvm_amdgcn_raw_buffer_load_i32x4(
+                src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
+
+            return bit_cast<int8x16_t>(tmp);
+#endif
+        }
+    }
+}
+
+template <typename T, index_t N>
+__device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src_thread_data,
+                                      int32x4_t dst_wave_buffer_resource,
+                                      index_t dst_thread_addr_offset,
+                                      index_t dst_wave_addr_offset)
+{
+    static_assert(
+        (is_same<T, double>::value && (N == 1 || N == 2)) ||
+            (is_same<T, float>::value && (N == 1 || N == 2 || N == 4)) ||
+            (is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, bhalf_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4)) ||
+            (is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
+        "wrong! not implemented");
+
+    if constexpr(is_same<T, double>::value)
+    {
+        // use fp32 store to mimic fp64 store
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_store_fp32x2(bit_cast<float2_t>(src_thread_data),
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                0);
+        }
+        else if constexpr(N == 2)
+        {
+            llvm_amdgcn_raw_buffer_store_fp32x4(bit_cast<float4_t>(src_thread_data),
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                0);
+        }
+    }
+    else if constexpr(is_same<T, float>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_store_fp32(src_thread_data,
+                                              dst_wave_buffer_resource,
+                                              dst_thread_addr_offset,
+                                              dst_wave_addr_offset,
+                                              0);
+        }
+        else if constexpr(N == 2)
+        {
+            llvm_amdgcn_raw_buffer_store_fp32x2(src_thread_data,
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                0);
+        }
+        else if constexpr(N == 4)
+        {
+            llvm_amdgcn_raw_buffer_store_fp32x4(src_thread_data,
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                0);
+        }
+    }
+    else if constexpr(is_same<T, half_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_store_fp16(src_thread_data,
+                                              dst_wave_buffer_resource,
+                                              dst_thread_addr_offset,
+                                              dst_wave_addr_offset,
+                                              0);
+        }
+        else if constexpr(N == 2)
+        {
+            llvm_amdgcn_raw_buffer_store_fp16x2(src_thread_data,
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                0);
+        }
+        else if constexpr(N == 4)
+        {
+            llvm_amdgcn_raw_buffer_store_fp16x4(src_thread_data,
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                0);
+        }
+        else if constexpr(N == 8)
+        {
+#if 0
+            vector_type<half_t, 8> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_store_fp16x4(tmp.AsType<half4_t>()[Number<0>{}],
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                0);
+
+            llvm_amdgcn_raw_buffer_store_fp16x4(tmp.AsType<half4_t>()[Number<1>{}],
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset + 4 * sizeof(half_t),
+                                                0);
+#else
+            llvm_amdgcn_raw_buffer_store_fp32x4(bit_cast<float4_t>(src_thread_data),
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                0);
+#endif
+        }
+    }
+    else if constexpr(is_same<T, bhalf_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_store_i16(src_thread_data,
+                                             dst_wave_buffer_resource,
+                                             dst_thread_addr_offset,
+                                             dst_wave_addr_offset,
+                                             0);
+        }
+        else if constexpr(N == 2)
+        {
+            llvm_amdgcn_raw_buffer_store_i16x2(src_thread_data,
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               0);
+        }
+        else if constexpr(N == 4)
+        {
+            llvm_amdgcn_raw_buffer_store_i16x4(src_thread_data,
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               0);
+        }
+        else if constexpr(N == 8)
+        {
+            vector_type<bhalf_t, 8> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_store_i16x4(tmp.AsType<bhalf4_t>()[Number<0>{}],
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               0);
+
+            llvm_amdgcn_raw_buffer_store_i16x4(tmp.AsType<bhalf4_t>()[Number<1>{}],
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset + 4 * sizeof(bhalf_t),
+                                               0);
+        }
+    }
+    else if constexpr(is_same<T, int32_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_store_i32(src_thread_data,
+                                             dst_wave_buffer_resource,
+                                             dst_thread_addr_offset,
+                                             dst_wave_addr_offset,
+                                             0);
+        }
+        else if constexpr(N == 2)
+        {
+            llvm_amdgcn_raw_buffer_store_i32x2(src_thread_data,
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               0);
+        }
+        else if constexpr(N == 4)
+        {
+            llvm_amdgcn_raw_buffer_store_i32x4(src_thread_data,
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               0);
+        }
+    }
+    else if constexpr(is_same<T, int8_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_store_i8(src_thread_data,
+                                            dst_wave_buffer_resource,
+                                            dst_thread_addr_offset,
+                                            dst_wave_addr_offset,
+                                            0);
+        }
+        else if constexpr(N == 2)
+        {
+#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
+            llvm_amdgcn_raw_buffer_store_i8x2(src_thread_data,
+                                              dst_wave_buffer_resource,
+                                              dst_thread_addr_offset,
+                                              dst_wave_addr_offset,
+                                              0);
+#else
+            llvm_amdgcn_raw_buffer_store_i16(bit_cast<int16_t>(src_thread_data),
+                                             dst_wave_buffer_resource,
+                                             dst_thread_addr_offset,
+                                             dst_wave_addr_offset,
+                                             0);
+#endif
+        }
+        else if constexpr(N == 4)
+        {
+#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
+            llvm_amdgcn_raw_buffer_store_i8x4(src_thread_data,
+                                              dst_wave_buffer_resource,
+                                              dst_thread_addr_offset,
+                                              dst_wave_addr_offset,
+                                              0);
+#else
+            llvm_amdgcn_raw_buffer_store_i32(bit_cast<int32_t>(src_thread_data),
+                                             dst_wave_buffer_resource,
+                                             dst_thread_addr_offset,
+                                             dst_wave_addr_offset,
+                                             0);
+#endif
+        }
+        else if constexpr(N == 8)
+        {
+            llvm_amdgcn_raw_buffer_store_i32x2(bit_cast<int32x2_t>(src_thread_data),
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               0);
+        }
+        else if constexpr(N == 16)
+        {
+            llvm_amdgcn_raw_buffer_store_i32x4(bit_cast<int32x4_t>(src_thread_data),
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               0);
+        }
+    }
+}
+
+template <typename T, index_t N>
+__device__ void amd_buffer_atomic_add_impl(const typename vector_type<T, N>::type src_thread_data,
+                                           int32x4_t dst_wave_buffer_resource,
+                                           index_t dst_thread_addr_offset,
+                                           index_t dst_wave_addr_offset)
+{
+    static_assert((is_same<T, float>::value && (N == 1 || N == 2 || N == 4)) ||
+                      (is_same<T, half_t>::value && (N == 2 || N == 4 || N == 8)) ||
+                      (is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4)),
+                  "wrong! not implemented");
+
+    if constexpr(is_same<T, float>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(src_thread_data,
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset,
+                                                   0);
+        }
+        else if constexpr(N == 2)
+        {
+            vector_type<float, 2> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(tmp.AsType<float>()[Number<0>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset,
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(tmp.AsType<float>()[Number<1>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + sizeof(float),
+                                                   0);
+        }
+        else if constexpr(N == 4)
+        {
+            vector_type<float, 4> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(tmp.AsType<float>()[Number<0>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset,
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(tmp.AsType<float>()[Number<1>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + sizeof(float),
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(tmp.AsType<float>()[Number<2>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + 2 * sizeof(float),
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(tmp.AsType<float>()[Number<3>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + 3 * sizeof(float),
+                                                   0);
+        }
+    }
+    else if constexpr(is_same<T, half_t>::value)
+    {
+        if constexpr(N == 2)
+        {
+            llvm_amdgcn_raw_buffer_atomic_add_fp16x2(src_thread_data,
+                                                     dst_wave_buffer_resource,
+                                                     dst_thread_addr_offset,
+                                                     dst_wave_addr_offset,
+                                                     0);
+        }
+        else if constexpr(N == 4)
+        {
+            vector_type<half_t, 4> tmp{src_thread_data};
+
+            static_for<0, 2, 1>{}([&](auto i) {
+                llvm_amdgcn_raw_buffer_atomic_add_fp16x2(tmp.AsType<half2_t>()[i],
+                                                         dst_wave_buffer_resource,
+                                                         dst_thread_addr_offset,
+                                                         dst_wave_addr_offset + i * sizeof(half2_t),
+                                                         0);
+            });
+        }
+        else if constexpr(N == 8)
+        {
+            vector_type<half_t, 8> tmp{src_thread_data};
+
+            static_for<0, 4, 1>{}([&](auto i) {
+                llvm_amdgcn_raw_buffer_atomic_add_fp16x2(tmp.AsType<half2_t>()[i],
+                                                         dst_wave_buffer_resource,
+                                                         dst_thread_addr_offset,
+                                                         dst_wave_addr_offset + i * sizeof(half2_t),
+                                                         0);
+            });
+        }
+    }
+    else if constexpr(is_same<T, int32_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_atomic_add_i32(src_thread_data,
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset,
+                                                  0);
+        }
+        else if constexpr(N == 2)
+        {
+            vector_type<int32_t, 2> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_atomic_add_i32(tmp.AsType<int32_t>()[Number<0>{}],
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset,
+                                                  0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_i32(tmp.AsType<int32_t>()[Number<1>{}],
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset + sizeof(int32_t),
+                                                  0);
+        }
+        else if constexpr(N == 4)
+        {
+            vector_type<int32_t, 4> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_atomic_add_i32(tmp.AsType<int32_t>()[Number<0>{}],
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset,
+                                                  0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_i32(tmp.AsType<int32_t>()[Number<1>{}],
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset + sizeof(int32_t),
+                                                  0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_i32(tmp.AsType<int32_t>()[Number<2>{}],
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset + 2 * sizeof(int32_t),
+                                                  0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_i32(tmp.AsType<int32_t>()[Number<3>{}],
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset + 3 * sizeof(int32_t),
+                                                  0);
+        }
+    }
+}
+
+template <typename T, index_t N>
+__device__ void amd_buffer_atomic_max_impl(const typename vector_type<T, N>::type src_thread_data,
+                                           int32x4_t dst_wave_buffer_resource,
+                                           index_t dst_thread_addr_offset,
+                                           index_t dst_wave_addr_offset)
+{
+    static_assert((is_same<T, double>::value && (N == 1 || N == 2 || N == 4)),
+                  "wrong! not implemented");
+    if constexpr(is_same<T, double>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(src_thread_data,
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset,
+                                                   0);
+        }
+        else if constexpr(N == 2)
+        {
+            vector_type<double, 2> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType<double>()[Number<0>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset,
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType<double>()[Number<1>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + sizeof(double),
+                                                   0);
+        }
+        else if constexpr(N == 4)
+        {
+            vector_type<double, 4> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType<double>()[Number<0>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset,
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType<double>()[Number<1>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + sizeof(double),
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType<double>()[Number<2>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + 2 * sizeof(double),
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType<double>()[Number<3>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + 3 * sizeof(double),
+                                                   0);
+        }
+    }
+}
+
+// buffer_load requires:
+//   1) p_src_wave must point to global memory space
+//   2) p_src_wave must be a wavewise pointer.
+// It is user's responsibility to make sure that is true.
+template <typename T, index_t N>
+__device__ typename vector_type_maker<T, N>::type::type
+amd_buffer_load_invalid_element_return_zero(const T* p_src_wave,
+                                            index_t src_thread_element_offset,
+                                            bool src_thread_element_valid,
+                                            index_t src_element_space_size)
+{
+    const int32x4_t src_wave_buffer_resource =
+        make_wave_buffer_resource(p_src_wave, src_element_space_size);
+
+    index_t src_thread_addr_offset = src_thread_element_offset * sizeof(T);
+
+    using vector_t = typename vector_type_maker<T, N>::type::type;
+    using scalar_t = typename scalar_type<vector_t>::type;
+
+    constexpr index_t vector_size = scalar_type<vector_t>::vector_size;
+
+#if CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK
+    uint32_t src_addr_shift = src_thread_element_valid ? 0 : 0x7fffffff;
+
+    return amd_buffer_load_impl<scalar_t, vector_size>(
+        src_wave_buffer_resource, src_addr_shift + src_thread_addr_offset, 0);
+#else
+    vector_t tmp = amd_buffer_load_impl<scalar_t, vector_size>(
+        src_wave_buffer_resource, src_thread_addr_offset, 0);
+
+    return src_thread_element_valid ? tmp : vector_t(0);
+#endif
+}
+
+// buffer_load requires:
+//   1) p_src_wave must point to global memory space
+//   2) p_src_wave must be a wavewise pointer.
+// It is user's responsibility to make sure that is true.
+template <typename T, index_t N>
+__device__ typename vector_type_maker<T, N>::type::type
+amd_buffer_load_invalid_element_return_customized_value(const T* p_src_wave,
+                                                        index_t src_thread_element_offset,
+                                                        bool src_thread_element_valid,
+                                                        index_t src_element_space_size,
+                                                        T customized_value)
+{
+    const int32x4_t src_wave_buffer_resource =
+        make_wave_buffer_resource(p_src_wave, src_element_space_size);
+
+    index_t src_thread_addr_offset = src_thread_element_offset * sizeof(T);
+
+    using vector_t = typename vector_type_maker<T, N>::type::type;
+    using scalar_t = typename scalar_type<vector_t>::type;
+
+    constexpr index_t vector_size = scalar_type<vector_t>::vector_size;
+
+    vector_t tmp = amd_buffer_load_impl<scalar_t, vector_size>(
+        src_wave_buffer_resource, src_thread_addr_offset, 0);
+
+    return src_thread_element_valid ? tmp : vector_t(customized_value);
+}
+
+// buffer_store requires:
+//   1) p_dst_wave must point to global memory
+//   2) p_dst_wave must be a wavewise pointer.
+// It is user's responsibility to make sure that is true.
+template <typename T, index_t N>
+__device__ void amd_buffer_store(const typename vector_type_maker<T, N>::type::type src_thread_data,
+                                 T* p_dst_wave,
+                                 const index_t dst_thread_element_offset,
+                                 const bool dst_thread_element_valid,
+                                 const index_t dst_element_space_size)
+{
+    const int32x4_t dst_wave_buffer_resource =
+        make_wave_buffer_resource(p_dst_wave, dst_element_space_size);
+
+    index_t dst_thread_addr_offset = dst_thread_element_offset * sizeof(T);
+
+    using vector_t                = typename vector_type_maker<T, N>::type::type;
+    using scalar_t                = typename scalar_type<vector_t>::type;
+    constexpr index_t vector_size = scalar_type<vector_t>::vector_size;
+
+#if CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK
+    uint32_t dst_addr_shift = dst_thread_element_valid ? 0 : 0x7fffffff;
+
+    amd_buffer_store_impl<scalar_t, vector_size>(
+        src_thread_data, dst_wave_buffer_resource, dst_addr_shift + dst_thread_addr_offset, 0);
+#else
+    if(dst_thread_element_valid)
+    {
+        amd_buffer_store_impl<scalar_t, vector_size>(
+            src_thread_data, dst_wave_buffer_resource, dst_thread_addr_offset, 0);
+    }
+#endif
+}
+
+// buffer_atomic_add requires:
+//   1) p_dst_wave must point to global memory
+//   2) p_dst_wave must be a wavewise pointer.
+// It is user's responsibility to make sure that is true.
+template <typename T, index_t N>
+__device__ void
+amd_buffer_atomic_add(const typename vector_type_maker<T, N>::type::type src_thread_data,
+                      T* p_dst_wave,
+                      const index_t dst_thread_element_offset,
+                      const bool dst_thread_element_valid,
+                      const index_t dst_element_space_size)
+{
+    const int32x4_t dst_wave_buffer_resource =
+        make_wave_buffer_resource(p_dst_wave, dst_element_space_size);
+
+    index_t dst_thread_addr_offset = dst_thread_element_offset * sizeof(T);
+
+    using vector_t                = typename vector_type_maker<T, N>::type::type;
+    using scalar_t                = typename scalar_type<vector_t>::type;
+    constexpr index_t vector_size = scalar_type<vector_t>::vector_size;
+
+#if CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_ADD_OOB_CHECK_OFFSET_TRICK
+    uint32_t dst_addr_shift = dst_thread_element_valid ? 0 : 0x7fffffff;
+
+    amd_buffer_atomic_add_impl<scalar_t, vector_size>(
+        src_thread_data, dst_wave_buffer_resource, dst_addr_shift + dst_thread_addr_offset, 0);
+#else
+    if(dst_thread_element_valid)
+    {
+        amd_buffer_atomic_add_impl<scalar_t, vector_size>(
+            src_thread_data, dst_wave_buffer_resource, dst_thread_addr_offset, 0);
+    }
+#endif
+}
+
+// buffer_atomic_max requires:
+//   1) p_dst_wave must point to global memory
+//   2) p_dst_wave must be a wavewise pointer.
+// It is user's responsibility to make sure that is true.
+template <typename T, index_t N>
+__device__ void
+amd_buffer_atomic_max(const typename vector_type_maker<T, N>::type::type src_thread_data,
+                      T* p_dst_wave,
+                      const index_t dst_thread_element_offset,
+                      const bool dst_thread_element_valid,
+                      const index_t dst_element_space_size)
+{
+    const int32x4_t dst_wave_buffer_resource =
+        make_wave_buffer_resource(p_dst_wave, dst_element_space_size);
+
+    index_t dst_thread_addr_offset = dst_thread_element_offset * sizeof(T);
+
+    using vector_t                = typename vector_type_maker<T, N>::type::type;
+    using scalar_t                = typename scalar_type<vector_t>::type;
+    constexpr index_t vector_size = scalar_type<vector_t>::vector_size;
+
+#if CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_MAX_OOB_CHECK_OFFSET_TRICK
+    uint32_t dst_addr_shift = dst_thread_element_valid ? 0 : 0x7fffffff;
+
+    amd_buffer_atomic_max_impl<scalar_t, vector_size>(
+        src_thread_data, dst_wave_buffer_resource, dst_addr_shift + dst_thread_addr_offset, 0);
+#else
+    if(dst_thread_element_valid)
+    {
+        amd_buffer_atomic_max_impl<scalar_t, vector_size>(
+            src_thread_data, dst_wave_buffer_resource, dst_thread_addr_offset, 0);
+    }
+#endif
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/amd_inline_asm.hpp b/3rdparty/composable_kernel/include/ck/utility/amd_inline_asm.hpp
new file mode 100644
index 0000000..82bf2a5
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/amd_inline_asm.hpp
@@ -0,0 +1,359 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_AMD_INLINE_ASM_HPP
+#define CK_AMD_INLINE_ASM_HPP
+
+#include "data_type.hpp"
+#include "c_style_pointer_cast.hpp"
+
+// TODO: deprecate all amd_assembly_outer_product_xxx
+
+namespace ck {
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+__device__ void amd_assembly_outer_product_1x2(float a, float b0, float b1, float& c0, float& c1)
+{
+    asm volatile("\n \
+            v_fmac_f32 %0, %2, %3 \n \
+            v_fmac_f32 %1, %2, %4 \n \
+            "
+                 : "=v"(c0), "=v"(c1)
+                 : "v"(a), "v"(b0), "v"(b1), "0"(c0), "1"(c1));
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+// c2 += inner_product(a, b2)
+// c3 += inner_product(a, b3)
+__device__ void amd_assembly_outer_product_1x4(
+    float a, float b0, float b1, float b2, float b3, float& c0, float& c1, float& c2, float& c3)
+{
+    asm volatile("\n \
+            v_fmac_f32 %0, %4, %5 \n \
+            v_fmac_f32 %1, %4, %6 \n \
+            v_fmac_f32 %2, %4, %7 \n \
+            v_fmac_f32 %3, %4, %8 \n \
+            "
+                 : "=v"(c0), "=v"(c1), "=v"(c2), "=v"(c3)
+                 : "v"(a), "v"(b0), "v"(b1), "v"(b2), "v"(b3), "0"(c0), "1"(c1), "2"(c2), "3"(c3));
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+__device__ void
+amd_assembly_outer_product_1x2(half2_t a, half2_t b0, half2_t b1, float& c0, float& c1)
+{
+    asm volatile("\n \
+            v_dot2_f32_f16 %0, %2, %3, %0\n \
+            v_dot2_f32_f16 %1, %2, %4, %1\n \
+            "
+                 : "=v"(c0), "=v"(c1)
+                 : "v"(a), "v"(b0), "v"(b1), "0"(c0), "1"(c1));
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+__device__ void
+amd_assembly_outer_product_1x2(half4_t a, half4_t b0, half4_t b1, float& c0, float& c1)
+{
+    // TODO remove pointer casting
+    const half2_t* p_a_half2  = c_style_pointer_cast<const half2_t*>(&a);
+    const half2_t* p_b0_half2 = c_style_pointer_cast<const half2_t*>(&b0);
+    const half2_t* p_b1_half2 = c_style_pointer_cast<const half2_t*>(&b1);
+
+    // do dot2 two times
+    asm volatile("\n \
+            v_dot2_f32_f16 %0, %2, %4, %0\n \
+            v_dot2_f32_f16 %1, %2, %6, %1\n \
+            v_dot2_f32_f16 %0, %3, %5, %0\n \
+            v_dot2_f32_f16 %1, %3, %7, %1\n \
+            "
+                 : "=v"(c0), "=v"(c1)
+                 : "v"(p_a_half2[0]),
+                   "v"(p_a_half2[1]),
+                   "v"(p_b0_half2[0]),
+                   "v"(p_b0_half2[1]),
+                   "v"(p_b1_half2[0]),
+                   "v"(p_b1_half2[1]),
+                   "0"(c0),
+                   "1"(c1));
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+// c2 += inner_product(a, b2)
+// c3 += inner_product(a, b3)
+__device__ void amd_assembly_outer_product_1x4(half2_t a,
+                                               half2_t b0,
+                                               half2_t b1,
+                                               half2_t b2,
+                                               half2_t b3,
+                                               float& c0,
+                                               float& c1,
+                                               float& c2,
+                                               float& c3)
+{
+    asm volatile("\n \
+            v_dot2_f32_f16 %0, %4, %5, %0\n \
+            v_dot2_f32_f16 %1, %4, %6, %1\n \
+            v_dot2_f32_f16 %2, %4, %7, %2\n \
+            v_dot2_f32_f16 %3, %4, %8, %3\n \
+            "
+                 : "=v"(c0), "=v"(c1), "=v"(c2), "=v"(c3)
+                 : "v"(a), "v"(b0), "v"(b1), "v"(b2), "v"(b3), "0"(c0), "1"(c1), "2"(c2), "3"(c3));
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+// c2 += inner_product(a, b2)
+// c3 += inner_product(a, b3)
+__device__ void amd_assembly_outer_product_1x4(half4_t a,
+                                               half4_t b0,
+                                               half4_t b1,
+                                               half4_t b2,
+                                               half4_t b3,
+                                               float& c0,
+                                               float& c1,
+                                               float& c2,
+                                               float& c3)
+{
+    // TODO remove pointer casting
+    const half2_t* p_a_half2  = c_style_pointer_cast<const half2_t*>(&a);
+    const half2_t* p_b0_half2 = c_style_pointer_cast<const half2_t*>(&b0);
+    const half2_t* p_b1_half2 = c_style_pointer_cast<const half2_t*>(&b1);
+    const half2_t* p_b2_half2 = c_style_pointer_cast<const half2_t*>(&b2);
+    const half2_t* p_b3_half2 = c_style_pointer_cast<const half2_t*>(&b3);
+
+    // do dot2 two times
+    asm volatile("\n \
+            v_dot2_f32_f16 %0, %4, %6,  %0\n \
+            v_dot2_f32_f16 %1, %4, %8,  %1\n \
+            v_dot2_f32_f16 %2, %4, %10, %2\n \
+            v_dot2_f32_f16 %3, %4, %12, %3\n \
+            v_dot2_f32_f16 %0, %5, %7,  %0\n \
+            v_dot2_f32_f16 %1, %5, %9,  %1\n \
+            v_dot2_f32_f16 %2, %5, %11, %2\n \
+            v_dot2_f32_f16 %3, %5, %13, %3\n \
+            "
+                 : "=v"(c0), "=v"(c1), "=v"(c2), "=v"(c3)
+                 : "v"(p_a_half2[0]),
+                   "v"(p_a_half2[1]),
+                   "v"(p_b0_half2[0]),
+                   "v"(p_b0_half2[1]),
+                   "v"(p_b1_half2[0]),
+                   "v"(p_b1_half2[1]),
+                   "v"(p_b2_half2[0]),
+                   "v"(p_b2_half2[1]),
+                   "v"(p_b3_half2[0]),
+                   "v"(p_b3_half2[1]),
+                   "0"(c0),
+                   "1"(c1),
+                   "2"(c2),
+                   "3"(c3));
+}
+
+__device__ void amd_assembly_outer_product_1x4(half8_t a,
+                                               half8_t b0,
+                                               half8_t b1,
+                                               half8_t b2,
+                                               half8_t b3,
+                                               float& c0,
+                                               float& c1,
+                                               float& c2,
+                                               float& c3)
+{
+
+    // TODO remove pointer casting
+    const half4_t* p_a_half4  = c_style_pointer_cast<const half4_t*>(&a);
+    const half4_t* p_b0_half4 = c_style_pointer_cast<const half4_t*>(&b0);
+    const half4_t* p_b1_half4 = c_style_pointer_cast<const half4_t*>(&b1);
+    const half4_t* p_b2_half4 = c_style_pointer_cast<const half4_t*>(&b2);
+    const half4_t* p_b3_half4 = c_style_pointer_cast<const half4_t*>(&b3);
+
+    amd_assembly_outer_product_1x4(
+        p_a_half4[0], p_b0_half4[0], p_b1_half4[0], p_b2_half4[0], p_b3_half4[0], c0, c1, c2, c3);
+
+    amd_assembly_outer_product_1x4(
+        p_a_half4[1], p_b0_half4[1], p_b1_half4[1], p_b2_half4[1], p_b3_half4[1], c0, c1, c2, c3);
+}
+
+__device__ void amd_assembly_outer_product_1x4(half16_t a,
+                                               half16_t b0,
+                                               half16_t b1,
+                                               half16_t b2,
+                                               half16_t b3,
+                                               float& c0,
+                                               float& c1,
+                                               float& c2,
+                                               float& c3)
+{
+    // TODO remove pointer casting
+    const half8_t* p_a_half8  = c_style_pointer_cast<const half8_t*>(&a);
+    const half8_t* p_b0_half8 = c_style_pointer_cast<const half8_t*>(&b0);
+    const half8_t* p_b1_half8 = c_style_pointer_cast<const half8_t*>(&b1);
+    const half8_t* p_b2_half8 = c_style_pointer_cast<const half8_t*>(&b2);
+    const half8_t* p_b3_half8 = c_style_pointer_cast<const half8_t*>(&b3);
+
+    amd_assembly_outer_product_1x4(
+        p_a_half8[0], p_b0_half8[0], p_b1_half8[0], p_b2_half8[0], p_b3_half8[0], c0, c1, c2, c3);
+
+    amd_assembly_outer_product_1x4(
+        p_a_half8[1], p_b0_half8[1], p_b1_half8[1], p_b2_half8[1], p_b3_half8[1], c0, c1, c2, c3);
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+__device__ void
+amd_assembly_outer_product_1x2(int8x4_t a, int8x4_t b0, int8x4_t b1, int32_t& c0, int32_t& c1)
+{
+#if 1
+    asm volatile("\n \
+            v_dot4_i32_i8 %0, %2, %3, %0\n \
+            v_dot4_i32_i8 %1, %2, %4, %1\n \
+            "
+                 : "=v"(c0), "=v"(c1)
+                 : "v"(bit_cast<int32_t>(a)),
+                   "v"(bit_cast<int32_t>(b0)),
+                   "v"(bit_cast<int32_t>(b1)),
+                   "0"(c0),
+                   "1"(c1));
+#else
+    c0 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b0), c0, false);
+    c1 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b1), c1, false);
+#endif
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+// c2 += inner_product(a, b2)
+// c3 += inner_product(a, b3)
+__device__ void amd_assembly_outer_product_1x4(int8x4_t a,
+                                               int8x4_t b0,
+                                               int8x4_t b1,
+                                               int8x4_t b2,
+                                               int8x4_t b3,
+                                               int32_t& c0,
+                                               int32_t& c1,
+                                               int32_t& c2,
+                                               int32_t& c3)
+{
+#if 1
+    asm volatile("\n \
+            v_dot4_i32_i8 %0, %4, %5, %0\n \
+            v_dot4_i32_i8 %1, %4, %6, %1\n \
+            v_dot4_i32_i8 %2, %4, %7, %2\n \
+            v_dot4_i32_i8 %3, %4, %8, %3\n \
+            "
+                 : "=v"(c0), "=v"(c1), "=v"(c2), "=v"(c3)
+                 : "v"(bit_cast<int32_t>(a)),
+                   "v"(bit_cast<int32_t>(b0)),
+                   "v"(bit_cast<int32_t>(b1)),
+                   "v"(bit_cast<int32_t>(b2)),
+                   "v"(bit_cast<int32_t>(b3)),
+                   "0"(c0),
+                   "1"(c1),
+                   "2"(c2),
+                   "3"(c3));
+#else
+    c0 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b0), c0, false);
+    c1 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b1), c1, false);
+    c2 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b2), c2, false);
+    c3 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b3), c3, false);
+#endif
+}
+
+__device__ void amd_assembly_outer_product_1x4(int8x8_t a,
+                                               int8x8_t b0,
+                                               int8x8_t b1,
+                                               int8x8_t b2,
+                                               int8x8_t b3,
+                                               int32_t& c0,
+                                               int32_t& c1,
+                                               int32_t& c2,
+                                               int32_t& c3)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    amd_assembly_outer_product_1x4(vector_type<int8_t, 8>{a}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 8>{b0}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 8>{b1}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 8>{b2}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 8>{b3}.AsType<int8x4_t>()[I0],
+                                   c0,
+                                   c1,
+                                   c2,
+                                   c3);
+
+    amd_assembly_outer_product_1x4(vector_type<int8_t, 8>{a}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 8>{b0}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 8>{b1}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 8>{b2}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 8>{b3}.AsType<int8x4_t>()[I1],
+                                   c0,
+                                   c1,
+                                   c2,
+                                   c3);
+}
+
+__device__ void amd_assembly_outer_product_1x4(int8x16_t a,
+                                               int8x16_t b0,
+                                               int8x16_t b1,
+                                               int8x16_t b2,
+                                               int8x16_t b3,
+                                               int32_t& c0,
+                                               int32_t& c1,
+                                               int32_t& c2,
+                                               int32_t& c3)
+
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    amd_assembly_outer_product_1x4(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 16>{b0}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 16>{b1}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 16>{b2}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 16>{b3}.AsType<int8x4_t>()[I0],
+                                   c0,
+                                   c1,
+                                   c2,
+                                   c3);
+
+    amd_assembly_outer_product_1x4(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 16>{b0}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 16>{b1}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 16>{b2}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 16>{b3}.AsType<int8x4_t>()[I1],
+                                   c0,
+                                   c1,
+                                   c2,
+                                   c3);
+
+    amd_assembly_outer_product_1x4(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I2],
+                                   vector_type<int8_t, 16>{b0}.AsType<int8x4_t>()[I2],
+                                   vector_type<int8_t, 16>{b1}.AsType<int8x4_t>()[I2],
+                                   vector_type<int8_t, 16>{b2}.AsType<int8x4_t>()[I2],
+                                   vector_type<int8_t, 16>{b3}.AsType<int8x4_t>()[I2],
+                                   c0,
+                                   c1,
+                                   c2,
+                                   c3);
+
+    amd_assembly_outer_product_1x4(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I3],
+                                   vector_type<int8_t, 16>{b0}.AsType<int8x4_t>()[I3],
+                                   vector_type<int8_t, 16>{b1}.AsType<int8x4_t>()[I3],
+                                   vector_type<int8_t, 16>{b2}.AsType<int8x4_t>()[I3],
+                                   vector_type<int8_t, 16>{b3}.AsType<int8x4_t>()[I3],
+                                   c0,
+                                   c1,
+                                   c2,
+                                   c3);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/amd_llvm_intrinsic.hpp b/3rdparty/composable_kernel/include/ck/utility/amd_llvm_intrinsic.hpp
new file mode 100644
index 0000000..01e77d7
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/amd_llvm_intrinsic.hpp
@@ -0,0 +1,14 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_AMD_LLVM_INTRINSIC_HPP
+#define CK_AMD_LLVM_INTRINSIC_HPP
+
+#include "data_type.hpp"
+
+namespace ck {
+
+__device__ int32_t llvm_amdgcn_readfirstlane_i32(int32_t i) __asm("llvm.amdgcn.readfirstlane");
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/amd_wmma.hpp b/3rdparty/composable_kernel/include/ck/utility/amd_wmma.hpp
new file mode 100644
index 0000000..752876a
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/amd_wmma.hpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_AMD_WMMA_HPP
+#define CK_AMD_WMMA_HPP
+
+#include "data_type.hpp"
+// TODO: Add arch limitation
+namespace ck {
+
+// wave32 only
+// src: fp16, dst: fp32
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_wmma_f32_16x16x16_f16_w32;
+
+template <>
+struct intrin_wmma_f32_16x16x16_f16_w32<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const half16_t& reg_a, const half16_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float8_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(
+            reg_a, reg_b, reg_c.template AsType<float8_t>()[Number<0>{}]);
+    }
+};
+
+// src: bf16, dst: fp32
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_wmma_f32_16x16x16_bf16_w32;
+
+template <>
+struct intrin_wmma_f32_16x16x16_bf16_w32<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf16_t& reg_a, const bhalf16_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float8_t>()(Number<0>{}) =
+            __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32(
+                reg_a, reg_b, reg_c.template AsType<float8_t>()[Number<0>{}]);
+    }
+};
+
+// src: fp16, dst: fp16
+template <index_t MPerWave, index_t NPerWave, index_t Opsel>
+struct intrin_wmma_f16_16x16x16_f16_w32;
+
+template <index_t Opsel>
+struct intrin_wmma_f16_16x16x16_f16_w32<16, 16, Opsel>
+{
+    template <class FloatC>
+    __device__ static void Run(const half16_t& reg_a, const half16_t& reg_b, FloatC& reg_c)
+    {
+        // opsel usage
+        // false: D0.[0:15] = result
+        // true : D0.[16:31]= result
+        reg_c.template AsType<half16_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f16_16x16x16_f16_w32(
+            reg_a, reg_b, reg_c.template AsType<half16_t>()[Number<0>{}], Opsel);
+    }
+};
+
+// src: bf16, dst: bf16
+template <index_t MPerWave, index_t NPerWave, index_t Opsel>
+struct intrin_wmma_bf16_16x16x16_bf16_w32;
+
+template <index_t Opsel>
+struct intrin_wmma_bf16_16x16x16_bf16_w32<16, 16, Opsel>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf16_t& reg_a, const bhalf16_t& reg_b, FloatC& reg_c)
+    {
+        // opsel usage
+        // false: D0.[0:15] = result
+        // true : D0.[16:31]= result
+        reg_c.template AsType<bhalf16_t>()(Number<0>{}) =
+            __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32(
+                reg_a, reg_b, reg_c.template AsType<bhalf16_t>()[Number<0>{}], Opsel);
+    }
+};
+
+// src: iu8, dst: i32
+template <index_t MPerWave, index_t NPerWave, bool neg_a, bool neg_b, bool clamp>
+struct intrin_wmma_i32_16x16x16_iu8_w32;
+
+template <bool neg_a, bool neg_b, bool clamp>
+struct intrin_wmma_i32_16x16x16_iu8_w32<16, 16, neg_a, neg_b, clamp>
+{
+    template <class FloatC>
+    __device__ static void Run(const int8x16_t& reg_a, const int8x16_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<int32x8_t>()(Number<0>{}) =
+            __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
+                neg_a,
+                bit_cast<int32x4_t>(reg_a),
+                neg_b,
+                bit_cast<int32x4_t>(reg_b),
+                reg_c.template AsType<int32x8_t>()[Number<0>{}],
+                clamp);
+    }
+};
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/amd_xdlops.hpp b/3rdparty/composable_kernel/include/ck/utility/amd_xdlops.hpp
new file mode 100644
index 0000000..b4be0cb
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/amd_xdlops.hpp
@@ -0,0 +1,320 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_AMD_XDLOPS_HPP
+#define CK_AMD_XDLOPS_HPP
+
+#include "data_type.hpp"
+
+namespace ck {
+
+// fp32
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x1f32;
+
+template <>
+struct intrin_mfma_f32_32x32x1f32<64, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float32_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x1f32(
+            reg_a, reg_b, reg_c.template AsType<float32_t>()[Number<0>{}], 1, 0, 0);
+        reg_c.template AsType<float32_t>()(Number<1>{}) = __builtin_amdgcn_mfma_f32_32x32x1f32(
+            reg_a, reg_b, reg_c.template AsType<float32_t>()[Number<1>{}], 1, 1, 0);
+    }
+};
+
+template <>
+struct intrin_mfma_f32_32x32x1f32<32, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float32_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x1f32(
+            reg_a, reg_b, reg_c.template AsType<float32_t>()[Number<0>{}], 1, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x2f32;
+
+template <>
+struct intrin_mfma_f32_32x32x2f32<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x2f32(
+            reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x4f32;
+
+template <>
+struct intrin_mfma_f32_16x16x4f32<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x4f32(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x1f32;
+
+template <>
+struct intrin_mfma_f32_16x16x1f32<16, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x1f32(
+            reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 2, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_4x4x1f32;
+
+template <>
+struct intrin_mfma_f32_4x4x1f32<4, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_4x4x1f32(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 4, 0, 0);
+    }
+};
+
+template <>
+struct intrin_mfma_f32_4x4x1f32<8, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_4x4x1f32(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 4, 0, 0);
+        reg_c.template AsType<float4_t>()(Number<1>{}) = __builtin_amdgcn_mfma_f32_4x4x1f32(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<1>{}], 4, 1, 0);
+    }
+};
+
+// fp16
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x4f16;
+
+template <>
+struct intrin_mfma_f32_32x32x4f16<64, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float32_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x4f16(
+            reg_a, reg_b, reg_c.template AsType<float32_t>()[Number<0>{}], 1, 0, 0);
+        reg_c.template AsType<float32_t>()(Number<1>{}) = __builtin_amdgcn_mfma_f32_32x32x4f16(
+            reg_a, reg_b, reg_c.template AsType<float32_t>()[Number<1>{}], 1, 1, 0);
+    }
+};
+
+template <>
+struct intrin_mfma_f32_32x32x4f16<32, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float32_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x4f16(
+            reg_a, reg_b, reg_c.template AsType<float32_t>()[Number<0>{}], 1, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x8f16;
+
+template <>
+struct intrin_mfma_f32_32x32x8f16<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x8f16(
+            reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x16f16;
+
+template <>
+struct intrin_mfma_f32_16x16x16f16<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x16f16(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x4f16;
+
+template <>
+struct intrin_mfma_f32_16x16x4f16<16, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x4f16(
+            reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 2, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_4x4x4f16;
+
+template <>
+struct intrin_mfma_f32_4x4x4f16<4, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_4x4x4f16(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 4, 0, 0);
+    }
+};
+
+template <>
+struct intrin_mfma_f32_4x4x4f16<8, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_4x4x4f16(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 4, 0, 0);
+        reg_c.template AsType<float4_t>()(Number<1>{}) = __builtin_amdgcn_mfma_f32_4x4x4f16(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<1>{}], 4, 1, 0);
+    }
+};
+
+// bfp16
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x8bf16_1k;
+
+template <>
+struct intrin_mfma_f32_32x32x8bf16_1k<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf4_t& reg_a, const bhalf4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x8bf16_1k(
+            reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x16bf16_1k;
+
+template <>
+struct intrin_mfma_f32_16x16x16bf16_1k<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf4_t& reg_a, const bhalf4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x4bf16;
+
+template <>
+struct intrin_mfma_f32_32x32x4bf16<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf2_t& reg_a, const bhalf2_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x4bf16(
+            reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x8bf16;
+
+template <>
+struct intrin_mfma_f32_16x16x8bf16<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf2_t& reg_a, const bhalf2_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x8bf16(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_i32_32x32x8i8;
+
+template <>
+struct intrin_mfma_i32_32x32x8i8<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const int8x4_t& reg_a, const int8x4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<int32x16_t>()(Number<0>{}) =
+            __builtin_amdgcn_mfma_i32_32x32x8i8(bit_cast<int32_t>(reg_a),
+                                                bit_cast<int32_t>(reg_b),
+                                                reg_c.template AsType<int32x16_t>()[Number<0>{}],
+                                                0,
+                                                0,
+                                                0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_i32_16x16x16i8;
+
+template <>
+struct intrin_mfma_i32_16x16x16i8<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const int8x4_t& reg_a, const int8x4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<int32x4_t>()(Number<0>{}) =
+            __builtin_amdgcn_mfma_i32_16x16x16i8(bit_cast<int32_t>(reg_a),
+                                                 bit_cast<int32_t>(reg_b),
+                                                 reg_c.template AsType<int32x4_t>()[Number<0>{}],
+                                                 0,
+                                                 0,
+                                                 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f64_16x16x4f64;
+
+template <>
+struct intrin_mfma_f64_16x16x4f64<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const double& reg_a, const double& reg_b, FloatC& reg_c)
+    {
+#ifdef __gfx90a__
+        reg_c.template AsType<double4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f64_16x16x4f64(
+            reg_a, reg_b, reg_c.template AsType<double4_t>()[Number<0>{}], 0, 0, 0);
+#else
+        ignore = reg_a;
+        ignore = reg_b;
+        ignore = reg_c;
+#endif
+    }
+};
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/array.hpp b/3rdparty/composable_kernel/include/ck/utility/array.hpp
new file mode 100644
index 0000000..370a457
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/array.hpp
@@ -0,0 +1,66 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_ARRAY_HPP
+#define CK_ARRAY_HPP
+
+#include "functional2.hpp"
+#include "sequence.hpp"
+
+namespace ck {
+
+template <typename TData, index_t NSize>
+struct Array
+{
+    using type      = Array;
+    using data_type = TData;
+
+    TData mData[NSize];
+
+    __host__ __device__ static constexpr index_t Size() { return NSize; }
+
+    __host__ __device__ constexpr const TData& At(index_t i) const { return mData[i]; }
+
+    __host__ __device__ constexpr TData& At(index_t i) { return mData[i]; }
+
+    __host__ __device__ constexpr const TData& operator[](index_t i) const { return At(i); }
+
+    __host__ __device__ constexpr TData& operator()(index_t i) { return At(i); }
+
+    template <typename T>
+    __host__ __device__ constexpr auto operator=(const T& a)
+    {
+        static_assert(T::Size() == Size(), "wrong! size not the same");
+
+        static_for<0, Size(), 1>{}([&](auto i) { operator()(i) = a[i]; });
+
+        return *this;
+    }
+};
+
+// empty Array
+template <typename TData>
+struct Array<TData, 0>
+{
+    using type      = Array;
+    using data_type = TData;
+
+    __host__ __device__ static constexpr index_t Size() { return 0; }
+};
+
+template <typename X, typename... Xs>
+__host__ __device__ constexpr auto make_array(X&& x, Xs&&... xs)
+{
+    using data_type = remove_cvref_t<X>;
+    return Array<data_type, sizeof...(Xs) + 1>{std::forward<X>(x), std::forward<Xs>(xs)...};
+}
+
+// make empty array
+template <typename X>
+__host__ __device__ constexpr auto make_array()
+{
+    return Array<X, 0>{};
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/array_multi_index.hpp b/3rdparty/composable_kernel/include/ck/utility/array_multi_index.hpp
new file mode 100644
index 0000000..9b8d5b9
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/array_multi_index.hpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_ARRAY_MULTI_INDEX_HPP
+#define CK_ARRAY_MULTI_INDEX_HPP
+
+#include "common_header.hpp"
+
+namespace ck {
+
+template <index_t N>
+using MultiIndex = Array<index_t, N>;
+
+template <typename... Xs>
+__host__ __device__ constexpr auto make_multi_index(Xs&&... xs)
+{
+    return make_array<index_t>(index_t{xs}...);
+}
+
+template <index_t NSize>
+__host__ __device__ constexpr auto make_zero_multi_index()
+{
+    return unpack([](auto... xs) { return make_multi_index(xs...); },
+                  typename uniform_sequence_gen<NSize, 0>::type{});
+}
+
+template <typename T>
+__host__ __device__ constexpr auto to_multi_index(const T& x)
+{
+    return unpack([](auto... ys) { return make_multi_index(ys...); }, x);
+}
+
+template <index_t NSize, typename X>
+__host__ __device__ constexpr auto operator+=(MultiIndex<NSize>& y, const X& x)
+{
+    static_assert(X::Size() == NSize, "wrong! size not the same");
+    static_for<0, NSize, 1>{}([&](auto i) { y(i) += x[i]; });
+    return y;
+}
+
+template <index_t NSize, typename X>
+__host__ __device__ constexpr auto operator-=(MultiIndex<NSize>& y, const X& x)
+{
+    static_assert(X::Size() == NSize, "wrong! size not the same");
+    static_for<0, NSize, 1>{}([&](auto i) { y(i) -= x[i]; });
+    return y;
+}
+
+template <index_t NSize, typename T>
+__host__ __device__ constexpr auto operator+(const MultiIndex<NSize>& a, const T& b)
+{
+    using type = MultiIndex<NSize>;
+    static_assert(T::Size() == NSize, "wrong! size not the same");
+    type r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = a[i] + b[i]; });
+    return r;
+}
+
+template <index_t NSize, typename T>
+__host__ __device__ constexpr auto operator-(const MultiIndex<NSize>& a, const T& b)
+{
+    using type = MultiIndex<NSize>;
+    static_assert(T::Size() == NSize, "wrong! size not the same");
+    type r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = a[i] - b[i]; });
+    return r;
+}
+
+template <index_t NSize, typename T>
+__host__ __device__ constexpr auto operator*(const MultiIndex<NSize>& a, const T& b)
+{
+    using type = MultiIndex<NSize>;
+    static_assert(T::Size() == NSize, "wrong! size not the same");
+    type r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = a[i] * b[i]; });
+    return r;
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/c_style_pointer_cast.hpp b/3rdparty/composable_kernel/include/ck/utility/c_style_pointer_cast.hpp
new file mode 100644
index 0000000..6e8b008
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/c_style_pointer_cast.hpp
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_C_STYLE_POINTER_CAST_HPP
+#define CK_C_STYLE_POINTER_CAST_HPP
+
+#include "type.hpp"
+#include "enable_if.hpp"
+
+namespace ck {
+
+template <typename PY,
+          typename PX,
+          typename enable_if<is_pointer_v<PY> && is_pointer_v<PX>, bool>::type = false>
+__host__ __device__ PY c_style_pointer_cast(PX p_x)
+{
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wold-style-cast"
+#pragma clang diagnostic ignored "-Wcast-align"
+    return (PY)p_x; // NOLINT(old-style-cast, cast-align)
+#pragma clang diagnostic pop
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/common_header.hpp b/3rdparty/composable_kernel/include/ck/utility/common_header.hpp
new file mode 100644
index 0000000..1378bbe
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/common_header.hpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/array.hpp"
+#include "ck/utility/container_helper.hpp"
+#include "ck/utility/statically_indexed_array.hpp"
+#include "ck/utility/container_element_picker.hpp"
+#include "ck/utility/multi_index.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/functional.hpp"
+#include "ck/utility/functional2.hpp"
+#include "ck/utility/functional3.hpp"
+#include "ck/utility/functional4.hpp"
+#include "ck/utility/enable_if.hpp"
+#include "ck/utility/ignore.hpp"
+#include "ck/utility/integral_constant.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/utility/number.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/sequence_helper.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/utility/tuple_helper.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/utility/magic_division.hpp"
+#include "ck/utility/c_style_pointer_cast.hpp"
+#include "ck/utility/is_known_at_compile_time.hpp"
+#include "ck/utility/transpose_vectors.hpp"
+#include "ck/utility/inner_product.hpp"
+#include "ck/utility/thread_group.hpp"
+#include "ck/utility/debug.hpp"
+
+#include "ck/utility/amd_buffer_addressing.hpp"
+#include "ck/utility/generic_memory_space_atomic.hpp"
+#include "ck/utility/get_id.hpp"
+#include "ck/utility/thread_group.hpp"
+#include "ck/utility/synchronization.hpp"
+#include "ck/utility/amd_address_space.hpp"
+#include "ck/utility/static_buffer.hpp"
+#include "ck/utility/dynamic_buffer.hpp"
+
+// TODO: remove this
+#if CK_USE_AMD_INLINE_ASM
+#include "ck/utility/amd_inline_asm.hpp"
+#endif
+
+#ifdef CK_USE_AMD_MFMA
+#include "ck/utility/amd_xdlops.hpp"
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/container_element_picker.hpp b/3rdparty/composable_kernel/include/ck/utility/container_element_picker.hpp
new file mode 100644
index 0000000..abc5185
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/container_element_picker.hpp
@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_CONTAINER_ELEMENT_PICKER_HPP
+#define CK_CONTAINER_ELEMENT_PICKER_HPP
+
+#include "functional2.hpp"
+#include "sequence.hpp"
+
+namespace ck {
+
+// Arr: Array or StaticallyIndexedArray
+// Picks: Sequence<...>
+template <typename Arr, typename Picks>
+struct ContainerElementPicker
+{
+    using type = ContainerElementPicker;
+#if 0
+    using data_type = typename Arr::data_type;
+#endif
+
+    __host__ __device__ constexpr ContainerElementPicker() = delete;
+
+    __host__ __device__ constexpr ContainerElementPicker(Arr& array) : mArray{array}
+    {
+        constexpr index_t imax =
+            reduce_on_sequence(Picks{}, math::maximize<index_t>{}, Number<0>{});
+
+        static_assert(imax < Arr::Size(), "wrong! exceeding # array element");
+    }
+
+    __host__ __device__ static constexpr auto Size() { return Picks::Size(); }
+
+    template <index_t I>
+    __host__ __device__ constexpr const auto& At(Number<I> i) const
+    {
+        static_assert(I < Size(), "wrong!");
+
+        constexpr auto IP = Picks{}[i];
+        return mArray[IP];
+    }
+
+    template <index_t I>
+    __host__ __device__ constexpr auto& At(Number<I> i)
+    {
+        static_assert(I < Size(), "wrong!");
+
+        constexpr auto IP = Picks{}[i];
+        return mArray(IP);
+    }
+
+    template <index_t I>
+    __host__ __device__ constexpr const auto& operator[](Number<I> i) const
+    {
+        return At(i);
+    }
+
+    template <index_t I>
+    __host__ __device__ constexpr auto& operator()(Number<I> i)
+    {
+        return At(i);
+    }
+
+    template <typename T>
+    __host__ __device__ constexpr auto operator=(const T& a)
+    {
+        static_assert(T::Size() == Size(), "wrong! size not the same");
+
+        static_for<0, Size(), 1>{}([&](auto i) { operator()(i) = a[i]; });
+
+        return *this;
+    }
+
+    private:
+    Arr& mArray;
+};
+
+// Arr: Array or StaticallyIndexedArray
+// Picks: Sequence<...>
+template <typename Arr, typename Picks>
+struct ConstantContainerElementPicker
+{
+    using type = ConstantContainerElementPicker;
+#if 0
+    using data_type = typename Arr::data_type;
+#endif
+
+    __host__ __device__ constexpr ConstantContainerElementPicker() = delete;
+
+    __host__ __device__ constexpr ConstantContainerElementPicker(const Arr& array) : mArray{array}
+    {
+        constexpr index_t imax =
+            reduce_on_sequence(Picks{}, math::maximize<index_t>{}, Number<0>{});
+
+        static_assert(imax < Arr::Size(), "wrong! exceeding # array element");
+    }
+
+    __host__ __device__ static constexpr auto Size() { return Picks::Size(); }
+
+    template <index_t I>
+    __host__ __device__ constexpr const auto& At(Number<I> i) const
+    {
+        static_assert(I < Size(), "wrong!");
+
+        constexpr auto IP = Picks{}[i];
+        return mArray[IP];
+    }
+
+    template <index_t I>
+    __host__ __device__ constexpr const auto& operator[](Number<I> i) const
+    {
+        return At(i);
+    }
+
+    private:
+    const Arr& mArray;
+};
+
+template <typename Arr, typename Picks, typename X>
+__host__ __device__ constexpr auto operator+=(ContainerElementPicker<Arr, Picks>& y, const X& x)
+{
+    using Y                 = ContainerElementPicker<Arr, Picks>;
+    constexpr index_t nsize = Y::Size();
+
+    static_assert(nsize == X::Size(), "wrong! size not the same");
+
+    static_for<0, nsize, 1>{}([&](auto i) { y(i) += x[i]; });
+
+    return y;
+}
+
+template <typename Arr, typename Picks, typename X>
+__host__ __device__ constexpr auto operator-=(ContainerElementPicker<Arr, Picks>& y, const X& x)
+{
+    using Y                 = ContainerElementPicker<Arr, Picks>;
+    constexpr index_t nsize = Y::Size();
+
+    static_assert(nsize == X::Size(), "wrong! size not the same");
+
+    static_for<0, nsize, 1>{}([&](auto i) { y(i) -= x[i]; });
+
+    return y;
+}
+
+template <typename Arr, typename Picks>
+__host__ __device__ constexpr auto pick_container_element(Arr& a, Picks)
+{
+    return ContainerElementPicker<Arr, Picks>(a);
+}
+
+template <typename Arr, typename Picks>
+__host__ __device__ constexpr auto pick_container_element(const Arr& a, Picks)
+{
+    return ConstantContainerElementPicker<Arr, Picks>(a);
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/container_helper.hpp b/3rdparty/composable_kernel/include/ck/utility/container_helper.hpp
new file mode 100644
index 0000000..c8b02bc
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/container_helper.hpp
@@ -0,0 +1,393 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_CONTAINER_HELPER_HPP
+#define CK_CONTAINER_HELPER_HPP
+
+#include "sequence.hpp"
+#include "sequence_helper.hpp"
+#include "array.hpp"
+#include "tuple.hpp"
+#include "tuple_helper.hpp"
+#include "statically_indexed_array.hpp"
+#include "container_element_picker.hpp"
+
+namespace ck {
+
+template <typename TData, index_t NSize>
+__host__ __device__ constexpr auto container_push_back(const Array<TData, NSize>& a, const TData& x)
+{
+    Array<TData, NSize + 1> r;
+
+    static_for<0, NSize, 1>{}([&r, &a ](auto i) constexpr { r(i) = a[i]; });
+
+    r(Number<NSize>{}) = x;
+
+    return r;
+}
+
+template <typename... Ts, typename T>
+__host__ __device__ constexpr auto container_push_front(const Tuple<Ts...>& a, const T& x)
+{
+    return container_concat(make_tuple(x), a);
+}
+
+template <typename... Ts, typename T>
+__host__ __device__ constexpr auto container_push_back(const Tuple<Ts...>& a, const T& x)
+{
+    return container_concat(a, make_tuple(x));
+}
+
+template <typename TData, index_t NSize, index_t... IRs>
+__host__ __device__ constexpr auto
+container_reorder_given_new2old(const Array<TData, NSize>& old_array, Sequence<IRs...> /*new2old*/)
+{
+    static_assert(NSize == sizeof...(IRs), "wrong! size not consistent");
+
+    static_assert(is_valid_sequence_map<Sequence<IRs...>>{}, "wrong! invalid reorder map");
+
+    return make_array(old_array[Number<IRs>{}]...);
+}
+
+template <typename TData, index_t NSize, index_t... IRs>
+__host__ __device__ constexpr auto
+container_reorder_given_old2new(const Array<TData, NSize>& old_array, Sequence<IRs...> old2new)
+{
+    return container_reorder_given_new2old(
+        old_array, typename sequence_map_inverse<decltype(old2new)>::type{});
+}
+
+template <typename... Ts, index_t... IRs>
+__host__ __device__ constexpr auto container_reorder_given_new2old(const Tuple<Ts...>& old_tuple,
+                                                                   Sequence<IRs...> /*new2old*/)
+{
+    static_assert(sizeof...(Ts) == sizeof...(IRs), "wrong! size not consistent");
+
+    static_assert(is_valid_sequence_map<Sequence<IRs...>>{}, "wrong! invalid reorder map");
+
+    return make_tuple(old_tuple[Number<IRs>{}]...);
+}
+
+template <typename... Ts, index_t... IRs>
+__host__ __device__ constexpr auto container_reorder_given_old2new(const Tuple<Ts...>& old_tuple,
+                                                                   Sequence<IRs...> old2new)
+{
+    return container_reorder_given_new2old(
+        old_tuple, typename sequence_map_inverse<decltype(old2new)>::type{});
+}
+
+template <index_t... Is, index_t... IRs>
+__host__ __device__ constexpr auto container_reorder_given_new2old(Sequence<Is...> /* old_seq */,
+                                                                   Sequence<IRs...> /*new2old*/)
+{
+    static_assert(sizeof...(Is) == sizeof...(IRs), "wrong! size not consistent");
+
+    static_assert(is_valid_sequence_map<Sequence<IRs...>>{}, "wrong! invalid reorder map");
+
+    return Sequence<Sequence<Is...>::At(Number<IRs>{})...>{};
+}
+
+template <index_t... Is, index_t... IRs>
+__host__ __device__ constexpr auto container_reorder_given_old2new(Sequence<Is...> old_seq,
+                                                                   Sequence<IRs...> /* old2new */)
+{
+    static_assert(sizeof...(Is) == sizeof...(IRs), "wrong! size not consistent");
+
+    static_assert(is_valid_sequence_map<Sequence<IRs...>>{}, "wrong! invalid reorder map");
+
+    constexpr auto new2old = typename sequence_map_inverse<Sequence<IRs...>>::type{};
+
+    return container_reorder_given_new2old(old_seq, new2old);
+}
+
+#if !CK_WORKAROUND_SWDEV_275126
+// rocm-4.1 compiler would crash for recursive lambda
+template <typename Container,
+          typename Reduce,
+          typename Init,
+          index_t IBegin = 0,
+          index_t IEnd   = Container::Size(),
+          index_t IStep  = 1>
+__host__ __device__ constexpr auto container_reduce(const Container& x,
+                                                    Reduce reduce,
+                                                    Init init,
+                                                    Number<IBegin> = Number<0>{},
+                                                    Number<IEnd>   = Number<Container::Size()>{},
+                                                    Number<IStep>  = Number<1>{})
+{
+    static_assert((IEnd - IBegin) % IStep == 0, "wrong!");
+
+    // f is recursive function, fs is a dummy of f
+    // i is index, y_old is current scan, r_old is current reduction
+    auto f = [&](auto fs, auto i, auto r_old) {
+        auto r_new = reduce(x[i], r_old);
+
+        if constexpr(i.value < IEnd - IStep)
+        {
+            // recursively call f/fs
+            return fs(fs, i + Number<IStep>{}, r_new);
+        }
+        else
+        {
+            return r_new;
+        }
+    };
+
+    // start recursion
+    return f(f, Number<IBegin>{}, init);
+}
+#else
+// i is index, y_old is current scan, r_old is current reduction
+template <typename Container,
+          typename Reduce,
+          typename ROld,
+          index_t I,
+          index_t IEnd,
+          index_t IStep>
+__host__ __device__ constexpr auto container_reduce_impl(
+    const Container& x, Reduce reduce, ROld r_old, Number<I> i, Number<IEnd>, Number<IStep>)
+{
+    auto r_new = reduce(x[i], r_old);
+
+    if constexpr(i.value < IEnd - IStep)
+    {
+        return container_reduce_impl(
+            x, reduce, r_new, i + Number<IStep>{}, Number<IEnd>{}, Number<IStep>{});
+    }
+    else
+    {
+        return r_new;
+    }
+}
+
+// rocm-4.1 compiler would crash for recursive lambda
+// container reduce with initial value
+template <typename Container,
+          typename Reduce,
+          typename Init,
+          index_t IBegin = 0,
+          index_t IEnd   = Container::Size(),
+          index_t IStep  = 1>
+__host__ __device__ constexpr auto container_reduce(const Container& x,
+                                                    Reduce reduce,
+                                                    Init init,
+                                                    Number<IBegin> = Number<0>{},
+                                                    Number<IEnd>   = Number<Container::Size()>{},
+                                                    Number<IStep>  = Number<1>{})
+{
+    static_assert((IEnd - IBegin) % IStep == 0, "wrong!");
+
+    if constexpr(IEnd > IBegin)
+    {
+        return container_reduce_impl(
+            x, reduce, init, Number<IBegin>{}, Number<IEnd>{}, Number<IStep>{});
+    }
+    else
+    {
+        return init;
+    }
+}
+#endif
+
+template <typename TData, index_t NSize, typename Reduce>
+__host__ __device__ constexpr auto
+container_reverse_inclusive_scan(const Array<TData, NSize>& x, Reduce f, TData init)
+{
+    Array<TData, NSize> y;
+
+    TData r = init;
+
+    static_for<NSize - 1, 0, -1>{}([&](auto i) {
+        r    = f(r, x[i]);
+        y(i) = r;
+    });
+
+    r              = f(r, x[Number<0>{}]);
+    y(Number<0>{}) = r;
+
+    return y;
+}
+
+template <typename TData, index_t NSize, typename Reduce>
+__host__ __device__ constexpr auto
+container_reverse_exclusive_scan(const Array<TData, NSize>& x, Reduce f, TData init)
+{
+    Array<TData, NSize> y;
+
+    TData r = init;
+
+    static_for<NSize - 1, 0, -1>{}([&](auto i) {
+        y(i) = r;
+        r    = f(r, x[i]);
+    });
+
+    y(Number<0>{}) = r;
+
+    return y;
+}
+
+template <index_t... Is, typename Reduce, index_t Init>
+__host__ __device__ constexpr auto
+container_reverse_exclusive_scan(const Sequence<Is...>& seq, Reduce f, Number<Init>)
+{
+    return reverse_exclusive_scan_sequence(seq, f, Number<Init>{});
+}
+
+#if !CK_WORKAROUND_SWDEV_275126
+// rocm4.1 compiler would crash with recursive lambda
+template <typename... Xs, typename Reduce, typename Init>
+__host__ __device__ constexpr auto
+container_reverse_exclusive_scan(const Tuple<Xs...>& x, Reduce reduce, Init init)
+{
+    constexpr index_t NSize = sizeof...(Xs);
+
+    // f is recursive function, fs is a dummy of f
+    // i is index, y_old is current scan, r_old is current reduction
+    auto f = [&](auto fs, auto i, auto y_old, auto r_old) {
+        auto r_new = reduce(x[i], r_old);
+
+        auto y_new = container_push_front(y_old, r_new);
+
+        if constexpr(i.value > 1)
+        {
+            // recursively call f/fs
+            return fs(fs, i - Number<1>{}, y_new, r_new);
+        }
+        else
+        {
+            return y_new;
+        }
+    };
+
+    // start recursion
+    return f(f, Number<NSize - 1>{}, make_tuple(init), init);
+}
+#else
+// i is index, y_old is current scan, r_old is current reduction
+template <typename... Xs, typename Reduce, index_t I, typename YOld, typename ROld>
+__host__ __device__ constexpr auto container_reverse_exclusive_scan_impl(
+    const Tuple<Xs...>& x, Reduce reduce, Number<I> i, YOld y_old, ROld r_old)
+{
+    auto r_new = reduce(x[i], r_old);
+
+    auto y_new = container_push_front(y_old, r_new);
+
+    if constexpr(i.value > 1)
+    {
+        // recursively call f/fs
+        return container_reverse_exclusive_scan_impl(x, reduce, i - Number<1>{}, y_new, r_new);
+    }
+    else
+    {
+        return y_new;
+    }
+}
+
+template <typename... Xs, typename Reduce, typename Init>
+__host__ __device__ constexpr auto
+container_reverse_exclusive_scan(const Tuple<Xs...>& x, Reduce reduce, Init init)
+{
+    constexpr index_t NSize = sizeof...(Xs);
+
+    return container_reverse_exclusive_scan_impl(
+        x, reduce, Number<NSize - 1>{}, make_tuple(init), init);
+}
+#endif
+
+// TODO: update to like container_reverse_exclusive_scan to deal with Tuple of Numebr<>
+template <typename... Xs, typename Reduce, typename TData>
+__host__ __device__ constexpr auto
+container_reverse_inclusive_scan(const Tuple<Xs...>& x, Reduce f, TData init)
+{
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> y;
+
+    TData r = init;
+
+    static_for<NSize - 1, 0, -1>{}([&](auto i) {
+        r    = f(r, x[i]);
+        y(i) = r;
+    });
+
+    r              = f(r, x[Number<0>{}]);
+    y(Number<0>{}) = r;
+
+    return y;
+}
+
+template <typename X, typename... Ys>
+__host__ __device__ constexpr auto container_concat(const X& x, const Ys&... ys)
+{
+    return container_concat(x, container_concat(ys...));
+}
+
+template <typename T, index_t NX, index_t NY>
+__host__ __device__ constexpr auto container_concat(const Array<T, NX>& ax, const Array<T, NY>& ay)
+{
+    return unpack2(
+        [&](auto&&... zs) { return make_array(std::forward<decltype(zs)>(zs)...); }, ax, ay);
+}
+
+template <typename... X, typename... Y>
+__host__ __device__ constexpr auto container_concat(const Tuple<X...>& tx, const Tuple<Y...>& ty)
+{
+    return unpack2(
+        [&](auto&&... zs) { return make_tuple(std::forward<decltype(zs)>(zs)...); }, tx, ty);
+}
+
+template <typename Container>
+__host__ __device__ constexpr auto container_concat(const Container& x)
+{
+    return x;
+}
+
+template <typename T, index_t N, index_t... Is>
+__host__ __device__ constexpr auto get_container_subset(const Array<T, N>& arr, Sequence<Is...>)
+{
+    static_assert(N >= sizeof...(Is), "wrong! size");
+
+    return make_array(arr[Number<Is>{}]...);
+}
+
+template <typename... Ts, index_t... Is>
+__host__ __device__ constexpr auto get_container_subset(const Tuple<Ts...>& tup, Sequence<Is...>)
+{
+    static_assert(sizeof...(Ts) >= sizeof...(Is), "wrong! size");
+
+    return make_tuple(tup[Number<Is>{}]...);
+}
+
+template <typename T, index_t N, index_t... Is>
+__host__ __device__ constexpr void
+set_container_subset(Array<T, N>& y, Sequence<Is...> picks, const Array<T, sizeof...(Is)>& x)
+{
+    static_assert(N >= sizeof...(Is), "wrong! size");
+
+    static_for<0, sizeof...(Is), 1>{}([&](auto i) { y(picks[i]) = x[i]; });
+}
+
+template <typename... Ys, index_t... Is, typename... Xs>
+__host__ __device__ constexpr void
+set_container_subset(Tuple<Ys...>& y, Sequence<Is...> picks, const Tuple<Xs...>& x)
+{
+    static_assert(sizeof...(Ys) >= sizeof...(Is) && sizeof...(Is) == sizeof...(Xs), "wrong! size");
+
+    static_for<0, sizeof...(Is), 1>{}([&](auto i) { y(picks[i]) = x[i]; });
+}
+
+template <index_t... Is>
+__host__ __device__ constexpr auto sequence_to_tuple_of_number(Sequence<Is...>)
+{
+    using Seq = Sequence<Is...>;
+
+    return generate_tuple(
+        [&](auto i) {
+            constexpr index_t tmp = Seq::At(i);
+            return Number<tmp>{};
+        },
+        Seq::Size());
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/data_type.hpp b/3rdparty/composable_kernel/include/ck/utility/data_type.hpp
new file mode 100644
index 0000000..40ee8b6
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/data_type.hpp
@@ -0,0 +1,1059 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/statically_indexed_array.hpp"
+
+namespace ck {
+
+using bhalf_t = ushort;
+using half_t  = _Float16;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using int4_t = _BitInt(4);
+#endif
+
+// vector_type
+template <typename T, index_t N>
+struct vector_type;
+
+// Caution: DO NOT REMOVE
+// intentionally have only declaration but no definition to cause compilation failure when trying to
+// instantiate this template. The purpose is to catch user's mistake when trying to make "vector of
+// vectors"
+template <typename T, index_t V, index_t N>
+struct vector_type<T __attribute__((ext_vector_type(V))), N>;
+
+// Caution: DO NOT REMOVE
+// intentionally have only declaration but no definition to cause compilation failure when trying to
+// instantiate this template. The purpose is to catch user's mistake when trying to make "vector of
+// vectors"
+template <typename T, index_t V, index_t N>
+struct vector_type<vector_type<T, V>, N>;
+
+// vector_type_maker
+// This is the right way to handle "vector of vectors": making a bigger vector instead
+template <typename T, index_t N>
+struct vector_type_maker
+{
+    using type = vector_type<T, N>;
+};
+
+template <typename T, index_t N0, index_t N1>
+struct vector_type_maker<T __attribute__((ext_vector_type(N1))), N0>
+{
+    using type = vector_type<T, N0 * N1>;
+};
+
+template <typename T, index_t N0, index_t N1>
+struct vector_type_maker<vector_type<T, N1>, N0>
+{
+    using type = vector_type<T, N0 * N1>;
+};
+
+template <typename T, index_t N>
+using vector_type_maker_t = typename vector_type_maker<T, N>::type;
+
+template <typename T, index_t N>
+__host__ __device__ constexpr auto make_vector_type(Number<N>)
+{
+    return typename vector_type_maker<T, N>::type{};
+}
+
+// scalar_type
+template <typename TV>
+struct scalar_type;
+
+// is_scalar_type
+template <typename TV>
+struct is_scalar_type
+{
+    static constexpr bool value = (scalar_type<remove_cvref_t<TV>>::vector_size == 1);
+};
+
+// has_same_scalar_type
+template <typename X, typename Y>
+using has_same_scalar_type = is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                     typename scalar_type<remove_cvref_t<Y>>::type>;
+
+template <typename T, index_t N>
+struct scalar_type<T __attribute__((ext_vector_type(N)))>
+{
+    using type                           = T;
+    static constexpr index_t vector_size = N;
+};
+
+template <typename T, index_t N>
+struct scalar_type<vector_type<T, N>>
+{
+    using type                           = T;
+    static constexpr index_t vector_size = N;
+};
+
+//
+template <>
+struct scalar_type<double>
+{
+    using type                           = double;
+    static constexpr index_t vector_size = 1;
+};
+
+template <>
+struct scalar_type<float>
+{
+    using type                           = float;
+    static constexpr index_t vector_size = 1;
+};
+
+template <>
+struct scalar_type<half_t>
+{
+    using type                           = half_t;
+    static constexpr index_t vector_size = 1;
+};
+
+template <>
+struct scalar_type<bhalf_t>
+{
+    using type                           = bhalf_t;
+    static constexpr index_t vector_size = 1;
+};
+
+template <>
+struct scalar_type<int32_t>
+{
+    using type                           = int32_t;
+    static constexpr index_t vector_size = 1;
+};
+
+template <>
+struct scalar_type<int8_t>
+{
+    using type                           = int8_t;
+    static constexpr index_t vector_size = 1;
+};
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+template <>
+struct scalar_type<int4_t>
+{
+    using type                           = int4_t;
+    static constexpr index_t vector_size = 1;
+};
+#endif
+
+//
+template <typename T>
+struct vector_type<T, 1>
+{
+    using d1_t = T;
+    using type = d1_t;
+
+    union
+    {
+        T d1_;
+        StaticallyIndexedArray<T, 1> d1x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value, "wrong!");
+
+        return data_.d1x1_;
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value, "wrong!");
+
+        return data_.d1x1_;
+    }
+};
+
+template <typename T>
+struct vector_type<T, 2>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+
+    using type = d2_t;
+
+    union
+    {
+        d2_t d2_;
+        StaticallyIndexedArray<d1_t, 2> d1x2_;
+        StaticallyIndexedArray<d2_t, 1> d2x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value, "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x2_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value, "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x2_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 4>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+
+    using type = d4_t;
+
+    union
+    {
+        d4_t d4_;
+        StaticallyIndexedArray<d1_t, 4> d1x4_;
+        StaticallyIndexedArray<d2_t, 2> d2x2_;
+        StaticallyIndexedArray<d4_t, 1> d4x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x4_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x2_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x4_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x2_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 8>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+
+    using type = d8_t;
+
+    union
+    {
+        d8_t d8_;
+        StaticallyIndexedArray<d1_t, 8> d1x8_;
+        StaticallyIndexedArray<d2_t, 4> d2x4_;
+        StaticallyIndexedArray<d4_t, 2> d4x2_;
+        StaticallyIndexedArray<d8_t, 1> d8x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x8_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x4_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x2_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x8_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x4_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x2_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 16>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+    typedef T d16_t __attribute__((ext_vector_type(16)));
+
+    using type = d16_t;
+
+    union
+    {
+        d16_t d16_;
+        StaticallyIndexedArray<d1_t, 16> d1x16_;
+        StaticallyIndexedArray<d2_t, 8> d2x8_;
+        StaticallyIndexedArray<d4_t, 4> d4x4_;
+        StaticallyIndexedArray<d8_t, 2> d8x2_;
+        StaticallyIndexedArray<d16_t, 1> d16x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x16_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x8_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x4_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x2_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x16_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x8_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x4_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x2_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 32>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+    typedef T d16_t __attribute__((ext_vector_type(16)));
+    typedef T d32_t __attribute__((ext_vector_type(32)));
+
+    using type = d32_t;
+
+    union
+    {
+        d32_t d32_;
+        StaticallyIndexedArray<d1_t, 32> d1x32_;
+        StaticallyIndexedArray<d2_t, 16> d2x16_;
+        StaticallyIndexedArray<d4_t, 8> d4x8_;
+        StaticallyIndexedArray<d8_t, 4> d8x4_;
+        StaticallyIndexedArray<d16_t, 2> d16x2_;
+        StaticallyIndexedArray<d32_t, 1> d32x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x32_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x16_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x8_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x4_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x2_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x32_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x16_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x8_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x4_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x2_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 64>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+    typedef T d16_t __attribute__((ext_vector_type(16)));
+    typedef T d32_t __attribute__((ext_vector_type(32)));
+    typedef T d64_t __attribute__((ext_vector_type(64)));
+
+    using type = d64_t;
+
+    union
+    {
+        d64_t d64_;
+        StaticallyIndexedArray<d1_t, 64> d1x64_;
+        StaticallyIndexedArray<d2_t, 32> d2x32_;
+        StaticallyIndexedArray<d4_t, 16> d4x16_;
+        StaticallyIndexedArray<d8_t, 8> d8x8_;
+        StaticallyIndexedArray<d16_t, 4> d16x4_;
+        StaticallyIndexedArray<d32_t, 2> d32x2_;
+        StaticallyIndexedArray<d64_t, 1> d64x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                          is_same<X, d64_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x64_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x32_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x16_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x8_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x4_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x2_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                          is_same<X, d64_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x64_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x32_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x16_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x8_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x4_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x2_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 128>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+    typedef T d16_t __attribute__((ext_vector_type(16)));
+    typedef T d32_t __attribute__((ext_vector_type(32)));
+    typedef T d64_t __attribute__((ext_vector_type(64)));
+    typedef T d128_t __attribute__((ext_vector_type(128)));
+
+    using type = d128_t;
+
+    union
+    {
+        d128_t d128_;
+        StaticallyIndexedArray<d1_t, 128> d1x128_;
+        StaticallyIndexedArray<d2_t, 64> d2x64_;
+        StaticallyIndexedArray<d4_t, 32> d4x32_;
+        StaticallyIndexedArray<d8_t, 16> d8x16_;
+        StaticallyIndexedArray<d16_t, 8> d16x8_;
+        StaticallyIndexedArray<d32_t, 4> d32x4_;
+        StaticallyIndexedArray<d64_t, 2> d64x2_;
+        StaticallyIndexedArray<d128_t, 1> d128x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                          is_same<X, d64_t>::value || is_same<X, d128_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x128_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x64_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x32_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x16_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x8_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x4_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x2_;
+        }
+        else if constexpr(is_same<X, d128_t>::value)
+        {
+            return data_.d128x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                          is_same<X, d64_t>::value || is_same<X, d128_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x128_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x64_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x32_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x16_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x8_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x4_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x2_;
+        }
+        else if constexpr(is_same<X, d128_t>::value)
+        {
+            return data_.d128x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 256>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+    typedef T d16_t __attribute__((ext_vector_type(16)));
+    typedef T d32_t __attribute__((ext_vector_type(32)));
+    typedef T d64_t __attribute__((ext_vector_type(64)));
+    typedef T d128_t __attribute__((ext_vector_type(128)));
+    typedef T d256_t __attribute__((ext_vector_type(256)));
+
+    using type = d256_t;
+
+    union
+    {
+        d256_t d256_;
+        StaticallyIndexedArray<d1_t, 256> d1x256_;
+        StaticallyIndexedArray<d2_t, 128> d2x128_;
+        StaticallyIndexedArray<d4_t, 64> d4x64_;
+        StaticallyIndexedArray<d8_t, 32> d8x32_;
+        StaticallyIndexedArray<d16_t, 16> d16x16_;
+        StaticallyIndexedArray<d32_t, 8> d32x8_;
+        StaticallyIndexedArray<d64_t, 4> d64x4_;
+        StaticallyIndexedArray<d128_t, 2> d128x2_;
+        StaticallyIndexedArray<d256_t, 1> d256x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(
+            is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value ||
+                is_same<X, d8_t>::value || is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                is_same<X, d64_t>::value || is_same<X, d128_t>::value || is_same<X, d256_t>::value,
+            "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x256_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x128_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x64_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x32_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x16_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x8_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x4_;
+        }
+        else if constexpr(is_same<X, d128_t>::value)
+        {
+            return data_.d128x2_;
+        }
+        else if constexpr(is_same<X, d256_t>::value)
+        {
+            return data_.d256x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(
+            is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value ||
+                is_same<X, d8_t>::value || is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                is_same<X, d64_t>::value || is_same<X, d128_t>::value || is_same<X, d256_t>::value,
+            "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x256_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x128_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x64_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x32_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x16_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x8_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x4_;
+        }
+        else if constexpr(is_same<X, d128_t>::value)
+        {
+            return data_.d128x2_;
+        }
+        else if constexpr(is_same<X, d256_t>::value)
+        {
+            return data_.d256x1_;
+        }
+    }
+};
+
+// fp64
+using double2_t = typename vector_type<double, 2>::type;
+using double4_t = typename vector_type<double, 4>::type;
+
+// fp32
+using float2_t  = typename vector_type<float, 2>::type;
+using float4_t  = typename vector_type<float, 4>::type;
+using float8_t  = typename vector_type<float, 8>::type;
+using float16_t = typename vector_type<float, 16>::type;
+using float32_t = typename vector_type<float, 32>::type;
+using float64_t = typename vector_type<float, 64>::type;
+
+// fp16
+using half2_t  = typename vector_type<half_t, 2>::type;
+using half4_t  = typename vector_type<half_t, 4>::type;
+using half8_t  = typename vector_type<half_t, 8>::type;
+using half16_t = typename vector_type<half_t, 16>::type;
+using half32_t = typename vector_type<half_t, 32>::type;
+using half64_t = typename vector_type<half_t, 64>::type;
+
+// bfp16
+using bhalf2_t  = typename vector_type<bhalf_t, 2>::type;
+using bhalf4_t  = typename vector_type<bhalf_t, 4>::type;
+using bhalf8_t  = typename vector_type<bhalf_t, 8>::type;
+using bhalf16_t = typename vector_type<bhalf_t, 16>::type;
+using bhalf32_t = typename vector_type<bhalf_t, 32>::type;
+using bhalf64_t = typename vector_type<bhalf_t, 64>::type;
+
+// i32
+using int32x2_t  = typename vector_type<int32_t, 2>::type;
+using int32x4_t  = typename vector_type<int32_t, 4>::type;
+using int32x8_t  = typename vector_type<int32_t, 8>::type;
+using int32x16_t = typename vector_type<int32_t, 16>::type;
+using int32x32_t = typename vector_type<int32_t, 32>::type;
+using int32x64_t = typename vector_type<int32_t, 64>::type;
+
+// i8
+using int8x2_t  = typename vector_type<int8_t, 2>::type;
+using int8x4_t  = typename vector_type<int8_t, 4>::type;
+using int8x8_t  = typename vector_type<int8_t, 8>::type;
+using int8x16_t = typename vector_type<int8_t, 16>::type;
+using int8x32_t = typename vector_type<int8_t, 32>::type;
+using int8x64_t = typename vector_type<int8_t, 64>::type;
+
+// Convert X to Y
+template <typename Y, typename X>
+__host__ __device__ constexpr Y type_convert(X x)
+{
+    static_assert(!std::is_reference_v<Y> && !std::is_reference_v<X>);
+
+    return static_cast<Y>(x);
+}
+
+// convert bfp16 to fp32
+template <>
+inline __host__ __device__ constexpr float type_convert<float, bhalf_t>(bhalf_t x)
+{
+    union
+    {
+        uint32_t int32;
+        float fp32;
+    } u = {uint32_t(x) << 16};
+
+    return u.fp32;
+}
+
+// convert fp32 to bfp16
+template <>
+inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, float>(float x)
+{
+    union
+    {
+        float fp32;
+        uint32_t int32;
+    } u = {x};
+
+    if(~u.int32 & 0x7f800000)
+    {
+        // When the exponent bits are not all 1s, then the value is zero, normal,
+        // or subnormal. We round the bfloat16 mantissa up by adding 0x7FFF, plus
+        // 1 if the least significant bit of the bfloat16 mantissa is 1 (odd).
+        // This causes the bfloat16's mantissa to be incremented by 1 if the 16
+        // least significant bits of the float mantissa are greater than 0x8000,
+        // or if they are equal to 0x8000 and the least significant bit of the
+        // bfloat16 mantissa is 1 (odd). This causes it to be rounded to even when
+        // the lower 16 bits are exactly 0x8000. If the bfloat16 mantissa already
+        // has the value 0x7f, then incrementing it causes it to become 0x00 and
+        // the exponent is incremented by one, which is the next higher FP value
+        // to the unrounded bfloat16 value. When the bfloat16 value is subnormal
+        // with an exponent of 0x00 and a mantissa of 0x7F, it may be rounded up
+        // to a normal value with an exponent of 0x01 and a mantissa of 0x00.
+        // When the bfloat16 value has an exponent of 0xFE and a mantissa of 0x7F,
+        // incrementing it causes it to become an exponent of 0xFF and a mantissa
+        // of 0x00, which is Inf, the next higher value to the unrounded value.
+        u.int32 += 0x7fff + ((u.int32 >> 16) & 1); // Round to nearest, round to even
+    }
+    else if(u.int32 & 0xffff)
+    {
+        // When all of the exponent bits are 1, the value is Inf or NaN.
+        // Inf is indicated by a zero mantissa. NaN is indicated by any nonzero
+        // mantissa bit. Quiet NaN is indicated by the most significant mantissa
+        // bit being 1. Signaling NaN is indicated by the most significant
+        // mantissa bit being 0 but some other bit(s) being 1. If any of the
+        // lower 16 bits of the mantissa are 1, we set the least significant bit
+        // of the bfloat16 mantissa, in order to preserve signaling NaN in case
+        // the bloat16's mantissa bits are all 0.
+        u.int32 |= 0x10000; // Preserve signaling NaN
+    }
+
+    return uint16_t(u.int32 >> 16);
+}
+
+template <typename T>
+struct NumericLimits
+{
+    __host__ __device__ static constexpr T Min() { return std::numeric_limits<T>::min(); }
+
+    __host__ __device__ static constexpr T Max() { return std::numeric_limits<T>::max(); }
+
+    __host__ __device__ static constexpr T Lowest() { return std::numeric_limits<T>::lowest(); }
+
+    __host__ __device__ static constexpr T QuietNaN()
+    {
+        return std::numeric_limits<T>::quiet_NaN();
+    }
+
+    __host__ __device__ static constexpr T Infinity() { return std::numeric_limits<T>::infinity(); }
+};
+
+template <>
+struct NumericLimits<half_t>
+{
+    static constexpr unsigned short binary_min    = 0x0400;
+    static constexpr unsigned short binary_max    = 0x7BFF;
+    static constexpr unsigned short binary_lowest = 0xFBFF;
+    static constexpr unsigned short binary_qnan   = 0x7FFF;
+
+    __host__ __device__ static constexpr half_t Min() { return bit_cast<half_t>(binary_min); }
+
+    __host__ __device__ static constexpr half_t Max() { return bit_cast<half_t>(binary_max); }
+
+    __host__ __device__ static constexpr half_t Lowest() { return bit_cast<half_t>(binary_lowest); }
+
+    __host__ __device__ static constexpr half_t QuietNaN() { return bit_cast<half_t>(binary_qnan); }
+};
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+template <>
+struct NumericLimits<int4_t>
+{
+    __host__ __device__ static constexpr int4_t Min() { return int4_t(-8); }
+
+    __host__ __device__ static constexpr int4_t Max() { return int4_t(7); }
+
+    __host__ __device__ static constexpr int4_t Lowest() { return int4_t(-8); }
+};
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/debug.hpp b/3rdparty/composable_kernel/include/ck/utility/debug.hpp
new file mode 100644
index 0000000..593bbb7
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/debug.hpp
@@ -0,0 +1,85 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef UTILITY_DEBUG_HPP
+#define UTILITY_DEBUG_HPP
+
+namespace ck {
+namespace debug {
+
+namespace detail {
+template <typename T, typename Enable = void>
+struct PrintAsType;
+
+template <typename T>
+struct PrintAsType<T, typename std::enable_if<std::is_floating_point<T>::value>::type>
+{
+    using type = float;
+    __host__ __device__ static void Print(const T& p) { printf("%.3f ", static_cast<type>(p)); }
+};
+
+template <>
+struct PrintAsType<ck::half_t, void>
+{
+    using type = float;
+    __host__ __device__ static void Print(const ck::half_t& p)
+    {
+        printf("%.3f ", static_cast<type>(p));
+    }
+};
+
+template <typename T>
+struct PrintAsType<T, typename std::enable_if<std::is_integral<T>::value>::type>
+{
+    using type = int;
+    __host__ __device__ static void Print(const T& p) { printf("%d ", static_cast<type>(p)); }
+};
+} // namespace detail
+
+// Print at runtime the data in shared memory in 128 bytes per row format given shared mem pointer
+// and the number of elements. Can optionally specify strides between elements and how many bytes'
+// worth of data per row.
+//
+// Usage example:
+//
+//   debug::print_shared(a_block_buf.p_data_, index_t(a_block_desc_k0_m_k1.GetElementSpaceSize()));
+//
+template <typename T, index_t element_stride = 1, index_t row_bytes = 128>
+__device__ void print_shared(T const* p_shared, index_t num_elements)
+{
+    constexpr index_t row_elements = row_bytes / sizeof(T);
+    static_assert((element_stride >= 1 && element_stride <= row_elements),
+                  "element_stride should between [1, row_elements]");
+
+    index_t wgid = blockIdx.x + blockIdx.y * gridDim.x + gridDim.x * gridDim.y * blockIdx.z;
+    index_t tid =
+        (threadIdx.z * (blockDim.x * blockDim.y)) + (threadIdx.y * blockDim.x) + threadIdx.x;
+
+    __syncthreads();
+
+    if(tid == 0)
+    {
+        printf("\nWorkgroup id %d, bytes per row %d, element stride %d\n\n",
+               wgid,
+               row_bytes,
+               element_stride);
+        for(index_t i = 0; i < num_elements; i += row_elements)
+        {
+            printf("elem %5d: ", i);
+            for(index_t j = 0; j < row_elements; j += element_stride)
+            {
+                detail::PrintAsType<T>::Print(p_shared[i + j]);
+            }
+
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    __syncthreads();
+}
+
+} // namespace debug
+} // namespace ck
+
+#endif // UTILITY_DEBUG_HPP
diff --git a/3rdparty/composable_kernel/include/ck/utility/dynamic_buffer.hpp b/3rdparty/composable_kernel/include/ck/utility/dynamic_buffer.hpp
new file mode 100644
index 0000000..c6f0d29
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/dynamic_buffer.hpp
@@ -0,0 +1,398 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/data_type.hpp"
+#include "enable_if.hpp"
+#include "c_style_pointer_cast.hpp"
+#include "amd_buffer_addressing.hpp"
+#include "generic_memory_space_atomic.hpp"
+
+namespace ck {
+
+// T may be scalar or vector
+// X may be scalar or vector
+// T and X have same scalar type
+// X contains multiple T
+template <AddressSpaceEnum BufferAddressSpace,
+          typename T,
+          typename ElementSpaceSize,
+          bool InvalidElementUseNumericalZeroValue>
+struct DynamicBuffer
+{
+    using type = T;
+
+    T* p_data_;
+    ElementSpaceSize element_space_size_;
+    T invalid_element_value_ = T{0};
+
+    __host__ __device__ constexpr DynamicBuffer(T* p_data, ElementSpaceSize element_space_size)
+        : p_data_{p_data}, element_space_size_{element_space_size}
+    {
+    }
+
+    __host__ __device__ constexpr DynamicBuffer(T* p_data,
+                                                ElementSpaceSize element_space_size,
+                                                T invalid_element_value)
+        : p_data_{p_data},
+          element_space_size_{element_space_size},
+          invalid_element_value_{invalid_element_value}
+    {
+    }
+
+    __host__ __device__ static constexpr AddressSpaceEnum GetAddressSpace()
+    {
+        return BufferAddressSpace;
+    }
+
+    __host__ __device__ constexpr const T& operator[](index_t i) const { return p_data_[i]; }
+
+    __host__ __device__ constexpr T& operator()(index_t i) { return p_data_[i]; }
+
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __host__ __device__ constexpr auto Get(index_t i, bool is_valid_element) const
+    {
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+#if CK_USE_AMD_BUFFER_LOAD
+        bool constexpr use_amd_buffer_addressing = true;
+#else
+        bool constexpr use_amd_buffer_addressing = false;
+#endif
+
+        if constexpr(GetAddressSpace() == AddressSpaceEnum::Global && use_amd_buffer_addressing)
+        {
+            constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
+
+            if constexpr(InvalidElementUseNumericalZeroValue)
+            {
+                return amd_buffer_load_invalid_element_return_zero<remove_cvref_t<T>, t_per_x>(
+                    p_data_, i, is_valid_element, element_space_size_);
+            }
+            else
+            {
+                return amd_buffer_load_invalid_element_return_customized_value<remove_cvref_t<T>,
+                                                                               t_per_x>(
+                    p_data_, i, is_valid_element, element_space_size_, invalid_element_value_);
+            }
+        }
+        else
+        {
+            if(is_valid_element)
+            {
+#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
+                X tmp;
+
+                __builtin_memcpy(&tmp, &(p_data_[i]), sizeof(X));
+
+                return tmp;
+#else
+                return *c_style_pointer_cast<const X*>(&p_data_[i]);
+#endif
+            }
+            else
+            {
+                if constexpr(InvalidElementUseNumericalZeroValue)
+                {
+                    return X{0};
+                }
+                else
+                {
+                    return X{invalid_element_value_};
+                }
+            }
+        }
+    }
+
+    template <InMemoryDataOperationEnum Op,
+              typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __host__ __device__ void Update(index_t i, bool is_valid_element, const X& x)
+    {
+        if constexpr(Op == InMemoryDataOperationEnum::Set)
+        {
+            this->template Set<X>(i, is_valid_element, x);
+        }
+        else if constexpr(Op == InMemoryDataOperationEnum::AtomicAdd)
+        {
+            this->template AtomicAdd<X>(i, is_valid_element, x);
+        }
+        else if constexpr(Op == InMemoryDataOperationEnum::AtomicMax)
+        {
+            this->template AtomicMax<X>(i, is_valid_element, x);
+        }
+        else if constexpr(Op == InMemoryDataOperationEnum::Add)
+        {
+            auto tmp = this->template Get<X>(i, is_valid_element);
+            this->template Set<X>(i, is_valid_element, x + tmp);
+            // tmp += x;
+            // this->template Set<X>(i, is_valid_element, tmp);
+        }
+    }
+
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __host__ __device__ void Set(index_t i, bool is_valid_element, const X& x)
+    {
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+#if CK_USE_AMD_BUFFER_STORE
+        bool constexpr use_amd_buffer_addressing = true;
+#else
+        bool constexpr use_amd_buffer_addressing      = false;
+#endif
+
+#if CK_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE
+        bool constexpr workaround_int8_ds_write_issue = true;
+#else
+        bool constexpr workaround_int8_ds_write_issue = false;
+#endif
+
+        if constexpr(GetAddressSpace() == AddressSpaceEnum::Global && use_amd_buffer_addressing)
+        {
+            constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
+
+            amd_buffer_store<remove_cvref_t<T>, t_per_x>(
+                x, p_data_, i, is_valid_element, element_space_size_);
+        }
+        else if constexpr(GetAddressSpace() == AddressSpaceEnum::Lds &&
+                          is_same<typename scalar_type<remove_cvref_t<T>>::type, int8_t>::value &&
+                          workaround_int8_ds_write_issue)
+        {
+            if(is_valid_element)
+            {
+                // HACK: compiler would lower IR "store<i8, 16> address_space(3)" into inefficient
+                // ISA, so I try to let compiler emit IR "store<i32, 4>" which would be lower to
+                // ds_write_b128
+                // TODO: remove this after compiler fix
+                static_assert((is_same<remove_cvref_t<T>, int8_t>::value &&
+                               is_same<remove_cvref_t<X>, int8_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x2_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x4_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x8_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x16_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8x4_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x4_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8x8_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x8_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8x16_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x16_t>::value),
+                              "wrong! not implemented for this combination, please add "
+                              "implementation");
+
+                if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
+                             is_same<remove_cvref_t<X>, int8_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int8_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int8_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x2_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int16_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int16_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x4_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int32_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int32_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x8_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int32x2_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int32x2_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x16_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int32x4_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int32x4_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8x4_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x4_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int32_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int32_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8x8_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x8_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int32x2_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int32x2_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8x16_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x16_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int32x4_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int32x4_t*>(&x);
+                }
+            }
+        }
+        else
+        {
+            if(is_valid_element)
+            {
+#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
+                X tmp = x;
+
+                __builtin_memcpy(&(p_data_[i]), &tmp, sizeof(X));
+#else
+                *c_style_pointer_cast<X*>(&p_data_[i]) = x;
+#endif
+            }
+        }
+    }
+
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __host__ __device__ void AtomicAdd(index_t i, bool is_valid_element, const X& x)
+    {
+        using scalar_t = typename scalar_type<remove_cvref_t<T>>::type;
+
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+        static_assert(GetAddressSpace() == AddressSpaceEnum::Global, "only support global mem");
+
+#if CK_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER && CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT
+        bool constexpr use_amd_buffer_addressing =
+            is_same_v<remove_cvref_t<scalar_t>, int32_t> ||
+            is_same_v<remove_cvref_t<scalar_t>, float> ||
+            (is_same_v<remove_cvref_t<scalar_t>, half_t> && scalar_per_x_vector % 2 == 0);
+#elif CK_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER && (!CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT)
+        bool constexpr use_amd_buffer_addressing = is_same_v<remove_cvref_t<scalar_t>, int32_t>;
+#elif(!CK_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER) && CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT
+        bool constexpr use_amd_buffer_addressing =
+            is_same_v<remove_cvref_t<scalar_t>, float> ||
+            (is_same_v<remove_cvref_t<scalar_t>, half_t> && scalar_per_x_vector % 2 == 0);
+#else
+        bool constexpr use_amd_buffer_addressing = false;
+#endif
+
+        if constexpr(use_amd_buffer_addressing)
+        {
+            constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
+
+            amd_buffer_atomic_add<remove_cvref_t<T>, t_per_x>(
+                x, p_data_, i, is_valid_element, element_space_size_);
+        }
+        else
+        {
+            if(is_valid_element)
+            {
+                atomic_add<X>(c_style_pointer_cast<X*>(&p_data_[i]), x);
+            }
+        }
+    }
+
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __host__ __device__ void AtomicMax(index_t i, bool is_valid_element, const X& x)
+    {
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+        static_assert(GetAddressSpace() == AddressSpaceEnum::Global, "only support global mem");
+
+#if CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64
+        using scalar_t                           = typename scalar_type<remove_cvref_t<T>>::type;
+        bool constexpr use_amd_buffer_addressing = is_same_v<remove_cvref_t<scalar_t>, double>;
+#else
+        bool constexpr use_amd_buffer_addressing = false;
+#endif
+
+        if constexpr(use_amd_buffer_addressing)
+        {
+            constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
+
+            amd_buffer_atomic_max<remove_cvref_t<T>, t_per_x>(
+                x, p_data_, i, is_valid_element, element_space_size_);
+        }
+        else if(is_valid_element)
+        {
+            atomic_max<X>(c_style_pointer_cast<X*>(&p_data_[i]), x);
+        }
+    }
+
+    __host__ __device__ static constexpr bool IsStaticBuffer() { return false; }
+
+    __host__ __device__ static constexpr bool IsDynamicBuffer() { return true; }
+};
+
+template <AddressSpaceEnum BufferAddressSpace, typename T, typename ElementSpaceSize>
+__host__ __device__ constexpr auto make_dynamic_buffer(T* p, ElementSpaceSize element_space_size)
+{
+    return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, true>{p, element_space_size};
+}
+
+template <
+    AddressSpaceEnum BufferAddressSpace,
+    typename T,
+    typename ElementSpaceSize,
+    typename X,
+    typename enable_if<is_same<remove_cvref_t<T>, remove_cvref_t<X>>::value, bool>::type = false>
+__host__ __device__ constexpr auto
+make_dynamic_buffer(T* p, ElementSpaceSize element_space_size, X invalid_element_value)
+{
+    return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, false>{
+        p, element_space_size, invalid_element_value};
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/enable_if.hpp b/3rdparty/composable_kernel/include/ck/utility/enable_if.hpp
new file mode 100644
index 0000000..297434b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/enable_if.hpp
@@ -0,0 +1,14 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+template <bool B, typename T = void>
+using enable_if = std::enable_if<B, T>;
+
+template <bool B, typename T = void>
+using enable_if_t = typename std::enable_if<B, T>::type;
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/functional.hpp b/3rdparty/composable_kernel/include/ck/utility/functional.hpp
new file mode 100644
index 0000000..08e7307
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/functional.hpp
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/integral_constant.hpp"
+#include "ck/utility/type.hpp"
+
+namespace ck {
+
+// TODO: right? wrong?
+struct forwarder
+{
+    template <typename T>
+    __host__ __device__ constexpr T&& operator()(T&& x) const
+    {
+        return static_cast<T&&>(x);
+    }
+};
+
+struct swallow
+{
+    template <typename... Ts>
+    __host__ __device__ constexpr swallow(Ts&&...)
+    {
+    }
+};
+
+template <typename T>
+struct logical_and
+{
+    constexpr bool operator()(const T& x, const T& y) const { return x && y; }
+};
+
+template <typename T>
+struct logical_or
+{
+    constexpr bool operator()(const T& x, const T& y) const { return x || y; }
+};
+
+template <typename T>
+struct logical_not
+{
+    constexpr bool operator()(const T& x) const { return !x; }
+};
+
+// Emulate if constexpr
+template <bool>
+struct static_if;
+
+template <>
+struct static_if<true>
+{
+    using Type = static_if<true>;
+
+    template <typename F>
+    __host__ __device__ constexpr auto operator()(F f) const
+    {
+        // This is a trick for compiler:
+        //   Pass forwarder to lambda "f" as "auto" argument, and make sure "f" will
+        //   use it,
+        //   this will make "f" a generic lambda, so that "f" won't be compiled
+        //   until being
+        //   instantiated here
+        f(forwarder{});
+        return Type{};
+    }
+
+    template <typename F>
+    __host__ __device__ static void Else(F)
+    {
+    }
+};
+
+template <>
+struct static_if<false>
+{
+    using Type = static_if<false>;
+
+    template <typename F>
+    __host__ __device__ constexpr auto operator()(F) const
+    {
+        return Type{};
+    }
+
+    template <typename F>
+    __host__ __device__ static void Else(F f)
+    {
+        // This is a trick for compiler:
+        //   Pass forwarder to lambda "f" as "auto" argument, and make sure "f" will
+        //   use it,
+        //   this will make "f" a generic lambda, so that "f" won't be compiled
+        //   until being
+        //   instantiated here
+        f(forwarder{});
+    }
+};
+
+template <bool predicate, class X, class Y>
+struct conditional;
+
+template <class X, class Y>
+struct conditional<true, X, Y>
+{
+    using type = X;
+};
+
+template <class X, class Y>
+struct conditional<false, X, Y>
+{
+    using type = Y;
+};
+
+template <bool predicate, class X, class Y>
+using conditional_t = typename conditional<predicate, X, Y>::type;
+
+// z = predicate ? x : y
+template <bool predicate, typename X, typename Y>
+constexpr auto conditional_expr(X&& x, Y&& y)
+{
+    if constexpr(predicate)
+    {
+        return std::forward<X>(x);
+    }
+    else
+    {
+        return std::forward<Y>(y);
+    }
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/functional2.hpp b/3rdparty/composable_kernel/include/ck/utility/functional2.hpp
new file mode 100644
index 0000000..6f125ca
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/functional2.hpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/functional.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+
+namespace detail {
+
+template <class>
+struct static_for_impl;
+
+template <index_t... Is>
+struct static_for_impl<Sequence<Is...>>
+{
+    template <class F>
+    __host__ __device__ constexpr void operator()(F f) const
+    {
+        swallow{(f(Number<Is>{}), 0)...};
+    }
+};
+
+} // namespace detail
+
+// F signature: F(Number<Iter>)
+template <index_t NBegin, index_t NEnd, index_t Increment>
+struct static_for
+{
+    __host__ __device__ constexpr static_for()
+    {
+        static_assert(Increment != 0 && (NEnd - NBegin) % Increment == 0,
+                      "Wrong! should satisfy (NEnd - NBegin) % Increment == 0");
+        static_assert((Increment > 0 && NBegin <= NEnd) || (Increment < 0 && NBegin >= NEnd),
+                      "wrongs! should (Increment > 0 && NBegin <= NEnd) || (Increment < 0 && "
+                      "NBegin >= NEnd)");
+    }
+
+    template <class F>
+    __host__ __device__ constexpr void operator()(F f) const
+    {
+        detail::static_for_impl<typename arithmetic_sequence_gen<NBegin, NEnd, Increment>::type>{}(
+            f);
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/functional3.hpp b/3rdparty/composable_kernel/include/ck/utility/functional3.hpp
new file mode 100644
index 0000000..06b67ef
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/functional3.hpp
@@ -0,0 +1,144 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/functional.hpp"
+#include "ck/utility/functional2.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/multi_index.hpp"
+
+namespace ck {
+
+namespace detail {
+
+// RemainLengths: Sequence<...>
+// Orders: Sequence<...>
+template <class RemainLengths, class Orders>
+struct static_ford_impl
+{
+    __host__ __device__ constexpr static_ford_impl()
+    {
+        static_assert(RemainLengths::GetSize() > 0, "wrong! should not get here");
+    }
+
+    // F signature: F(Sequence<...>)
+    // CurrentOrderedId: Sequence<...>
+    template <class F, class CurrentOrderedId>
+    __host__ __device__ constexpr void operator()(F f, CurrentOrderedId) const
+    {
+        static_for<0, RemainLengths::Front(), 1>{}([=](auto I) {
+            static_ford_impl<decltype(RemainLengths::PopFront()), Orders>{}(
+                f, CurrentOrderedId::PushBack(I));
+        });
+    }
+};
+
+template <class Orders>
+struct static_ford_impl<Sequence<>, Orders>
+{
+    // F signature: F(Sequence<...>)
+    // OrderedId: Sequence<...>
+    template <class F, class OrderedId>
+    __host__ __device__ constexpr void operator()(F f, OrderedId) const
+    {
+        // retrive unordered Id
+        f(OrderedId::ReorderGivenOld2New(Orders{}));
+    }
+};
+
+// RemainLengths: Sequence<...>
+// Orders: Sequence<...>
+template <class RemainLengths, class Orders>
+struct ford_impl
+{
+    __host__ __device__ constexpr ford_impl()
+    {
+        static_assert(RemainLengths::GetSize() > 0, "wrong! should not get here");
+    }
+
+    // F signature: F(Array<...> multi_id)
+    // CurrentOrderdId: Array<...>
+    template <class F, class CurrentOrderedId>
+    __host__ __device__ constexpr void operator()(F f, CurrentOrderedId current_ordered_id) const
+    {
+        for(index_t i = 0; i < RemainLengths::Front(); ++i)
+        {
+            ford_impl<decltype(RemainLengths::PopFront()), Orders>{}(
+                f, container_push_back(current_ordered_id, i));
+        }
+    }
+};
+
+template <class Orders>
+struct ford_impl<Sequence<>, Orders>
+{
+    // F signature: F(Array<...> multi_id)
+    // CurrentOrderdId: Array<...>
+    template <class F, class CurrentOrderedId>
+    __host__ __device__ constexpr void operator()(F f, CurrentOrderedId current_ordered_id) const
+    {
+        // retrive unordered Id
+        f(container_reorder_given_old2new(current_ordered_id, Orders{}));
+    }
+};
+
+} // namespace detail
+
+// Lengths is Sequence<...>, it is the length of each dimension for
+// N-dimensional loop
+// Orders is Sequence<...>, it is the order of dimension in which static_ford
+// will loop over each
+// dimension
+template <class Lengths,
+          class Orders = typename arithmetic_sequence_gen<0, Lengths::GetSize(), 1>::type>
+struct static_ford
+{
+    __host__ __device__ constexpr static_ford()
+    {
+        static_assert(Lengths::GetSize() > 0, "wrong! Lengths is empty");
+        static_assert(Lengths::GetSize() == Orders::GetSize(), "wrong! inconsistent size");
+    }
+
+    // F signature: F(Sequence<...> multi_id)
+    // multi_id is the unordered multi-index
+    template <class F>
+    __host__ __device__ constexpr void operator()(F f) const
+    {
+        constexpr auto ordered_lengths = Lengths::ReorderGivenNew2Old(Orders{});
+        detail::static_ford_impl<decltype(ordered_lengths), Orders>{}(f, Sequence<>{});
+    }
+};
+
+// Lengths is Sequence<...>, it is the length of each dimension for
+// N-dimensional loop
+// Orders is Sequence<...>, it is the order of dimension in which ford will loop
+// over each
+// dimension
+template <class Lengths,
+          class Orders = typename arithmetic_sequence_gen<0, Lengths::GetSize(), 1>::type>
+struct ford
+{
+    __host__ __device__ constexpr ford()
+    {
+        static_assert(Lengths::GetSize() > 0, "wrong! Lengths is empty");
+        static_assert(Lengths::GetSize() == Orders::GetSize(), "wrong! inconsistent size");
+    }
+
+    // F signature: F(Array<...> multi_id)
+    // multi_id is the unordered multi-index
+    template <class F>
+    __host__ __device__ constexpr void operator()(F f) const
+    {
+        constexpr auto ordered_lengths = Lengths::ReorderGivenNew2Old(Orders{});
+
+        for(index_t i = 0; i < ordered_lengths.Front(); ++i)
+        {
+            detail::ford_impl<decltype(ordered_lengths.PopFront()), Orders>{}(f,
+                                                                              make_multi_index(i));
+        }
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/functional4.hpp b/3rdparty/composable_kernel/include/ck/utility/functional4.hpp
new file mode 100644
index 0000000..6eeaf15
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/functional4.hpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_FUNCTIONAL4_HPP
+#define CK_FUNCTIONAL4_HPP
+
+#include "sequence.hpp"
+#include "tuple.hpp"
+#include "array.hpp"
+
+namespace ck {
+
+namespace detail {
+
+template <typename Indices>
+struct unpack_impl;
+
+template <index_t... Is>
+struct unpack_impl<Sequence<Is...>>
+{
+    template <typename F, typename X>
+    __host__ __device__ constexpr auto operator()(F&& f, X&& x) const
+    {
+        return std::forward<F>(f)(std::forward<X>(x).At(Number<Is>{})...);
+    }
+};
+
+template <typename Seq0, typename Seq1>
+struct unpack2_impl;
+
+// TODO: remove this, after properly implementing unpack that takes any number of containers
+template <index_t... Is, index_t... Js>
+struct unpack2_impl<Sequence<Is...>, Sequence<Js...>>
+{
+    template <typename F, typename X, typename Y>
+    __host__ __device__ constexpr auto operator()(F&& f, X&& x, Y&& y) const
+    {
+        return std::forward<F>(f)(std::forward<X>(x).At(Number<Is>{})...,
+                                  std::forward<Y>(y).At(Number<Js>{})...);
+    }
+};
+
+} // namespace detail
+
+template <typename F, typename X>
+__host__ __device__ constexpr auto unpack(F&& f, X&& x)
+{
+    using X_ = remove_reference_t<X>;
+    return detail::unpack_impl<typename arithmetic_sequence_gen<0, X_::Size(), 1>::type>{}(
+        std::forward<F>(f), std::forward<X>(x));
+}
+
+// TODO: properly implement unpack that takes any number of containers
+template <typename F, typename X, typename Y>
+__host__ __device__ constexpr auto unpack2(F&& f, X&& x, Y&& y)
+{
+    using X_ = remove_reference_t<X>;
+    using Y_ = remove_reference_t<Y>;
+    return detail::unpack2_impl<typename arithmetic_sequence_gen<0, X_::Size(), 1>::type,
+                                typename arithmetic_sequence_gen<0, Y_::Size(), 1>::type>{}(
+        std::forward<F>(f), std::forward<X>(x), std::forward<Y>(y));
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/generic_memory_space_atomic.hpp b/3rdparty/composable_kernel/include/ck/utility/generic_memory_space_atomic.hpp
new file mode 100644
index 0000000..6a1ca96
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/generic_memory_space_atomic.hpp
@@ -0,0 +1,123 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "data_type.hpp"
+
+namespace ck {
+
+// Caution: DO NOT REMOVE
+// intentionally have only declaration but no definition to cause compilation failure when trying to
+// instantiate this template. The purpose is to make the implementation of atomic_add explicit for
+// each datatype.
+template <typename X>
+__device__ X atomic_add(X* p_dst, const X& x);
+
+template <>
+__device__ int32_t atomic_add<int32_t>(int32_t* p_dst, const int32_t& x)
+{
+    return atomicAdd(p_dst, x);
+}
+
+template <>
+__device__ uint32_t atomic_add<uint32_t>(uint32_t* p_dst, const uint32_t& x)
+{
+    return atomicAdd(p_dst, x);
+}
+
+template <>
+__device__ float atomic_add<float>(float* p_dst, const float& x)
+{
+    return atomicAdd(p_dst, x);
+}
+
+template <>
+__device__ double atomic_add<double>(double* p_dst, const double& x)
+{
+    return atomicAdd(p_dst, x);
+}
+
+template <>
+__device__ float2_t atomic_add<float2_t>(float2_t* p_dst, const float2_t& x)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    const vector_type<float, 2> vx{x};
+    vector_type<float, 2> vy{0};
+
+    vy.template AsType<float>()(I0) =
+        atomicAdd(c_style_pointer_cast<float*>(p_dst), vx.template AsType<float>()[I0]);
+    vy.template AsType<float>()(I1) =
+        atomicAdd(c_style_pointer_cast<float*>(p_dst) + 1, vx.template AsType<float>()[I1]);
+
+    return vy.template AsType<float2_t>()[I0];
+}
+
+template <>
+__device__ double2_t atomic_add<double2_t>(double2_t* p_dst, const double2_t& x)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    const vector_type<double, 2> vx{x};
+    vector_type<double, 2> vy{0};
+
+    vy.template AsType<double>()(I0) =
+        atomicAdd(c_style_pointer_cast<double*>(p_dst), vx.template AsType<double>()[I0]);
+    vy.template AsType<double>()(I1) =
+        atomicAdd(c_style_pointer_cast<double*>(p_dst) + 1, vx.template AsType<double>()[I1]);
+
+    return vy.template AsType<double2_t>()[I0];
+}
+
+// Caution: DO NOT REMOVE
+// intentionally have only declaration but no definition to cause compilation failure when trying to
+// instantiate this template. The purpose is to make the implementation of atomic_max explicit for
+// each datatype.
+
+template <typename X>
+__device__ X atomic_max(X* p_dst, const X& x);
+
+template <>
+__device__ int32_t atomic_max<int32_t>(int32_t* p_dst, const int32_t& x)
+{
+    return atomicMax(p_dst, x);
+}
+
+template <>
+__device__ uint32_t atomic_max<uint32_t>(uint32_t* p_dst, const uint32_t& x)
+{
+    return atomicMax(p_dst, x);
+}
+
+template <>
+__device__ float atomic_max<float>(float* p_dst, const float& x)
+{
+    return atomicMax(p_dst, x);
+}
+
+template <>
+__device__ double atomic_max<double>(double* p_dst, const double& x)
+{
+    return atomicMax(p_dst, x);
+}
+
+template <>
+__device__ float2_t atomic_max<float2_t>(float2_t* p_dst, const float2_t& x)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    const vector_type<float, 2> vx{x};
+    vector_type<float, 2> vy{0};
+
+    vy.template AsType<float>()(I0) =
+        atomicMax(c_style_pointer_cast<float*>(p_dst), vx.template AsType<float>()[I0]);
+    vy.template AsType<float>()(I1) =
+        atomicMax(c_style_pointer_cast<float*>(p_dst) + 1, vx.template AsType<float>()[I1]);
+
+    return vy.template AsType<float2_t>()[I0];
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/get_id.hpp b/3rdparty/composable_kernel/include/ck/utility/get_id.hpp
new file mode 100644
index 0000000..44ff438
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/get_id.hpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+
+namespace ck {
+
+__host__ __device__ constexpr index_t get_warp_size()
+{
+    // warpSize is defined by HIP
+    return warpSize;
+}
+
+__device__ index_t get_thread_local_1d_id() { return threadIdx.x; }
+
+__device__ index_t get_thread_global_1d_id() { return blockIdx.x * blockDim.x + threadIdx.x; }
+
+__device__ index_t get_warp_local_1d_id() { return threadIdx.x / get_warp_size(); }
+
+__device__ index_t get_block_1d_id() { return blockIdx.x; }
+
+__device__ index_t get_grid_size() { return gridDim.x; }
+
+__device__ index_t get_block_size() { return blockDim.x; }
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/ignore.hpp b/3rdparty/composable_kernel/include/ck/utility/ignore.hpp
new file mode 100644
index 0000000..ac33cbf
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/ignore.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+// https://en.cppreference.com/w/cpp/utility/tuple/ignore
+
+namespace ck {
+
+namespace detail {
+struct ignore_t
+{
+    template <typename T>
+    constexpr void operator=(T&&) const noexcept
+    {
+    }
+};
+} // namespace detail
+
+inline constexpr detail::ignore_t ignore;
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/inner_product.hpp b/3rdparty/composable_kernel/include/ck/utility/inner_product.hpp
new file mode 100644
index 0000000..5da1ec9
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/inner_product.hpp
@@ -0,0 +1,211 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "data_type.hpp"
+
+namespace ck {
+
+template <typename TA, typename TB, typename TC>
+__device__ void inner_product(const TA& a, const TB& b, TC& c);
+
+template <>
+__device__ void inner_product<float, float, float>(const float& a, const float& b, float& c)
+{
+#if CK_USE_AMD_INNER_PRODUCT_INLINE_ASM && defined(CK_USE_AMD_V_MAC_F32)
+    asm volatile("\n \
+            v_mac_f32 %0, %1, %2 \n \
+            "
+                 : "=v"(c)
+                 : "v"(a), "v"(b), "0"(c));
+#elif CK_USE_AMD_INNER_PRODUCT_INLINE_ASM && defined(CK_USE_AMD_V_FMAC_F32)
+    asm volatile("\n \
+            v_fmac_f32 %0, %1, %2 \n \
+            "
+                 : "=v"(c)
+                 : "v"(a), "v"(b), "0"(c));
+#else
+    c += a * b;
+#endif
+}
+
+template <>
+__device__ void
+inner_product<float2_t, float2_t, float>(const float2_t& a, const float2_t& b, float& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    inner_product(vector_type<float, 2>{a}.AsType<float>()[I0],
+                  vector_type<float, 2>{b}.AsType<float>()[I0],
+                  c);
+
+    inner_product(vector_type<float, 2>{a}.AsType<float>()[I1],
+                  vector_type<float, 2>{b}.AsType<float>()[I1],
+                  c);
+}
+
+template <>
+__device__ void
+inner_product<float4_t, float4_t, float>(const float4_t& a, const float4_t& b, float& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    inner_product(vector_type<float, 4>{a}.AsType<float>()[I0],
+                  vector_type<float, 4>{b}.AsType<float>()[I0],
+                  c);
+
+    inner_product(vector_type<float, 4>{a}.AsType<float>()[I1],
+                  vector_type<float, 4>{b}.AsType<float>()[I1],
+                  c);
+
+    inner_product(vector_type<float, 4>{a}.AsType<float>()[I2],
+                  vector_type<float, 4>{b}.AsType<float>()[I2],
+                  c);
+
+    inner_product(vector_type<float, 4>{a}.AsType<float>()[I3],
+                  vector_type<float, 4>{b}.AsType<float>()[I3],
+                  c);
+}
+
+template <>
+__device__ void inner_product<half_t, half_t, float>(const half_t& a, const half_t& b, float& c)
+{
+  c+=static_cast<float>(a*b);
+}
+
+template <>
+__device__ void inner_product<half2_t, half2_t, float>(const half2_t& a, const half2_t& b, float& c)
+{
+#if defined(CK_USE_AMD_V_DOT2_F32_F16)
+#if CK_USE_AMD_INNER_PRODUCT_INLINE_ASM
+    asm volatile("\n \
+            v_dot2_f32_f16 %0, %1, %2, %0\n \
+            "
+                 : "=v"(c)
+                 : "v"(a), "v"(b), "0"(c));
+#else
+    c = __builtin_amdgcn_sdot2(a, b, c, false);
+#endif
+#else
+    const vector_type<half_t, 2> a_vector{a};
+    const vector_type<half_t, 2> b_vector{b};
+
+    static_for<0, 2, 1>{}([&](auto i) {
+        c += type_convert<int32_t>(a_vector.AsType<half_t>()[i]) *
+             type_convert<int32_t>(b_vector.AsType<half_t>()[i]);
+    });
+#endif
+}
+
+template <>
+__device__ void inner_product<half4_t, half4_t, float>(const half4_t& a, const half4_t& b, float& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    inner_product(vector_type<half_t, 4>{a}.AsType<half2_t>()[I0],
+                  vector_type<half_t, 4>{b}.AsType<half2_t>()[I0],
+                  c);
+
+    inner_product(vector_type<half_t, 4>{a}.AsType<half2_t>()[I1],
+                  vector_type<half_t, 4>{b}.AsType<half2_t>()[I1],
+                  c);
+}
+
+template <>
+__device__ void inner_product<half8_t, half8_t, float>(const half8_t& a, const half8_t& b, float& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    inner_product(vector_type<half_t, 8>{a}.AsType<half2_t>()[I0],
+                  vector_type<half_t, 8>{b}.AsType<half2_t>()[I0],
+                  c);
+
+    inner_product(vector_type<half_t, 8>{a}.AsType<half2_t>()[I1],
+                  vector_type<half_t, 8>{b}.AsType<half2_t>()[I1],
+                  c);
+
+    inner_product(vector_type<half_t, 8>{a}.AsType<half2_t>()[I2],
+                  vector_type<half_t, 8>{b}.AsType<half2_t>()[I2],
+                  c);
+
+    inner_product(vector_type<half_t, 8>{a}.AsType<half2_t>()[I3],
+                  vector_type<half_t, 8>{b}.AsType<half2_t>()[I3],
+                  c);
+}
+
+template <>
+__device__ void
+inner_product<int8x4_t, int8x4_t, int32_t>(const int8x4_t& a, const int8x4_t& b, int32_t& c)
+{
+#if defined(CK_USE_AMD_V_DOT4_I32_I8)
+#if CK_USE_AMD_INNER_PRODUCT_INLINE_ASM
+    asm volatile("\n \
+            v_dot4_i32_i8 %0, %1, %2, %0\n \
+            "
+                 : "=v"(c)
+                 : "v"(bit_cast<int32_t>(a)), "v"(bit_cast<int32_t>(b)), "0"(c));
+#else
+    c = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b), c, false);
+#endif
+#else
+    const vector_type<int8_t, 4> a_vector{a};
+    const vector_type<int8_t, 4> b_vector{b};
+
+    static_for<0, 4, 1>{}([&](auto i) {
+        c += type_convert<int32_t>(a_vector.AsType<int8_t>()[i]) *
+             type_convert<int32_t>(b_vector.AsType<int8_t>()[i]);
+    });
+#endif
+}
+
+template <>
+__device__ void
+inner_product<int8x8_t, int8x8_t, int32_t>(const int8x8_t& a, const int8x8_t& b, int32_t& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    inner_product(vector_type<int8_t, 8>{a}.AsType<int8x4_t>()[I0],
+                  vector_type<int8_t, 8>{b}.AsType<int8x4_t>()[I0],
+                  c);
+
+    inner_product(vector_type<int8_t, 8>{a}.AsType<int8x4_t>()[I1],
+                  vector_type<int8_t, 8>{b}.AsType<int8x4_t>()[I1],
+                  c);
+}
+
+template <>
+__device__ void
+inner_product<int8x16_t, int8x16_t, int32_t>(const int8x16_t& a, const int8x16_t& b, int32_t& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    inner_product(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I0],
+                  vector_type<int8_t, 16>{b}.AsType<int8x4_t>()[I0],
+                  c);
+
+    inner_product(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I1],
+                  vector_type<int8_t, 16>{b}.AsType<int8x4_t>()[I1],
+                  c);
+
+    inner_product(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I2],
+                  vector_type<int8_t, 16>{b}.AsType<int8x4_t>()[I2],
+                  c);
+
+    inner_product(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I3],
+                  vector_type<int8_t, 16>{b}.AsType<int8x4_t>()[I3],
+                  c);
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/integral_constant.hpp b/3rdparty/composable_kernel/include/ck/utility/integral_constant.hpp
new file mode 100644
index 0000000..9aab4e2
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/integral_constant.hpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+template <class T, T v>
+struct integral_constant
+{
+    static constexpr T value = v;
+    typedef T value_type;
+    typedef integral_constant type;
+    __host__ __device__ constexpr operator value_type() const noexcept { return value; }
+    __host__ __device__ constexpr value_type operator()() const noexcept { return value; }
+};
+
+template <typename TX, TX X, typename TY, TY Y>
+__host__ __device__ constexpr auto operator+(integral_constant<TX, X>, integral_constant<TY, Y>)
+{
+    return integral_constant<decltype(X + Y), X + Y>{};
+}
+
+template <typename TX, TX X, typename TY, TY Y>
+__host__ __device__ constexpr auto operator-(integral_constant<TX, X>, integral_constant<TY, Y>)
+{
+    static_assert(Y <= X, "wrong!");
+    return integral_constant<decltype(X - Y), X - Y>{};
+}
+
+template <typename TX, TX X, typename TY, TY Y>
+__host__ __device__ constexpr auto operator*(integral_constant<TX, X>, integral_constant<TY, Y>)
+{
+    return integral_constant<decltype(X * Y), X * Y>{};
+}
+
+template <typename TX, TX X, typename TY, TY Y>
+__host__ __device__ constexpr auto operator/(integral_constant<TX, X>, integral_constant<TY, Y>)
+{
+    static_assert(Y > 0, "wrong!");
+    return integral_constant<decltype(X / Y), X / Y>{};
+}
+
+template <typename TX, TX X, typename TY, TY Y>
+__host__ __device__ constexpr auto operator%(integral_constant<TX, X>, integral_constant<TY, Y>)
+{
+    static_assert(Y > 0, "wrong!");
+    return integral_constant<decltype(X % Y), X % Y>{};
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/is_known_at_compile_time.hpp b/3rdparty/composable_kernel/include/ck/utility/is_known_at_compile_time.hpp
new file mode 100644
index 0000000..8198154
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/is_known_at_compile_time.hpp
@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "integral_constant.hpp"
+#include "sequence.hpp"
+#include "tuple.hpp"
+
+namespace ck {
+
+template <typename T>
+struct is_known_at_compile_time;
+
+template <>
+struct is_known_at_compile_time<index_t>
+{
+    static constexpr bool value = false;
+};
+
+template <>
+struct is_known_at_compile_time<long_index_t>
+{
+    static constexpr bool value = false;
+};
+
+template <typename T, T X>
+struct is_known_at_compile_time<integral_constant<T, X>>
+{
+    static constexpr bool value = true;
+};
+
+template <index_t... Is>
+struct is_known_at_compile_time<Sequence<Is...>>
+{
+    static constexpr bool value = true;
+};
+
+template <typename... Ts>
+struct is_known_at_compile_time<Tuple<Ts...>>
+{
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return container_reduce(
+            Tuple<Ts...>{},
+            [](auto x, bool r) {
+                return is_known_at_compile_time<remove_cvref_t<decltype(x)>>::value & r;
+            },
+            true);
+    }
+
+    static constexpr bool value = IsKnownAtCompileTime();
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/magic_division.hpp b/3rdparty/composable_kernel/include/ck/utility/magic_division.hpp
new file mode 100644
index 0000000..a5e8e92
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/magic_division.hpp
@@ -0,0 +1,160 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "integral_constant.hpp"
+#include "number.hpp"
+#include "type.hpp"
+#include "tuple.hpp"
+
+namespace ck {
+
+// magic number division
+// Caution:
+//   1. For uint32_t as dividend: magic number division implementation being used would produce
+//   correct result if the dividend is uint32_t and its value is within 31-bit value range.
+//   2. For int32_t as dividendd: magic number division for int32_t dividened has not been
+//   implemented, the int32_t dividend would be bit-wise interpreted as uint32_t and magic number
+//   division implementation for uint32_t is then used. Therefore, dividend value need to be
+//   non-negative.
+// TODO:
+//   1. Implement magic number divison for int32_t
+//   2. Implement magic number divison for unit32_t with 32-bit value range
+struct MagicDivision
+{
+    // uint32_t
+    __host__ __device__ static constexpr auto CalculateMagicNumbers(uint32_t divisor)
+    {
+        // WARNING: magic division is only applicable for division inside this range.
+        // You should use the return value of CalculateMagicNumbers, if division is not inside this
+        // range. The "else" logic below is to quiet down run-time error.
+        if(divisor >= 1 && divisor <= INT32_MAX)
+        {
+            uint32_t shift = 0;
+            for(shift = 0; shift < 32; ++shift)
+            {
+                if((1U << shift) >= divisor)
+                {
+                    break;
+                }
+            }
+
+            uint64_t one        = 1;
+            uint64_t multiplier = ((one << 32) * ((one << shift) - divisor)) / divisor + 1;
+            // assert(multiplier <= 0xffffffffUL);
+
+            return make_tuple(uint32_t(multiplier), shift);
+        }
+        else
+        {
+            return make_tuple(uint32_t(0), uint32_t(0));
+        }
+    }
+
+    __host__ __device__ static constexpr uint32_t CalculateMagicMultiplier(uint32_t divisor)
+    {
+        auto tmp = CalculateMagicNumbers(divisor);
+
+        return tmp[Number<0>{}];
+    }
+
+    __host__ __device__ static constexpr uint32_t CalculateMagicShift(uint32_t divisor)
+    {
+        auto tmp = CalculateMagicNumbers(divisor);
+
+        return tmp[Number<1>{}];
+    }
+
+    // integral_constant<uint32_t, .>
+    template <uint32_t Divisor>
+    __host__ __device__ static constexpr auto
+        CalculateMagicNumbers(integral_constant<uint32_t, Divisor>)
+    {
+        constexpr auto tmp = CalculateMagicNumbers(uint32_t{Divisor});
+
+        constexpr uint32_t multiplier = tmp[Number<0>{}];
+        constexpr uint32_t shift      = tmp[Number<1>{}];
+
+        return make_tuple(integral_constant<uint32_t, multiplier>{},
+                          integral_constant<uint32_t, shift>{});
+    }
+
+    template <uint32_t Divisor>
+    __host__ __device__ static constexpr auto
+        CalculateMagicMultiplier(integral_constant<uint32_t, Divisor>)
+    {
+        constexpr uint32_t multiplier = CalculateMagicMultiplier(uint32_t{Divisor});
+
+        return integral_constant<uint32_t, multiplier>{};
+    }
+
+    template <uint32_t Divisor>
+    __host__ __device__ static constexpr auto
+        CalculateMagicShift(integral_constant<uint32_t, Divisor>)
+    {
+        constexpr uint32_t shift = CalculateMagicShift(uint32_t{Divisor});
+
+        return integral_constant<uint32_t, shift>{};
+    }
+
+    // integral_constant<int32_t, .>
+    template <int32_t Divisor>
+    __host__ __device__ static constexpr auto
+        CalculateMagicNumbers(integral_constant<int32_t, Divisor>)
+    {
+        return CalculateMagicNumbers(integral_constant<uint32_t, Divisor>{});
+    }
+
+    template <int32_t Divisor>
+    __host__ __device__ static constexpr auto
+        CalculateMagicMultiplier(integral_constant<int32_t, Divisor>)
+    {
+        return CalculateMagicMultiplier(integral_constant<uint32_t, Divisor>{});
+    }
+
+    template <int32_t Divisor>
+    __host__ __device__ static constexpr auto
+        CalculateMagicShift(integral_constant<int32_t, Divisor>)
+    {
+        return CalculateMagicShift(integral_constant<uint32_t, Divisor>{});
+    }
+
+    // magic division for uint32_t
+    __device__ static constexpr uint32_t
+    DoMagicDivision(uint32_t dividend, uint32_t multiplier, uint32_t shift)
+    {
+        uint32_t tmp = __umulhi(dividend, multiplier);
+        return (tmp + dividend) >> shift;
+    }
+
+    __host__ static constexpr uint32_t
+    DoMagicDivision(uint32_t dividend, uint32_t multiplier, uint32_t shift)
+    {
+        uint32_t tmp = static_cast<uint64_t>(dividend) * multiplier >> 32;
+        return (tmp + dividend) >> shift;
+    }
+
+    // magic division for int32_t
+    // HACK: use dividend_i32 as if it's uint32_t, dividend_i32 need to be
+    // non-negative for result to be correct
+    // TODO: figure out how to do magic number divison for int32_t as dividended
+    __device__ static constexpr int32_t
+    DoMagicDivision(int32_t dividend_i32, uint32_t multiplier, uint32_t shift)
+    {
+        uint32_t dividend_u32 = bit_cast<uint32_t>(dividend_i32);
+        uint32_t tmp          = __umulhi(dividend_u32, multiplier);
+        return (tmp + dividend_u32) >> shift;
+    }
+
+    __host__ static constexpr int32_t
+    DoMagicDivision(int32_t dividend_i32, uint32_t multiplier, uint32_t shift)
+    {
+        uint32_t dividend_u32 = bit_cast<uint32_t>(dividend_i32);
+        uint32_t tmp          = static_cast<uint64_t>(dividend_u32) * multiplier >> 32;
+        return (tmp + dividend_u32) >> shift;
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/math.hpp b/3rdparty/composable_kernel/include/ck/utility/math.hpp
new file mode 100644
index 0000000..12203bd
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/math.hpp
@@ -0,0 +1,240 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "integral_constant.hpp"
+#include "number.hpp"
+#include "type.hpp"
+#include "enable_if.hpp"
+
+namespace ck {
+namespace math {
+
+template <typename T, T s>
+struct scales
+{
+    __host__ __device__ constexpr T operator()(T a) const { return s * a; }
+};
+
+template <typename T>
+struct plus
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const { return a + b; }
+};
+
+template <typename T>
+struct minus
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const { return a - b; }
+};
+
+struct multiplies
+{
+    template <typename A, typename B>
+    __host__ __device__ constexpr auto operator()(const A& a, const B& b) const
+    {
+        return a * b;
+    }
+};
+
+template <typename T>
+struct maximize
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const { return a >= b ? a : b; }
+};
+
+template <typename T>
+struct minimize
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const { return a <= b ? a : b; }
+};
+
+template <typename T>
+struct integer_divide_ceiler
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const
+    {
+        static_assert(is_same<T, index_t>{} || is_same<T, int>{}, "wrong type");
+
+        return (a + b - Number<1>{}) / b;
+    }
+};
+
+template <typename X, typename Y>
+__host__ __device__ constexpr auto integer_divide_floor(X x, Y y)
+{
+    return x / y;
+}
+
+template <typename X, typename Y>
+__host__ __device__ constexpr auto integer_divide_ceil(X x, Y y)
+{
+    return (x + y - Number<1>{}) / y;
+}
+
+template <typename X, typename Y>
+__host__ __device__ constexpr auto integer_least_multiple(X x, Y y)
+{
+    return y * integer_divide_ceil(x, y);
+}
+
+template <typename T>
+__host__ __device__ constexpr T max(T x)
+{
+    return x;
+}
+
+template <typename T>
+__host__ __device__ constexpr T max(T x, T y)
+{
+    return x > y ? x : y;
+}
+
+template <index_t X>
+__host__ __device__ constexpr index_t max(Number<X>, index_t y)
+{
+    return X > y ? X : y;
+}
+
+template <index_t Y>
+__host__ __device__ constexpr index_t max(index_t x, Number<Y>)
+{
+    return x > Y ? x : Y;
+}
+
+template <typename X, typename... Ys>
+__host__ __device__ constexpr auto max(X x, Ys... ys)
+{
+    static_assert(sizeof...(Ys) > 0, "not enough argument");
+
+    return max(x, max(ys...));
+}
+
+template <typename T>
+__host__ __device__ constexpr T min(T x)
+{
+    return x;
+}
+
+template <typename T>
+__host__ __device__ constexpr T min(T x, T y)
+{
+    return x < y ? x : y;
+}
+
+template <index_t X>
+__host__ __device__ constexpr index_t min(Number<X>, index_t y)
+{
+    return X < y ? X : y;
+}
+
+template <index_t Y>
+__host__ __device__ constexpr index_t min(index_t x, Number<Y>)
+{
+    return x < Y ? x : Y;
+}
+
+template <typename X, typename... Ys>
+__host__ __device__ constexpr auto min(X x, Ys... ys)
+{
+    static_assert(sizeof...(Ys) > 0, "not enough argument");
+
+    return min(x, min(ys...));
+}
+
+template <typename T>
+__host__ __device__ constexpr T clamp(const T& x, const T& lowerbound, const T& upperbound)
+{
+    return min(max(x, lowerbound), upperbound);
+}
+
+// disallow implicit type casting
+template <typename T>
+__device__ T exp(T x);
+
+// TODO: add f16 support using v_exp_f16
+
+template <>
+__device__ float exp<float>(float x)
+{
+    return __expf(x);
+}
+
+template <>
+__device__ double exp<double>(double x)
+{
+    return exp(x);
+}
+
+// greatest common divisor, aka highest common factor
+__host__ __device__ constexpr index_t gcd(index_t x, index_t y)
+{
+    if(x < 0)
+    {
+        return gcd(-x, y);
+    }
+    else if(y < 0)
+    {
+        return gcd(x, -y);
+    }
+    else if(x == y || x == 0)
+    {
+        return y;
+    }
+    else if(y == 0)
+    {
+        return x;
+    }
+    else if(x > y)
+    {
+        return gcd(x % y, y);
+    }
+    else
+    {
+        return gcd(x, y % x);
+    }
+}
+
+template <index_t X, index_t Y>
+__host__ __device__ constexpr auto gcd(Number<X>, Number<Y>)
+{
+    constexpr auto r = gcd(X, Y);
+
+    return Number<r>{};
+}
+
+template <typename X, typename... Ys, typename enable_if<sizeof...(Ys) >= 2, bool>::type = false>
+__host__ __device__ constexpr auto gcd(X x, Ys... ys)
+{
+    return gcd(x, gcd(ys...));
+}
+
+// least common multiple
+template <typename X, typename Y>
+__host__ __device__ constexpr auto lcm(X x, Y y)
+{
+    return (x * y) / gcd(x, y);
+}
+
+template <typename X, typename... Ys, typename enable_if<sizeof...(Ys) >= 2, bool>::type = false>
+__host__ __device__ constexpr auto lcm(X x, Ys... ys)
+{
+    return lcm(x, lcm(ys...));
+}
+
+template <typename T>
+struct equal
+{
+    __host__ __device__ constexpr bool operator()(T x, T y) const { return x == y; }
+};
+
+template <typename T>
+struct less
+{
+    __host__ __device__ constexpr bool operator()(T x, T y) const { return x < y; }
+};
+
+} // namespace math
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/math_v2.hpp b/3rdparty/composable_kernel/include/ck/utility/math_v2.hpp
new file mode 100644
index 0000000..dc97666
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/math_v2.hpp
@@ -0,0 +1,164 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cmath>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/type.hpp"
+
+namespace ck {
+namespace math {
+
+// math functions for the host,  some are implemented by calling C++ std functions
+
+static inline __host__ float abs(float x) { return std::abs(x); };
+
+static inline __host__ double abs(double x) { return std::abs(x); };
+
+static inline __host__ int8_t abs(int8_t x)
+{
+    int8_t sgn = x >> (8 - 1);
+
+    return (x ^ sgn) - sgn;
+};
+
+static inline __host__ int32_t abs(int32_t x)
+{
+    int32_t sgn = x >> (32 - 1);
+
+    return (x ^ sgn) - sgn;
+};
+
+static inline __host__ half_t abs(half_t x)
+{
+    uint16_t xx = ck::bit_cast<uint16_t>(x);
+
+    uint16_t abs_xx = xx & 0x7fff;
+
+    half_t abs_x = ck::bit_cast<half_t>(abs_xx);
+
+    return abs_x;
+};
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+static inline __host__ int4_t abs(int4_t x)
+{
+    int4_t sgn = x >> (4 - 1);
+    return (x ^ sgn) - sgn;
+}
+#endif
+
+static inline __host__ bool isnan(float x) { return std::isnan(x); };
+
+static inline __host__ bool isnan(double x) { return std::isnan(x); };
+
+static inline __host__ bool isnan(int8_t x)
+{
+    (void)x;
+    return false;
+};
+
+static inline __host__ bool isnan(int32_t x)
+{
+    (void)x;
+    return false;
+};
+
+static inline __host__ bool isnan(half_t x)
+{
+    uint16_t xx = ck::bit_cast<uint16_t>(x);
+
+    return (xx & 0x7FFF) > 0x7C00;
+};
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+static inline __host__ bool isnan(int4_t x)
+{
+    (void)x;
+    return false;
+};
+#endif
+
+static inline __host__ float sqrt(float x) { return std::sqrt(x); };
+
+static inline __host__ double sqrt(double x) { return std::sqrt(x); };
+
+// math functions for the HIP kernel,  some are implemented by calling hip builtin functions
+
+static inline __device__ float abs(float x) { return ::abs(x); };
+
+static inline __device__ double abs(double x) { return ::abs(x); };
+
+static inline __device__ int8_t abs(int8_t x)
+{
+    int8_t sgn = x >> (8 - 1);
+
+    return (x ^ sgn) - sgn;
+};
+
+static inline __device__ int32_t abs(int32_t x)
+{
+    int32_t sgn = x >> (32 - 1);
+
+    return (x ^ sgn) - sgn;
+};
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+static inline __device__ int4_t abs(int4_t x)
+{
+    int4_t sgn = x >> (4 - 1);
+
+    return (x ^ sgn) - sgn;
+};
+#endif
+
+static inline __device__ half_t abs(half_t x)
+{
+    uint16_t xx = ck::bit_cast<uint16_t>(x);
+
+    uint16_t abs_xx = xx & 0x7fff;
+
+    half_t abs_x = ck::bit_cast<half_t>(abs_xx);
+
+    return abs_x;
+};
+
+static inline __device__ bool isnan(float x) { return ::isnan(x); };
+
+static inline __device__ bool isnan(double x) { return ::isnan(x); };
+
+static inline __device__ bool isnan(int8_t x)
+{
+    (void)x;
+    return false;
+};
+
+static inline __device__ bool isnan(int32_t x)
+{
+    (void)x;
+    return false;
+};
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+static inline __device__ bool isnan(int4_t x)
+{
+    (void)x;
+    return false;
+};
+#endif
+
+static inline __device__ bool isnan(half_t x)
+{
+    uint16_t xx = ck::bit_cast<uint16_t>(x);
+
+    return (xx & 0x7FFF) > 0x7C00;
+};
+
+static inline __device__ float sqrt(float x) { return ::sqrtf(x); };
+
+static inline __device__ double sqrt(double x) { return ::sqrt(x); };
+
+} // namespace math
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/multi_index.hpp b/3rdparty/composable_kernel/include/ck/utility/multi_index.hpp
new file mode 100644
index 0000000..1d544c0
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/multi_index.hpp
@@ -0,0 +1,12 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "common_header.hpp"
+
+#if CK_EXPERIMENTAL_USE_DYNAMICALLY_INDEXED_MULTI_INDEX
+#include "array_multi_index.hpp"
+#else
+#include "statically_indexed_array_multi_index.hpp"
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/number.hpp b/3rdparty/composable_kernel/include/ck/utility/number.hpp
new file mode 100644
index 0000000..f3ca6b6
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/number.hpp
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_NUMBER_HPP
+#define CK_NUMBER_HPP
+
+#include "integral_constant.hpp"
+
+namespace ck {
+
+template <index_t N>
+using Number = integral_constant<index_t, N>;
+
+template <index_t N>
+using LongNumber = integral_constant<long_index_t, N>;
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/print.hpp b/3rdparty/composable_kernel/include/ck/utility/print.hpp
new file mode 100644
index 0000000..eed1ca4
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/print.hpp
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_PRINT_HPP
+#define CK_PRINT_HPP
+
+#include "array.hpp"
+#include "statically_indexed_array.hpp"
+#include "container_helper.hpp"
+#include "sequence.hpp"
+
+namespace ck {
+
+template <typename T>
+__host__ __device__ void print_array(const char* s, T a)
+{
+    constexpr index_t nsize = a.Size();
+
+    printf("%s size %d, {", s, nsize);
+    static_for<0, nsize, 1>{}([&a](auto i) constexpr { printf("%d, ", int32_t{a[i]}); });
+    printf("}\n");
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/reduction_common.hpp b/3rdparty/composable_kernel/include/ck/utility/reduction_common.hpp
new file mode 100644
index 0000000..aceef7b
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/reduction_common.hpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_enums.hpp"
+
+namespace ck {
+
+struct float_equal_one
+{
+    template <class T>
+    __host__ __device__ inline bool operator()(T x)
+    {
+        return x <= static_cast<T>(1.0f) and x >= static_cast<T>(1.0f);
+    };
+};
+
+struct float_equal_zero
+{
+    template <class T>
+    __host__ __device__ inline bool operator()(T x)
+    {
+        return x <= static_cast<T>(0.0f) and x >= static_cast<T>(0.0f);
+    };
+};
+
+template <index_t N>
+static constexpr __device__ index_t get_shift()
+{
+    return (get_shift<N / 2>() + 1);
+};
+
+template <>
+constexpr __device__ index_t get_shift<1>()
+{
+    return (0);
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/reduction_enums.hpp b/3rdparty/composable_kernel/include/ck/utility/reduction_enums.hpp
new file mode 100644
index 0000000..6785633
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/reduction_enums.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+enum struct ReduceTensorOp
+{
+    ADD   = 0,
+    MUL   = 1,
+    MIN   = 2,
+    MAX   = 3,
+    AMAX  = 4,
+    AVG   = 5,
+    NORM1 = 6,
+    NORM2 = 7,
+    // MUL_NO_ZEROS = 8,
+};
+
+enum struct NanPropagation
+{
+    NOT_PROPAGATE_NAN = 0,
+    PROPAGATE_NAN     = 1,
+};
+
+enum struct ReduceTensorIndices
+{
+    NO_INDICES        = 0,
+    FLATTENED_INDICES = 1,
+};
+
+enum struct IndicesType
+{
+    INDICES_32BIT = 0,
+    INDICES_64BIT = 1,
+    INDICES_16BIT = 2,
+    INDICES_8BIT  = 3,
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/reduction_functions_accumulate.hpp b/3rdparty/composable_kernel/include/ck/utility/reduction_functions_accumulate.hpp
new file mode 100644
index 0000000..724e559
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/reduction_functions_accumulate.hpp
@@ -0,0 +1,115 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+namespace ck {
+namespace detail {
+
+// Check for NaN; guarantee NaNs are NOT propagated to result (i.e., ignore NaNs)
+template <typename ReduceOperation, typename AccDataType>
+struct AccumulateWithNanIgnore
+{
+    __device__ static inline void Calculate(AccDataType& accuVal, AccDataType currVal)
+    {
+        if(!ck::math::isnan(currVal))
+        {
+            ReduceOperation{}(accuVal, currVal);
+        }
+    };
+};
+
+template <bool PropagateNan, typename ReduceOperation, typename AccDataType>
+struct AccumulateWithNanCheck;
+
+// Does not check for NaN; does not guarantee NaNs be propagated to result
+// e.g., given that max(a, b) = a > b ? a : b
+// then  max(NaN, 1) returns 1
+//       max(1, NaN) returns NaN
+// since any comparison involving NaNs returns false
+template <typename ReduceOperation, typename AccDataType>
+struct AccumulateWithNanCheck<false, ReduceOperation, AccDataType>
+{
+    // cppcheck-suppress constParameter
+    __host__ __device__ static inline void Calculate(AccDataType& accuVal, AccDataType currVal)
+    {
+        ReduceOperation{}(accuVal, currVal);
+    };
+};
+
+// Check for NaN; guarantees NaNs be propagated to result
+template <typename ReduceOperation, typename AccDataType>
+struct AccumulateWithNanCheck<true, ReduceOperation, AccDataType>
+{
+    __host__ __device__ static inline void Calculate(AccDataType& accuVal, AccDataType currVal)
+    {
+        using ck::math::isnan;
+
+        if(isnan(currVal))
+        {
+            accuVal = currVal;
+        }
+        else
+        {
+            ReduceOperation{}(accuVal, currVal);
+        };
+    };
+};
+
+template <bool PropagateNan, typename ReduceOperation, typename AccDataType, typename IndexDataType>
+struct AccumulateWithIndexAndNanCheck;
+
+template <typename ReduceOperation, typename AccDataType, typename IndexDataType>
+struct AccumulateWithIndexAndNanCheck<false, ReduceOperation, AccDataType, IndexDataType>
+{
+    __host__ __device__ static inline void
+    // cppcheck-suppress constParameter
+    Calculate(AccDataType& accuVal,
+              AccDataType currVal,
+              IndexDataType& accuIndex,
+              IndexDataType currIndex)
+    {
+        bool changed = false;
+
+        ReduceOperation{}(accuVal, currVal, changed);
+
+        if(changed)
+            accuIndex = currIndex;
+    };
+};
+
+template <typename ReduceOperation, typename AccDataType, typename IndexDataType>
+struct AccumulateWithIndexAndNanCheck<true, ReduceOperation, AccDataType, IndexDataType>
+{
+    // The method is called when the ReduceOperation is indexable and the user asked for indices
+    __host__ __device__ static inline void Calculate(AccDataType& accuVal,
+                                                     AccDataType currVal,
+                                                     IndexDataType& accuIndex,
+                                                     IndexDataType currIndex)
+    {
+        using ck::math::isnan;
+
+        if(isnan(currVal))
+        {
+            accuVal   = currVal;
+            accuIndex = currIndex;
+        }
+        else
+        {
+            bool changed = false;
+
+            ReduceOperation{}(accuVal, currVal, changed);
+
+            if(changed)
+                accuIndex = currIndex;
+        }
+    };
+};
+
+} // namespace detail
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/reduction_operator.hpp b/3rdparty/composable_kernel/include/ck/utility/reduction_operator.hpp
new file mode 100644
index 0000000..25ae8fd
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/reduction_operator.hpp
@@ -0,0 +1,292 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/type.hpp"
+
+namespace ck {
+
+namespace reduce {
+
+// Every binary operator used in reduction is represented by a templated functor class. Each functor
+// class must provide at least
+// three members:
+// 1) GetIdentityValue() -- the interface to return the "identity element" for the binary
+// operator, "identity element" is the unique
+//                    element in the algebraic space that doesn't affect the value of other elements
+//                    when operated against them, and the concept is similar to zero vector in
+//                    vector space
+//                    (http://pages.cs.wisc.edu/~matthewb/pages/notes/pdf/linearalgebra/VectorSpaces.pdf).
+// 2) IsCompatibleInMemoryDataOperation() -- return true if the reduction task corresponding to this
+// operator can use the InMemoryDataOperation to finalize, or else it return false
+// 3) operator() -- the first argument of the operator must be both an input & output, and the
+//                  corresponding variable usually stores
+//                  the accumulated result of many operator() calls; the second argument is only an
+//                  input. For indexable binary
+//                  operator, the second version of operator() has third argument (which is an
+//                  output) to indicate whether the
+//                  accumulated value (the first argument) has changed, in which case the recorded
+//                  accumulated index also need be
+//                  changed.
+
+struct Add
+{
+    template <typename T>
+    __host__ __device__ static constexpr T GetIdentityValue()
+    {
+        return type_convert<T>(0.0f);
+    };
+
+    __host__ __device__ static constexpr bool
+    IsCompatibleInMemoryDataOperation(InMemoryDataOperationEnum operation)
+    {
+        return operation == InMemoryDataOperationEnum::AtomicAdd ||
+               operation == InMemoryDataOperationEnum::Set;
+    };
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, int32_t>::value,
+                      "The data type is not supported by the Add accumulator!");
+
+        a = a + b;
+    }
+};
+
+struct SquaredAdd
+{
+    template <class T>
+    __host__ __device__ static constexpr T GetIdentityValue()
+    {
+        return type_convert<T>(0.0f);
+    };
+
+    __host__ __device__ static constexpr bool
+    IsCompatibleInMemoryDataOperation(InMemoryDataOperationEnum operation)
+    {
+        return operation == InMemoryDataOperationEnum::AtomicAdd ||
+               operation == InMemoryDataOperationEnum::Set;
+    };
+
+    template <class T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the SquaredAdd accumulator!");
+
+        a = a + b * b;
+    }
+};
+
+struct Mul
+{
+    template <typename T>
+    __host__ __device__ static constexpr T GetIdentityValue()
+    {
+        return type_convert<T>(1.0f);
+    };
+
+    __host__ __device__ static constexpr bool
+    IsCompatibleInMemoryDataOperation(InMemoryDataOperationEnum operation)
+    {
+        return operation == InMemoryDataOperationEnum::Set;
+    };
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, int32_t>::value,
+                      "The data type is not supported by the Mul accumulator!");
+
+        a = a * b;
+    }
+};
+
+struct Max
+{
+    template <typename T>
+    __host__ __device__ static constexpr T GetIdentityValue()
+    {
+        return NumericLimits<T>::Lowest();
+    };
+
+    __host__ __device__ static constexpr bool
+    IsCompatibleInMemoryDataOperation(InMemoryDataOperationEnum operation)
+    {
+        // ToChange: atomic_max to be added
+        return operation == InMemoryDataOperationEnum::Set;
+    };
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the Max accumulator!");
+
+        if(a < b)
+            a = b;
+    }
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b, bool& changed) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the Max accumulator!");
+
+        if(a < b)
+        {
+            a       = b;
+            changed = true;
+        }
+    }
+};
+
+struct Min
+{
+    template <typename T>
+    __host__ __device__ static constexpr T GetIdentityValue()
+    {
+        return NumericLimits<T>::Max();
+    };
+
+    __host__ __device__ static constexpr bool
+    IsCompatibleInMemoryDataOperation(InMemoryDataOperationEnum operation)
+    {
+        // ToChange: atomic_min to be added
+        return operation == InMemoryDataOperationEnum::Set;
+    };
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the Min accumulator!");
+
+        if(a > b)
+            a = b;
+    }
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b, bool& changed) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the Min accumulator!");
+
+        if(a > b)
+        {
+            a       = b;
+            changed = true;
+        }
+    }
+};
+
+struct AMax
+{
+    template <typename T>
+    __host__ __device__ static constexpr T GetIdentityValue()
+    {
+        return type_convert<T>(0.0f);
+    };
+
+    __host__ __device__ static constexpr bool
+    IsCompatibleInMemoryDataOperation(InMemoryDataOperationEnum operation)
+    {
+        // ToChange: atomic_max to be added
+        return operation == InMemoryDataOperationEnum::Set;
+    };
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the AMax accumulator!");
+
+        if(a < b)
+            a = b;
+    }
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b, bool& changed) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the AMax accumulator!");
+
+        if(a < b)
+        {
+            a       = b;
+            changed = true;
+        }
+    }
+};
+
+template <typename T>
+constexpr T GetIdentityValueForInMemoryDataOperation(InMemoryDataOperationEnum operation)
+{
+    T result = ck::type_convert<T>(0.0f);
+
+    if(operation == InMemoryDataOperationEnum::AtomicMax)
+        result = ck::NumericLimits<T>::Lowest();
+
+    return (result);
+};
+
+template <InMemoryDataOperationEnum Operation, typename DataType>
+struct InMemoryDataOperatonSupportedOnDataType
+{
+    static constexpr bool value = false;
+};
+
+template <typename DataType>
+struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::AtomicAdd, DataType>
+{
+    static constexpr bool value =
+        is_same<DataType, float>::value || is_same<DataType, double>::value;
+};
+
+template <typename DataType>
+struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::AtomicMax, DataType>
+{
+    static constexpr bool value =
+        is_same<DataType, float>::value || is_same<DataType, double>::value;
+};
+
+template <typename DataType>
+struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::Set, DataType>
+{
+    static constexpr bool value =
+        is_same<DataType, float>::value || is_same<DataType, double>::value ||
+        is_same<DataType, half_t>::value || is_same<DataType, bhalf_t>::value ||
+        is_same<DataType, int8_t>::value || is_same<DataType, int32_t>::value;
+};
+
+template <typename DataType>
+struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::Add, DataType>
+{
+    static constexpr bool value =
+        is_same<DataType, float>::value || is_same<DataType, double>::value ||
+        is_same<DataType, half_t>::value || is_same<DataType, int8_t>::value ||
+        is_same<DataType, int32_t>::value;
+};
+
+} // namespace reduce
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/sequence.hpp b/3rdparty/composable_kernel/include/ck/utility/sequence.hpp
new file mode 100644
index 0000000..97b5972
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/sequence.hpp
@@ -0,0 +1,899 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/integral_constant.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/utility/functional.hpp"
+#include "ck/utility/math.hpp"
+
+namespace ck {
+
+template <index_t, index_t, index_t>
+struct static_for;
+
+template <index_t...>
+struct Sequence;
+
+template <typename Seq, index_t I>
+struct sequence_split;
+
+template <typename>
+struct sequence_reverse;
+
+template <typename>
+struct sequence_map_inverse;
+
+template <typename>
+struct is_valid_sequence_map;
+
+template <index_t I, index_t... Is>
+__host__ __device__ constexpr auto sequence_pop_front(Sequence<I, Is...>);
+
+template <typename Seq>
+__host__ __device__ constexpr auto sequence_pop_back(Seq);
+
+template <index_t... Is>
+struct Sequence
+{
+    using Type      = Sequence;
+    using data_type = index_t;
+
+    static constexpr index_t mSize = sizeof...(Is);
+
+    __host__ __device__ static constexpr auto Size() { return Number<mSize>{}; }
+
+    __host__ __device__ static constexpr auto GetSize() { return Size(); }
+
+    __host__ __device__ static constexpr index_t At(index_t I)
+    {
+        // the last dummy element is to prevent compiler complain about empty array, when mSize = 0
+        const index_t mData[mSize + 1] = {Is..., 0};
+        return mData[I];
+    }
+
+    template <index_t I>
+    __host__ __device__ static constexpr auto At(Number<I>)
+    {
+        static_assert(I < mSize, "wrong! I too large");
+
+        return Number<At(I)>{};
+    }
+
+    template <index_t I>
+    __host__ __device__ static constexpr auto Get(Number<I>)
+    {
+        return At(Number<I>{});
+    }
+
+    template <typename I>
+    __host__ __device__ constexpr auto operator[](I i) const
+    {
+        return At(i);
+    }
+
+    template <index_t... IRs>
+    __host__ __device__ static constexpr auto ReorderGivenNew2Old(Sequence<IRs...> /*new2old*/)
+    {
+        static_assert(sizeof...(Is) == sizeof...(IRs),
+                      "wrong! reorder map should have the same size as Sequence to be rerodered");
+
+        static_assert(is_valid_sequence_map<Sequence<IRs...>>::value, "wrong! invalid reorder map");
+
+        return Sequence<Type::At(Number<IRs>{})...>{};
+    }
+
+    // MapOld2New is Sequence<...>
+    template <typename MapOld2New>
+    __host__ __device__ static constexpr auto ReorderGivenOld2New(MapOld2New)
+    {
+        static_assert(MapOld2New::Size() == Size(),
+                      "wrong! reorder map should have the same size as Sequence to be rerodered");
+
+        static_assert(is_valid_sequence_map<MapOld2New>::value, "wrong! invalid reorder map");
+
+        return ReorderGivenNew2Old(typename sequence_map_inverse<MapOld2New>::type{});
+    }
+
+    __host__ __device__ static constexpr auto Reverse()
+    {
+        return typename sequence_reverse<Type>::type{};
+    }
+
+    __host__ __device__ static constexpr auto Front()
+    {
+        static_assert(mSize > 0, "wrong!");
+        return At(Number<0>{});
+    }
+
+    __host__ __device__ static constexpr auto Back()
+    {
+        static_assert(mSize > 0, "wrong!");
+        return At(Number<mSize - 1>{});
+    }
+
+    __host__ __device__ static constexpr auto PopFront() { return sequence_pop_front(Type{}); }
+
+    __host__ __device__ static constexpr auto PopBack() { return sequence_pop_back(Type{}); }
+
+    template <index_t... Xs>
+    __host__ __device__ static constexpr auto PushFront(Sequence<Xs...>)
+    {
+        return Sequence<Xs..., Is...>{};
+    }
+
+    template <index_t... Xs>
+    __host__ __device__ static constexpr auto PushFront(Number<Xs>...)
+    {
+        return Sequence<Xs..., Is...>{};
+    }
+
+    template <index_t... Xs>
+    __host__ __device__ static constexpr auto PushBack(Sequence<Xs...>)
+    {
+        return Sequence<Is..., Xs...>{};
+    }
+
+    template <index_t... Xs>
+    __host__ __device__ static constexpr auto PushBack(Number<Xs>...)
+    {
+        return Sequence<Is..., Xs...>{};
+    }
+
+    template <index_t... Ns>
+    __host__ __device__ static constexpr auto Extract(Number<Ns>...)
+    {
+        return Sequence<Type::At(Number<Ns>{})...>{};
+    }
+
+    template <index_t... Ns>
+    __host__ __device__ static constexpr auto Extract(Sequence<Ns...>)
+    {
+        return Sequence<Type::At(Number<Ns>{})...>{};
+    }
+
+    template <index_t I, index_t X>
+    __host__ __device__ static constexpr auto Modify(Number<I>, Number<X>)
+    {
+        static_assert(I < Size(), "wrong!");
+
+        using seq_split          = sequence_split<Type, I>;
+        constexpr auto seq_left  = typename seq_split::left_type{};
+        constexpr auto seq_right = typename seq_split::right_type{}.PopFront();
+
+        return seq_left.PushBack(Number<X>{}).PushBack(seq_right);
+    }
+
+    template <typename F>
+    __host__ __device__ static constexpr auto Transform(F f)
+    {
+        return Sequence<f(Is)...>{};
+    }
+
+    __host__ __device__ static void Print()
+    {
+        printf("{");
+        printf("size %d, ", index_t{Size()});
+        static_for<0, Size(), 1>{}([&](auto i) { printf("%d ", At(i).value); });
+        printf("}");
+    }
+};
+
+// merge sequence
+template <typename Seq, typename... Seqs>
+struct sequence_merge
+{
+    using type = typename sequence_merge<Seq, typename sequence_merge<Seqs...>::type>::type;
+};
+
+template <index_t... Xs, index_t... Ys>
+struct sequence_merge<Sequence<Xs...>, Sequence<Ys...>>
+{
+    using type = Sequence<Xs..., Ys...>;
+};
+
+template <typename Seq>
+struct sequence_merge<Seq>
+{
+    using type = Seq;
+};
+
+// generate sequence
+template <index_t NSize, typename F>
+struct sequence_gen
+{
+    template <index_t IBegin, index_t NRemain, typename G>
+    struct sequence_gen_impl
+    {
+        static constexpr index_t NRemainLeft  = NRemain / 2;
+        static constexpr index_t NRemainRight = NRemain - NRemainLeft;
+        static constexpr index_t IMiddle      = IBegin + NRemainLeft;
+
+        using type = typename sequence_merge<
+            typename sequence_gen_impl<IBegin, NRemainLeft, G>::type,
+            typename sequence_gen_impl<IMiddle, NRemainRight, G>::type>::type;
+    };
+
+    template <index_t I, typename G>
+    struct sequence_gen_impl<I, 1, G>
+    {
+        static constexpr index_t Is = G{}(Number<I>{});
+        using type                  = Sequence<Is>;
+    };
+
+    template <index_t I, typename G>
+    struct sequence_gen_impl<I, 0, G>
+    {
+        using type = Sequence<>;
+    };
+
+    using type = typename sequence_gen_impl<0, NSize, F>::type;
+};
+
+// arithmetic sequence
+template <index_t IBegin, index_t IEnd, index_t Increment>
+struct arithmetic_sequence_gen
+{
+    struct F
+    {
+        __host__ __device__ constexpr index_t operator()(index_t i) const
+        {
+            return i * Increment + IBegin;
+        }
+    };
+
+    using type0 = typename sequence_gen<(IEnd - IBegin) / Increment, F>::type;
+    using type1 = Sequence<>;
+
+    static constexpr bool kHasContent =
+        (Increment > 0 && IBegin < IEnd) || (Increment < 0 && IBegin > IEnd);
+
+    using type = typename conditional<kHasContent, type0, type1>::type;
+};
+
+// uniform sequence
+template <index_t NSize, index_t I>
+struct uniform_sequence_gen
+{
+    struct F
+    {
+        __host__ __device__ constexpr index_t operator()(index_t) const { return I; }
+    };
+
+    using type = typename sequence_gen<NSize, F>::type;
+};
+
+// reverse inclusive scan (with init) sequence
+template <typename, typename, index_t>
+struct sequence_reverse_inclusive_scan;
+
+template <index_t I, index_t... Is, typename Reduce, index_t Init>
+struct sequence_reverse_inclusive_scan<Sequence<I, Is...>, Reduce, Init>
+{
+    using old_scan = typename sequence_reverse_inclusive_scan<Sequence<Is...>, Reduce, Init>::type;
+
+    static constexpr index_t new_reduce = Reduce{}(I, old_scan{}.Front());
+
+    using type = typename sequence_merge<Sequence<new_reduce>, old_scan>::type;
+};
+
+template <index_t I, typename Reduce, index_t Init>
+struct sequence_reverse_inclusive_scan<Sequence<I>, Reduce, Init>
+{
+    using type = Sequence<Reduce{}(I, Init)>;
+};
+
+template <typename Reduce, index_t Init>
+struct sequence_reverse_inclusive_scan<Sequence<>, Reduce, Init>
+{
+    using type = Sequence<>;
+};
+
+// split sequence
+template <typename Seq, index_t I>
+struct sequence_split
+{
+    static constexpr index_t NSize = Seq{}.Size();
+
+    using range0 = typename arithmetic_sequence_gen<0, I, 1>::type;
+    using range1 = typename arithmetic_sequence_gen<I, NSize, 1>::type;
+
+    using left_type  = decltype(Seq::Extract(range0{}));
+    using right_type = decltype(Seq::Extract(range1{}));
+};
+
+// reverse sequence
+template <typename Seq>
+struct sequence_reverse
+{
+    static constexpr index_t NSize = Seq{}.Size();
+
+    using seq_split = sequence_split<Seq, NSize / 2>;
+    using type      = typename sequence_merge<
+        typename sequence_reverse<typename seq_split::right_type>::type,
+        typename sequence_reverse<typename seq_split::left_type>::type>::type;
+};
+
+template <index_t I>
+struct sequence_reverse<Sequence<I>>
+{
+    using type = Sequence<I>;
+};
+
+template <index_t I0, index_t I1>
+struct sequence_reverse<Sequence<I0, I1>>
+{
+    using type = Sequence<I1, I0>;
+};
+
+#if 1
+template <typename Reduce, typename Seq, typename... Seqs>
+struct sequence_reduce
+{
+    using type = typename sequence_reduce<Reduce,
+                                          Seq,
+                                          typename sequence_reduce<Reduce, Seqs...>::type>::type;
+};
+
+template <typename Reduce, index_t... Xs, index_t... Ys>
+struct sequence_reduce<Reduce, Sequence<Xs...>, Sequence<Ys...>>
+{
+    using type = Sequence<Reduce{}(Xs, Ys)...>;
+};
+
+template <typename Reduce, typename Seq>
+struct sequence_reduce<Reduce, Seq>
+{
+    using type = Seq;
+};
+#endif
+
+template <typename Values, typename Ids, typename Compare>
+struct sequence_sort_impl
+{
+    template <typename LeftValues,
+              typename LeftIds,
+              typename RightValues,
+              typename RightIds,
+              typename MergedValues,
+              typename MergedIds,
+              typename Comp>
+    struct sorted_sequence_merge_impl
+    {
+        static constexpr bool choose_left = LeftValues::Front() < RightValues::Front();
+
+        static constexpr index_t chosen_value =
+            choose_left ? LeftValues::Front() : RightValues::Front();
+        static constexpr index_t chosen_id = choose_left ? LeftIds::Front() : RightIds::Front();
+
+        using new_merged_values = decltype(MergedValues::PushBack(Number<chosen_value>{}));
+        using new_merged_ids    = decltype(MergedIds::PushBack(Number<chosen_id>{}));
+
+        using new_left_values =
+            typename conditional<choose_left, decltype(LeftValues::PopFront()), LeftValues>::type;
+        using new_left_ids =
+            typename conditional<choose_left, decltype(LeftIds::PopFront()), LeftIds>::type;
+
+        using new_right_values =
+            typename conditional<choose_left, RightValues, decltype(RightValues::PopFront())>::type;
+        using new_right_ids =
+            typename conditional<choose_left, RightIds, decltype(RightIds::PopFront())>::type;
+
+        using merge = sorted_sequence_merge_impl<new_left_values,
+                                                 new_left_ids,
+                                                 new_right_values,
+                                                 new_right_ids,
+                                                 new_merged_values,
+                                                 new_merged_ids,
+                                                 Comp>;
+        // this is output
+        using merged_values = typename merge::merged_values;
+        using merged_ids    = typename merge::merged_ids;
+    };
+
+    template <typename LeftValues,
+              typename LeftIds,
+              typename MergedValues,
+              typename MergedIds,
+              typename Comp>
+    struct sorted_sequence_merge_impl<LeftValues,
+                                      LeftIds,
+                                      Sequence<>,
+                                      Sequence<>,
+                                      MergedValues,
+                                      MergedIds,
+                                      Comp>
+    {
+        using merged_values = typename sequence_merge<MergedValues, LeftValues>::type;
+        using merged_ids    = typename sequence_merge<MergedIds, LeftIds>::type;
+    };
+
+    template <typename RightValues,
+              typename RightIds,
+              typename MergedValues,
+              typename MergedIds,
+              typename Comp>
+    struct sorted_sequence_merge_impl<Sequence<>,
+                                      Sequence<>,
+                                      RightValues,
+                                      RightIds,
+                                      MergedValues,
+                                      MergedIds,
+                                      Comp>
+    {
+        using merged_values = typename sequence_merge<MergedValues, RightValues>::type;
+        using merged_ids    = typename sequence_merge<MergedIds, RightIds>::type;
+    };
+
+    template <typename LeftValues,
+              typename LeftIds,
+              typename RightValues,
+              typename RightIds,
+              typename Comp>
+    struct sorted_sequence_merge
+    {
+        using merge = sorted_sequence_merge_impl<LeftValues,
+                                                 LeftIds,
+                                                 RightValues,
+                                                 RightIds,
+                                                 Sequence<>,
+                                                 Sequence<>,
+                                                 Comp>;
+
+        using merged_values = typename merge::merged_values;
+        using merged_ids    = typename merge::merged_ids;
+    };
+
+    static constexpr index_t nsize = Values::Size();
+
+    using split_unsorted_values = sequence_split<Values, nsize / 2>;
+    using split_unsorted_ids    = sequence_split<Ids, nsize / 2>;
+
+    using left_unsorted_values = typename split_unsorted_values::left_type;
+    using left_unsorted_ids    = typename split_unsorted_ids::left_type;
+    using left_sort          = sequence_sort_impl<left_unsorted_values, left_unsorted_ids, Compare>;
+    using left_sorted_values = typename left_sort::sorted_values;
+    using left_sorted_ids    = typename left_sort::sorted_ids;
+
+    using right_unsorted_values = typename split_unsorted_values::right_type;
+    using right_unsorted_ids    = typename split_unsorted_ids::right_type;
+    using right_sort = sequence_sort_impl<right_unsorted_values, right_unsorted_ids, Compare>;
+    using right_sorted_values = typename right_sort::sorted_values;
+    using right_sorted_ids    = typename right_sort::sorted_ids;
+
+    using merged_sorted = sorted_sequence_merge<left_sorted_values,
+                                                left_sorted_ids,
+                                                right_sorted_values,
+                                                right_sorted_ids,
+                                                Compare>;
+
+    using sorted_values = typename merged_sorted::merged_values;
+    using sorted_ids    = typename merged_sorted::merged_ids;
+};
+
+template <index_t ValueX, index_t ValueY, index_t IdX, index_t IdY, typename Compare>
+struct sequence_sort_impl<Sequence<ValueX, ValueY>, Sequence<IdX, IdY>, Compare>
+{
+    static constexpr bool choose_x = Compare{}(ValueX, ValueY);
+
+    using sorted_values =
+        typename conditional<choose_x, Sequence<ValueX, ValueY>, Sequence<ValueY, ValueX>>::type;
+    using sorted_ids = typename conditional<choose_x, Sequence<IdX, IdY>, Sequence<IdY, IdX>>::type;
+};
+
+template <index_t Value, index_t Id, typename Compare>
+struct sequence_sort_impl<Sequence<Value>, Sequence<Id>, Compare>
+{
+    using sorted_values = Sequence<Value>;
+    using sorted_ids    = Sequence<Id>;
+};
+
+template <typename Compare>
+struct sequence_sort_impl<Sequence<>, Sequence<>, Compare>
+{
+    using sorted_values = Sequence<>;
+    using sorted_ids    = Sequence<>;
+};
+
+template <typename Values, typename Compare>
+struct sequence_sort
+{
+    using unsorted_ids = typename arithmetic_sequence_gen<0, Values::Size(), 1>::type;
+    using sort         = sequence_sort_impl<Values, unsorted_ids, Compare>;
+
+    // this is output
+    using type                = typename sort::sorted_values;
+    using sorted2unsorted_map = typename sort::sorted_ids;
+};
+
+template <typename Values, typename Less, typename Equal>
+struct sequence_unique_sort
+{
+    template <typename RemainValues,
+              typename RemainIds,
+              typename UniquifiedValues,
+              typename UniquifiedIds,
+              typename Eq>
+    struct sorted_sequence_uniquify_impl
+    {
+        static constexpr index_t current_value = RemainValues::Front();
+        static constexpr index_t current_id    = RemainIds::Front();
+
+        static constexpr bool is_unique_value = (current_value != UniquifiedValues::Back());
+
+        using new_remain_values = decltype(RemainValues::PopFront());
+        using new_remain_ids    = decltype(RemainIds::PopFront());
+
+        using new_uniquified_values =
+            typename conditional<is_unique_value,
+                                 decltype(UniquifiedValues::PushBack(Number<current_value>{})),
+                                 UniquifiedValues>::type;
+
+        using new_uniquified_ids =
+            typename conditional<is_unique_value,
+                                 decltype(UniquifiedIds::PushBack(Number<current_id>{})),
+                                 UniquifiedIds>::type;
+
+        using uniquify = sorted_sequence_uniquify_impl<new_remain_values,
+                                                       new_remain_ids,
+                                                       new_uniquified_values,
+                                                       new_uniquified_ids,
+                                                       Eq>;
+
+        // this is output
+        using uniquified_values = typename uniquify::uniquified_values;
+        using uniquified_ids    = typename uniquify::uniquified_ids;
+    };
+
+    template <typename UniquifiedValues, typename UniquifiedIds, typename Eq>
+    struct sorted_sequence_uniquify_impl<Sequence<>,
+                                         Sequence<>,
+                                         UniquifiedValues,
+                                         UniquifiedIds,
+                                         Eq>
+    {
+        using uniquified_values = UniquifiedValues;
+        using uniquified_ids    = UniquifiedIds;
+    };
+
+    template <typename SortedValues, typename SortedIds, typename Eq>
+    struct sorted_sequence_uniquify
+    {
+        using uniquify = sorted_sequence_uniquify_impl<decltype(SortedValues::PopFront()),
+                                                       decltype(SortedIds::PopFront()),
+                                                       Sequence<SortedValues::Front()>,
+                                                       Sequence<SortedIds::Front()>,
+                                                       Eq>;
+
+        using uniquified_values = typename uniquify::uniquified_values;
+        using uniquified_ids    = typename uniquify::uniquified_ids;
+    };
+
+    using sort          = sequence_sort<Values, Less>;
+    using sorted_values = typename sort::type;
+    using sorted_ids    = typename sort::sorted2unsorted_map;
+
+    using uniquify = sorted_sequence_uniquify<sorted_values, sorted_ids, Equal>;
+
+    // this is output
+    using type                = typename uniquify::uniquified_values;
+    using sorted2unsorted_map = typename uniquify::uniquified_ids;
+};
+
+template <typename SeqMap>
+struct is_valid_sequence_map : is_same<typename arithmetic_sequence_gen<0, SeqMap::Size(), 1>::type,
+                                       typename sequence_sort<SeqMap, math::less<index_t>>::type>
+{
+};
+
+template <typename SeqMap>
+struct sequence_map_inverse
+{
+    template <typename X2Y, typename WorkingY2X, index_t XBegin, index_t XRemain>
+    struct sequence_map_inverse_impl
+    {
+        static constexpr auto new_y2x =
+            WorkingY2X::Modify(X2Y::At(Number<XBegin>{}), Number<XBegin>{});
+
+        using type =
+            typename sequence_map_inverse_impl<X2Y, decltype(new_y2x), XBegin + 1, XRemain - 1>::
+                type;
+    };
+
+    template <typename X2Y, typename WorkingY2X, index_t XBegin>
+    struct sequence_map_inverse_impl<X2Y, WorkingY2X, XBegin, 0>
+    {
+        using type = WorkingY2X;
+    };
+
+    using type =
+        typename sequence_map_inverse_impl<SeqMap,
+                                           typename uniform_sequence_gen<SeqMap::Size(), 0>::type,
+                                           0,
+                                           SeqMap::Size()>::type;
+};
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr bool operator==(Sequence<Xs...>, Sequence<Ys...>)
+{
+    return ((Xs == Ys) && ...);
+}
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr auto operator+(Sequence<Xs...>, Sequence<Ys...>)
+{
+    static_assert(sizeof...(Xs) == sizeof...(Ys), "wrong! inconsistent size");
+
+    return Sequence<(Xs + Ys)...>{};
+}
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr auto operator-(Sequence<Xs...>, Sequence<Ys...>)
+{
+    static_assert(sizeof...(Xs) == sizeof...(Ys), "wrong! inconsistent size");
+
+    return Sequence<(Xs - Ys)...>{};
+}
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr auto operator*(Sequence<Xs...>, Sequence<Ys...>)
+{
+    static_assert(sizeof...(Xs) == sizeof...(Ys), "wrong! inconsistent size");
+
+    return Sequence<(Xs * Ys)...>{};
+}
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr auto operator/(Sequence<Xs...>, Sequence<Ys...>)
+{
+    static_assert(sizeof...(Xs) == sizeof...(Ys), "wrong! inconsistent size");
+
+    return Sequence<(Xs / Ys)...>{};
+}
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr auto operator%(Sequence<Xs...>, Sequence<Ys...>)
+{
+    static_assert(sizeof...(Xs) == sizeof...(Ys), "wrong! inconsistent size");
+
+    return Sequence<(Xs % Ys)...>{};
+}
+
+template <index_t... Xs, index_t Y>
+__host__ __device__ constexpr auto operator+(Sequence<Xs...>, Number<Y>)
+{
+    return Sequence<(Xs + Y)...>{};
+}
+
+template <index_t... Xs, index_t Y>
+__host__ __device__ constexpr auto operator-(Sequence<Xs...>, Number<Y>)
+{
+    return Sequence<(Xs - Y)...>{};
+}
+
+template <index_t... Xs, index_t Y>
+__host__ __device__ constexpr auto operator*(Sequence<Xs...>, Number<Y>)
+{
+    return Sequence<(Xs * Y)...>{};
+}
+
+template <index_t... Xs, index_t Y>
+__host__ __device__ constexpr auto operator/(Sequence<Xs...>, Number<Y>)
+{
+    return Sequence<(Xs / Y)...>{};
+}
+
+template <index_t... Xs, index_t Y>
+__host__ __device__ constexpr auto operator%(Sequence<Xs...>, Number<Y>)
+{
+    return Sequence<(Xs % Y)...>{};
+}
+
+template <index_t Y, index_t... Xs>
+__host__ __device__ constexpr auto operator+(Number<Y>, Sequence<Xs...>)
+{
+    return Sequence<(Y + Xs)...>{};
+}
+
+template <index_t Y, index_t... Xs>
+__host__ __device__ constexpr auto operator-(Number<Y>, Sequence<Xs...>)
+{
+    return Sequence<(Y - Xs)...>{};
+}
+
+template <index_t Y, index_t... Xs>
+__host__ __device__ constexpr auto operator*(Number<Y>, Sequence<Xs...>)
+{
+    return Sequence<(Y * Xs)...>{};
+}
+
+template <index_t Y, index_t... Xs>
+__host__ __device__ constexpr auto operator/(Number<Y>, Sequence<Xs...>)
+{
+    return Sequence<(Y / Xs)...>{};
+}
+
+template <index_t Y, index_t... Xs>
+__host__ __device__ constexpr auto operator%(Number<Y>, Sequence<Xs...>)
+{
+    return Sequence<(Y % Xs)...>{};
+}
+
+template <index_t I, index_t... Is>
+__host__ __device__ constexpr auto sequence_pop_front(Sequence<I, Is...>)
+{
+    return Sequence<Is...>{};
+}
+
+template <typename Seq>
+__host__ __device__ constexpr auto sequence_pop_back(Seq)
+{
+    static_assert(Seq::Size() > 0, "wrong! cannot pop an empty Sequence!");
+    return sequence_pop_front(Seq::Reverse()).Reverse();
+}
+
+template <typename... Seqs>
+__host__ __device__ constexpr auto merge_sequences(Seqs...)
+{
+    return typename sequence_merge<Seqs...>::type{};
+}
+
+template <typename F, index_t... Xs>
+__host__ __device__ constexpr auto transform_sequences(F f, Sequence<Xs...>)
+{
+    return Sequence<f(Xs)...>{};
+}
+
+template <typename F, index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr auto transform_sequences(F f, Sequence<Xs...>, Sequence<Ys...>)
+{
+    static_assert(Sequence<Xs...>::mSize == Sequence<Ys...>::mSize, "Dim not the same");
+
+    return Sequence<f(Xs, Ys)...>{};
+}
+
+template <typename F, index_t... Xs, index_t... Ys, index_t... Zs>
+__host__ __device__ constexpr auto
+transform_sequences(F f, Sequence<Xs...>, Sequence<Ys...>, Sequence<Zs...>)
+{
+    static_assert(Sequence<Xs...>::mSize == Sequence<Ys...>::mSize &&
+                      Sequence<Xs...>::mSize == Sequence<Zs...>::mSize,
+                  "Dim not the same");
+
+    return Sequence<f(Xs, Ys, Zs)...>{};
+}
+
+template <typename Seq, typename Reduce, index_t Init>
+__host__ __device__ constexpr auto reverse_inclusive_scan_sequence(Seq, Reduce, Number<Init>)
+{
+    return typename sequence_reverse_inclusive_scan<Seq, Reduce, Init>::type{};
+}
+
+template <typename Seq, typename Reduce, index_t Init>
+__host__ __device__ constexpr auto reverse_exclusive_scan_sequence(Seq, Reduce, Number<Init>)
+{
+    return reverse_inclusive_scan_sequence(Seq::PopFront(), Reduce{}, Number<Init>{})
+        .PushBack(Number<Init>{});
+}
+
+template <typename Seq, typename Reduce, index_t Init>
+__host__ __device__ constexpr auto inclusive_scan_sequence(Seq, Reduce, Number<Init>)
+{
+    return reverse_inclusive_scan_sequence(Seq{}.Reverse(), Reduce{}, Number<Init>{}).Reverse();
+}
+
+template <typename Seq, index_t... Is>
+__host__ __device__ constexpr auto pick_sequence_elements_by_ids(Seq, Sequence<Is...> /* ids */)
+{
+    return Sequence<Seq::At(Number<Is>{})...>{};
+}
+
+#if 1
+namespace detail {
+template <typename WorkSeq, typename RemainSeq, typename RemainMask>
+struct pick_sequence_elements_by_mask_impl
+{
+    using new_work_seq = typename conditional<RemainMask::Front(),
+                                              decltype(WorkSeq::PushBack(RemainSeq::Front())),
+                                              WorkSeq>::type;
+
+    using type =
+        typename pick_sequence_elements_by_mask_impl<new_work_seq,
+                                                     decltype(RemainSeq::PopFront()),
+                                                     decltype(RemainMask::PopFront())>::type;
+};
+
+template <typename WorkSeq>
+struct pick_sequence_elements_by_mask_impl<WorkSeq, Sequence<>, Sequence<>>
+{
+    using type = WorkSeq;
+};
+
+} // namespace detail
+
+template <typename Seq, typename Mask>
+__host__ __device__ constexpr auto pick_sequence_elements_by_mask(Seq, Mask)
+{
+    static_assert(Seq::Size() == Mask::Size(), "wrong!");
+
+    return typename detail::pick_sequence_elements_by_mask_impl<Sequence<>, Seq, Mask>::type{};
+}
+
+namespace detail {
+template <typename WorkSeq, typename RemainValues, typename RemainIds>
+struct modify_sequence_elements_by_ids_impl
+{
+    using new_work_seq = decltype(WorkSeq::Modify(RemainIds::Front(), RemainValues::Front()));
+
+    using type =
+        typename modify_sequence_elements_by_ids_impl<new_work_seq,
+                                                      decltype(RemainValues::PopFront()),
+                                                      decltype(RemainIds::PopFront())>::type;
+};
+
+template <typename WorkSeq>
+struct modify_sequence_elements_by_ids_impl<WorkSeq, Sequence<>, Sequence<>>
+{
+    using type = WorkSeq;
+};
+} // namespace detail
+
+template <typename Seq, typename Values, typename Ids>
+__host__ __device__ constexpr auto modify_sequence_elements_by_ids(Seq, Values, Ids)
+{
+    static_assert(Values::Size() == Ids::Size() && Seq::Size() >= Values::Size(), "wrong!");
+
+    return typename detail::modify_sequence_elements_by_ids_impl<Seq, Values, Ids>::type{};
+}
+#endif
+
+template <typename Seq, typename Reduce, index_t Init>
+__host__ __device__ constexpr index_t
+reduce_on_sequence(Seq, Reduce f, Number<Init> /*initial_value*/)
+{
+    index_t result = Init;
+
+    for(index_t i = 0; i < Seq::Size(); ++i)
+    {
+        result = f(result, Seq::At(i));
+    }
+
+    return result;
+}
+
+// TODO: a generic any_of for any container
+template <typename Seq, typename F>
+__host__ __device__ constexpr bool sequence_any_of(Seq, F f)
+{
+    bool flag = false;
+
+    for(index_t i = 0; i < Seq::Size(); ++i)
+    {
+        flag = flag || f(Seq::At(i));
+    }
+
+    return flag;
+}
+
+// TODO: a generic all_of for any container
+template <typename Seq, typename F>
+__host__ __device__ constexpr bool sequence_all_of(Seq, F f)
+{
+    bool flag = true;
+
+    for(index_t i = 0; i < Seq::Size(); ++i)
+    {
+        flag = flag && f(Seq::At(i));
+    }
+
+    return flag;
+}
+
+template <typename Sx, typename Sy>
+using sequence_merge_t = typename sequence_merge<Sx, Sy>::type;
+
+template <index_t NSize, index_t I>
+using uniform_sequence_gen_t = typename uniform_sequence_gen<NSize, I>::type;
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/sequence_helper.hpp b/3rdparty/composable_kernel/include/ck/utility/sequence_helper.hpp
new file mode 100644
index 0000000..db25c27
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/sequence_helper.hpp
@@ -0,0 +1,37 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/tuple.hpp"
+
+namespace ck {
+
+template <index_t... Is>
+__host__ __device__ constexpr auto make_sequence(Number<Is>...)
+{
+    return Sequence<Is...>{};
+}
+
+// F returns index_t
+template <typename F, index_t N>
+__host__ __device__ constexpr auto generate_sequence(F, Number<N>)
+{
+    return typename sequence_gen<N, F>::type{};
+}
+
+// F returns Number<>
+template <typename F, index_t N>
+__host__ __device__ constexpr auto generate_sequence_v2(F&& f, Number<N>)
+{
+    return unpack([&f](auto&&... xs) { return make_sequence(f(xs)...); },
+                  typename arithmetic_sequence_gen<0, N, 1>::type{});
+}
+
+template <index_t... Is>
+__host__ __device__ constexpr auto to_sequence(Tuple<Number<Is>...>)
+{
+    return Sequence<Is...>{};
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/span.hpp b/3rdparty/composable_kernel/include/ck/utility/span.hpp
new file mode 100644
index 0000000..1e50121
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/span.hpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstddef>
+#include <array>
+#include <type_traits>
+
+namespace ck {
+
+template <typename T>
+class span
+{
+    public:
+    using element_type    = T;
+    using value_type      = std::remove_cv_t<element_type>;
+    using size_type       = std::size_t;
+    using difference_type = std::ptrdiff_t;
+    using pointer         = element_type*;
+    using const_pointer   = const element_type*;
+    using reference       = element_type&;
+    using const_reference = const element_type&;
+    using iterator        = pointer;
+    using const_iterator  = pointer;
+
+    constexpr span() : span(nullptr, size_type{0}) {}
+
+    constexpr span(pointer first, size_type count) : ptr_(first), size_(count) {}
+
+    constexpr span(pointer first, pointer last) : span(first, last - first) {}
+
+    template <std::size_t N>
+    constexpr span(element_type (&arr)[N]) noexcept : span(arr, N)
+    {
+    }
+
+    template <std::size_t N>
+    constexpr span(std::array<value_type, N>& arr) noexcept : span(arr.data(), N)
+    {
+    }
+
+    template <typename Container>
+    constexpr span(const Container& container) : span(container.data(), container.size())
+    {
+    }
+
+    constexpr iterator begin() const noexcept { return ptr_; }
+    constexpr const_iterator cbegin() const noexcept { return begin(); }
+
+    constexpr iterator end() const noexcept { return begin() + size(); }
+    constexpr const_iterator cend() const noexcept { return end(); }
+
+    constexpr reference front() const { return *begin(); }
+    constexpr reference back() const { return *(--end()); }
+
+    constexpr reference operator[](size_type idx) const { return *(begin() + idx); }
+    constexpr pointer data() const noexcept { return ptr_; }
+
+    constexpr size_type size() const noexcept { return size_; }
+
+    private:
+    pointer ptr_;
+    size_type size_;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/static_buffer.hpp b/3rdparty/composable_kernel/include/ck/utility/static_buffer.hpp
new file mode 100644
index 0000000..dd25c96
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/static_buffer.hpp
@@ -0,0 +1,195 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "statically_indexed_array.hpp"
+
+namespace ck {
+
+// static buffer for scalar
+template <AddressSpaceEnum AddressSpace,
+          typename T,
+          index_t N,
+          bool InvalidElementUseNumericalZeroValue> // TODO remove this bool, no longer needed
+struct StaticBuffer : public StaticallyIndexedArray<T, N>
+{
+    using type = T;
+    using base = StaticallyIndexedArray<T, N>;
+
+    __host__ __device__ constexpr StaticBuffer() : base{} {}
+
+    template <typename... Ys>
+    __host__ __device__ constexpr StaticBuffer& operator=(const Tuple<Ys...>& y)
+    {
+        static_assert(base::Size() == sizeof...(Ys), "wrong! size not the same");
+        StaticBuffer& x = *this;
+        static_for<0, base::Size(), 1>{}([&](auto i) { x(i) = y[i]; });
+        return x;
+    }
+
+    __host__ __device__ constexpr StaticBuffer& operator=(const T& y)
+    {
+        StaticBuffer& x = *this;
+        static_for<0, base::Size(), 1>{}([&](auto i) { x(i) = y; });
+        return x;
+    }
+
+    __host__ __device__ static constexpr AddressSpaceEnum GetAddressSpace() { return AddressSpace; }
+
+    __host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
+
+    __host__ __device__ static constexpr bool IsDynamicBuffer() { return false; }
+
+    // read access
+    template <index_t I>
+    __host__ __device__ constexpr const T& operator[](Number<I> i) const
+    {
+        return base::operator[](i);
+    }
+
+    // write access
+    template <index_t I>
+    __host__ __device__ constexpr T& operator()(Number<I> i)
+    {
+        return base::operator()(i);
+    }
+
+    __host__ __device__ void Set(T x)
+    {
+        static_for<0, N, 1>{}([&](auto i) { operator()(i) = T{x}; });
+    }
+
+    __host__ __device__ void Clear() { Set(T{0}); }
+};
+
+// static buffer for vector
+template <AddressSpaceEnum AddressSpace,
+          typename S,
+          index_t NumOfVector,
+          index_t ScalarPerVector,
+          bool InvalidElementUseNumericalZeroValue, // TODO remove this bool, no longer needed,
+          typename enable_if<is_scalar_type<S>::value, bool>::type = false>
+struct StaticBufferTupleOfVector
+    : public StaticallyIndexedArray<vector_type<S, ScalarPerVector>, NumOfVector>
+{
+    using V    = typename vector_type<S, ScalarPerVector>::type;
+    using base = StaticallyIndexedArray<vector_type<S, ScalarPerVector>, NumOfVector>;
+
+    static constexpr auto s_per_v   = Number<ScalarPerVector>{};
+    static constexpr auto num_of_v_ = Number<NumOfVector>{};
+    static constexpr auto s_per_buf = s_per_v * num_of_v_;
+
+    __host__ __device__ constexpr StaticBufferTupleOfVector() : base{} {}
+
+    __host__ __device__ static constexpr AddressSpaceEnum GetAddressSpace() { return AddressSpace; }
+
+    __host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
+
+    __host__ __device__ static constexpr bool IsDynamicBuffer() { return false; }
+
+    __host__ __device__ static constexpr index_t Size() { return s_per_buf; };
+
+    // Get S
+    // i is offset of S
+    template <index_t I>
+    __host__ __device__ constexpr const S& operator[](Number<I> i) const
+    {
+        constexpr auto i_v = i / s_per_v;
+        constexpr auto i_s = i % s_per_v;
+
+        return base::operator[](i_v).template AsType<S>()[i_s];
+    }
+
+    // Set S
+    // i is offset of S
+    template <index_t I>
+    __host__ __device__ constexpr S& operator()(Number<I> i)
+    {
+        constexpr auto i_v = i / s_per_v;
+        constexpr auto i_s = i % s_per_v;
+
+        return base::operator()(i_v).template AsType<S>()(i_s);
+    }
+
+    // Get X
+    // i is offset of S, not X. i should be aligned to X
+    template <typename X,
+              index_t I,
+              typename enable_if<has_same_scalar_type<S, X>::value, bool>::type = false>
+    __host__ __device__ constexpr auto GetAsType(Number<I> i) const
+    {
+        constexpr auto s_per_x = Number<scalar_type<remove_cvref_t<X>>::vector_size>{};
+
+        static_assert(s_per_v % s_per_x == 0, "wrong! V must  one or multiple X");
+        static_assert(i % s_per_x == 0, "wrong!");
+
+        constexpr auto i_v = i / s_per_v;
+        constexpr auto i_x = (i % s_per_v) / s_per_x;
+
+        return base::operator[](i_v).template AsType<X>()[i_x];
+    }
+
+    // Set X
+    // i is offset of S, not X. i should be aligned to X
+    template <typename X,
+              index_t I,
+              typename enable_if<has_same_scalar_type<S, X>::value, bool>::type = false>
+    __host__ __device__ constexpr void SetAsType(Number<I> i, X x)
+    {
+        constexpr auto s_per_x = Number<scalar_type<remove_cvref_t<X>>::vector_size>{};
+
+        static_assert(s_per_v % s_per_x == 0, "wrong! V must contain one or multiple X");
+        static_assert(i % s_per_x == 0, "wrong!");
+
+        constexpr auto i_v = i / s_per_v;
+        constexpr auto i_x = (i % s_per_v) / s_per_x;
+
+        base::operator()(i_v).template AsType<X>()(i_x) = x;
+    }
+
+    // Get read access to vector_type V
+    // i is offset of S, not V. i should be aligned to V
+    template <index_t I>
+    __host__ __device__ constexpr const auto& GetVectorTypeReference(Number<I> i) const
+    {
+        static_assert(i % s_per_v == 0, "wrong!");
+
+        constexpr auto i_v = i / s_per_v;
+
+        return base::operator[](i_v);
+    }
+
+    // Get write access to vector_type V
+    // i is offset of S, not V. i should be aligned to V
+    template <index_t I>
+    __host__ __device__ constexpr auto& GetVectorTypeReference(Number<I> i)
+    {
+        static_assert(i % s_per_v == 0, "wrong!");
+
+        constexpr auto i_v = i / s_per_v;
+
+        return base::operator()(i_v);
+    }
+
+    __host__ __device__ void Clear()
+    {
+        constexpr index_t NumScalars = NumOfVector * ScalarPerVector;
+
+        static_for<0, NumScalars, 1>{}([&](auto i) { SetAsType(i, S{0}); });
+    }
+};
+
+template <AddressSpaceEnum AddressSpace, typename T, index_t N>
+__host__ __device__ constexpr auto make_static_buffer(Number<N>)
+{
+    return StaticBuffer<AddressSpace, T, N, true>{};
+}
+
+template <AddressSpaceEnum AddressSpace, typename T, long_index_t N>
+__host__ __device__ constexpr auto make_static_buffer(LongNumber<N>)
+{
+    return StaticBuffer<AddressSpace, T, N, true>{};
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/statically_indexed_array.hpp b/3rdparty/composable_kernel/include/ck/utility/statically_indexed_array.hpp
new file mode 100644
index 0000000..3438776
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/statically_indexed_array.hpp
@@ -0,0 +1,105 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_STATICALLY_INDEXED_ARRAY_HPP
+#define CK_STATICALLY_INDEXED_ARRAY_HPP
+
+#include "functional2.hpp"
+#include "sequence.hpp"
+#include "tuple.hpp"
+
+namespace ck {
+
+namespace detail {
+template <typename X, typename Y>
+struct tuple_concat;
+
+template <typename... Xs, typename... Ys>
+struct tuple_concat<Tuple<Xs...>, Tuple<Ys...>>
+{
+    using type = Tuple<Xs..., Ys...>;
+};
+
+template <typename T, index_t N>
+struct StaticallyIndexedArrayImpl
+{
+    using type =
+        typename tuple_concat<typename StaticallyIndexedArrayImpl<T, N / 2>::type,
+                              typename StaticallyIndexedArrayImpl<T, N - N / 2>::type>::type;
+};
+
+template <typename T>
+struct StaticallyIndexedArrayImpl<T, 0>
+{
+    using type = Tuple<>;
+};
+
+template <typename T>
+struct StaticallyIndexedArrayImpl<T, 1>
+{
+    using type = Tuple<T>;
+};
+} // namespace detail
+
+template <typename T, index_t N>
+using StaticallyIndexedArray = typename detail::StaticallyIndexedArrayImpl<T, N>::type;
+
+template <typename X, typename... Xs>
+__host__ __device__ constexpr auto make_statically_indexed_array(const X& x, const Xs&... xs)
+{
+    return StaticallyIndexedArray<X, sizeof...(Xs) + 1>(x, static_cast<X>(xs)...);
+}
+
+// make empty StaticallyIndexedArray
+template <typename X>
+__host__ __device__ constexpr auto make_statically_indexed_array()
+{
+    return StaticallyIndexedArray<X, 0>();
+}
+
+template <typename T, index_t N>
+struct StaticallyIndexedArray_v2
+{
+    __host__ __device__ constexpr StaticallyIndexedArray_v2() = default;
+
+    __host__ __device__ static constexpr index_t Size() { return N; }
+
+    // read access
+    template <index_t I>
+    __host__ __device__ constexpr const auto& At(Number<I>) const
+    {
+        static_assert(I < N, "wrong! out of range");
+
+        return data_[I];
+    }
+
+    // write access
+    template <index_t I>
+    __host__ __device__ constexpr auto& At(Number<I>)
+    {
+        static_assert(I < N, "wrong! out of range");
+
+        return data_[I];
+    }
+
+    // read access
+    template <index_t I>
+    __host__ __device__ constexpr const auto& operator[](Number<I> i) const
+    {
+        return At(i);
+    }
+
+    // write access
+    template <index_t I>
+    __host__ __device__ constexpr auto& operator()(Number<I> i)
+    {
+        return At(i);
+    }
+
+    __host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
+
+    T data_[N];
+};
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/statically_indexed_array_multi_index.hpp b/3rdparty/composable_kernel/include/ck/utility/statically_indexed_array_multi_index.hpp
new file mode 100644
index 0000000..21b2941
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/statically_indexed_array_multi_index.hpp
@@ -0,0 +1,162 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_STATICALLY_INDEXED_ARRAY_MULTI_INDEX_HPP
+#define CK_STATICALLY_INDEXED_ARRAY_MULTI_INDEX_HPP
+
+#include "common_header.hpp"
+
+namespace ck {
+
+template <index_t N>
+using MultiIndex = StaticallyIndexedArray<index_t, N>;
+
+template <typename... Xs>
+__host__ __device__ constexpr auto make_multi_index(Xs&&... xs)
+{
+    return make_statically_indexed_array<index_t>(index_t{xs}...);
+}
+
+template <index_t NSize>
+__host__ __device__ constexpr auto make_zero_multi_index()
+{
+    return unpack([](auto... xs) { return make_multi_index(xs...); },
+                  typename uniform_sequence_gen<NSize, 0>::type{});
+}
+
+template <typename T>
+__host__ __device__ constexpr auto to_multi_index(const T& x)
+{
+    return unpack([](auto... ys) { return make_multi_index(ys...); }, x);
+}
+
+// Here should use MultiIndex<NSize>, instead of Tuple<Ys...>, although the former
+// is the alias of the latter. This is because compiler cannot infer the NSize if
+// using MultiIndex<NSize>
+// TODO: how to fix this?
+template <
+    typename... Ys,
+    typename X,
+    enable_if_t<!std::is_integral<X>::value && !std::is_floating_point<X>::value, bool> = false>
+__host__ __device__ constexpr auto operator+=(Tuple<Ys...>& y, const X& x)
+{
+    static_assert(X::Size() == sizeof...(Ys), "wrong! size not the same");
+    constexpr index_t NSize = sizeof...(Ys);
+    static_for<0, NSize, 1>{}([&](auto i) { y(i) += x[i]; });
+    return y;
+}
+
+template <
+    typename... Ys,
+    typename X,
+    enable_if_t<!std::is_integral<X>::value && !std::is_floating_point<X>::value, bool> = false>
+__host__ __device__ constexpr auto operator-=(Tuple<Ys...>& y, const X& x)
+{
+    static_assert(X::Size() == sizeof...(Ys), "wrong! size not the same");
+    constexpr index_t NSize = sizeof...(Ys);
+    static_for<0, NSize, 1>{}([&](auto i) { y(i) -= x[i]; });
+    return y;
+}
+
+template <
+    typename... Xs,
+    typename Y,
+    enable_if_t<!std::is_integral<Y>::value && !std::is_floating_point<Y>::value, bool> = false>
+__host__ __device__ constexpr auto operator+(const Tuple<Xs...>& x, const Y& y)
+{
+    static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = x[i] + y[i]; });
+    return r;
+}
+
+template <
+    typename... Xs,
+    typename Y,
+    enable_if_t<!std::is_integral<Y>::value && !std::is_floating_point<Y>::value, bool> = false>
+__host__ __device__ constexpr auto operator-(const Tuple<Xs...>& x, const Y& y)
+{
+    static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = x[i] - y[i]; });
+    return r;
+}
+
+template <
+    typename... Xs,
+    typename Y,
+    enable_if_t<!std::is_integral<Y>::value && !std::is_floating_point<Y>::value, bool> = false>
+__host__ __device__ constexpr auto operator*(const Tuple<Xs...>& x, const Y& y)
+{
+    static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = x[i] * y[i]; });
+    return r;
+}
+
+// MultiIndex = scalar * MultiIndex
+template <typename... Xs,
+          typename Y,
+          enable_if_t<std::is_integral<Y>::value || std::is_floating_point<Y>::value, bool> = false>
+__host__ __device__ constexpr auto operator*(Y a, const Tuple<Xs...>& x)
+{
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = a * x[i]; });
+    return r;
+}
+
+// MultiIndex = MultiIndex * scalar
+template <typename... Xs,
+          typename Y,
+          enable_if_t<std::is_integral<Y>::value || std::is_floating_point<Y>::value, bool> = false>
+__host__ __device__ constexpr auto operator*(const Tuple<Xs...>& x, Y a)
+{
+    return a * x;
+}
+
+namespace mathext {
+
+template <typename... Xs>
+__host__ __device__ constexpr auto exp(const Tuple<Xs...>& x)
+{
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = math::exp(x[i]); });
+    return r;
+}
+
+template <typename... Xs, typename Y>
+__host__ __device__ constexpr auto max(const Tuple<Xs...>& x, const Y& y)
+{
+    static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = math::max(x[i], y[i]); });
+    return r;
+}
+
+} // namespace mathext
+
+template <typename... Xs>
+__host__ __device__ void print_multi_index(const Tuple<Xs...>& x)
+{
+    printf("{");
+    printf("MultiIndex, ");
+    printf("size %d,", index_t{sizeof...(Xs)});
+    static_for<0, sizeof...(Xs), 1>{}(
+        [&](auto i) { printf("%d ", static_cast<index_t>(x.At(i))); });
+    printf("}");
+}
+
+} // namespace ck
+#endif
diff --git a/3rdparty/composable_kernel/include/ck/utility/synchronization.hpp b/3rdparty/composable_kernel/include/ck/utility/synchronization.hpp
new file mode 100644
index 0000000..ec8b689
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/synchronization.hpp
@@ -0,0 +1,32 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+
+namespace ck {
+
+__device__ void block_sync_lds()
+{
+#if CK_EXPERIMENTAL_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM
+    asm volatile("\
+    s_waitcnt lgkmcnt(0) \n \
+    s_barrier \
+    " ::);
+#else
+    __syncthreads();
+#endif
+}
+__device__ void s_nop()
+{
+#if 1
+    asm volatile("\
+    s_nop 0 \n \
+    " ::);
+#else
+    __builtin_amdgcn_s_barrier();
+#endif
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/thread_group.hpp b/3rdparty/composable_kernel/include/ck/utility/thread_group.hpp
new file mode 100644
index 0000000..d469dec
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/thread_group.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "get_id.hpp"
+
+namespace ck {
+
+template <index_t ThreadPerBlock>
+struct ThisThreadBlock
+{
+    static constexpr index_t kNumThread_ = ThreadPerBlock;
+
+    __device__ static constexpr index_t GetNumOfThread() { return kNumThread_; }
+
+    __device__ static constexpr bool IsBelong() { return true; }
+
+    __device__ static index_t GetThreadId() { return get_thread_local_1d_id(); }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/transpose_vectors.hpp b/3rdparty/composable_kernel/include/ck/utility/transpose_vectors.hpp
new file mode 100644
index 0000000..2b0075d
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/transpose_vectors.hpp
@@ -0,0 +1,165 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "statically_indexed_array.hpp"
+#include "data_type.hpp"
+
+namespace ck {
+
+template <typename S,
+          index_t NX,
+          index_t NY,
+          typename enable_if<is_scalar_type<S>::value, bool>::type = false>
+struct transpose_vectors;
+
+// transpose fp16 2x2
+__device__ void transpose_fp16_2x2(const half2_t& x0, const half2_t& x1, half2_t& y0, half2_t& y1)
+{
+#if 0
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    const vector_type<half_t, 2> vx0{x0}, vx1{x1};
+    vector_type<half_t, 2> vy0, vy1;
+
+    vy0.template AsType<half_t>()(I0) = vx0.template AsType<half_t>()[I0];
+    vy0.template AsType<half_t>()(I1) = vx1.template AsType<half_t>()[I0];
+
+    vy1.template AsType<half_t>()(I0) = vx0.template AsType<half_t>()[I1];
+    vy1.template AsType<half_t>()(I1) = vx1.template AsType<half_t>()[I1];
+
+    y0 = vy0.template AsType<half2_t>()[I0];
+    y1 = vy1.template AsType<half2_t>()[I0];
+#else
+    constexpr int32_t m0 = 0x05040100;
+    constexpr int32_t m1 = 0x07060302;
+
+    // ex: v_perm_b32(0x 11 22 33 44, 0x 55 66 77 88, 0x 05 01 04 00) -> 0x33774488
+    //                   -- -- -- --     -- -- -- --      -  -  -  -
+    //             index  7  6  5  4      3  2  1  0     33 77 44 88
+    // index is reversed because of little endianness (least significant bits first)
+    y0 = bit_cast<half2_t>(__builtin_amdgcn_perm(bit_cast<int32_t>(x1), bit_cast<int32_t>(x0), m0));
+    y1 = bit_cast<half2_t>(__builtin_amdgcn_perm(bit_cast<int32_t>(x1), bit_cast<int32_t>(x0), m1));
+#endif
+}
+
+template <index_t NX, index_t NY>
+struct transpose_vectors<half_t, NX, NY>
+{
+    // we got [NY * NX] amount of S data to be transposed
+    static constexpr index_t s_per_x = NY;
+    static constexpr index_t s_per_y = NX;
+
+    using S  = half_t;
+    using VX = vector_type<half_t, s_per_x>;
+    using VY = vector_type<half_t, s_per_y>;
+
+    __device__ void operator()(const StaticallyIndexedArray<const VX&, NX>& vx_tuple,
+                               StaticallyIndexedArray<VY&, NY>& vy_tuple)
+    {
+        static constexpr auto I1 = Number<1>{};
+        static constexpr auto I2 = Number<2>{};
+
+        static_assert((NX % 2 == 0 && NY % 2 == 0), "wrong!");
+
+        // loop over 2x2 tile and transpose data from vx_tuple into vy_tuple
+        static_for<0, NY, 2>{}([&](auto iy) {
+            static_for<0, NX, 2>{}([&](auto ix) {
+                // reference to 2 half2_t data from vx_tuple
+                const auto& x_s2_0 = vx_tuple[ix].template AsType<half2_t>()[iy / I2];
+                const auto& x_s2_1 = vx_tuple[ix + I1].template AsType<half2_t>()[iy / I2];
+
+                // reference to 2 half2_t data from vy_tuple
+                auto& y_s2_0 = vy_tuple(iy).template AsType<half2_t>()(ix / I2);
+                auto& y_s2_1 = vy_tuple(iy + I1).template AsType<half2_t>()(ix / I2);
+
+                // transpose
+                transpose_fp16_2x2(x_s2_0, x_s2_1, y_s2_0, y_s2_1);
+            });
+        });
+    }
+};
+
+// transpose int8 4x4
+__device__ void transpose_int8_4x4(const int8x4_t& x0,
+                                   const int8x4_t& x1,
+                                   const int8x4_t& x2,
+                                   const int8x4_t& x3,
+                                   int8x4_t& y0,
+                                   int8x4_t& y1,
+                                   int8x4_t& y2,
+                                   int8x4_t& y3)
+{
+    int32_t t0, t1;
+    int32_t z0, z1, z2, z3;
+    constexpr int32_t m0 = 0x05010400;
+    constexpr int32_t m1 = 0x05040100;
+    constexpr int32_t m2 = 0x07060302;
+    constexpr int32_t m3 = 0x07030602;
+
+    // ex: v_perm_b32(0x 11 22 33 44, 0x 55 66 77 88, 0x 05 01 04 00) -> 0x33774488
+    //                   -- -- -- --     -- -- -- --      -  -  -  -
+    //             index  7  6  5  4      3  2  1  0     33 77 44 88
+    // index is reversed because of little endianness (least significant bits first)
+    t0 = __builtin_amdgcn_perm(bit_cast<int32_t>(x1), bit_cast<int32_t>(x0), m0);
+    t1 = __builtin_amdgcn_perm(bit_cast<int32_t>(x3), bit_cast<int32_t>(x2), m0);
+    z0 = __builtin_amdgcn_perm(bit_cast<int32_t>(t1), bit_cast<int32_t>(t0), m1);
+    z1 = __builtin_amdgcn_perm(bit_cast<int32_t>(t1), bit_cast<int32_t>(t0), m2);
+    t0 = __builtin_amdgcn_perm(bit_cast<int32_t>(x1), bit_cast<int32_t>(x0), m3);
+    t1 = __builtin_amdgcn_perm(bit_cast<int32_t>(x3), bit_cast<int32_t>(x2), m3);
+    z2 = __builtin_amdgcn_perm(bit_cast<int32_t>(t1), bit_cast<int32_t>(t0), m1);
+    z3 = __builtin_amdgcn_perm(bit_cast<int32_t>(t1), bit_cast<int32_t>(t0), m2);
+
+    y0 = bit_cast<int8x4_t>(z0);
+    y1 = bit_cast<int8x4_t>(z1);
+    y2 = bit_cast<int8x4_t>(z2);
+    y3 = bit_cast<int8x4_t>(z3);
+}
+
+template <index_t NX, index_t NY>
+struct transpose_vectors<int8_t, NX, NY>
+{
+    // we got [NY * NX] amount of S data to be transposed
+    static constexpr index_t s_per_x = NY;
+    static constexpr index_t s_per_y = NX;
+
+    using S  = int8_t;
+    using VX = vector_type<int8_t, s_per_x>;
+    using VY = vector_type<int8_t, s_per_y>;
+
+    __device__ void operator()(const StaticallyIndexedArray<const VX&, NX>& vx_tuple,
+                               StaticallyIndexedArray<VY&, NY>& vy_tuple)
+    {
+        static constexpr auto I1 = Number<1>{};
+        static constexpr auto I2 = Number<2>{};
+        static constexpr auto I3 = Number<3>{};
+        static constexpr auto I4 = Number<4>{};
+
+        static_assert((NX % 4 == 0 && NY % 4 == 0), "wrong!");
+
+        // loop over 4x4 tile and transpose data from vx_tuple into vy_tuple
+        static_for<0, NY, 4>{}([&](auto iy) {
+            static_for<0, NX, 4>{}([&](auto ix) {
+                // reference to 4 int8 data from vx_tuple
+                const auto& x_s4_0 = vx_tuple[ix].template AsType<int8x4_t>()[iy / I4];
+                const auto& x_s4_1 = vx_tuple[ix + I1].template AsType<int8x4_t>()[iy / I4];
+                const auto& x_s4_2 = vx_tuple[ix + I2].template AsType<int8x4_t>()[iy / I4];
+                const auto& x_s4_3 = vx_tuple[ix + I3].template AsType<int8x4_t>()[iy / I4];
+
+                // reference to 4 int8 data from vy_tuple
+                auto& y_s4_0 = vy_tuple(iy).template AsType<int8x4_t>()(ix / I4);
+                auto& y_s4_1 = vy_tuple(iy + I1).template AsType<int8x4_t>()(ix / I4);
+                auto& y_s4_2 = vy_tuple(iy + I2).template AsType<int8x4_t>()(ix / I4);
+                auto& y_s4_3 = vy_tuple(iy + I3).template AsType<int8x4_t>()(ix / I4);
+
+                // transpose
+                transpose_int8_4x4(x_s4_0, x_s4_1, x_s4_2, x_s4_3, y_s4_0, y_s4_1, y_s4_2, y_s4_3);
+            });
+        });
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/tuple.hpp b/3rdparty/composable_kernel/include/ck/utility/tuple.hpp
new file mode 100644
index 0000000..d8664be
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/tuple.hpp
@@ -0,0 +1,221 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/integral_constant.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/utility/enable_if.hpp"
+
+namespace ck {
+
+namespace detail {
+
+template <index_t>
+struct TupleElementKey
+{
+    __host__ __device__ constexpr TupleElementKey() = default;
+};
+
+template <typename Key, typename Data>
+struct TupleElementKeyData
+{
+    using DataType = Data;
+
+#if 0 // workaround compiler complaint about implicitly-deleted default constructor
+    __host__ __device__ constexpr TupleElementKeyData() = default;
+#else
+    __host__ __device__ constexpr TupleElementKeyData() : mData{} {}
+#endif
+
+    template <typename T,
+              typename enable_if<!is_same<remove_cvref_t<T>, TupleElementKeyData>::value,
+                                 bool>::type = false>
+    __host__ __device__ constexpr TupleElementKeyData(T&& v) : mData(std::forward<T>(v))
+    {
+    }
+
+    DataType mData;
+};
+
+// for read access of tuple element
+template <typename Key, typename Data>
+__host__ __device__ constexpr const Data&
+get_tuple_element_data_reference(const TupleElementKeyData<Key, Data>& x)
+{
+    return static_cast<const Data&>(x.mData);
+}
+
+// for write access of tuple element
+template <typename Key, typename Data>
+__host__ __device__ constexpr Data&
+get_tuple_element_data_reference(TupleElementKeyData<Key, Data>& x)
+{
+    return x.mData;
+}
+
+// TODO: not sure the use of reference is correct
+template <typename Key, typename Data>
+__host__ __device__ constexpr Data&&
+get_tuple_element_data_reference(TupleElementKeyData<Key, Data>&& x)
+{
+    return static_cast<Data&&>(x.mData);
+}
+
+// for infering type of tuple element
+template <typename Key, typename Data>
+__host__ __device__ constexpr Data get_tuple_element_data(const TupleElementKeyData<Key, Data>& x)
+{
+    return std::forward(x.mData);
+}
+
+template <typename Indices, typename... Xs>
+struct TupleImpl;
+
+template <index_t... Is, typename... Xs>
+struct TupleImpl<Sequence<Is...>, Xs...> : TupleElementKeyData<TupleElementKey<Is>, Xs>...
+{
+    __host__ __device__ constexpr TupleImpl() = default;
+
+    template <typename Y,
+              typename enable_if<sizeof...(Is) == 1 && sizeof...(Xs) == 1 &&
+                                     !is_same<remove_cvref_t<Y>, TupleImpl>::value,
+                                 bool>::type = false>
+    __host__ __device__ constexpr TupleImpl(Y&& y)
+        : TupleElementKeyData<TupleElementKey<Is>, Xs>(std::forward<Y>(y))...
+    {
+    }
+
+    template <typename... Ys, typename enable_if<sizeof...(Ys) >= 2, bool>::type = false>
+    __host__ __device__ constexpr TupleImpl(Ys&&... ys)
+        : TupleElementKeyData<TupleElementKey<Is>, Xs>(std::forward<Ys>(ys))...
+    {
+        static_assert(sizeof...(Is) == sizeof...(Xs) && sizeof...(Is) == sizeof...(Ys),
+                      "wrong! inconsistent size");
+    }
+
+    __host__ __device__ static constexpr index_t Size() { return sizeof...(Xs); }
+
+    template <index_t I>
+    __host__ __device__ constexpr const auto& GetElementDataByKey(TupleElementKey<I>) const
+    {
+        return get_tuple_element_data_reference<TupleElementKey<I>>(*this);
+    }
+
+    template <index_t I>
+    __host__ __device__ constexpr auto& GetElementDataByKey(TupleElementKey<I>)
+    {
+        return get_tuple_element_data_reference<TupleElementKey<I>>(*this);
+    }
+};
+
+} // namespace detail
+
+template <typename... Xs>
+struct Tuple : detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(Xs), 1>::type, Xs...>
+{
+    using base =
+        detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(Xs), 1>::type, Xs...>;
+
+    __host__ __device__ constexpr Tuple() = default;
+
+    template <typename Y,
+              typename enable_if<sizeof...(Xs) == 1 && !is_same<remove_cvref_t<Y>, Tuple>::value,
+                                 bool>::type = false>
+    __host__ __device__ constexpr Tuple(Y&& y) : base(std::forward<Y>(y))
+    {
+    }
+
+    template <typename... Ys,
+              typename enable_if<sizeof...(Ys) == sizeof...(Xs) && sizeof...(Ys) >= 2, bool>::type =
+                  false>
+    __host__ __device__ constexpr Tuple(Ys&&... ys) : base(std::forward<Ys>(ys)...)
+    {
+    }
+
+    __host__ __device__ static constexpr index_t Size() { return sizeof...(Xs); }
+
+    // read access
+    template <index_t I>
+    __host__ __device__ constexpr const auto& At(Number<I>) const
+    {
+        static_assert(I < base::Size(), "wrong! out of range");
+        return base::GetElementDataByKey(detail::TupleElementKey<I>{});
+    }
+
+    // write access
+    template <index_t I>
+    __host__ __device__ constexpr auto& At(Number<I>)
+    {
+        static_assert(I < base::Size(), "wrong! out of range");
+        return base::GetElementDataByKey(detail::TupleElementKey<I>{});
+    }
+
+    // read access
+    template <index_t I>
+    __host__ __device__ constexpr const auto& operator[](Number<I> i) const
+    {
+        return At(i);
+    }
+
+    // write access
+    template <index_t I>
+    __host__ __device__ constexpr auto& operator()(Number<I> i)
+    {
+        return At(i);
+    }
+
+    template <typename T>
+    __host__ __device__ constexpr auto operator=(const T& a)
+    {
+        static_assert(T::Size() == Size(), "wrong! size not the same");
+
+        static_for<0, Size(), 1>{}([&](auto i) { operator()(i) = a[i]; });
+
+        return *this;
+    }
+
+    __host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
+};
+
+template <>
+struct Tuple<>
+{
+    __host__ __device__ constexpr Tuple() = default;
+
+    __host__ __device__ static constexpr index_t Size() { return 0; }
+
+    template <typename T>
+    __host__ __device__ constexpr auto operator=(const T&)
+    {
+        return *this;
+    }
+
+    __host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
+};
+
+template <index_t I, typename TTuple>
+struct tuple_element
+{
+    // type should keep the cv/ref qualifier of original tuple element
+    using type = decltype(detail::get_tuple_element_data<detail::TupleElementKey<I>>(TTuple{}));
+};
+
+template <index_t I, typename TTuple>
+using tuple_element_t = typename tuple_element<I, TTuple>::type;
+
+template <typename... Xs>
+__host__ __device__ constexpr auto make_tuple(Xs&&... xs)
+{
+    return Tuple<remove_cvref_t<Xs>...>(std::forward<Xs>(xs)...);
+}
+
+// https://en.cppreference.com/w/cpp/utility/tuple/tie
+template <typename... Args>
+constexpr Tuple<Args&...> tie(Args&... args) noexcept
+{
+    return {args...};
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/tuple_helper.hpp b/3rdparty/composable_kernel/include/ck/utility/tuple_helper.hpp
new file mode 100644
index 0000000..6f5b142
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/tuple_helper.hpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "functional4.hpp"
+#include "tuple.hpp"
+
+namespace ck {
+
+template <typename F, index_t N>
+__host__ __device__ constexpr auto generate_tuple(F&& f, Number<N>)
+{
+    return unpack([&f](auto&&... xs) { return make_tuple(f(xs)...); },
+                  typename arithmetic_sequence_gen<0, N, 1>::type{});
+}
+
+template <typename F, index_t N>
+__host__ __device__ constexpr auto generate_tie(F&& f, Number<N>)
+{
+    return unpack([&f](auto&&... xs) { return tie(f(xs)...); },
+                  typename arithmetic_sequence_gen<0, N, 1>::type{});
+}
+
+// tx and ty are tuple of references, return type of will tuple of referennce (not rvalue)
+template <typename... X, typename... Y>
+__host__ __device__ constexpr auto concat_tuple_of_reference(const Tuple<X&...>& tx,
+                                                             const Tuple<Y&...>& ty)
+{
+    return unpack2(
+        [&](auto&&... zs) { return Tuple<decltype(zs)...>{std::forward<decltype(zs)>(zs)...}; },
+        tx,
+        ty);
+}
+
+namespace detail {
+
+template <typename F, typename X, index_t... Is>
+__host__ __device__ constexpr auto transform_tuples_impl(F f, const X& x, Sequence<Is...>)
+{
+    return make_tuple(f(x.At(Number<Is>{}))...);
+}
+
+template <typename F, typename X, typename Y, index_t... Is>
+__host__ __device__ constexpr auto
+transform_tuples_impl(F f, const X& x, const Y& y, Sequence<Is...>)
+{
+    return make_tuple(f(x.At(Number<Is>{}), y.At(Number<Is>{}))...);
+}
+
+template <typename F, typename X, typename Y, typename Z, index_t... Is>
+__host__ __device__ constexpr auto
+transform_tuples_impl(F f, const X& x, const Y& y, const Z& z, Sequence<Is...>)
+{
+    return make_tuple(f(x.At(Number<Is>{}), y.At(Number<Is>{}), z.At(Number<Is>{}))...);
+}
+
+} // namespace detail
+
+template <typename F, typename X>
+__host__ __device__ constexpr auto transform_tuples(F f, const X& x)
+{
+    return detail::transform_tuples_impl(
+        f, x, typename arithmetic_sequence_gen<0, X::Size(), 1>::type{});
+}
+
+template <typename F, typename X, typename Y>
+__host__ __device__ constexpr auto transform_tuples(F f, const X& x, const Y& y)
+{
+    return detail::transform_tuples_impl(
+        f, x, y, typename arithmetic_sequence_gen<0, X::Size(), 1>::type{});
+}
+
+template <typename F, typename X, typename Y, typename Z>
+__host__ __device__ constexpr auto transform_tuples(F f, const X& x, const Y& y, const Z& z)
+{
+    return detail::transform_tuples_impl(
+        f, x, y, z, typename arithmetic_sequence_gen<0, X::Size(), 1>::type{});
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/include/ck/utility/type.hpp b/3rdparty/composable_kernel/include/ck/utility/type.hpp
new file mode 100644
index 0000000..90b9df2
--- /dev/null
+++ b/3rdparty/composable_kernel/include/ck/utility/type.hpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/integral_constant.hpp"
+#include "ck/utility/enable_if.hpp"
+
+namespace ck {
+
+template <typename X, typename Y>
+struct is_same : public integral_constant<bool, false>
+{
+};
+
+template <typename X>
+struct is_same<X, X> : public integral_constant<bool, true>
+{
+};
+
+template <typename X, typename Y>
+inline constexpr bool is_same_v = is_same<X, Y>::value;
+
+template <typename T>
+using remove_reference_t = typename std::remove_reference<T>::type;
+
+template <typename T>
+using remove_cv_t = typename std::remove_cv<T>::type;
+
+template <typename T>
+using remove_cvref_t = remove_cv_t<std::remove_reference_t<T>>;
+
+template <typename T>
+using remove_pointer_t = typename std::remove_pointer<T>::type;
+
+template <typename T>
+inline constexpr bool is_pointer_v = std::is_pointer<T>::value;
+
+template <typename Y, typename X, typename enable_if<sizeof(X) == sizeof(Y), bool>::type = false>
+__host__ __device__ constexpr Y bit_cast(const X& x)
+{
+#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_BIT_CAST
+    Y y;
+
+    __builtin_memcpy(&y, &x, sizeof(X));
+
+    return y;
+#else
+    union AsType
+    {
+        X x;
+        Y y;
+    };
+
+    return AsType{x}.y;
+#endif
+}
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/CMakeLists.txt b/3rdparty/composable_kernel/library/CMakeLists.txt
new file mode 100644
index 0000000..90873fd
--- /dev/null
+++ b/3rdparty/composable_kernel/library/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_subdirectory(src/tensor_operation_instance/gpu)
+add_subdirectory(src/utility)
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
new file mode 100644
index 0000000..46a1fa5
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
@@ -0,0 +1,138 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct ReferenceBatchedGemm : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_g_m_k,
+                 const Tensor<BDataType>& b_g_k_n,
+                 Tensor<CDataType>& c_g_m_n,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_g_m_k_{a_g_m_k},
+              b_g_k_n_{b_g_k_n},
+              c_g_m_n_{c_g_m_n},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_g_m_k_;
+        const Tensor<BDataType>& b_g_k_n_;
+        Tensor<CDataType>& c_g_m_n_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceBatchedGemm::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_gmk_gkn_gmn = [&](auto g, auto m, auto n) {
+                const int K = arg.a_g_m_k_.mDesc.GetLengths()[2];
+
+                AccDataType v_acc = 0;
+
+                for(int k = 0; k < K; ++k)
+                {
+                    ADataType v_a;
+                    BDataType v_b;
+
+                    arg.a_element_op_(v_a, arg.a_g_m_k_(g, m, k));
+                    arg.b_element_op_(v_b, arg.b_g_k_n_(g, k, n));
+
+                    v_acc +=
+                        ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
+                }
+
+                AccDataType v_c;
+
+                arg.c_element_op_(v_c, v_acc);
+
+                arg.c_g_m_n_(g, m, n) = ck::type_convert<CDataType>(v_c);
+            };
+
+            make_ParallelTensorFunctor(f_gmk_gkn_gmn,
+                                       arg.c_g_m_n_.mDesc.GetLengths()[0],
+                                       arg.c_g_m_n_.mDesc.GetLengths()[1],
+                                       arg.c_g_m_n_.mDesc.GetLengths()[2])(
+                std::thread::hardware_concurrency());
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_g_m_k,
+                             const Tensor<BDataType>& b_g_k_n,
+                             Tensor<CDataType>& c_g_m_n,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{a_g_m_k, b_g_k_n, c_g_m_n, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceBatchedGemm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp
new file mode 100644
index 0000000..0b621e8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp
@@ -0,0 +1,412 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <array>
+#include <algorithm>
+#include <thread>
+
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/ignore.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct ReferenceBatchNormBwd : public device::DeviceBatchNormBwd<XDataType,
+                                                                 DxDataType,
+                                                                 DyDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 DscaleDbiasDataType,
+                                                                 MeanVarDataType,
+                                                                 DyElementwiseOp,
+                                                                 Rank,
+                                                                 NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> dxStrides,
+                 const std::array<index_t, Rank> dyStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                 const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                 const std::array<index_t, NumInvariantDim> bnDscaleDbiasStrides,
+                 const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const DyDataType* p_dy,
+                 const ScaleDataType* p_scale,
+                 const MeanVarDataType* p_savedMean,
+                 const MeanVarDataType* p_savedInvVar,
+                 double epsilon,
+                 const DyElementwiseOp dy_elementwise_op,
+                 DxDataType* p_dx,
+                 DscaleDbiasDataType* p_dscale,
+                 DscaleDbiasDataType* p_dbias)
+            : reduceDims_(reduceDims),
+              bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnDscaleDbiasStrides_(bnDscaleDbiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              p_dy_(p_dy),
+              p_scale_(p_scale),
+              p_savedMean_(p_savedMean),
+              p_savedInvVar_(p_savedInvVar),
+              dy_elementwise_op_(dy_elementwise_op),
+              p_dx_(p_dx),
+              p_dscale_(p_dscale),
+              p_dbias_(p_dbias)
+        {
+            using ck::host_common::get_index_set;
+
+            if(std::any_of(
+                   reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+                throw std::runtime_error("Invalid reduce dimensions!");
+
+            // get invariant_dims[] and invariant_lengths[]
+            for(int dim = 0, i = 0; dim < Rank; dim++)
+                if(std::none_of(
+                       reduceDims.begin(), reduceDims.end(), [&](int d) { return d == dim; }))
+                {
+                    invariantDims_[i]     = dim;
+                    invariant_lengths_[i] = xyLengths[dim];
+                    i++;
+                };
+
+            // get reduce_lengths_[]
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims[j];
+                reduce_lengths_[i++] = xyLengths[dim];
+            };
+
+            for(int i = 0; i < NumInvariantDim; i++)
+                if(invariant_lengths_[i] != bnScaleBiasMeanVarLengths_[i])
+                    throw std::runtime_error("Invalid lengths parameters!");
+
+            for(int j = 0, i = 0; j < NumInvariantDim; j++)
+            {
+                int dim                  = invariantDims_[j];
+                x_invariant_strides_[i]  = xStrides[dim];
+                dy_invariant_strides_[i] = dyStrides[dim];
+                dx_invariant_strides_[i] = dxStrides[dim];
+                i++;
+            };
+
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim               = reduceDims_[j];
+                x_reduce_strides_[i]  = xStrides[dim];
+                dy_reduce_strides_[i] = dyStrides[dim];
+                dx_reduce_strides_[i] = dxStrides[dim];
+                i++;
+            };
+
+            reduceSize_ = std::accumulate(
+                reduce_lengths_.begin(), reduce_lengths_.end(), 1, std::multiplies<size_t>{});
+
+            invariant_index_set_ = get_index_set<NumInvariantDim>(invariant_lengths_);
+            reduce_index_set_    = get_index_set<NumBatchNormReduceDim>(reduce_lengths_);
+
+            epsilon_ = type_convert<AccDataType>(epsilon);
+
+            haveSavedMeanInvVar_ = (p_savedMean != nullptr && p_savedInvVar != nullptr);
+        }
+
+        std::array<int, NumBatchNormReduceDim> reduceDims_;
+        std::array<int, NumInvariantDim> invariantDims_;
+        std::array<index_t, NumInvariantDim> invariant_lengths_;
+        std::array<index_t, NumBatchNormReduceDim> reduce_lengths_;
+
+        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths_;
+        const std::array<index_t, NumInvariantDim> bnScaleStrides_;
+        const std::array<index_t, NumInvariantDim> bnDscaleDbiasStrides_;
+        const std::array<index_t, NumInvariantDim> bnMeanVarStrides_;
+
+        std::array<index_t, NumInvariantDim> x_invariant_strides_;
+        std::array<index_t, NumInvariantDim> dy_invariant_strides_;
+        std::array<index_t, NumInvariantDim> dx_invariant_strides_;
+        std::array<index_t, NumBatchNormReduceDim> x_reduce_strides_;
+        std::array<index_t, NumBatchNormReduceDim> dy_reduce_strides_;
+        std::array<index_t, NumBatchNormReduceDim> dx_reduce_strides_;
+
+        const XDataType* p_x_;
+        const DyDataType* p_dy_;
+        const ScaleDataType* p_scale_;
+        const MeanVarDataType* p_savedMean_;
+        const MeanVarDataType* p_savedInvVar_;
+        const DyElementwiseOp dy_elementwise_op_;
+
+        DxDataType* p_dx_;
+        DscaleDbiasDataType* p_dscale_;
+        DscaleDbiasDataType* p_dbias_;
+
+        bool haveSavedMeanInvVar_;
+
+        std::vector<std::array<index_t, NumInvariantDim>> invariant_index_set_;
+        std::vector<std::array<index_t, NumBatchNormReduceDim>> reduce_index_set_;
+
+        AccDataType epsilon_;
+        size_t reduceSize_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            using ck::host_common::get_offset_from_index;
+
+            auto thread_reduce_func = [&](auto invariant_index) {
+                size_t x_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.x_invariant_strides_, invariant_index);
+                size_t dy_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.dy_invariant_strides_, invariant_index);
+                size_t dx_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.dx_invariant_strides_, invariant_index);
+
+                AccDataType mean     = type_convert<AccDataType>(0.0f);
+                AccDataType variance = type_convert<AccDataType>(0.0f);
+                AccDataType invVar;
+                int32_t curr_count = 0;
+
+                if(arg.haveSavedMeanInvVar_)
+                {
+                    size_t mean_invVar_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                        arg.bnMeanVarStrides_, invariant_index);
+
+                    mean =
+                        type_convert<AccDataType>(arg.p_savedMean_[mean_invVar_invariant_offset]);
+                    invVar =
+                        type_convert<AccDataType>(arg.p_savedInvVar_[mean_invVar_invariant_offset]);
+                }
+                else
+                {
+                    // compute mean, variance using welford method
+                    for(const auto& reduce_index : arg.reduce_index_set_)
+                    {
+                        size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                            arg.x_reduce_strides_, reduce_index);
+
+                        auto x_offset = x_invariant_offset + x_reduce_offset;
+
+                        curr_count++;
+
+                        AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                        AccDataType delta = x - mean;
+
+                        mean += delta / curr_count;
+
+                        AccDataType delta2 = x - mean;
+
+                        variance += delta * delta2;
+                    };
+
+                    // actual variance
+                    variance = variance / curr_count;
+
+                    // inv-variance defined as 1/sqrt(epsilon+variance)
+                    invVar =
+                        type_convert<AccDataType>(1.0f) / ck::math::sqrt(arg.epsilon_ + variance);
+                };
+
+                AccDataType dbias =
+                    type_convert<AccDataType>(0.0f); // Sum on reduced dimensions of dy
+                AccDataType dscale =
+                    type_convert<AccDataType>(0.0f); // Sum on reduced dimensions of dy * norm_x
+
+                // 1) calculate dy * (x - mean) * inv-variance
+                // 2) calculate sum(dy) on reduced dimensions
+                // 3) calculate sum(dy * norm_x) on reduced dimensions
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+                    size_t dy_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.dy_reduce_strides_, reduce_index);
+
+                    auto x_offset  = x_invariant_offset + x_reduce_offset;
+                    auto dy_offset = dy_invariant_offset + dy_reduce_offset;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType norm_x = (x - mean) * invVar;
+                    AccDataType dy     = type_convert<AccDataType>(arg.p_dy_[dy_offset]);
+
+                    arg.dy_elementwise_op_(dy, dy);
+
+                    dbias += dy;
+                    dscale += norm_x * dy;
+                };
+
+                size_t dscale_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.bnDscaleDbiasStrides_, invariant_index);
+                size_t dbias_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.bnDscaleDbiasStrides_, invariant_index);
+
+                arg.p_dscale_[dscale_offset] = type_convert<DscaleDbiasDataType>(dscale);
+                arg.p_dbias_[dbias_offset]   = type_convert<DscaleDbiasDataType>(dbias);
+
+                size_t scale_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnScaleStrides_, invariant_index);
+
+                AccDataType scale = type_convert<AccDataType>(arg.p_scale_[scale_offset]);
+
+                AccDataType multiplier = type_convert<AccDataType>(1.0f) /
+                                         type_convert<AccDataType>(arg.reduceSize_) * invVar *
+                                         scale;
+
+                // 1) calculate tmp = dscale * (x - mean) * inv-variance
+                // 2) calculate dx = 1/reduceSize * inv-variance * scale * (reduceSize * dy - dbias
+                // - tmp)
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+                    size_t dy_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.dy_reduce_strides_, reduce_index);
+                    size_t dx_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.dx_reduce_strides_, reduce_index);
+
+                    auto x_offset  = x_invariant_offset + x_reduce_offset;
+                    auto dy_offset = dy_invariant_offset + dy_reduce_offset;
+                    auto dx_offset = dx_invariant_offset + dx_reduce_offset;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType norm_x = (x - mean) * invVar;
+                    AccDataType dy     = type_convert<AccDataType>(arg.p_dy_[dy_offset]);
+
+                    arg.dy_elementwise_op_(dy, dy);
+
+                    AccDataType tmpVal = norm_x * dscale;
+
+                    AccDataType dx = multiplier * (type_convert<AccDataType>(arg.reduceSize_) * dy -
+                                                   dbias - tmpVal);
+
+                    arg.p_dx_[dx_offset] = type_convert<DxDataType>(dx);
+                };
+            };
+
+            std::size_t num_thread = std::thread::hardware_concurrency();
+            std::size_t work_per_thread =
+                (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+            std::vector<joinable_thread> threads(num_thread);
+
+            for(std::size_t it = 0; it < num_thread; ++it)
+            {
+                std::size_t i_begin = it * work_per_thread;
+                std::size_t i_end   = std::min(static_cast<size_t>((it + 1) * work_per_thread),
+                                             arg.invariant_index_set_.size());
+
+                auto f = [=] {
+                    for(std::size_t i = i_begin; i < i_end; ++i)
+                    {
+                        thread_reduce_func(arg.invariant_index_set_[i]);
+                    }
+                };
+
+                threads[it] = joinable_thread(f);
+            }
+
+            return (0.0f);
+        };
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        };
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        (void)p_arg;
+
+        return (true);
+    };
+
+    std::unique_ptr<device::BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> xyLengths,
+                        const std::array<index_t, Rank> xStrides,
+                        const std::array<index_t, Rank> dxStrides,
+                        const std::array<index_t, Rank> dyStrides,
+                        const std::array<int, NumBatchNormReduceDim> reduceDims,
+                        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                        const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                        const std::array<index_t, NumInvariantDim> bnDscaleDbiasStrides,
+                        const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                        const void* p_x,
+                        const void* p_dy,
+                        const void* p_scale,
+                        const void* p_savedMean,
+                        const void* p_savedInvVar,
+                        double epsilon,
+                        const DyElementwiseOp dy_elementwise_op,
+                        void* p_dx,
+                        void* p_dscale,
+                        void* p_dbias) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          dxStrides,
+                                          dyStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnDscaleDbiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const DyDataType*>(p_dy),
+                                          static_cast<const ScaleDataType*>(p_scale),
+                                          static_cast<const MeanVarDataType*>(p_savedMean),
+                                          static_cast<const MeanVarDataType*>(p_savedInvVar),
+                                          epsilon,
+                                          dy_elementwise_op,
+                                          static_cast<DxDataType*>(p_dx),
+                                          static_cast<DscaleDbiasDataType*>(p_dscale),
+                                          static_cast<DscaleDbiasDataType*>(p_dbias));
+    };
+
+    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "Reference_BatchNorm_Backward" << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp
new file mode 100644
index 0000000..dd0db31
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp
@@ -0,0 +1,368 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <array>
+#include <algorithm>
+#include <thread>
+
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/ignore.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct ReferenceBatchNormFwd : public device::DeviceBatchNormFwd<XDataType,
+                                                                 YDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 BiasDataType,
+                                                                 MeanVarDataType,
+                                                                 YElementwiseOp,
+                                                                 Rank,
+                                                                 NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> yStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                 const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                 const std::array<index_t, NumInvariantDim> bnBiasStrides,
+                 const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const ScaleDataType* bnScale,
+                 const BiasDataType* bnBias,
+                 double epsilon,
+                 const YElementwiseOp y_elementwise_op,
+                 YDataType* p_y,
+                 MeanVarDataType* resultSaveMean,
+                 MeanVarDataType* resultSaveInvVariance,
+                 double averageFactor,
+                 MeanVarDataType* resultRunningMean,
+                 MeanVarDataType* resultRunningVariance)
+            : reduceDims_(reduceDims),
+              bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnBiasStrides_(bnBiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              bnScale_(bnScale),
+              bnBias_(bnBias),
+              y_elementwise_op_(y_elementwise_op),
+              p_y_(p_y),
+              resultSaveMean_(resultSaveMean),
+              resultSaveInvVariance_(resultSaveInvVariance),
+              resultRunningMean_(resultRunningMean),
+              resultRunningVariance_(resultRunningVariance)
+        {
+            using ck::host_common::get_index_set;
+
+            if(std::any_of(
+                   reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+                throw std::runtime_error("Invalid reduce dimensions!");
+
+            // get invariant_dims[] and invariant_lengths[]
+            for(int dim = 0, i = 0; dim < Rank; dim++)
+                if(std::none_of(
+                       reduceDims.begin(), reduceDims.end(), [&](int d) { return d == dim; }))
+                {
+                    invariantDims_[i]     = dim;
+                    invariant_lengths_[i] = xyLengths[dim];
+                    i++;
+                };
+
+            // get reduce_lengths_[]
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims[j];
+                reduce_lengths_[i++] = xyLengths[dim];
+            };
+
+            for(int i = 0; i < NumInvariantDim; i++)
+                if(invariant_lengths_[i] != bnScaleBiasMeanVarLengths_[i])
+                    throw std::runtime_error("Invalid lengths parameters!");
+
+            for(int j = 0, i = 0; j < NumInvariantDim; j++)
+            {
+                int dim                 = invariantDims_[j];
+                x_invariant_strides_[i] = xStrides[dim];
+                y_invariant_strides_[i] = yStrides[dim];
+                i++;
+            };
+
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims_[j];
+                x_reduce_strides_[i] = xStrides[dim];
+                y_reduce_strides_[i] = yStrides[dim];
+                i++;
+            };
+
+            invariant_index_set_ = get_index_set<NumInvariantDim>(invariant_lengths_);
+            reduce_index_set_    = get_index_set<NumBatchNormReduceDim>(reduce_lengths_);
+
+            epsilon_       = type_convert<AccDataType>(epsilon);
+            averageFactor_ = type_convert<AccDataType>(averageFactor);
+
+            resultSave    = (resultSaveMean != nullptr && resultSaveInvVariance != nullptr);
+            resultRunning = (resultRunningMean != nullptr && resultRunningVariance != nullptr);
+        }
+
+        std::array<int, NumBatchNormReduceDim> reduceDims_;
+        std::array<int, NumInvariantDim> invariantDims_;
+        std::array<index_t, NumInvariantDim> invariant_lengths_;
+        std::array<index_t, NumBatchNormReduceDim> reduce_lengths_;
+
+        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths_;
+        const std::array<index_t, NumInvariantDim> bnScaleStrides_;
+        const std::array<index_t, NumInvariantDim> bnBiasStrides_;
+        const std::array<index_t, NumInvariantDim> bnMeanVarStrides_;
+
+        std::array<index_t, NumInvariantDim> x_invariant_strides_;
+        std::array<index_t, NumInvariantDim> y_invariant_strides_;
+        std::array<index_t, NumBatchNormReduceDim> x_reduce_strides_;
+        std::array<index_t, NumBatchNormReduceDim> y_reduce_strides_;
+
+        const XDataType* p_x_;
+        const ScaleDataType* bnScale_;
+        const BiasDataType* bnBias_;
+        const YElementwiseOp y_elementwise_op_;
+        YDataType* p_y_;
+
+        MeanVarDataType* resultSaveMean_;
+        MeanVarDataType* resultSaveInvVariance_;
+        MeanVarDataType* resultRunningMean_;
+        MeanVarDataType* resultRunningVariance_;
+
+        bool resultSave, resultRunning;
+
+        std::vector<std::array<index_t, NumInvariantDim>> invariant_index_set_;
+        std::vector<std::array<index_t, NumBatchNormReduceDim>> reduce_index_set_;
+
+        AccDataType averageFactor_;
+        AccDataType epsilon_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            using ck::host_common::get_offset_from_index;
+
+            auto thread_reduce_func = [&](auto invariant_index) {
+                size_t x_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.x_invariant_strides_, invariant_index);
+                size_t y_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.y_invariant_strides_, invariant_index);
+                AccDataType mean     = type_convert<AccDataType>(0.0f);
+                AccDataType variance = type_convert<AccDataType>(0.0f);
+                int32_t curr_count   = 0;
+
+                // compute mean, variance using welford method
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+
+                    auto x_offset = x_invariant_offset + x_reduce_offset;
+
+                    curr_count++;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType delta = x - mean;
+
+                    mean += delta / curr_count;
+
+                    AccDataType delta2 = x - mean;
+
+                    variance += delta * delta2;
+                };
+
+                // actual variance
+                variance = variance / curr_count;
+
+                // inv-variance defined as 1/sqrt(epsilon+variance)
+                AccDataType invVariance =
+                    type_convert<AccDataType>(1.0f) / ck::math::sqrt(arg.epsilon_ + variance);
+
+                // save the mean/inv-variance if required
+                if(arg.resultSave)
+                {
+                    size_t offset = get_offset_from_index<NumInvariantDim>(arg.bnMeanVarStrides_,
+                                                                           invariant_index);
+
+                    arg.resultSaveMean_[offset]        = type_convert<MeanVarDataType>(mean);
+                    arg.resultSaveInvVariance_[offset] = type_convert<MeanVarDataType>(invVariance);
+                };
+
+                // update the moving average if required
+                if(arg.resultRunning)
+                {
+                    size_t offset = get_offset_from_index<NumInvariantDim>(arg.bnMeanVarStrides_,
+                                                                           invariant_index);
+
+                    AccDataType oneMinusAverageFactor =
+                        type_convert<AccDataType>(1.0) - arg.averageFactor_;
+                    arg.resultRunningMean_[offset] = type_convert<MeanVarDataType>(
+                        type_convert<AccDataType>(arg.resultRunningMean_[offset]) *
+                            oneMinusAverageFactor +
+                        mean * arg.averageFactor_);
+                    arg.resultRunningVariance_[offset] = type_convert<MeanVarDataType>(
+                        arg.resultRunningVariance_[offset] * oneMinusAverageFactor +
+                        variance * arg.averageFactor_);
+                };
+
+                size_t scale_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnScaleStrides_, invariant_index);
+                size_t bias_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnBiasStrides_, invariant_index);
+
+                AccDataType scale = type_convert<AccDataType>(arg.bnScale_[scale_offset]);
+                AccDataType bias  = type_convert<AccDataType>(arg.bnBias_[bias_offset]);
+
+                // Normalization
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+                    size_t y_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.y_reduce_strides_, reduce_index);
+
+                    auto x_offset = x_invariant_offset + x_reduce_offset;
+                    auto y_offset = y_invariant_offset + y_reduce_offset;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType norm_x = (x - mean) * invVariance;
+
+                    AccDataType y = scale * norm_x + bias;
+
+                    arg.y_elementwise_op_(y, y);
+
+                    arg.p_y_[y_offset] = type_convert<YDataType>(y);
+                };
+            };
+
+            std::size_t num_thread = std::thread::hardware_concurrency();
+            std::size_t work_per_thread =
+                (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+            std::vector<joinable_thread> threads(num_thread);
+
+            for(std::size_t it = 0; it < num_thread; ++it)
+            {
+                std::size_t i_begin = it * work_per_thread;
+                std::size_t i_end   = std::min(static_cast<size_t>((it + 1) * work_per_thread),
+                                             arg.invariant_index_set_.size());
+
+                auto f = [=] {
+                    for(std::size_t i = i_begin; i < i_end; ++i)
+                    {
+                        thread_reduce_func(arg.invariant_index_set_[i]);
+                    }
+                };
+
+                threads[it] = joinable_thread(f);
+            }
+
+            return (0.0f);
+        };
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        };
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        (void)p_arg;
+
+        return (true);
+    };
+
+    std::unique_ptr<device::BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, 4> xyLengths,
+                        const std::array<index_t, 4> xStrides,
+                        const std::array<index_t, 4> yStrides,
+                        const std::array<int, 3> reduceDims,
+                        const std::array<index_t, 1> bnScaleBiasMeanVarLengths,
+                        const std::array<index_t, 1> bnScaleStrides,
+                        const std::array<index_t, 1> bnBiasStrides,
+                        const std::array<index_t, 1> bnMeanVarStrides,
+                        const void* p_x,
+                        const void* bnScale,
+                        const void* bnBias,
+                        double epsilon,
+                        const YElementwiseOp y_elementwise_op,
+                        void* p_y,
+                        void* resultSaveMean,
+                        void* resultSaveInvVariance,
+                        double averageFactor,
+                        void* resultRunningMean,
+                        void* resultRunningVariance) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnBiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const ScaleDataType*>(bnScale),
+                                          static_cast<const BiasDataType*>(bnBias),
+                                          epsilon,
+                                          y_elementwise_op,
+                                          static_cast<YDataType*>(p_y),
+                                          static_cast<MeanVarDataType*>(resultSaveMean),
+                                          static_cast<MeanVarDataType*>(resultSaveInvVariance),
+                                          averageFactor,
+                                          static_cast<MeanVarDataType*>(resultRunningMean),
+                                          static_cast<MeanVarDataType*>(resultRunningVariance));
+    };
+
+    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "Reference_BatchNorm_Forward" << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp
new file mode 100644
index 0000000..463c655
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp
@@ -0,0 +1,300 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+#include <array>
+#include <algorithm>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct ReferenceBatchNormInfer : public device::DeviceBatchNormInfer<XDataType,
+                                                                     YDataType,
+                                                                     AccDataType,
+                                                                     ScaleDataType,
+                                                                     BiasDataType,
+                                                                     MeanVarDataType,
+                                                                     YElementwiseOp,
+                                                                     Rank,
+                                                                     NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> yStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                 const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                 const std::array<index_t, NumInvariantDim> bnBiasStrides,
+                 const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const ScaleDataType* bnScale,
+                 const BiasDataType* bnBias,
+                 double epsilon,
+                 const YElementwiseOp y_elementwise_op,
+                 const MeanVarDataType* estimatedMean,
+                 const MeanVarDataType* estimatedVariance,
+                 YDataType* p_y)
+            : reduceDims_(reduceDims),
+              bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnBiasStrides_(bnBiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              bnScale_(bnScale),
+              bnBias_(bnBias),
+              y_elementwise_op_(y_elementwise_op),
+              estimatedMean_(estimatedMean),
+              estimatedVariance_(estimatedVariance),
+              p_y_(p_y)
+        {
+            using ck::host_common::get_index_set;
+
+            if(std::any_of(
+                   reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+                throw std::runtime_error("Invalid reduce dimensions!");
+
+            // get invariant_dims[] and invariant_lengths[]
+            for(int dim = 0, i = 0; dim < Rank; dim++)
+                if(std::none_of(
+                       reduceDims.begin(), reduceDims.end(), [&](int d) { return d == dim; }))
+                {
+                    invariantDims_[i]     = dim;
+                    invariant_lengths_[i] = xyLengths[dim];
+                    i++;
+                };
+
+            // get reduce_lengths_[]
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims[j];
+                reduce_lengths_[i++] = xyLengths[dim];
+            };
+
+            // check invariant_lengths_ and bnScaleBiasMeanVarLengths
+            for(int i = 0; i < NumInvariantDim; i++)
+                if(invariant_lengths_[i] != bnScaleBiasMeanVarLengths_[i])
+                    throw std::runtime_error("Invalid lengths parameters!");
+
+            for(int j = 0, i = 0; j < NumInvariantDim; j++)
+            {
+                int dim                 = invariantDims_[j];
+                x_invariant_strides_[i] = xStrides[dim];
+                y_invariant_strides_[i] = yStrides[dim];
+                i++;
+            };
+
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims_[j];
+                x_reduce_strides_[i] = xStrides[dim];
+                y_reduce_strides_[i] = yStrides[dim];
+                i++;
+            };
+
+            invariant_index_set_ = get_index_set<NumInvariantDim>(invariant_lengths_);
+            reduce_index_set_    = get_index_set<NumBatchNormReduceDim>(reduce_lengths_);
+
+            epsilon_ = type_convert<AccDataType>(epsilon);
+        }
+
+        std::array<int, NumBatchNormReduceDim> reduceDims_;
+        std::array<int, NumInvariantDim> invariantDims_;
+        std::array<index_t, NumInvariantDim> invariant_lengths_;
+        std::array<index_t, NumBatchNormReduceDim> reduce_lengths_;
+
+        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths_;
+        const std::array<index_t, NumInvariantDim> bnScaleStrides_;
+        const std::array<index_t, NumInvariantDim> bnBiasStrides_;
+        const std::array<index_t, NumInvariantDim> bnMeanVarStrides_;
+
+        std::array<index_t, NumInvariantDim> x_invariant_strides_;
+        std::array<index_t, NumInvariantDim> y_invariant_strides_;
+        std::array<index_t, NumBatchNormReduceDim> x_reduce_strides_;
+        std::array<index_t, NumBatchNormReduceDim> y_reduce_strides_;
+
+        const XDataType* p_x_;
+        const ScaleDataType* bnScale_;
+        const BiasDataType* bnBias_;
+        const YElementwiseOp y_elementwise_op_;
+
+        const MeanVarDataType* estimatedMean_;
+        const MeanVarDataType* estimatedVariance_;
+
+        YDataType* p_y_;
+
+        std::vector<std::array<index_t, NumInvariantDim>> invariant_index_set_;
+        std::vector<std::array<index_t, NumBatchNormReduceDim>> reduce_index_set_;
+
+        AccDataType epsilon_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            using ck::host_common::get_offset_from_index;
+
+            auto thread_reduce_func = [&](auto invariant_index) {
+                size_t x_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.x_invariant_strides_, invariant_index);
+                size_t y_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.y_invariant_strides_, invariant_index);
+
+                size_t mean_variance_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnMeanVarStrides_, invariant_index);
+
+                AccDataType mean     = arg.estimatedMean_[mean_variance_offset];
+                AccDataType variance = arg.estimatedVariance_[mean_variance_offset];
+
+                // inv-variance defined as 1/sqrt(epsilon+variance)
+                AccDataType invVariance =
+                    type_convert<AccDataType>(1.0f) / std::sqrt(arg.epsilon_ + variance);
+
+                size_t scale_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnScaleStrides_, invariant_index);
+                size_t bias_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnBiasStrides_, invariant_index);
+
+                AccDataType scale = type_convert<AccDataType>(arg.bnScale_[scale_offset]);
+                AccDataType bias  = type_convert<AccDataType>(arg.bnBias_[bias_offset]);
+
+                // normalization
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+                    size_t y_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.y_reduce_strides_, reduce_index);
+
+                    auto x_offset = x_invariant_offset + x_reduce_offset;
+                    auto y_offset = y_invariant_offset + y_reduce_offset;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType norm_x = (x - mean) * invVariance;
+
+                    AccDataType y = scale * norm_x + bias;
+
+                    arg.y_elementwise_op_(y, y);
+
+                    arg.p_y_[y_offset] = type_convert<YDataType>(y);
+                };
+            };
+
+            std::size_t num_thread = std::thread::hardware_concurrency();
+            std::size_t work_per_thread =
+                (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+            std::vector<joinable_thread> threads(num_thread);
+
+            for(std::size_t it = 0; it < num_thread; ++it)
+            {
+                std::size_t i_begin = it * work_per_thread;
+                std::size_t i_end   = std::min(static_cast<size_t>((it + 1) * work_per_thread),
+                                             arg.invariant_index_set_.size());
+
+                auto f = [=] {
+                    for(std::size_t i = i_begin; i < i_end; ++i)
+                    {
+                        thread_reduce_func(arg.invariant_index_set_[i]);
+                    }
+                };
+
+                threads[it] = joinable_thread(f);
+            }
+
+            return (0.0f);
+        };
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        };
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        (void)p_arg;
+
+        return (true);
+    };
+
+    std::unique_ptr<device::BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> xyLengths,
+                        const std::array<index_t, Rank> xStrides,
+                        const std::array<index_t, Rank> yStrides,
+                        const std::array<int, NumBatchNormReduceDim> reduceDims,
+                        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                        const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                        const std::array<index_t, NumInvariantDim> bnBiasStrides,
+                        const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                        const void* p_x,
+                        const void* bnScale,
+                        const void* bnBias,
+                        double epsilon,
+                        const YElementwiseOp y_elementwise_op,
+                        const void* estimatedMean,
+                        const void* estimatedVariance,
+                        void* p_y) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnBiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const ScaleDataType*>(bnScale),
+                                          static_cast<const BiasDataType*>(bnBias),
+                                          epsilon,
+                                          y_elementwise_op,
+                                          static_cast<const MeanVarDataType*>(estimatedMean),
+                                          static_cast<const MeanVarDataType*>(estimatedVariance),
+                                          static_cast<YDataType*>(p_y));
+    };
+
+    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "Reference_BatchNorm_Infer<" << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp
new file mode 100644
index 0000000..b0149d8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp
@@ -0,0 +1,184 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// FIXME: support arbitrary elementwise operation for A/B/C
+template <
+    typename ADataType,
+    typename BDataType,
+    typename CDataType,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    enable_if_t<
+        is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+        bool> = false>
+struct ReferenceCGemm : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_m_k_real,
+                 const Tensor<ADataType>& a_m_k_imag,
+                 const Tensor<BDataType>& b_k_n_real,
+                 const Tensor<BDataType>& b_k_n_imag,
+                 Tensor<CDataType>& c_m_n_real,
+                 Tensor<CDataType>& c_m_n_imag,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_m_k_real_{a_m_k_real},
+              a_m_k_imag_{a_m_k_imag},
+              b_k_n_real_{b_k_n_real},
+              b_k_n_imag_{b_k_n_imag},
+              c_m_n_real_{c_m_n_real},
+              c_m_n_imag_{c_m_n_imag},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_m_k_real_;
+        const Tensor<ADataType>& a_m_k_imag_;
+        const Tensor<BDataType>& b_k_n_real_;
+        const Tensor<BDataType>& b_k_n_imag_;
+        Tensor<CDataType>& c_m_n_real_;
+        Tensor<CDataType>& c_m_n_imag_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceCGemm::Argument;
+
+        float Run(const Argument& arg)
+        {
+            const std::size_t K = arg.a_m_k_real_.mDesc.GetLengths()[1];
+
+            if(K != arg.a_m_k_imag_.mDesc.GetLengths()[1])
+            {
+                throw std::runtime_error("wrong! Incompatible real and imag sizes in CGEMM");
+            }
+
+            auto f_mk_kn_mn_real = [&](auto m, auto n) {
+                float v_c_real = 0;
+
+                for(std::size_t k = 0; k < K; ++k)
+                {
+                    float v_a_real = ck::type_convert<float>(arg.a_m_k_real_(m, k));
+                    float v_a_imag = ck::type_convert<float>(arg.a_m_k_imag_(m, k));
+                    float v_b_real = ck::type_convert<float>(arg.b_k_n_real_(k, n));
+                    float v_b_imag = ck::type_convert<float>(arg.b_k_n_imag_(k, n));
+
+                    v_c_real += v_a_real * v_b_real - v_a_imag * v_b_imag;
+                }
+
+                arg.c_m_n_real_(m, n) = ck::type_convert<CDataType>(v_c_real);
+            };
+
+            auto f_mk_kn_mn_imag = [&](auto m, auto n) {
+                float v_c_imag = 0;
+
+                for(std::size_t k = 0; k < K; ++k)
+                {
+                    float v_a_real = ck::type_convert<float>(arg.a_m_k_real_(m, k));
+                    float v_a_imag = ck::type_convert<float>(arg.a_m_k_imag_(m, k));
+                    float v_b_real = ck::type_convert<float>(arg.b_k_n_real_(k, n));
+                    float v_b_imag = ck::type_convert<float>(arg.b_k_n_imag_(k, n));
+
+                    v_c_imag += v_a_real * v_b_imag + v_a_imag * v_b_real;
+                }
+
+                arg.c_m_n_imag_(m, n) = ck::type_convert<CDataType>(v_c_imag);
+            };
+
+            make_ParallelTensorFunctor(f_mk_kn_mn_real,
+                                       arg.c_m_n_real_.mDesc.GetLengths()[0],
+                                       arg.c_m_n_real_.mDesc.GetLengths()[1])(
+                std::thread::hardware_concurrency());
+            make_ParallelTensorFunctor(f_mk_kn_mn_imag,
+                                       arg.c_m_n_imag_.mDesc.GetLengths()[0],
+                                       arg.c_m_n_imag_.mDesc.GetLengths()[1])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_m_k_real,
+                             const Tensor<ADataType>& a_m_k_imag,
+                             const Tensor<BDataType>& b_k_n_real,
+                             const Tensor<BDataType>& b_k_n_imag,
+                             Tensor<CDataType>& c_m_n_real,
+                             Tensor<CDataType>& c_m_n_imag,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{a_m_k_real,
+                        a_m_k_imag,
+                        b_k_n_real,
+                        b_k_n_imag,
+                        c_m_n_real,
+                        c_m_n_imag,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceCGemm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
new file mode 100644
index 0000000..225f7b7
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
@@ -0,0 +1,378 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// input descriptor in [G, N, C, Do, Ho, Wo] order
+// weight descriptor in [G, K, C, Z, Y, X] order
+// output descriptor in [G, N, K, Di, Hi, Wi] order
+// phyiscal layout is irrelavent
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
+struct ReferenceConvBwdData : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(Tensor<InDataType>& input,
+                 const Tensor<WeiDataType>& weight,
+                 const Tensor<OutDataType>& output,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : input_{input},
+              weight_{weight},
+              output_{output},
+              conv_strides_{conv_filter_strides},
+              conv_dilations_{conv_filter_dilations},
+              in_left_pads_{input_left_pads},
+              in_right_pads_{input_right_pads},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+        }
+
+        Tensor<InDataType>& input_;
+        const Tensor<WeiDataType>& weight_;
+        const Tensor<OutDataType>& output_;
+
+        std::vector<index_t> conv_strides_;
+        std::vector<index_t> conv_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceConvBwdData::Argument;
+
+        float Run(const Argument& arg)
+        {
+            if(!(arg.input_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.weight_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.output_.GetNumOfDimension() == NDimSpatial + 3))
+            {
+                throw std::runtime_error("wrong! inconsistent dimension");
+            }
+
+            if constexpr(NDimSpatial == 1)
+            {
+                auto f_ncw = [&](auto g, auto n, auto c, auto wi) {
+                    std::size_t K  = arg.weight_.GetLengths()[1];
+                    std::size_t X  = arg.weight_.GetLengths()[3];
+                    std::size_t Wo = arg.output_.GetLengths()[3];
+
+                    float v_acc = 0;
+
+                    for(std::size_t x = 0; x < X; ++x)
+                    {
+                        auto w_tmp = static_cast<ck::long_index_t>(wi) +
+                                     static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
+                                     static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]);
+
+                        if(w_tmp % arg.conv_strides_[0] == 0)
+                        {
+                            auto wo = static_cast<ck::long_index_t>(w_tmp) /
+                                      static_cast<ck::long_index_t>(arg.conv_strides_[0]);
+
+                            if(wo >= 0 && ck::type_convert<std::size_t>(wo) < Wo)
+                            {
+                                for(std::size_t k = 0; k < K; ++k)
+                                {
+                                    float v_out = 0;
+                                    float v_wei = 0;
+
+                                    arg.out_element_op_(
+                                        v_out, ck::type_convert<float>(arg.output_(g, n, k, wo)));
+
+                                    arg.wei_element_op_(
+                                        v_wei, ck::type_convert<float>(arg.weight_(g, k, c, x)));
+
+                                    v_acc += v_out * v_wei;
+                                }
+                            }
+                        }
+                    }
+
+                    float v_in;
+
+                    arg.in_element_op_(v_in, v_acc);
+
+                    arg.input_(g, n, c, wi) = ck::type_convert<InDataType>(v_acc);
+                };
+
+                make_ParallelTensorFunctor(f_ncw,
+                                           arg.input_.GetLengths()[0],
+                                           arg.input_.GetLengths()[1],
+                                           arg.input_.GetLengths()[2],
+                                           arg.input_.GetLengths()[3])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 2)
+            {
+                auto f_nchw = [&](auto g, auto n, auto c, auto hi, auto wi) {
+                    std::size_t K = arg.weight_.GetLengths()[1];
+                    std::size_t Y = arg.weight_.GetLengths()[3];
+                    std::size_t X = arg.weight_.GetLengths()[4];
+
+                    std::size_t Ho = arg.output_.GetLengths()[3];
+                    std::size_t Wo = arg.output_.GetLengths()[4];
+
+                    float v_acc = 0;
+
+                    for(std::size_t y = 0; y < Y; ++y)
+                    {
+                        auto h_tmp = static_cast<ck::long_index_t>(hi) +
+                                     static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
+                                     static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]);
+                        if(h_tmp % arg.conv_strides_[0] == 0)
+                        {
+                            auto ho = static_cast<ck::long_index_t>(h_tmp) /
+                                      static_cast<ck::long_index_t>(arg.conv_strides_[0]);
+                            if(ho >= 0 && ck::type_convert<std::size_t>(ho) < Ho)
+                            {
+                                for(std::size_t x = 0; x < X; ++x)
+                                {
+                                    auto w_tmp =
+                                        static_cast<ck::long_index_t>(wi) +
+                                        static_cast<ck::long_index_t>(arg.in_left_pads_[1]) -
+                                        static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]);
+                                    if(w_tmp % arg.conv_strides_[1] == 0)
+                                    {
+                                        auto wo =
+                                            static_cast<ck::long_index_t>(w_tmp) /
+                                            static_cast<ck::long_index_t>(arg.conv_strides_[1]);
+                                        if(wo >= 0 && ck::type_convert<std::size_t>(wo) < Wo)
+                                        {
+                                            for(std::size_t k = 0; k < K; ++k)
+                                            {
+                                                float v_out = 0;
+                                                float v_wei = 0;
+
+                                                arg.out_element_op_(
+                                                    v_out,
+                                                    ck::type_convert<float>(
+                                                        arg.output_(g, n, k, ho, wo)));
+
+                                                arg.wei_element_op_(
+                                                    v_wei,
+                                                    ck::type_convert<float>(
+                                                        arg.weight_(g, k, c, y, x)));
+
+                                                v_acc += v_out * v_wei;
+                                            }
+                                        }
+                                    }
+                                }
+                            }
+                        }
+                    }
+
+                    float v_in;
+
+                    arg.in_element_op_(v_in, v_acc);
+
+                    arg.input_(g, n, c, hi, wi) = ck::type_convert<InDataType>(v_acc);
+                };
+
+                make_ParallelTensorFunctor(f_nchw,
+                                           arg.input_.GetLengths()[0],
+                                           arg.input_.GetLengths()[1],
+                                           arg.input_.GetLengths()[2],
+                                           arg.input_.GetLengths()[3],
+                                           arg.input_.GetLengths()[4])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                auto f_ncdhw = [&](auto g, auto n, auto c, auto di, auto hi, auto wi) {
+                    std::size_t K = arg.weight_.GetLengths()[1];
+                    std::size_t Z = arg.weight_.GetLengths()[3];
+                    std::size_t Y = arg.weight_.GetLengths()[4];
+                    std::size_t X = arg.weight_.GetLengths()[5];
+
+                    std::size_t Do = arg.output_.GetLengths()[3];
+                    std::size_t Ho = arg.output_.GetLengths()[4];
+                    std::size_t Wo = arg.output_.GetLengths()[5];
+
+                    float v_acc = 0;
+
+                    for(std::size_t z = 0; z < Z; ++z)
+                    {
+                        auto d_tmp = static_cast<ck::long_index_t>(di) +
+                                     static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
+                                     static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]);
+                        if(d_tmp % arg.conv_strides_[0] == 0)
+                        {
+                            auto do_ = static_cast<ck::long_index_t>(d_tmp) /
+                                       static_cast<ck::long_index_t>(arg.conv_strides_[0]);
+                            if(do_ >= 0 && ck::type_convert<std::size_t>(do_) < Do)
+                            {
+                                for(std::size_t y = 0; y < Y; ++y)
+                                {
+                                    auto h_tmp =
+                                        static_cast<ck::long_index_t>(hi) +
+                                        static_cast<ck::long_index_t>(arg.in_left_pads_[1]) -
+                                        static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]);
+                                    if(h_tmp % arg.conv_strides_[1] == 0)
+                                    {
+                                        auto ho =
+                                            static_cast<ck::long_index_t>(h_tmp) /
+                                            static_cast<ck::long_index_t>(arg.conv_strides_[1]);
+                                        if(ho >= 0 && ck::type_convert<std::size_t>(ho) < Ho)
+                                        {
+                                            for(std::size_t x = 0; x < X; ++x)
+                                            {
+                                                auto w_tmp = static_cast<ck::long_index_t>(wi) +
+                                                             static_cast<ck::long_index_t>(
+                                                                 arg.in_left_pads_[2]) -
+                                                             static_cast<ck::long_index_t>(
+                                                                 x * arg.conv_dilations_[2]);
+
+                                                if(w_tmp % arg.conv_strides_[2] == 0)
+                                                {
+                                                    auto wo = static_cast<ck::long_index_t>(w_tmp) /
+                                                              static_cast<ck::long_index_t>(
+                                                                  arg.conv_strides_[2]);
+                                                    if(wo >= 0 &&
+                                                       ck::type_convert<std::size_t>(wo) < Wo)
+                                                    {
+                                                        for(std::size_t k = 0; k < K; ++k)
+                                                        {
+                                                            float v_out = 0;
+                                                            float v_wei = 0;
+
+                                                            arg.out_element_op_(
+                                                                v_out,
+                                                                ck::type_convert<float>(arg.output_(
+                                                                    g, n, k, do_, ho, wo)));
+
+                                                            arg.wei_element_op_(
+                                                                v_wei,
+                                                                ck::type_convert<float>(
+                                                                    arg.weight_(g, k, c, z, y, x)));
+
+                                                            v_acc += v_out * v_wei;
+                                                        }
+                                                    }
+                                                }
+                                            }
+                                        }
+                                    }
+                                }
+                            }
+                        }
+                    }
+
+                    float v_in;
+
+                    arg.in_element_op_(v_in, v_acc);
+
+                    arg.input_(g, n, c, di, hi, wi) = ck::type_convert<InDataType>(v_acc);
+                };
+
+                make_ParallelTensorFunctor(f_ncdhw,
+                                           arg.input_.GetLengths()[0],
+                                           arg.input_.GetLengths()[1],
+                                           arg.input_.GetLengths()[2],
+                                           arg.input_.GetLengths()[3],
+                                           arg.input_.GetLengths()[4],
+                                           arg.input_.GetLengths()[5])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(Tensor<InDataType>& input,
+                             const Tensor<WeiDataType>& weight,
+                             const Tensor<OutDataType>& output,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{input,
+                        weight,
+                        output,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceConvBwdData"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp
new file mode 100644
index 0000000..7d62158
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp
@@ -0,0 +1,327 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// input descriptor in [G, N, C, Do, Ho, Wo] order
+// weight descriptor in [G, K, C, Z, Y, X] order
+// output descriptor in [G, N, K, Di, Hi, Wi] order
+// phyiscal layout is irrelavent
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
+struct ReferenceConvBwdWeight : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<InDataType>& in_n_c_hi_wi,
+                 Tensor<WeiDataType>& wei_k_c_y_x,
+                 const Tensor<OutDataType>& out_n_k_ho_wo,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : input_{in_n_c_hi_wi},
+              weight_{wei_k_c_y_x},
+              output_{out_n_k_ho_wo},
+              conv_strides_{conv_filter_strides},
+              conv_dilations_{conv_filter_dilations},
+              in_left_pads_{input_left_pads},
+              in_right_pads_{input_right_pads},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+        }
+
+        const Tensor<InDataType>& input_;
+        Tensor<WeiDataType>& weight_;
+        const Tensor<OutDataType>& output_;
+
+        std::vector<index_t> conv_strides_;
+        std::vector<index_t> conv_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceConvBwdWeight::Argument;
+
+        float Run(const Argument& arg)
+        {
+            if(!(arg.input_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.weight_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.output_.GetNumOfDimension() == NDimSpatial + 3))
+            {
+                throw std::runtime_error("wrong! inconsistent dimension");
+            }
+
+            if constexpr(NDimSpatial == 1)
+            {
+                auto f_kcx = [&](auto g, auto k, auto c, auto x) {
+                    float v_acc = 0;
+
+                    for(std::size_t n = 0; n < arg.output_.GetLengths()[1]; ++n)
+                    {
+                        for(std::size_t wo = 0; wo < arg.output_.GetLengths()[3]; ++wo)
+                        {
+                            auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[0]) +
+                                      static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+
+                            if(wi >= 0 &&
+                               ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3])
+                            {
+                                float v_out;
+                                float v_in;
+
+                                arg.out_element_op_(
+                                    v_out, ck::type_convert<float>(arg.output_(g, n, k, wo)));
+
+                                arg.in_element_op_(
+                                    v_in, ck::type_convert<float>(arg.input_(g, n, c, wi)));
+
+                                v_acc += v_out * v_in;
+                            }
+                        }
+                    }
+
+                    float v_wei;
+
+                    arg.wei_element_op_(v_wei, v_acc);
+
+                    arg.weight_(g, k, c, x) = ck::type_convert<WeiDataType>(v_wei);
+                };
+
+                make_ParallelTensorFunctor(f_kcx,
+                                           arg.weight_.GetLengths()[0],
+                                           arg.weight_.GetLengths()[1],
+                                           arg.weight_.GetLengths()[2],
+                                           arg.weight_.GetLengths()[3])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 2)
+            {
+                auto f_kcyx = [&](auto g, auto k, auto c, auto y, auto x) {
+                    std::size_t N = arg.output_.GetLengths()[1];
+
+                    std::size_t Ho = arg.output_.GetLengths()[3];
+                    std::size_t Wo = arg.output_.GetLengths()[4];
+
+                    float v_acc = 0;
+
+                    for(std::size_t n = 0; n < N; ++n)
+                    {
+                        for(std::size_t ho = 0; ho < Ho; ++ho)
+                        {
+                            auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) +
+                                      static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+
+                            for(std::size_t wo = 0; wo < Wo; ++wo)
+                            {
+                                auto wi =
+                                    static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) +
+                                    static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]) -
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
+
+                                if(hi >= 0 &&
+                                   ck::type_convert<std::size_t>(hi) < arg.input_.GetLengths()[3] &&
+                                   wi >= 0 &&
+                                   ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4])
+                                {
+                                    float v_out;
+                                    float v_in;
+
+                                    arg.out_element_op_(
+                                        v_out,
+                                        ck::type_convert<float>(arg.output_(g, n, k, ho, wo)));
+
+                                    arg.in_element_op_(
+                                        v_in, ck::type_convert<float>(arg.input_(g, n, c, hi, wi)));
+
+                                    v_acc += v_out * v_in;
+                                }
+                            }
+                        }
+                    }
+
+                    float v_wei;
+
+                    arg.wei_element_op_(v_wei, v_acc);
+
+                    arg.weight_(g, k, c, y, x) = ck::type_convert<WeiDataType>(v_wei);
+                };
+
+                make_ParallelTensorFunctor(f_kcyx,
+                                           arg.weight_.GetLengths()[0],
+                                           arg.weight_.GetLengths()[1],
+                                           arg.weight_.GetLengths()[2],
+                                           arg.weight_.GetLengths()[3],
+                                           arg.weight_.GetLengths()[4])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                auto f_kczyx = [&](auto g, auto k, auto c, auto z, auto y, auto x) {
+                    float v_acc = 0;
+
+                    for(std::size_t n = 0; n < arg.output_.GetLengths()[1]; ++n)
+                    {
+                        for(std::size_t do_ = 0; do_ < arg.output_.GetLengths()[3]; ++do_)
+                        {
+                            auto di = static_cast<ck::long_index_t>(do_ * arg.conv_strides_[0]) +
+                                      static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+                            for(std::size_t ho = 0; ho < arg.output_.GetLengths()[4]; ++ho)
+                            {
+                                auto hi =
+                                    static_cast<ck::long_index_t>(ho * arg.conv_strides_[1]) +
+                                    static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]) -
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
+                                for(std::size_t wo = 0; wo < arg.output_.GetLengths()[5]; ++wo)
+                                {
+                                    auto wi =
+                                        static_cast<ck::long_index_t>(wo * arg.conv_strides_[2]) +
+                                        static_cast<ck::long_index_t>(x * arg.conv_dilations_[2]) -
+                                        static_cast<ck::long_index_t>(arg.in_left_pads_[2]);
+
+                                    if(di >= 0 &&
+                                       ck::type_convert<std::size_t>(di) <
+                                           arg.input_.GetLengths()[3] &&
+                                       hi >= 0 &&
+                                       ck::type_convert<std::size_t>(hi) <
+                                           arg.input_.GetLengths()[4] &&
+                                       wi >= 0 &&
+                                       ck::type_convert<std::size_t>(wi) <
+                                           arg.input_.GetLengths()[5])
+                                    {
+                                        float v_out;
+                                        float v_in;
+
+                                        arg.out_element_op_(v_out,
+                                                            ck::type_convert<float>(
+                                                                arg.output_(g, n, k, do_, ho, wo)));
+
+                                        arg.in_element_op_(v_in,
+                                                           ck::type_convert<float>(
+                                                               arg.input_(g, n, c, di, hi, wi)));
+
+                                        v_acc += v_out * v_in;
+                                    }
+                                }
+                            }
+                        }
+                    }
+
+                    float v_wei;
+
+                    arg.wei_element_op_(v_wei, v_acc);
+
+                    arg.weight_(g, k, c, z, y, x) = ck::type_convert<WeiDataType>(v_wei);
+                };
+
+                make_ParallelTensorFunctor(f_kczyx,
+                                           arg.weight_.GetLengths()[0],
+                                           arg.weight_.GetLengths()[1],
+                                           arg.weight_.GetLengths()[2],
+                                           arg.weight_.GetLengths()[3],
+                                           arg.weight_.GetLengths()[4],
+                                           arg.weight_.GetLengths()[5])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<InDataType>& in_n_c_hi_wi,
+                             Tensor<WeiDataType>& wei_k_c_y_x,
+                             const Tensor<OutDataType>& out_n_k_ho_wo,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{in_n_c_hi_wi,
+                        wei_k_c_y_x,
+                        out_n_k_ho_wo,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceConvBwdWeight"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
new file mode 100644
index 0000000..b8d47d2
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
@@ -0,0 +1,339 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <type_traits>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+//
+// @brief      Reference implementation for forward convolution.
+//
+// @paragraph
+//             Tensor descriptor in GNCHW/GKCXY/GNKHW dimensional order
+//             Supports both GNCHW/NGCHW as well as GNHWC/NHWGC physical layout
+//             as long as dimensions in tensor descriptor is in GNCHW order
+//
+// @tparam     InDataType               Input tensor data type.
+// @tparam     WeiDataType              Weights tensor data type.
+// @tparam     OutDataType              Output tensor data type.
+// @tparam     InElementwiseOperation   Functor for input tensor elementwise
+//                                      operation.
+// @tparam     WeiElementwiseOperation  Functor for weights tensor elementwise
+//                                      operation.
+// @tparam     NDimSpatial  Number of spatial dimensions.
+//
+// input descriptor in [G, N, C, Do, Ho, Wo] order
+// weight descriptor in [G, K, C, Z, Y, X] order
+// output descriptor in [G, N, K, Di, Hi, Wi] order
+// phyiscal layout is irrelavent
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
+struct ReferenceConvFwd : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<InDataType>& input,
+                 const Tensor<WeiDataType>& weight,
+                 Tensor<OutDataType>& output,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : input_{input},
+              weight_{weight},
+              output_{output},
+              conv_strides_{conv_filter_strides},
+              conv_dilations_{conv_filter_dilations},
+              in_left_pads_{input_left_pads},
+              in_right_pads_{input_right_pads},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+        }
+
+        const Tensor<InDataType>& input_;
+        const Tensor<WeiDataType>& weight_;
+        Tensor<OutDataType>& output_;
+
+        std::vector<index_t> conv_strides_;
+        std::vector<index_t> conv_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceConvFwd::Argument;
+
+        float Run(const Argument& arg)
+        {
+            if(!(arg.input_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.weight_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.output_.GetNumOfDimension() == NDimSpatial + 3))
+            {
+                throw std::runtime_error("wrong! inconsistent dimension");
+            }
+
+            if constexpr(NDimSpatial == 1)
+            {
+                auto func = [&](auto g, auto n, auto k, auto wo) {
+                    float v_acc = 0;
+
+                    for(std::size_t c = 0; c < arg.weight_.GetLengths()[2]; ++c)
+                    {
+                        for(std::size_t x = 0; x < arg.weight_.GetLengths()[3]; ++x)
+                        {
+                            auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[0]) +
+                                      static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+
+                            if(wi >= 0 &&
+                               ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3])
+                            {
+                                float v_in;
+                                float v_wei;
+
+                                arg.in_element_op_(
+                                    v_in, ck::type_convert<float>(arg.input_(g, n, c, wi)));
+
+                                arg.wei_element_op_(
+                                    v_wei, ck::type_convert<float>(arg.weight_(g, k, c, x)));
+
+                                v_acc += v_in * v_wei;
+                            }
+                        }
+                    }
+
+                    float v_out;
+
+                    arg.out_element_op_(v_out, v_acc);
+
+                    arg.output_(g, n, k, wo) = ck::type_convert<OutDataType>(v_out);
+                };
+
+                make_ParallelTensorFunctor(func,
+                                           arg.output_.GetLengths()[0],
+                                           arg.output_.GetLengths()[1],
+                                           arg.output_.GetLengths()[2],
+                                           arg.output_.GetLengths()[3])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 2)
+            {
+                auto func = [&](auto g, auto n, auto k, auto ho, auto wo) {
+                    float v_acc = 0;
+
+                    for(std::size_t c = 0; c < arg.weight_.GetLengths()[2]; ++c)
+                    {
+                        for(std::size_t y = 0; y < arg.weight_.GetLengths()[3]; ++y)
+                        {
+                            auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) +
+                                      static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+
+                            for(std::size_t x = 0; x < arg.weight_.GetLengths()[4]; ++x)
+                            {
+                                auto wi =
+                                    static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) +
+                                    static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]) -
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
+
+                                if(hi >= 0 &&
+                                   ck::type_convert<std::size_t>(hi) < arg.input_.GetLengths()[3] &&
+                                   wi >= 0 &&
+                                   ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4])
+                                {
+                                    float v_in;
+                                    float v_wei;
+
+                                    arg.in_element_op_(
+                                        v_in, ck::type_convert<float>(arg.input_(g, n, c, hi, wi)));
+
+                                    arg.wei_element_op_(
+                                        v_wei, ck::type_convert<float>(arg.weight_(g, k, c, y, x)));
+
+                                    v_acc += v_in * v_wei;
+                                }
+                            }
+                        }
+                    }
+
+                    float v_out;
+
+                    arg.out_element_op_(v_out, v_acc);
+
+                    arg.output_(g, n, k, ho, wo) = ck::type_convert<OutDataType>(v_out);
+                };
+
+                make_ParallelTensorFunctor(func,
+                                           arg.output_.GetLengths()[0],
+                                           arg.output_.GetLengths()[1],
+                                           arg.output_.GetLengths()[2],
+                                           arg.output_.GetLengths()[3],
+                                           arg.output_.GetLengths()[4])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                auto func = [&](auto g, auto n, auto k, auto d_o, auto ho, auto wo) {
+                    float v_acc = 0;
+
+                    for(std::size_t c = 0; c < arg.weight_.GetLengths()[2]; ++c)
+                    {
+                        for(std::size_t z = 0; z < arg.weight_.GetLengths()[3]; ++z)
+                        {
+                            auto di = static_cast<ck::long_index_t>(d_o * arg.conv_strides_[0]) +
+                                      static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+                            for(std::size_t y = 0; y < arg.weight_.GetLengths()[4]; ++y)
+                            {
+                                auto hi =
+                                    static_cast<ck::long_index_t>(ho * arg.conv_strides_[1]) +
+                                    static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]) -
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
+                                for(std::size_t x = 0; x < arg.weight_.GetLengths()[5]; ++x)
+                                {
+                                    auto wi =
+                                        static_cast<ck::long_index_t>(wo * arg.conv_strides_[2]) +
+                                        static_cast<ck::long_index_t>(x * arg.conv_dilations_[2]) -
+                                        static_cast<ck::long_index_t>(arg.in_left_pads_[2]);
+                                    if(di >= 0 &&
+                                       ck::type_convert<std::size_t>(di) <
+                                           arg.input_.GetLengths()[3] &&
+                                       hi >= 0 &&
+                                       ck::type_convert<std::size_t>(hi) <
+                                           arg.input_.GetLengths()[4] &&
+                                       wi >= 0 &&
+                                       ck::type_convert<std::size_t>(wi) <
+                                           arg.input_.GetLengths()[5])
+                                    {
+                                        float v_in;
+                                        float v_wei;
+
+                                        arg.in_element_op_(v_in,
+                                                           ck::type_convert<float>(
+                                                               arg.input_(g, n, c, di, hi, wi)));
+
+                                        arg.wei_element_op_(
+                                            v_wei,
+                                            ck::type_convert<float>(arg.weight_(g, k, c, z, y, x)));
+
+                                        v_acc += v_in * v_wei;
+                                    }
+                                }
+                            }
+                        }
+                    }
+
+                    float v_out;
+
+                    arg.out_element_op_(v_out, v_acc);
+
+                    arg.output_(g, n, k, d_o, ho, wo) = ck::type_convert<OutDataType>(v_out);
+                };
+
+                make_ParallelTensorFunctor(func,
+                                           arg.output_.GetLengths()[0],
+                                           arg.output_.GetLengths()[1],
+                                           arg.output_.GetLengths()[2],
+                                           arg.output_.GetLengths()[3],
+                                           arg.output_.GetLengths()[4],
+                                           arg.output_.GetLengths()[5])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override
+    {
+        return NDimSpatial >= 1 && NDimSpatial <= 3;
+    }
+
+    static auto MakeArgument(const Tensor<InDataType>& input,
+                             const Tensor<WeiDataType>& weight,
+                             Tensor<OutDataType>& output,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{input,
+                        weight,
+                        output,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceConvFwd"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
new file mode 100644
index 0000000..be22003
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
@@ -0,0 +1,192 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// out[N, Ho, Wo, K] =
+//     activate(in[N, Hi, Wi, C] * wei[K, Y, X, C] + bias[K])
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct ReferenceConvFwd_Bias_Activation : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<InDataType>& in_n_c_hi_wi,
+                 const Tensor<WeiDataType>& wei_k_c_y_x,
+                 Tensor<OutDataType>& out_n_k_ho_wo,
+                 const Tensor<OutDataType>& bias_k,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : in_n_c_hi_wi_{in_n_c_hi_wi},
+              wei_k_c_y_x_{wei_k_c_y_x},
+              out_n_k_ho_wo_{out_n_k_ho_wo},
+              bias_k_{bias_k},
+              conv_strides_{conv_filter_strides},
+              conv_dilations_{conv_filter_dilations},
+              in_left_pads_{input_left_pads},
+              in_right_pads_{input_right_pads},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+        }
+
+        const Tensor<InDataType>& in_n_c_hi_wi_;
+        const Tensor<WeiDataType>& wei_k_c_y_x_;
+        Tensor<OutDataType>& out_n_k_ho_wo_;
+        const Tensor<OutDataType>& bias_k_;
+
+        std::vector<index_t> conv_strides_;
+        std::vector<index_t> conv_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceConvFwd_Bias_Activation::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
+                float v_acc = 0;
+
+                for(std::size_t c = 0; c < arg.wei_k_c_y_x_.mDesc.GetLengths()[1]; ++c)
+                {
+                    for(std::size_t y = 0; y < arg.wei_k_c_y_x_.mDesc.GetLengths()[2]; ++y)
+                    {
+                        auto hi = ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[0]) +
+                                  ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[0]) -
+                                  ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
+                        for(std::size_t x = 0; x < arg.wei_k_c_y_x_.mDesc.GetLengths()[3]; ++x)
+                        {
+                            auto wi =
+                                ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[1]) +
+                                ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[1]) -
+                                ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]);
+                            if(hi >= 0 &&
+                               ck::type_convert<std::size_t>(hi) <
+                                   arg.in_n_c_hi_wi_.mDesc.GetLengths()[2] &&
+                               wi >= 0 &&
+                               ck::type_convert<std::size_t>(wi) <
+                                   arg.in_n_c_hi_wi_.mDesc.GetLengths()[3])
+                            {
+                                float v_in;
+                                float v_wei;
+
+                                arg.in_element_op_(
+                                    v_in,
+                                    static_cast<const float>(arg.in_n_c_hi_wi_(n, c, hi, wi)));
+                                arg.wei_element_op_(
+                                    v_wei, static_cast<const float>(arg.wei_k_c_y_x_(k, c, y, x)));
+
+                                v_acc += v_in * v_wei;
+                            }
+                        }
+                    }
+                }
+
+                float v_out;
+
+                arg.out_element_op_(v_out, v_acc, static_cast<float>(arg.bias_k_(k)));
+
+                arg.out_n_k_ho_wo_(n, k, ho, wo) = v_out;
+            };
+
+            make_ParallelTensorFunctor(f_nchw,
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[0],
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[1],
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[2],
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[3])(
+                std::thread::hardware_concurrency());
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<InDataType>& in_n_c_hi_wi,
+                             const Tensor<WeiDataType>& wei_k_c_y_x,
+                             Tensor<OutDataType>& out_n_k_ho_wo,
+                             const Tensor<OutDataType>& bias_k,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{in_n_c_hi_wi,
+                        wei_k_c_y_x,
+                        out_n_k_ho_wo,
+                        bias_k,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceConvFwd_Bias_Activation"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
new file mode 100644
index 0000000..f949f27
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
@@ -0,0 +1,200 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// out[N, Ho, Wo, K] =
+//     activate(in[N, Hi, Wi, C] * wei[K, Y, X, C] + bias[K]) + residual[N, Ho, Wo, K]
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct ReferenceConvFwd_Bias_Activation_Add : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<InDataType>& in_n_c_hi_wi,
+                 const Tensor<WeiDataType>& wei_k_c_y_x,
+                 Tensor<OutDataType>& out_n_k_ho_wo,
+                 const Tensor<OutDataType>& bias_k,
+                 const Tensor<OutDataType>& resi_n_k_ho_wo,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : in_n_c_hi_wi_{in_n_c_hi_wi},
+              wei_k_c_y_x_{wei_k_c_y_x},
+              out_n_k_ho_wo_{out_n_k_ho_wo},
+              bias_k_{bias_k},
+              resi_n_k_ho_wo_{resi_n_k_ho_wo},
+              conv_strides_{conv_filter_strides},
+              conv_dilations_{conv_filter_dilations},
+              in_left_pads_{input_left_pads},
+              in_right_pads_{input_right_pads},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+        }
+
+        const Tensor<InDataType>& in_n_c_hi_wi_;
+        const Tensor<WeiDataType>& wei_k_c_y_x_;
+        Tensor<OutDataType>& out_n_k_ho_wo_;
+        const Tensor<OutDataType>& bias_k_;
+        const Tensor<OutDataType>& resi_n_k_ho_wo_;
+
+        std::vector<index_t> conv_strides_;
+        std::vector<index_t> conv_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceConvFwd_Bias_Activation_Add::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
+                float v_acc = 0;
+
+                for(std::size_t c = 0; c < arg.wei_k_c_y_x_.mDesc.GetLengths()[1]; ++c)
+                {
+                    for(std::size_t y = 0; y < arg.wei_k_c_y_x_.mDesc.GetLengths()[2]; ++y)
+                    {
+                        auto hi = ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[0]) +
+                                  ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[0]) -
+                                  ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
+                        for(std::size_t x = 0; x < arg.wei_k_c_y_x_.mDesc.GetLengths()[3]; ++x)
+                        {
+                            auto wi =
+                                ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[1]) +
+                                ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[1]) -
+                                ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]);
+                            if(hi >= 0 &&
+                               ck::type_convert<std::size_t>(hi) <
+                                   arg.in_n_c_hi_wi_.mDesc.GetLengths()[2] &&
+                               wi >= 0 &&
+                               ck::type_convert<std::size_t>(wi) <
+                                   arg.in_n_c_hi_wi_.mDesc.GetLengths()[3])
+                            {
+                                float v_in;
+                                float v_wei;
+
+                                arg.in_element_op_(
+                                    v_in,
+                                    static_cast<const float>(arg.in_n_c_hi_wi_(n, c, hi, wi)));
+                                arg.wei_element_op_(
+                                    v_wei, static_cast<const float>(arg.wei_k_c_y_x_(k, c, y, x)));
+
+                                v_acc += v_in * v_wei;
+                            }
+                        }
+                    }
+                }
+
+                float v_out;
+
+                arg.out_element_op_(v_out,
+                                    v_acc,
+                                    static_cast<const float>(arg.bias_k_(k)),
+                                    static_cast<const float>(arg.resi_n_k_ho_wo_(n, k, ho, wo)));
+
+                arg.out_n_k_ho_wo_(n, k, ho, wo) = v_out;
+            };
+
+            make_ParallelTensorFunctor(f_nchw,
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[0],
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[1],
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[2],
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[3])(
+                std::thread::hardware_concurrency());
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<InDataType>& in_n_c_hi_wi,
+                             const Tensor<WeiDataType>& wei_k_c_y_x,
+                             Tensor<OutDataType>& out_n_k_ho_wo,
+                             const Tensor<OutDataType>& bias_k,
+                             const Tensor<OutDataType>& resi_n_k_ho_wo,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{in_n_c_hi_wi,
+                        wei_k_c_y_x,
+                        out_n_k_ho_wo,
+                        bias_k,
+                        resi_n_k_ho_wo,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceConvFwd_Bias_Activation_Add"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
new file mode 100644
index 0000000..6728bb1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct ReferenceGemm : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_m_k,
+                 const Tensor<BDataType>& b_k_n,
+                 Tensor<CDataType>& c_m_n,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_m_k_{a_m_k},
+              b_k_n_{b_k_n},
+              c_m_n_{c_m_n},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_m_k_;
+        const Tensor<BDataType>& b_k_n_;
+        Tensor<CDataType>& c_m_n_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceGemm::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_mk_kn_mn = [&](auto m, auto n) {
+                const int K = arg.a_m_k_.mDesc.GetLengths()[1];
+
+                AccDataType v_acc = 0;
+
+                for(int k = 0; k < K; ++k)
+                {
+                    ADataType v_a;
+                    BDataType v_b;
+
+                    arg.a_element_op_(v_a, arg.a_m_k_(m, k));
+                    arg.b_element_op_(v_b, arg.b_k_n_(k, n));
+
+                    v_acc +=
+                        ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
+                }
+
+                AccDataType v_c;
+
+                arg.c_element_op_(v_c, v_acc);
+
+                arg.c_m_n_(m, n) = ck::type_convert<CDataType>(v_c);
+            };
+
+            make_ParallelTensorFunctor(
+                f_mk_kn_mn, arg.c_m_n_.mDesc.GetLengths()[0], arg.c_m_n_.mDesc.GetLengths()[1])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_m_k,
+                             const Tensor<BDataType>& b_k_n,
+                             Tensor<CDataType>& c_m_n,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceGemm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_2d.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_2d.hpp
new file mode 100644
index 0000000..c77d22f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_2d.hpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename ADataType,
+          typename BDataType,
+          typename C0DataType,
+          typename CDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct ReferenceGemmBias2D : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_m_k,
+                 const Tensor<BDataType>& b_k_n,
+                 const Tensor<C0DataType>& c0_m_n,
+                 Tensor<CDataType>& c_m_n,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_m_k_{a_m_k},
+              b_k_n_{b_k_n},
+              c0_m_n_{c0_m_n},
+              c_m_n_{c_m_n},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_m_k_;
+        const Tensor<BDataType>& b_k_n_;
+        const Tensor<CDataType>& c0_m_n_;
+        Tensor<CDataType>& c_m_n_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceGemmBias2D::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_mk_kn_mn = [&](auto m, auto n) {
+                const int K = arg.a_m_k_.mDesc.GetLengths()[1];
+
+                AccDataType a   = 0;
+                AccDataType b   = 0;
+                AccDataType acc = 0;
+
+                for(int k = 0; k < K; ++k)
+                {
+                    arg.a_element_op_(a, ck::type_convert<AccDataType>(arg.a_m_k_(m, k)));
+                    arg.b_element_op_(b, ck::type_convert<AccDataType>(arg.b_k_n_(k, n)));
+                    acc += a * b;
+                }
+
+                CDataType cast_acc = static_cast<CDataType>(acc);
+                arg.c_element_op_(arg.c_m_n_(m, n), cast_acc, arg.c0_m_n_(m, n));
+            };
+
+            make_ParallelTensorFunctor(
+                f_mk_kn_mn, arg.c_m_n_.mDesc.GetLengths()[0], arg.c_m_n_.mDesc.GetLengths()[1])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_m_k,
+                             const Tensor<BDataType>& b_k_n,
+                             const Tensor<C0DataType>& c0_m_n,
+                             Tensor<CDataType>& c_m_n,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{a_m_k, b_k_n, c0_m_n, c_m_n, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceGemmBias2D"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation.hpp
new file mode 100644
index 0000000..7dfc3c1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation.hpp
@@ -0,0 +1,140 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct ReferenceGemmBiasActivation : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_m_k,
+                 const Tensor<BDataType>& b_k_n,
+                 Tensor<CDataType>& c_m_n,
+                 const Tensor<CDataType>& c0_n,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_m_k_{a_m_k},
+              b_k_n_{b_k_n},
+              c_m_n_{c_m_n},
+              c0_n_{c0_n},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_m_k_;
+        const Tensor<BDataType>& b_k_n_;
+        Tensor<CDataType>& c_m_n_;
+        const Tensor<CDataType>& c0_n_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceGemmBiasActivation::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_mk_kn_mn = [&](auto m, auto n) {
+                const int K = arg.a_m_k_.mDesc.GetLengths()[1];
+
+                float v_acc = 0;
+
+                for(int k = 0; k < K; ++k)
+                {
+                    float v_a;
+                    float v_b;
+
+                    arg.a_element_op_(v_a, static_cast<const float>(arg.a_m_k_(m, k)));
+                    arg.b_element_op_(v_b, static_cast<const float>(arg.b_k_n_(k, n)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                float v_c;
+
+                arg.c_element_op_(v_c, v_acc, static_cast<float>(arg.c0_n_(n)));
+
+                arg.c_m_n_(m, n) = v_c;
+            };
+
+            make_ParallelTensorFunctor(
+                f_mk_kn_mn, arg.c_m_n_.mDesc.GetLengths()[0], arg.c_m_n_.mDesc.GetLengths()[1])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_m_k,
+                             const Tensor<BDataType>& b_k_n,
+                             Tensor<CDataType>& c_m_n,
+                             const Tensor<CDataType>& c0_n,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{a_m_k, b_k_n, c_m_n, c0_n, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceGemmBiasActivation"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation_add.hpp
new file mode 100644
index 0000000..99102a4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation_add.hpp
@@ -0,0 +1,148 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct ReferenceGemmBiasActivationAdd : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_m_k,
+                 const Tensor<BDataType>& b_k_n,
+                 Tensor<CDataType>& c_m_n,
+                 const Tensor<CDataType>& c0_n,
+                 const Tensor<CDataType>& c1_m_n,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_m_k_{a_m_k},
+              b_k_n_{b_k_n},
+              c_m_n_{c_m_n},
+              c0_n_{c0_n},
+              c1_m_n_{c1_m_n},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_m_k_;
+        const Tensor<BDataType>& b_k_n_;
+        Tensor<CDataType>& c_m_n_;
+        const Tensor<CDataType>& c0_n_;
+        const Tensor<CDataType>& c1_m_n_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceGemmBiasActivationAdd::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_mk_kn_mn = [&](auto m, auto n) {
+                const int K = arg.a_m_k_.mDesc.GetLengths()[1];
+
+                float v_acc = 0;
+
+                for(int k = 0; k < K; ++k)
+                {
+                    float v_a;
+                    float v_b;
+
+                    arg.a_element_op_(v_a, static_cast<const float>(arg.a_m_k_(m, k)));
+                    arg.b_element_op_(v_b, static_cast<const float>(arg.b_k_n_(k, n)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                float v_c;
+
+                arg.c_element_op_(v_c,
+                                  v_acc,
+                                  static_cast<float>(arg.c0_n_(n)),
+                                  static_cast<float>(arg.c1_m_n_(m, n)));
+
+                arg.c_m_n_(m, n) = v_c;
+            };
+
+            make_ParallelTensorFunctor(
+                f_mk_kn_mn, arg.c_m_n_.mDesc.GetLengths()[0], arg.c_m_n_.mDesc.GetLengths()[1])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_m_k,
+                             const Tensor<BDataType>& b_k_n,
+                             Tensor<CDataType>& c_m_n,
+                             const Tensor<CDataType>& c0_n,
+                             const Tensor<CDataType>& c1_m_n,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{
+            a_m_k, b_k_n, c_m_n, c0_n, c1_m_n, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceGemmBiasActivationAdd"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp
new file mode 100644
index 0000000..28132aa
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp
@@ -0,0 +1,236 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// D = Layernorm(acc_element_op(A * B + broadcast(bias)) + add) * broadcast(gamma) + broadcast(beta)
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename C0DataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename CElementwiseOperation>
+struct ReferenceGemmLayernorm : public device::BaseOperator
+{
+    using ReferenceGemmInstance = ReferenceGemm<ADataType,
+                                                BDataType,
+                                                AccDataType,
+                                                AccDataType,
+                                                AElementwiseOperation,
+                                                BElementwiseOperation,
+                                                element_wise::PassThrough>;
+
+    template <typename InDataType, typename OutDataType, typename ComputeDataType>
+    static void RunLayernorm(Tensor<OutDataType>& result,
+                             const Tensor<ComputeDataType>& acc, // MxN
+                             const Tensor<InDataType>& gamma,    // 1xN
+                             const Tensor<InDataType>& beta,     // 1xN
+                             const InDataType epsilon = 1e-5)
+    {
+        assert(acc.mDesc.GetLengths()[1] == gamma.mDesc.GetLengths()[0] &&
+               acc.mDesc.GetLengths()[1] == beta.mDesc.GetLengths()[0]);
+
+        size_t M = acc.mDesc.GetLengths()[0];
+        size_t N = acc.mDesc.GetLengths()[1];
+
+        Tensor<ComputeDataType> avg_acc_sq({M});
+        Tensor<ComputeDataType> avg_acc({M});
+        Tensor<ComputeDataType> acc_layernorm(acc);
+
+        // reduce N dim
+        for(size_t i = 0; i < M; i++)
+        {
+            ComputeDataType sum_acc_sq = 0;
+            ComputeDataType sum_acc    = 0;
+            for(size_t j = 0; j < N; j++)
+            {
+                sum_acc_sq += acc_layernorm(i, j) * acc_layernorm(i, j);
+                sum_acc += acc_layernorm(i, j);
+            }
+            avg_acc_sq(i) = sum_acc_sq / N;
+            avg_acc(i)    = sum_acc / N;
+        }
+
+        // normalize
+        acc_layernorm.ForEach([&](auto& self, auto idx) {
+            self(idx[0], idx[1]) =
+                (self(idx[0], idx[1]) - avg_acc(idx[0])) /
+                sqrt(avg_acc_sq(idx[0]) - avg_acc(idx[0]) * avg_acc(idx[0]) + epsilon);
+        });
+
+        // affine
+        acc_layernorm.ForEach([&](auto& self, auto idx) {
+            self(idx[0], idx[1]) = self(idx[0], idx[1]) * gamma(idx[1]) + beta(idx[1]);
+        });
+
+        // cast
+        result = acc_layernorm.template CopyAsType<OutDataType>();
+    }
+
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_m_k,
+                 const Tensor<BDataType>& b_k_n,
+                 Tensor<CDataType>& c_m_n,
+                 const Tensor<C0DataType>& c0_n_bias,  // 1xN
+                 const Tensor<C0DataType>& c0_m_n_add, // MxN
+                 const Tensor<C0DataType>& c0_n_gamma, // 1xN
+                 const Tensor<C0DataType>& c0_n_beta,  // 1xN
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 AccElementwiseOperation acc_element_op,
+                 CElementwiseOperation c_element_op,
+                 const CDataType epsilon = 1e-5)
+            : a_m_k_{a_m_k},
+              b_k_n_{b_k_n},
+              c_m_n_{c_m_n},
+              c0_n_bias_{c0_n_bias},
+              c0_m_n_add_{c0_m_n_add},
+              c0_n_gamma_{c0_n_gamma},
+              c0_n_beta_{c0_n_beta},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              c_element_op_{c_element_op},
+              epsilon_{epsilon}
+        {
+        }
+
+        const Tensor<ADataType>& a_m_k_;
+        const Tensor<BDataType>& b_k_n_;
+        Tensor<CDataType>& c_m_n_;
+        const Tensor<C0DataType>& c0_n_bias_;
+        const Tensor<C0DataType>& c0_m_n_add_;
+        const Tensor<C0DataType>& c0_n_gamma_;
+        const Tensor<C0DataType>& c0_n_beta_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        CElementwiseOperation c_element_op_;
+
+        const CDataType epsilon_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        // using Argument = ReferenceGemm::Argument;
+
+        float Run(const Argument& arg)
+        {
+            Tensor<AccDataType> acc_m_n(arg.c_m_n_.mDesc);
+            acc_m_n.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0});
+
+            auto ref_gemm     = ReferenceGemmInstance{};
+            auto ref_invoker  = ref_gemm.MakeInvoker();
+            auto ref_argument = ref_gemm.MakeArgument(arg.a_m_k_,
+                                                      arg.b_k_n_,
+                                                      acc_m_n,
+                                                      arg.a_element_op_,
+                                                      arg.b_element_op_,
+                                                      element_wise::PassThrough{});
+
+            // gemm
+            ref_invoker.Run(ref_argument);
+
+            // activation(acc + bias)
+            acc_m_n.ForEach([&](auto& self, auto idx) {
+                AccDataType out;
+                arg.acc_element_op_(out, acc_m_n(idx[0], idx[1]) + arg.c0_n_bias_(idx[1]));
+                self(idx[0], idx[1]) = out;
+            });
+
+            // add from other layers
+            acc_m_n.ForEach([&](auto& self, auto idx) {
+                self(idx[0], idx[1]) += arg.c0_m_n_add_(idx[0], idx[1]);
+            });
+
+            // layernorm
+            RunLayernorm(arg.c_m_n_, acc_m_n, arg.c0_n_gamma_, arg.c0_n_beta_);
+
+            // elementwise op
+            arg.c_m_n_.ForEach([&](auto& self, auto idx) {
+                arg.c_element_op_(self(idx[0], idx[1]), self(idx[0], idx[1]));
+            });
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_m_k,
+                             const Tensor<BDataType>& b_k_n,
+                             Tensor<CDataType>& c_m_n,
+                             const Tensor<C0DataType>& c0_n_bias,  // 1xN
+                             const Tensor<C0DataType>& c0_m_n_add, // 1xN
+                             const Tensor<C0DataType>& c0_n_gamma, // 1xN
+                             const Tensor<C0DataType>& c0_n_beta,  // 1xN
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             AccElementwiseOperation acc_element_op,
+                             CElementwiseOperation c_element_op,
+                             const CDataType epsilon = 1e-5)
+    {
+        return Argument{a_m_k,
+                        b_k_n,
+                        c_m_n,
+                        c0_n_bias,
+                        c0_m_n_add,
+                        c0_n_gamma,
+                        c0_n_beta,
+                        a_element_op,
+                        b_element_op,
+                        acc_element_op,
+                        c_element_op,
+                        epsilon};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceGemmLayernorm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
new file mode 100644
index 0000000..fedd4dc
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
@@ -0,0 +1,191 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+#include <algorithm>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename AccElementwiseOperation>
+struct ReferenceGroupnorm : public device::BaseOperator
+{
+    // x = [N, H, W, G, C]
+    // y = [N, H, W, G, C]
+    // reduce dim [H, W, C], mean, var = [N, G]
+    // gamma, beta = [G, C]
+    // beta: [G, C]
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<XDataType>& x,
+                 const Tensor<GammaDataType>& gamma,
+                 const Tensor<BetaDataType>& beta,
+                 Tensor<YDataType>& y,
+                 AccElementwiseOperation acc_elementwise_op,
+                 const std::vector<index_t> lengths,
+                 AccDataType epsilon)
+            : x_(x),
+              gamma_(gamma),
+              beta_(beta),
+              y_(y),
+              acc_elementwise_op_(acc_elementwise_op),
+              lengths_(lengths),
+              epsilon_(epsilon)
+        {
+        }
+
+        const Tensor<XDataType> x_;
+        const Tensor<XDataType> gamma_;
+        const Tensor<XDataType> beta_;
+        Tensor<YDataType>& y_;
+        AccElementwiseOperation acc_elementwise_op_;
+        std::vector<index_t> lengths_;
+        AccDataType epsilon_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            int N = arg.lengths_[0];
+            int H = arg.lengths_[1];
+            int W = arg.lengths_[2];
+            int G = arg.lengths_[3];
+            int C = arg.lengths_[4];
+
+            Tensor<AccDataType> mean({N, G});
+            Tensor<AccDataType> var({N, G});
+
+            // Compute mean & var in [H, W, C] by Welford Algorithm
+            // TODO - parallel for each HWC
+            // TODO - address calculation
+            for(int n = 0; n < N; ++n)
+            {
+                for(int g = 0; g < G; ++g)
+                {
+                    AccDataType mean_val = type_convert<AccDataType>(0.0f);
+                    AccDataType var_val  = type_convert<AccDataType>(0.0f);
+                    int32_t curr_count   = 0;
+
+                    for(int h = 0; h < H; ++h)
+                    {
+                        for(int w = 0; w < W; ++w)
+                        {
+                            for(int c = 0; c < C; ++c)
+                            {
+                                curr_count++;
+                                AccDataType x = type_convert<AccDataType>(arg.x_(n, h, w, g, c));
+                                AccDataType delta = x - mean_val;
+                                mean_val += delta / curr_count;
+                                AccDataType delta2 = x - mean_val;
+                                var_val += delta * delta2;
+                            }
+                        }
+                    }
+
+                    mean(n, g) = mean_val;
+                    var(n, g)  = var_val / curr_count;
+                }
+            }
+
+            // Normalization
+            for(int n = 0; n < N; ++n)
+            {
+                for(int h = 0; h < H; ++h)
+                {
+                    for(int w = 0; w < W; ++w)
+                    {
+                        for(int g = 0; g < G; ++g)
+                        {
+                            for(int c = 0; c < C; ++c)
+                            {
+                                AccDataType x = type_convert<AccDataType>(arg.x_(n, h, w, g, c));
+                                AccDataType gamma    = type_convert<AccDataType>(arg.gamma_(g, c));
+                                AccDataType beta     = type_convert<AccDataType>(arg.beta_(g, c));
+                                AccDataType mean_val = type_convert<AccDataType>(mean(n, g));
+                                AccDataType var_val  = type_convert<AccDataType>(var(n, g));
+                                AccDataType y        = gamma * (x - mean_val) /
+                                                    ck::math::sqrt(arg.epsilon_ + var_val) +
+                                                beta;
+                                arg.acc_elementwise_op_(y, y);
+                                arg.y_(n, h, w, g, c) = type_convert<YDataType>(y);
+                            }
+                        }
+                    }
+                }
+            }
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
+        if(p_arg_->lengths_.size() != 5)
+            return false;
+
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<XDataType>& x,
+                             const Tensor<GammaDataType>& gamma,
+                             const Tensor<BetaDataType>& beta,
+                             Tensor<YDataType>& y,
+                             AccElementwiseOperation acc_elementwise_op,
+                             const std::vector<index_t> lengths,
+                             AccDataType epsilon)
+    {
+        return Argument{x, gamma, beta, y, acc_elementwise_op, lengths, epsilon};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceLayernorm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
new file mode 100644
index 0000000..680d94f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
@@ -0,0 +1,171 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+#include <algorithm>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename AccElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+struct ReferenceLayernorm : public device::BaseOperator
+{
+    // TODO - support generic layernorm
+    static_assert((Rank == 2 && NumReduceDim == 1), "Only support 2D version so far");
+
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<XDataType>& x_m_n,
+                 const Tensor<GammaDataType>& gamma_n,
+                 const Tensor<BetaDataType>& beta_n,
+                 Tensor<YDataType>& y_m_n,
+                 AccElementwiseOperation acc_elementwise_op,
+                 const std::vector<index_t> lengths,
+                 const std::vector<index_t> reduceDims,
+                 AccDataType epsilon)
+            : x_m_n_(x_m_n),
+              gamma_n_(gamma_n),
+              beta_n_(beta_n),
+              y_m_n_(y_m_n),
+              acc_elementwise_op_(acc_elementwise_op),
+              lengths_(lengths),
+              reduceDims_(reduceDims),
+              epsilon_(epsilon)
+        {
+        }
+
+        const Tensor<XDataType> x_m_n_;
+        const Tensor<XDataType> gamma_n_;
+        const Tensor<XDataType> beta_n_;
+        Tensor<YDataType>& y_m_n_;
+        AccElementwiseOperation acc_elementwise_op_;
+        std::vector<index_t> lengths_;
+        std::vector<index_t> reduceDims_;
+        AccDataType epsilon_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            int M = arg.lengths_[0];
+            int N = arg.lengths_[1];
+
+            Tensor<AccDataType> mean({M});
+            Tensor<AccDataType> var({M});
+
+            for(int m = 0; m < M; ++m)
+            {
+                mean(m) = 0;
+                var(m)  = 0;
+
+                for(int n = 0; n < N; ++n)
+                {
+                    auto x_val = ck::type_convert<AccDataType>(arg.x_m_n_(m, n));
+                    mean(m) += x_val;
+                    var(m) += x_val * x_val;
+                }
+
+                mean(m) = mean(m) / N;
+                var(m)  = (var(m) / N) - (mean(m) * mean(m));
+            }
+
+            for(int m = 0; m < M; ++m)
+            {
+                for(int n = 0; n < N; ++n)
+                {
+                    auto x_val = ck::type_convert<AccDataType>(arg.x_m_n_(m, n));
+                    auto y_val = (x_val - mean(m)) / sqrt(var(m) + arg.epsilon_);
+                    y_val      = (y_val * arg.gamma_n_(n)) + arg.beta_n_(n);
+                    arg.acc_elementwise_op_(y_val, y_val);
+                    arg.y_m_n_(m, n) = ck::type_convert<YDataType>(y_val);
+                }
+            }
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
+
+        // TODO - support generic layernorm
+        if(p_arg_->lengths_.size() != 2)
+            return false;
+
+        if(p_arg_->reduceDims_.size() != 1)
+            return false;
+
+        if(p_arg_->reduceDims_[0] != 1)
+            return false;
+
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<XDataType>& x_m_n,
+                             const Tensor<GammaDataType>& gamma_n,
+                             const Tensor<BetaDataType>& beta_n,
+                             Tensor<YDataType>& y_m_n,
+                             AccElementwiseOperation acc_elementwise_op,
+                             const std::vector<index_t> lengths,
+                             const std::vector<index_t> reduceDims,
+                             AccDataType epsilon)
+    {
+        return Argument{
+            x_m_n, gamma_n, beta_n, y_m_n, acc_elementwise_op, lengths, reduceDims, epsilon};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceLayernorm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp
new file mode 100644
index 0000000..4839eb8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp
@@ -0,0 +1,175 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+#include <algorithm>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename InDataType, typename OutDataType, typename AccDataType>
+struct ReferenceSoftmax : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<InDataType>& in,
+                 Tensor<OutDataType>& out,
+                 AccDataType alpha,
+                 AccDataType beta,
+                 const std::vector<index_t> sm_reduce_dims)
+            : in_(in), out_(out), alpha_(alpha), beta_(beta), sm_reduce_dims_(sm_reduce_dims)
+        {
+            // std::cout << "debug: scalar dims: ";
+            for(size_t i = 0; i < in.mDesc.GetNumOfDimension(); i++)
+            {
+                if(std::find(sm_reduce_dims.begin(), sm_reduce_dims.end(), i) ==
+                   sm_reduce_dims.end())
+                {
+                    sm_scalar_dims_.push_back(i);
+                    // std::cout << i << ", ";
+                }
+            }
+            // std::cout << std::endl;
+        }
+
+        const Tensor<InDataType>& in_;
+        Tensor<OutDataType>& out_;
+        AccDataType alpha_;
+        AccDataType beta_;
+        std::vector<index_t> sm_reduce_dims_;
+        std::vector<index_t> sm_scalar_dims_; // dim after internal max/sum reduction
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            std::vector<size_t> scalar_lengths;
+            for(index_t dim : arg.sm_scalar_dims_)
+            {
+                scalar_lengths.push_back(arg.in_.mDesc.GetLengths()[dim]);
+            }
+            // max and sum reduction with final reduced values of dim=0 is a scalar so give it
+            // appropriate lengths of {1}
+            if(arg.sm_scalar_dims_.size() == 0)
+            {
+                scalar_lengths.push_back(1);
+            }
+
+            Tensor<AccDataType> reduce_max(scalar_lengths);
+            reduce_max.GenerateTensorValue(
+                GeneratorTensor_1<AccDataType>{std::numeric_limits<AccDataType>::lowest()});
+            Tensor<AccDataType> reduce_sum(scalar_lengths);
+            reduce_sum.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0});
+
+            // when final reduced values is of dim=0, the index will be transformed into empty
+            // std::vector which is actually a valid input for Tensor::operator(std::vector) and
+            // internally accesses 0'th element
+            auto to_sm_scalar_idx = [&](auto idx) {
+                std::vector<size_t> sm_scalar_idx;
+                for(index_t dim : arg.sm_scalar_dims_)
+                {
+                    sm_scalar_idx.push_back(idx[dim]);
+                }
+                return sm_scalar_idx;
+            };
+
+            arg.in_.ForEach([&](auto& self, auto idx) {
+                reduce_max(to_sm_scalar_idx(idx)) = std::max(
+                    reduce_max(to_sm_scalar_idx(idx)), ck::type_convert<AccDataType>(self(idx)));
+            });
+
+            // LogRangeAsType<float>(std::cout << "reduce_max: ", reduce_max.mData, ",") <<
+            // std::endl;
+
+            Tensor<AccDataType> in_stable(arg.in_.mDesc);
+            in_stable.ForEach([&](auto& self, auto idx) {
+                // numerator = exp(x - max(x))
+                self(idx) = std::exp(ck::type_convert<AccDataType>(arg.in_(idx)) -
+                                     reduce_max(to_sm_scalar_idx(idx)));
+            });
+
+            // LogRangeAsType<float>(std::cout << "in_stable: ", in_stable.mData, ",") << std::endl;
+
+            in_stable.ForEach([&](auto& self, auto idx) {
+                // denominator = sum(exp(x - max(x)))
+                reduce_sum(to_sm_scalar_idx(idx)) += self(idx);
+            });
+
+            // LogRangeAsType<float>(std::cout << "reduce_sum: ", reduce_sum.mData, ",") <<
+            // std::endl;
+
+            arg.out_.ForEach([&](auto& self, auto idx) {
+                AccDataType temp_result =
+                    arg.alpha_ * in_stable(idx) / reduce_sum(to_sm_scalar_idx(idx)) +
+                    arg.beta_ * self(idx);
+                self(idx) = ck::type_convert<OutDataType>(temp_result);
+            });
+
+            // LogRangeAsType<float>(std::cout << "out: ", arg.out_.mData, ",") << std::endl;
+            // reduction along reduce dims
+            // LogRangeAsType<float>(std::cout << "reduce_max: ", reduce_max.mData, ",") <<
+            // std::endl; LogRangeAsType<float>(std::cout << "reduce_sum: ", reduce_sum.mData, ",")
+            // << std::endl;
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<InDataType>& in,
+                             Tensor<OutDataType>& out,
+                             AccDataType alpha,
+                             AccDataType beta,
+                             const std::vector<index_t> sm_reduce_dims)
+    {
+        return Argument{in, out, alpha, beta, sm_reduce_dims};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceSoftmax"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp
new file mode 100644
index 0000000..b6a9b0f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp
@@ -0,0 +1,205 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+#include <algorithm>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename EmbType,
+          typename IndexType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename OutType>
+struct ReferenceSparseEmbedding3ForwardLayernorm : public device::BaseOperator
+{
+    struct Argument : public device::BaseArgument
+    {
+        Argument(Tensor<OutType>& output,
+                 const Tensor<EmbType>& emb_a,
+                 const Tensor<EmbType>& emb_b,
+                 const Tensor<EmbType>& emb_c,
+                 const Tensor<IndexType>& index_a,
+                 const Tensor<IndexType>& index_b,
+                 const Tensor<IndexType>& index_c,
+                 const Tensor<GammaDataType>& gamma,
+                 const Tensor<BetaDataType>& beta,
+                 ck::index_t NumRows,
+                 ck::index_t EmbeddingDim,
+                 ck::index_t IndexLength,
+                 AccDataType epsilon)
+            : output_(output),
+              emb_a_(emb_a),
+              emb_b_(emb_b),
+              emb_c_(emb_c),
+              index_a_(index_a),
+              index_b_(index_b),
+              index_c_(index_c),
+              gamma_(gamma),
+              beta_(beta),
+              NumRows_(NumRows),
+              EmbeddingDim_(EmbeddingDim),
+              IndexLength_(IndexLength),
+              epsilon_(epsilon)
+        {
+        }
+        Tensor<OutType>& output_;
+        const Tensor<EmbType> emb_a_;
+        const Tensor<EmbType> emb_b_;
+        const Tensor<EmbType> emb_c_;
+        const Tensor<IndexType> index_a_;
+        const Tensor<IndexType> index_b_;
+        const Tensor<IndexType> index_c_;
+        const Tensor<GammaDataType> gamma_;
+        const Tensor<BetaDataType> beta_;
+        ck::index_t NumRows_;
+        ck::index_t EmbeddingDim_;
+        ck::index_t IndexLength_;
+        AccDataType epsilon_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            ck::index_t D = arg.EmbeddingDim_;
+            ck::index_t L = arg.IndexLength_;
+            ck::index_t E = arg.NumRows_;
+
+            Tensor<AccDataType> accumulator({L, D});
+
+            Tensor<AccDataType> mean({L});
+            Tensor<AccDataType> var({L});
+
+            accumulator.SetZero();
+
+            auto f_emb_per_row = [&](auto idx) {
+                IndexType idx_a = arg.index_a_(idx);
+                IndexType idx_b = arg.index_b_(idx);
+                IndexType idx_c = arg.index_c_(idx);
+
+                if(!((idx_a < E) && (idx_b < E) && (idx_c < E)))
+                {
+                    throw(std::runtime_error("wrong! out of range"));
+                }
+
+                for(auto d = 0; d < D; d++)
+                {
+                    auto v_a = ck::type_convert<AccDataType>(arg.emb_a_(idx_a, d));
+                    auto v_b = ck::type_convert<AccDataType>(arg.emb_b_(idx_b, d));
+                    auto v_c = ck::type_convert<AccDataType>(arg.emb_c_(idx_c, d));
+
+                    accumulator(idx, d) += v_a + v_b + v_c;
+                }
+            };
+            make_ParallelTensorFunctor(f_emb_per_row, L)(std::thread::hardware_concurrency());
+
+            // layernorm
+            for(auto idx = 0; idx < L; ++idx)
+            {
+                mean(idx) = 0;
+                var(idx)  = 0;
+
+                for(auto d = 0; d < D; ++d)
+                {
+                    auto x_val = accumulator(idx, d);
+                    mean(idx) += x_val;
+                    var(idx) += x_val * x_val;
+                }
+
+                mean(idx) = mean(idx) / D;
+                var(idx)  = (var(idx) / D) - (mean(idx) * mean(idx));
+            }
+
+            for(auto idx = 0; idx < L; ++idx)
+            {
+                for(auto d = 0; d < D; ++d)
+                {
+                    auto x_val          = accumulator(idx, d);
+                    auto y_val          = (x_val - mean(idx)) / sqrt(var(idx) + arg.epsilon_);
+                    y_val               = (y_val * arg.gamma_(d)) + arg.beta_(d);
+                    arg.output_(idx, d) = ck::type_convert<OutType>(y_val);
+                }
+            }
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(Tensor<OutType>& output,
+                             const Tensor<EmbType>& emb_a,
+                             const Tensor<EmbType>& emb_b,
+                             const Tensor<EmbType>& emb_c,
+                             const Tensor<IndexType>& index_a,
+                             const Tensor<IndexType>& index_b,
+                             const Tensor<IndexType>& index_c,
+                             const Tensor<GammaDataType>& gamma,
+                             const Tensor<BetaDataType>& beta,
+                             ck::index_t NumRows,
+                             ck::index_t EmbeddingDim,
+                             ck::index_t IndexLength,
+                             AccDataType epsilon)
+    {
+        return Argument(output,
+                        emb_a,
+                        emb_b,
+                        emb_c,
+                        index_a,
+                        index_b,
+                        index_c,
+                        gamma,
+                        beta,
+                        NumRows,
+                        EmbeddingDim,
+                        IndexLength,
+                        epsilon);
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceSparseEmbedding3ForwardLayernorm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp
new file mode 100644
index 0000000..df4fca6
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp
@@ -0,0 +1,125 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef NAIVE_CONV_FWD_HPP
+#define NAIVE_CONV_FWD_HPP
+
+namespace ck {
+namespace ref {
+
+/*
+ * \brief naive implementation of 3D convolution. Layout is (NDHWC, KZYXC, NDHWK).
+ *
+ * \param N number of batches
+ * \param K number of filters
+ * \param C number of channels of weight
+ * \param (Di, Hi, Wi) depth, height and width dimension of data
+ * \param (Z, Y, X) depth, height and width dimensions of weights
+ * \param (Do, Ho, Wo) depth, height and width dimension of output
+ * \param (stride_z, stride_y, stride_x) strides
+ * \param (dilation_z, dilation_y, dilation_x) dilations
+ * \param (pad_z, pad_y, pad_x) pads
+ */
+template <typename TIn,
+          typename TWei,
+          typename TOut,
+          typename TAcc,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+__global__ void naive_conv_fwd_ndhwc_kzyxc_ndhwk(const TIn* __restrict__ p_in,
+                                                 const TWei* __restrict__ p_wei,
+                                                 TOut* __restrict__ p_out,
+                                                 index_t N,
+                                                 index_t K,
+                                                 index_t C,
+                                                 index_t Di,
+                                                 index_t Hi,
+                                                 index_t Wi,
+                                                 index_t Z,
+                                                 index_t Y,
+                                                 index_t X,
+                                                 index_t Do,
+                                                 index_t Ho,
+                                                 index_t Wo,
+                                                 index_t stride_z,
+                                                 index_t stride_y,
+                                                 index_t stride_x,
+                                                 index_t dilation_z,
+                                                 index_t dilation_y,
+                                                 index_t dilation_x,
+                                                 index_t pad_z,
+                                                 index_t pad_y,
+                                                 index_t pad_x)
+{
+    const index_t tid                = blockIdx.x * blockDim.x + threadIdx.x;
+    const index_t num_threads        = blockDim.x * gridDim.x;
+    const long_index_t output_length = N * Do * Ho * Wo * K;
+
+    const index_t out_strides[] = {Do * Ho * Wo * K, Ho * Wo * K, Wo * K, K};
+    const index_t in_strides[]  = {Di * Hi * Wi * C, Hi * Wi * C, Wi * C, C};
+    const index_t wei_strides[] = {Z * Y * X * C, Y * X * C, X * C, C};
+
+    constexpr auto in_op  = InElementwiseOperation{};
+    constexpr auto wei_op = WeiElementwiseOperation{};
+    constexpr auto out_op = OutElementwiseOperation{};
+
+    TIn in_val;
+    TWei wei_val;
+    TOut out_val;
+
+    for(long_index_t ii = tid; ii < output_length; ii += num_threads)
+    {
+        const index_t n  = ii / out_strides[0];
+        index_t k        = ii - n * out_strides[0];
+        const index_t dO = k / out_strides[1];
+        k -= dO * out_strides[1];
+        const index_t ho = k / out_strides[2];
+        k -= ho * out_strides[2];
+        const index_t wo = k / out_strides[3];
+        k -= wo * out_strides[3];
+
+        TAcc acc = static_cast<TAcc>(0);
+
+        const TIn* in_n   = p_in + static_cast<long_index_t>(n) * in_strides[0];
+        const TWei* wei_k = p_wei + static_cast<long_index_t>(k) * wei_strides[0];
+
+        for(index_t z = 0; z < Z; ++z)
+        {
+            index_t di          = stride_z * dO - pad_z + dilation_z * z;
+            const TIn* in_n_di  = in_n + di * in_strides[1];
+            const TWei* wei_k_z = wei_k + z * wei_strides[1];
+
+            for(index_t y = 0; y < Y; ++y)
+            {
+                index_t hi            = stride_y * ho - pad_y + dilation_y * y;
+                const TIn* in_n_di_hi = in_n_di + hi * in_strides[2];
+                const TWei* wei_k_z_y = wei_k_z + y * wei_strides[2];
+
+                for(index_t x = 0; x < X; ++x)
+                {
+                    index_t wi               = stride_x * wo - pad_x + dilation_x * x;
+                    const TIn* in_n_di_hi_wi = in_n_di_hi + wi * in_strides[3];
+                    const TWei* wei_k_z_y_x  = wei_k_z_y + x * wei_strides[3];
+
+                    if(di >= 0 && di < Di && hi >= 0 && hi < Hi && wi >= 0 && wi < Wi)
+                    {
+                        for(index_t c = 0; c < C; ++c)
+                        {
+                            in_op(in_val, in_n_di_hi_wi[c]);
+                            wei_op(wei_val, wei_k_z_y_x[c]);
+                            acc += in_val * wei_val;
+                        }
+                    }
+                }
+            }
+        }
+
+        out_op(out_val, static_cast<TOut>(acc));
+        p_out[ii] = out_val;
+    }
+}
+} // namespace ref
+} // namespace ck
+
+#endif
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp
new file mode 100644
index 0000000..20df1b3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <type_traits>
+
+#include "ck/utility/functional2.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <typename BaseOp, typename NewOpInstances>
+void add_device_operation_instances(std::vector<std::unique_ptr<BaseOp>>& op_instances,
+                                    const NewOpInstances& new_op_instances)
+{
+    ck::static_for<0, std::tuple_size_v<NewOpInstances>, 1>{}([&](auto i) {
+        const auto new_op_instance = std::get<i>(new_op_instances);
+
+        using NewOpInstance = remove_cvref_t<decltype(new_op_instance)>;
+
+        static_assert(std::is_base_of_v<BaseOp, NewOpInstance>,
+                      "wrong! NewOpInstance should be derived from BaseOp");
+
+        op_instances.push_back(std::make_unique<NewOpInstance>(new_op_instance));
+    });
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
new file mode 100644
index 0000000..91980a9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
@@ -0,0 +1,114 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// aliasing, for commonly used data type
+using F64  = double;
+using F32  = float;
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using I32  = int32_t;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using F16_Tuple     = ck::Tuple<F16>;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using F32_Tuple     = ck::Tuple<F32>;
+using I32_Tuple     = ck::Tuple<I32>;
+using I32_F32_Tuple = ck::Tuple<I32, F32>;
+
+// GEMM layout
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using Row_Tuple     = ck::Tuple<Row>;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+// Conv layout
+//
+using NWC   = ck::tensor_layout::convolution::NWC;
+using NHWC  = ck::tensor_layout::convolution::NHWC;
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+using KXC   = ck::tensor_layout::convolution::KXC;
+using KYXC  = ck::tensor_layout::convolution::KYXC;
+using KZYXC = ck::tensor_layout::convolution::KZYXC;
+
+using NWK   = ck::tensor_layout::convolution::NWK;
+using NHWK  = ck::tensor_layout::convolution::NHWK;
+using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+//
+using GNWC   = ck::tensor_layout::convolution::GNWC;
+using GNHWC  = ck::tensor_layout::convolution::GNHWC;
+using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
+
+using GKXC   = ck::tensor_layout::convolution::GKXC;
+using GKYXC  = ck::tensor_layout::convolution::GKYXC;
+using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
+
+using GNWK   = ck::tensor_layout::convolution::GNWK;
+using GNHWK  = ck::tensor_layout::convolution::GNHWK;
+using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
+
+//
+using NWGC   = ck::tensor_layout::convolution::NWGC;
+using NHWGC  = ck::tensor_layout::convolution::NHWGC;
+using NDHWGC = ck::tensor_layout::convolution::NDHWGC;
+
+using KXGC   = ck::tensor_layout::convolution::KXGC;
+using KYXGC  = ck::tensor_layout::convolution::KYXGC;
+using KZYXGC = ck::tensor_layout::convolution::KZYXGC;
+
+using NWGK   = ck::tensor_layout::convolution::NWGK;
+using NHWGK  = ck::tensor_layout::convolution::NHWGK;
+using NDHWGK = ck::tensor_layout::convolution::NDHWGK;
+
+//
+using GK          = ck::tensor_layout::convolution::G_K;
+using GK_Tuple    = ck::Tuple<GK>;
+using GK_GK_Tuple = ck::Tuple<GK, GK>;
+
+// pointwise functor
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using Relu           = ck::tensor_operation::element_wise::Relu;
+using Scale          = ck::tensor_operation::element_wise::Scale;
+using Bilinear       = ck::tensor_operation::element_wise::Bilinear;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+using AddFastGelu    = ck::tensor_operation::element_wise::AddFastGelu;
+using FastGelu       = ck::tensor_operation::element_wise::FastGelu;
+
+template <typename Activation>
+using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;
+
+template <typename Activation>
+using Add_Activation_Mul_Clamp =
+    ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<Activation>;
+
+template <typename Activation>
+using Activation_Mul2_Clamp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<Activation>;
+
+template <typename Activation>
+using Add_Activation_Mul2_Clamp =
+    ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<Activation>;
+
+template <typename DeviceOp, typename Tag = void>
+struct DeviceOperationInstanceFactory;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp
new file mode 100644
index 0000000..0655fd9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp
@@ -0,0 +1,259 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Col,
+                                                  Row,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Col,
+                                                  Col,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Row,
+                                                  Row,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Row,
+                                                  Col,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatchedGemm<
+    ALayout,
+    BLayout,
+    CLayout,
+    ADataType,
+    BDataType,
+    CDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceBatchedGemm<ALayout,
+                                       BLayout,
+                                       CLayout,
+                                       ADataType,
+                                       BDataType,
+                                       CDataType,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
+                     is_same_v<CDataType, float>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                          is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
+                          is_same_v<CDataType, bhalf_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
+                          is_same_v<CDataType, int8_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp
new file mode 100644
index 0000000..495c5f8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp
@@ -0,0 +1,139 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
+using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultipleDGemmMultipleD<Row,
+                                                                        Col,
+                                                                        ck::Tuple<Row>,
+                                                                        Row,
+                                                                        ck::Tuple<Row>,
+                                                                        Row,
+                                                                        F16,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        CDE0ElementOp,
+                                                                        PassThrough,
+                                                                        CDE1ElementOp>>>&
+        instances);
+
+void add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultipleDGemmMultipleD<Row,
+                                                                        Col,
+                                                                        ck::Tuple<Row>,
+                                                                        Col,
+                                                                        ck::Tuple<Row>,
+                                                                        Row,
+                                                                        F16,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        CDE0ElementOp,
+                                                                        PassThrough,
+                                                                        CDE1ElementOp>>>&
+        instances);
+
+template <typename A0Layout,
+          typename B0Layout,
+          typename D0sLayout,
+          typename B1Layout,
+          typename D1sLayout,
+          typename E1Layout,
+          typename A0DataType,
+          typename B0DataType,
+          typename D0sDataType,
+          typename B1DataType,
+          typename D1sDataType,
+          typename E1DataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmMultipleDGemmMultipleD<A0Layout,
+                                                                          B0Layout,
+                                                                          D0sLayout,
+                                                                          B1Layout,
+                                                                          D1sLayout,
+                                                                          E1Layout,
+                                                                          A0DataType,
+                                                                          B0DataType,
+                                                                          D0sDataType,
+                                                                          B1DataType,
+                                                                          D1sDataType,
+                                                                          E1DataType,
+                                                                          PassThrough,
+                                                                          PassThrough,
+                                                                          CDE0ElementOp,
+                                                                          PassThrough,
+                                                                          CDE1ElementOp>>
+{
+    using DeviceOp = DeviceBatchedGemmMultipleDGemmMultipleD<A0Layout,
+                                                             B0Layout,
+                                                             D0sLayout,
+                                                             B1Layout,
+                                                             D1sLayout,
+                                                             E1Layout,
+                                                             A0DataType,
+                                                             B0DataType,
+                                                             D0sDataType,
+                                                             B1DataType,
+                                                             D1sDataType,
+                                                             E1DataType,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             CDE0ElementOp,
+                                                             PassThrough,
+                                                             CDE1ElementOp>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<A0DataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<E1DataType, half_t>)
+        {
+            if constexpr(is_same_v<A0Layout, Row> && is_same_v<B0Layout, Col> &&
+                         is_same_v<B1Layout, Row> && is_same_v<E1Layout, Row>)
+            {
+                add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<A0Layout, Row> && is_same_v<B0Layout, Col> &&
+                              is_same_v<B1Layout, Col> && is_same_v<E1Layout, Row>)
+            {
+                add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp
new file mode 100644
index 0000000..a6dcfa3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp
@@ -0,0 +1,113 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
+                                                      Col,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
+                                                      Col,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+template <typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmGemm<ALayout,
+                                                        B0Layout,
+                                                        B1Layout,
+                                                        CLayout,
+                                                        ADataType,
+                                                        B0DataType,
+                                                        B1DataType,
+                                                        CDataType,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>
+{
+    using DeviceOp = DeviceBatchedGemmGemm<ALayout,
+                                           B0Layout,
+                                           B1Layout,
+                                           CLayout,
+                                           ADataType,
+                                           B0DataType,
+                                           B1DataType,
+                                           CDataType,
+                                           PassThrough,
+                                           PassThrough,
+                                           PassThrough,
+                                           PassThrough,
+                                           PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<B0Layout, Col> &&
+                         is_same_v<B1Layout, Row> && is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<B0Layout, Col> &&
+                              is_same_v<B1Layout, Col> && is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
new file mode 100644
index 0000000..8a0b1b1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
@@ -0,0 +1,121 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemm<Row,
+                                                             Col,
+                                                             Row,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             Scale,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             false>>>& instances);
+
+void add_device_batched_gemm_masking_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemm<Row,
+                                                             Col,
+                                                             Row,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             Scale,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             true>>>& instances);
+
+template <typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          bool MaskOutUpperTriangle>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm<ALayout,
+                                                               B0Layout,
+                                                               B1Layout,
+                                                               CLayout,
+                                                               ADataType,
+                                                               B0DataType,
+                                                               B1DataType,
+                                                               CDataType,
+                                                               PassThrough,
+                                                               PassThrough,
+                                                               Scale,
+                                                               PassThrough,
+                                                               PassThrough,
+                                                               MaskOutUpperTriangle>>
+{
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemm<ALayout,
+                                                  B0Layout,
+                                                  B1Layout,
+                                                  CLayout,
+                                                  ADataType,
+                                                  B0DataType,
+                                                  B1DataType,
+                                                  CDataType,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  Scale,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  MaskOutUpperTriangle>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<B0Layout, Col> &&
+                         is_same_v<B1Layout, Row> && is_same_v<CLayout, Row>)
+            {
+                if constexpr(MaskOutUpperTriangle)
+                {
+                    add_device_batched_gemm_masking_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+                        op_ptrs);
+                }
+                else
+                {
+                    add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+                        op_ptrs);
+                }
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp
new file mode 100644
index 0000000..89df1a7
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp
@@ -0,0 +1,185 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            ck::Tuple<>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            Scale,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            ck::Tuple<>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            Scale,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+
+void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            ck::Tuple<>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            Scale,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            ck::Tuple<>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            Scale,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+
+template <typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          MaskingSpecialization MaskingSpec>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      ADataType,
+                                                                      B0DataType,
+                                                                      B1DataType,
+                                                                      CDataType,
+                                                                      ck::Tuple<>,
+                                                                      ck::Tuple<>,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      Scale,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      MaskingSpec>>
+{
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         ADataType,
+                                                         B0DataType,
+                                                         B1DataType,
+                                                         CDataType,
+                                                         ck::Tuple<>,
+                                                         ck::Tuple<>,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Scale,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         MaskingSpec>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
+                          is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp
new file mode 100644
index 0000000..c84ffcf
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp
@@ -0,0 +1,124 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_batchnorm_backward_rank_4_3_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<F16, F32, F32, F32, F16, F32, F32, PassThrough, 4, 3>>>&);
+
+// FP32
+void add_device_batchnorm_backward_rank_4_3_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<F32, F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&);
+
+// BF16
+void add_device_batchnorm_backward_rank_4_3_bf16_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<BF16, F32, F32, F32, BF16, F32, F32, PassThrough, 4, 3>>>&);
+
+// FP64
+void add_device_batchnorm_backward_rank_4_3_f64_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<F64, F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&);
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
+                                                     DxDataType,
+                                                     DyDataType,
+                                                     AccDataType,
+                                                     ScaleDataType,
+                                                     DscaleDbiasDataType,
+                                                     MeanVarDataType,
+                                                     DyElementwiseOp,
+                                                     Rank,
+                                                     NumReduceDim>>
+{
+    using DeviceOp = DeviceBatchNormBwd<XDataType,
+                                        DxDataType,
+                                        DyDataType,
+                                        AccDataType,
+                                        ScaleDataType,
+                                        DscaleDbiasDataType,
+                                        MeanVarDataType,
+                                        DyElementwiseOp,
+                                        Rank,
+                                        NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<XDataType, F16> && is_same_v<DxDataType, F32> &&
+                     is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> &&
+                     is_same_v<ScaleDataType, F16> && is_same_v<DscaleDbiasDataType, F32> &&
+                     is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_backward_rank_4_3_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F32> && is_same_v<DxDataType, F32> &&
+                          is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> &&
+                          is_same_v<ScaleDataType, F32> && is_same_v<DscaleDbiasDataType, F32> &&
+                          is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_backward_rank_4_3_f32_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, BF16> && is_same_v<DxDataType, F32> &&
+                          is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> &&
+                          is_same_v<ScaleDataType, BF16> && is_same_v<DscaleDbiasDataType, F32> &&
+                          is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_backward_rank_4_3_bf16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F64> && is_same_v<DxDataType, F64> &&
+                          is_same_v<DyDataType, F64> && is_same_v<AccDataType, F64> &&
+                          is_same_v<ScaleDataType, F64> && is_same_v<DscaleDbiasDataType, F64> &&
+                          is_same_v<MeanVarDataType, F64>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_backward_rank_4_3_f64_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp
new file mode 100644
index 0000000..8e40d60
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp
@@ -0,0 +1,117 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_batchnorm_forward_rank_4_3_f16_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F16, F16, F32, F16, F16, F32, PassThrough, 4, 3>>>&);
+
+// FP32
+void add_device_batchnorm_forward_rank_4_3_f32_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&);
+
+// BF16
+void add_device_batchnorm_forward_rank_4_3_bf16_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<BF16, BF16, F32, BF16, BF16, F32, PassThrough, 4, 3>>>&);
+
+// FP64
+void add_device_batchnorm_forward_rank_4_3_f64_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&);
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                     YDataType,
+                                                     AccDataType,
+                                                     ScaleDataType,
+                                                     BiasDataType,
+                                                     MeanVarDataType,
+                                                     YElementwiseOp,
+                                                     Rank,
+                                                     NumReduceDim>>
+{
+    using DeviceOp = DeviceBatchNormFwd<XDataType,
+                                        YDataType,
+                                        AccDataType,
+                                        ScaleDataType,
+                                        BiasDataType,
+                                        MeanVarDataType,
+                                        YElementwiseOp,
+                                        Rank,
+                                        NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<XDataType, F16> && is_same_v<YDataType, F16> &&
+                     is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, F16> &&
+                     is_same_v<BiasDataType, F16> && is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_forward_rank_4_3_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F32> && is_same_v<YDataType, F32> &&
+                          is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, F32> &&
+                          is_same_v<BiasDataType, F32> && is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_forward_rank_4_3_f32_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, BF16> && is_same_v<YDataType, BF16> &&
+                          is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, BF16> &&
+                          is_same_v<BiasDataType, BF16> && is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_forward_rank_4_3_bf16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F64> && is_same_v<YDataType, F64> &&
+                          is_same_v<AccDataType, F64> && is_same_v<ScaleDataType, F64> &&
+                          is_same_v<BiasDataType, F64> && is_same_v<MeanVarDataType, F64>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_forward_rank_4_3_f64_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
new file mode 100644
index 0000000..a0cea7e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
@@ -0,0 +1,128 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances);
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances);
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances);
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances);
+
+// Contraction + Bilinear
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD<
+    NumDimM,
+    NumDimN,
+    NumDimK,
+    ADataType,
+    BDataType,
+    ck::Tuple<DDataType>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::Bilinear>>
+{
+    using DeviceOp = DeviceContractionMultipleD<NumDimM,
+                                                NumDimN,
+                                                NumDimK,
+                                                ADataType,
+                                                BDataType,
+                                                ck::Tuple<DDataType>,
+                                                EDataType,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::Bilinear>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
+                     is_same_v<DDataType, float> && is_same_v<EDataType, float>)
+        {
+            if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
+            {
+                add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
+                    op_ptrs);
+                add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
+                    op_ptrs);
+                add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance(
+                    op_ptrs);
+                add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
new file mode 100644
index 0000000..e921ecd
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
@@ -0,0 +1,127 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances);
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances);
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances);
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances);
+
+// Contraction + Scale
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD<
+    NumDimM,
+    NumDimN,
+    NumDimK,
+    ADataType,
+    BDataType,
+    ck::Tuple<>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::Scale>>
+{
+    using DeviceOp = DeviceContractionMultipleD<NumDimM,
+                                                NumDimN,
+                                                NumDimK,
+                                                ADataType,
+                                                BDataType,
+                                                ck::Tuple<>,
+                                                EDataType,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::Scale>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
+                     is_same_v<EDataType, float>)
+        {
+            if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
+            {
+                add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
+                    op_ptrs);
+                add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
+                    op_ptrs);
+                add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
+                    op_ptrs);
+                add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
new file mode 100644
index 0000000..ec5d18f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
@@ -0,0 +1,309 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// conv1d backward data
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<1,
+                                                  NWC,
+                                                  KXC,
+                                                  NWK,
+                                                  BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvBwdData<1, NWC, KXC, NWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvBwdData<1, NWC, KXC, NWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<1,
+                                                  NWC,
+                                                  KXC,
+                                                  NWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+// conv2d backward data
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F16,
+                                                  F16,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F32,
+                                                  F32,
+                                                  F32,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+// conv2d dl
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F16,
+                                                  F16,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F32,
+                                                  F32,
+                                                  F32,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+// conv3d backward data
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  F16,
+                                                  F16,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  F32,
+                                                  F32,
+                                                  F32,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBwdData<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceConvBwdData<NumDimSpatial,
+                                       InLayout,
+                                       WeiLayout,
+                                       OutLayout,
+                                       InDataType,
+                                       WeiDataType,
+                                       OutDataType,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, NWC> && is_same_v<WeiLayout, KXC> &&
+                     is_same_v<OutLayout, NWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
+                          is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
+                add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+                add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
+                add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
+                          is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
new file mode 100644
index 0000000..62f28c9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
@@ -0,0 +1,128 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// conv2d forward
+void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvFwd<2, NHWC, KYXC, NHWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvFwd<2,
+                                              NHWC,
+                                              KYXC,
+                                              NHWK,
+                                              BF16,
+                                              BF16,
+                                              BF16,
+                                              PassThrough,
+                                              PassThrough,
+                                              PassThrough>>>& instances);
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvFwd<2, NHWC, KYXC, NHWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvFwd<2, NHWC, KYXC, NHWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvFwd<2,
+                                              NHWC,
+                                              KYXC,
+                                              NHWK,
+                                              int8_t,
+                                              int8_t,
+                                              int8_t,
+                                              PassThrough,
+                                              PassThrough,
+                                              PassThrough>>>& instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceConvFwd<NumDimSpatial,
+                                                InLayout,
+                                                WeiLayout,
+                                                OutLayout,
+                                                InDataType,
+                                                WeiDataType,
+                                                OutDataType,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceConvFwd<NumDimSpatial,
+                                   InLayout,
+                                   WeiLayout,
+                                   OutLayout,
+                                   InDataType,
+                                   WeiDataType,
+                                   OutDataType,
+                                   ck::tensor_operation::element_wise::PassThrough,
+                                   ck::tensor_operation::element_wise::PassThrough,
+                                   ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
+                     is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+                add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp
new file mode 100644
index 0000000..141af55
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp
@@ -0,0 +1,54 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Normalize                            = ck::tensor_operation::element_wise::Normalize;
+using DeviceNormalizeFromMeanMeanSquarePtr = ck::tensor_operation::device::DeviceElementwiseBasePtr<
+    Tuple<half_t, float, float, half_t, half_t>,
+    Tuple<half_t>,
+    Normalize,
+    2>;
+
+void add_device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances(
+    std::vector<DeviceNormalizeFromMeanMeanSquarePtr>& instances);
+
+template <typename InputType,
+          typename MeanType,
+          typename MeanSquareType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename OutputType>
+auto get_device_normalize_from_mean_meansquare_instances()
+{
+    std::vector<DeviceNormalizeFromMeanMeanSquarePtr> op_ptrs;
+
+    if constexpr(is_same<InputType, half_t>::value && is_same<MeanType, float>::value &&
+                 is_same<MeanSquareType, float>::value && is_same<GammaDataType, half_t>::value &&
+                 is_same<BetaDataType, half_t>::value && is_same<OutputType, half_t>::value)
+    {
+        ck::tensor_operation::device::instance::
+            add_device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances(op_ptrs);
+    }
+
+    return op_ptrs;
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp
new file mode 100644
index 0000000..682f546
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using DeviceGemmAddAddMeanSquareMeanPtr = ck::tensor_operation::device::DeviceGemmReducePtr<1, 2>;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+    std::vector<DeviceGemmAddAddMeanSquareMeanPtr>&);
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+    std::vector<DeviceGemmAddAddMeanSquareMeanPtr>&);
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+    std::vector<DeviceGemmAddAddMeanSquareMeanPtr>&);
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+    std::vector<DeviceGemmAddAddMeanSquareMeanPtr>&);
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+auto get_device_gemm_add_add_mean_squaremean_instances()
+{
+    std::vector<DeviceGemmAddAddMeanSquareMeanPtr> op_ptrs;
+
+    if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
+                 is_same<CDataType, half_t>::value)
+    {
+        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+                    op_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+                    op_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+                    op_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+                    op_ptrs);
+        }
+    }
+
+    return op_ptrs;
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp
new file mode 100644
index 0000000..c87ae15
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_elementwise_normalization_rank_2_1_f16_instances(
+    std::vector<std::unique_ptr<DeviceElementwiseNormalization<ck::Tuple<F16, F16>,
+                                                               F16,
+                                                               F16,
+                                                               F32,
+                                                               F16,
+                                                               element_wise::Add,
+                                                               PassThrough,
+                                                               2,
+                                                               1>>>&);
+
+template <typename InDataTypeTuple,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElementwiseNormalization<
+    InDataTypeTuple,
+    GammaDataType,
+    BetaDataType,
+    F32,
+    YDataType,
+    ck::tensor_operation::element_wise::Add,
+    ck::tensor_operation::element_wise::PassThrough,
+    Rank,
+    NumReduceDim>>
+{
+    using DeviceOp = DeviceElementwiseNormalization<InDataTypeTuple,
+                                                    GammaDataType,
+                                                    BetaDataType,
+                                                    F32,
+                                                    YDataType,
+                                                    ck::tensor_operation::element_wise::Add,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    Rank,
+                                                    NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<GammaDataType, F16> && is_same_v<BetaDataType, F16> &&
+                     is_same_v<YDataType, F16>)
+        {
+            if constexpr(Rank == 2 && NumReduceDim == 1)
+            {
+                add_device_elementwise_normalization_rank_2_1_f16_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
new file mode 100644
index 0000000..e230507
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
@@ -0,0 +1,385 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <memory>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+
+        instances);
+
+void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+
+        DeviceGemm<Col, Row, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f64_f64_f64_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGemm<ALayout,
+                                             BLayout,
+                                             CLayout,
+                                             ADataType,
+                                             BDataType,
+                                             CDataType,
+                                             ck::tensor_operation::element_wise::PassThrough,
+                                             ck::tensor_operation::element_wise::PassThrough,
+                                             ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemm<ALayout,
+                                BLayout,
+                                CLayout,
+                                ADataType,
+                                BDataType,
+                                CDataType,
+                                ck::tensor_operation::element_wise::PassThrough,
+                                ck::tensor_operation::element_wise::PassThrough,
+                                ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
+                     is_same_v<CDataType, float>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+                add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+                add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+                add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+                add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+                add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+                add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+                add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+                add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                          is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+                add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+                add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+                add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+                add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+                add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+                add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, ck::bhalf_t> &&
+                          is_same_v<CDataType, ck::bhalf_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
+                          is_same_v<CDataType, int8_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(op_ptrs);
+                add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(op_ptrs);
+                add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(op_ptrs);
+                add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(op_ptrs);
+                add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp
new file mode 100644
index 0000000..09d8e8b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp
@@ -0,0 +1,155 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>&);
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>&);
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>&);
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>&);
+
+// GEMM + Add + Add + FastGelu
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
+    ALayout,
+    BLayout,
+    ck::Tuple<D0Layout, D1Layout>,
+    ELayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<D0DataType, D1DataType>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::AddAddFastGelu>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         ck::Tuple<D0Layout, D1Layout>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         ck::Tuple<D0DataType, D1DataType>,
+                                         EDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::AddAddFastGelu>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
new file mode 100644
index 0000000..554437f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>&);
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>&);
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>&);
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>&);
+
+// GEMM + Add + FastGelu
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                      BLayout,
+                                                      ck::Tuple<D0Layout>,
+                                                      ELayout,
+                                                      ADataType,
+                                                      BDataType,
+                                                      ck::Tuple<D0DataType>,
+                                                      EDataType,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      AddFastGelu>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         ck::Tuple<D0Layout>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         ck::Tuple<D0DataType>,
+                                         EDataType,
+                                         PassThrough,
+                                         PassThrough,
+                                         AddFastGelu>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp
new file mode 100644
index 0000000..ef70504
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp
@@ -0,0 +1,148 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances);
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances);
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances);
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances);
+
+// GEMM + Bilinear
+template <typename ALayout,
+          typename BLayout,
+          typename DLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
+    ALayout,
+    BLayout,
+    ck::Tuple<DLayout>,
+    ELayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<DDataType>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::Bilinear>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         ck::Tuple<DLayout>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         ck::Tuple<DDataType>,
+                                         EDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::Bilinear>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<DDataType, half_t> && is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<DLayout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<DLayout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<DLayout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<DLayout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp
new file mode 100644
index 0000000..fbc5df9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp
@@ -0,0 +1,138 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>&);
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>&);
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>&);
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>&);
+
+// GEMM + FastGelu
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                                                        BLayout,
+                                                                                        Empty_Tuple,
+                                                                                        ELayout,
+                                                                                        ADataType,
+                                                                                        BDataType,
+                                                                                        Empty_Tuple,
+                                                                                        EDataType,
+                                                                                        PassThrough,
+                                                                                        PassThrough,
+                                                                                        FastGelu>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         Empty_Tuple,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         Empty_Tuple,
+                                         EDataType,
+                                         PassThrough,
+                                         PassThrough,
+                                         FastGelu>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
new file mode 100644
index 0000000..8986a79
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGemmSplitK<ALayout,
+                                                   BLayout,
+                                                   CLayout,
+                                                   ADataType,
+                                                   BDataType,
+                                                   CDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemmSplitK<ALayout,
+                                      BLayout,
+                                      CLayout,
+                                      ADataType,
+                                      BDataType,
+                                      CDataType,
+                                      ck::tensor_operation::element_wise::PassThrough,
+                                      ck::tensor_operation::element_wise::PassThrough,
+                                      ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
+                     is_same_v<CDataType, float>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                          is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
new file mode 100644
index 0000000..81b2b4f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
@@ -0,0 +1,90 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// conv2d backward data
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  GNHWK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  GNHWC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+template <ck::index_t NumDimSpatial,
+          typename OutLayout,
+          typename WeiLayout,
+          typename InLayout,
+          typename OutDataType,
+          typename WeiDataType,
+          typename InDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<
+        NumDimSpatial,
+        OutLayout,
+        WeiLayout,
+        Empty_Tuple,
+        InLayout,
+        OutDataType,
+        WeiDataType,
+        Empty_Tuple,
+        InDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp =
+        DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
+                                          OutLayout,
+                                          WeiLayout,
+                                          Empty_Tuple,
+                                          InLayout,
+                                          OutDataType,
+                                          WeiDataType,
+                                          Empty_Tuple,
+                                          InDataType,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
+        {
+            if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
+                         is_same_v<OutDataType, F16>)
+            {
+                add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
new file mode 100644
index 0000000..ef6920e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
@@ -0,0 +1,235 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// conv1d backward weight
+void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
+                                                           GNWC,
+                                                           GKXC,
+                                                           GNWK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
+                                                           GNWC,
+                                                           GKXC,
+                                                           GNWK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
+                                                           GNWC,
+                                                           GKXC,
+                                                           GNWK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+// conv2d backward weight
+void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           GNHWC,
+                                                           GKYXC,
+                                                           GNHWK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           GNHWC,
+                                                           GKYXC,
+                                                           GNHWK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           GNHWC,
+                                                           GKYXC,
+                                                           GNHWK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+// conv3d backward weight
+void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvBwdWeight<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGroupedConvBwdWeight<NumDimSpatial,
+                                                InLayout,
+                                                WeiLayout,
+                                                OutLayout,
+                                                InDataType,
+                                                WeiDataType,
+                                                OutDataType,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, GNWC> &&
+                     is_same_v<WeiLayout, GKXC> && is_same_v<OutLayout, GNWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
+                    op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
+                          is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
+                    op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC> &&
+                          is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, GNDHWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
new file mode 100644
index 0000000..ee38b73
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
@@ -0,0 +1,396 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// grouped conv1d forward, GNWC/GKXC/GNWK
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+// grouped conv2d forward, GNHWC/GKYXC/GNHWK
+void add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+// grouped conv2d forward, NHWGC/GKYXC/NHWGK
+void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+// grouped conv3d forward, GNDHWC/GKZYXC/GNDHWK
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    Empty_Tuple,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    Empty_Tuple,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   Empty_Tuple,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   Empty_Tuple,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, GNWC> &&
+                     is_same_v<WeiLayout, GKXC> && is_same_v<OutLayout, GNWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances(op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
+                          is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
+                add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
+                add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances(op_ptrs);
+                add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
+                          is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                // no instance
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                // no instance
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                // no instance
+            }
+        }
+        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC> &&
+                          is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, GNDHWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
new file mode 100644
index 0000000..c64598d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedGemm<
+    ALayout,
+    BLayout,
+    Empty_Tuple,
+    ELayout,
+    ADataType,
+    BDataType,
+    Empty_Tuple,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGroupedGemm<ALayout,
+                                       BLayout,
+                                       Empty_Tuple,
+                                       ELayout,
+                                       ADataType,
+                                       BDataType,
+                                       Empty_Tuple,
+                                       EDataType,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/normalization.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/normalization.hpp
new file mode 100644
index 0000000..55c67b7
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/normalization.hpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_normalization_rank_2_1_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, PassThrough, 2, 1>>>&);
+
+void add_device_normalization_rank_4_3_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, PassThrough, 4, 3>>>&);
+
+void add_device_normalization_rank_5_3_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, PassThrough, 5, 3>>>&);
+
+// FP32
+void add_device_normalization_rank_2_1_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, PassThrough, 2, 1>>>&);
+
+void add_device_normalization_rank_4_3_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&);
+
+void add_device_normalization_rank_5_3_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, PassThrough, 5, 3>>>&);
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormalization<
+    XDataType,
+    GammaDataType,
+    BetaDataType,
+    F32,
+    YDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    Rank,
+    NumReduceDim>>
+{
+    using DeviceOp = DeviceNormalization<XDataType,
+                                         GammaDataType,
+                                         BetaDataType,
+                                         F32,
+                                         YDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         Rank,
+                                         NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F16> &&
+                     is_same_v<BetaDataType, F16> && is_same_v<YDataType, F16>)
+        {
+            if constexpr(Rank == 2 && NumReduceDim == 1)
+            {
+                add_device_normalization_rank_2_1_f16_instances(op_ptrs);
+            }
+            else if constexpr(Rank == 4 && NumReduceDim == 3)
+            {
+                add_device_normalization_rank_4_3_f16_instances(op_ptrs);
+            }
+            else if constexpr(Rank == 5 && NumReduceDim == 3)
+            {
+                add_device_normalization_rank_5_3_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F32> && is_same_v<GammaDataType, F32> &&
+                          is_same_v<BetaDataType, F32> && is_same_v<YDataType, F32>)
+        {
+            if constexpr(Rank == 2 && NumReduceDim == 1)
+            {
+                add_device_normalization_rank_2_1_f32_instances(op_ptrs);
+            }
+            else if constexpr(Rank == 4 && NumReduceDim == 3)
+            {
+                add_device_normalization_rank_4_3_f32_instances(op_ptrs);
+            }
+            else if constexpr(Rank == 5 && NumReduceDim == 3)
+            {
+                add_device_normalization_rank_5_3_f32_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp
new file mode 100644
index 0000000..eda81a2
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp
@@ -0,0 +1,114 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// grouped conv2d forward, GNHWC/GKYXC/GNHWK
+void add_device_conv2d_bias_perchannel_quantization_int8_instances(
+    std::vector<
+        std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                      GNHWC,
+                                                      GKYXC,
+                                                      GK_GK_Tuple,
+                                                      GNHWK,
+                                                      int8_t,
+                                                      int8_t,
+                                                      I32_F32_Tuple,
+                                                      int8_t,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      Add_Activation_Mul2_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_bias_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_GK_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Activation_Mul2_Clamp<Relu>>>>&
+        instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    DsLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    DsDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Add_Activation_Mul2_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   DsLayout,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   DsDataType,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Add_Activation_Mul2_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_GK_Tuple> &&
+                     is_same_v<OutLayout, GNHWK>)
+        {
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<DsDataType, I32_F32_Tuple> && is_same_v<OutDataType, int8_t>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                    add_device_conv2d_bias_perchannel_quantization_int8_instances(op_ptrs);
+                else if constexpr(is_same_v<Activation, Relu>)
+                    add_device_conv2d_bias_relu_perchannel_quantization_int8_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp
new file mode 100644
index 0000000..1138402
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp
@@ -0,0 +1,114 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// grouped conv2d forward, GNHWC/GKYXC/GNHWK
+void add_device_conv2d_bias_perlayer_quantization_int8_instances(
+    std::vector<
+        std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                      GNHWC,
+                                                      GKYXC,
+                                                      GK_Tuple,
+                                                      GNHWK,
+                                                      int8_t,
+                                                      int8_t,
+                                                      I32_Tuple,
+                                                      int8_t,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      Add_Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_bias_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Activation_Mul_Clamp<Relu>>>>&
+        instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    DsLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    DsDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Add_Activation_Mul_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   DsLayout,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   DsDataType,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Add_Activation_Mul_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_Tuple> &&
+                     is_same_v<OutLayout, GNHWK>)
+        {
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<DsDataType, I32_Tuple> && is_same_v<OutDataType, int8_t>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                    add_device_conv2d_bias_perlayer_quantization_int8_instances(op_ptrs);
+                else if constexpr(is_same_v<Activation, Relu>)
+                    add_device_conv2d_bias_relu_perlayer_quantization_int8_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp
new file mode 100644
index 0000000..1a67ce5
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp
@@ -0,0 +1,113 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// grouped conv2d forward, GNHWC/GKYXC/GNHWK
+void add_device_conv2d_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul2_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul2_Clamp<Relu>>>>&
+        instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    DsLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    DsDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Activation_Mul2_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   GK_Tuple,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   F32_Tuple,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Activation_Mul2_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_Tuple> &&
+                     is_same_v<OutLayout, GNHWK>)
+        {
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<OutDataType, int8_t>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                    add_device_conv2d_perchannel_quantization_int8_instances(op_ptrs);
+                else if constexpr(is_same_v<Activation, Relu>)
+                    add_device_conv2d_relu_perchannel_quantization_int8_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp
new file mode 100644
index 0000000..410be4a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp
@@ -0,0 +1,110 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// grouped conv2d forward, GNHWC/GKYXC/GNHWK
+void add_device_conv2d_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul_Clamp<Relu>>>>&
+        instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    Empty_Tuple,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    Empty_Tuple,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Activation_Mul_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   Empty_Tuple,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   Empty_Tuple,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Activation_Mul_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
+        {
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<OutDataType, int8_t>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                    add_device_conv2d_perlayer_quantization_int8_instances(op_ptrs);
+                else if constexpr(is_same_v<Activation, Relu>)
+                    add_device_conv2d_relu_perlayer_quantization_int8_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
new file mode 100644
index 0000000..550a7b0
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp"
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
new file mode 100644
index 0000000..90cfe83
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
@@ -0,0 +1,123 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using reduce_configuration_1_instances_blockwise = std::tuple<
+    // clang-format off
+    // BlockSize | MThreadClusterSize | KThreadClusterSize
+    ReductionConfiguration_1<256, 128, 2>,
+    ReductionConfiguration_1<256, 64, 4>,
+    ReductionConfiguration_1<256, 32, 8>,
+    ReductionConfiguration_1<256, 16, 16>,
+    ReductionConfiguration_1<256, 8, 32>,
+    ReductionConfiguration_1<256, 4, 64>,
+    ReductionConfiguration_1<256, 2, 128>,
+    ReductionConfiguration_1<256, 1, 256>
+    // clang-format on
+    >;
+
+#ifdef QUICK_REDUCE_TEST
+using reduce_configuration_2_instances_blockwise = std::tuple<
+    // clang-format off
+    // InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
+    ReductionConfiguration_2<0, 2, 2, 2, 1>,
+    ReductionConfiguration_2<0, 1, 1, 2, 1>,
+    ReductionConfiguration_2<1, 2, 1, 1, 2>,
+    ReductionConfiguration_2<0, 1, 1, 3, 1>,
+    ReductionConfiguration_2<1, 1, 1, 1, 3>
+    // clang-format on
+    >;
+#else
+using reduce_configuration_2_instances_blockwise = std::tuple<
+    // clang-format off
+    // InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
+    ReductionConfiguration_2<0, 4, 4, 8, 1>,
+    ReductionConfiguration_2<0, 4, 4, 4, 1>,
+    ReductionConfiguration_2<0, 2, 2, 2, 1>,
+
+    ReductionConfiguration_2<1, 4, 1, 1, 8>,
+    ReductionConfiguration_2<1, 4, 1, 1, 4>,
+    ReductionConfiguration_2<1, 2, 1, 1, 2>,
+
+    // special instances
+    ReductionConfiguration_2<0, 1, 1, 3, 1>,
+    ReductionConfiguration_2<0, 1, 1, 5, 1>,
+    ReductionConfiguration_2<0, 1, 1, 7, 1>,
+    ReductionConfiguration_2<0, 1, 1, 11, 1>,
+
+    ReductionConfiguration_2<1, 1, 1, 1, 3>,
+    ReductionConfiguration_2<1, 1, 1, 1, 5>,
+    ReductionConfiguration_2<1, 1, 1, 1, 7>,
+    ReductionConfiguration_2<1, 1, 1, 1, 11>
+    // clang-format on
+    >;
+#endif
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          int Rank,
+          int NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          bool PropagateNan,
+          bool OutputIndex>
+void add_device_reduce_instance_blockwise(
+    std::vector<DeviceReducePtr<Rank, NumReduceDim, InElementwiseOp, AccElementwiseOp>>&
+        device_op_instances)
+{
+    static_for<0, std::tuple_size<reduce_configuration_1_instances_blockwise>::value, 1>{}(
+        [&](auto i) {
+            using cfg1 = remove_cvref_t<decltype(
+                std::get<i.value>(reduce_configuration_1_instances_blockwise{}))>;
+
+            static_for<0, std::tuple_size<reduce_configuration_2_instances_blockwise>::value, 1>{}(
+                [&](auto j) {
+                    using cfg2 = remove_cvref_t<decltype(
+                        std::get<j.value>(reduce_configuration_2_instances_blockwise{}))>;
+
+                    using ReduceOpInstance =
+                        DeviceReduceMultiBlock<InDataType,
+                                               AccDataType,
+                                               OutDataType,
+                                               Rank,
+                                               NumReduceDim,
+                                               ReduceOperation,
+                                               InElementwiseOp,
+                                               AccElementwiseOp,
+                                               InMemoryDataOperationEnum::Set,
+                                               PropagateNan,
+                                               OutputIndex,
+                                               false, // HaveIndexInputIfOutputIndex
+                                               cfg1::BlockSize_,
+                                               cfg1::MThreadClusterSize_,
+                                               cfg1::KThreadClusterSize_,
+                                               cfg2::MThreadSliceSize_,
+                                               cfg2::KThreadSliceSize_,
+                                               cfg2::InSrcVectorDim_,
+                                               cfg2::InSrcVectorSize_,
+                                               cfg2::OutDstVectorSize_>;
+
+                    device_op_instances.push_back(
+                        std::make_unique<ReduceOpInstance>(ReduceOpInstance{}));
+                });
+        });
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp
new file mode 100644
index 0000000..521d93e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp
new file mode 100644
index 0000000..fe3fd6c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp
new file mode 100644
index 0000000..52a2b69
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp
new file mode 100644
index 0000000..ee4fee4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp
new file mode 100644
index 0000000..3abdb7f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp
new file mode 100644
index 0000000..b0dbcf3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp
new file mode 100644
index 0000000..7bbf3df
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp
new file mode 100644
index 0000000..559f322
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp
new file mode 100644
index 0000000..28c9610
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp
new file mode 100644
index 0000000..5080d28
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp
new file mode 100644
index 0000000..0d24d15
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp
new file mode 100644
index 0000000..c806e80
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp
new file mode 100644
index 0000000..b7c046e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp
new file mode 100644
index 0000000..771bec1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp
new file mode 100644
index 0000000..c1fe8ad
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp
new file mode 100644
index 0000000..6bc0662
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp
new file mode 100644
index 0000000..6f80051
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp
new file mode 100644
index 0000000..c771ac4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp
new file mode 100644
index 0000000..b9ddbb9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp
new file mode 100644
index 0000000..390a719
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp
new file mode 100644
index 0000000..2a9ddbc
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp
new file mode 100644
index 0000000..5746884
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp
new file mode 100644
index 0000000..ad0f235
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp
new file mode 100644
index 0000000..c7d9527
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp
new file mode 100644
index 0000000..ec56229
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp
new file mode 100644
index 0000000..48f66da
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp
new file mode 100644
index 0000000..fabfa5b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp
new file mode 100644
index 0000000..e08faec
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp
new file mode 100644
index 0000000..a1e692a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp
new file mode 100644
index 0000000..e9654e8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp
new file mode 100644
index 0000000..7824421
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp
new file mode 100644
index 0000000..df323d4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp
new file mode 100644
index 0000000..8c08e5e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <int BlockSize, int MThreadClusterSize, int KThreadClusterSize>
+struct ReductionConfiguration_1
+{
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize, "Invalid Configuration!");
+
+    static constexpr int BlockSize_          = BlockSize;
+    static constexpr int MThreadClusterSize_ = MThreadClusterSize;
+    static constexpr int KThreadClusterSize_ = KThreadClusterSize;
+};
+
+template <int InSrcVectorDim,
+          int InSrcVectorSize,
+          int OutDstVectorSize,
+          int MThreadSliceSize,
+          int KThreadSliceSize>
+struct ReductionConfiguration_2
+{
+    static constexpr int InSrcVectorDim_   = InSrcVectorDim;
+    static constexpr int InSrcVectorSize_  = InSrcVectorSize;
+    static constexpr int OutDstVectorSize_ = OutDstVectorSize;
+    static constexpr int MThreadSliceSize_ = MThreadSliceSize;
+    static constexpr int KThreadSliceSize_ = KThreadSliceSize;
+};
+
+using ReduceAdd  = ck::reduce::Add;
+using ReduceMin  = ck::reduce::Min;
+using ReduceMax  = ck::reduce::Max;
+using ReduceAMax = ck::reduce::AMax;
+
+using UnarySquare = ck::tensor_operation::element_wise::UnarySquare;
+using UnarySqrt   = ck::tensor_operation::element_wise::UnarySqrt;
+using UnaryDivide = ck::tensor_operation::element_wise::UnaryDivide;
+using UnaryAbs    = ck::tensor_operation::element_wise::UnaryAbs;
+
+#define QUICK_REDUCE_TEST 1
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp
new file mode 100644
index 0000000..acf55d0
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp
@@ -0,0 +1,123 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using reduce_configuration_1_instances_multiblock_atomic_add = std::tuple<
+    // clang-format off
+    // BlockSize | MThreadClusterSize | KThreadClusterSize
+    ReductionConfiguration_1<256, 128, 2>,
+    ReductionConfiguration_1<256, 64, 4>,
+    ReductionConfiguration_1<256, 32, 8>,
+    ReductionConfiguration_1<256, 16, 16>,
+    ReductionConfiguration_1<256, 8, 32>,
+    ReductionConfiguration_1<256, 4, 64>,
+    ReductionConfiguration_1<256, 2, 128>,
+    ReductionConfiguration_1<256, 1, 256>
+    // clang-format on
+    >;
+
+#ifdef QUICK_REDUCE_TEST
+using reduce_configuration_2_instances_multiblock_atomic_add = std::tuple<
+    // clang-format off
+    // InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
+    ReductionConfiguration_2<0, 2, 2, 2, 1>,
+    ReductionConfiguration_2<0, 1, 1, 2, 1>,
+    ReductionConfiguration_2<1, 2, 1, 1, 2>,
+    ReductionConfiguration_2<0, 1, 1, 3, 1>,
+    ReductionConfiguration_2<1, 1, 1, 1, 3>
+    // clang-format on
+    >;
+#else
+using reduce_configuration_2_instances_multiblock_atomic_add = std::tuple<
+    // clang-format off
+    // InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
+    ReductionConfiguration_2<0, 4, 4, 8, 1>,
+    ReductionConfiguration_2<0, 4, 4, 4, 1>,
+    ReductionConfiguration_2<0, 2, 2, 2, 1>,
+
+    ReductionConfiguration_2<1, 4, 1, 1, 8>,
+    ReductionConfiguration_2<1, 4, 1, 1, 4>,
+    ReductionConfiguration_2<1, 2, 1, 1, 2>,
+
+    // special instances
+    ReductionConfiguration_2<0, 1, 1, 3, 1>,
+    ReductionConfiguration_2<0, 1, 1, 5, 1>,
+    ReductionConfiguration_2<0, 1, 1, 7, 1>,
+    ReductionConfiguration_2<0, 1, 1, 11, 1>,
+
+    ReductionConfiguration_2<1, 1, 1, 1, 3>,
+    ReductionConfiguration_2<1, 1, 1, 1, 5>,
+    ReductionConfiguration_2<1, 1, 1, 1, 7>,
+    ReductionConfiguration_2<1, 1, 1, 1, 11>
+    // clang-format on
+    >;
+#endif
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          int Rank,
+          int NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          bool PropagateNan,
+          bool OutputIndex>
+void add_device_reduce_instance_multiblock_atomic_add(
+    std::vector<DeviceReducePtr<Rank, NumReduceDim, InElementwiseOp, AccElementwiseOp>>&
+        device_op_instances)
+{
+    static_for<0,
+               std::tuple_size<reduce_configuration_1_instances_multiblock_atomic_add>::value,
+               1>{}([&](auto i) {
+        using cfg1 = remove_cvref_t<decltype(
+            std::get<i.value>(reduce_configuration_1_instances_multiblock_atomic_add{}))>;
+
+        static_for<0,
+                   std::tuple_size<reduce_configuration_2_instances_multiblock_atomic_add>::value,
+                   1>{}([&](auto j) {
+            using cfg2 = remove_cvref_t<decltype(
+                std::get<j.value>(reduce_configuration_2_instances_multiblock_atomic_add{}))>;
+
+            using ReduceOpInstance = DeviceReduceMultiBlock<InDataType,
+                                                            AccDataType,
+                                                            OutDataType,
+                                                            Rank,
+                                                            NumReduceDim,
+                                                            ReduceOperation,
+                                                            InElementwiseOp,
+                                                            AccElementwiseOp,
+                                                            InMemoryDataOperationEnum::AtomicAdd,
+                                                            PropagateNan,
+                                                            OutputIndex,
+                                                            false, // HaveIndexInputIfOutputIndex
+                                                            cfg1::BlockSize_,
+                                                            cfg1::MThreadClusterSize_,
+                                                            cfg1::KThreadClusterSize_,
+                                                            cfg2::MThreadSliceSize_,
+                                                            cfg2::KThreadSliceSize_,
+                                                            cfg2::InSrcVectorDim_,
+                                                            cfg2::InSrcVectorSize_,
+                                                            cfg2::OutDstVectorSize_>;
+
+            device_op_instances.push_back(std::make_unique<ReduceOpInstance>(ReduceOpInstance{}));
+        });
+    });
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp
new file mode 100644
index 0000000..f5102f4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp
new file mode 100644
index 0000000..ec51311
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp
new file mode 100644
index 0000000..3a3d53b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp
new file mode 100644
index 0000000..bbf4398
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp
new file mode 100644
index 0000000..55147a6
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp
new file mode 100644
index 0000000..4bff06c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp
new file mode 100644
index 0000000..daffa1a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp
new file mode 100644
index 0000000..52c4171
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp
new file mode 100644
index 0000000..2f358b0
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp
new file mode 100644
index 0000000..84c99dc
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp
new file mode 100644
index 0000000..dfcc8dd
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp
@@ -0,0 +1,100 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+#ifdef QUICK_REDUCE_TEST
+using reduce_configuration_2_instances_threadwise = std::tuple<
+    // clang-format off
+    // InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
+    ReductionConfiguration_2<0, 2, 2, 2, 1>,
+    ReductionConfiguration_2<0, 1, 1, 2, 1>,
+    ReductionConfiguration_2<1, 2, 1, 1, 2>,
+    ReductionConfiguration_2<0, 1, 1, 3, 1>,
+    ReductionConfiguration_2<1, 1, 1, 1, 3>
+    // clang-format on
+    >;
+#else
+using reduce_configuration_2_instances_threadwise = std::tuple<
+    // clang-format off
+    // InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
+    ReductionConfiguration_2<0, 4, 4, 8, 1>,
+    ReductionConfiguration_2<0, 4, 4, 4, 1>,
+    ReductionConfiguration_2<0, 2, 2, 2, 1>,
+
+    ReductionConfiguration_2<1, 4, 1, 1, 8>,
+    ReductionConfiguration_2<1, 4, 1, 1, 4>,
+    ReductionConfiguration_2<1, 2, 1, 1, 2>,
+
+    // special instances
+    ReductionConfiguration_2<0, 1, 1, 3, 1>,
+    ReductionConfiguration_2<0, 1, 1, 5, 1>,
+    ReductionConfiguration_2<0, 1, 1, 7, 1>,
+    ReductionConfiguration_2<0, 1, 1, 11, 1>,
+
+    ReductionConfiguration_2<1, 1, 1, 1, 3>,
+    ReductionConfiguration_2<1, 1, 1, 1, 5>,
+    ReductionConfiguration_2<1, 1, 1, 1, 7>,
+    ReductionConfiguration_2<1, 1, 1, 1, 11>
+    // clang-format on
+    >;
+#endif
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          int Rank,
+          int NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          bool PropagateNan,
+          bool OutputIndex>
+void add_device_reduce_instance_threadwise(
+    std::vector<DeviceReducePtr<Rank, NumReduceDim, InElementwiseOp, AccElementwiseOp>>&
+        device_op_instances)
+{
+    using cfg1 = ReductionConfiguration_1<256, 256, 1>;
+
+    static_for<0, std::tuple_size<reduce_configuration_2_instances_threadwise>::value, 1>{}(
+        [&](auto j) {
+            using cfg2 = remove_cvref_t<decltype(
+                std::get<j.value>(reduce_configuration_2_instances_threadwise{}))>;
+
+            using ReduceOpInstance = DeviceReduceThreadWise<InDataType,
+                                                            AccDataType,
+                                                            OutDataType,
+                                                            Rank,
+                                                            NumReduceDim,
+                                                            ReduceOperation,
+                                                            InElementwiseOp,
+                                                            AccElementwiseOp,
+                                                            PropagateNan,
+                                                            OutputIndex,
+                                                            false, // HaveIndexInputIfOutputIndex
+                                                            cfg1::BlockSize_,
+                                                            cfg2::MThreadSliceSize_,
+                                                            cfg2::KThreadSliceSize_,
+                                                            cfg2::InSrcVectorDim_,
+                                                            cfg2::InSrcVectorSize_,
+                                                            cfg2::OutDstVectorSize_>;
+
+            device_op_instances.push_back(std::make_unique<ReduceOpInstance>(ReduceOpInstance{}));
+        });
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp
new file mode 100644
index 0000000..4168508
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp
new file mode 100644
index 0000000..317006e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp
new file mode 100644
index 0000000..fc7718d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp
new file mode 100644
index 0000000..e661638
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp
new file mode 100644
index 0000000..a9441b8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp
new file mode 100644
index 0000000..6820ace
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp
new file mode 100644
index 0000000..ab3d4e6
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp
new file mode 100644
index 0000000..ee08c96
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp
new file mode 100644
index 0000000..1007ca2
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp
new file mode 100644
index 0000000..1d562c4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp
new file mode 100644
index 0000000..5aac638
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp
new file mode 100644
index 0000000..7a3c764
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp
new file mode 100644
index 0000000..4685d7b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp
new file mode 100644
index 0000000..1de338f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp
new file mode 100644
index 0000000..e86c41a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp
new file mode 100644
index 0000000..2ca9008
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp
new file mode 100644
index 0000000..38380e7
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp
new file mode 100644
index 0000000..04c5f3e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp
new file mode 100644
index 0000000..fef5d40
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp
new file mode 100644
index 0000000..2416f61
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp
new file mode 100644
index 0000000..fbd0285
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp
new file mode 100644
index 0000000..103b85a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp
new file mode 100644
index 0000000..e01f590
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp
new file mode 100644
index 0000000..14a7459
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp
new file mode 100644
index 0000000..7dfd806
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp
new file mode 100644
index 0000000..7670a27
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp
new file mode 100644
index 0000000..8bb85f3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp
new file mode 100644
index 0000000..a005ba8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp
new file mode 100644
index 0000000..9e8c07e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp
new file mode 100644
index 0000000..a69f88f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp
new file mode 100644
index 0000000..734b31c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp
new file mode 100644
index 0000000..237bd96
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp
new file mode 100644
index 0000000..36eb092
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <type_traits>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>&);
+void add_device_softmax_f16_f16_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>&);
+
+void add_device_softmax_f32_f32_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>&);
+void add_device_softmax_f32_f32_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>&);
+
+void add_device_softmax_i8_i8_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>&);
+void add_device_softmax_i8_i8_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>&);
+
+template <typename InDataType, typename AccDataType, typename OutDataType, index_t Rank>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::
+        DeviceSoftmax<InDataType, AccDataType, OutDataType, PassThrough, PassThrough, Rank>>
+{
+    using DeviceOp =
+        DeviceSoftmax<InDataType, AccDataType, OutDataType, PassThrough, PassThrough, Rank>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(std::is_same_v<InDataType, F16> && std::is_same_v<AccDataType, F32> &&
+                     std::is_same_v<OutDataType, F16>)
+        {
+            if constexpr(Rank == 3)
+                add_device_softmax_f16_f16_rank3_instances(op_ptrs);
+            else if constexpr(Rank == 4)
+                add_device_softmax_f16_f16_rank4_instances(op_ptrs);
+        }
+        else if constexpr(std::is_same_v<InDataType, F32> && std::is_same_v<AccDataType, F32> &&
+                          std::is_same_v<OutDataType, F32>)
+        {
+            if constexpr(Rank == 3)
+                add_device_softmax_f32_f32_rank3_instances(op_ptrs);
+            else if constexpr(Rank == 4)
+                add_device_softmax_f32_f32_rank4_instances(op_ptrs);
+        }
+        else if constexpr(std::is_same_v<InDataType, I8> && std::is_same_v<AccDataType, F32> &&
+                          std::is_same_v<OutDataType, I8>)
+        {
+            if constexpr(Rank == 3)
+                add_device_softmax_i8_i8_rank3_instances(op_ptrs);
+            else if constexpr(Rank == 4)
+                add_device_softmax_i8_i8_rank4_instances(op_ptrs);
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.hpp
new file mode 100644
index 0000000..83f52fc
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
+void add_device_softmax_f16_f16_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp
new file mode 100644
index 0000000..046ff57
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp
new file mode 100644
index 0000000..8e6a226
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp
new file mode 100644
index 0000000..518fa5f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp
new file mode 100644
index 0000000..10016cd
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp
new file mode 100644
index 0000000..cdd5a3c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp
new file mode 100644
index 0000000..a8be272
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp
new file mode 100644
index 0000000..ec8296f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp
new file mode 100644
index 0000000..b3877c4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp
@@ -0,0 +1,39 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <index_t Rank, index_t Reduce>
+using device_softmax_f16_f16_instances = std::tuple<
+    // clang-format off
+    //                InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
+    // fallback kernel
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               1,              1>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  4,                 64,                1,                8,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,                8,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               16,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               32,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,                8,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               16,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               32,              1,               8,              8>,
+    // Reduction on middle dimensions
+    // InSrcVectorDim is 0 since we want to coalesce reads on M dimension
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                4,              0,               1,              1>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                4,              0,               8,              4>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.hpp
new file mode 100644
index 0000000..a6d9a35
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
+void add_device_softmax_f32_f32_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp
new file mode 100644
index 0000000..6621a2c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp
new file mode 100644
index 0000000..3dfac98
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp
new file mode 100644
index 0000000..6d2a0c9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp
new file mode 100644
index 0000000..97dd3dc
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp
new file mode 100644
index 0000000..58f8760
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp
new file mode 100644
index 0000000..df8d31f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp
new file mode 100644
index 0000000..1bd7732
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp
new file mode 100644
index 0000000..16f129d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp
@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <index_t Rank, index_t Reduce>
+using device_softmax_f32_f32_instances = std::tuple<
+    // clang-format off
+    //                InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               1,               1>, // fallback kernel
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  4,                 64,                1,                8,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,                8,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               16,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               32,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,                8,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               16,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               32,              1,               4,               4>,
+    // Reduction on middle dimensions
+    // InSrcVectorDim is 0 since we want to coalesce reads on M dimension
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                4,              0,               1,               1>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                4,              0,               4,               4>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.hpp
new file mode 100644
index 0000000..f80f712
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
+void add_device_softmax_i8_i8_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.hpp
new file mode 100644
index 0000000..6f9952e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.hpp
new file mode 100644
index 0000000..2cbd13a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.hpp
new file mode 100644
index 0000000..7b12522
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank3_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.hpp
new file mode 100644
index 0000000..54d477f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank4_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.hpp
new file mode 100644
index 0000000..4ffc44e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.hpp
new file mode 100644
index 0000000..08cbb81
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.hpp
new file mode 100644
index 0000000..187d034
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp
new file mode 100644
index 0000000..7fc9ed6
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <index_t Rank, index_t Reduce>
+using device_softmax_i8_i8_instances = std::tuple<
+    // clang-format off
+    //                InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
+    // fallback kernel
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,               16,              1,               1,              1>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,               16,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  4,                 64,                1,               16,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               16,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               32,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               64,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               16,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               32,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               64,              1,              16,             16>,
+    // Reduction on middle dimensions
+    // InSrcVectorDim is 0 since we want to coalesce reads on M dimension
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                8,              0,               1,              1>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                 32,                  8,               32,                8,              0,              16,              8>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp
new file mode 100644
index 0000000..03be6e2
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp
@@ -0,0 +1,8 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.hpp"
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/algorithm.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/algorithm.hpp
new file mode 100644
index 0000000..86f04dd
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/algorithm.hpp
@@ -0,0 +1,43 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <iterator>
+#include <type_traits>
+#include <utility>
+
+namespace ck {
+namespace ranges {
+template <typename InputRange, typename OutputIterator>
+auto copy(InputRange&& range, OutputIterator iter)
+    -> decltype(std::copy(std::begin(std::forward<InputRange>(range)),
+                          std::end(std::forward<InputRange>(range)),
+                          iter))
+{
+    return std::copy(std::begin(std::forward<InputRange>(range)),
+                     std::end(std::forward<InputRange>(range)),
+                     iter);
+}
+
+template <typename T, typename OutputRange>
+auto fill(OutputRange&& range, const T& init)
+    -> std::void_t<decltype(std::fill(std::begin(std::forward<OutputRange>(range)),
+                                      std::end(std::forward<OutputRange>(range)),
+                                      init))>
+{
+    std::fill(std::begin(std::forward<OutputRange>(range)),
+              std::end(std::forward<OutputRange>(range)),
+              init);
+}
+
+template <typename InputRange, typename OutputIterator, typename UnaryOperation>
+auto transform(InputRange&& range, OutputIterator iter, UnaryOperation unary_op)
+    -> decltype(std::transform(std::begin(range), std::end(range), iter, unary_op))
+{
+    return std::transform(std::begin(range), std::end(range), iter, unary_op);
+}
+
+} // namespace ranges
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/check_err.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/check_err.hpp
new file mode 100644
index 0000000..a89d03d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/check_err.hpp
@@ -0,0 +1,218 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cmath>
+#include <cstdlib>
+#include <iostream>
+#include <iomanip>
+#include <iterator>
+#include <limits>
+#include <type_traits>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/host_utility/io.hpp"
+
+#include "ck/library/utility/ranges.hpp"
+
+namespace ck {
+namespace utils {
+
+template <typename Range, typename RefRange>
+typename std::enable_if<
+    std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
+        std::is_floating_point_v<ranges::range_value_t<Range>> &&
+        !std::is_same_v<ranges::range_value_t<Range>, half_t>,
+    bool>::type
+check_err(const Range& out,
+          const RefRange& ref,
+          const std::string& msg = "Error: Incorrect results!",
+          double rtol            = 1e-5,
+          double atol            = 3e-6)
+{
+    if(out.size() != ref.size())
+    {
+        std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
+                  << std::endl;
+        return false;
+    }
+
+    bool res{true};
+    int err_count  = 0;
+    double err     = 0;
+    double max_err = std::numeric_limits<double>::min();
+    for(std::size_t i = 0; i < ref.size(); ++i)
+    {
+        const double o = *std::next(std::begin(out), i);
+        const double r = *std::next(std::begin(ref), i);
+        err            = std::abs(o - r);
+        if(err > atol + rtol * std::abs(r) || !std::isfinite(o) || !std::isfinite(r))
+        {
+            max_err = err > max_err ? err : max_err;
+            err_count++;
+            if(err_count < 5)
+            {
+                std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
+                          << "] != ref[" << i << "]: " << o << " != " << r << std::endl;
+            }
+            res = false;
+        }
+    }
+    if(!res)
+    {
+        std::cerr << std::setw(12) << std::setprecision(7) << "max err: " << max_err << std::endl;
+    }
+    return res;
+}
+
+template <typename Range, typename RefRange>
+typename std::enable_if<
+    std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
+        std::is_same_v<ranges::range_value_t<Range>, bhalf_t>,
+    bool>::type
+check_err(const Range& out,
+          const RefRange& ref,
+          const std::string& msg = "Error: Incorrect results!",
+          double rtol            = 1e-3,
+          double atol            = 1e-3)
+{
+    if(out.size() != ref.size())
+    {
+        std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
+                  << std::endl;
+        return false;
+    }
+
+    bool res{true};
+    int err_count = 0;
+    double err    = 0;
+    // TODO: This is a hack. We should have proper specialization for bhalf_t data type.
+    double max_err = std::numeric_limits<float>::min();
+    for(std::size_t i = 0; i < ref.size(); ++i)
+    {
+        const double o = type_convert<float>(*std::next(std::begin(out), i));
+        const double r = type_convert<float>(*std::next(std::begin(ref), i));
+        err            = std::abs(o - r);
+        if(err > atol + rtol * std::abs(r) || !std::isfinite(o) || !std::isfinite(r))
+        {
+            max_err = err > max_err ? err : max_err;
+            err_count++;
+            if(err_count < 5)
+            {
+                std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
+                          << "] != ref[" << i << "]: " << o << " != " << r << std::endl;
+            }
+            res = false;
+        }
+    }
+    if(!res)
+    {
+        std::cerr << std::setw(12) << std::setprecision(7) << "max err: " << max_err << std::endl;
+    }
+    return res;
+}
+
+template <typename Range, typename RefRange>
+typename std::enable_if<
+    std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
+        std::is_same_v<ranges::range_value_t<Range>, half_t>,
+    bool>::type
+check_err(const Range& out,
+          const RefRange& ref,
+          const std::string& msg = "Error: Incorrect results!",
+          double rtol            = 1e-3,
+          double atol            = 1e-3)
+{
+    if(out.size() != ref.size())
+    {
+        std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
+                  << std::endl;
+        return false;
+    }
+
+    bool res{true};
+    int err_count  = 0;
+    double err     = 0;
+    double max_err = std::numeric_limits<ranges::range_value_t<Range>>::min();
+    for(std::size_t i = 0; i < ref.size(); ++i)
+    {
+        const double o = type_convert<float>(*std::next(std::begin(out), i));
+        const double r = type_convert<float>(*std::next(std::begin(ref), i));
+        err            = std::abs(o - r);
+        if(err > atol + rtol * std::abs(r) || !std::isfinite(o) || !std::isfinite(r))
+        {
+            max_err = err > max_err ? err : max_err;
+            err_count++;
+            if(err_count < 5)
+            {
+                std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
+                          << "] != ref[" << i << "]: " << o << " != " << r << std::endl;
+            }
+            res = false;
+        }
+    }
+    if(!res)
+    {
+        std::cerr << std::setw(12) << std::setprecision(7) << "max err: " << max_err << std::endl;
+    }
+    return res;
+}
+
+template <typename Range, typename RefRange>
+std::enable_if_t<(std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
+                  std::is_integral_v<ranges::range_value_t<Range>> &&
+                  !std::is_same_v<ranges::range_value_t<Range>, bhalf_t>)
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+                     || std::is_same_v<ranges::range_value_t<Range>, int4_t>
+#endif
+                 ,
+                 bool>
+check_err(const Range& out,
+          const RefRange& ref,
+          const std::string& msg = "Error: Incorrect results!",
+          double                 = 0,
+          double atol            = 0)
+{
+    if(out.size() != ref.size())
+    {
+        std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
+                  << std::endl;
+        return false;
+    }
+
+    bool res{true};
+    int err_count   = 0;
+    int64_t err     = 0;
+    int64_t max_err = std::numeric_limits<int64_t>::min();
+    for(std::size_t i = 0; i < ref.size(); ++i)
+    {
+        const int64_t o = *std::next(std::begin(out), i);
+        const int64_t r = *std::next(std::begin(ref), i);
+        err             = std::abs(o - r);
+
+        if(err > atol)
+        {
+            max_err = err > max_err ? err : max_err;
+            err_count++;
+            if(err_count < 5)
+            {
+                std::cerr << msg << " out[" << i << "] != ref[" << i << "]: " << o << " != " << r
+                          << std::endl;
+            }
+            res = false;
+        }
+    }
+    if(!res)
+    {
+        std::cerr << "max err: " << max_err << std::endl;
+    }
+    return res;
+}
+
+} // namespace utils
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/conv_common.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/conv_common.hpp
new file mode 100644
index 0000000..6fad9f7
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/conv_common.hpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+
+template <typename... InDesc,
+          typename... WeiDesc,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename LeftPads,
+          typename RightPads>
+constexpr auto get_convolution_output_default_4d_tensor_descriptor(
+    const ck::TensorDescriptor<InDesc...>& in_desc,
+    const ck::TensorDescriptor<WeiDesc...>& wei_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations conv_dilations,
+    const LeftPads& left_pads,
+    const RightPads& right_pads)
+{
+    using namespace ck;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    assert(in_desc.GetNumOfDimension() == 4);
+    assert(wei_desc.GetNumOfDimension() == 4);
+    assert(in_desc.GetLength(I1) == wei_desc.GetLength(I1));
+
+    const auto N  = in_desc.GetLength(I0);
+    const auto Hi = in_desc.GetLength(I2);
+    const auto Wi = in_desc.GetLength(I3);
+
+    const auto K = wei_desc.GetLength(I0);
+    const auto Y = wei_desc.GetLength(I2);
+    const auto X = wei_desc.GetLength(I3);
+
+    const auto LeftPadH = left_pads[I0];
+    const auto LeftPadW = left_pads[I1];
+
+    const auto RightPadH = right_pads[I0];
+    const auto RightPadW = right_pads[I1];
+
+    const auto YEff = (Y - I1) * conv_dilations[I0] + I1;
+    const auto XEff = (X - I1) * conv_dilations[I1] + I1;
+
+    const auto Ho = (Hi + LeftPadH + RightPadH - YEff) / conv_strides[I0] + I1;
+    const auto Wo = (Wi + LeftPadW + RightPadW - XEff) / conv_strides[I1] + I1;
+
+    return make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho, Wo));
+}
+
+template <class InDesc, class WeiDesc, class OutDesc>
+constexpr std::size_t
+calculate_convolution_flops(const InDesc&, const WeiDesc& wei_desc, const OutDesc& out_desc)
+{
+    using namespace ck;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    const index_t N  = out_desc.GetLength(I0);
+    const index_t K  = out_desc.GetLength(I1);
+    const index_t Ho = out_desc.GetLength(I2);
+    const index_t Wo = out_desc.GetLength(I3);
+
+    const index_t C = wei_desc.GetLength(I1);
+    const index_t Y = wei_desc.GetLength(I2);
+    const index_t X = wei_desc.GetLength(I3);
+
+    return std::size_t(2) * N * K * Ho * Wo * C * Y * X;
+}
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp
new file mode 100644
index 0000000..2b4f63b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp
@@ -0,0 +1,355 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace utils {
+namespace conv {
+
+namespace detail {
+
+template <typename OldLayout>
+std::vector<std::size_t> get_layout_transpose_gnchw_to_old()
+{
+    // HACK: NHWC/KYXC/NHWK, which is treated as GNHWC/GKYXC/GNHWK by this function,
+    // is used by some legacy kernel. New kernel should use GNHWK/GKYXC/GNHWK
+    // TODO: remove this branch after removing legacy kernel
+    if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NWC> ||
+                 ck::is_same_v<OldLayout, ck::tensor_layout::convolution::KXC> ||
+                 ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NWK>)
+    {
+        return {0, 1, 3, 2};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NHWC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::KYXC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NHWK>)
+    {
+        return {0, 1, 4, 2, 3};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NDHWC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::KZYXC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NDHWK>)
+    {
+        return {0, 1, 5, 2, 3, 4};
+    }
+    // separate from legacy code above
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNCW> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GKCX> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNKW>)
+    {
+        return {0, 1, 2, 3};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNCHW> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GKCYX> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNKHW>)
+    {
+        return {0, 1, 2, 3, 4};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNCDHW> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GKCZYX> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNKDHW>)
+    {
+        return {0, 1, 2, 3, 4, 5};
+    }
+    if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNWC> ||
+                 ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GKXC> ||
+                 ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNWK>)
+    {
+        return {0, 1, 3, 2};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNHWC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GKYXC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNHWK>)
+    {
+        return {0, 1, 4, 2, 3};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNDHWC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GKZYXC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNDHWK>)
+    {
+        return {0, 1, 5, 2, 3, 4};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NWGC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::KXGC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NWGK>)
+    {
+        return {2, 0, 3, 1};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NHWGC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::KYXGC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NHWGK>)
+    {
+        return {3, 0, 4, 1, 2};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NDHWGC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::KZYXGC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NDHWGK>)
+    {
+        return {4, 0, 5, 1, 2, 3};
+    }
+    else
+    {
+        printf("%s\n", __func__);
+        throw std::runtime_error("wrong! unsupported layout");
+    }
+}
+
+} // namespace detail
+
+// make tensor descriptor for packed input tensor, and order the dimension in the order of GNCHW
+// regardless of physical layout
+template <typename InLayout>
+HostTensorDescriptor
+make_input_host_tensor_descriptor_g_n_c_wis_packed(const ck::utils::conv::ConvParam& param)
+{
+    std::vector<std::size_t> physical_lengths;
+
+    // HACK: NHWC/KYXC/NHWK, which is treated as GNHWC/GKYXC/GNHWK by this function,
+    // is used by some legacy kernel. New kernel should use GNHWK/GKYXC/GNHWK
+    // TODO: remove this branch after removing legacy kernel
+    if constexpr(ck::is_same_v<InLayout, ck::tensor_layout::convolution::NWC> ||
+                 ck::is_same_v<InLayout, ck::tensor_layout::convolution::NHWC> ||
+                 ck::is_same_v<InLayout, ck::tensor_layout::convolution::NDHWC>)
+    {
+        if(param.G_ != 1)
+        {
+            throw std::runtime_error("wrong! G != 1");
+        }
+
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.input_spatial_lengths_.begin(),
+                                param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    // separate from legacy code above
+    else if constexpr(ck::is_same_v<InLayout, ck::tensor_layout::convolution::GNCW> ||
+                      ck::is_same_v<InLayout, ck::tensor_layout::convolution::GNCHW> ||
+                      ck::is_same_v<InLayout, ck::tensor_layout::convolution::GNCDHW>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.end(),
+                                param.input_spatial_lengths_.begin(),
+                                param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<InLayout, ck::tensor_layout::convolution::GNWC> ||
+                      ck::is_same_v<InLayout, ck::tensor_layout::convolution::GNHWC> ||
+                      ck::is_same_v<InLayout, ck::tensor_layout::convolution::GNDHWC>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.input_spatial_lengths_.begin(),
+                                param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<InLayout, ck::tensor_layout::convolution::NWGC> ||
+                      ck::is_same_v<InLayout, ck::tensor_layout::convolution::NHWGC> ||
+                      ck::is_same_v<InLayout, ck::tensor_layout::convolution::NDHWGC>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 1,
+                                param.input_spatial_lengths_.begin(),
+                                param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else
+    {
+        printf("%s\n", __func__);
+        printf("%s\n", InLayout::name);
+        throw std::runtime_error("wrong! unsupported layout");
+    }
+
+    return transpose_host_tensor_descriptor_given_new2old(
+        HostTensorDescriptor(physical_lengths),
+        detail::get_layout_transpose_gnchw_to_old<InLayout>());
+}
+
+// make tensor descriptor for packed weight tensor, and order the dimension in the order of GKCYX
+// regardless of physical layout
+template <typename WeiLayout>
+HostTensorDescriptor
+make_weight_host_tensor_descriptor_g_k_c_xs_packed(const ck::utils::conv::ConvParam& param)
+{
+    std::vector<std::size_t> physical_lengths;
+
+    // HACK: NHWC/KYXC/NHWK, which is treated as GNHWC/GKYXC/GNHWK by this function,
+    // is used by some legacy kernel. New kernel should use GNHWK/GKYXC/GNHWK
+    // TODO: remove this branch after removing legacy kernel
+    if constexpr(ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KXC> ||
+                 ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KYXC> ||
+                 ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KZYXC>)
+    {
+        if(param.G_ != 1)
+        {
+            throw std::runtime_error("wrong! G != 1");
+        }
+
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    // separate from legacy code above
+    else if constexpr(ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KXC> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KYXC> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KZYXC>)
+    {
+        if(param.G_ != 1)
+        {
+            throw std::runtime_error("wrong! G != 1");
+        }
+
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.end(),
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::GKCX> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::GKCYX> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::GKCZYX>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.end(),
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::GKXC> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::GKYXC> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::GKZYXC>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KXGC> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KYXGC> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KZYXGC>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 1,
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else
+    {
+        printf("%s\n", __func__);
+        printf("%s\n", WeiLayout::name);
+        throw std::runtime_error("wrong! unsupported layout");
+    }
+
+    return transpose_host_tensor_descriptor_given_new2old(
+        HostTensorDescriptor(physical_lengths),
+        detail::get_layout_transpose_gnchw_to_old<WeiLayout>());
+}
+
+// make tensor descriptor for packed output tensor, and order the dimension in the order of GNKHW
+// regardless of physical layout
+template <typename OutLayout>
+HostTensorDescriptor
+make_output_host_tensor_descriptor_g_n_k_wos_packed(const ck::utils::conv::ConvParam& param)
+{
+    std::vector<std::size_t> physical_lengths;
+
+    // HACK: NHWC/KYXC/NHWK, which is treated as GNHWC/GKYXC/GNHWK by this function,
+    // is used by some legacy kernel. New kernel should use GNHWK/GKYXC/GNHWK
+    // TODO: remove this branch after removing legacy kernel
+    if constexpr(ck::is_same_v<OutLayout, ck::tensor_layout::convolution::NWK> ||
+                 ck::is_same_v<OutLayout, ck::tensor_layout::convolution::NHWK> ||
+                 ck::is_same_v<OutLayout, ck::tensor_layout::convolution::NDHWK>)
+    {
+        if(param.G_ != 1)
+        {
+            throw std::runtime_error("wrong! G != 1");
+        }
+
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.K_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.output_spatial_lengths_.begin(),
+                                param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    // separate from legacy code above
+    else if constexpr(ck::is_same_v<OutLayout, ck::tensor_layout::convolution::GNKW> ||
+                      ck::is_same_v<OutLayout, ck::tensor_layout::convolution::GNKHW> ||
+                      ck::is_same_v<OutLayout, ck::tensor_layout::convolution::GNKDHW>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.K_)};
+
+        physical_lengths.insert(physical_lengths.end(),
+                                param.output_spatial_lengths_.begin(),
+                                param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<OutLayout, ck::tensor_layout::convolution::GNWK> ||
+                      ck::is_same_v<OutLayout, ck::tensor_layout::convolution::GNHWK> ||
+                      ck::is_same_v<OutLayout, ck::tensor_layout::convolution::GNDHWK>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.K_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.output_spatial_lengths_.begin(),
+                                param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<OutLayout, ck::tensor_layout::convolution::NWGK> ||
+                      ck::is_same_v<OutLayout, ck::tensor_layout::convolution::NHWGK> ||
+                      ck::is_same_v<OutLayout, ck::tensor_layout::convolution::NDHWGK>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.K_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 1,
+                                param.output_spatial_lengths_.begin(),
+                                param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else
+    {
+        printf("%s\n", __func__);
+        printf("%s\n", OutLayout::name);
+        throw std::runtime_error("wrong! unsupported layout");
+    }
+
+    return transpose_host_tensor_descriptor_given_new2old(
+        HostTensorDescriptor(physical_lengths),
+        detail::get_layout_transpose_gnchw_to_old<OutLayout>());
+}
+
+} // namespace conv
+} // namespace utils
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/convolution_parameter.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/convolution_parameter.hpp
new file mode 100644
index 0000000..f4a2b56
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/convolution_parameter.hpp
@@ -0,0 +1,101 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <numeric>
+#include <iterator>
+#include <vector>
+
+#include "ck/ck.hpp"
+
+#include "ck/library/utility/numeric.hpp"
+
+namespace ck {
+namespace utils {
+namespace conv {
+
+struct ConvParam
+{
+    ConvParam();
+    ConvParam(ck::index_t n_dim,
+              ck::index_t group_count,
+              ck::index_t n_batch,
+              ck::index_t n_out_channels,
+              ck::index_t n_in_channels,
+              const std::vector<ck::index_t>& filters_len,
+              const std::vector<ck::index_t>& input_len,
+              const std::vector<ck::index_t>& strides,
+              const std::vector<ck::index_t>& dilations,
+              const std::vector<ck::index_t>& left_pads,
+              const std::vector<ck::index_t>& right_pads);
+
+    ck::index_t num_dim_spatial_;
+    ck::index_t G_;
+    ck::index_t N_;
+    ck::index_t K_;
+    ck::index_t C_;
+
+    std::vector<ck::index_t> filter_spatial_lengths_;
+    std::vector<ck::index_t> input_spatial_lengths_;
+    std::vector<ck::index_t> output_spatial_lengths_;
+
+    std::vector<ck::index_t> conv_filter_strides_;
+    std::vector<ck::index_t> conv_filter_dilations_;
+
+    std::vector<ck::index_t> input_left_pads_;
+    std::vector<ck::index_t> input_right_pads_;
+
+    std::vector<ck::index_t> GetOutputSpatialLengths() const;
+
+    std::size_t GetFlops() const;
+
+    template <typename InDataType>
+    std::size_t GetInputByte() const
+    {
+        // sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
+        return sizeof(InDataType) *
+               (G_ * N_ * C_ *
+                ck::accumulate_n<std::size_t>(
+                    std::begin(input_spatial_lengths_), num_dim_spatial_, 1, std::multiplies<>()));
+    }
+
+    template <typename WeiDataType>
+    std::size_t GetWeightByte() const
+    {
+        // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
+        return sizeof(WeiDataType) *
+               (G_ * K_ * C_ *
+                ck::accumulate_n<std::size_t>(
+                    std::begin(filter_spatial_lengths_), num_dim_spatial_, 1, std::multiplies<>()));
+    }
+
+    template <typename OutDataType>
+    std::size_t GetOutputByte() const
+    {
+        // sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
+        return sizeof(OutDataType) * (G_ * N_ * K_ *
+                                      std::accumulate(std::begin(output_spatial_lengths_),
+                                                      std::end(output_spatial_lengths_),
+                                                      static_cast<std::size_t>(1),
+                                                      std::multiplies<std::size_t>()));
+    }
+
+    template <typename InDataType, typename WeiDataType, typename OutDataType>
+    std::size_t GetByte() const
+    {
+        return GetInputByte<InDataType>() + GetWeightByte<WeiDataType>() +
+               GetOutputByte<OutDataType>();
+    }
+};
+
+std::string get_conv_param_parser_helper_msg();
+
+ConvParam parse_conv_param(int num_dim_spatial, int arg_idx, char* const argv[]);
+
+} // namespace conv
+} // namespace utils
+} // namespace ck
+
+std::ostream& operator<<(std::ostream& os, const ck::utils::conv::ConvParam& p);
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/device_memory.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/device_memory.hpp
new file mode 100644
index 0000000..3c4ece4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/device_memory.hpp
@@ -0,0 +1,43 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+
+template <typename T>
+__global__ void set_buffer_value(T* p, T x, uint64_t buffer_element_size)
+{
+    for(uint64_t i = threadIdx.x; i < buffer_element_size; i += blockDim.x)
+    {
+        p[i] = x;
+    }
+}
+
+struct DeviceMem
+{
+    DeviceMem() = delete;
+    DeviceMem(std::size_t mem_size);
+    void* GetDeviceBuffer() const;
+    std::size_t GetBufferSize() const;
+    void ToDevice(const void* p) const;
+    void FromDevice(void* p) const;
+    void SetZero() const;
+    template <typename T>
+    void SetValue(T x) const;
+    ~DeviceMem();
+
+    void* mpDeviceBuf;
+    std::size_t mMemSize;
+};
+
+template <typename T>
+void DeviceMem::SetValue(T x) const
+{
+    if(mMemSize % sizeof(T) != 0)
+    {
+        throw std::runtime_error("wrong! not entire DeviceMem will be set");
+    }
+
+    set_buffer_value<T><<<1, 1024>>>(static_cast<T*>(mpDeviceBuf), x, mMemSize / sizeof(T));
+}
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/fill.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/fill.hpp
new file mode 100644
index 0000000..54d58f3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/fill.hpp
@@ -0,0 +1,118 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cmath>
+#include <iterator>
+#include <random>
+#include <type_traits>
+#include <utility>
+
+#include "ck/utility/data_type.hpp"
+
+namespace ck {
+namespace utils {
+
+template <typename T>
+struct FillUniformDistribution
+{
+    float a_{-5.f};
+    float b_{5.f};
+
+    template <typename ForwardIter>
+    void operator()(ForwardIter first, ForwardIter last) const
+    {
+        std::mt19937 gen(11939);
+        std::uniform_real_distribution<float> dis(a_, b_);
+        std::generate(first, last, [&dis, &gen]() { return ck::type_convert<T>(dis(gen)); });
+    }
+
+    template <typename ForwardRange>
+    auto operator()(ForwardRange&& range) const
+        -> std::void_t<decltype(std::declval<const FillUniformDistribution&>()(
+            std::begin(std::forward<ForwardRange>(range)),
+            std::end(std::forward<ForwardRange>(range))))>
+    {
+        (*this)(std::begin(std::forward<ForwardRange>(range)),
+                std::end(std::forward<ForwardRange>(range)));
+    }
+};
+
+// Normally FillUniformDistributionIntegerValue should use std::uniform_int_distribution as below.
+// However this produces segfaults in std::mt19937 which look like inifite loop.
+//      template <typename T>
+//      struct FillUniformDistributionIntegerValue
+//      {
+//          int a_{-5};
+//          int b_{5};
+//
+//          template <typename ForwardIter>
+//          void operator()(ForwardIter first, ForwardIter last) const
+//          {
+//              std::mt19937 gen(11939);
+//              std::uniform_int_distribution<int> dis(a_, b_);
+//              std::generate(
+//                  first, last, [&dis, &gen]() { return ck::type_convert<T>(dis(gen)); });
+//          }
+//      };
+
+// Workaround for uniform_int_distribution not working as expected. See note above.<
+template <typename T>
+struct FillUniformDistributionIntegerValue
+{
+    float a_{-5.f};
+    float b_{5.f};
+
+    template <typename ForwardIter>
+    void operator()(ForwardIter first, ForwardIter last) const
+    {
+        std::mt19937 gen(11939);
+        std::uniform_real_distribution<float> dis(a_, b_);
+        std::generate(
+            first, last, [&dis, &gen]() { return ck::type_convert<T>(std::round(dis(gen))); });
+    }
+
+    template <typename ForwardRange>
+    auto operator()(ForwardRange&& range) const
+        -> std::void_t<decltype(std::declval<const FillUniformDistributionIntegerValue&>()(
+            std::begin(std::forward<ForwardRange>(range)),
+            std::end(std::forward<ForwardRange>(range))))>
+    {
+        (*this)(std::begin(std::forward<ForwardRange>(range)),
+                std::end(std::forward<ForwardRange>(range)));
+    }
+};
+
+template <typename T>
+struct FillMonotonicSeq
+{
+    T init_value_{0};
+    T step_{1};
+
+    template <typename ForwardIter>
+    void operator()(ForwardIter first, ForwardIter last) const
+    {
+        std::generate(first, last, [=, n = init_value_]() mutable {
+            auto tmp = n;
+            n += step_;
+            return tmp;
+        });
+    }
+};
+
+template <typename T>
+struct FillConstant
+{
+    T value_{0};
+
+    template <typename ForwardIter>
+    void operator()(ForwardIter first, ForwardIter last) const
+    {
+        std::fill(first, last, value_);
+    }
+};
+
+} // namespace utils
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/host_common_util.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/host_common_util.hpp
new file mode 100644
index 0000000..6f4466e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/host_common_util.hpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <array>
+#include <iostream>
+#include <fstream>
+#include <string>
+#include <algorithm>
+
+#include "ck/ck.hpp"
+
+namespace ck {
+
+namespace host_common {
+
+template <typename T>
+static inline void dumpBufferToFile(const char* fileName, T* data, size_t dataNumItems)
+{
+    std::ofstream outFile(fileName, std::ios::binary);
+    if(outFile)
+    {
+        outFile.write(reinterpret_cast<char*>(data), dataNumItems * sizeof(T));
+        outFile.close();
+        std::cout << "Write output to file " << fileName << std::endl;
+    }
+    else
+    {
+        std::cout << "Could not open file " << fileName << " for writing" << std::endl;
+    }
+};
+
+template <typename T>
+static inline T getSingleValueFromString(const std::string& valueStr)
+{
+    std::istringstream iss(valueStr);
+
+    T val;
+
+    iss >> val;
+
+    return (val);
+};
+
+template <typename T>
+static inline std::vector<T> getTypeValuesFromString(const char* cstr_values)
+{
+    std::string valuesStr(cstr_values);
+
+    std::vector<T> values;
+    std::size_t pos = 0;
+    std::size_t new_pos;
+
+    new_pos = valuesStr.find(',', pos);
+    while(new_pos != std::string::npos)
+    {
+        const std::string sliceStr = valuesStr.substr(pos, new_pos - pos);
+
+        T val = getSingleValueFromString<T>(sliceStr);
+
+        values.push_back(val);
+
+        pos     = new_pos + 1;
+        new_pos = valuesStr.find(',', pos);
+    };
+
+    std::string sliceStr = valuesStr.substr(pos);
+    T val                = getSingleValueFromString<T>(sliceStr);
+
+    values.push_back(val);
+
+    return (values);
+}
+
+template <int NDim>
+static inline std::vector<std::array<index_t, NDim>>
+get_index_set(const std::array<index_t, NDim>& dim_lengths)
+{
+    static_assert(NDim >= 1, "NDim >= 1 is required to use this function!");
+
+    if constexpr(NDim == 1)
+    {
+        std::vector<std::array<index_t, NDim>> index_set;
+
+        for(int i = 0; i < dim_lengths[0]; i++)
+        {
+            std::array<index_t, 1> index{i};
+
+            index_set.push_back(index);
+        };
+
+        return index_set;
+    }
+    else
+    {
+        std::vector<std::array<index_t, NDim>> index_set;
+        std::array<index_t, NDim - 1> partial_dim_lengths;
+
+        std::copy(dim_lengths.begin() + 1, dim_lengths.end(), partial_dim_lengths.begin());
+
+        std::vector<std::array<index_t, NDim - 1>> partial_index_set;
+
+        partial_index_set = get_index_set<NDim - 1>(partial_dim_lengths);
+
+        for(index_t i = 0; i < dim_lengths[0]; i++)
+            for(const auto& partial_index : partial_index_set)
+            {
+                std::array<index_t, NDim> index;
+
+                index[0] = i;
+
+                std::copy(partial_index.begin(), partial_index.end(), index.begin() + 1);
+
+                index_set.push_back(index);
+            };
+
+        return index_set;
+    };
+};
+
+template <int NDim>
+static inline size_t get_offset_from_index(const std::array<index_t, NDim>& strides,
+                                           const std::array<index_t, NDim>& index)
+{
+    size_t offset = 0;
+
+    for(int i = 0; i < NDim; i++)
+        offset += index[i] * strides[i];
+
+    return (offset);
+};
+
+} // namespace host_common
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/host_conv.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/host_conv.hpp
new file mode 100644
index 0000000..8348a30
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/host_conv.hpp
@@ -0,0 +1,152 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "host_tensor.hpp"
+#include "conv_common.hpp"
+
+template <typename TIn,
+          typename TWei,
+          typename TOut,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads>
+void host_conv_nchw_kcyx_nkhw(const Tensor<TIn>& in,
+                              const Tensor<TWei>& wei,
+                              Tensor<TOut>& out,
+                              const ConvStrides& conv_strides,
+                              const ConvDilations& conv_dilations,
+                              const InLeftPads& in_left_pads,
+                              const InRightPads&)
+{
+    constexpr auto I0 = ck::Number<0>{};
+    constexpr auto I1 = ck::Number<1>{};
+
+    auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
+        float v = 0;
+        for(int c = 0; c < wei.mDesc.GetLengths()[1]; ++c)
+        {
+            for(int y = 0; y < wei.mDesc.GetLengths()[2]; ++y)
+            {
+                int hi = ho * conv_strides[I0] + y * conv_dilations[I0] - in_left_pads[I0];
+                for(int x = 0; x < wei.mDesc.GetLengths()[3]; ++x)
+                {
+                    int wi = wo * conv_strides[I1] + x * conv_dilations[I1] - in_left_pads[I1];
+                    if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
+                       wi < in.mDesc.GetLengths()[3])
+                    {
+                        v += ck::type_convert<float>(in(n, c, hi, wi)) *
+                             ck::type_convert<float>(wei(k, c, y, x));
+                    }
+                }
+            }
+        }
+        out(n, k, ho, wo) = ck::type_convert<TOut>(v);
+    };
+
+    make_ParallelTensorFunctor(f_nchw,
+                               out.mDesc.GetLengths()[0],
+                               out.mDesc.GetLengths()[1],
+                               out.mDesc.GetLengths()[2],
+                               out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
+}
+
+template <typename TIn,
+          typename TWei,
+          typename TOut,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads>
+void host_conv3d_ndhwc_kzyxc_ndhwk(const Tensor<TIn>& in,
+                                   const Tensor<TWei>& wei,
+                                   Tensor<TOut>& out,
+                                   const ConvStrides& conv_strides,
+                                   const ConvDilations& conv_dilations,
+                                   const InLeftPads& in_left_pads,
+                                   const InRightPads&)
+{
+    using namespace ck;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    const auto Di     = in.mDesc.GetLengths()[1];
+    const auto Hi     = in.mDesc.GetLengths()[2];
+    const auto Wi     = in.mDesc.GetLengths()[3];
+    const auto Z      = wei.mDesc.GetLengths()[1];
+    const auto Y      = wei.mDesc.GetLengths()[2];
+    const auto X      = wei.mDesc.GetLengths()[3];
+    const auto C      = wei.mDesc.GetLengths()[4];
+
+    auto f_ndhwc = [&](auto n, auto do_tmp, auto ho_tmp, auto wo_tmp, auto k) {
+        // do__ must be converted to signed integer, otherwise zmin might be wrong in cases
+        // negative values.
+        const int do_ = static_cast<int>(do_tmp);
+        const int ho  = static_cast<int>(ho_tmp);
+        const int wo  = static_cast<int>(wo_tmp);
+        const int zmin =
+            std::max(0,
+                     (in_left_pads[I0] - do_ * conv_strides[I0] + conv_dilations[I0] - 1) /
+                         conv_dilations[I0]);
+        const int ymin =
+            std::max(0,
+                     (in_left_pads[I1] - ho * conv_strides[I1] + conv_dilations[I1] - 1) /
+                         conv_dilations[I1]);
+        const int xmin =
+            std::max(0,
+                     (in_left_pads[I2] - wo * conv_strides[I2] + conv_dilations[I2] - 1) /
+                         conv_dilations[I2]);
+        const int zmax =
+            std::min(Z, (in_left_pads[I0] - do_ * conv_strides[I0] + Di) / conv_dilations[I0]);
+        const int ymax =
+            std::min(Y, (in_left_pads[I1] - ho * conv_strides[I1] + Hi) / conv_dilations[I1]);
+        const int xmax =
+            std::min(X, (in_left_pads[I2] - wo * conv_strides[I2] + Wi) / conv_dilations[I2]);
+        const int di_min = do_ * conv_strides[I0] + zmin * conv_dilations[I0] - in_left_pads[I0];
+        const int hi_min = ho * conv_strides[I1] + ymin * conv_dilations[I1] - in_left_pads[I1];
+        const int wi_min = wo * conv_strides[I2] + xmin * conv_dilations[I2] - in_left_pads[I2];
+
+        double v = 0;
+
+        const TIn* in_n   = in.mData.data() + n * Di * Hi * Wi * C;
+        const TWei* wei_k = wei.mData.data() + k * Z * Y * X * C;
+
+        int di = di_min;
+        for(int z = zmin; z < zmax; ++z, di += conv_dilations[I0])
+        {
+            const TIn* in_n_di  = in_n + di * Hi * Wi * C;
+            const TWei* wei_k_z = wei_k + z * Y * X * C;
+            int hi              = hi_min;
+
+            for(int y = ymin; y < ymax; ++y, hi += conv_dilations[I1])
+            {
+                const TIn* in_n_di_hi = in_n_di + hi * Wi * C;
+                const TWei* wei_k_z_y = wei_k_z + y * X * C;
+                int wi                = wi_min;
+
+                for(int x = xmin; x < xmax; ++x, wi += conv_dilations[I2])
+                {
+                    const TIn* in_n_di_hi_wi = in_n_di_hi + wi * C;
+                    const TWei* wei_k_z_y_x  = wei_k_z_y + x * C;
+
+                    for(int c = 0; c < C; ++c)
+                    {
+                        v += static_cast<const double>(in_n_di_hi_wi[c]) *
+                             static_cast<const double>(wei_k_z_y_x[c]);
+                    }
+                }
+            }
+        }
+
+        out(n, do_, ho, wo, k) = v;
+    };
+
+    make_ParallelTensorFunctor(f_ndhwc,
+                               out.mDesc.GetLengths()[0],
+                               out.mDesc.GetLengths()[1],
+                               out.mDesc.GetLengths()[2],
+                               out.mDesc.GetLengths()[3],
+                               out.mDesc.GetLengths()[4])(std::thread::hardware_concurrency() - 4);
+}
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/host_gemm.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/host_gemm.hpp
new file mode 100644
index 0000000..44036d0
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/host_gemm.hpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "host_tensor.hpp"
+
+template <typename AType,
+          typename BType,
+          typename CType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+void host_gemm_mk_kn_mn(const Tensor<AType>& a_m_k,
+                        const Tensor<BType>& b_k_n,
+                        Tensor<CType>& c_m_n,
+                        const AElementwiseOperation& a_element_op,
+                        const BElementwiseOperation& b_element_op,
+                        const CElementwiseOperation& c_element_op)
+{
+    auto f_mk_kn_mn = [&](auto m, auto n) {
+        const int K = a_m_k.mDesc.GetLengths()[1];
+
+        float v_acc = 0;
+
+        for(int k = 0; k < K; ++k)
+        {
+            float v_a;
+            float v_b;
+
+            a_element_op(v_a, static_cast<const float>(a_m_k(m, k)));
+            b_element_op(v_b, static_cast<const float>(b_k_n(k, n)));
+
+            v_acc += v_a * v_b;
+        }
+
+        float v_c;
+
+        c_element_op(v_c, v_acc);
+
+        c_m_n(m, n) = v_c;
+    };
+
+    make_ParallelTensorFunctor(f_mk_kn_mn,
+                               c_m_n.mDesc.GetLengths()[0],
+                               c_m_n.mDesc.GetLengths()[1])(std::thread::hardware_concurrency());
+}
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/host_reduction.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/host_reduction.hpp
new file mode 100644
index 0000000..7c0c969
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/host_reduction.hpp
@@ -0,0 +1,374 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <array>
+#include <functional>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+template <int NDim>
+static void get_all_indexes(const std::array<size_t, NDim>& dimLengths,
+                            std::vector<std::array<size_t, NDim>>& indexes)
+{
+    static_assert(NDim >= 1, "NDim >= 1 is required to use this function!");
+
+    if constexpr(NDim == 1)
+    {
+        for(size_t i = 0; i < dimLengths[0]; i++)
+        {
+            std::array<size_t, 1> index{i};
+
+            indexes.push_back(index);
+        };
+    }
+    else
+    {
+        std::array<size_t, NDim - 1> partial_dim_lengths;
+
+        for(int i = 0; i < NDim - 1; i++)
+            partial_dim_lengths[i] = dimLengths[i + 1];
+
+        std::vector<std::array<size_t, NDim - 1>> partial_indexes;
+
+        get_all_indexes<NDim - 1>(partial_dim_lengths, partial_indexes);
+
+        for(size_t i = 0; i < dimLengths[0]; i++)
+            for(const auto& index : partial_indexes)
+            {
+                std::array<size_t, NDim> extIndex;
+
+                extIndex[0] = i;
+
+                for(int k = 0; k < NDim - 1; k++)
+                    extIndex[k + 1] = index[k];
+
+                indexes.push_back(extIndex);
+            };
+    };
+};
+
+template <int NDim>
+static size_t get_offset_from_index(const std::array<size_t, NDim>& strides,
+                                    const std::array<size_t, NDim>& index)
+{
+    size_t offset = 0;
+
+    for(int i = 0; i < NDim; i++)
+        offset += strides[i] * index[i];
+
+    return (offset);
+};
+
+template <int NDim>
+static size_t get_offset_from_index(const std::vector<size_t>& strides,
+                                    const std::array<size_t, NDim>& index)
+{
+    size_t offset = 0;
+
+    for(int i = 0; i < NDim; i++)
+        offset += strides[i] * index[i];
+
+    return (offset);
+};
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          int Rank,
+          int NumReduceDim,
+          bool PropagateNan,
+          bool OutputIndex>
+struct ReductionHost
+{
+    using IndexDataType = int32_t;
+
+    static constexpr int NumInvariantDim = Rank - NumReduceDim;
+
+    std::vector<size_t> outStrides;
+
+    IndexDataType divider;
+
+    std::array<size_t, NumReduceDim> reduceLengths;
+    std::array<size_t, NumReduceDim> reduceStrides;
+    std::array<size_t, NumInvariantDim> invariantLengths;
+    std::array<size_t, NumInvariantDim> invariantStrides;
+
+    std::vector<std::array<size_t, NumReduceDim>> reduce_dim_indexes;
+    std::vector<std::array<size_t, NumInvariantDim>> invariant_dim_indexes;
+
+    ReductionHost(HostTensorDescriptor& inDesc,
+                  HostTensorDescriptor& outDesc,
+                  const std::array<int, NumInvariantDim> invariantDims,
+                  const std::array<int, NumReduceDim> reduceDims)
+    {
+        // this->outLengths = to_int_vector(outDesc.GetLengths());
+        this->outStrides = outDesc.GetStrides();
+
+        int product = 1;
+
+        for(int i = 0; i < NumReduceDim; i++)
+        {
+            reduceLengths[i] = inDesc.GetLengths()[reduceDims[i]];
+            reduceStrides[i] = inDesc.GetStrides()[reduceDims[i]];
+            product *= inDesc.GetLengths()[reduceDims[i]];
+        };
+
+        divider = product;
+
+        for(int i = 0; i < NumInvariantDim; i++)
+        {
+            invariantLengths[i] = inDesc.GetLengths()[invariantDims[i]];
+            invariantStrides[i] = inDesc.GetStrides()[invariantDims[i]];
+        };
+
+        reduce_dim_indexes.clear();
+        get_all_indexes<NumReduceDim>(reduceLengths, reduce_dim_indexes);
+
+        if constexpr(NumInvariantDim > 0)
+        {
+            invariant_dim_indexes.clear();
+            get_all_indexes<NumInvariantDim>(invariantLengths, invariant_dim_indexes);
+        };
+    };
+
+    void Run(float alpha,
+             const InDataType* in_data,
+             float beta,
+             OutDataType* out_data,
+             IndexDataType* out_indices,
+             InElementwiseOperation in_elementwise_op,
+             AccElementwiseOperation acc_elementwise_op)
+    {
+        if constexpr(OutputIndex)
+        {
+            RunImpl_with_index(
+                alpha, in_data, beta, out_data, out_indices, in_elementwise_op, acc_elementwise_op);
+        }
+        else
+        {
+            RunImpl_no_index(alpha, in_data, beta, out_data, in_elementwise_op, acc_elementwise_op);
+        };
+    };
+
+    void RunImpl_with_index(float alpha,
+                            const InDataType* in_data,
+                            float beta,
+                            OutDataType* out_data,
+                            IndexDataType* out_indices,
+                            InElementwiseOperation in_elementwise_op,
+                            AccElementwiseOperation acc_elementwise_op)
+    {
+        using ck::float_equal_one;
+        using ck::float_equal_zero;
+        using ck::type_convert;
+
+        using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
+                                                                        ReduceOperation,
+                                                                        AccDataType,
+                                                                        IndexDataType>;
+
+        if constexpr(NumInvariantDim == 0)
+        {
+            AccDataType accuVal     = ReduceOperation::template GetIdentityValue<AccDataType>();
+            IndexDataType accuIndex = 0;
+
+            for(std::size_t i = 0; i < reduce_dim_indexes.size(); i++)
+            {
+                auto offset_reduce =
+                    get_offset_from_index<NumReduceDim>(reduceStrides, reduce_dim_indexes[i]);
+
+                auto currVal = type_convert<AccDataType>(in_data[offset_reduce]);
+
+                in_elementwise_op(currVal, currVal);
+
+                auto currIndex = static_cast<IndexDataType>(i);
+
+                Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+            };
+
+            acc_elementwise_op(accuVal, accuVal);
+
+            if(!float_equal_one{}(alpha))
+                accuVal *= type_convert<AccDataType>(alpha);
+
+            if(!float_equal_zero{}(beta))
+                accuVal += type_convert<AccDataType>(out_data[0]) * type_convert<AccDataType>(beta);
+
+            out_data[0]    = type_convert<OutDataType>(accuVal);
+            out_indices[0] = accuIndex;
+        }
+        else
+        {
+            auto thread_reduce_func = [&](auto invariant_index) {
+                AccDataType accuVal     = ReduceOperation::template GetIdentityValue<AccDataType>();
+                IndexDataType accuIndex = 0;
+
+                auto offset_invariant =
+                    get_offset_from_index<NumInvariantDim>(invariantStrides, invariant_index);
+
+                for(std::size_t i = 0; i < reduce_dim_indexes.size(); i++)
+                {
+                    auto offset_reduce =
+                        get_offset_from_index<NumReduceDim>(reduceStrides, reduce_dim_indexes[i]);
+
+                    auto currVal =
+                        type_convert<AccDataType>(in_data[offset_invariant + offset_reduce]);
+
+                    in_elementwise_op(currVal, currVal);
+
+                    auto currIndex = static_cast<IndexDataType>(i);
+
+                    Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+                };
+
+                acc_elementwise_op(accuVal, accuVal);
+
+                if(!float_equal_one{}(alpha))
+                    accuVal *= type_convert<AccDataType>(alpha);
+
+                auto dst_offset =
+                    get_offset_from_index<NumInvariantDim>(outStrides, invariant_index);
+
+                if(!float_equal_zero{}(beta))
+                    accuVal += type_convert<AccDataType>(out_data[dst_offset]) *
+                               type_convert<AccDataType>(beta);
+
+                out_data[dst_offset]    = type_convert<OutDataType>(accuVal);
+                out_indices[dst_offset] = accuIndex;
+            };
+
+            std::size_t num_thread = 1;
+            std::size_t work_per_thread =
+                (invariant_dim_indexes.size() + num_thread - 1) / num_thread;
+
+            std::vector<joinable_thread> threads(num_thread);
+
+            for(std::size_t it = 0; it < num_thread; ++it)
+            {
+                std::size_t iw_begin = it * work_per_thread;
+                std::size_t iw_end =
+                    std::min((it + 1) * work_per_thread, invariant_dim_indexes.size());
+
+                auto f = [=] {
+                    for(std::size_t iw = iw_begin; iw < iw_end; ++iw)
+                    {
+                        thread_reduce_func(invariant_dim_indexes[iw]);
+                    }
+                };
+
+                threads[it] = joinable_thread(f);
+            }
+        };
+    };
+
+    void RunImpl_no_index(float alpha,
+                          const InDataType* in_data,
+                          float beta,
+                          OutDataType* out_data,
+                          InElementwiseOperation in_elementwise_op,
+                          AccElementwiseOperation acc_elementwise_op)
+    {
+        using ck::float_equal_one;
+        using ck::float_equal_zero;
+        using ck::type_convert;
+
+        using Accumulation =
+            ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+        if constexpr(NumInvariantDim == 0)
+        {
+            AccDataType accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+            for(const auto& reduce_index : reduce_dim_indexes)
+            {
+                auto offset_reduce =
+                    get_offset_from_index<NumReduceDim>(reduceStrides, reduce_index);
+
+                auto currVal = type_convert<AccDataType>(in_data[offset_reduce]);
+
+                in_elementwise_op(currVal, currVal);
+
+                Accumulation::Calculate(accuVal, currVal);
+            };
+
+            acc_elementwise_op(accuVal, accuVal);
+
+            if(!float_equal_one{}(alpha))
+                accuVal *= type_convert<AccDataType>(alpha);
+
+            if(!float_equal_zero{}(beta))
+                accuVal += type_convert<AccDataType>(out_data[0]) * type_convert<AccDataType>(beta);
+
+            out_data[0] = type_convert<OutDataType>(accuVal);
+        }
+        else
+        {
+            auto thread_reduce_func = [&](auto invariant_index) {
+                AccDataType accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+                auto offset_invariant =
+                    get_offset_from_index<NumInvariantDim>(invariantStrides, invariant_index);
+
+                for(const auto& reduce_index : reduce_dim_indexes)
+                {
+                    auto offset_reduce =
+                        get_offset_from_index<NumReduceDim>(reduceStrides, reduce_index);
+
+                    auto currVal =
+                        type_convert<AccDataType>(in_data[offset_invariant + offset_reduce]);
+
+                    in_elementwise_op(currVal, currVal);
+
+                    Accumulation::Calculate(accuVal, currVal);
+                };
+
+                acc_elementwise_op(accuVal, accuVal);
+
+                if(!float_equal_one{}(alpha))
+                    accuVal *= type_convert<AccDataType>(alpha);
+
+                auto dst_offset =
+                    get_offset_from_index<NumInvariantDim>(outStrides, invariant_index);
+
+                if(!float_equal_zero{}(beta))
+                    accuVal += type_convert<AccDataType>(out_data[dst_offset]) *
+                               type_convert<AccDataType>(beta);
+
+                out_data[dst_offset] = type_convert<OutDataType>(accuVal);
+            };
+
+            std::size_t num_thread = 1;
+            std::size_t work_per_thread =
+                (invariant_dim_indexes.size() + num_thread - 1) / num_thread;
+
+            std::vector<joinable_thread> threads(num_thread);
+
+            for(std::size_t it = 0; it < num_thread; ++it)
+            {
+                std::size_t iw_begin = it * work_per_thread;
+                std::size_t iw_end =
+                    std::min((it + 1) * work_per_thread, invariant_dim_indexes.size());
+
+                auto f = [=] {
+                    for(std::size_t iw = iw_begin; iw < iw_end; ++iw)
+                    {
+                        thread_reduce_func(invariant_dim_indexes[iw]);
+                    }
+                };
+
+                threads[it] = joinable_thread(f);
+            }
+        };
+    };
+};
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/host_tensor.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/host_tensor.hpp
new file mode 100644
index 0000000..a8c7fd0
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/host_tensor.hpp
@@ -0,0 +1,472 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cassert>
+#include <iostream>
+#include <numeric>
+#include <thread>
+#include <utility>
+#include <vector>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/span.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/ranges.hpp"
+
+template <typename Range>
+std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim)
+{
+    bool first = true;
+    for(auto&& v : range)
+    {
+        if(first)
+            first = false;
+        else
+            os << delim;
+        os << v;
+    }
+    return os;
+}
+
+template <typename T, typename Range>
+std::ostream& LogRangeAsType(std::ostream& os, Range&& range, std::string delim)
+{
+    bool first = true;
+    for(auto&& v : range)
+    {
+        if(first)
+            first = false;
+        else
+            os << delim;
+        os << static_cast<T>(v);
+    }
+    return os;
+}
+
+template <typename F, typename T, std::size_t... Is>
+auto call_f_unpack_args_impl(F f, T args, std::index_sequence<Is...>)
+{
+    return f(std::get<Is>(args)...);
+}
+
+template <typename F, typename T>
+auto call_f_unpack_args(F f, T args)
+{
+    constexpr std::size_t N = std::tuple_size<T>{};
+
+    return call_f_unpack_args_impl(f, args, std::make_index_sequence<N>{});
+}
+
+template <typename F, typename T, std::size_t... Is>
+auto construct_f_unpack_args_impl(T args, std::index_sequence<Is...>)
+{
+    return F(std::get<Is>(args)...);
+}
+
+template <typename F, typename T>
+auto construct_f_unpack_args(F, T args)
+{
+    constexpr std::size_t N = std::tuple_size<T>{};
+
+    return construct_f_unpack_args_impl<F>(args, std::make_index_sequence<N>{});
+}
+
+struct HostTensorDescriptor
+{
+    HostTensorDescriptor() = default;
+
+    void CalculateStrides();
+
+    template <typename X, typename = std::enable_if_t<std::is_convertible_v<X, std::size_t>>>
+    HostTensorDescriptor(const std::initializer_list<X>& lens) : mLens(lens.begin(), lens.end())
+    {
+        this->CalculateStrides();
+    }
+
+    template <typename Lengths,
+              typename = std::enable_if_t<
+                  std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t>>>
+    HostTensorDescriptor(const Lengths& lens) : mLens(lens.begin(), lens.end())
+    {
+        this->CalculateStrides();
+    }
+
+    template <typename X,
+              typename Y,
+              typename = std::enable_if_t<std::is_convertible_v<X, std::size_t> &&
+                                          std::is_convertible_v<Y, std::size_t>>>
+    HostTensorDescriptor(const std::initializer_list<X>& lens,
+                         const std::initializer_list<Y>& strides)
+        : mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
+    {
+    }
+
+    template <typename Lengths,
+              typename Strides,
+              typename = std::enable_if_t<
+                  std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t> &&
+                  std::is_convertible_v<ck::ranges::range_value_t<Strides>, std::size_t>>>
+    HostTensorDescriptor(const Lengths& lens, const Strides& strides)
+        : mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
+    {
+    }
+
+    std::size_t GetNumOfDimension() const;
+    std::size_t GetElementSize() const;
+    std::size_t GetElementSpaceSize() const;
+
+    const std::vector<std::size_t>& GetLengths() const;
+    const std::vector<std::size_t>& GetStrides() const;
+
+    template <typename... Is>
+    std::size_t GetOffsetFromMultiIndex(Is... is) const
+    {
+        assert(sizeof...(Is) == this->GetNumOfDimension());
+        std::initializer_list<std::size_t> iss{static_cast<std::size_t>(is)...};
+        return std::inner_product(iss.begin(), iss.end(), mStrides.begin(), std::size_t{0});
+    }
+
+    std::size_t GetOffsetFromMultiIndex(std::vector<std::size_t> iss) const
+    {
+        return std::inner_product(iss.begin(), iss.end(), mStrides.begin(), std::size_t{0});
+    }
+
+    friend std::ostream& operator<<(std::ostream& os, const HostTensorDescriptor& desc);
+
+    private:
+    std::vector<std::size_t> mLens;
+    std::vector<std::size_t> mStrides;
+};
+
+template <typename New2Old>
+HostTensorDescriptor transpose_host_tensor_descriptor_given_new2old(const HostTensorDescriptor& a,
+                                                                    const New2Old& new2old)
+{
+    std::vector<std::size_t> new_lengths(a.GetNumOfDimension());
+    std::vector<std::size_t> new_strides(a.GetNumOfDimension());
+
+    for(std::size_t i = 0; i < a.GetNumOfDimension(); i++)
+    {
+        new_lengths[i] = a.GetLengths()[new2old[i]];
+        new_strides[i] = a.GetStrides()[new2old[i]];
+    }
+
+    return HostTensorDescriptor(new_lengths, new_strides);
+}
+
+struct joinable_thread : std::thread
+{
+    template <typename... Xs>
+    joinable_thread(Xs&&... xs) : std::thread(std::forward<Xs>(xs)...)
+    {
+    }
+
+    joinable_thread(joinable_thread&&) = default;
+    joinable_thread& operator=(joinable_thread&&) = default;
+
+    ~joinable_thread()
+    {
+        if(this->joinable())
+            this->join();
+    }
+};
+
+template <typename F, typename... Xs>
+struct ParallelTensorFunctor
+{
+    F mF;
+    static constexpr std::size_t NDIM = sizeof...(Xs);
+    std::array<std::size_t, NDIM> mLens;
+    std::array<std::size_t, NDIM> mStrides;
+    std::size_t mN1d;
+
+    ParallelTensorFunctor(F f, Xs... xs) : mF(f), mLens({static_cast<std::size_t>(xs)...})
+    {
+        mStrides.back() = 1;
+        std::partial_sum(mLens.rbegin(),
+                         mLens.rend() - 1,
+                         mStrides.rbegin() + 1,
+                         std::multiplies<std::size_t>());
+        mN1d = mStrides[0] * mLens[0];
+    }
+
+    std::array<std::size_t, NDIM> GetNdIndices(std::size_t i) const
+    {
+        std::array<std::size_t, NDIM> indices;
+
+        for(std::size_t idim = 0; idim < NDIM; ++idim)
+        {
+            indices[idim] = i / mStrides[idim];
+            i -= indices[idim] * mStrides[idim];
+        }
+
+        return indices;
+    }
+
+    void operator()(std::size_t num_thread = 1) const
+    {
+        std::size_t work_per_thread = (mN1d + num_thread - 1) / num_thread;
+
+        std::vector<joinable_thread> threads(num_thread);
+
+        for(std::size_t it = 0; it < num_thread; ++it)
+        {
+            std::size_t iw_begin = it * work_per_thread;
+            std::size_t iw_end   = std::min((it + 1) * work_per_thread, mN1d);
+
+            auto f = [=] {
+                for(std::size_t iw = iw_begin; iw < iw_end; ++iw)
+                {
+                    call_f_unpack_args(mF, GetNdIndices(iw));
+                }
+            };
+            threads[it] = joinable_thread(f);
+        }
+    }
+};
+
+template <typename F, typename... Xs>
+auto make_ParallelTensorFunctor(F f, Xs... xs)
+{
+    return ParallelTensorFunctor<F, Xs...>(f, xs...);
+}
+
+template <typename T>
+struct Tensor
+{
+    using Descriptor = HostTensorDescriptor;
+    using Data       = std::vector<T>;
+
+    template <typename X>
+    Tensor(std::initializer_list<X> lens) : mDesc(lens), mData(mDesc.GetElementSpaceSize())
+    {
+    }
+
+    template <typename X, typename Y>
+    Tensor(std::initializer_list<X> lens, std::initializer_list<Y> strides)
+        : mDesc(lens, strides), mData(mDesc.GetElementSpaceSize())
+    {
+    }
+
+    template <typename Lengths>
+    Tensor(const Lengths& lens) : mDesc(lens), mData(mDesc.GetElementSpaceSize())
+    {
+    }
+
+    template <typename Lengths, typename Strides>
+    Tensor(const Lengths& lens, const Strides& strides)
+        : mDesc(lens, strides), mData(GetElementSpaceSize())
+    {
+    }
+
+    Tensor(const Descriptor& desc) : mDesc(desc), mData(mDesc.GetElementSpaceSize()) {}
+
+    template <typename OutT>
+    Tensor<OutT> CopyAsType() const
+    {
+        Tensor<OutT> ret(mDesc);
+
+        ck::ranges::transform(
+            mData, ret.mData.begin(), [](auto value) { return ck::type_convert<OutT>(value); });
+
+        return ret;
+    }
+
+    Tensor()              = delete;
+    Tensor(const Tensor&) = default;
+    Tensor(Tensor&&)      = default;
+
+    ~Tensor() = default;
+
+    Tensor& operator=(const Tensor&) = default;
+    Tensor& operator=(Tensor&&) = default;
+
+    template <typename FromT>
+    explicit Tensor(const Tensor<FromT>& other) : Tensor(other.template CopyAsType<T>())
+    {
+    }
+
+    decltype(auto) GetLengths() const { return mDesc.GetLengths(); }
+
+    decltype(auto) GetStrides() const { return mDesc.GetStrides(); }
+
+    std::size_t GetNumOfDimension() const { return mDesc.GetNumOfDimension(); }
+
+    std::size_t GetElementSize() const { return mDesc.GetElementSize(); }
+
+    std::size_t GetElementSpaceSize() const { return mDesc.GetElementSpaceSize(); }
+
+    std::size_t GetElementSpaceSizeInBytes() const { return sizeof(T) * GetElementSpaceSize(); }
+
+    void SetZero() { ck::ranges::fill<T>(mData, 0); }
+
+    template <typename F>
+    void ForEach_impl(F&& f, std::vector<size_t>& idx, size_t rank)
+    {
+        if(rank == mDesc.GetNumOfDimension())
+        {
+            f(*this, idx);
+            return;
+        }
+        // else
+        for(size_t i = 0; i < mDesc.GetLengths()[rank]; i++)
+        {
+            idx[rank] = i;
+            ForEach_impl(std::forward<F>(f), idx, rank + 1);
+        }
+    }
+
+    template <typename F>
+    void ForEach(F&& f)
+    {
+        std::vector<size_t> idx(mDesc.GetNumOfDimension(), 0);
+        ForEach_impl(std::forward<F>(f), idx, size_t(0));
+    }
+
+    template <typename F>
+    void ForEach_impl(const F&& f, std::vector<size_t>& idx, size_t rank) const
+    {
+        if(rank == mDesc.GetNumOfDimension())
+        {
+            f(*this, idx);
+            return;
+        }
+        // else
+        for(size_t i = 0; i < mDesc.GetLengths()[rank]; i++)
+        {
+            idx[rank] = i;
+            ForEach_impl(std::forward<const F>(f), idx, rank + 1);
+        }
+    }
+
+    template <typename F>
+    void ForEach(const F&& f) const
+    {
+        std::vector<size_t> idx(mDesc.GetNumOfDimension(), 0);
+        ForEach_impl(std::forward<const F>(f), idx, size_t(0));
+    }
+
+    template <typename G>
+    void GenerateTensorValue(G g, std::size_t num_thread = 1)
+    {
+        switch(mDesc.GetNumOfDimension())
+        {
+        case 1: {
+            auto f = [&](auto i) { (*this)(i) = g(i); };
+            make_ParallelTensorFunctor(f, mDesc.GetLengths()[0])(num_thread);
+            break;
+        }
+        case 2: {
+            auto f = [&](auto i0, auto i1) { (*this)(i0, i1) = g(i0, i1); };
+            make_ParallelTensorFunctor(f, mDesc.GetLengths()[0], mDesc.GetLengths()[1])(num_thread);
+            break;
+        }
+        case 3: {
+            auto f = [&](auto i0, auto i1, auto i2) { (*this)(i0, i1, i2) = g(i0, i1, i2); };
+            make_ParallelTensorFunctor(
+                f, mDesc.GetLengths()[0], mDesc.GetLengths()[1], mDesc.GetLengths()[2])(num_thread);
+            break;
+        }
+        case 4: {
+            auto f = [&](auto i0, auto i1, auto i2, auto i3) {
+                (*this)(i0, i1, i2, i3) = g(i0, i1, i2, i3);
+            };
+            make_ParallelTensorFunctor(f,
+                                       mDesc.GetLengths()[0],
+                                       mDesc.GetLengths()[1],
+                                       mDesc.GetLengths()[2],
+                                       mDesc.GetLengths()[3])(num_thread);
+            break;
+        }
+        case 5: {
+            auto f = [&](auto i0, auto i1, auto i2, auto i3, auto i4) {
+                (*this)(i0, i1, i2, i3, i4) = g(i0, i1, i2, i3, i4);
+            };
+            make_ParallelTensorFunctor(f,
+                                       mDesc.GetLengths()[0],
+                                       mDesc.GetLengths()[1],
+                                       mDesc.GetLengths()[2],
+                                       mDesc.GetLengths()[3],
+                                       mDesc.GetLengths()[4])(num_thread);
+            break;
+        }
+        case 6: {
+            auto f = [&](auto i0, auto i1, auto i2, auto i3, auto i4, auto i5) {
+                (*this)(i0, i1, i2, i3, i4) = g(i0, i1, i2, i3, i4, i5);
+            };
+            make_ParallelTensorFunctor(f,
+                                       mDesc.GetLengths()[0],
+                                       mDesc.GetLengths()[1],
+                                       mDesc.GetLengths()[2],
+                                       mDesc.GetLengths()[3],
+                                       mDesc.GetLengths()[4],
+                                       mDesc.GetLengths()[5])(num_thread);
+            break;
+        }
+        default: throw std::runtime_error("unspported dimension");
+        }
+    }
+
+    template <typename... Is>
+    T& operator()(Is... is)
+    {
+        return mData[mDesc.GetOffsetFromMultiIndex(is...)];
+    }
+
+    template <typename... Is>
+    const T& operator()(Is... is) const
+    {
+        return mData[mDesc.GetOffsetFromMultiIndex(is...)];
+    }
+
+    T& operator()(std::vector<std::size_t> idx)
+    {
+        return mData[mDesc.GetOffsetFromMultiIndex(idx)];
+    }
+
+    const T& operator()(std::vector<std::size_t> idx) const
+    {
+        return mData[mDesc.GetOffsetFromMultiIndex(idx)];
+    }
+
+    typename Data::iterator begin() { return mData.begin(); }
+
+    typename Data::iterator end() { return mData.end(); }
+
+    typename Data::pointer data() { return mData.data(); }
+
+    typename Data::const_iterator begin() const { return mData.begin(); }
+
+    typename Data::const_iterator end() const { return mData.end(); }
+
+    typename Data::const_pointer data() const { return mData.data(); }
+
+    typename Data::size_type size() const { return mData.size(); }
+
+    template <typename U = T>
+    auto AsSpan() const
+    {
+        constexpr std::size_t FromSize = sizeof(T);
+        constexpr std::size_t ToSize   = sizeof(U);
+
+        using Element = std::add_const_t<std::remove_reference_t<U>>;
+        return ck::span<Element>{reinterpret_cast<Element*>(data()), size() * FromSize / ToSize};
+    }
+
+    template <typename U = T>
+    auto AsSpan()
+    {
+        constexpr std::size_t FromSize = sizeof(T);
+        constexpr std::size_t ToSize   = sizeof(U);
+
+        using Element = std::remove_reference_t<U>;
+        return ck::span<Element>{reinterpret_cast<Element*>(data()), size() * FromSize / ToSize};
+    }
+
+    Descriptor mDesc;
+    Data mData;
+};
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/host_tensor_generator.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/host_tensor_generator.hpp
new file mode 100644
index 0000000..4259862
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/host_tensor_generator.hpp
@@ -0,0 +1,190 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cmath>
+#include <numeric>
+#include <random>
+
+#include "ck/ck.hpp"
+
+template <typename T>
+struct GeneratorTensor_0
+{
+    template <typename... Is>
+    T operator()(Is...)
+    {
+        return T{0};
+    }
+};
+
+template <typename T>
+struct GeneratorTensor_1
+{
+    T value = 1;
+
+    template <typename... Is>
+    T operator()(Is...)
+    {
+        return value;
+    }
+};
+
+template <>
+struct GeneratorTensor_1<ck::bhalf_t>
+{
+    float value = 1.0;
+
+    template <typename... Is>
+    ck::bhalf_t operator()(Is...)
+    {
+        return ck::type_convert<ck::bhalf_t>(value);
+    }
+};
+
+template <>
+struct GeneratorTensor_1<int8_t>
+{
+    int8_t value = 1;
+
+    template <typename... Is>
+    int8_t operator()(Is...)
+    {
+        return value;
+    }
+};
+
+template <typename T>
+struct GeneratorTensor_2
+{
+    int min_value = 0;
+    int max_value = 1;
+
+    template <typename... Is>
+    T operator()(Is...)
+    {
+        return static_cast<T>((std::rand() % (max_value - min_value)) + min_value);
+    }
+};
+
+template <>
+struct GeneratorTensor_2<ck::bhalf_t>
+{
+    int min_value = 0;
+    int max_value = 1;
+
+    template <typename... Is>
+    ck::bhalf_t operator()(Is...)
+    {
+        float tmp = (std::rand() % (max_value - min_value)) + min_value;
+        return ck::type_convert<ck::bhalf_t>(tmp);
+    }
+};
+
+template <>
+struct GeneratorTensor_2<int8_t>
+{
+    int min_value = 0;
+    int max_value = 1;
+
+    template <typename... Is>
+    int8_t operator()(Is...)
+    {
+        return (std::rand() % (max_value - min_value)) + min_value;
+    }
+};
+
+template <typename T>
+struct GeneratorTensor_3
+{
+    float min_value = 0;
+    float max_value = 1;
+
+    template <typename... Is>
+    T operator()(Is...)
+    {
+        float tmp = float(std::rand()) / float(RAND_MAX);
+
+        return static_cast<T>(min_value + tmp * (max_value - min_value));
+    }
+};
+
+template <>
+struct GeneratorTensor_3<ck::bhalf_t>
+{
+    float min_value = 0;
+    float max_value = 1;
+
+    template <typename... Is>
+    ck::bhalf_t operator()(Is...)
+    {
+        float tmp = float(std::rand()) / float(RAND_MAX);
+
+        float fp32_tmp = min_value + tmp * (max_value - min_value);
+
+        return ck::type_convert<ck::bhalf_t>(fp32_tmp);
+    }
+};
+
+template <typename T>
+struct GeneratorTensor_4
+{
+    std::default_random_engine generator;
+    std::normal_distribution<float> distribution;
+
+    GeneratorTensor_4(float mean, float stddev) : generator(1), distribution(mean, stddev){};
+
+    template <typename... Is>
+    T operator()(Is...)
+    {
+        float tmp = distribution(generator);
+
+        return ck::type_convert<T>(tmp);
+    }
+};
+
+struct GeneratorTensor_Checkboard
+{
+    template <typename... Ts>
+    float operator()(Ts... Xs) const
+    {
+        std::array<ck::index_t, sizeof...(Ts)> dims = {static_cast<ck::index_t>(Xs)...};
+        return std::accumulate(dims.begin(),
+                               dims.end(),
+                               true,
+                               [](bool init, ck::index_t x) -> int { return init != (x % 2); })
+                   ? 1
+                   : -1;
+    }
+};
+
+template <ck::index_t Dim>
+struct GeneratorTensor_Sequential
+{
+    template <typename... Ts>
+    float operator()(Ts... Xs) const
+    {
+        std::array<ck::index_t, sizeof...(Ts)> dims = {{static_cast<ck::index_t>(Xs)...}};
+        return dims[Dim];
+    }
+};
+
+template <typename T, size_t NumEffectiveDim = 2>
+struct GeneratorTensor_Diagonal
+{
+    T value{1};
+
+    template <typename... Ts>
+    T operator()(Ts... Xs) const
+    {
+        std::array<ck::index_t, sizeof...(Ts)> dims = {{static_cast<ck::index_t>(Xs)...}};
+        size_t start_dim                            = dims.size() - NumEffectiveDim;
+        bool pred                                   = true;
+        for(size_t i = start_dim + 1; i < dims.size(); i++)
+        {
+            pred &= (dims[start_dim] == dims[i]);
+        }
+        return pred ? value : T{0};
+    }
+};
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/iterator.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/iterator.hpp
new file mode 100644
index 0000000..9fdc88e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/iterator.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iterator>
+#include <utility>
+
+#include "ck/utility/type.hpp"
+
+namespace ck {
+
+template <typename T>
+using iter_value_t = typename std::iterator_traits<remove_cvref_t<T>>::value_type;
+
+template <typename T>
+using iter_reference_t = decltype(*std::declval<T&>());
+
+template <typename T>
+using iter_difference_t = typename std::iterator_traits<remove_cvref_t<T>>::difference_type;
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/literals.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/literals.hpp
new file mode 100644
index 0000000..a73a2ea
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/literals.hpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace literals {
+// [P0330] Literal Suffix for (signed) size_t (C++23)
+// ref: https://wg21.link/p0330r8
+inline constexpr std::size_t operator""_uz(unsigned long long size)
+{
+    return static_cast<std::size_t>(size);
+}
+
+inline constexpr std::size_t operator""_zu(unsigned long long size)
+{
+    return static_cast<std::size_t>(size);
+}
+} // namespace literals
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/numeric.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/numeric.hpp
new file mode 100644
index 0000000..70a7e87
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/numeric.hpp
@@ -0,0 +1,16 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iterator>
+#include <numeric>
+
+namespace ck {
+template <typename T, typename ForwardIterator, typename Size, typename BinaryOperation>
+auto accumulate_n(ForwardIterator first, Size count, T init, BinaryOperation op)
+    -> decltype(std::accumulate(first, std::next(first, count), init, op))
+{
+    return std::accumulate(first, std::next(first, count), init, op);
+}
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/op_instance_engine.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/op_instance_engine.hpp
new file mode 100644
index 0000000..78812e8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/op_instance_engine.hpp
@@ -0,0 +1,249 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <iostream>
+#include <limits>
+#include <memory>
+#include <stdexcept>
+#include <tuple>
+#include <utility>
+#include <vector>
+
+#include "ck/utility/functional2.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace utils {
+
+struct ProfileBestConfig
+{
+    std::string best_op_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_tflops     = std::numeric_limits<float>::max();
+    float best_gb_per_sec = std::numeric_limits<float>::max();
+};
+
+/**
+ * @brief      This class describes an operation instance(s).
+ *
+ *             Op instance defines a particular specializations of operator
+ *             template. Thanks to this specific input/output data types, data
+ *             layouts and modifying elementwise operations it is able to create
+ *             it's input/output tensors, provide pointers to instances which
+ *             can execute it and all operation specific parameters.
+ */
+template <typename OutDataType, typename... InArgTypes>
+class OpInstance
+{
+    public:
+    template <typename T>
+    using TensorPtr      = std::unique_ptr<Tensor<T>>;
+    using InTensorsTuple = std::tuple<TensorPtr<InArgTypes>...>;
+    using DeviceMemPtr   = std::unique_ptr<DeviceMem>;
+    using DeviceBuffers  = std::vector<DeviceMemPtr>;
+
+    OpInstance()                  = default;
+    OpInstance(const OpInstance&) = default;
+    OpInstance& operator=(const OpInstance&) = default;
+    virtual ~OpInstance(){};
+
+    virtual InTensorsTuple GetInputTensors() const         = 0;
+    virtual TensorPtr<OutDataType> GetOutputTensor() const = 0;
+    virtual std::unique_ptr<tensor_operation::device::BaseInvoker>
+    MakeInvokerPointer(tensor_operation::device::BaseOperator*) const = 0;
+    virtual std::unique_ptr<tensor_operation::device::BaseArgument>
+    MakeArgumentPointer(tensor_operation::device::BaseOperator*,
+                        const DeviceBuffers&,
+                        const DeviceMemPtr&) const = 0;
+    virtual std::size_t GetFlops() const           = 0;
+    virtual std::size_t GetBtype() const           = 0;
+};
+
+/**
+ * @brief      A generic operation instance run engine.
+ */
+template <typename OutDataType, typename... InArgTypes>
+class OpInstanceRunEngine
+{
+    public:
+    using OpInstanceT = OpInstance<InArgTypes..., OutDataType>;
+    template <typename T>
+    using TensorPtr        = std::unique_ptr<Tensor<T>>;
+    using DeviceMemPtr     = std::unique_ptr<DeviceMem>;
+    using InTensorsTuple   = std::tuple<TensorPtr<InArgTypes>...>;
+    using DeviceBuffers    = std::vector<DeviceMemPtr>;
+    using InArgsTypesTuple = std::tuple<InArgTypes...>;
+
+    OpInstanceRunEngine() = delete;
+
+    template <typename ReferenceOp = std::function<void()>>
+    OpInstanceRunEngine(const OpInstanceT& op_instance,
+                        const ReferenceOp& reference_op = ReferenceOp{},
+                        bool do_verification            = true)
+        : op_instance_{op_instance}
+    {
+        in_tensors_ = op_instance_.GetInputTensors();
+        out_tensor_ = op_instance_.GetOutputTensor();
+
+        if constexpr(std::is_invocable_v<ReferenceOp,
+                                         const Tensor<InArgTypes>&...,
+                                         Tensor<OutDataType>&>)
+        {
+            if(do_verification)
+            {
+                ref_output_ = op_instance_.GetOutputTensor();
+                CallRefOpUnpackArgs(reference_op, std::make_index_sequence<kNInArgs_>{});
+            }
+        }
+        AllocateDeviceInputTensors(std::make_index_sequence<kNInArgs_>{});
+        out_device_buffer_ = std::make_unique<DeviceMem>(sizeof(OutDataType) *
+                                                         out_tensor_->mDesc.GetElementSpaceSize());
+        out_device_buffer_->SetZero();
+    }
+
+    virtual ~OpInstanceRunEngine(){};
+
+    template <typename OpInstancePtr>
+    bool Test(const std::vector<OpInstancePtr>& op_ptrs)
+    {
+        bool res{true};
+        for(auto& op_ptr : op_ptrs)
+        {
+            auto invoker  = op_instance_.MakeInvokerPointer(op_ptr.get());
+            auto argument = op_instance_.MakeArgumentPointer(
+                op_ptr.get(), in_device_buffers_, out_device_buffer_);
+            if(op_ptr->IsSupportedArgument(argument.get()))
+            {
+                std::cout << "Testing instance: " << op_ptr->GetTypeString() << std::endl;
+                invoker->Run(argument.get());
+                out_device_buffer_->FromDevice(out_tensor_->mData.data());
+                if(!ref_output_)
+                {
+                    throw std::runtime_error(
+                        "OpInstanceRunEngine::Test: Reference value not availabe."
+                        " You have to provide reference function.");
+                }
+                // TODO: enable flexible use of custom check_error functions
+                bool inst_res = CheckErr(out_tensor_->mData, ref_output_->mData);
+                std::cout << (inst_res ? "SUCCESS" : "FAILURE") << std::endl;
+                res = res && inst_res;
+                out_device_buffer_->SetZero();
+            }
+            else
+            {
+                std::cout << "Given conv problem is not supported by instance: \n\t>>>>"
+                          << op_ptr->GetTypeString() << std::endl;
+            }
+        }
+        return res;
+    }
+
+    template <typename OpInstancePtr>
+    ProfileBestConfig Profile(const std::vector<OpInstancePtr>& op_ptrs,
+                              bool time_kernel     = false,
+                              bool do_verification = false,
+                              bool do_log          = false)
+    {
+        ProfileBestConfig best_config;
+
+        for(auto& op_ptr : op_ptrs)
+        {
+            auto invoker  = op_instance_.MakeInvokerPointer(op_ptr.get());
+            auto argument = op_instance_.MakeArgumentPointer(
+                op_ptr.get(), in_device_buffers_, out_device_buffer_);
+            if(op_ptr->IsSupportedArgument(argument.get()))
+            {
+                std::string op_name = op_ptr->GetTypeString();
+                float avg_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+                std::size_t flops     = op_instance_.GetFlops();
+                std::size_t num_btype = op_instance_.GetBtype();
+                float tflops          = static_cast<float>(flops) / 1.E9 / avg_time;
+                float gb_per_sec      = num_btype / 1.E6 / avg_time;
+
+                std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                          << " GB/s, " << op_name << std::endl;
+
+                if(avg_time < best_config.best_avg_time)
+                {
+                    best_config.best_op_name    = op_name;
+                    best_config.best_tflops     = tflops;
+                    best_config.best_gb_per_sec = gb_per_sec;
+                    best_config.best_avg_time   = avg_time;
+                }
+
+                if(do_verification)
+                {
+                    out_device_buffer_->FromDevice(out_tensor_->mData.data());
+                    if(!ref_output_)
+                    {
+                        throw std::runtime_error(
+                            "OpInstanceRunEngine::Profile: Reference value not availabe."
+                            " You have to provide reference function.");
+                    }
+                    // TODO: enable flexible use of custom check_error functions
+                    CheckErr(out_tensor_->mData, ref_output_->mData);
+
+                    if(do_log) {}
+                }
+                out_device_buffer_->SetZero();
+            }
+        }
+        return best_config;
+    }
+
+    void SetAtol(double a) { atol_ = a; }
+    void SetRtol(double r) { rtol_ = r; }
+
+    private:
+    template <typename F, std::size_t... Is>
+    void CallRefOpUnpackArgs(const F& f, std::index_sequence<Is...>) const
+    {
+        f(*std::get<Is>(in_tensors_)..., *ref_output_);
+    }
+
+    template <std::size_t... Is>
+    void AllocateDeviceInputTensors(std::index_sequence<Is...>)
+    {
+        (AllocateDeviceInputTensorsImpl<Is>(), ...);
+    }
+
+    template <std::size_t Index>
+    void AllocateDeviceInputTensorsImpl()
+    {
+        const auto& ts = std::get<Index>(in_tensors_);
+        in_device_buffers_
+            .emplace_back(
+                std::make_unique<DeviceMem>(sizeof(std::tuple_element_t<Index, InArgsTypesTuple>) *
+                                            ts->mDesc.GetElementSpaceSize()))
+            ->ToDevice(ts->mData.data());
+    }
+
+    static constexpr std::size_t kNInArgs_ = std::tuple_size_v<InTensorsTuple>;
+    const OpInstanceT& op_instance_;
+    double rtol_{1e-5};
+    double atol_{1e-8};
+
+    InTensorsTuple in_tensors_;
+    TensorPtr<OutDataType> out_tensor_;
+    TensorPtr<OutDataType> ref_output_;
+
+    DeviceBuffers in_device_buffers_;
+    DeviceMemPtr out_device_buffer_;
+
+    template <typename T>
+    bool CheckErr(const std::vector<T>& dev_out, const std::vector<T>& ref_out) const
+    {
+        return ck::utils::check_err(dev_out, ref_out, "Error: incorrect results!", rtol_, atol_);
+    }
+};
+
+} // namespace utils
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/include/ck/library/utility/ranges.hpp b/3rdparty/composable_kernel/library/include/ck/library/utility/ranges.hpp
new file mode 100644
index 0000000..55c322f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/include/ck/library/utility/ranges.hpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iterator>
+#include <type_traits>
+#include <utility>
+
+#include "ck/library/utility/iterator.hpp"
+
+namespace ck {
+namespace ranges {
+
+template <typename R>
+using iterator_t = decltype(std::begin(std::declval<R&>()));
+
+template <typename R>
+using sentinel_t = decltype(std::end(std::declval<R&>()));
+
+template <typename R>
+using range_size_t = decltype(std::size(std::declval<R&>()));
+
+template <typename R>
+using range_difference_t = ck::iter_difference_t<ranges::iterator_t<R>>;
+
+template <typename R>
+using range_value_t = iter_value_t<ranges::iterator_t<R>>;
+
+template <typename R>
+using range_reference_t = iter_reference_t<ranges::iterator_t<R>>;
+
+template <typename T, typename = void>
+struct is_range : std::false_type
+{
+};
+
+template <typename T>
+struct is_range<
+    T,
+    std::void_t<decltype(std::begin(std::declval<T&>())), decltype(std::end(std::declval<T&>()))>>
+    : std::true_type
+{
+};
+
+template <typename T>
+inline constexpr bool is_range_v = is_range<T>::value;
+
+template <typename T, typename = void>
+struct is_sized_range : std::false_type
+{
+};
+
+template <typename T>
+struct is_sized_range<T, std::void_t<decltype(std::size(std::declval<T&>()))>>
+    : std::bool_constant<is_range_v<T>>
+{
+};
+} // namespace ranges
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/CMakeLists.txt
new file mode 100644
index 0000000..9b7cd3a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/CMakeLists.txt
@@ -0,0 +1,69 @@
+function(add_instance_library INSTANCE_NAME)
+    message("adding instance ${INSTANCE_NAME}")
+    add_library(${INSTANCE_NAME} OBJECT ${ARGN})
+    target_compile_features(${INSTANCE_NAME} PUBLIC)
+    set_target_properties(${INSTANCE_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    clang_tidy_check(${INSTANCE_NAME})
+endfunction(add_instance_library INSTANCE_NAME)
+
+
+file(GLOB dir_list LIST_DIRECTORIES true *)
+set(CK_DEVICE_INSTANCES)
+FOREACH(subdir_path ${dir_list})
+set(target_dir)
+IF(IS_DIRECTORY "${subdir_path}")
+    get_filename_component(target_dir ${subdir_path} NAME)
+    add_subdirectory(${target_dir})
+    list(APPEND CK_DEVICE_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
+ENDIF()
+ENDFOREACH()
+
+add_library(device_operations STATIC ${CK_DEVICE_INSTANCES})
+add_library(composablekernels::device_operations ALIAS device_operations)
+
+
+set(DEV_OPS_INC_DIRS
+    ${PROJECT_SOURCE_DIR}/include/ck/
+    ${PROJECT_SOURCE_DIR}/library/include/ck/
+)
+
+target_compile_features(device_operations PUBLIC)
+set_target_properties(device_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
+target_include_directories(device_operations PUBLIC
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/utility>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_description>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/problem_transform>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/device>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/device/impl>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/grid>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/block>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/warp>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/thread>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/element>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/utility>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/reduce>
+)
+
+#once new arches are enabled make this an option on the main cmake file
+# and pass down here to be exported
+target_compile_options(device_operations PRIVATE
+    --offload-arch=gfx908
+    --offload-arch=gfx90a
+)
+
+target_compile_options(device_operations PRIVATE --gpu-max-threads-per-block=1024)
+
+# install(TARGETS device_operations LIBRARY DESTINATION lib)
+rocm_install(TARGETS device_operations
+        EXPORT device_operationsTargets)
+
+rocm_install(DIRECTORY ${DEV_OPS_INC_DIRS} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/ck)
+rocm_install(EXPORT device_operationsTargets
+        FILE composable_kerneldevice_operationsTargets.cmake
+        NAMESPACE composable_kernel::
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt
new file mode 100644
index 0000000..0f2a739
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt
@@ -0,0 +1,18 @@
+add_instance_library(device_batched_gemm_instance
+   device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
+   device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
+   device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
+   device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
+   device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
+   device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
+   device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
+   device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp
+   device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
+   device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
+   device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
+   device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp
+   device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
+   device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
+   device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
+   device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
new file mode 100644
index 0000000..cc87874
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //##################| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##################|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##################|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp
new file mode 100644
index 0000000..04200cf
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
new file mode 100644
index 0000000..7b86f3c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //#################|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //#################|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //#################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
new file mode 100644
index 0000000..2afb1af
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
@@ -0,0 +1,63 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //#################|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //#################|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //#################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
new file mode 100644
index 0000000..68d7689
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK|          LoopScheduler|                    Pipeline|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar| Prefetch|                       |                            |
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector| Stage   |                       |                            |
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+        // pipeline v1, 2 waves
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+        // pipeline v2, 1 wave
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
new file mode 100644
index 0000000..737e5bf
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK|          LoopScheduler|                    Pipeline|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar| Prefetch|                       |                            |
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector| Stage   |                       |                            |
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+        // pipeline v1, 2 waves
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+        // pipeline v2, 1 wave
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
new file mode 100644
index 0000000..e09d017
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
@@ -0,0 +1,110 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK|          LoopScheduler|                    Pipeline|
+        //#################|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar| Prefetch|                       |                            |
+        //#################|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector| Stage   |                       |                            |
+        //#################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,   256,     4,  8,   16,   16,    1,    8,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,   128,     4,  8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    64,     4,  8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,   256,     4,  8,   16,   16,    1,    8,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,   128,     4,  8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    64,     4,  8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,   256,     4,  8,   16,   16,    1,    8,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,   128,     4,  8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    64,     4,  8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
new file mode 100644
index 0000000..984d66e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK|          LoopScheduler|                    Pipeline|
+        //#################|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar| Prefetch|                       |                            |
+        //#################|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector| Stage   |                       |                            |
+        //#################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES         
+        // pipeline v2, 1 wave
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
new file mode 100644
index 0000000..12cada9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp
new file mode 100644
index 0000000..13f1988
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
new file mode 100644
index 0000000..2ca1adc
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
new file mode 100644
index 0000000..fe5de52
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
@@ -0,0 +1,63 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
new file mode 100644
index 0000000..5b55c8e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using AData   = int8_t;
+using BData   = int8_t;
+using CData   = int8_t;
+using AccData = int32_t;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|          AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer|   ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer|  BBlockTransfer|  BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|           Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|    ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|        DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|       SrcScalar|       DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|               |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave|  Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|     PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |       PerVector|    PerVector_K1|          |                |       PerVector|
+        //##########|               |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                 |               |               |               |               |                 |          |                |               |               |              |                |                |          |                |                |
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,               16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,               16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              1,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              2,              16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Col,
+                                                  Row,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp
new file mode 100644
index 0000000..9517e45
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using AData   = int8_t;
+using BData   = int8_t;
+using CData   = int8_t;
+using AccData = int32_t;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|          AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|           Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|               |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|               |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,               16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,               16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Col,
+                                                  Col,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
new file mode 100644
index 0000000..43b9124
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using AData   = int8_t;
+using BData   = int8_t;
+using CData   = int8_t;
+using AccData = int32_t;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|          AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer|  ABlockTransfer|  ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer|  BBlockTransfer|  BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|           Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|       SrcScalar|       DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|       SrcScalar|       DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|               |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |       PerVector|    PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |       PerVector|    PerVector_K1|          |                |       PerVector|
+        //##########|               |      |      |        |        |        |        |            |            |            |      |      |      |      |    |     |     |     |     |                |               |               |               |                |                |          |                |               |               |              |                |                |          |                |                |
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,               16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,               16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,               16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,               16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,               16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,               16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,               16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,               16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Row,
+                                                  Row,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
new file mode 100644
index 0000000..326500f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
@@ -0,0 +1,71 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using AData   = int8_t;
+using BData   = int8_t;
+using CData   = int8_t;
+using AccData = int32_t;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|          AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer|  ABlockTransfer|  ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer|  BBlockTransfer|  BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|           Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|       SrcScalar|       DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|       SrcScalar|       DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|               |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |       PerVector|    PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |       PerVector|    PerVector_K1|          |                |       PerVector|
+        //##########|               |      |      |        |        |        |        |            |            |            |      |      |      |      |    |     |     |     |     |                |               |               |               |                |                |          |                |               |               |              |                |                |          |                |                |
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Row,
+                                                  Col,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt
new file mode 100644
index 0000000..d0e9b26
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt
@@ -0,0 +1,4 @@
+add_instance_library(device_batched_gemm_add_relu_gemm_add_instance
+    device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
+    device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
new file mode 100644
index 0000000..e1bfa88
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough   = ck::tensor_operation::element_wise::PassThrough;
+using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
+using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
+    std::tuple<
+        // clang-format off
+        //##################################################| A0Layout| B0Layout|       D0Layout| B1Layout|      D1sLayout| E1Layout| A0Data| B0Data| Acc0DataType|     D0DataType| B1Data| Acc1CData| CShuffle|        D1sData| E1Data|          A0|          B0|          CDE0|          B1|          CDE1| PadGemm0M| PadGemm0N| PadGemm0K| PadGemm1N| PadGemm1K|NumGemm0K| Block|  Gemm0| Gemm0| Gemm0| Gemm1| Gemm1|A0K1|B0K1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|   C1Shuffle|   C1Shuffle| CDE1BlockTransferClusterLengths| CDE1BlockTransfer|
+        //##################################################|         |         |               |         |               |         |   Type|   Type|         Type|           Type|   Type|      Type| DataType|           Type|   Type| Elementwise| Elementwise|   Elementwise| Elementwise|   Elementwise|          |          |          |          |          | Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|  ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MWaveMPerXdl|   ScalarPerVector|
+        //##################################################|         |         |               |         |               |         |       |       |             |               |       |          |         |               |       |   Operation|   Operation|     Operation|   Operation|     Operation|          |          |          |          |          |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per|Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|            _NBlock_NWaveNPerXdl|     _NWaveNPerXdl|
+        //##################################################|         |         |               |         |               |         |       |       |             |               |       |          |         |               |       |            |            |              |            |              |          |          |          |          |          |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|               |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                                |                  |
+        // no padding
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,                  S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,                  S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,                  S<1, 32, 1, 8>,               8>,
+        // Padded fallback kernel
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultipleDGemmMultipleD<Row,
+                                                                        Col,
+                                                                        ck::Tuple<Row>,
+                                                                        Row,
+                                                                        ck::Tuple<Row>,
+                                                                        Row,
+                                                                        F16,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        CDE0ElementOp,
+                                                                        PassThrough,
+                                                                        CDE1ElementOp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
new file mode 100644
index 0000000..f59b742
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough   = ck::tensor_operation::element_wise::PassThrough;
+using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
+using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instances =
+    std::tuple<
+        // clang-format off
+        //##################################################| A0Layout| B0Layout|       D0Layout| B1Layout|      D1sLayout| E1Layout| A0Data| B0Data| Acc0DataType|     D0DataType| B1Data| Acc1CData| CShuffle|        D1sData| E1Data|          A0|          B0|          CDE0|          B1|          CDE1| PadGemm0M| PadGemm0N| PadGemm0K| PadGemm1N| PadGemm1K| NumGemm0K| Block|  Gemm0| Gemm0| Gemm0| Gemm1| Gemm1| A0K1| B0K1|B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|   C1Shuffle|   C1Shuffle| CDE1BlockTransferClusterLengths| CDE1BlockTransfer|
+        //##################################################|         |         |               |         |               |         |   Type|   Type|         Type|           Type|   Type|      Type| DataType|           Type|   Type| Elementwise| Elementwise|   Elementwise| Elementwise|   Elementwise|          |          |          |          |          |  Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|     |     |    |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MWaveMPerXdl|   ScalarPerVector|
+        //##################################################|         |         |               |         |               |         |       |       |             |               |       |          |         |               |       |   Operation|   Operation|     Operation|   Operation|     Operation|          |          |          |          |          |     Stage|      |  Block| Block| Block| Block| Block|     |     |    |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|            _NBlock_NWaveNPerXdl|     _NWaveNPerXdl|
+        //##################################################|         |         |               |         |               |         |       |       |             |               |       |          |         |               |       |            |            |              |            |              |          |          |          |          |          |          |      |       |      |      |      |      |     |     |    |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                                |                  |
+        // no padding
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,    256,   128,    32,   128,    32,    8,   8,    4,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,    128,   128,    64,    64,    32,    8,   8,    4,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,    128,   128,    32,    64,    32,    8,   8,    4,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,    128,   128,    64,   128,    32,    8,   8,    4,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,    128,   128,    32,   128,    32,    8,   8,    4,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,     64,   256,    32,   128,    32,    8,   8,    4,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           8,                  S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,     64,   256,    32,    64,    32,    8,   8,    4,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           4,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,     64,   256,    64,   128,    32,    8,   8,    4,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           8,                  S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,     64,   256,    64,    64,    32,    8,   8,    4,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           4,                  S<1, 32, 1, 8>,               8>,
+        // Padded fallback kernel                                                                                                                                                                                                       
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,         1,   256,    128,   128,    64,   128,    32,    8,   8,    4,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,         1,   256,    128,    64,    32,   128,    32,    8,   8,    4,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,                  S<1, 32, 1, 8>,               8>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultipleDGemmMultipleD<Row,
+                                                                        Col,
+                                                                        ck::Tuple<Row>,
+                                                                        Col,
+                                                                        ck::Tuple<Row>,
+                                                                        Row,
+                                                                        F16,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        CDE0ElementOp,
+                                                                        PassThrough,
+                                                                        CDE1ElementOp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt
new file mode 100644
index 0000000..865a31e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt
@@ -0,0 +1,4 @@
+add_instance_library(device_batched_gemm_gemm_instance
+    device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
+    device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
new file mode 100644
index 0000000..9b96194
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances = std::tuple<
+    // clang-format off
+        //################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################|        |         |         |        |  Type|   Type|   Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|        |         |         |        |      |       |       |      |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |
+        //################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    2,   32,   32,     1,     8,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8>,
+        // Padded fallback kernel
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
+                                                      Col,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
new file mode 100644
index 0000000..0713dfc
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instances = std::tuple<
+    // clang-format off
+        //################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################|        |         |         |        |  Type|   Type|   Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|        |         |         |        |      |       |       |      |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |
+        //################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    4,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    4,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    4,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    4,   32,   32,     1,     8,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    4,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    4,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    4,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    4,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    4,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           8,               S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    4,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           4,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    4,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           8,               S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    4,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           4,               S<1, 32, 1, 8>,               8>,
+        // Padded fallback kernel
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    64,   128,    32,   8,   8,    4,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,    64,    32,   128,    32,   8,   8,    4,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
+                                                      Col,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
new file mode 100644
index 0000000..db3719c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
@@ -0,0 +1,7 @@
+add_instance_library(device_batched_gemm_reduce_instance
+    device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
+    device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
+    device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
+    device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
+)
+
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
new file mode 100644
index 0000000..521c3d9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Identity, Identity>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|      Reduce|      ReduceInEleOp|      ReduceAccEleOp|        Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise|   Operation|                   |                    |    MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|            |                   |                    |     Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |            |            |            |            |                   |                    |              |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   4,   4,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
new file mode 100644
index 0000000..231d612
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Identity, Identity>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|      Reduce|      ReduceInEleOp|      ReduceAccEleOp|        Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise|   Operation|                   |                    |    MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|            |                   |                    |     Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |            |            |            |            |                   |                    |              |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
new file mode 100644
index 0000000..165bc39
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Identity, Identity>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|      Reduce|      ReduceInEleOp|      ReduceAccEleOp|        Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise|   Operation|                   |                    |    MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|            |                   |                    |     Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |            |            |            |            |                   |                    |              |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
new file mode 100644
index 0000000..832fc3b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Identity, Identity>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|      Reduce|      ReduceInEleOp|      ReduceAccEleOp|        Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise|   Operation|                   |                    |    MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|            |                   |                    |     Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |            |            |            |            |                   |                    |              |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt
new file mode 100644
index 0000000..29fce56
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt
@@ -0,0 +1,4 @@
+add_instance_library(device_batched_gemm_softmax_gemm_instance
+    device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
+)
+
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
new file mode 100644
index 0000000..99e8712
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+template <bool Masking>
+using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
+    std::tuple<
+        // clang-format off
+        //#######################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|    MaskOut|
+        //#######################################|        |         |         |        |  Type|   Type|   Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|      Upper|
+        //#######################################|        |         |         |        |      |       |       |      |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|   Triangle|
+        //#######################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |           |
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    2,   32,   32,     1,     8,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8,   Masking>,
+        // Padded fallback kernel
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>
+        // clang-format on
+        >;
+
+template <bool Masking>
+using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances =
+    std::tuple<
+        // clang-format off
+        //#######################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|    MaskOut|
+        //#######################################|        |         |         |        |  Type|   Type|   Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|      Upper|
+        //#######################################|        |         |         |        |      |       |       |      |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|   Triangle|
+        //#######################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |           |
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    256,   128,    40,    64,    32,   4,   4,    2,   32,   32,     2,     4,     2,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    256,   128,    40,   128,    32,   4,   4,    2,   32,   32,     2,     4,     4,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   256,    40,    64,    32,   4,   4,    2,   32,   32,     1,     8,     2,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   256,    40,   128,    32,   4,   4,    2,   32,   32,     1,     8,     4,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    40,    64,    32,   4,   4,    2,   32,   32,     1,     4,     2,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    40,   128,    32,   4,   4,    2,   32,   32,     1,     4,     4,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemm<Row,
+                                                             Col,
+                                                             Row,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             Scale,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             false>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
+            false>{});
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances<
+            false>{});
+}
+
+void add_device_batched_gemm_masking_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemm<Row,
+                                                             Col,
+                                                             Row,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             Scale,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             true>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
+            true>{});
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances<
+            true>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt
new file mode 100644
index 0000000..76121ff
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_instance_library(device_batched_gemm_softmax_gemm_permute_instance
+    device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
+    device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
+)
+
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
new file mode 100644
index 0000000..53ad7ba
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
@@ -0,0 +1,133 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          MaskingSpecialization MaskingSpec>
+using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances =
+    std::tuple<
+        // clang-format off
+        // #############################################|  NumDimG| NumDimM| NumDimN| NumDimK| NumDimO|  AData|  B0Data|  B1Data|  CData| Acc0BiasData| Acc1BiasData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM|   ATensorSpec|  B0TensorSpec|  B1TensorSpec|   CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer| MaskingSpec|
+        // #############################################|         |        |        |        |        |   Type|    Type|    Type|   Type|         Type|         Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization|              |              |              |              | Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            |
+        // #############################################|         |        |        |        |        |       |        |        |       |             |             |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |              |              |              |              |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            |
+        // #############################################|         |        |        |        |        |       |        |        |       |             |             |        |         |            |            |            |            |            |               |              |              |              |              |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |            |
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    2,   32,   32,     1,     8,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        // Padded fallback kernel
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            ck::Tuple<>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            Scale,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskOutUpperTriangle>{});
+}
+
+void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            ck::Tuple<>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            Scale,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskDisabled>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
new file mode 100644
index 0000000..21da689
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
@@ -0,0 +1,133 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          MaskingSpecialization MaskingSpec>
+using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
+    std::tuple<
+        // clang-format off
+        // #############################################|  NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| AData| B0Data| B1Data| CData| Acc0BiasData| Acc1BiasData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM|   ATensorSpec|  B0TensorSpec|  B1TensorSpec|   CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer| MaskingSpec|
+        // #############################################|         |        |        |        |        |  Type|   Type|   Type|  Type|         Type|         Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization|              |              |              |              | Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            |
+        // #############################################|         |        |        |        |        |      |       |       |      |             |             |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |              |              |              |              |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            |
+        // #############################################|         |        |        |        |        |      |       |       |      |             |             |        |         |            |            |            |            |            |               |              |              |              |              |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |            |
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    2,   32,   32,     1,     8,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        // Padded fallback kernel
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            ck::Tuple<>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            Scale,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskOutUpperTriangle>{});
+}
+
+void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            ck::Tuple<>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            Scale,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskDisabled>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt
new file mode 100644
index 0000000..d12a2f2
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt
@@ -0,0 +1,10 @@
+add_instance_library(device_batchnorm_instance
+    device_batchnorm_forward_f16_instance.cpp
+    device_batchnorm_forward_f32_instance.cpp
+    device_batchnorm_forward_bf16_instance.cpp
+    device_batchnorm_forward_f64_instance.cpp
+    device_batchnorm_backward_f16_instance.cpp
+    device_batchnorm_backward_f32_instance.cpp
+    device_batchnorm_backward_bf16_instance.cpp
+    device_batchnorm_backward_f64_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp
new file mode 100644
index 0000000..b62c8b9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp
@@ -0,0 +1,146 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_bf16_blockwise_instances =
+     std::tuple <
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, DscaleDbiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcVectorSize, DscaleDbiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+     >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_bf16_multiblock_instances =
+     std::tuple <
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcDstVectorSize, BiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_backward_rank_4_3_bf16_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<BF16, F32, F32, F32, BF16, F32, F32, PassThrough, 4, 3>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_bf16_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_bf16_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp
new file mode 100644
index 0000000..d05b8b5
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_f16_blockwise_instances =
+     std::tuple <
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, DscaleDbiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcVectorSize, DscaleDbiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+     >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_f16_multiblock_instances =
+     std::tuple <
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcDstVectorSize, BiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_backward_rank_4_3_f16_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormBwd<F16, F32, F32, F32, F16, F32, F32, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_f16_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_f16_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp
new file mode 100644
index 0000000..e3ef95d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_f32_blockwise_instances = std::tuple<
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, DscaleDbiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcVectorSize, DscaleDbiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+    >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_f32_multiblock_instances =
+     std::tuple <
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcDstVectorSize, BiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_backward_rank_4_3_f32_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormBwd<F32, F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_f32_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_f32_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp
new file mode 100644
index 0000000..41be396
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_f64_blockwise_instances = std::tuple<
+      // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, DscaleDbiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcVectorSize, DscaleDbiasDstVectorSize, MeanVarSrcVectorSize 
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+    >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_f64_multiblock_instances =
+     std::tuple <
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcDstVectorSize, BiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_backward_rank_4_3_f64_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormBwd<F64, F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_f64_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_f64_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp
new file mode 100644
index 0000000..cd1e05b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_bf16_blockwise_instances =
+     std::tuple <
+        // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize 
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+     >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_bf16_multiblock_instances =
+     std::tuple <
+        // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_forward_rank_4_3_bf16_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<BF16, BF16, F32, BF16, BF16, F32, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_bf16_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_bf16_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp
new file mode 100644
index 0000000..073dd58
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_f16_blockwise_instances =
+     std::tuple <
+        // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize 
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+     >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_f16_multiblock_instances =
+     std::tuple <
+        // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_forward_rank_4_3_f16_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F16, F16, F32, F16, F16, F32, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_f16_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_f16_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp
new file mode 100644
index 0000000..be63bd4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_f32_blockwise_instances = std::tuple<
+    // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+    >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_f32_multiblock_instances =
+     std::tuple <
+        // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_forward_rank_4_3_f32_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_f32_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_f32_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp
new file mode 100644
index 0000000..fe87091
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_f64_blockwise_instances = std::tuple<
+    // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+    >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_f64_multiblock_instances =
+     std::tuple <
+        // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_forward_rank_4_3_f64_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_f64_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_f64_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt
new file mode 100644
index 0000000..ffd6a6a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt
@@ -0,0 +1,7 @@
+add_instance_library(device_contraction_bilinear_instance
+    device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp
+    device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp
+    device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp
+    device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp
+)
+
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp
new file mode 100644
index 0000000..ebbff88
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32       = float;
+using F32_Tuple = ck::Tuple<F32>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// k/k/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |          |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp
new file mode 100644
index 0000000..980383f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32       = float;
+using F32_Tuple = ck::Tuple<F32>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// k/n/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |          |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   1,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   1,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   1,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   1,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp
new file mode 100644
index 0000000..2d4b6e3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32       = float;
+using F32_Tuple = ck::Tuple<F32>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// m/k/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|  CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|          _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|          _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |          |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                              |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   1,   4,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   1,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   1,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   1,   4,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   1,   4,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   1,   4,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp
new file mode 100644
index 0000000..7caa469
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32       = float;
+using F32_Tuple = ck::Tuple<F32>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// m/n/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |          |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   1,   1,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   1,   1,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   1,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   1,   1,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   1,   1,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   1,   1,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt
new file mode 100644
index 0000000..7ad6605
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt
@@ -0,0 +1,7 @@
+add_instance_library(device_contraction_scale_instance
+    device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
+    device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
+    device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
+    device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
+)
+
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
new file mode 100644
index 0000000..5118d0d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
@@ -0,0 +1,78 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32         = float;
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] = E[m0, m1, n0, n1]
+// k/k/n are the fast changing dimension for A/B/E
+using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
new file mode 100644
index 0000000..655d4f0
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32         = float;
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] = E[m0, m1, n0, n1]
+// k/n/n are the fast changing dimension for A/B/E
+using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   1,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   1,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   1,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   1,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
new file mode 100644
index 0000000..a9d20be
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32         = float;
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] = E[m0, m1, n0, n1]
+// m/k/n are the fast changing dimension for A/B/E
+using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|      DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|  CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|        Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|          _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |            |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|          _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |            |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                              |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   1,   4,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   1,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   1,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   1,   4,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   1,   4,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   1,   4,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
new file mode 100644
index 0000000..a68f5c9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32         = float;
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] = E[m0, m1, n0, n1]
+// m/n/n are the fast changing dimension for A/B/E
+using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|      DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|        Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |            |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |            |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   1,   1,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   1,   1,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   1,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   1,   1,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   1,   1,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   1,   1,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000..75a3670
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_conv1d_bwd_data_instance
+   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp
+   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp
+   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp
+   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp
new file mode 100644
index 0000000..5a5c838
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = bhalf_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NWC = ck::tensor_layout::convolution::NWC;
+using KXC = ck::tensor_layout::convolution::KXC;
+using NWK = ck::tensor_layout::convolution::NWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_bf16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<1,
+                                                  NWC,
+                                                  KXC,
+                                                  NWK,
+                                                  BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances{});
+    add_device_operation_instances(
+        instances, device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp
new file mode 100644
index 0000000..e0f3d61
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp
@@ -0,0 +1,95 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NWC = ck::tensor_layout::convolution::NWC;
+using KXC = ck::tensor_layout::convolution::KXC;
+using NWK = ck::tensor_layout::convolution::NWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,   8,  32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,   8,  32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,   8,  32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,   8,  32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,       
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,   8,  32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,   8,  32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,   8,  32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,   8,  32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvBwdData<1, NWC, KXC, NWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp
new file mode 100644
index 0000000..30537d9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp
@@ -0,0 +1,94 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NWC = ck::tensor_layout::convolution::NWC;
+using KXC = ck::tensor_layout::convolution::KXC;
+using NWK = ck::tensor_layout::convolution::NWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|        ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|      Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                    |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                 |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvBwdData<1, NWC, KXC, NWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances{});
+    add_device_operation_instances(
+        instances, device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp
new file mode 100644
index 0000000..190c39b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp
@@ -0,0 +1,99 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NWC = ck::tensor_layout::convolution::NWC;
+using KXC = ck::tensor_layout::convolution::KXC;
+using NWK = ck::tensor_layout::convolution::NWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_int8_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|    ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |    |     |     | Wave| Wave|  Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |    |     |     |     |     |                 |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  16,   32,   32,    4,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  16,   32,   32,    2,    4,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  16,   32,   32,    4,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  16,   32,   32,    2,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  16,   32,   32,    2,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  16,   32,   32,    2,    1,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  16,   32,   32,    1,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  16,   32,   32,    2,    1,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  16,   32,   32,    1,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  16,   32,   32,    2,    1,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  16,   32,   32,    1,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<1,
+                                                  NWC,
+                                                  KXC,
+                                                  NWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances{});
+    add_device_operation_instances(
+        instances, device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000..281453b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt
@@ -0,0 +1,10 @@
+add_instance_library(device_conv2d_bwd_data_instance
+   device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
+   device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
+   device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
+   device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
+
+   device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp
+   device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp
+   device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp
+) 
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp
new file mode 100644
index 0000000..e14cd55
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //#########################|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|        Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //#########################|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|            Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //#########################|           |           |            |            |            |    Operation|     Operation|     Operation|     Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //#########################|           |           |            |            |            |             |              |              |                   |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+    DeviceConvNdBwdDataNwcKxcNwk_Dl<          2, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDataDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 1, 8, 2>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 8, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //#########################|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|                     Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //#########################|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|                         Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //#########################|           |           |            |            |            |    Operation|     Operation|     Operation|                  Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //#########################|           |           |            |            |            |             |              |              |                                |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+    DeviceConvNdBwdDataNwcKxcNwk_Dl<          2, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 1, 8, 2>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 8, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  InDataType,
+                                                  WeiDataType,
+                                                  OutDataType,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp
new file mode 100644
index 0000000..f001b83
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using InDataType  = float;
+using WeiDataType = float;
+using OutDataType = float;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances = std::tuple<
+    // clang-format off
+        //#########################|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|        Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //#########################|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|            Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //#########################|           |           |            |            |            |    Operation|     Operation|     Operation|     Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //#########################|           |           |            |            |            |             |              |              |                   |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+    DeviceConvNdBwdDataNwcKxcNwk_Dl<          2, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDataDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 1, 8, 1>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //#########################|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|                     Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //#########################|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|                         Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //#########################|           |           |            |            |            |    Operation|     Operation|     Operation|                  Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //#########################|           |           |            |            |            |             |              |              |                                |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+    DeviceConvNdBwdDataNwcKxcNwk_Dl<          2, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 1, 8, 1>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  InDataType,
+                                                  WeiDataType,
+                                                  OutDataType,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp
new file mode 100644
index 0000000..83ba6a1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using InDataType  = int8_t;
+using WeiDataType = int8_t;
+using OutDataType = int8_t;
+using AccDataType = int32_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances = std::tuple<
+    // clang-format off
+        //#########################|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|        Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //#########################|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|            Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //#########################|           |           |            |            |            |    Operation|     Operation|     Operation|     Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //#########################|           |           |            |            |            |             |              |              |                   |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+    DeviceConvNdBwdDataNwcKxcNwk_Dl<          2, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDataDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<1, 1, 8, 4>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 8, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances = std::tuple<
+    // clang-format off
+        //#########################|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|                     Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //#########################|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|                         Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //#########################|           |           |            |            |            |    Operation|     Operation|     Operation|                  Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //#########################|           |           |            |            |            |             |              |              |                                |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+    DeviceConvNdBwdDataNwcKxcNwk_Dl<          2, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<1, 1, 8, 4>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 8, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  InDataType,
+                                                  WeiDataType,
+                                                  OutDataType,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
new file mode 100644
index 0000000..1da9a81
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
@@ -0,0 +1,157 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+// FIXME: retire dedicated 2D version
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = bhalf_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances = std::tuple<
+    // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|        ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|      Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                    |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+using device_conv_dedidecate_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances =
+    std::tuple<
+        // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                                 |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_conv_dedidecate_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
new file mode 100644
index 0000000..7c33df5
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+// FIXME: retire dedicated 2D version
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,   8,  32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,   8,  32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,   8,  32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,   8,  32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,       
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,   8,  32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,   8,  32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,   8,  32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,   8,  32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,   8,  32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+#if !CK_WORKAROUND_SWDEV_325164
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,   8,  32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,   8,  32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+#endif
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,   8,  32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,   8,  32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,   8,  32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,   8,  32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,   8,  32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances =
+    std::tuple<
+        // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                                 |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F16,
+                                                  F16,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
new file mode 100644
index 0000000..a5f8629
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
@@ -0,0 +1,155 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+// FIXME: retire dedicated 2D version
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|        ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|      Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                    |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                 |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances = std::tuple<
+    // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|        ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|      Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                    |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances =
+    std::tuple<
+        // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                                 |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F32,
+                                                  F32,
+                                                  F32,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
new file mode 100644
index 0000000..8076d6d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
@@ -0,0 +1,153 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+// FIXME: retire dedicated 2D version
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|    ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |    |     |     | Wave| Wave|  Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |    |     |     |     |     |                 |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  16,   32,   32,    4,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  16,   32,   32,    2,    4,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  16,   32,   32,    4,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  16,   32,   32,    2,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  16,   32,   32,    2,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  16,   32,   32,    2,    1,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  16,   32,   32,    1,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  16,   32,   32,    2,    1,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  16,   32,   32,    1,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  16,   32,   32,    2,    1,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  16,   32,   32,    1,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances = std::tuple<
+    // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|        ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|      Specialization|      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                    |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances =
+    std::tuple<
+        // clang-format off
+        //#####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //#####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|    ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //#####################################################################|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |    |     |     | Wave| Wave|  Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //#####################################################################|       |        |        |        |            |            |            |                                 |      |      |      |      |    |     |     |     |     |                 |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  16,   32,   32,    4,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  16,   32,   32,    2,    4,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  16,   32,   32,    4,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  16,   32,   32,    2,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  16,   32,   32,    2,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  16,   32,   32,    2,    1,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  16,   32,   32,    1,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  16,   32,   32,    2,    1,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  16,   32,   32,    1,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  16,   32,   32,    2,    1,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  16,   32,   32,    1,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt
new file mode 100644
index 0000000..5b64685
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt
@@ -0,0 +1,7 @@
+add_instance_library(device_conv2d_fwd_instance
+   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
+   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
+   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
+   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
+   device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
new file mode 100644
index 0000000..33503b9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
@@ -0,0 +1,155 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto ConvFwdOddC =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
+
+// arbitrary conv
+using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// 1x1, pad 0
+using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// 1x1, stride 1, pad 0
+using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_odd_c_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   128,    64,     2,  4,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   256,    64,     2,  4,   32,   32,    4,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,   128,    64,     2,  4,   32,   32,    2,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,    64,    64,     2,  4,   32,   32,    1,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvFwd<2, NHWC, KYXC, NHWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_odd_c_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
new file mode 100644
index 0000000..c5e4bd1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
@@ -0,0 +1,128 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_bf16_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvFwd<2,
+                                              NHWC,
+                                              KYXC,
+                                              NHWK,
+                                              BF16,
+                                              BF16,
+                                              BF16,
+                                              PassThrough,
+                                              PassThrough,
+                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_bf16_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
new file mode 100644
index 0000000..f43d13e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
@@ -0,0 +1,120 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvFwd<2, NHWC, KYXC, NHWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_f16_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
new file mode 100644
index 0000000..0ce6b04
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
@@ -0,0 +1,119 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvFwd<2, NHWC, KYXC, NHWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_f32_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
new file mode 100644
index 0000000..76ab318
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
@@ -0,0 +1,125 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData|  WeiData| OutData|  AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|     Type|    Type|     Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |         |        |         |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |         |        |         |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_int8_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData|  WeiData| OutData|  AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|     Type|    Type|     Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |         |        |         |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |         |        |         |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData|  WeiData| OutData|  AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|     Type|    Type|     Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |         |        |         |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |         |        |         |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvFwd<2,
+                                              NHWC,
+                                              KYXC,
+                                              NHWK,
+                                              int8_t,
+                                              int8_t,
+                                              int8_t,
+                                              PassThrough,
+                                              PassThrough,
+                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_int8_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt
new file mode 100644
index 0000000..670cd94
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_instance_library(device_conv2d_fwd_bias_relu_instance
+   device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
new file mode 100644
index 0000000..f8c2550
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
@@ -0,0 +1,154 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddRelu     = ck::tensor_operation::element_wise::AddRelu;
+
+static constexpr auto MemorySet = ck::InMemoryDataOperationEnum::Set;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto ConvFwdOddC =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
+
+// arbitrary conv
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|           Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|  GlobalMemory| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################################|       |        |        |        |   Operation|   Operation|   Operation| DataOperation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################################|       |        |        |        |            |            |            |              |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// 1x1, pad 0
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|           Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|  GlobalMemory| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################################|       |        |        |        |   Operation|   Operation|   Operation| DataOperation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################################|       |        |        |        |            |            |            |              |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// 1x1, stride 1, pad 0
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|           Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|  GlobalMemory| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################################|       |        |        |        |   Operation|   Operation|   Operation| DataOperation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################################|       |        |        |        |            |            |            |              |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// Odd C
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_odd_c_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|           Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|  GlobalMemory| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################################|       |        |        |        |   Operation|   Operation|   Operation| DataOperation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################################|       |        |        |        |            |            |            |              |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   128,    64,     2,  4,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   256,    64,     2,  4,   32,   32,    4,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,   128,    64,     2,  4,   32,   32,    2,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,    64,    64,     2,  4,   32,   32,    1,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<DeviceConvFwdBiasActivationPtr<PassThrough, PassThrough, AddRelu>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_odd_c_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt
new file mode 100644
index 0000000..68d5f58
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt
@@ -0,0 +1,4 @@
+add_instance_library(device_conv2d_fwd_bias_relu_add_instance
+   device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
+)
+
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
new file mode 100644
index 0000000..fe71524
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
@@ -0,0 +1,154 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto ConvFwdOddC =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
+
+// arbitrary conv
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //##############################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##############################################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// 1x1, pad 0
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##############################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##############################################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// 1x1, stride 1, pad 0
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##############################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##############################################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// Odd C
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_odd_c_f16_instances = std::tuple<
+    // clang-format off
+        //##############################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##############################################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   128,    64,     2,  4,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   256,    64,     2,  4,   32,   32,    4,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,   128,    64,     2,  4,   32,   32,    2,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,    64,    64,     2,  4,   32,   32,    1,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_odd_c_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000..db92208
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_conv3d_bwd_data_instance
+   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp
+   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp
+   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp
+   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp
new file mode 100644
index 0000000..04ce7c0
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = bhalf_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+using KZYXC = ck::tensor_layout::convolution::KZYXC;
+using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_bf16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances{});
+    add_device_operation_instances(
+        instances, device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp
new file mode 100644
index 0000000..0251d91
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+using KZYXC = ck::tensor_layout::convolution::KZYXC;
+using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,   8,  32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,   8,  32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,   8,  32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,   8,  32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,       
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,   8,  32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,   8,  32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,   8,  32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,   8,  32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  F16,
+                                                  F16,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp
new file mode 100644
index 0000000..c297572
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp
@@ -0,0 +1,101 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+using KZYXC = ck::tensor_layout::convolution::KZYXC;
+using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|        ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|      Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                    |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                 |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  F32,
+                                                  F32,
+                                                  F32,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances{});
+    add_device_operation_instances(
+        instances, device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp
new file mode 100644
index 0000000..fc86d73
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp
@@ -0,0 +1,99 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+using KZYXC = ck::tensor_layout::convolution::KZYXC;
+using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_int8_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|    ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |    |     |     | Wave| Wave|  Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |    |     |     |     |     |                 |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  16,   32,   32,    4,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  16,   32,   32,    2,    4,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  16,   32,   32,    4,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  16,   32,   32,    2,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  16,   32,   32,    2,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  16,   32,   32,    2,    1,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  16,   32,   32,    1,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  16,   32,   32,    2,    1,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  16,   32,   32,    1,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  16,   32,   32,    2,    1,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  16,   32,   32,    1,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances{});
+    add_device_operation_instances(
+        instances, device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt
new file mode 100644
index 0000000..47516b4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_instance_library(device_elementwise_instance
+    device_normalize_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp
new file mode 100644
index 0000000..baddecf
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using inputType      = F16;
+using MeanType       = F32;
+using SquareMeanType = F32;
+using GammaDataType  = F16;
+using BetaDataType   = F16;
+using outputType     = F16;
+
+using Normalize = ck::tensor_operation::element_wise::Normalize;
+using device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances = std::tuple<
+    // clang-format off
+    //###################|<in, mean, square_mean, gamma, beta>| <out>|  functor| NDim| MPerThread| <in, mean, square_mean, gamma, beta ScalarPerVector>| <out ScalarPerVector>|
+    DeviceElementwise<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   8,       Sequence<8, 1, 1, 8, 8>,      Sequence<8>                >,
+    DeviceElementwise<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   4,       Sequence<4, 1, 1, 4, 4>,      Sequence<4>                >,
+    DeviceElementwise<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   2,       Sequence<2, 1, 1, 2, 2>,      Sequence<2>                >,
+    DeviceElementwise<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   1,       Sequence<1, 1, 1, 1, 1>,      Sequence<1>                >
+    // clang-format on
+    >;
+
+void add_device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances(
+    std::vector<DeviceElementwiseBasePtr<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, 2>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt
new file mode 100644
index 0000000..0c7cc2c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_instance_library(device_elementwise_normalization_instance
+    device_elementwise_normalization_f16_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp
new file mode 100644
index 0000000..7f15372
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp
@@ -0,0 +1,54 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Add  = ck::tensor_operation::element_wise::Add;
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+template <typename XElementwise, typename YElementwise, index_t Rank, index_t Reduce>
+// clang-format off
+using device_elementwise_normalization_f16_instances =
+    std::tuple <
+        // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>, // fallback kernel
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 1024, 1, 1024, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 2, 1, 2, 1, 2, 2>
+    >;
+// clang-format on
+
+void add_device_elementwise_normalization_rank_2_1_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwiseNormalization<ck::Tuple<F16, F16>, F16, F16, F32, F16, Add, Pass, 2, 1>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_elementwise_normalization_f16_instances<Add, Pass, 2, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
new file mode 100644
index 0000000..e20d592
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
@@ -0,0 +1,43 @@
+add_instance_library(device_gemm_instance
+   device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp
+   device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp
+   device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp
+   device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp
+   device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp
+   device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp
+   device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp
+   device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp
+   device_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
+   device_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
+   device_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp
+   device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
+   device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp
+   device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp
+   device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp
+   device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp
+   device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp
+   device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp
+   device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp
+   device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp
+   device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp
+   device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp
+   device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..5d2f18e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_dl_f16_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        // #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // #########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // #########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // #########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..01e3b37
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_dl_f16_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        // #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // #########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // #########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        // #########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..804e86a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_dl_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        // #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // #########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // #########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        // #########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..159fa90
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_dl_f16_f16_f16_mk_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+        // clang-format on
+        >;
+
+void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..d8e7798
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_dl_f32_f32_f32_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                    |                    |                 |                   |
+        DeviceGemmDl<   F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 1>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<2, 1, 4, 1>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 1>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f32_f32_f32_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..0034ac5
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_dl_f32_f32_f32_km_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl<   F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 1>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,        S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+        // clang-format on
+        >;
+
+void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f32_f32_f32_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..0b540b8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_dl_f32_f32_f32_mk_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl<   F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,        S<1, 1, 1, 1>,      S<2, 1, 4, 1>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 1>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+        // clang-format on
+        >;
+
+void add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f32_f32_f32_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..4f6ff51
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_dl_f32_f32_f32_mk_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl<   F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,        S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,        S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+        // clang-format on
+        >;
+
+void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f32_f32_f32_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..a420824
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_dl_i8_i8_i8_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_i8_i8_i8_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..06fab7f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_dl_i8_i8_i8_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_i8_i8_i8_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..b6d72fa
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_dl_i8_i8_i8_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                |                   |
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_i8_i8_i8_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..67d2e3c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_dl_i8_i8_i8_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_i8_i8_i8_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..03eebf4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,100 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                     Version|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..5d8de04
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..7b12b7c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..730ffd4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..619473f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..8e06f9d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,111 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..f9458b7
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,111 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..77a03b7
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,111 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..ef8d7d4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..cb65cc7
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   1,   1,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   1,   1,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   1,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   1,   1,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   1,   1,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   1,   1,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..5b1014e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|  CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|          _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|          _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                              |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   1,   4,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   1,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   1,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   1,   4,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   1,   4,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   1,   4,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..e6f6add
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   4,   1,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   4,   1,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   4,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   4,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   4,   1,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   4,   1,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..80b3d03
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,64 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..93b3df1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //#####################| ALayout| BLayout| CLayout|  AData|  BData|  CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |   Type|   Type|   Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |       |       |       |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |       |       |       |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,   4,   4,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,   4,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,   4,   4,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,   4,   4,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,   4,   4,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..f10365d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //#####################| ALayout| BLayout| CLayout|  AData|  BData|  CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |   Type|   Type|   Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |       |       |       |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |       |       |       |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,   4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,   4,  16,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,   4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,   4,  16,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,   4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,   4,  16,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,   4,  16,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,   4,  16,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..a7a9eb6
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //#####################| ALayout| BLayout| CLayout|  AData|  BData|  CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |   Type|   Type|   Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |       |       |       |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |       |       |       |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,  16,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,  16,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,  16,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,  16,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,  16,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,  16,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,  16,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,  16,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..9fb45b0
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //#####################| ALayout| BLayout| CLayout|  AData|  BData|  CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |   Type|   Type|   Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |       |       |       |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |       |       |       |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    64,  16,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    64,  16,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    64,  16,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 2>,             16>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..18a7867
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,110 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault   = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_f16_f16_f16_km_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler| Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |         |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |         |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |         |
+        // pipeline v1, 1 wave
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+// irregular tile size
+using device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        // pipeline v1, 1 wave
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   64,     16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   64,     16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   64,     16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f16_f16_f16_km_kn_mn_instances{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..cef6070
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,110 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault   = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_f16_f16_f16_km_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        // pipeline v1, 1 wave
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+// irregular tile size
+using device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        // pipeline v1, 1 wave
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmMNPadding,   64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmMNPadding,   64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmMNPadding,   64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f16_f16_f16_km_nk_mn_instances{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..1be70d6
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault   = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        // pipeline v1, 1 wave
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,   256,     4,  8,   16,   16,    1,    8,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,   128,     4,  8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,    64,     4,  8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,   256,     4,  8,   16,   16,    1,    8,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,   128,     4,  8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,    64,     4,  8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,   256,     4,  8,   16,   16,    1,    8,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,   128,     4,  8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,    64,     4,  8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+// irregular tile size
+using device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        // pipeline v1, 1 wave
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..6b8455f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,130 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault   = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        // pipeline v1, 1 wave
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        // pipeline v1, 1 wave
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   144,     8,  8,   16,   16,    2,    9,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8,  8, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   144,     4,  8,   16,   16,    2,    9,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   144,     8,  8,   16,   16,    2,    9,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8,  8, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   144,     4,  8,   16,   16,    2,    9,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   144,     8,  8,   16,   16,    2,    9,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8,  8, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   144,     4,  8,   16,   16,    2,    9,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..b9e28e3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_f32_f32_f32_km_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f32_f32_f32_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..2b1a5a5
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_f32_f32_f32_km_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f32_f32_f32_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..301d3b5
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..cd16f35
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..3916669
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp
@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_f64_f64_f64_km_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  2,   16,   16,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  2,   16,   16,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              2,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f64_f64_f64_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..0a62303
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp
@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_f64_f64_f64_km_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  2,   16,   16,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  2,   16,   16,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f64_f64_f64_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f64_f64_f64_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..5ef8d08
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp
@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  2,   16,   16,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  2,   16,   16,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..c9557ba
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  2,   16,   16,    4,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  2,   16,   16,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  2,   16,   16,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  2,   16,   16,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  2,   16,   16,    2,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  2,   16,   16,    4,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  2,   16,   16,    2,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt
new file mode 100644
index 0000000..bbf81a5
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_gemm_add_add_fastgelu_instance
+   device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
+   device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
+   device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
+   device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000..463e086
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
@@ -0,0 +1,153 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[k, m], b[k, n], d0[m, n], d1[m, n]
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+// irregular tile size
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000..b71ff1b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
@@ -0,0 +1,153 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[k, m], b[n, k], d0[m, n], d1[m, n]
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+// irregular tile size
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,      PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000..9060c9b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
@@ -0,0 +1,153 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+// irregular tile size
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000..81cf01d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
@@ -0,0 +1,144 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[m, k], b[n, k], d0[m, n], d1[m ,n]
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES         
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+// irregular tile size
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt
new file mode 100644
index 0000000..0beb10e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_gemm_add_fastgelu_instance
+   device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
+   device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
+   device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
+   device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
new file mode 100644
index 0000000..4da85cc
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0)
+// outout: e[m, n]
+// input: a[k, m], b[k, n], d0[m, n]
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,      PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
new file mode 100644
index 0000000..ab83e4b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[k, m], b[n, k], d0[m, n], d1[m, n]
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,      PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
new file mode 100644
index 0000000..a4cd3fa
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
new file mode 100644
index 0000000..207e76f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
@@ -0,0 +1,127 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[m, k], b[n, k], d0[m, n], d1[m ,n]
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES         
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt
new file mode 100644
index 0000000..ccada3a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_gemm_bias_add_reduce_instance
+    device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
+    device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
+    device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
+    device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..e8747af
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[k, m] * b[k, n]
+using device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|          C1|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |                   |   Operation|   Operation|   Operation|   Operation|          |                   |                    |   Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |                   |            |            |            |            |          |                   |                    |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+    std::vector<DeviceGemmReducePtr<1, ReduceOps::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..ed54c3a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[k, m] * b[n, k]
+using device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|          C1|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |                   |   Operation|   Operation|   Operation|   Operation|          |                   |                    |   Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |                   |            |            |            |            |          |                   |                    |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+    std::vector<DeviceGemmReducePtr<1, ReduceOps::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..da7eae6
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[m, k] * b[n, k]
+using device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|          C1|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |                   |   Operation|   Operation|   Operation|   Operation|          |                   |                    |   Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |                   |            |            |            |            |          |                   |                    |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+    std::vector<DeviceGemmReducePtr<1, ReduceOps::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..3434509
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[m, k] * b[n, k]
+using device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data|  GemmAcc| CShuffle| ReduceAcc|         DData|           A|           B|           C|          C1|         Dxs|    DxsInEleOp|    DxsAccEleOp|           D|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type|  Type|  Type| DataType| DataType|  DataType|    Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise|      Reduce|              |               |  MemoryData| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |              |   Operation|   Operation|   Operation|   Operation|   Operation|              |               |   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |              |            |            |            |            |            |              |               |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+    std::vector<DeviceGemmReducePtr<1, ReduceOps::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt
new file mode 100644
index 0000000..cb1b3a4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_gemm_bilinear_instance
+   device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
+   device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
+   device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
+   device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
new file mode 100644
index 0000000..55461df
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
@@ -0,0 +1,105 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<ck::half_t>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e[m, n] = bilinear(a[k, m] * b[k, n], d[m, n])
+using device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances = std::tuple<
+    // clang-format off
+        // no padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+
+        // M/N/K Padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
new file mode 100644
index 0000000..405e699
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
@@ -0,0 +1,105 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<ck::half_t>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e[m, n] = bilinear(a[k, m] * b[n, k], d[m, n])
+using device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances = std::tuple<
+    // clang-format off
+        // no padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+
+        // M/N/K Padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
new file mode 100644
index 0000000..9af31b3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
@@ -0,0 +1,105 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<ck::half_t>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e[m, n] = bilinear(a[m, k] * b[k, n], d[m, n])
+using device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances = std::tuple<
+    // clang-format off
+        // no padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+
+        // M/N/K padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
new file mode 100644
index 0000000..841b7a1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
@@ -0,0 +1,142 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<ck::half_t>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e[m, n] = bilinear(a[m, k] * b[n, k], d[m, n])
+using device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances = std::tuple<
+    // clang-format off
+        // no padding
+        // N % 8 == 0 && K % 8 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // M/N/K padding
+        // N % 8 == 0 && K % 8 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // M/N/K padding
+        // N % 4 == 0 && K % 4 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+
+        // M/N/K padding
+        // N % 8 == 0 && K % 1 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>
+
+    // clang-format on
+    >;
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt
new file mode 100644
index 0000000..17d27ab
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_gemm_fastgelu_instance
+   device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
+   device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
+   device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..9f7f643
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,135 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b))
+// outout: e[m, n]
+// input: a[k, m], b[k, n]
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,      PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..c8e9f35
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,135 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b))
+// outout: e[m, n]
+// input: a[k, m], b[k, n]
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..5f804d4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,135 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b))
+// outout: e[m, n]
+// input: a[m, k], b[k, n]
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,      PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..60cb138
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,126 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b))
+// outout: e[m, n]
+// input: a[m, k], b[n, k]
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES         
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,      PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt
new file mode 100644
index 0000000..2b2cf8c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_gemm_reduce_instance
+    device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
+    device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
+    device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
+    device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..59e2b2d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[k, m] * b[k, n]
+using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //###########################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //###########################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|          |                   |                    |   Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |            |            |            |          |                   |                    |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+    // clang-format on
+    >;
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..bb09bf8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[k, m] * b[n, k]
+using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //###########################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //###########################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|          |                   |                    |   Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |            |            |            |          |                   |                    |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+    // clang-format on
+    >;
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..0a3b566
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[m, k] * b[n, k]
+using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //###########################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //###########################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|          |                   |                    |   Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |            |            |            |          |                   |                    |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+    // clang-format on
+    >;
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..2b17e47
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[m, k] * b[n, k]
+using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //###########################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //###########################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|          |                   |                    |   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |            |            |            |          |                   |                    |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>
+    // clang-format on
+    >;
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt
new file mode 100644
index 0000000..6b33622
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt
@@ -0,0 +1,10 @@
+add_instance_library(device_gemm_splitk_instance
+   device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp
+   device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp
+   device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp
+   device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp
+   device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
+   device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp
+   device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..e178d3b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              1,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..52be9fe
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              1,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..9b5ff40
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..7fc35c4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,66 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..f27b219
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              4,      true,           1,           1,                   S<1, 16, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              1,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..b9a1095
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              1,              4,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..44e5f59
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              4,      true,           1,           1,                   S<1, 16, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..f3a9063
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,66 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 4>,               4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt
new file mode 100644
index 0000000..3808e02
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_instance_library(device_grouped_conv1d_bwd_weight_instance
+    device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
+    device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
+    device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
new file mode 100644
index 0000000..05ba449
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
@@ -0,0 +1,106 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = bhalf_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNWC = ck::tensor_layout::convolution::GNWC;
+using GKXC = ck::tensor_layout::convolution::GKXC;
+using GNWK = ck::tensor_layout::convolution::GNWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_grouped_conv1d_bwd_weight_xdl_c_shuffle_gnwc_gkxc_gnwk_bf16_f32_bf16_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|   BBlockTransfer|  BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                 |                |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 32, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+        // clang-format on
+        >;
+
+using device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_1x1_s1_p0_bf16_f32_bf16_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|   BBlockTransfer|  BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                 |                |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 32, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
+                                                           GNWC,
+                                                           GKXC,
+                                                           GNWK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv1d_bwd_weight_xdl_c_shuffle_gnwc_gkxc_gnwk_bf16_f32_bf16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_1x1_s1_p0_bf16_f32_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
new file mode 100644
index 0000000..7a610a7
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
@@ -0,0 +1,104 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNWC = ck::tensor_layout::convolution::GNWC;
+using GKXC = ck::tensor_layout::convolution::GKXC;
+using GNWK = ck::tensor_layout::convolution::GNWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_grouped_conv1d_bwd_weight_xdl_c_shuffle_gnwc_gkxc_gnwk_f16_default_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+using device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
+                                                           GNWC,
+                                                           GKXC,
+                                                           GNWK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv1d_bwd_weight_xdl_c_shuffle_gnwc_gkxc_gnwk_f16_default_instances{});
+    add_device_operation_instances(
+        instances, device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
new file mode 100644
index 0000000..90e074f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
@@ -0,0 +1,103 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNWC = ck::tensor_layout::convolution::GNWC;
+using GKXC = ck::tensor_layout::convolution::GKXC;
+using GNWK = ck::tensor_layout::convolution::GNWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_grouped_conv1d_bwd_weight_xdl_c_shuffle_gnwc_gkxc_gnwk_f32_default_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 64, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 16, 4>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  4>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 16, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+        // clang-format on
+        >;
+
+using device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_1x1_s1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 64, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,    32,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    32,    64,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+    // clang-format on
+    >;
+
+void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
+                                                           GNWC,
+                                                           GKXC,
+                                                           GNWK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv1d_bwd_weight_xdl_c_shuffle_gnwc_gkxc_gnwk_f32_default_instances{});
+    add_device_operation_instances(
+        instances, device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt
new file mode 100644
index 0000000..1d90593
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_grouped_conv1d_fwd_instance
+   device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
+   device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
+   device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
+   device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
new file mode 100644
index 0000000..74aebf1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
@@ -0,0 +1,129 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNWC = ck::tensor_layout::convolution::GNWC;
+using GKXC = ck::tensor_layout::convolution::GKXC;
+using GNWK = ck::tensor_layout::convolution::GNWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// Compilation parameters for in[g, n, wi, c] * wei[g, k, x, c] = out[g, n, wo, k]
+using device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
new file mode 100644
index 0000000..361ea8f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
@@ -0,0 +1,129 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNWC = ck::tensor_layout::convolution::GNWC;
+using GKXC = ck::tensor_layout::convolution::GKXC;
+using GNWK = ck::tensor_layout::convolution::GNWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// Compilation parameters for in[g, n, wi, c] * wei[g, k, x, c] = out[g, n, wo, k]
+using device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
new file mode 100644
index 0000000..3145b71
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
@@ -0,0 +1,128 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNWC = ck::tensor_layout::convolution::GNWC;
+using GKXC = ck::tensor_layout::convolution::GKXC;
+using GNWK = ck::tensor_layout::convolution::GNWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// Compilation parameters for in[g, n, wi, c] * wei[g, k, x, c] = out[g, n, wo, k]
+using device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp
new file mode 100644
index 0000000..cde93f9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp
@@ -0,0 +1,125 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNWC = ck::tensor_layout::convolution::GNWC;
+using GKXC = ck::tensor_layout::convolution::GKXC;
+using GNWK = ck::tensor_layout::convolution::GNWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// Compilation parameters for in[g, n, wi, c] * wei[g, k, x, c] = out[g, n, wo, k]
+using device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances = std::tuple<
+    // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+                                                                                                                                                                                                                
+        // Filter1x1Stride1Pad0                                                                                                                                                                                 
+        //########################################|  NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       1,   GNWC,   GKXC, Empty_Tuple,   GNWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000..3b2968d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_instance_library(device_grouped_conv2d_bwd_data_instance
+   device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
new file mode 100644
index 0000000..3d604d4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
@@ -0,0 +1,99 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNHWC = ck::tensor_layout::convolution::GNHWC;
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+using GNHWK = ck::tensor_layout::convolution::GNHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+using device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances = std::tuple<
+    // clang-format off
+        // 1. Default
+        // ##############################################|    NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+        // ##############################################| Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+        // ##############################################|        |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+        // ##############################################|        |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+
+        // 2. Filter1x1Stride1Pad0
+        // ##############################################|    NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise|              ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+        // ##############################################| Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|               DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+        // ##############################################|        |        |        |            |        |      |      |        |         |            |      |             |             |               |                                 |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+        // ##############################################|        |        |        |            |        |      |      |        |         |            |      |             |             |               |                                 |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>
+    // clang-format on
+    >;
+
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  GNHWK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  GNHWC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
new file mode 100644
index 0000000..4009121
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_grouped_conv2d_bwd_weight_instance
+    device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
+    device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
+    device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
+)
+
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
new file mode 100644
index 0000000..ede21f1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
@@ -0,0 +1,106 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = bhalf_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNHWC = ck::tensor_layout::convolution::GNHWC;
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+using GNHWK = ck::tensor_layout::convolution::GNHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_grouped_conv2d_bwd_weight_xdl_c_shuffle_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|   BBlockTransfer|  BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                 |                |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 32, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+        // clang-format on
+        >;
+
+using device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_1x1_s1_p0_bf16_f32_bf16_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|   BBlockTransfer|  BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                 |                |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 32, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           GNHWC,
+                                                           GKYXC,
+                                                           GNHWK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv2d_bwd_weight_xdl_c_shuffle_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_1x1_s1_p0_bf16_f32_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
new file mode 100644
index 0000000..99e5566
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
@@ -0,0 +1,105 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNHWC = ck::tensor_layout::convolution::GNHWC;
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+using GNHWK = ck::tensor_layout::convolution::GNHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_grouped_conv2d_bwd_weight_xdl_c_shuffle_gnhwc_gkyxc_gnhwk_f16_default_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+using device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           GNHWC,
+                                                           GKYXC,
+                                                           GNHWK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv2d_bwd_weight_xdl_c_shuffle_gnhwc_gkyxc_gnhwk_f16_default_instances{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
new file mode 100644
index 0000000..15871a2
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
@@ -0,0 +1,104 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNHWC = ck::tensor_layout::convolution::GNHWC;
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+using GNHWK = ck::tensor_layout::convolution::GNHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_grouped_conv2d_bwd_weight_xdl_c_shuffle_gnhwc_gkyxc_gnhwk_f32_default_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 64, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 16, 4>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  4>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 16, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+        // clang-format on
+        >;
+
+using device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_1x1_s1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 64, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,    32,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    32,    64,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+    // clang-format on
+    >;
+
+void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           GNHWC,
+                                                           GKYXC,
+                                                           GNHWK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv2d_bwd_weight_xdl_c_shuffle_gnhwc_gkyxc_gnhwk_f32_default_instances{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
new file mode 100644
index 0000000..5ef1b68
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
@@ -0,0 +1,13 @@
+add_instance_library(device_grouped_conv2d_fwd_instance
+   # GNHWC, GKYXC, GNHWK
+   device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
+   device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
+   device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
+   device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instance.cpp
+   # NHWGC, GKYXC, NHWGK
+   device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
+   #dl 
+   device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
+   device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
+   device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
new file mode 100644
index 0000000..fc18b3c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
@@ -0,0 +1,103 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using AccDataType = float;
+using OutDataType = ck::half_t;
+
+using Empty_Tuple = ck::Tuple<>;
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+using InLayout  = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout = ck::tensor_layout::convolution::GKYXC;
+using OutLayout = ck::tensor_layout::convolution::GNHWK;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+static constexpr auto Filter1x1Pad0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+static constexpr auto Filter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances = std::tuple<
+    // clang-format off
+           // ########################################|        NDim|     InData|     WeiData|    MultpleD|     OutData|     AccData| InLayout| WeiLayout|   MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+           // ########################################|     Spatial|       Type|        Type|        Type|        Type|        Type|         |          |      Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+           // ########################################|            |           |            |            |            |            |         |          |            |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+           // ########################################|            |           |            |            |            |            |         |          |            |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_Filter1x1Pad0_instances = std::tuple<
+    // clang-format off
+           // ########################################|        NDim|     InData|     WeiData|    MultpleD|     OutData|     AccData| InLayout| WeiLayout|   MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+           // ########################################|     Spatial|       Type|        Type|        Type|        Type|        Type|         |          |      Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+           // ########################################|            |           |            |            |            |            |         |          |            |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+           // ########################################|            |           |            |            |            |            |         |          |            |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,  Filter1x1Pad0,    GemmPadingSpec,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_Filter1x1Stride1Pad0_instances =
+    std::tuple<
+        // clang-format off
+           // ########################################|        NDim|     InData|     WeiData|    MultpleD|     OutData|     AccData| InLayout| WeiLayout|   MultipleD| OutLayout|           In|           Wei|           Out|          Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+           // ########################################|     Spatial|       Type|        Type|        Type|        Type|        Type|         |          |      Layout|          |  Elementwise|   Elementwise|   Elementwise|              Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+           // ########################################|            |           |            |            |            |            |         |          |            |          |    Operation|     Operation|     Operation|       Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+           // ########################################|            |           |            |            |            |            |         |          |            |          |             |              |              |                     |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp, Filter1x1Stride1Pad0,    GemmPadingSpec,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              InLayout,
+                                                              WeiLayout,
+                                                              Empty_Tuple,
+                                                              OutLayout,
+                                                              InDataType,
+                                                              WeiDataType,
+                                                              Empty_Tuple,
+                                                              OutDataType,
+                                                              InElementOp,
+                                                              WeiElementOp,
+                                                              OutElementOp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances{});
+
+    add_device_operation_instances(
+        instances, device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_Filter1x1Pad0_instances{});
+
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_Filter1x1Stride1Pad0_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
new file mode 100644
index 0000000..648b396
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
@@ -0,0 +1,107 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using InDataType  = float;
+using WeiDataType = float;
+using AccDataType = float;
+using OutDataType = float;
+
+using Empty_Tuple = ck::Tuple<>;
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+using InLayout  = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout = ck::tensor_layout::convolution::GKYXC;
+using OutLayout = ck::tensor_layout::convolution::GNHWK;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+static constexpr auto Filter1x1Pad0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+static constexpr auto Filter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances = std::tuple<
+    // clang-format off
+           // clang-format off
+           // ########################################|        NDim|     InData|     WeiData|    MultpleD|     OutData|     AccData| InLayout| WeiLayout|   MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+           // ########################################|     Spatial|       Type|        Type|        Type|        Type|        Type|         |          |      Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+           // ########################################|            |           |            |            |            |            |         |          |            |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+           // ########################################|            |           |            |            |            |            |         |          |            |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_Filter1x1Pad0_instances = std::tuple<
+    // clang-format off
+           // clang-format off
+           // ########################################|        NDim|     InData|     WeiData|    MultpleD|     OutData|     AccData| InLayout| WeiLayout|   MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+           // ########################################|     Spatial|       Type|        Type|        Type|        Type|        Type|         |          |      Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+           // ########################################|            |           |            |            |            |            |         |          |            |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+           // ########################################|            |           |            |            |            |            |         |          |            |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,  Filter1x1Pad0,    GemmPadingSpec,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_Filter1x1Stride1Pad0_instances =
+    std::tuple<
+        // clang-format off
+           // clang-format off
+           // ########################################|        NDim|     InData|     WeiData|    MultpleD|     OutData|     AccData| InLayout| WeiLayout|   MultipleD| OutLayout|           In|           Wei|           Out|          Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+           // ########################################|     Spatial|       Type|        Type|        Type|        Type|        Type|         |          |      Layout|          |  Elementwise|   Elementwise|   Elementwise|              Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+           // ########################################|            |           |            |            |            |            |         |          |            |          |    Operation|     Operation|     Operation|       Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+           // ########################################|            |           |            |            |            |            |         |          |            |          |             |              |              |                     |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp, Filter1x1Stride1Pad0,    GemmPadingSpec,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              InLayout,
+                                                              WeiLayout,
+                                                              Empty_Tuple,
+                                                              OutLayout,
+                                                              InDataType,
+                                                              WeiDataType,
+                                                              Empty_Tuple,
+                                                              OutDataType,
+                                                              InElementOp,
+                                                              WeiElementOp,
+                                                              OutElementOp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances{});
+
+    add_device_operation_instances(
+        instances, device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_Filter1x1Pad0_instances{});
+
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_Filter1x1Stride1Pad0_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp
new file mode 100644
index 0000000..1cb5d06
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp
@@ -0,0 +1,104 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using InDataType  = int8_t;
+using WeiDataType = int8_t;
+using AccDataType = int32_t;
+using OutDataType = int8_t;
+
+using Empty_Tuple = ck::Tuple<>;
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+using InLayout  = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout = ck::tensor_layout::convolution::GKYXC;
+using OutLayout = ck::tensor_layout::convolution::GNHWK;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+static constexpr auto Filter1x1Pad0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+static constexpr auto Filter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances = std::tuple<
+    // clang-format off
+           // ########################################|        NDim|     InData|     WeiData|    MultpleD|     OutData|     AccData| InLayout| WeiLayout|   MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+           // ########################################|     Spatial|       Type|        Type|        Type|        Type|        Type|         |          |      Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+           // ########################################|            |           |            |            |            |            |         |          |            |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+           // ########################################|            |           |            |            |            |            |         |          |            |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_Filter1x1Pad0_instances = std::tuple<
+    // clang-format off
+           // ########################################|        NDim|     InData|     WeiData|    MultpleD|     OutData|     AccData| InLayout| WeiLayout|   MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+           // ########################################|     Spatial|       Type|        Type|        Type|        Type|        Type|         |          |      Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+           // ########################################|            |           |            |            |            |            |         |          |            |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+           // ########################################|            |           |            |            |            |            |         |          |            |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,  Filter1x1Pad0,    GemmPadingSpec,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_Filter1x1Stride1Pad0_instances =
+    std::tuple<
+        // clang-format off
+           // ########################################|        NDim|     InData|     WeiData|    MultpleD|     OutData|     AccData| InLayout| WeiLayout|   MultipleD| OutLayout|           In|           Wei|           Out|          Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+           // ########################################|     Spatial|       Type|        Type|        Type|        Type|        Type|         |          |      Layout|          |  Elementwise|   Elementwise|   Elementwise|              Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+           // ########################################|            |           |            |            |            |            |         |          |            |          |    Operation|     Operation|     Operation|       Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+           // ########################################|            |           |            |            |            |            |         |          |            |          |             |              |              |                     |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp, Filter1x1Stride1Pad0,    GemmPadingSpec,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              InLayout,
+                                                              WeiLayout,
+                                                              Empty_Tuple,
+                                                              OutLayout,
+                                                              InDataType,
+                                                              WeiDataType,
+                                                              Empty_Tuple,
+                                                              OutDataType,
+                                                              InElementOp,
+                                                              WeiElementOp,
+                                                              OutElementOp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances{});
+
+    add_device_operation_instances(
+        instances, device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_Filter1x1Pad0_instances{});
+
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_Filter1x1Stride1Pad0_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
new file mode 100644
index 0000000..29f3310
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
@@ -0,0 +1,156 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNHWC = ck::tensor_layout::convolution::GNHWC;
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+using GNHWK = ck::tensor_layout::convolution::GNHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto ConvFwdOddC =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
+using device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // OddC
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   256,    64,    32,   8,   8,   32,   32,    4,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,    64,    64,    32,   8,   8,   32,   32,    1,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
new file mode 100644
index 0000000..6a4a3d2
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
@@ -0,0 +1,156 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNHWC = ck::tensor_layout::convolution::GNHWC;
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+using GNHWK = ck::tensor_layout::convolution::GNHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto ConvFwdOddC =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// Compilation parameters for in[g, n, hi ,wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
+using device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // OddC
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   256,    64,    32,   8,   8,   32,   32,    4,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,    64,    64,    32,   8,   8,   32,   32,    1,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
new file mode 100644
index 0000000..1fec35f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
@@ -0,0 +1,128 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNHWC = ck::tensor_layout::convolution::GNHWC;
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+using GNHWK = ck::tensor_layout::convolution::GNHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
+using device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instance.cpp
new file mode 100644
index 0000000..59b0121
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instance.cpp
@@ -0,0 +1,125 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNHWC = ck::tensor_layout::convolution::GNHWC;
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+using GNHWK = ck::tensor_layout::convolution::GNHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
+using device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances = std::tuple<
+    // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+                                                                                                                                                                                                                
+        // Filter1x1Stride1Pad0                                                                                                                                                                                 
+        //########################################|  NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, Empty_Tuple,  GNHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
new file mode 100644
index 0000000..8aca730
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
@@ -0,0 +1,156 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWGC = ck::tensor_layout::convolution::NHWGC;
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+using NHWGK = ck::tensor_layout::convolution::NHWGK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto ConvFwdOddC =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+using device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // OddC
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   256,   256,    64,    32,   8,   8,   32,   32,    4,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  NHWGC,  GKYXC, Empty_Tuple,  NHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    ConvFwdOddC, GemmMNKPadding,        1,   128,    64,    64,    32,   8,   8,   32,   32,    1,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
new file mode 100644
index 0000000..04cad43
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_instance_library(device_grouped_conv3d_bwd_weight_instance
+    device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
+    device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
+    device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
new file mode 100644
index 0000000..e48db4a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
@@ -0,0 +1,106 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = bhalf_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
+using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
+using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+using device_grouped_conv3d_bwd_weight_xdl_c_shuffle_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|   BBlockTransfer|  BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                 |                |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 32, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,               ConvBwdWeightDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+        // clang-format on
+        >;
+
+using device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_1x1_s1_p0_bf16_f32_bf16_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|   BBlockTransfer|  BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                 |                |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 32, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdWeightFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv3d_bwd_weight_xdl_c_shuffle_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_1x1_s1_p0_bf16_f32_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
new file mode 100644
index 0000000..1655850
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
@@ -0,0 +1,106 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
+using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
+using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+using device_grouped_conv3d_bwd_weight_xdl_c_shuffle_gndhwc_gkzyxc_gndhwk_f16_default_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+using device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_1x1_s1_p0_f16_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv3d_bwd_weight_xdl_c_shuffle_gndhwc_gkzyxc_gndhwk_f16_default_instances{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
new file mode 100644
index 0000000..aba46b7
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
@@ -0,0 +1,105 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
+using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
+using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+using device_grouped_conv3d_bwd_weight_xdl_c_shuffle_gndhwc_gkzyxc_gndhwk_f32_default_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 64, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 16, 4>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  4>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 16, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,              ConvBwdWeightDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+        // clang-format on
+        >;
+
+using device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_1x1_s1_p0_f32_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 64, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,   128,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,   128,    32,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    64,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvBwdWeightFilter1x1Stride1Pad0,    64,    32,    64,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv3d_bwd_weight_xdl_c_shuffle_gndhwc_gkzyxc_gndhwk_f32_default_instances{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
new file mode 100644
index 0000000..78eedca
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_grouped_conv3d_fwd_instance
+   device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
+   device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
+   device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
+   device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
new file mode 100644
index 0000000..b4ae8b6
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
@@ -0,0 +1,129 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
+using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
+using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, wo, k]
+using device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
new file mode 100644
index 0000000..061674b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
@@ -0,0 +1,129 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
+using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
+using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, wo, k]
+using device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
new file mode 100644
index 0000000..ed7e547
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
@@ -0,0 +1,128 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
+using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
+using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, wo, k]
+using device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
new file mode 100644
index 0000000..bf5fa30
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
@@ -0,0 +1,125 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
+using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
+using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, wo, k]
+using device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances = std::tuple<
+    // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+                                                                                                                                                                                                                
+        // Filter1x1Stride1Pad0                                                                                                                                                                                 
+        //########################################|  NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, GNDHWC, GKZYXC, Empty_Tuple, GNDHWK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt
new file mode 100644
index 0000000..82beb2a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_grouped_gemm_instance
+   device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
+   device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
+   device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000..5f2097b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,94 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[k, m] * b[k, n] = e[m, n]
+using device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    16,    64,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 4>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              2,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances{});
+    add_device_operation_instances(
+        instances, device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000..677bd1a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,94 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[k, m] * b[n, k] = e[m, n]
+using device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    16,    64,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 4>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              2,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances{});
+    add_device_operation_instances(
+        instances, device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000..95a1a87
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,94 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[m, k] * b[k, n] = e[m, n]
+using device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    16,    64,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              2,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances{});
+    add_device_operation_instances(
+        instances, device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000..a103406
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,91 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[m, k] * b[n, k] = e[m, n]
+using device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    16,    64,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              2,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances{});
+    add_device_operation_instances(
+        instances, device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt
new file mode 100644
index 0000000..aa0cc11
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt
@@ -0,0 +1,4 @@
+add_instance_library(device_normalization_instance
+    device_normalization_f16_instance.cpp
+    device_normalization_f32_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_normalization_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_normalization_f16_instance.cpp
new file mode 100644
index 0000000..8994d9d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_normalization_f16_instance.cpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+// clang-format off
+using device_normalization_f16_instances =
+    std::tuple <
+        // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>, // fallback kernel
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 2, 1, 2, 1, 2, 2>
+    >;
+// clang-format on
+
+void add_device_normalization_rank_2_1_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 2, 1>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 2, 1>{});
+}
+
+void add_device_normalization_rank_4_3_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 4, 3>{});
+}
+
+void add_device_normalization_rank_5_3_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 5, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 5, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_normalization_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_normalization_f32_instance.cpp
new file mode 100644
index 0000000..4a7e1fd
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_normalization_f32_instance.cpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_layernorm_f32_instances = std::tuple<
+    // clang-format off
+        // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>
+    // clang-format on
+    >;
+
+void add_device_normalization_rank_2_1_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Pass, 2, 1>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_layernorm_f32_instances<Pass, 2, 1>{});
+}
+
+void add_device_normalization_rank_4_3_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Pass, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_layernorm_f32_instances<Pass, 4, 3>{});
+}
+
+void add_device_normalization_rank_5_3_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Pass, 5, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_layernorm_f32_instances<Pass, 5, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt
new file mode 100644
index 0000000..9f826af
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_quantization_instance
+    device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
+    device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
+    device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
+    device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
new file mode 100644
index 0000000..e87e987
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
@@ -0,0 +1,74 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_xdl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_bias_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_GK_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Mul2_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_int8_32Ds_instances<GK_GK_Tuple,
+                                                                     I32_F32_Tuple,
+                                                                     Add_Mul2_Clamp,
+                                                                     ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_int8_32Ds_instances<GK_GK_Tuple,
+                                                                     I32_F32_Tuple,
+                                                                     Add_Mul2_Clamp,
+                                                                     ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_int8_32Ds_instances<GK_GK_Tuple,
+                                                                     I32_F32_Tuple,
+                                                                     Add_Mul2_Clamp,
+                                                                     ConvFwd1x1S1P0>{});
+}
+
+void add_device_conv2d_bias_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_GK_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Relu_Mul2_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_int8_32Ds_instances<GK_GK_Tuple,
+                                                                     I32_F32_Tuple,
+                                                                     Add_Relu_Mul2_Clamp,
+                                                                     ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_int8_32Ds_instances<GK_GK_Tuple,
+                                                                     I32_F32_Tuple,
+                                                                     Add_Relu_Mul2_Clamp,
+                                                                     ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_int8_32Ds_instances<GK_GK_Tuple,
+                                                                     I32_F32_Tuple,
+                                                                     Add_Relu_Mul2_Clamp,
+                                                                     ConvFwd1x1S1P0>{});
+}
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
new file mode 100644
index 0000000..06eed76
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_xdl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_bias_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, I32_Tuple, Add_Mul_Clamp, ConvFwdDefault>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, I32_Tuple, Add_Mul_Clamp, ConvFwd1x1P0>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, I32_Tuple, Add_Mul_Clamp, ConvFwd1x1S1P0>{});
+}
+
+void add_device_conv2d_bias_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Relu_Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_int8_32Ds_instances<GK_Tuple,
+                                                                     I32_Tuple,
+                                                                     Add_Relu_Mul_Clamp,
+                                                                     ConvFwdDefault>{});
+
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, I32_Tuple, Add_Relu_Mul_Clamp, ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_conv2d_int8_32Ds_instances<GK_Tuple,
+                                                                     I32_Tuple,
+                                                                     Add_Relu_Mul_Clamp,
+                                                                     ConvFwd1x1S1P0>{});
+}
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_int8_instance.hpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_int8_instance.hpp
new file mode 100644
index 0000000..6904e26
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_int8_instance.hpp
@@ -0,0 +1,111 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Empty_Tuple = ck::Tuple<>;
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using GNHWC       = ck::tensor_layout::convolution::GNHWC;
+using GKYXC       = ck::tensor_layout::convolution::GKYXC;
+using GNHWK       = ck::tensor_layout::convolution::GNHWK;
+using GK          = ck::tensor_layout::convolution::G_K;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Relu        = ck::tensor_operation::element_wise::Relu;
+
+using GK_Tuple      = ck::Tuple<GK>;
+using GK_GK_Tuple   = ck::Tuple<GK, GK>;
+using I32_Tuple     = ck::Tuple<int32_t>;
+using F32_Tuple     = ck::Tuple<float>;
+using I32_F32_Tuple = ck::Tuple<int32_t, float>;
+
+using Mul_Clamp      = ck::tensor_operation::element_wise::Activation_Mul_Clamp<PassThrough>;
+using Relu_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Relu>;
+
+using Add_Mul_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<PassThrough>;
+using Add_Relu_Mul_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<Relu>;
+
+using Mul2_Clamp      = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<PassThrough>;
+using Relu_Mul2_Clamp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<Relu>;
+
+using Add_Mul2_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<PassThrough>;
+using Add_Relu_Mul2_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<Relu>;
+
+static constexpr ck::index_t NDimSpatial = 2;
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+template <typename DsLayout,
+          typename DsDatatype,
+          typename OutElementOp,
+          ConvolutionForwardSpecialization ConvSpec>
+// clang-format off
+using device_conv2d_int8_instances =
+    std::tuple <
+        //########################################|  NumDim|      A|      B|       Ds|      E|  AData|  BData| AccData| CShuffle|         Ds|  EData|           A|           B|          CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|   Layout| Layout|   Type|   Type|    Type| DataType|   DataType|   Type| Elementwise| Elementwise|  Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |         |       |       |       |        |         |           |       |   Operation|   Operation|    Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |         |       |       |       |        |         |           |       |            |            |             |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,               16>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,               16>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,               16>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,               16>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 2>,               16>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,               16>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,    64,    64,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 2>,               16>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,               16>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,               16>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,   128,    32,    64,  16,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 2>,               16>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,    32,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,               16>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,    64,    64,    32,    64,  16,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 2>,               16>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,    64,    32,    64,    64,  16,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 2>,               16>
+    >;
+// clang-format on
+
+// for conv + multiple of 32 bit Ds. bit of Ds will affect the ScalarPerVector of C
+template <typename DsLayout,
+          typename DsDatatype,
+          typename OutElementOp,
+          ConvolutionForwardSpecialization ConvSpec>
+// clang-format off
+using device_conv2d_int8_32Ds_instances =
+    std::tuple <
+        //########################################|  NumDim|      A|      B|       Ds|      E|  AData|  BData| AccData| CShuffle|         Ds|  EData|           A|           B|          CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|   Layout| Layout|   Type|   Type|    Type| DataType|   DataType|   Type| Elementwise| Elementwise|  Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |         |       |       |       |        |         |           |       |   Operation|   Operation|    Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |         |       |       |       |        |         |           |       |            |            |             |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 2>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,    64,    64,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 2>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,   128,    32,    64,  16,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 2>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,    32,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,    64,    64,    32,    64,  16,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 2>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       2,  GNHWC,  GKYXC, DsLayout,  GNHWK, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,    64,    32,    64,    64,  16,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 2>,               8>
+    >;
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
new file mode 100644
index 0000000..5f1aa0c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_xdl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Mul2_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Mul2_Clamp, ConvFwdDefault>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Mul2_Clamp, ConvFwd1x1P0>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Mul2_Clamp, ConvFwd1x1S1P0>{});
+}
+
+void add_device_conv2d_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Relu_Mul2_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Relu_Mul2_Clamp, ConvFwdDefault>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Relu_Mul2_Clamp, ConvFwd1x1P0>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Relu_Mul2_Clamp, ConvFwd1x1S1P0>{});
+}
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
new file mode 100644
index 0000000..83435d8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_xdl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_instances<Empty_Tuple, Empty_Tuple, Mul_Clamp, ConvFwdDefault>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_instances<Empty_Tuple, Empty_Tuple, Mul_Clamp, ConvFwd1x1P0>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_instances<Empty_Tuple, Empty_Tuple, Mul_Clamp, ConvFwd1x1S1P0>{});
+}
+
+void add_device_conv2d_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Relu_Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_instances<Empty_Tuple, Empty_Tuple, Relu_Mul_Clamp, ConvFwdDefault>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_instances<Empty_Tuple, Empty_Tuple, Relu_Mul_Clamp, ConvFwd1x1P0>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_instances<Empty_Tuple, Empty_Tuple, Relu_Mul_Clamp, ConvFwd1x1S1P0>{});
+}
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt
new file mode 100644
index 0000000..31ae722
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt
@@ -0,0 +1,76 @@
+add_instance_library(device_reduce_instance
+   device_reduce_instance_blockwise_f16_f16_f16_min.cpp
+   device_reduce_instance_blockwise_f16_f16_f16_max.cpp
+   device_reduce_instance_blockwise_f16_f16_f16_amax.cpp
+   device_reduce_instance_blockwise_f16_f32_f16_add.cpp
+   device_reduce_instance_blockwise_f16_f32_f16_avg.cpp
+   device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp
+   device_reduce_instance_blockwise_f32_f32_f32_add.cpp
+   device_reduce_instance_blockwise_f32_f32_f32_avg.cpp
+   device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp
+   device_reduce_instance_blockwise_f32_f32_f32_min.cpp
+   device_reduce_instance_blockwise_f32_f32_f32_max.cpp
+   device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
+   device_reduce_instance_blockwise_f32_f64_f32_add.cpp
+   device_reduce_instance_blockwise_f32_f64_f32_avg.cpp
+   device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp
+   device_reduce_instance_blockwise_f64_f64_f64_add.cpp
+   device_reduce_instance_blockwise_f64_f64_f64_avg.cpp
+   device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp
+   device_reduce_instance_blockwise_f64_f64_f64_min.cpp
+   device_reduce_instance_blockwise_f64_f64_f64_max.cpp
+   device_reduce_instance_blockwise_f64_f64_f64_amax.cpp
+   device_reduce_instance_blockwise_i8_i32_i8_add.cpp
+   device_reduce_instance_blockwise_i8_i32_i8_avg.cpp
+   device_reduce_instance_blockwise_i8_i8_i8_min.cpp   
+   device_reduce_instance_blockwise_i8_i8_i8_max.cpp   
+   device_reduce_instance_blockwise_i8_i8_i8_amax.cpp   
+   device_reduce_instance_blockwise_b16_f32_b16_add.cpp
+   device_reduce_instance_blockwise_b16_f32_b16_avg.cpp
+   device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp
+   device_reduce_instance_blockwise_b16_f32_b16_min.cpp
+   device_reduce_instance_blockwise_b16_f32_b16_max.cpp
+   device_reduce_instance_blockwise_b16_f32_b16_amax.cpp
+   device_reduce_instance_threadwise_f16_f16_f16_min.cpp
+   device_reduce_instance_threadwise_f16_f16_f16_max.cpp
+   device_reduce_instance_threadwise_f16_f16_f16_amax.cpp
+   device_reduce_instance_threadwise_f16_f32_f16_add.cpp
+   device_reduce_instance_threadwise_f16_f32_f16_avg.cpp
+   device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp
+   device_reduce_instance_threadwise_f32_f32_f32_add.cpp
+   device_reduce_instance_threadwise_f32_f32_f32_avg.cpp
+   device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp
+   device_reduce_instance_threadwise_f32_f32_f32_min.cpp
+   device_reduce_instance_threadwise_f32_f32_f32_max.cpp
+   device_reduce_instance_threadwise_f32_f32_f32_amax.cpp
+   device_reduce_instance_threadwise_f32_f64_f32_add.cpp
+   device_reduce_instance_threadwise_f32_f64_f32_avg.cpp
+   device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp
+   device_reduce_instance_threadwise_f64_f64_f64_add.cpp
+   device_reduce_instance_threadwise_f64_f64_f64_avg.cpp
+   device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp
+   device_reduce_instance_threadwise_f64_f64_f64_min.cpp
+   device_reduce_instance_threadwise_f64_f64_f64_max.cpp
+   device_reduce_instance_threadwise_f64_f64_f64_amax.cpp
+   device_reduce_instance_threadwise_i8_i32_i8_add.cpp
+   device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
+   device_reduce_instance_threadwise_i8_i8_i8_min.cpp
+   device_reduce_instance_threadwise_i8_i8_i8_max.cpp
+   device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
+   device_reduce_instance_threadwise_b16_f32_b16_add.cpp
+   device_reduce_instance_threadwise_b16_f32_b16_avg.cpp
+   device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp
+   device_reduce_instance_threadwise_b16_f32_b16_min.cpp
+   device_reduce_instance_threadwise_b16_f32_b16_max.cpp
+   device_reduce_instance_threadwise_b16_f32_b16_amax.cpp
+   device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp
+   device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp
+   device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp
+   device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp
+   device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp
+   device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp
+   device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp
+   device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp
+   device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp
+   device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp
new file mode 100644
index 0000000..1909183
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp
new file mode 100644
index 0000000..ec30201
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp
new file mode 100644
index 0000000..89f3e58
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp
new file mode 100644
index 0000000..f1bdd19
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp
new file mode 100644
index 0000000..58e9c56
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp
new file mode 100644
index 0000000..e5012c6
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp
new file mode 100644
index 0000000..0970cb9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp
new file mode 100644
index 0000000..6ee179a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp
new file mode 100644
index 0000000..e53b403
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp
new file mode 100644
index 0000000..cab5738
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp
new file mode 100644
index 0000000..7d2a4fa
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp
new file mode 100644
index 0000000..e08b64f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp
new file mode 100644
index 0000000..89cabf3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
new file mode 100644
index 0000000..1e602c1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp
new file mode 100644
index 0000000..489b4bc
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp
new file mode 100644
index 0000000..04e2c5b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp
new file mode 100644
index 0000000..5c0e536
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp
new file mode 100644
index 0000000..899dfcd
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp
new file mode 100644
index 0000000..5624337
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp
new file mode 100644
index 0000000..2f3067c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp
new file mode 100644
index 0000000..2648e7d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp
new file mode 100644
index 0000000..f67ae2e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp
new file mode 100644
index 0000000..6f8e078
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp
new file mode 100644
index 0000000..69fecf7
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp
new file mode 100644
index 0000000..129a4f0
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp
new file mode 100644
index 0000000..21babc4
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp
new file mode 100644
index 0000000..b85b3e2
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp
new file mode 100644
index 0000000..24a8293
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp
new file mode 100644
index 0000000..73e60fa
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp
new file mode 100644
index 0000000..72e649d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp
new file mode 100644
index 0000000..a7e053a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp
new file mode 100644
index 0000000..0e3abd3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp
new file mode 100644
index 0000000..4b32456
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp
new file mode 100644
index 0000000..3298587
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp
new file mode 100644
index 0000000..729d4fd
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp
new file mode 100644
index 0000000..e3e36e3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp
new file mode 100644
index 0000000..e7580e7
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp
new file mode 100644
index 0000000..1e6feb0
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp
new file mode 100644
index 0000000..669c4d3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp
new file mode 100644
index 0000000..335a547
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp
new file mode 100644
index 0000000..e95e839
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp
new file mode 100644
index 0000000..2549815
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp
new file mode 100644
index 0000000..7262b8a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp
new file mode 100644
index 0000000..c526a74
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp
new file mode 100644
index 0000000..4c7252e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp
new file mode 100644
index 0000000..618900a
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp
new file mode 100644
index 0000000..ce747cb
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp
new file mode 100644
index 0000000..06f622b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp
new file mode 100644
index 0000000..708eb58
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp
new file mode 100644
index 0000000..c8a62fa
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp
new file mode 100644
index 0000000..ce20921
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp
new file mode 100644
index 0000000..29251a8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp
new file mode 100644
index 0000000..734fa9f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp
new file mode 100644
index 0000000..d7a0e2b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp
new file mode 100644
index 0000000..8b97f30
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp
new file mode 100644
index 0000000..53d01e3
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp
new file mode 100644
index 0000000..125d054
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp
new file mode 100644
index 0000000..fb86a2b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp
new file mode 100644
index 0000000..49af083
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp
new file mode 100644
index 0000000..30cc1b1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp
new file mode 100644
index 0000000..24f8a9b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp
new file mode 100644
index 0000000..a26702f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp
new file mode 100644
index 0000000..34fe326
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp
new file mode 100644
index 0000000..74b15ed
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp
new file mode 100644
index 0000000..6576249
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp
new file mode 100644
index 0000000..5e74295
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp
new file mode 100644
index 0000000..6fdea6c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp
new file mode 100644
index 0000000..317d573
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp
new file mode 100644
index 0000000..29f95eb
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp
new file mode 100644
index 0000000..aa9f47c
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
new file mode 100644
index 0000000..54a9dd1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
new file mode 100644
index 0000000..4ef5717
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp
new file mode 100644
index 0000000..140a3c1
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp
new file mode 100644
index 0000000..317b4ad
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
new file mode 100644
index 0000000..fc13261
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
@@ -0,0 +1,26 @@
+add_instance_library(device_softmax_instance
+    device_softmax_i8_i8_instance.cpp
+    device_softmax_i8_i8_instance_rank3_reduce1.cpp
+    device_softmax_i8_i8_instance_rank3_reduce2.cpp
+    device_softmax_i8_i8_instance_rank3_reduce3.cpp
+    device_softmax_i8_i8_instance_rank4_reduce1.cpp
+    device_softmax_i8_i8_instance_rank4_reduce2.cpp
+    device_softmax_i8_i8_instance_rank4_reduce3.cpp
+    device_softmax_i8_i8_instance_rank4_reduce4.cpp
+    device_softmax_f16_f16_instance.cpp
+    device_softmax_f16_f16_instance_rank3_reduce1.cpp
+    device_softmax_f16_f16_instance_rank3_reduce2.cpp
+    device_softmax_f16_f16_instance_rank3_reduce3.cpp
+    device_softmax_f16_f16_instance_rank4_reduce1.cpp
+    device_softmax_f16_f16_instance_rank4_reduce2.cpp
+    device_softmax_f16_f16_instance_rank4_reduce3.cpp
+    device_softmax_f16_f16_instance_rank4_reduce4.cpp
+    device_softmax_f32_f32_instance.cpp
+    device_softmax_f32_f32_instance_rank3_reduce1.cpp
+    device_softmax_f32_f32_instance_rank3_reduce2.cpp
+    device_softmax_f32_f32_instance_rank3_reduce3.cpp
+    device_softmax_f32_f32_instance_rank4_reduce1.cpp
+    device_softmax_f32_f32_instance_rank4_reduce2.cpp
+    device_softmax_f32_f32_instance_rank4_reduce3.cpp
+    device_softmax_f32_f32_instance_rank4_reduce4.cpp
+)
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.cpp
new file mode 100644
index 0000000..14d2764
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances)
+{
+    add_device_softmax_f16_f16_rank3_reduce1_instances(instances);
+    add_device_softmax_f16_f16_rank3_reduce2_instances(instances);
+    add_device_softmax_f16_f16_rank3_reduce3_instances(instances);
+}
+
+void add_device_softmax_f16_f16_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances)
+{
+    add_device_softmax_f16_f16_rank4_reduce1_instances(instances);
+    add_device_softmax_f16_f16_rank4_reduce2_instances(instances);
+    add_device_softmax_f16_f16_rank4_reduce3_instances(instances);
+    add_device_softmax_f16_f16_rank4_reduce4_instances(instances);
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp
new file mode 100644
index 0000000..fa334b9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 3;
+
+void add_device_softmax_f16_f16_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp
new file mode 100644
index 0000000..1c9d37d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 3;
+
+void add_device_softmax_f16_f16_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp
new file mode 100644
index 0000000..5fbdab5
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 3;
+
+void add_device_softmax_f16_f16_rank3_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp
new file mode 100644
index 0000000..7dd8640
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_f16_f16_rank4_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp
new file mode 100644
index 0000000..b32fe68
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_f16_f16_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp
new file mode 100644
index 0000000..c05048e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_f16_f16_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp
new file mode 100644
index 0000000..6a23570
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_f16_f16_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.cpp
new file mode 100644
index 0000000..e5bec5e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances)
+{
+    add_device_softmax_f32_f32_rank3_reduce1_instances(instances);
+    add_device_softmax_f32_f32_rank3_reduce2_instances(instances);
+    add_device_softmax_f32_f32_rank3_reduce3_instances(instances);
+}
+
+void add_device_softmax_f32_f32_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances)
+{
+    add_device_softmax_f32_f32_rank4_reduce1_instances(instances);
+    add_device_softmax_f32_f32_rank4_reduce2_instances(instances);
+    add_device_softmax_f32_f32_rank4_reduce3_instances(instances);
+    add_device_softmax_f32_f32_rank4_reduce4_instances(instances);
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp
new file mode 100644
index 0000000..57d3f18
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 3;
+
+void add_device_softmax_f32_f32_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp
new file mode 100644
index 0000000..fae3a4d
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 3;
+
+void add_device_softmax_f32_f32_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp
new file mode 100644
index 0000000..b6fb70e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 3;
+
+void add_device_softmax_f32_f32_rank3_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp
new file mode 100644
index 0000000..33c7b6f
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_f32_f32_rank4_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp
new file mode 100644
index 0000000..c22aa57
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_f32_f32_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp
new file mode 100644
index 0000000..55f3d2b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_f32_f32_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp
new file mode 100644
index 0000000..fb0bcf5
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_f32_f32_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.cpp
new file mode 100644
index 0000000..608cfcf
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances)
+{
+    add_device_softmax_i8_i8_rank3_reduce1_instances(instances);
+    add_device_softmax_i8_i8_rank3_reduce2_instances(instances);
+    add_device_softmax_i8_i8_rank3_reduce3_instances(instances);
+}
+
+void add_device_softmax_i8_i8_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances)
+{
+    add_device_softmax_i8_i8_rank4_reduce1_instances(instances);
+    add_device_softmax_i8_i8_rank4_reduce2_instances(instances);
+    add_device_softmax_i8_i8_rank4_reduce3_instances(instances);
+    add_device_softmax_i8_i8_rank4_reduce4_instances(instances);
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.cpp
new file mode 100644
index 0000000..15552db
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 3;
+
+void add_device_softmax_i8_i8_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.cpp
new file mode 100644
index 0000000..6767402
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 3;
+
+void add_device_softmax_i8_i8_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.cpp
new file mode 100644
index 0000000..4b33da9
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 3;
+
+void add_device_softmax_i8_i8_rank3_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.cpp
new file mode 100644
index 0000000..fe3b823
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_i8_i8_rank4_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.cpp
new file mode 100644
index 0000000..8ecdf87
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_i8_i8_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.cpp
new file mode 100644
index 0000000..3563135
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_i8_i8_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.cpp b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.cpp
new file mode 100644
index 0000000..aa21a0b
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_i8_i8_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/library/src/utility/CMakeLists.txt b/3rdparty/composable_kernel/library/src/utility/CMakeLists.txt
new file mode 100644
index 0000000..7f6a59e
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/utility/CMakeLists.txt
@@ -0,0 +1,28 @@
+## utility
+set(UTILITY_SOURCE
+    device_memory.cpp
+    host_tensor.cpp
+    convolution_parameter.cpp
+)
+
+add_library(utility STATIC ${UTILITY_SOURCE})
+add_library(composable_kernel::utility ALIAS utility)
+
+target_include_directories(utility PUBLIC
+    "$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck>"
+    "$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/utility>"
+)
+
+rocm_install(
+    TARGETS utility
+    EXPORT utilityTargets
+)
+
+rocm_install(
+    EXPORT utilityTargets
+    FILE composable_kernelutilityTargets.cmake
+    NAMESPACE composable_kernel::
+    DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+)
+
+clang_tidy_check(utility)
diff --git a/3rdparty/composable_kernel/library/src/utility/convolution_parameter.cpp b/3rdparty/composable_kernel/library/src/utility/convolution_parameter.cpp
new file mode 100644
index 0000000..c8712d2
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/utility/convolution_parameter.cpp
@@ -0,0 +1,171 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/host_utility/io.hpp"
+
+#include "ck/library/utility/convolution_parameter.hpp"
+
+namespace ck {
+namespace utils {
+namespace conv {
+
+ConvParam::ConvParam(ck::index_t n_dim,
+                     ck::index_t group_count,
+                     ck::index_t n_batch,
+                     ck::index_t n_out_channels,
+                     ck::index_t n_in_channels,
+                     const std::vector<ck::index_t>& filters_len,
+                     const std::vector<ck::index_t>& input_len,
+                     const std::vector<ck::index_t>& strides,
+                     const std::vector<ck::index_t>& dilations,
+                     const std::vector<ck::index_t>& left_pads,
+                     const std::vector<ck::index_t>& right_pads)
+    : num_dim_spatial_(n_dim),
+      G_(group_count),
+      N_(n_batch),
+      K_(n_out_channels),
+      C_(n_in_channels),
+      filter_spatial_lengths_(filters_len),
+      input_spatial_lengths_(input_len),
+      output_spatial_lengths_(num_dim_spatial_),
+      conv_filter_strides_(strides),
+      conv_filter_dilations_(dilations),
+      input_left_pads_(left_pads),
+      input_right_pads_(right_pads)
+{
+    if(static_cast<ck::index_t>(filter_spatial_lengths_.size()) != num_dim_spatial_ ||
+       static_cast<ck::index_t>(input_spatial_lengths_.size()) != num_dim_spatial_ ||
+       static_cast<ck::index_t>(conv_filter_strides_.size()) != num_dim_spatial_ ||
+       static_cast<ck::index_t>(conv_filter_dilations_.size()) != num_dim_spatial_ ||
+       static_cast<ck::index_t>(input_left_pads_.size()) != num_dim_spatial_ ||
+       static_cast<ck::index_t>(input_right_pads_.size()) != num_dim_spatial_)
+    {
+        throw(
+            std::runtime_error("ConvParam::ConvParam: "
+                               "parameter size is different from number of declared dimensions!"));
+    }
+
+    for(ck::index_t i = 0; i < num_dim_spatial_; ++i)
+    {
+        // XEff = (X - 1) * conv_dilation_w + 1;
+        // Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+        const ck::index_t x_eff = (filter_spatial_lengths_[i] - 1) * conv_filter_dilations_[i] + 1;
+
+        output_spatial_lengths_[i] =
+            (input_spatial_lengths_[i] + input_left_pads_[i] + input_right_pads_[i] - x_eff) /
+                conv_filter_strides_[i] +
+            1;
+    }
+}
+
+ConvParam::ConvParam()
+    : ConvParam::ConvParam(2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1})
+{
+}
+
+std::vector<ck::index_t> ConvParam::GetOutputSpatialLengths() const
+{
+    return output_spatial_lengths_;
+}
+
+std::size_t ConvParam::GetFlops() const
+{
+    // 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product>
+    return static_cast<std::size_t>(2) * G_ * N_ * K_ * C_ *
+           ck::accumulate_n<std::size_t>(
+               std::begin(output_spatial_lengths_), num_dim_spatial_, 1, std::multiplies<>()) *
+           ck::accumulate_n<std::size_t>(
+               std::begin(filter_spatial_lengths_), num_dim_spatial_, 1, std::multiplies<>());
+}
+
+std::string get_conv_param_parser_helper_msg()
+{
+    std::string msg;
+
+    msg += "Following arguments (depending on number of spatial dims):\n"
+           " Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)\n"
+           " G, N, K, C, \n"
+           " <filter spatial dimensions>, (ie Y, X for 2D)\n"
+           " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
+           " <strides>, (ie Sy, Sx for 2D)\n"
+           " <dilations>, (ie Dy, Dx for 2D)\n"
+           " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
+           " <right padding>, (ie RightPy, RightPx for 2D)\n";
+
+    return msg;
+}
+
+ck::utils::conv::ConvParam parse_conv_param(int num_dim_spatial, int arg_idx, char* const argv[])
+{
+    const ck::index_t G = std::stoi(argv[arg_idx++]);
+    const ck::index_t N = std::stoi(argv[arg_idx++]);
+    const ck::index_t K = std::stoi(argv[arg_idx++]);
+    const ck::index_t C = std::stoi(argv[arg_idx++]);
+
+    std::vector<ck::index_t> filter_spatial_lengths(num_dim_spatial);
+    std::vector<ck::index_t> input_spatial_lengths(num_dim_spatial);
+    std::vector<ck::index_t> conv_filter_strides(num_dim_spatial);
+    std::vector<ck::index_t> conv_filter_dilations(num_dim_spatial);
+    std::vector<ck::index_t> input_left_pads(num_dim_spatial);
+    std::vector<ck::index_t> input_right_pads(num_dim_spatial);
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        input_left_pads[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        input_right_pads[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    return ck::utils::conv::ConvParam{num_dim_spatial,
+                                      G,
+                                      N,
+                                      K,
+                                      C,
+                                      filter_spatial_lengths,
+                                      input_spatial_lengths,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads};
+}
+} // namespace conv
+} // namespace utils
+} // namespace ck
+
+std::ostream& operator<<(std::ostream& os, const ck::utils::conv::ConvParam& p)
+{
+    os << "ConvParam {"
+       << "\nnum_dim_spatial: " << p.num_dim_spatial_ << "\nG: " << p.G_ << "\nN: " << p.N_
+       << "\nK: " << p.K_ << "\nC: " << p.C_
+       << "\nfilter_spatial_lengths: " << p.filter_spatial_lengths_
+       << "\ninput_spatial_lengths: " << p.input_spatial_lengths_
+       << "\nconv_filter_strides: " << p.conv_filter_strides_
+       << "\nconv_filter_dilations: " << p.conv_filter_dilations_
+       << "\ninput_left_pads: " << p.input_left_pads_
+       << "\ninput_right_pads: " << p.input_right_pads_ << "}\n";
+
+    return os;
+}
diff --git a/3rdparty/composable_kernel/library/src/utility/device_memory.cpp b/3rdparty/composable_kernel/library/src/utility/device_memory.cpp
new file mode 100644
index 0000000..90f9433
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/utility/device_memory.cpp
@@ -0,0 +1,29 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/host_utility/hip_check_error.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+
+DeviceMem::DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
+{
+    hip_check_error(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
+}
+
+void* DeviceMem::GetDeviceBuffer() const { return mpDeviceBuf; }
+
+std::size_t DeviceMem::GetBufferSize() const { return mMemSize; }
+
+void DeviceMem::ToDevice(const void* p) const
+{
+    hip_check_error(hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
+}
+
+void DeviceMem::FromDevice(void* p) const
+{
+    hip_check_error(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
+}
+
+void DeviceMem::SetZero() const { hip_check_error(hipMemset(mpDeviceBuf, 0, mMemSize)); }
+
+DeviceMem::~DeviceMem() { hip_check_error(hipFree(mpDeviceBuf)); }
diff --git a/3rdparty/composable_kernel/library/src/utility/host_tensor.cpp b/3rdparty/composable_kernel/library/src/utility/host_tensor.cpp
new file mode 100644
index 0000000..e34fbc8
--- /dev/null
+++ b/3rdparty/composable_kernel/library/src/utility/host_tensor.cpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cassert>
+
+#include "ck/library/utility/host_tensor.hpp"
+
+void HostTensorDescriptor::CalculateStrides()
+{
+    mStrides.clear();
+    mStrides.resize(mLens.size(), 0);
+    if(mStrides.empty())
+        return;
+
+    mStrides.back() = 1;
+    std::partial_sum(
+        mLens.rbegin(), mLens.rend() - 1, mStrides.rbegin() + 1, std::multiplies<std::size_t>());
+}
+
+std::size_t HostTensorDescriptor::GetNumOfDimension() const { return mLens.size(); }
+
+std::size_t HostTensorDescriptor::GetElementSize() const
+{
+    assert(mLens.size() == mStrides.size());
+    return std::accumulate(
+        mLens.begin(), mLens.end(), std::size_t{1}, std::multiplies<std::size_t>());
+}
+
+std::size_t HostTensorDescriptor::GetElementSpaceSize() const
+{
+    std::size_t space = 1;
+    for(std::size_t i = 0; i < mLens.size(); ++i)
+    {
+        if(mLens[i] == 0)
+            continue;
+
+        space += (mLens[i] - 1) * mStrides[i];
+    }
+    return space;
+}
+
+const std::vector<std::size_t>& HostTensorDescriptor::GetLengths() const { return mLens; }
+
+const std::vector<std::size_t>& HostTensorDescriptor::GetStrides() const { return mStrides; }
+
+std::ostream& operator<<(std::ostream& os, const HostTensorDescriptor& desc)
+{
+    os << "dim " << desc.GetNumOfDimension() << ", ";
+
+    os << "lengths {";
+    LogRange(os, desc.GetLengths(), ", ");
+    os << "}, ";
+
+    os << "strides {";
+    LogRange(os, desc.GetStrides(), ", ");
+    os << "}";
+
+    return os;
+}
diff --git a/3rdparty/composable_kernel/profiler/CMakeLists.txt b/3rdparty/composable_kernel/profiler/CMakeLists.txt
new file mode 100644
index 0000000..bdd7125
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/CMakeLists.txt
@@ -0,0 +1,5 @@
+include_directories(BEFORE
+    ${CMAKE_CURRENT_LIST_DIR}/include
+)
+
+add_subdirectory(src)
diff --git a/3rdparty/composable_kernel/profiler/README.md b/3rdparty/composable_kernel/profiler/README.md
new file mode 100644
index 0000000..bfd6a3a
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/README.md
@@ -0,0 +1,48 @@
+## Profile GEMM kernels
+```bash
+#arg1: tensor operation (gemm=GEMM)
+#arg2: data type (0=fp32, 1=fp16)
+#arg3: matrix layout (0=NN, 1=NT, 2=TN, 3=TT)
+#arg4: verification (0=no, 1=yes)
+#arg5: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg6: print matrix value (0=no, 1=yes)
+#arg7: run kernel # of times (>1)
+#arg8 to 13: M, N, K, StrideA, StrideB, StrideC
+
+################        op  datatype  layout  verify  init  log  repeat  M___ N___ K___  StrideA StrideB StrideC
+./bin/ckProfiler      gemm         1       1       1     1    0       5  3840 4096 4096     4096    4096    4096
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```bash
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+....
+Best Perf: 1.1933 ms, 107.977 TFlops, 79.0848 GB/s
+```
+
+## Profile 2d forward convolution kernels
+```bash
+#arg1: tensor operation (conv=Convolution)
+#arg2: data type (0=fp32, 1=fp16)
+#arg3: input tensor layout (0=NCHW, 1=NHWC)
+#arg4: weight tensor layout (0=KCYX, 1=KYXC)
+#arg5: output tensor layout (0=NKHW, 1=NHWK)
+#arg6: verification (0=no, 1=yes)
+#arg7: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg8: print matrix value (0=no, 1=yes)
+#arg9: run kernel # of times (>1)
+#arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx
+ ################          op datatype  in_layout   wei_layout  out_layout  verify  init  log  repeat  N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
+ ./bin/ckProfiler  conv2d_fwd        1          1            1           1       1     1    0       5  128  256  192 3 3   71   71     2 2       1 1      1 1       1 1
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
+wei_k_c_y_x: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
+out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
+....
+Best Perf: 1.42509 ms, 102.988 TFlops, 234.086 GB/s
+```
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/data_type_enum.hpp b/3rdparty/composable_kernel/profiler/include/profiler/data_type_enum.hpp
new file mode 100644
index 0000000..afcd6fe
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/data_type_enum.hpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+enum struct DataTypeEnum
+{
+    Half     = 0,
+    Float    = 1,
+    Int32    = 2,
+    Int8     = 3,
+    Int8x4   = 4,
+    BFloat16 = 5,
+    Double   = 6,
+    Unknown  = 100,
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/data_type_enum_helper.hpp b/3rdparty/composable_kernel/profiler/include/profiler/data_type_enum_helper.hpp
new file mode 100644
index 0000000..d9bd5e1
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/data_type_enum_helper.hpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma
+
+#include "ck/utility/data_type.hpp"
+#include "profiler/data_type_enum.hpp"
+
+namespace ck {
+
+template <DataTypeEnum DataTypeEnum>
+struct get_datatype_from_enum;
+
+template <>
+struct get_datatype_from_enum<DataTypeEnum::Int8>
+{
+    using type = int8_t;
+};
+
+template <>
+struct get_datatype_from_enum<DataTypeEnum::Int32>
+{
+    using type = int32_t;
+};
+
+template <>
+struct get_datatype_from_enum<DataTypeEnum::Half>
+{
+    using type = half_t;
+};
+
+template <>
+struct get_datatype_from_enum<DataTypeEnum::Float>
+{
+    using type = float;
+};
+
+template <>
+struct get_datatype_from_enum<DataTypeEnum::Double>
+{
+    using type = double;
+};
+
+template <typename T>
+struct get_datatype_enum_from_type;
+
+template <>
+struct get_datatype_enum_from_type<int8_t>
+{
+    static constexpr DataTypeEnum value = DataTypeEnum::Int8;
+};
+
+template <>
+struct get_datatype_enum_from_type<int32_t>
+{
+    static constexpr DataTypeEnum value = DataTypeEnum::Int32;
+};
+
+template <>
+struct get_datatype_enum_from_type<half_t>
+{
+    static constexpr DataTypeEnum value = DataTypeEnum::Half;
+};
+
+template <>
+struct get_datatype_enum_from_type<float>
+{
+    static constexpr DataTypeEnum value = DataTypeEnum::Float;
+};
+
+template <>
+struct get_datatype_enum_from_type<double>
+{
+    static constexpr DataTypeEnum value = DataTypeEnum::Double;
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp
new file mode 100644
index 0000000..b162542
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp
@@ -0,0 +1,360 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename A0Layout,
+          typename B0Layout,
+          typename D0sLayout,
+          typename B1Layout,
+          typename D1sLayout,
+          typename E1Layout,
+          typename A0DataType,
+          typename B0DataType,
+          typename D0sDataType,
+          typename B1DataType,
+          typename D1sDataType,
+          typename E1DataType>
+bool profile_batched_gemm_add_relu_gemm_add_impl(bool do_verification,
+                                                 int init_method,
+                                                 bool do_log,
+                                                 bool time_kernel,
+                                                 int M,
+                                                 int N,
+                                                 int K,
+                                                 int O,
+                                                 int BatchCount    = 1,
+                                                 int StrideA0      = -1,
+                                                 int StrideB0      = -1,
+                                                 int StrideD0      = -1,
+                                                 int StrideB1      = -1,
+                                                 int StrideD1      = -1,
+                                                 int StrideE1      = -1,
+                                                 int BatchStrideA0 = -1,
+                                                 int BatchStrideB0 = -1,
+                                                 int BatchStrideD0 = -1,
+                                                 int BatchStrideB1 = -1,
+                                                 int BatchStrideD1 = -1,
+                                                 int BatchStrideE1 = -1)
+
+{
+    using Row = tensor_layout::gemm::RowMajor;
+    using Col = tensor_layout::gemm::ColumnMajor;
+
+    using PassThrough = tensor_operation::element_wise::PassThrough;
+
+    using A0ElementOp   = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
+    using B1ElementOp   = PassThrough;
+    using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
+
+    using D0DataType = remove_cvref_t<tuple_element_t<0, D0sDataType>>;
+
+    using D0Layout   = remove_cvref_t<tuple_element_t<0, D0sLayout>>;
+    using D1DataType = remove_cvref_t<tuple_element_t<0, D1sDataType>>;
+    using D1Layout   = remove_cvref_t<tuple_element_t<0, D1sLayout>>;
+
+    // for reference
+    using RefAcc0DataType = float;
+    using RefAcc1DataType = float;
+
+    bool pass = true;
+
+    const int DefaultStrideA0 = ck::is_same_v<A0Layout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? O : M;
+    const int DefaultStrideE1 = ck::is_same_v<E1Layout, Row> ? O : M;
+
+    StrideA0 = (StrideA0 < 0) ? DefaultStrideA0 : StrideA0;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideD0 = (StrideD0 < 0) ? DefaultStrideD0 : StrideD0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideD1 = (StrideD1 < 0) ? DefaultStrideD1 : StrideD1;
+    StrideE1 = (StrideE1 < 0) ? DefaultStrideE1 : StrideE1;
+
+    const int DefaultBatchStrideA0 = (ck::is_same_v<A0Layout, Col> ? K : M) * StrideA0;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideD0 = (ck::is_same_v<D0Layout, Col> ? N : M) * StrideD0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideD1 = (ck::is_same_v<D1Layout, Col> ? O : M) * StrideD1;
+    const int DefaultBatchStrideE1 = (ck::is_same_v<E1Layout, Col> ? O : M) * StrideE1;
+
+    BatchStrideA0 = BatchStrideA0 < 0 ? DefaultBatchStrideA0 : BatchStrideA0;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideD0 = BatchStrideD0 < 0 ? DefaultBatchStrideD0 : BatchStrideD0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideD1 = BatchStrideD1 < 0 ? DefaultBatchStrideD1 : BatchStrideD1;
+    BatchStrideE1 = BatchStrideE1 < 0 ? DefaultBatchStrideE1 : BatchStrideE1;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    // E_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<A0DataType> a0_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA0, BatchStrideA0, A0Layout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<D0DataType> d0_g_m_n(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideD0, BatchStrideD0, D0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<D1DataType> d1_g_m_o(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideD1, BatchStrideD1, D1Layout{}));
+    Tensor<E1DataType> e1_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideE1, BatchStrideE1, E1Layout{}));
+    Tensor<E1DataType> e1_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideE1, BatchStrideE1, E1Layout{}));
+
+    // Host verification: Output of Gemm0 is input A of Gemm1
+    Tensor<RefAcc0DataType> c0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+    Tensor<RefAcc0DataType> e0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+    Tensor<RefAcc1DataType> c1_g_m_o(f_host_tensor_descriptor(BatchCount, M, O, O, M * O, Row{}));
+
+    std::cout << "a0_g_m_k: " << a0_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "d0_g_m_n: " << d0_g_m_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "d1_g_m_o: " << d1_g_m_o.mDesc << std::endl;
+    std::cout << "e1_g_m_o: " << e1_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_g_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 3});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 3});
+        d0_g_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 3});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 3});
+        d1_g_m_o.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-2, 3});
+        break;
+    default:
+        a0_g_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        d0_g_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        d1_g_m_o.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    DeviceMem a0_g_m_k_device_buf(sizeof(A0DataType) * a0_g_m_k.mDesc.GetElementSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
+    DeviceMem d0_g_m_n_device_buf(sizeof(D0DataType) * d0_g_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
+    DeviceMem d1_g_m_o_device_buf(sizeof(D1DataType) * d1_g_m_o.mDesc.GetElementSpaceSize());
+    DeviceMem e1_g_m_o_device_buf(sizeof(E1DataType) *
+                                  e1_g_m_o_device_result.mDesc.GetElementSize());
+
+    a0_g_m_k_device_buf.ToDevice(a0_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    d0_g_m_n_device_buf.ToDevice(d0_g_m_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+    d1_g_m_o_device_buf.ToDevice(d1_g_m_o.mData.data());
+
+    auto a0_element_op   = A0ElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto cde0_element_op = CDE0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto cde1_element_op = CDE1ElementOp{};
+
+    using DeviceOp =
+        tensor_operation::device::DeviceBatchedGemmMultipleDGemmMultipleD<A0Layout,
+                                                                          B0Layout,
+                                                                          D0sLayout,
+                                                                          B1Layout,
+                                                                          D1sLayout,
+                                                                          E1Layout,
+                                                                          A0DataType,
+                                                                          B0DataType,
+                                                                          D0sDataType,
+                                                                          B1DataType,
+                                                                          D1sDataType,
+                                                                          E1DataType,
+                                                                          A0ElementOp,
+                                                                          B0ElementOp,
+                                                                          CDE0ElementOp,
+                                                                          B1ElementOp,
+                                                                          CDE1ElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    if(do_verification)
+    {
+        // Ref Gemm0
+        using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<A0DataType,
+                                                                                    B0DataType,
+                                                                                    RefAcc0DataType,
+                                                                                    RefAcc0DataType,
+                                                                                    A0ElementOp,
+                                                                                    B0ElementOp,
+                                                                                    PassThrough>;
+
+        // Ref Gemm1
+        using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<RefAcc0DataType,
+                                                                                    B1DataType,
+                                                                                    RefAcc1DataType,
+                                                                                    RefAcc1DataType,
+                                                                                    PassThrough,
+                                                                                    B1ElementOp,
+                                                                                    PassThrough>;
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a0_g_m_k, b0_g_k_n, c0_g_m_n, a0_element_op, b0_element_op, PassThrough{});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        // cde0_elementwise
+        e0_g_m_n.ForEach(
+            [&](auto&, auto idx) { cde0_element_op(e0_g_m_n(idx), c0_g_m_n(idx), d0_g_m_n(idx)); });
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            e0_g_m_n, b1_g_n_o, c1_g_m_o, PassThrough{}, b1_element_op, PassThrough{});
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // cde1_elementwise
+        e1_g_m_o_host_result.ForEach([&](auto&, auto idx) {
+            cde1_element_op(e1_g_m_o_host_result(idx), c1_g_m_o(idx), d1_g_m_o(idx));
+        });
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<A0DataType*>(a0_g_m_k_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+            std::array<const void*, 1>{d0_g_m_n_device_buf.GetDeviceBuffer()},
+            static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+            std::array<const void*, 1>{d1_g_m_o_device_buf.GetDeviceBuffer()},
+            static_cast<E1DataType*>(e1_g_m_o_device_buf.GetDeviceBuffer()),
+            M,
+            N,
+            K,
+            O,
+            BatchCount,
+            StrideA0,
+            StrideB0,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideB1,
+            std::array<ck::index_t, 1>{StrideD1},
+            StrideE1,
+            BatchStrideA0,
+            BatchStrideB0,
+            std::array<ck::index_t, 1>{BatchStrideD0},
+            BatchStrideB1,
+            std::array<ck::index_t, 1>{BatchStrideD1},
+            BatchStrideE1,
+            a0_element_op,
+            b0_element_op,
+            cde0_element_op,
+            b1_element_op,
+            cde1_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+            std::size_t num_btype =
+                (sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(D0DataType) * N +
+                 sizeof(B1DataType) * N * O + sizeof(E1DataType) * M * O + sizeof(D1DataType) * O) *
+                BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e1_g_m_o_device_buf.FromDevice(e1_g_m_o_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(e1_g_m_o_device_result, e1_g_m_o_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(
+                        std::cout << "e1_g_m_o_host_result : ", e1_g_m_o_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "e1_g_m_o_device_result : ", e1_g_m_o_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp
new file mode 100644
index 0000000..1583c6d
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp
@@ -0,0 +1,319 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout>
+bool profile_batched_gemm_gemm_impl(bool do_verification,
+                                    int init_method,
+                                    bool do_log,
+                                    bool time_kernel,
+                                    int M,
+                                    int N,
+                                    int K,
+                                    int O,
+                                    int BatchCount    = 1,
+                                    int StrideA       = -1,
+                                    int StrideB0      = -1,
+                                    int StrideB1      = -1,
+                                    int StrideC       = -1,
+                                    int BatchStrideA  = -1,
+                                    int BatchStrideB0 = -1,
+                                    int BatchStrideB1 = -1,
+                                    int BatchStrideC  = -1)
+
+{
+
+    using Row           = tensor_layout::gemm::RowMajor;
+    using Col           = tensor_layout::gemm::ColumnMajor;
+    using PassThrough   = tensor_operation::element_wise::PassThrough;
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using B1ElementOp   = PassThrough;
+    using Acc0ElementOp = PassThrough;
+    using CElementOp    = PassThrough;
+    using AccDataType   = float;
+
+    // Ref Gemm0
+    using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+    // Ref Gemm
+    using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+    bool pass = true;
+
+    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
+
+    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
+
+    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    // C_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<CDataType> c_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    // Host verification: Output of Gemm0 is input A of Gemm1
+    Tensor<ADataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 3});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 3});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 3});
+        break;
+    case 2:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) * c_g_m_o_device_result.mDesc.GetElementSize());
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    using DeviceOp = tensor_operation::device::DeviceBatchedGemmGemm<ALayout,
+                                                                     B0Layout,
+                                                                     B1Layout,
+                                                                     CLayout,
+                                                                     ADataType,
+                                                                     B0DataType,
+                                                                     B1DataType,
+                                                                     CDataType,
+                                                                     AElementOp,
+                                                                     B0ElementOp,
+                                                                     Acc0ElementOp,
+                                                                     B1ElementOp,
+                                                                     CElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // early fail when no instances are found
+    if(op_ptrs.size() == 0)
+    {
+        return false;
+    }
+
+    if(do_verification)
+    {
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, PassThrough{});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            acc0_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+            static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+            M,
+            N,
+            K,
+            O,
+            BatchCount,
+            StrideA,
+            StrideB0,
+            StrideB1,
+            StrideC,
+            BatchStrideA,
+            BatchStrideB0,
+            BatchStrideB1,
+            BatchStrideC,
+            a_element_op,
+            b0_element_op,
+            acc0_element_op,
+            b1_element_op,
+            c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                                    BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(c_g_m_o_device_result, c_g_m_o_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a_g_m_k: ", a_g_m_k.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b0_g_k_n : ", b0_g_k_n.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b1_g_n_o : ", b1_g_n_o.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_g_m_o_host_result : ", c_g_m_o_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_g_m_o_device_result : ", c_g_m_o_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_impl.hpp
new file mode 100644
index 0000000..c07d7c0
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_impl.hpp
@@ -0,0 +1,233 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_batched_gemm_impl(int do_verification,
+                               int init_method,
+                               bool do_log,
+                               bool time_kernel,
+                               int M,
+                               int N,
+                               int K,
+                               int BatchStrideA,
+                               int BatchStrideB,
+                               int BatchStrideC,
+                               int StrideA,
+                               int StrideB,
+                               int StrideC,
+                               int BatchCount)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<BDataType> b_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB, BatchStrideB, BLayout{}));
+    Tensor<CDataType> c_g_m_n_host_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_n_device_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, BatchStrideC, CLayout{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    if(do_verification)
+    {
+        using ReferenceBatchedGemmInstance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                             BDataType,
+                                                             CDataType,
+                                                             float,
+                                                             AElementOp,
+                                                             BElementOp,
+                                                             CElementOp>;
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_g_m_k.mData.data());
+    b_device_buf.ToDevice(b_g_k_n.mData.data());
+    c_device_buf.ToDevice(c_g_m_n_device_result.mData.data());
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemm<ALayout,
+                                                                     BLayout,
+                                                                     CLayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     CElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                        static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                        M,
+                                        N,
+                                        K,
+                                        StrideA,
+                                        StrideB,
+                                        StrideC,
+                                        BatchStrideA,
+                                        BatchStrideB,
+                                        BatchStrideC,
+                                        BatchCount,
+                                        ck::tensor_operation::element_wise::PassThrough{},
+                                        ck::tensor_operation::element_wise::PassThrough{},
+                                        ck::tensor_operation::element_wise::PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init C to zero before profiling next kernel
+            c_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
+
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                     sizeof(CDataType) * M * N) *
+                                    BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(c_g_m_n_device_result, c_g_m_n_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp
new file mode 100644
index 0000000..45b7b77
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp
@@ -0,0 +1,362 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32                 = float;
+using F16                 = ck::half_t;
+using ReducePtrsGlobal    = ck::Tuple<F32*, F32*>;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Identity, Identity>;
+
+using DeviceGemmReduceNoOpPtr =
+    ck::tensor_operation::device::DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>;
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ReduceDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_batched_gemm_reduce_impl(int do_verification,
+                                      int init_method,
+                                      bool do_log,
+                                      bool time_kernel,
+                                      int M,
+                                      int N,
+                                      int K,
+                                      int StrideA,
+                                      int StrideB,
+                                      int StrideC,
+                                      int BatchCount)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {row * stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {col * stride, 1_uz, stride});
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(f_host_tensor_descriptor(BatchCount, M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_g_k_n(f_host_tensor_descriptor(BatchCount, K, N, StrideB, BLayout{}));
+
+    Tensor<CDataType> c_g_m_n_host_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> d0_g_m_host_result({BatchCount, M});
+    Tensor<ReduceDataType> d1_g_m_host_result({BatchCount, M});
+
+    Tensor<CDataType> c_g_m_n_device_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> d0_g_m_device_result({BatchCount, M});
+    Tensor<ReduceDataType> d1_g_m_device_result({BatchCount, M});
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
+    std::cout << "d0_g_m: " << d0_g_m_host_result.mDesc << std::endl;
+    std::cout << "d1_g_m: " << d1_g_m_host_result.mDesc << std::endl;
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        std::srand(0);
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        break;
+    default:
+        std::srand(0);
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+    }
+
+    using AElementOp            = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp            = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp            = ck::tensor_operation::element_wise::PassThrough;
+    using ReduceOp0             = ck::reduce::Add;
+    using ReduceOp1             = ck::reduce::Add;
+    using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
+
+    auto a_element_op                     = AElementOp{};
+    auto b_element_op                     = BElementOp{};
+    auto c_element_op                     = CElementOp{};
+    std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
+
+    const auto reduce0_op = ReduceOp0{};
+    const auto reduce1_op = ReduceOp1{};
+
+    auto passthrough                            = UnaryIdenticElementOp{};
+    auto square                                 = UnarySquareElementOp{};
+    std::array<void*, 2> reduce_in_element_ops  = {&passthrough, &square};
+    std::array<void*, 2> reduce_out_element_ops = {&passthrough, &passthrough};
+
+    if(do_verification)
+    {
+        using ReferenceBatchedGemmInstance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                             BDataType,
+                                                             CDataType,
+                                                             float,
+                                                             AElementOp,
+                                                             BElementOp,
+                                                             CElementOp>;
+
+        using ReduceAccDataType = ReduceDataType;
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        for(int batch = 0; batch < BatchCount; ++batch)
+        {
+            for(int m = 0; m < M; ++m)
+            {
+                auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
+
+                for(int n = 0; n < N; ++n)
+                {
+                    ReduceAccDataType d0_val =
+                        ck::type_convert<ReduceAccDataType>(c_g_m_n_host_result(batch, m, n));
+                    ReduceAccDataType d1_val;
+
+                    square(d1_val, d0_val);
+                    reduce0_op(reduce0_acc, d0_val);
+                    reduce1_op(reduce1_acc, d1_val);
+                }
+
+                d0_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce0_acc);
+                d1_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce1_acc);
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
+                                 d0_g_m_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
+                                 d1_g_m_device_result.mDesc.GetElementSpaceSize());
+
+    std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
+                                      reduce1_device_buf.GetDeviceBuffer()};
+
+    a_device_buf.ToDevice(a_g_m_k.mData.data());
+    b_device_buf.ToDevice(b_g_k_n.mData.data());
+
+    // add device GEMM instances
+    std::vector<ck::tensor_operation::device::instance::DeviceGemmReduceNoOpPtr> gemm_ptrs;
+
+    if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
+                 is_same<CDataType, half_t>::value)
+    {
+        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
+                    gemm_ptrs);
+        }
+    }
+
+    if(gemm_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device GEMM instance found");
+    }
+
+    std::string best_gemm_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device GEMM instances
+    for(auto& gemm_ptr : gemm_ptrs)
+    {
+        auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                          b_device_buf.GetDeviceBuffer(),
+                                                          nullptr,
+                                                          {},
+                                                          c_device_buf.GetDeviceBuffer(),
+                                                          p_reduces,
+                                                          M,
+                                                          N,
+                                                          K,
+                                                          StrideA,
+                                                          StrideB,
+                                                          StrideC,
+                                                          {},
+                                                          gemm_element_ops,
+                                                          {},
+                                                          reduce_in_element_ops,
+                                                          reduce_out_element_ops,
+                                                          BatchCount);
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // init DO, D1 to 0
+            reduce0_device_buf.SetZero();
+            reduce1_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::string gemm_name = gemm_ptr->GetTypeString();
+
+            std::size_t flop      = std::size_t(2) * BatchCount * M * N * K;
+            std::size_t num_btype = sizeof(ADataType) * BatchCount * M * K +
+                                    sizeof(BDataType) * BatchCount * K * N +
+                                    sizeof(CDataType) * BatchCount * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << gemm_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_gemm_name  = gemm_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
+                reduce0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
+                reduce1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
+
+                bool c_error  = ck::utils::check_err(c_g_m_n_device_result, c_g_m_n_host_result);
+                bool d0_error = ck::utils::check_err(d0_g_m_device_result, d0_g_m_host_result);
+                bool d1_error = ck::utils::check_err(d1_g_m_device_result, d1_g_m_host_result);
+
+                pass = pass && (c_error == true);
+                pass = pass && (d0_error == true);
+                pass = pass && (d1_error == true);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "d0_host: ", d0_g_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d0_device: ", d0_g_m_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "d1_host: ", d1_g_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d1_device: ", d1_g_m_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << "does not support this GEMM problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp
new file mode 100644
index 0000000..f5ec235
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp
@@ -0,0 +1,347 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout,
+          bool MaskOutUpperTriangle>
+bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
+                                            int init_method,
+                                            bool do_log,
+                                            bool time_kernel,
+                                            int M,
+                                            int N,
+                                            int K,
+                                            int O,
+                                            int BatchCount    = 1,
+                                            int StrideA       = -1,
+                                            int StrideB0      = -1,
+                                            int StrideB1      = -1,
+                                            int StrideC       = -1,
+                                            int BatchStrideA  = -1,
+                                            int BatchStrideB0 = -1,
+                                            int BatchStrideB1 = -1,
+                                            int BatchStrideC  = -1,
+                                            float alpha       = -1.f)
+
+{
+
+    using Row           = tensor_layout::gemm::RowMajor;
+    using Col           = tensor_layout::gemm::ColumnMajor;
+    using PassThrough   = tensor_operation::element_wise::PassThrough;
+    using Scale         = tensor_operation::element_wise::Scale;
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = Scale;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+    using AccDataType   = float;
+
+    // Ref Gemm0: various type in, fp32 out
+    using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+    // Ref Softmax: fp32 in, various type out
+    using ReferenceSoftmaxInstance =
+        tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+    // Ref Gemm1: various type in, various type out
+    using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+    bool pass = true;
+
+    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
+
+    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
+
+    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    // C_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<CDataType> c_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    // Host verification: Output of Gemm0 is input A of Gemm1
+    Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+    Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
+
+    std::srand(1); // work around test flakiness
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        // Still unsure whether this kind of deterministic floating point accurary issue is expected
+        // or not. May want to try exact same approach as the GPU kernel in the host reference
+        // GEMM+Softmax+GEMM function to see if the accuracy discrepancy goes away. Until then,
+        // shrink the input value range as it is less likely to produce errors of around ~1e-3.
+        // a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        // b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        // b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        break;
+    case 2:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) * c_g_m_o_device_result.mDesc.GetElementSize());
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+
+    if(alpha < 0)
+    {
+        alpha = 1.f / std::sqrt(K); // usually 1 / sqrt(head_dim)
+    }
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    using DeviceOp = tensor_operation::device::DeviceBatchedGemmSoftmaxGemm<ALayout,
+                                                                            B0Layout,
+                                                                            B1Layout,
+                                                                            CLayout,
+                                                                            ADataType,
+                                                                            B0DataType,
+                                                                            B1DataType,
+                                                                            CDataType,
+                                                                            AElementOp,
+                                                                            B0ElementOp,
+                                                                            Acc0ElementOp,
+                                                                            B1ElementOp,
+                                                                            CElementOp,
+                                                                            MaskOutUpperTriangle>;
+
+    // get device op instances
+    const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, Scale{alpha});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        // mask out upper triangle
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(MaskOutUpperTriangle && idx[1] < idx[2])
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+            static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+            M,
+            N,
+            K,
+            O,
+            BatchCount,
+            StrideA,
+            StrideB0,
+            StrideB1,
+            StrideC,
+            BatchStrideA,
+            BatchStrideB0,
+            BatchStrideB1,
+            BatchStrideC,
+            a_element_op,
+            b0_element_op,
+            acc0_element_op,
+            b1_element_op,
+            c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                                    BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(c_g_m_o_device_result, c_g_m_o_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a_g_m_k: ", a_g_m_k.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b0_g_k_n : ", b0_g_k_n.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b1_g_n_o : ", b1_g_n_o.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_g_m_o_host_result : ", c_g_m_o_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_g_m_o_device_result : ", c_g_m_o_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp
new file mode 100644
index 0000000..91c28f2
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp
@@ -0,0 +1,367 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasesDataType,
+          typename Acc1BiasesDataType,
+          tensor_operation::device::MaskingSpecialization MaskingSpec>
+bool profile_batched_gemm_softmax_gemm_permute_impl(bool do_verification,
+                                                    int init_method,
+                                                    bool do_log,
+                                                    bool time_kernel,
+                                                    int M,
+                                                    int N,
+                                                    int K,
+                                                    int O,
+                                                    int G0,
+                                                    int G1,
+                                                    float alpha = -1.f)
+
+{
+
+    using PassThrough   = tensor_operation::element_wise::PassThrough;
+    using Scale         = tensor_operation::element_wise::Scale;
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = Scale;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+    using AccDataType   = float;
+    using tensor_operation::device::MaskingSpecialization;
+
+    // Ref Gemm0: various type in, fp32 out
+    using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+    // Ref Softmax: fp32 in, various type out
+    using ReferenceSoftmaxInstance =
+        tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+    // Ref Gemm1: various type in, various type out
+    using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+    bool pass = true;
+
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+    const int BatchCount = G0 * G1;
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
+    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
+
+    std::srand(1); // work around test flakiness
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        // Still unsure whether this kind of deterministic floating point accurary issue is expected
+        // or not. May want to try exact same approach as the GPU kernel in the host reference
+        // GEMM+Softmax+GEMM function to see if the accuracy discrepancy goes away. Until then,
+        // shrink the input value range as it is less likely to produce errors of around ~1e-3.
+        // a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        // b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        // b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) *
+                           c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
+
+    if(alpha < 0)
+    {
+        alpha = 1.f / std::sqrt(K); // usually 1 / sqrt(head_dim)
+    }
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    using DeviceOp = tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                                   1,
+                                                                                   1,
+                                                                                   1,
+                                                                                   1,
+                                                                                   ADataType,
+                                                                                   B0DataType,
+                                                                                   B1DataType,
+                                                                                   CDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   ck::Tuple<>,
+                                                                                   AElementOp,
+                                                                                   B0ElementOp,
+                                                                                   Acc0ElementOp,
+                                                                                   B1ElementOp,
+                                                                                   CElementOp,
+                                                                                   MaskingSpec>;
+
+    // get device op instances
+    const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, Scale{alpha});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        // mask out upper triangle
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle && idx[1] < idx[2])
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // permute
+        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+        });
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+            static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+            {}, // std::array<void*, 1> p_acc0_biases;
+            {}, // std::array<void*, 1> p_acc1_biases;
+            a_gs_ms_ks_lengths,
+            a_gs_ms_ks_strides,
+            b0_gs_ns_ks_lengths,
+            b0_gs_ns_ks_strides,
+            b1_gs_os_ns_lengths,
+            b1_gs_os_ns_strides,
+            c_gs_ms_os_lengths,
+            c_gs_ms_os_strides,
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
+            a_element_op,
+            b0_element_op,
+            acc0_element_op,
+            b1_element_op,
+            c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                                    BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+                // default absolute error and relative error is 0.001
+                double rtol = 1e-3;
+                double atol = 1e-3;
+
+                // when BF16 is taken, set absolute error and relative error to 0.01
+                if(std::is_same_v<ADataType, ck::bhalf_t> &&
+                   std::is_same_v<B0DataType, ck::bhalf_t> &&
+                   std::is_same_v<B1DataType, ck::bhalf_t> &&
+                   std::is_same_v<CDataType, ck::bhalf_t>)
+                {
+                    rtol = 1e-2;
+                    atol = 1e-2;
+                }
+
+                pass = pass & ck::utils::check_err(c_gs_ms_os_device_result,
+                                                   c_gs_ms_os_host_result,
+                                                   "Error: Incorrect results!",
+                                                   rtol,
+                                                   atol);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a_gs_ms_ks: ", a_gs_ms_ks.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b0_gs_ns_ks : ", b0_gs_ns_ks.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b1_gs_os_ns : ", b1_gs_os_ns.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_gs_ms_os_host_result : ", c_gs_ms_os_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_gs_ms_os_device_result : ",
+                                          c_gs_ms_os_device_result.mData,
+                                          ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_batchnorm_backward_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_batchnorm_backward_impl.hpp
new file mode 100644
index 0000000..79d8862
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_batchnorm_backward_impl.hpp
@@ -0,0 +1,390 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <stdexcept>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+bool profile_batchnorm_backward_impl(bool do_verification,
+                                     int init_method,
+                                     bool do_dumpout,
+                                     bool time_kernel,
+                                     const std::vector<size_t> inOutLengths,
+                                     const std::vector<int> reduceDims,
+                                     bool haveSavedMeanInvVar,
+                                     double epsilon)
+{
+    if(inOutLengths.size() != Rank || reduceDims.size() != NumBatchNormReduceDim)
+    {
+        throw std::runtime_error("Invalid tensor lengths or number of reduce dimensions!");
+    };
+
+    std::vector<size_t> scaleBiasMeanVarLengths;
+
+    // used for calculating the effective transferred bytes by each operation
+    size_t total_length;
+    size_t invariant_length = 1;
+
+    total_length =
+        std::accumulate(inOutLengths.begin(), inOutLengths.end(), 1, std::multiplies<size_t>{});
+
+    if(std::any_of(reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+        throw std::runtime_error("Invalid reduce dimensions!");
+
+    for(int dim = 0; dim < Rank; dim++)
+    {
+        if(std::none_of(reduceDims.begin(), reduceDims.end(), [&](int d) { return dim == d; }))
+        {
+            scaleBiasMeanVarLengths.push_back(inOutLengths[dim]);
+            invariant_length *= inOutLengths[dim];
+        };
+    }
+
+    // input data of the batchnorm backward algorithm
+    Tensor<XDataType> x(inOutLengths);
+    Tensor<DyDataType> dy(inOutLengths);
+    Tensor<ScaleDataType> bnScale(scaleBiasMeanVarLengths);
+
+    Tensor<MeanVarDataType> savedMean(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> savedInvVar(scaleBiasMeanVarLengths);
+    // savedVariance is only used for initializing savedInvVar
+    Tensor<MeanVarDataType> savedVariance(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm backward algorithm
+    Tensor<DxDataType> dx_ref(inOutLengths);
+    Tensor<DxDataType> dx(inOutLengths);
+
+    Tensor<DscaleDbiasDataType> dscale(scaleBiasMeanVarLengths);
+    Tensor<DscaleDbiasDataType> dbias(scaleBiasMeanVarLengths);
+
+    Tensor<DscaleDbiasDataType> dscale_ref(scaleBiasMeanVarLengths);
+    Tensor<DscaleDbiasDataType> dbias_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(haveSavedMeanInvVar)
+    {
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 1.0f;
+        const float noise_stddev = 0.0001f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+
+        // initialize the savedMean to be values with tiny variation to the mean of the x values
+        savedMean.GenerateTensorValue(GeneratorTensor_4<MeanVarDataType>{x_mean, noise_stddev},
+                                      num_thread);
+
+        // initialize the variance to be values with tiny variation to the variance of the x values
+        savedVariance.GenerateTensorValue(
+            GeneratorTensor_4<MeanVarDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+
+        auto it_src       = savedVariance.mData.begin();
+        auto it_dst       = savedInvVar.mData.begin();
+        float tmp_epsilon = std::numeric_limits<float>::epsilon();
+
+        while(it_src != savedVariance.mData.end())
+        {
+            *it_dst = type_convert<AccDataType>(
+                1.0f / std::sqrtf(type_convert<float>(*it_src) + tmp_epsilon));
+
+            it_src++;
+            it_dst++;
+        };
+    }
+    else
+    {
+        const float x_mean   = 0.0f;
+        const float x_stddev = 1.0f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            dy.GenerateTensorValue(GeneratorTensor_0<DyDataType>{}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_0<ScaleDataType>{}, num_thread);
+            break;
+        case 1:
+            dy.GenerateTensorValue(GeneratorTensor_1<DyDataType>{1}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_1<ScaleDataType>{1}, num_thread);
+            break;
+        case 2:
+            dy.GenerateTensorValue(GeneratorTensor_2<DyDataType>{-2, 2}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_2<ScaleDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            dy.GenerateTensorValue(GeneratorTensor_3<DyDataType>{-0.2f, 0.2f}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_3<ScaleDataType>{-0.5f, 0.5f}, num_thread);
+        }
+    };
+
+    // input data of the batchnorm backward algorithm
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem dy_dev(sizeof(DyDataType) * dy.mDesc.GetElementSpaceSize());
+
+    DeviceMem bnScale_dev(sizeof(ScaleDataType) * bnScale.mDesc.GetElementSpaceSize());
+
+    DeviceMem savedMean_dev(sizeof(MeanVarDataType) * savedMean.mDesc.GetElementSpaceSize());
+    DeviceMem savedInvVar_dev(sizeof(MeanVarDataType) * savedInvVar.mDesc.GetElementSpaceSize());
+
+    // output data of the batchnorm backward algorithm
+    DeviceMem dx_dev(sizeof(DxDataType) * dx.mDesc.GetElementSpaceSize());
+
+    DeviceMem dscale_dev(sizeof(DscaleDbiasDataType) * dscale.mDesc.GetElementSpaceSize());
+    DeviceMem dbias_dev(sizeof(DscaleDbiasDataType) * dbias.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    dy_dev.ToDevice(dy.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+
+    if(haveSavedMeanInvVar)
+    {
+        savedMean_dev.ToDevice(savedMean.mData.data());
+        savedInvVar_dev.ToDevice(savedInvVar.mData.data());
+    };
+
+    std::array<index_t, Rank> arrInOutLengths;
+    std::array<index_t, Rank> arrInOutStrides;
+    std::array<index_t, Rank - NumBatchNormReduceDim> arrScaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumBatchNormReduceDim> arrScaleBiasMeanVarStrides;
+    std::array<int, NumBatchNormReduceDim> arrReduceDims;
+
+    std::copy(inOutLengths.begin(), inOutLengths.end(), arrInOutLengths.begin());
+    std::copy(inOutStrides.begin(), inOutStrides.end(), arrInOutStrides.begin());
+    std::copy(scaleBiasMeanVarLengths.begin(),
+              scaleBiasMeanVarLengths.end(),
+              arrScaleBiasMeanVarLengths.begin());
+    std::copy(scaleBiasMeanVarStrides.begin(),
+              scaleBiasMeanVarStrides.end(),
+              arrScaleBiasMeanVarStrides.begin());
+
+    std::copy(reduceDims.begin(), reduceDims.end(), arrReduceDims.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    // add device batchnorm-backward instances
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
+                                                                      DxDataType,
+                                                                      DxDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      DscaleDbiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThroughOp,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferenceBatchNormBwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormBwd<XDataType,
+                                                              DxDataType,
+                                                              DyDataType,
+                                                              AccDataType,
+                                                              ScaleDataType,
+                                                              DscaleDbiasDataType,
+                                                              MeanVarDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumBatchNormReduceDim>;
+
+        auto batchNormBwd_ref = ReferenceBatchNormBwdInstance{};
+
+        auto argument_ptr_ref = batchNormBwd_ref.MakeArgumentPointer(
+            arrInOutLengths,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrReduceDims,
+            arrScaleBiasMeanVarLengths,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            x.mData.data(),
+            dy.mData.data(),
+            bnScale.mData.data(),
+            haveSavedMeanInvVar ? savedMean.mData.data() : nullptr,
+            haveSavedMeanInvVar ? savedInvVar.mData.data() : nullptr,
+            epsilon,
+            PassThroughOp{},
+            dx_ref.mData.data(),
+            dscale_ref.mData.data(),
+            dbias_ref.mData.data());
+
+        if(!batchNormBwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reference instance, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormBwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+    }
+
+    int num_kernel = 0;
+    bool pass      = true;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            arrInOutLengths,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrReduceDims,
+            arrScaleBiasMeanVarLengths,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            x_dev.GetDeviceBuffer(),
+            dy_dev.GetDeviceBuffer(),
+            bnScale_dev.GetDeviceBuffer(),
+            haveSavedMeanInvVar ? savedMean_dev.GetDeviceBuffer() : nullptr,
+            haveSavedMeanInvVar ? savedInvVar_dev.GetDeviceBuffer() : nullptr,
+            epsilon,
+            PassThroughOp{},
+            dx_dev.GetDeviceBuffer(),
+            dscale_dev.GetDeviceBuffer(),
+            dbias_dev.GetDeviceBuffer());
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            num_kernel++;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString()
+                          << " skipped due to unsupported argument: " << std::endl;
+            }
+
+            continue;
+        };
+
+        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+        DeviceMem workspace_dev(workspace_sz);
+
+        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        size_t num_bytes = 0;
+
+        // inputing of x, dy, scale, outputing of dx, dscale, dbias
+        num_bytes += total_length * (sizeof(XDataType) + sizeof(DyDataType) + sizeof(DxDataType)) +
+                     invariant_length * sizeof(DscaleDbiasDataType) * 2;
+
+        // inputting of savedMean, savedInvVariance
+        if(haveSavedMeanInvVar)
+            num_bytes += invariant_length * sizeof(MeanVarDataType) * 2;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            using ck::utils::check_err;
+            bool single_pass = true;
+
+            dx_dev.FromDevice(dx.mData.data());
+            dscale_dev.FromDevice(dscale.data());
+            dbias_dev.FromDevice(dbias.data());
+
+            // clang-format off
+            single_pass = single_pass && ck::utils::check_err(dx.mData, dx_ref.mData, "dx result:", 5e-4, 5e-4);
+            single_pass = single_pass && ck::utils::check_err(dscale.mData, dscale_ref.mData, "dScale result:", 3e-3, 3e-3);
+            single_pass = single_pass && ck::utils::check_err(dbias.mData, dbias_ref.mData, "dBias result:", 3e-3, 3e-3);
+            // clang-format on
+
+            pass = pass && single_pass;
+        };
+
+        if(do_dumpout)
+        {
+            using ck::host_common::dumpBufferToFile;
+
+            // clang-format off
+            dumpBufferToFile("dump_x.bin", x.mData.data(), x.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dy.bin", dy.mData.data(), dy.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dx.bin", dx.mData.data(), dx.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dx_ref.bin", dx_ref.mData.data(), dx_ref.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dscale.bin", dscale.mData.data(), dscale.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dscale_ref.bin", dscale_ref.mData.data(), dscale_ref.mDesc.GetElementSize());
+            // clang-format off
+        };
+    }
+
+    if(time_kernel)
+    {
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_batchnorm_forward_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_batchnorm_forward_impl.hpp
new file mode 100644
index 0000000..82fe75b
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_batchnorm_forward_impl.hpp
@@ -0,0 +1,412 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <stdexcept>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+bool profile_batchnorm_forward_impl(int do_verification,
+                                    int init_method,
+                                    bool do_dumpout,
+                                    bool time_kernel,
+                                    const std::vector<size_t> inOutLengths,
+                                    const std::vector<int> reduceDims,
+                                    bool updateMovingAverage,
+                                    bool saveMeanAndInvVariance,
+                                    double averageFactor,
+                                    double epsilon)
+{
+    if(inOutLengths.size() != Rank || reduceDims.size() != NumBatchNormReduceDim)
+    {
+        throw std::runtime_error("Invalid tensor lengths or number of reduce dimensions!");
+    };
+
+    std::vector<size_t> scaleBiasMeanVarLengths;
+
+    // used for calculating the effective transferred bytes by each operation
+    size_t total_length;
+    size_t invariant_length = 1;
+
+    total_length =
+        std::accumulate(inOutLengths.begin(), inOutLengths.end(), 1, std::multiplies<size_t>{});
+
+    if(std::any_of(reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+        throw std::runtime_error("Invalid reduce dimensions!");
+
+    for(int dim = 0; dim < Rank; dim++)
+    {
+        if(std::none_of(reduceDims.begin(), reduceDims.end(), [&](int d) { return dim == d; }))
+        {
+            scaleBiasMeanVarLengths.push_back(inOutLengths[dim]);
+            invariant_length *= inOutLengths[dim];
+        };
+    }
+
+    // input data of the batchnorm forward algorithm
+    Tensor<XDataType> x(inOutLengths);
+    Tensor<ScaleDataType> bnScale(scaleBiasMeanVarLengths);
+    Tensor<BiasDataType> bnBias(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm forward algorithm
+    Tensor<YDataType> y_ref(inOutLengths);
+    Tensor<YDataType> y(inOutLengths);
+
+    Tensor<MeanVarDataType> resultSaveMean_ref(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> resultSaveInvVariance_ref(scaleBiasMeanVarLengths);
+
+    Tensor<MeanVarDataType> resultRunningMean_ref(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> resultRunningVariance_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(updateMovingAverage)
+    {
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 1.0f;
+        const float noise_stddev = 0.04f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+
+        // initialize the runningMean to be values with tiny variation to the mean of the x
+        // values
+        resultRunningMean_ref.GenerateTensorValue(
+            GeneratorTensor_4<MeanVarDataType>{x_mean, noise_stddev}, num_thread);
+
+        // initialize the runningVariance to be values with tiny variation to the variance of
+        // the x values
+        resultRunningVariance_ref.GenerateTensorValue(
+            GeneratorTensor_4<MeanVarDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+    }
+    else
+    {
+        if constexpr(ck::is_same_v<XDataType, int8_t>)
+            x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5}, num_thread);
+        else
+            x.GenerateTensorValue(GeneratorTensor_3<XDataType>{-1.0f, 1.0f}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            bnScale.GenerateTensorValue(GeneratorTensor_0<ScaleDataType>{}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_0<BiasDataType>{}, num_thread);
+            break;
+        case 1:
+            bnScale.GenerateTensorValue(GeneratorTensor_1<ScaleDataType>{1}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_1<BiasDataType>{0}, num_thread);
+            break;
+        case 2:
+            bnScale.GenerateTensorValue(GeneratorTensor_2<ScaleDataType>{-5, 5}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            bnScale.GenerateTensorValue(GeneratorTensor_3<ScaleDataType>{-1.0f, 1.0f}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_3<BiasDataType>{-1.0f, 1.0f}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(XDataType) * y.mDesc.GetElementSpaceSize());
+    DeviceMem bnScale_dev(sizeof(ScaleDataType) * bnScale.mDesc.GetElementSpaceSize());
+    DeviceMem bnBias_dev(sizeof(BiasDataType) * bnBias.mDesc.GetElementSpaceSize());
+
+    // mean_dev or resultSaveMean_dev
+    DeviceMem resultSaveMean_dev(sizeof(MeanVarDataType) *
+                                 resultSaveMean_ref.mDesc.GetElementSpaceSize());
+    // meansquare_dev or resultSaveInvVariance_dev
+    DeviceMem resultSaveInvVariance_dev(sizeof(MeanVarDataType) *
+                                        resultSaveInvVariance_ref.mDesc.GetElementSpaceSize());
+    // resultRunningMean_dev
+    DeviceMem resultRunningMean_dev(sizeof(MeanVarDataType) *
+                                    resultRunningMean_ref.mDesc.GetElementSpaceSize());
+    // resultRunningVariance_dev
+    DeviceMem resultRunningVariance_dev(sizeof(MeanVarDataType) *
+                                        resultRunningVariance_ref.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+    bnBias_dev.ToDevice(bnBias.mData.data());
+
+    if(updateMovingAverage)
+    {
+        resultRunningMean_dev.ToDevice(resultRunningMean_ref.mData.data());
+        resultRunningVariance_dev.ToDevice(resultRunningVariance_ref.mData.data());
+    };
+
+    // used for storing the device result for verification when updateMovingAverage is enabled
+    Tensor<MeanVarDataType> resultRunningMean(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> resultRunningVariance(scaleBiasMeanVarLengths);
+
+    // used for storing the device result for verification when saveMeanAndInvVariance is enabled
+    Tensor<MeanVarDataType> resultSaveMean(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> resultSaveInvVariance(scaleBiasMeanVarLengths);
+
+    std::array<index_t, Rank> arrInOutLengths;
+    std::array<index_t, Rank> arrInOutStrides;
+    std::array<index_t, Rank - NumBatchNormReduceDim> arrScaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumBatchNormReduceDim> arrScaleBiasMeanVarStrides;
+    std::array<int, NumBatchNormReduceDim> arrReduceDims;
+
+    std::copy(inOutLengths.begin(), inOutLengths.end(), arrInOutLengths.begin());
+    std::copy(inOutStrides.begin(), inOutStrides.end(), arrInOutStrides.begin());
+    std::copy(scaleBiasMeanVarLengths.begin(),
+              scaleBiasMeanVarLengths.end(),
+              arrScaleBiasMeanVarLengths.begin());
+    std::copy(scaleBiasMeanVarStrides.begin(),
+              scaleBiasMeanVarStrides.end(),
+              arrScaleBiasMeanVarStrides.begin());
+
+    std::copy(reduceDims.begin(), reduceDims.end(), arrReduceDims.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    // add device batchnorm-forward instances
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                                      YDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      BiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThroughOp,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferenceBatchNormFwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormFwd<XDataType,
+                                                              YDataType,
+                                                              AccDataType,
+                                                              ScaleDataType,
+                                                              BiasDataType,
+                                                              MeanVarDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumBatchNormReduceDim>;
+
+        auto batchNormFwd_ref = ReferenceBatchNormFwdInstance{};
+
+        auto argument_ptr_ref = batchNormFwd_ref.MakeArgumentPointer(
+            arrInOutLengths,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrReduceDims,
+            arrScaleBiasMeanVarLengths,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            x.mData.data(),
+            bnScale.mData.data(),
+            bnBias.mData.data(),
+            epsilon,
+            PassThroughOp{},
+            y_ref.mData.data(),
+            saveMeanAndInvVariance ? resultSaveMean_ref.mData.data() : nullptr,
+            saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr,
+            averageFactor,
+            updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr,
+            updateMovingAverage ? resultRunningVariance_ref.mData.data() : nullptr);
+
+        if(!batchNormFwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reference instance, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormFwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+    }
+
+    int num_kernel = 0;
+    bool pass      = true;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            arrInOutLengths,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrReduceDims,
+            arrScaleBiasMeanVarLengths,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            x_dev.GetDeviceBuffer(),
+            bnScale_dev.GetDeviceBuffer(),
+            bnBias_dev.GetDeviceBuffer(),
+            epsilon,
+            PassThroughOp{},
+            y_dev.GetDeviceBuffer(),
+            saveMeanAndInvVariance ? resultSaveMean_dev.GetDeviceBuffer() : nullptr,
+            saveMeanAndInvVariance ? resultSaveInvVariance_dev.GetDeviceBuffer() : nullptr,
+            averageFactor,
+            updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
+            updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr);
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            num_kernel++;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString()
+                          << " skipped due to unsupported argument: " << std::endl;
+            }
+
+            continue;
+        };
+
+        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+        DeviceMem workspace_dev(workspace_sz);
+
+        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        size_t num_bytes = 0;
+
+        // inputing of x, scale, bias, outputing of y
+        num_bytes += total_length * (sizeof(XDataType) + sizeof(YDataType)) +
+                     invariant_length * (sizeof(ScaleDataType) + sizeof(BiasDataType));
+
+        // outputing of mean, inv-variance
+        num_bytes += saveMeanAndInvVariance ? invariant_length * sizeof(MeanVarDataType) * 2 : 0;
+
+        // updating of moving mean, variance
+        num_bytes += updateMovingAverage ? invariant_length * sizeof(MeanVarDataType) * 4 : 0;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            using ck::utils::check_err;
+            bool single_pass;
+
+            y_dev.FromDevice(y.mData.data());
+
+            if constexpr(ck::is_same_v<YDataType, ck::bhalf_t>)
+                single_pass = check_err(y.mData, y_ref.mData, "y results", 1e-2, 1e-2);
+            else
+                single_pass = check_err(y.mData, y_ref.mData, "y results", 4e-3, 4e-3);
+
+            if(updateMovingAverage)
+            {
+                resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
+                resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
+
+                // clang-format off
+                single_pass = single_pass && check_err(resultRunningMean.mData, resultRunningMean_ref.mData, "average mean results", 1.5e-5, 1.5e-5);
+                single_pass = single_pass && check_err(resultRunningVariance.mData, resultRunningVariance_ref.mData, "average variance results", 1e-5, 1e-5);
+                // clang-format on
+            };
+
+            if(saveMeanAndInvVariance)
+            {
+                resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
+                resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
+
+                // clang-format off
+                single_pass = single_pass && check_err(resultSaveMean.mData, resultSaveMean_ref.mData, "mean results", 3e-5, 3e-5);
+                single_pass = single_pass && check_err(resultSaveInvVariance.mData, resultSaveInvVariance_ref.mData, "inv-variance results", 7e-5, 7e-5);
+                // clang-format on
+            };
+
+            pass = pass && single_pass;
+        };
+
+        if(do_dumpout)
+        {
+            using ck::host_common::dumpBufferToFile;
+
+            // clang-format off
+            dumpBufferToFile("dump_x.bin", x.mData.data(), x.mDesc.GetElementSize());
+            dumpBufferToFile("dump_y.bin", y.mData.data(), y.mDesc.GetElementSize());
+            dumpBufferToFile("dump_y_ref.bin", y_ref.mData.data(), y_ref.mDesc.GetElementSize());
+            // clang-format off
+
+            if(saveMeanAndInvVariance)
+            {
+                // clang-format off
+                dumpBufferToFile("dump_mean.bin", resultSaveMean.mData.data(), resultSaveMean.mDesc.GetElementSize());
+                dumpBufferToFile("dump_mean_ref.bin", resultSaveMean_ref.mData.data(), resultSaveMean_ref.mDesc.GetElementSize()); 
+                dumpBufferToFile("dump_invvar.bin", resultSaveInvVariance.mData.data(), resultSaveInvVariance.mDesc.GetElementSize());
+                dumpBufferToFile("dump_invvar_ref.bin", resultSaveInvVariance_ref.mData.data(), resultSaveInvVariance_ref.mDesc.GetElementSize());
+                // clang-format on
+            };
+        };
+    }
+
+    if(time_kernel)
+    {
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_conv_bwd_data_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_conv_bwd_data_impl.hpp
new file mode 100644
index 0000000..86d394d
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_conv_bwd_data_impl.hpp
@@ -0,0 +1,248 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename DataType>
+void show_data_nhwc_layout(Tensor<DataType>& nhwc)
+{
+    std::cout << "[";
+    for(int n = 0; n < ck::type_convert<int>(nhwc.mDesc.GetLengths()[0]); n++)
+    {
+        std::cout << "[";
+        for(int hi = 0; hi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[2]); hi++)
+        {
+            std::cout << "[";
+            for(int wi = 0; wi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[3]); wi++)
+            {
+                std::cout << "[";
+                for(int c = 0; c < ck::type_convert<int>(nhwc.mDesc.GetLengths()[1]); c++)
+                {
+                    std::cout << static_cast<float>(nhwc(n, c, hi, wi)) << "  ";
+                }
+                std::cout << "]";
+            }
+            std::cout << "]";
+        }
+        std::cout << "]";
+    }
+    std::cout << "]";
+}
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+bool profile_conv_bwd_data_impl(int do_verification,
+                                int init_method,
+                                bool do_log,
+                                bool time_kernel,
+                                const ck::utils::conv::ConvParam& conv_param)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    Tensor<InDataType> input_host_result(in_g_n_c_wis_desc);
+    Tensor<InDataType> input_device_result(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> output(out_g_n_k_wos_desc);
+
+    std::cout << "input: " << input_host_result.mDesc << std::endl;
+    std::cout << "weight: " << weight.mDesc << std::endl;
+    std::cout << "output: " << output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        output.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        output.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+        weight.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * output.mDesc.GetElementSpaceSize());
+
+    out_device_buf.ToDevice(output.mData.data());
+    wei_device_buf.ToDevice(weight.mData.data());
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
+                                                                         InDataType,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         OutElementOp>{};
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(input_host_result,
+                                                  weight,
+                                                  output,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  InElementOp{},
+                                                  WeiElementOp{},
+                                                  OutElementOp{});
+        ref_invoker.Run(ref_argument);
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceConvBwdData<NDimSpatial,
+                                                                     InLayout,
+                                                                     WeiLayout,
+                                                                     OutLayout,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device Conv instances
+    bool pass = true;
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                        conv_param.N_,
+                                        conv_param.K_,
+                                        conv_param.C_,
+                                        conv_param.input_spatial_lengths_,
+                                        conv_param.filter_spatial_lengths_,
+                                        conv_param.output_spatial_lengths_,
+                                        conv_param.conv_filter_strides_,
+                                        conv_param.conv_filter_dilations_,
+                                        conv_param.input_left_pads_,
+                                        conv_param.input_right_pads_,
+                                        in_element_op,
+                                        wei_element_op,
+                                        out_element_op);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // for conv bwd data, some input tensor element are zero, but not written by kernel,
+            // need to set zero
+            in_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s" << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                in_device_buf.FromDevice(input_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(input_device_result, input_host_result);
+
+                if(do_log)
+                {
+                    std::cout << "in : ";
+                    show_data_nhwc_layout(output);
+                    std::cout << std::endl;
+
+                    std::cout << "wei: ";
+                    show_data_nhwc_layout(weight);
+                    std::cout << std::endl;
+
+                    std::cout << "out_host  : ";
+                    show_data_nhwc_layout(input_host_result);
+                    std::cout << std::endl;
+
+                    std::cout << "out_device: ";
+                    show_data_nhwc_layout(input_device_result);
+                    std::cout << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp
new file mode 100644
index 0000000..1aebef8
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp
@@ -0,0 +1,278 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using DeviceConvFwdBiasReluAddPtr =
+    DeviceConvFwdBiasActivationAddPtr<ck::tensor_operation::element_wise::PassThrough,
+                                      ck::tensor_operation::element_wise::PassThrough,
+                                      ck::tensor_operation::element_wise::AddReluAdd>;
+
+void add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<DeviceConvFwdBiasReluAddPtr>&);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <int NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout>
+void profile_conv_fwd_bias_relu_add_impl(int do_verification,
+                                         int init_method,
+                                         bool do_log,
+                                         bool time_kernel,
+                                         ck::index_t N,
+                                         ck::index_t K,
+                                         ck::index_t C,
+                                         std::vector<ck::index_t> input_spatial_lengths,
+                                         std::vector<ck::index_t> filter_spatial_lengths,
+                                         std::vector<ck::index_t> output_spatial_lengths,
+                                         std::vector<ck::index_t> conv_filter_strides,
+                                         std::vector<ck::index_t> conv_filter_dilations,
+                                         std::vector<ck::index_t> input_left_pads,
+                                         std::vector<ck::index_t> input_right_pads)
+{
+    const ck::index_t Y = filter_spatial_lengths[0];
+    const ck::index_t X = filter_spatial_lengths[1];
+
+    const ck::index_t Hi = input_spatial_lengths[0];
+    const ck::index_t Wi = input_spatial_lengths[1];
+
+    const ck::index_t Ho = output_spatial_lengths[0];
+    const ck::index_t Wo = output_spatial_lengths[1];
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
+            using namespace ck::literals;
+
+            if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
+                         is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
+                         is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, H * W, W, 1_uz});
+            }
+            else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
+                              is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
+                              is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
+            }
+        };
+
+    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
+    Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
+    Tensor<OutDataType> out_n_k_ho_wo_host_result(
+        f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+    Tensor<OutDataType> out_n_k_ho_wo_device_result(
+        f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+
+    // bias: assume contiguous 1d vector
+    Tensor<OutDataType> bias_k({K});
+
+    // residual: assume same layout as output tensor
+    Tensor<OutDataType> resi_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
+    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo_host_result.mDesc << std::endl;
+    std::cout << "bias_k: " << bias_k.mDesc << std::endl;
+    std::cout << "resi_n_k_ho_wo: " << resi_n_k_ho_wo.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        bias_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        resi_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        break;
+    default:
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        bias_k.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+        resi_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+    }
+
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(do_verification)
+    {
+        using ReferenceConvFwdInstance =
+            ck::tensor_operation::host::ReferenceConvFwd_Bias_Activation_Add<InDataType,
+                                                                             WeiDataType,
+                                                                             OutDataType,
+                                                                             InElementOp,
+                                                                             WeiElementOp,
+                                                                             OutElementOp>;
+
+        auto ref_conv    = ReferenceConvFwdInstance{};
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
+                                                  wei_k_c_y_x,
+                                                  out_n_k_ho_wo_host_result,
+                                                  bias_k,
+                                                  resi_n_k_ho_wo,
+                                                  conv_filter_strides,
+                                                  conv_filter_dilations,
+                                                  input_left_pads,
+                                                  input_right_pads,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_k_ho_wo_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(OutDataType) * bias_k.mDesc.GetElementSpaceSize());
+    DeviceMem resi_device_buf(sizeof(OutDataType) * resi_n_k_ho_wo.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+    wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
+    bias_device_buf.ToDevice(bias_k.mData.data());
+    resi_device_buf.ToDevice(resi_n_k_ho_wo.mData.data());
+
+    using DeviceConvFwdBiasReluAddPtr = ck::tensor_operation::device::
+        DeviceConvFwdBiasActivationAddPtr<InElementOp, WeiElementOp, OutElementOp>;
+
+    // add device operator instances
+    std::vector<DeviceConvFwdBiasReluAddPtr> op_ptrs;
+
+    if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
+                 ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
+                 ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
+    {
+        ck::tensor_operation::device::instance::
+            add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+    }
+
+    if(op_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device Conv instance found");
+    }
+
+    std::string best_conv_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device Conv instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<const InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<const WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()),
+            static_cast<const OutDataType*>(resi_device_buf.GetDeviceBuffer()),
+            N,
+            K,
+            C,
+            input_spatial_lengths,
+            filter_spatial_lengths,
+            output_spatial_lengths,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            in_element_op,
+            wei_element_op,
+            out_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string conv_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
+
+            std::size_t num_btype =
+                sizeof(InDataType) * (N * C * Hi * Wi) + sizeof(WeiDataType) * (K * C * Y * X) +
+                sizeof(OutDataType) * (N * K * Ho * Wo) + sizeof(OutDataType) * (K) +
+                sizeof(OutDataType) * (N * K * Ho * Wo);
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << conv_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_conv_name  = conv_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
+
+                ck::utils::check_err(out_n_k_ho_wo_device_result, out_n_k_ho_wo_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "wei: ", wei_k_c_y_x.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "out_host  : ", out_n_k_ho_wo_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "out_device: ", out_n_k_ho_wo_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp
new file mode 100644
index 0000000..2bac144
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp
@@ -0,0 +1,266 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using DeviceConvFwdBiasReluPtr =
+    DeviceConvFwdBiasActivationPtr<ck::tensor_operation::element_wise::PassThrough,
+                                   ck::tensor_operation::element_wise::PassThrough,
+                                   ck::tensor_operation::element_wise::AddRelu>;
+
+void add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<DeviceConvFwdBiasReluPtr>&);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <int NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout>
+void profile_conv_fwd_bias_relu_impl(int do_verification,
+                                     int init_method,
+                                     bool do_log,
+                                     bool time_kernel,
+                                     ck::index_t N,
+                                     ck::index_t K,
+                                     ck::index_t C,
+                                     std::vector<ck::index_t> input_spatial_lengths,
+                                     std::vector<ck::index_t> filter_spatial_lengths,
+                                     std::vector<ck::index_t> output_spatial_lengths,
+                                     std::vector<ck::index_t> conv_filter_strides,
+                                     std::vector<ck::index_t> conv_filter_dilations,
+                                     std::vector<ck::index_t> input_left_pads,
+                                     std::vector<ck::index_t> input_right_pads)
+{
+    const ck::index_t Y = filter_spatial_lengths[0];
+    const ck::index_t X = filter_spatial_lengths[1];
+
+    const ck::index_t Hi = input_spatial_lengths[0];
+    const ck::index_t Wi = input_spatial_lengths[1];
+
+    const ck::index_t Ho = output_spatial_lengths[0];
+    const ck::index_t Wo = output_spatial_lengths[1];
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
+            using namespace ck::literals;
+
+            if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
+                         is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
+                         is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, H * W, W, 1_uz});
+            }
+            else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
+                              is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
+                              is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
+            }
+        };
+
+    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
+    Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
+    Tensor<OutDataType> out_n_k_ho_wo_host_result(
+        f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+    Tensor<OutDataType> out_n_k_ho_wo_device_result(
+        f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+
+    // bias: assume contiguous 1d vector
+    Tensor<OutDataType> bias_k({K});
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
+    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo_host_result.mDesc << std::endl;
+    std::cout << "bias_k: " << bias_k.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        bias_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        break;
+    default:
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        bias_k.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+    }
+
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(do_verification)
+    {
+        using ReferenceConvFwdInstance =
+            ck::tensor_operation::host::ReferenceConvFwd_Bias_Activation<InDataType,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         OutElementOp>;
+
+        auto ref_conv    = ReferenceConvFwdInstance{};
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
+                                                  wei_k_c_y_x,
+                                                  out_n_k_ho_wo_host_result,
+                                                  bias_k,
+                                                  conv_filter_strides,
+                                                  conv_filter_dilations,
+                                                  input_left_pads,
+                                                  input_right_pads,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+        ref_invoker.Run(ref_argument);
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_k_ho_wo_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(OutDataType) * bias_k.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+    wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
+    bias_device_buf.ToDevice(bias_k.mData.data());
+
+    using DeviceConvFwdBiasReluPtr = ck::tensor_operation::device::
+        DeviceConvFwdBiasActivationPtr<InElementOp, WeiElementOp, OutElementOp>;
+
+    // add device operator instances
+    std::vector<DeviceConvFwdBiasReluPtr> op_ptrs;
+
+    if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
+                 ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
+                 ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
+    {
+        ck::tensor_operation::device::instance::
+            add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+    }
+
+    if(op_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device Conv instance found");
+    }
+
+    std::string best_conv_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device Conv instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<const InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<const WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()),
+            N,
+            K,
+            C,
+            input_spatial_lengths,
+            filter_spatial_lengths,
+            output_spatial_lengths,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            in_element_op,
+            wei_element_op,
+            out_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string conv_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
+
+            std::size_t num_btype =
+                sizeof(InDataType) * (N * C * Hi * Wi) + sizeof(WeiDataType) * (K * C * Y * X) +
+                sizeof(OutDataType) * (N * K * Ho * Wo) + sizeof(OutDataType) * (K);
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << conv_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_conv_name  = conv_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
+
+                ck::utils::check_err(out_n_k_ho_wo_device_result, out_n_k_ho_wo_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "wei: ", wei_k_c_y_x.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "out_host  : ", out_n_k_ho_wo_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "out_device: ", out_n_k_ho_wo_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_conv_fwd_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_conv_fwd_impl.hpp
new file mode 100644
index 0000000..1f3ba8f
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_conv_fwd_impl.hpp
@@ -0,0 +1,221 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/convolution_forward.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+bool profile_conv_fwd_impl(int do_verification,
+                           int init_method,
+                           bool do_log,
+                           bool time_kernel,
+                           const ck::utils::conv::ConvParam& conv_param)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    Tensor<InDataType> input(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> host_output(out_g_n_k_wos_desc);
+    Tensor<OutDataType> device_output(out_g_n_k_wos_desc);
+
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "weight: " << weight.mDesc << std::endl;
+    std::cout << "output: " << host_output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        weight.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * device_output.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(input.mData.data());
+    wei_device_buf.ToDevice(weight.mData.data());
+
+    // run reference op
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>{};
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(input,
+                                                  weight,
+                                                  host_output,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        // init host output to zero
+        host_output.SetZero();
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceConvFwd<NDimSpatial,
+                                                                 InLayout,
+                                                                 WeiLayout,
+                                                                 OutLayout,
+                                                                 InDataType,
+                                                                 WeiDataType,
+                                                                 OutDataType,
+                                                                 InElementOp,
+                                                                 WeiElementOp,
+                                                                 OutElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    bool pass = true;
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                        conv_param.N_,
+                                        conv_param.K_,
+                                        conv_param.C_,
+                                        conv_param.input_spatial_lengths_,
+                                        conv_param.filter_spatial_lengths_,
+                                        conv_param.GetOutputSpatialLengths(),
+                                        conv_param.conv_filter_strides_,
+                                        conv_param.conv_filter_dilations_,
+                                        conv_param.input_left_pads_,
+                                        conv_param.input_right_pads_,
+                                        in_element_op,
+                                        wei_element_op,
+                                        out_element_op);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init output to zero before profiling next kernel
+            out_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+            float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                out_device_buf.FromDevice(device_output.mData.data());
+
+                pass = pass & ck::utils::check_err(device_output, host_output);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "input : ", input.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "weight: ", weight.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "host_output  : ", host_output.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "device_output: ", device_output.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_convnd_bwd_data_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_convnd_bwd_data_impl.hpp
new file mode 100644
index 0000000..1e69ebc
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_convnd_bwd_data_impl.hpp
@@ -0,0 +1,486 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/conv_util.hpp"
+#include "ck/library/host_tensor/device_memory.hpp"
+#include "ck/library/host_tensor/host_tensor.hpp"
+#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
+
+using F16  = ck::half_t;
+using F32  = float;
+using BF16 = ck::bhalf_t;
+using INT8 = int8_t;
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using DeviceConvBwdDataNoOpPtr =
+    DeviceConvBwdDataPtr<ck::tensor_operation::element_wise::PassThrough,
+                         ck::tensor_operation::element_wise::PassThrough,
+                         ck::tensor_operation::element_wise::PassThrough>;
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(
+    std::vector<DeviceConvBwdDataNoOpPtr>&);
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(
+    std::vector<DeviceConvBwdDataNoOpPtr>&);
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(
+    std::vector<DeviceConvBwdDataNoOpPtr>&);
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
+    std::vector<DeviceConvBwdDataNoOpPtr>&);
+
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<DeviceConvBwdDataNoOpPtr>&);
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<DeviceConvBwdDataNoOpPtr>&);
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
+    std::vector<DeviceConvBwdDataNoOpPtr>&);
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<DeviceConvBwdDataNoOpPtr>&);
+
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
+    std::vector<DeviceConvBwdDataNoOpPtr>&);
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
+    std::vector<DeviceConvBwdDataNoOpPtr>&);
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
+    std::vector<DeviceConvBwdDataNoOpPtr>&);
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
+    std::vector<DeviceConvBwdDataNoOpPtr>&);
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+using DeviceConvBwdDataNoOpPtr = ck::tensor_operation::device::instance::DeviceConvBwdDataNoOpPtr;
+
+template <typename InLayout>
+HostTensorDescriptor get_input_host_tensor_descriptor(const std::vector<std::size_t>& dims,
+                                                      int num_dim_spatial = 2)
+{
+    namespace tl = ck::tensor_layout::convolution;
+
+    switch(num_dim_spatial)
+    {
+    case 3: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
+    }
+    case 2: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
+    }
+    case 1: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
+    }
+    default: {
+        throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+    }
+    }
+}
+template <typename WeiLayout>
+HostTensorDescriptor get_filters_host_tensor_descriptor(const std::vector<std::size_t>& dims,
+                                                        int num_dim_spatial = 2)
+{
+    namespace tl = ck::tensor_layout::convolution;
+
+    switch(num_dim_spatial)
+    {
+    case 3: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
+    }
+    case 2: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
+    }
+    case 1: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
+    }
+    default: {
+        throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+    }
+    }
+}
+template <typename OutLayout>
+HostTensorDescriptor get_output_host_ensor_descriptor(const std::vector<std::size_t>& dims,
+                                                      int num_dim_spatial = 2)
+{
+    namespace tl = ck::tensor_layout::convolution;
+
+    switch(num_dim_spatial)
+    {
+    case 3: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
+    }
+    case 2: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
+    }
+    case 1: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
+    }
+    default: {
+        throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+    }
+    }
+}
+template <typename InDataType, typename WeiDataType, typename OutDataType>
+void get_device_conv_bwd_data_op_ptr(
+    InDataType, WeiDataType, OutDataType, std::vector<DeviceConvBwdDataNoOpPtr>&, int)
+{
+    std::cout << "can not find device conv bwd data" << std::endl;
+    exit(1);
+}
+template <>
+void get_device_conv_bwd_data_op_ptr(
+    F32, F32, F32, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
+{
+    switch(num_dim_spatial)
+    {
+    case 1:
+        ck::tensor_operation::device::instance::
+            add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(conv_ptrs);
+        break;
+    case 2:
+        ck::tensor_operation::device::instance::
+            add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
+        break;
+    case 3:
+        ck::tensor_operation::device::instance::
+            add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(conv_ptrs);
+        break;
+    default: break;
+    }
+}
+template <>
+void get_device_conv_bwd_data_op_ptr(
+    F16, F16, F16, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
+{
+    switch(num_dim_spatial)
+    {
+    case 1:
+        ck::tensor_operation::device::instance::
+            add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(conv_ptrs);
+        break;
+    case 2:
+        ck::tensor_operation::device::instance::
+            add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
+        break;
+    case 3:
+        ck::tensor_operation::device::instance::
+            add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(conv_ptrs);
+        break;
+    default: break;
+    }
+}
+template <>
+void get_device_conv_bwd_data_op_ptr(
+    BF16, BF16, BF16, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
+{
+    switch(num_dim_spatial)
+    {
+    case 1:
+        ck::tensor_operation::device::instance::
+            add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(conv_ptrs);
+        break;
+    case 2:
+        ck::tensor_operation::device::instance::
+            add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
+        break;
+    case 3:
+        ck::tensor_operation::device::instance::
+            add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(conv_ptrs);
+        break;
+    default: break;
+    }
+}
+template <>
+void get_device_conv_bwd_data_op_ptr(
+    INT8, INT8, INT8, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
+{
+    switch(num_dim_spatial)
+    {
+    case 1:
+        ck::tensor_operation::device::instance::
+            add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(conv_ptrs);
+        break;
+    case 2:
+        ck::tensor_operation::device::instance::
+            add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
+        break;
+    case 3:
+        ck::tensor_operation::device::instance::
+            add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(conv_ptrs);
+        break;
+    default: break;
+    }
+}
+
+template <typename T>
+static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
+{
+    float max_diff = 1e-6;
+
+    for(std::size_t i = 0; i < ref.mData.size(); ++i)
+    {
+        float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
+        if(max_diff < diff)
+        {
+            return false;
+        }
+    }
+    return true;
+}
+template <typename DataType>
+void show_data_nhwc_layout(Tensor<DataType>& nhwc)
+{
+    std::cout << "[";
+    for(int n = 0; n < ck::type_convert<int>(nhwc.mDesc.GetLengths()[0]); n++)
+    {
+        std::cout << "[";
+        for(int hi = 0; hi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[2]); hi++)
+        {
+            std::cout << "[";
+            for(int wi = 0; wi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[3]); wi++)
+            {
+                std::cout << "[";
+                for(int c = 0; c < ck::type_convert<int>(nhwc.mDesc.GetLengths()[1]); c++)
+                {
+                    std::cout << static_cast<float>(nhwc(n, c, hi, wi)) << "  ";
+                }
+                std::cout << "]";
+            }
+            std::cout << "]";
+        }
+        std::cout << "]";
+    }
+    std::cout << "]";
+}
+
+template <int NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout>
+bool profile_convnd_bwd_data_impl(int do_verification,
+                                  int init_method,
+                                  bool do_log,
+                                  bool time_kernel,
+                                  ck::index_t N,
+                                  ck::index_t K,
+                                  ck::index_t C,
+                                  const std::vector<ck::index_t>& input_spatial_lengths,
+                                  const std::vector<ck::index_t>& filter_spatial_lengths,
+                                  const std::vector<ck::index_t>& output_spatial_lengths,
+                                  const std::vector<ck::index_t>& conv_filter_strides,
+                                  const std::vector<ck::index_t>& conv_filter_dilations,
+                                  const std::vector<ck::index_t>& input_left_pads,
+                                  const std::vector<ck::index_t>& input_right_pads)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    std::vector<std::size_t> input_dims{static_cast<std::size_t>(N), static_cast<std::size_t>(C)};
+    input_dims.insert(
+        std::end(input_dims), std::begin(input_spatial_lengths), std::end(input_spatial_lengths));
+
+    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(K), static_cast<std::size_t>(C)};
+    filter_dims.insert(std::end(filter_dims),
+                       std::begin(filter_spatial_lengths),
+                       std::end(filter_spatial_lengths));
+
+    std::vector<std::size_t> output_dims{static_cast<std::size_t>(N), static_cast<std::size_t>(K)};
+    output_dims.insert(std::end(output_dims),
+                       std::begin(output_spatial_lengths),
+                       std::end(output_spatial_lengths));
+
+    Tensor<InDataType> input_host_result(
+        get_input_host_tensor_descriptor<InLayout>(input_dims, NDimSpatial));
+    Tensor<InDataType> input_device_result(
+        get_input_host_tensor_descriptor<InLayout>(input_dims, NDimSpatial));
+    Tensor<WeiDataType> weights(
+        get_filters_host_tensor_descriptor<WeiLayout>(filter_dims, NDimSpatial));
+    Tensor<OutDataType> output(
+        get_output_host_ensor_descriptor<OutLayout>(output_dims, NDimSpatial));
+
+    std::cout << "input: " << input_host_result.mDesc << std::endl;
+    std::cout << "weights: " << weights.mDesc << std::endl;
+    std::cout << "output: " << output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        output.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        weights.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        output.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
+        weights.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input_device_result.mDesc.GetElementSpace());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * weights.mDesc.GetElementSpace());
+    DeviceMem out_device_buf(sizeof(OutDataType) * output.mDesc.GetElementSpace());
+
+    out_device_buf.ToDevice(output.mData.data());
+    wei_device_buf.ToDevice(weights.mData.data());
+
+    // reset input to zero
+    in_device_buf.SetZero();
+
+    if(do_verification)
+    {
+        auto RunReference = [&](auto& ref_conv) {
+            auto ref_invoker = ref_conv.MakeInvoker();
+
+            auto ref_argument = ref_conv.MakeArgument(input_host_result,
+                                                      weights,
+                                                      output,
+                                                      conv_filter_strides,
+                                                      conv_filter_dilations,
+                                                      input_left_pads,
+                                                      input_right_pads,
+                                                      InElementOp{},
+                                                      WeiElementOp{},
+                                                      OutElementOp{});
+            ref_invoker.Run(ref_argument);
+        };
+
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<InDataType,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         AccDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         OutElementOp,
+                                                                         NDimSpatial>();
+        RunReference(ref_conv);
+    }
+
+    // add device Conv instances
+    std::vector<DeviceConvBwdDataNoOpPtr> conv_ptrs;
+    get_device_conv_bwd_data_op_ptr(
+        InDataType{}, WeiDataType{}, OutDataType{}, conv_ptrs, NDimSpatial);
+
+    if(conv_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device Conv instance found");
+    }
+
+    std::string best_conv_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device Conv instances
+    bool success = true;
+    for(auto& conv_ptr : conv_ptrs)
+    {
+        auto argument_ptr = conv_ptr->MakeArgumentPointer(
+            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            N,
+            K,
+            C,
+            input_spatial_lengths,
+            filter_spatial_lengths,
+            output_spatial_lengths,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            in_element_op,
+            wei_element_op,
+            out_element_op);
+
+        auto invoker_ptr = conv_ptr->MakeInvokerPointer();
+
+        if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string conv_name = conv_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop =
+                ck::utils::conv::get_flops(N, C, K, filter_spatial_lengths, output_spatial_lengths);
+            std::size_t num_btype =
+                ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>(
+                    N, C, K, input_spatial_lengths, filter_spatial_lengths, output_spatial_lengths);
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s" << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_conv_name  = conv_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                in_device_buf.FromDevice(input_device_result.mData.data());
+
+                if(!check_out(input_host_result, input_device_result))
+                {
+                    std::cout << "Fail Info: " << conv_ptr->GetTypeString() << std::endl;
+
+                    success = false;
+                }
+                else
+                {
+                    std::cout << "Pass Info: " << conv_ptr->GetTypeString() << std::endl;
+                }
+
+                success = ck::utils::check_err(input_host_result, input_device_result);
+
+                if(do_log)
+                {
+                    std::cout << "in : ";
+                    show_data_nhwc_layout(output);
+                    std::cout << std::endl;
+
+                    std::cout << "wei: ";
+                    show_data_nhwc_layout(weights);
+                    std::cout << std::endl;
+
+                    std::cout << "out_host  : ";
+                    show_data_nhwc_layout(input_host_result);
+                    std::cout << std::endl;
+
+                    std::cout << "out_device: ";
+                    show_data_nhwc_layout(input_device_result);
+                    std::cout << std::endl;
+                }
+            }
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
+    return success;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_convnd_bwd_weight_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_convnd_bwd_weight_impl.hpp
new file mode 100644
index 0000000..e37c887
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_convnd_bwd_weight_impl.hpp
@@ -0,0 +1,474 @@
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_backward_weight.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/conv_util.hpp"
+#include "ck/library/host_tensor/device_memory.hpp"
+#include "ck/library/host_tensor/host_tensor.hpp"
+#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp"
+
+using F16  = ck::half_t;
+using F32  = float;
+using BF16 = ck::bhalf_t;
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using DeviceConvndBwdWeightNoOpPtr =
+    DeviceConvBwdWeightPtr<ck::tensor_operation::element_wise::PassThrough,
+                           ck::tensor_operation::element_wise::PassThrough,
+                           ck::tensor_operation::element_wise::PassThrough>;
+
+void add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_f32_instances(
+    std::vector<DeviceConvndBwdWeightNoOpPtr>&);
+void add_device_convnd_bwd_weight_xdl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<DeviceConvndBwdWeightNoOpPtr>&);
+void add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
+    std::vector<DeviceConvndBwdWeightNoOpPtr>&);
+
+void add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_f16_instances(
+    std::vector<DeviceConvndBwdWeightNoOpPtr>&);
+void add_device_convnd_bwd_weight_xdl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<DeviceConvndBwdWeightNoOpPtr>&);
+void add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
+    std::vector<DeviceConvndBwdWeightNoOpPtr>&);
+
+void add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_bf16_instances(
+    std::vector<DeviceConvndBwdWeightNoOpPtr>&);
+void add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_bf16_instances(
+    std::vector<DeviceConvndBwdWeightNoOpPtr>&);
+void add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
+    std::vector<DeviceConvndBwdWeightNoOpPtr>&);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+using DeviceConvndBwdWeightNoOpPtr =
+    ck::tensor_operation::device::instance::DeviceConvndBwdWeightNoOpPtr;
+
+template <typename InLayout>
+HostTensorDescriptor get_input_host_tensor_descriptor(const std::vector<std::size_t>& dims,
+                                                      int num_dim_spatial = 2)
+{
+    namespace tl = ck::tensor_layout::convolution;
+
+    switch(num_dim_spatial)
+    {
+    case 3: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
+    }
+    case 2: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
+    }
+    case 1: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
+    }
+    default: {
+        throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+    }
+    }
+}
+
+template <typename WeiLayout>
+HostTensorDescriptor get_filters_host_tensor_descriptor(const std::vector<std::size_t>& dims,
+                                                        int num_dim_spatial = 2)
+{
+    namespace tl = ck::tensor_layout::convolution;
+
+    switch(num_dim_spatial)
+    {
+    case 3: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
+    }
+    case 2: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
+    }
+    case 1: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
+    }
+    default: {
+        throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+    }
+    }
+}
+
+template <typename OutLayout>
+HostTensorDescriptor get_output_host_ensor_descriptor(const std::vector<std::size_t>& dims,
+                                                      int num_dim_spatial = 2)
+{
+    namespace tl = ck::tensor_layout::convolution;
+
+    switch(num_dim_spatial)
+    {
+    case 3: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
+    }
+    case 2: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
+    }
+    case 1: {
+        return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
+    }
+    default: {
+        throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+    }
+    }
+}
+
+template <typename InDataType, typename WeiDataType, typename OutDataType>
+void get_device_conv_bwd_weight_op_ptr(
+    InDataType, WeiDataType, OutDataType, std::vector<DeviceConvndBwdWeightNoOpPtr>&, int)
+{
+    std::cout << "can not find device conv bwd weight" << std::endl;
+    exit(1);
+}
+
+template <>
+void get_device_conv_bwd_weight_op_ptr(
+    F32, F32, F32, std::vector<DeviceConvndBwdWeightNoOpPtr>& conv_ptrs, int num_dim_spatial)
+{
+    switch(num_dim_spatial)
+    {
+    case 1:
+        ck::tensor_operation::device::instance::
+            add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_f32_instances(conv_ptrs);
+        break;
+    case 2:
+        ck::tensor_operation::device::instance::
+            add_device_convnd_bwd_weight_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
+        break;
+    case 3:
+        ck::tensor_operation::device::instance::
+            add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_f32_instances(conv_ptrs);
+        break;
+    default: break;
+    }
+}
+
+template <>
+void get_device_conv_bwd_weight_op_ptr(
+    F16, F16, F16, std::vector<DeviceConvndBwdWeightNoOpPtr>& conv_ptrs, int num_dim_spatial)
+{
+    switch(num_dim_spatial)
+    {
+    case 1:
+        ck::tensor_operation::device::instance::
+            add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_f16_instances(conv_ptrs);
+        break;
+    case 2:
+        ck::tensor_operation::device::instance::
+            add_device_convnd_bwd_weight_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
+        break;
+    case 3:
+        ck::tensor_operation::device::instance::
+            add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_f16_instances(conv_ptrs);
+        break;
+    default: break;
+    }
+}
+
+template <>
+void get_device_conv_bwd_weight_op_ptr(
+    BF16, BF16, BF16, std::vector<DeviceConvndBwdWeightNoOpPtr>& conv_ptrs, int num_dim_spatial)
+{
+    switch(num_dim_spatial)
+    {
+    case 1:
+        ck::tensor_operation::device::instance::
+            add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_bf16_instances(conv_ptrs);
+        break;
+    case 2:
+        ck::tensor_operation::device::instance::
+            add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
+        break;
+    case 3:
+        ck::tensor_operation::device::instance::
+            add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(conv_ptrs);
+        break;
+    default: break;
+    }
+}
+
+template <typename DataType>
+void show_data_nhwc_layout(Tensor<DataType>& nhwc)
+{
+    std::cout << "[";
+    for(int n = 0; n < ck::type_convert<int>(nhwc.mDesc.GetLengths()[0]); n++)
+    {
+        std::cout << "[";
+        for(int hi = 0; hi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[2]); hi++)
+        {
+            std::cout << "[";
+            for(int wi = 0; wi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[3]); wi++)
+            {
+                std::cout << "[";
+                for(int c = 0; c < ck::type_convert<int>(nhwc.mDesc.GetLengths()[1]); c++)
+                {
+                    std::cout << static_cast<float>(nhwc(n, c, hi, wi)) << "  ";
+                }
+                std::cout << "]";
+            }
+            std::cout << "]";
+        }
+        std::cout << "]";
+    }
+    std::cout << "]";
+}
+
+template <int NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout>
+bool profile_convnd_bwd_weight_impl(int do_verification,
+                                    int init_method,
+                                    bool do_log,
+                                    bool time_kernel,
+                                    ck::index_t N,
+                                    ck::index_t K,
+                                    ck::index_t C,
+                                    std::vector<ck::index_t> input_spatial_lengths,
+                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                    std::vector<ck::index_t> output_spatial_lengths,
+                                    std::vector<ck::index_t> conv_filter_strides,
+                                    std::vector<ck::index_t> conv_filter_dilations,
+                                    std::vector<ck::index_t> input_left_pads,
+                                    std::vector<ck::index_t> input_right_pads,
+                                    ck::index_t split_k)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    std::vector<std::size_t> input_dims{static_cast<std::size_t>(N), static_cast<std::size_t>(C)};
+    input_dims.insert(
+        std::end(input_dims), std::begin(input_spatial_lengths), std::end(input_spatial_lengths));
+
+    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(K), static_cast<std::size_t>(C)};
+    filter_dims.insert(std::end(filter_dims),
+                       std::begin(filter_spatial_lengths),
+                       std::end(filter_spatial_lengths));
+
+    std::vector<std::size_t> output_dims{static_cast<std::size_t>(N), static_cast<std::size_t>(K)};
+    output_dims.insert(std::end(output_dims),
+                       std::begin(output_spatial_lengths),
+                       std::end(output_spatial_lengths));
+
+    Tensor<InDataType> input(get_input_host_tensor_descriptor<InLayout>(input_dims, NDimSpatial));
+    Tensor<WeiDataType> weights_host_result(
+        get_filters_host_tensor_descriptor<WeiLayout>(filter_dims, NDimSpatial));
+    Tensor<WeiDataType> weights_device_result(
+        get_filters_host_tensor_descriptor<WeiLayout>(filter_dims, NDimSpatial));
+    Tensor<OutDataType> output(
+        get_output_host_ensor_descriptor<OutLayout>(output_dims, NDimSpatial));
+
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "weights: " << weights_host_result.mDesc << std::endl;
+    std::cout << "output: " << output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        input.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
+        output.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 2});
+        break;
+    default:
+        input.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
+        output.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpace());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * weights_device_result.mDesc.GetElementSpace());
+    DeviceMem out_device_buf(sizeof(OutDataType) * output.mDesc.GetElementSpace());
+
+    in_device_buf.ToDevice(input.mData.data());
+    out_device_buf.ToDevice(output.mData.data());
+
+    // reset input to zero
+    wei_device_buf.SetZero();
+
+    if(do_verification)
+    {
+        auto RunReference = [&](auto& ref_conv) {
+            auto ref_invoker = ref_conv.MakeInvoker();
+
+            auto ref_argument = ref_conv.MakeArgument(input,
+                                                      weights_host_result,
+                                                      output,
+                                                      conv_filter_strides,
+                                                      conv_filter_dilations,
+                                                      input_left_pads,
+                                                      input_right_pads,
+                                                      InElementOp{},
+                                                      WeiElementOp{},
+                                                      OutElementOp{});
+            ref_invoker.Run(ref_argument);
+        };
+
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdWeight<InDataType,
+                                                                           WeiDataType,
+                                                                           OutDataType,
+                                                                           InElementOp,
+                                                                           WeiElementOp,
+                                                                           OutElementOp,
+                                                                           NDimSpatial>();
+        RunReference(ref_conv);
+    }
+
+    // add device Conv instances
+    std::vector<DeviceConvndBwdWeightNoOpPtr> conv_ptrs;
+    get_device_conv_bwd_weight_op_ptr(
+        InDataType{}, WeiDataType{}, OutDataType{}, conv_ptrs, NDimSpatial);
+
+    if(conv_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device Conv instance found");
+    }
+
+    std::string best_conv_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device Conv instances
+    bool success = true;
+    for(auto& conv_ptr : conv_ptrs)
+    {
+        // using atomic, so need to reset input, setzero is done in invoker
+        // if(split_k > 1)
+        //{
+        //    wei_device_buf.SetZero();
+        //}
+
+        auto argument_ptr = conv_ptr->MakeArgumentPointer(
+            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            N,
+            K,
+            C,
+            input_spatial_lengths,
+            filter_spatial_lengths,
+            output_spatial_lengths,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            in_element_op,
+            wei_element_op,
+            out_element_op,
+            split_k);
+
+        if(!conv_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::cout << "wrong! device_conv with the specified compilation parameters does "
+                         "not support this Conv problem"
+                      << std::endl;
+            continue;
+        }
+
+        auto invoker_ptr      = conv_ptr->MakeInvokerPointer();
+        std::string conv_name = conv_ptr->GetTypeString();
+        float ave_time        = 0;
+
+        if(std::is_same<InDataType, ck::bhalf_t>::value && split_k > 1)
+        {
+            // alloc work space
+            size_t bwd_weight_workspace_size = conv_ptr->GetWorkSpaceSize(argument_ptr.get());
+            if(bwd_weight_workspace_size <= 0)
+            {
+                printf("wrong work space size\n");
+                exit(1);
+            }
+            DeviceMem wei_work_space_device_buf(bwd_weight_workspace_size);
+            wei_work_space_device_buf.SetZero();
+            conv_ptr->SetWorkSpacePointer(argument_ptr.get(),
+                                          wei_work_space_device_buf.GetDeviceBuffer());
+            ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+        }
+        else
+        {
+            ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+        }
+
+        std::size_t flop =
+            ck::utils::conv::get_flops(N, C, K, filter_spatial_lengths, output_spatial_lengths);
+        std::size_t num_btype = ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>(
+            N, C, K, input_spatial_lengths, filter_spatial_lengths, output_spatial_lengths);
+
+        float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s" << std::endl;
+
+        if(tflops > best_tflops)
+        {
+            best_conv_name  = conv_name;
+            best_tflops     = tflops;
+            best_ave_time   = ave_time;
+            best_gb_per_sec = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            wei_device_buf.FromDevice(weights_device_result.mData.data());
+
+            success = ck::utils::check_err(weights_host_result, weights_device_result);
+
+            if(success == false)
+            {
+                std::cout << "Fail Info: " << conv_ptr->GetTypeString() << std::endl;
+            }
+            else
+            {
+                std::cout << "Pass Info: " << conv_ptr->GetTypeString() << std::endl;
+            }
+
+            if(do_log)
+            {
+                std::cout << "in : ";
+                show_data_nhwc_layout(output);
+                std::cout << std::endl;
+
+                std::cout << "wei: ";
+                show_data_nhwc_layout(weights_host_result);
+                std::cout << std::endl;
+
+                std::cout << "out  : ";
+                show_data_nhwc_layout(input);
+                std::cout << std::endl;
+
+                std::cout << "wei_device: ";
+                show_data_nhwc_layout(weights_device_result);
+                std::cout << std::endl;
+            }
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
+    return success;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
new file mode 100644
index 0000000..7707e16
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
@@ -0,0 +1,266 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
+void host_elementwise2D(HostTensorC& C,
+                        const HostTensorA& A,
+                        const HostTensorB& B,
+                        const std::vector<std::size_t>& shape,
+                        Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(std::size_t m = 0; m < shape[0]; ++m)
+        for(std::size_t n = 0; n < shape[1]; ++n)
+        {
+            auto a_val  = A(m, n);
+            auto b_val  = B(m, n);
+            ctype c_val = 0;
+            functor(c_val, a_val, b_val);
+            C(m, n) = c_val;
+        }
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType>
+bool profile_elementwise_layernorm_impl(int do_verification,
+                                        int init_method,
+                                        bool do_log,
+                                        bool time_kernel,
+                                        std::vector<index_t> length)
+{
+    using Add         = ck::tensor_operation::element_wise::Add;
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    if(length.size() != 2)
+        return false;
+
+    index_t M      = length[0];
+    index_t N      = length[1];
+    index_t Stride = N;
+
+    constexpr int Rank         = 2;
+    constexpr int NumReduceDim = 1;
+
+    std::vector<index_t> reduce_dim      = {1};
+    std::vector<index_t> gammaBetaLength = {N};
+    std::vector<index_t> gammaBetaStride = {0, 1};
+
+    auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
+        using namespace ck::literals;
+
+        return HostTensorDescriptor({row, col}, {stride, 1_uz});
+    };
+
+    Tensor<ADataType> a(length);
+    Tensor<BDataType> b(length);
+    Tensor<GammaDataType> gamma(gammaBetaLength);
+    Tensor<BetaDataType> beta(gammaBetaLength);
+    Tensor<YDataType> y(length);
+    Tensor<YDataType> host_y(length);
+
+    switch(init_method)
+    {
+    case 0:
+        a.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
+        b.GenerateTensorValue(GeneratorTensor_1<BDataType>{});
+        gamma.GenerateTensorValue(GeneratorTensor_1<GammaDataType>{});
+        beta.GenerateTensorValue(GeneratorTensor_1<BetaDataType>{});
+        break;
+    case 1:
+        a.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
+        beta.GenerateTensorValue(GeneratorTensor_2<BetaDataType>{-5, 5});
+        break;
+    default:
+        a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0, 1});
+        b.GenerateTensorValue(GeneratorTensor_3<BDataType>{0, 1});
+        gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-0.5, 0.5});
+        beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a_dev(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_dev(sizeof(ADataType) * b.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    a_dev.ToDevice(a.mData.data());
+    b_dev.ToDevice(b.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    std::array<const void*, 2> input = {a_dev.GetDeviceBuffer(), b_dev.GetDeviceBuffer()};
+
+    // add device normalization instances
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwiseNormalization<
+        ck::Tuple<ADataType, BDataType>,
+        GammaDataType,
+        BetaDataType,
+        AccDataType,
+        YDataType,
+        Add,
+        PassThrough,
+        2,
+        1>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using XDataType             = ADataType;
+        std::vector<std::size_t> mn = {static_cast<unsigned long>(M),
+                                       static_cast<unsigned long>(N)};
+        Tensor<XDataType> x(f_host_tensor_descriptor2d(M, N, Stride));
+        host_elementwise2D<Tensor<ADataType>, Tensor<BDataType>, Tensor<XDataType>, Add>(
+            x, a, b, mn, Add{});
+
+        using ReferenceInstance = ck::tensor_operation::host::ReferenceLayernorm<XDataType,
+                                                                                 GammaDataType,
+                                                                                 BetaDataType,
+                                                                                 YDataType,
+                                                                                 AccDataType,
+                                                                                 PassThrough,
+                                                                                 Rank,
+                                                                                 NumReduceDim>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, {M, N}, {1}, 1e-4);
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            length,
+            {
+                std::vector<ck::index_t>{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()},
+                std::vector<ck::index_t>{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()},
+            },
+            gammaBetaStride,
+            gammaBetaStride,
+            std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+            reduce_dim,
+            1e-4,
+            input,
+            gamma_dev.GetDeviceBuffer(),
+            beta_dev.GetDeviceBuffer(),
+            y_dev.GetDeviceBuffer(),
+            Add{},
+            PassThrough{});
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            continue;
+        }
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes = a.mDesc.GetElementSize() * sizeof(ADataType) +
+                                b.mDesc.GetElementSize() * sizeof(BDataType) +
+                                gamma.mDesc.GetElementSize() * sizeof(GammaDataType) +
+                                beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
+                                y.mDesc.GetElementSize() * sizeof(YDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            y_dev.FromDevice(y.mData.data());
+
+            bool pass =
+                ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(std::cout << "a  : ", a.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "b  : ", b.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "host_y  : ", host_y.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "y  : ", y.mData, ",") << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", length, ",") << ", ";
+        std::cout << "num_kernel = " << num_kernel << ", best perf = " << best_avg_time << " ms, "
+                  << best_gb_per_sec << " GB/s, " << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is tested" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp
new file mode 100644
index 0000000..3cc2ea3
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp
@@ -0,0 +1,242 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout>
+bool profile_gemm_add_add_fastgelu_impl(int do_verification,
+                                        int init_method,
+                                        bool /*do_log*/,
+                                        bool time_kernel,
+                                        int M,
+                                        int N,
+                                        int K,
+                                        int StrideA,
+                                        int StrideB,
+                                        int StrideD0,
+                                        int StrideD1,
+                                        int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+    using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = AddAddFastGelu;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddAddFastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
new file mode 100644
index 0000000..d53a658
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
@@ -0,0 +1,232 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename ELayout>
+bool profile_gemm_add_fastgelu_impl(int do_verification,
+                                    int init_method,
+                                    bool /*do_log*/,
+                                    bool time_kernel,
+                                    int M,
+                                    int N,
+                                    int K,
+                                    int StrideA,
+                                    int StrideB,
+                                    int StrideD0,
+                                    int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = AddFastGelu;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddFastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp
new file mode 100644
index 0000000..b4ec78c
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp
@@ -0,0 +1,384 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32                 = float;
+using F16                 = ck::half_t;
+using ReducePtrsGlobal    = ck::Tuple<F32*, F32*>;
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using DeviceGemmBiasAddReduceNoOpPtr =
+    ck::tensor_operation::device::DeviceGemmReducePtr<1, ReducePtrsGlobal::Size()>;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+    std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+    std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+    std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+    std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename BiasDataType,
+          typename D0DataType,
+          typename ReduceDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+void profile_gemm_bias_add_reduce_impl(int do_verification,
+                                       int init_method,
+                                       bool do_log,
+                                       bool time_kernel,
+                                       int M,
+                                       int N,
+                                       int K,
+                                       int StrideA,
+                                       int StrideB,
+                                       int StrideC,
+                                       int StrideD0)
+{
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<BiasDataType> bias_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> reduce0_m_host_result({M});
+    Tensor<ReduceDataType> reduce1_m_host_result({M});
+
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> reduce0_m_device_result({M});
+    Tensor<ReduceDataType> reduce1_m_device_result({M});
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+    std::cout << "reduce0_m: " << reduce0_m_host_result.mDesc << std::endl;
+    std::cout << "reduce1_m: " << reduce1_m_host_result.mDesc << std::endl;
+
+    std::size_t num_thread = 1;
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        std::srand(0);
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        bias_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        break;
+    default:
+        std::srand(0);
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+        bias_n.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5}, num_thread);
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+    }
+
+    using PassThrough           = ck::tensor_operation::element_wise::PassThrough;
+    using AElementOp            = PassThrough;
+    using BElementOp            = PassThrough;
+    using CElementOp            = PassThrough;
+    using D0ElementOp           = PassThrough;
+    using ReduceOp0             = ck::reduce::Add;
+    using ReduceOp1             = ck::reduce::Add;
+    using UnaryDivElementOp     = ck::tensor_operation::element_wise::UnaryDivide;
+    using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
+
+    auto a_element_op                     = AElementOp{};
+    auto b_element_op                     = BElementOp{};
+    auto c_element_op                     = CElementOp{};
+    std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
+
+    auto d0_element_op    = D0ElementOp{};
+    const auto reduce0_op = ReduceOp0{};
+    const auto reduce1_op = ReduceOp1{};
+
+    auto passthrough                            = UnaryIdenticElementOp{};
+    auto square                                 = UnarySquareElementOp{};
+    auto div                                    = UnaryDivElementOp{N};
+    std::array<void*, 2> reduce_in_element_ops  = {&passthrough, &square};
+    std::array<void*, 2> reduce_out_element_ops = {&div, &div};
+
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                ReduceDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        using ReduceAccDataType = ReduceDataType;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType acc = static_cast<ReduceAccDataType>(c_m_n_host_result(m, n)) +
+                                        static_cast<ReduceAccDataType>(bias_n(n));
+
+                ReduceAccDataType d0 = static_cast<ReduceAccDataType>(d0_m_n(m, n));
+                c_element_op(acc, acc);
+                d0_element_op(d0, d0);
+                acc += d0;
+                c_m_n_host_result(m, n) = static_cast<CDataType>(acc);
+            }
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+            auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType d0_val =
+                    ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
+                ReduceAccDataType d1_val;
+
+                square(d1_val, d0_val);
+                reduce0_op(reduce0_acc, d0_val);
+                reduce1_op(reduce1_acc, d1_val);
+            }
+
+            div(reduce0_acc, reduce0_acc);
+            div(reduce1_acc, reduce1_acc);
+            reduce0_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce0_acc);
+            reduce1_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce1_acc);
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
+                                 reduce0_m_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
+                                 reduce1_m_device_result.mDesc.GetElementSpaceSize());
+
+    std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
+                                      reduce1_device_buf.GetDeviceBuffer()};
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    bias_device_buf.ToDevice(bias_n.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+
+    // add device GEMM instances
+    std::vector<ck::tensor_operation::device::instance::DeviceGemmBiasAddReduceNoOpPtr> gemm_ptrs;
+
+    if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
+                 is_same<CDataType, half_t>::value)
+    {
+        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+                    gemm_ptrs);
+        }
+    }
+
+    if(gemm_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device GEMM instance found");
+    }
+
+    std::string best_gemm_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device GEMM instances
+    for(auto& gemm_ptr : gemm_ptrs)
+    {
+        auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                          b_device_buf.GetDeviceBuffer(),
+                                                          bias_device_buf.GetDeviceBuffer(),
+                                                          {d0_device_buf.GetDeviceBuffer()},
+                                                          c_device_buf.GetDeviceBuffer(),
+                                                          p_reduces,
+                                                          M,
+                                                          N,
+                                                          K,
+                                                          StrideA,
+                                                          StrideB,
+                                                          StrideC,
+                                                          {StrideD0},
+                                                          gemm_element_ops,
+                                                          {&d0_element_op},
+                                                          reduce_in_element_ops,
+                                                          reduce_out_element_ops);
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // init DO, D1 to 0
+            reduce0_device_buf.SetZero();
+            reduce1_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::string gemm_name = gemm_ptr->GetTypeString();
+
+            std::size_t flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
+
+            std::size_t num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                   sizeof(CDataType) * M * N + sizeof(BiasDataType) * M * N +
+                                   sizeof(D0DataType) * M * N + sizeof(ReduceDataType) * M +
+                                   sizeof(ReduceDataType) * M;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << gemm_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_gemm_name  = gemm_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+                reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
+                reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
+
+                ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+                ck::utils::check_err(reduce0_m_device_result, reduce0_m_host_result);
+                ck::utils::check_err(reduce1_m_device_result, reduce1_m_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host: ", c_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d0_host: ", reduce0_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d0_device: ", reduce0_m_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d1_host: ", reduce1_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d1_device: ", reduce1_m_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << "does not support this GEMM problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_bilinear_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_bilinear_impl.hpp
new file mode 100644
index 0000000..31bae28
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_bilinear_impl.hpp
@@ -0,0 +1,234 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename DDataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename DLayout,
+          typename ELayout>
+bool profile_gemm_bilinear_impl(int do_verification,
+                                int init_method,
+                                bool /*do_log*/,
+                                bool time_kernel,
+                                int M,
+                                int N,
+                                int K,
+                                int StrideA,
+                                int StrideB,
+                                int StrideD,
+                                int StrideE,
+                                float alpha,
+                                float beta)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = Bilinear;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{alpha, beta};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<DLayout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<DDataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Bilinear>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_m_n_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d_m_n_device_buf.ToDevice(d_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_fastgelu_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_fastgelu_impl.hpp
new file mode 100644
index 0000000..f9a544c
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_fastgelu_impl.hpp
@@ -0,0 +1,222 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout>
+bool profile_gemm_fastgelu_impl(int do_verification,
+                                int init_method,
+                                bool /*do_log*/,
+                                bool time_kernel,
+                                int M,
+                                int N,
+                                int K,
+                                int StrideA,
+                                int StrideB,
+                                int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = FastGelu;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::FastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        std::array<const void*, 0>{},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        std::array<ck::index_t, 0>{},
+                                                        StrideE,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_impl.hpp
new file mode 100644
index 0000000..9b16410
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_impl.hpp
@@ -0,0 +1,254 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType>
+int profile_gemm_impl(int do_verification,
+                      int init_method,
+                      bool do_log,
+                      bool time_kernel,
+                      int M,
+                      int N,
+                      int K,
+                      int StrideA,
+                      int StrideB,
+                      int StrideC)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemm<ALayout,
+                                                              BLayout,
+                                                              CLayout,
+                                                              ADataType,
+                                                              BDataType,
+                                                              CDataType,
+                                                              AElementOp,
+                                                              BElementOp,
+                                                              CElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // Run reference op
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        auto ref_op      = ReferenceGemmInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        auto ref_argument = ref_op.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                        static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                        M,
+                                        N,
+                                        K,
+                                        StrideA,
+                                        StrideB,
+                                        StrideC,
+                                        a_element_op,
+                                        b_element_op,
+                                        c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init C to zero before profiling next kernel
+            c_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    if constexpr(is_same<CDataType, float>::value)
+    {
+        std::cout << "Best Perf for datatype = f32";
+    }
+    else if constexpr(is_same<CDataType, half_t>::value)
+    {
+        std::cout << "Best Perf for datatype = f16";
+    }
+    else if constexpr(is_same<CDataType, bhalf_t>::value)
+    {
+        std::cout << "Best Perf for datatype = bf16";
+    }
+    else if constexpr(is_same<CDataType, int8_t>::value)
+    {
+        std::cout << "Best Perf for datatype = int8";
+    }
+
+    if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " ALayout =  RowMajor";
+    }
+    else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " ALayout =  ColumnMajor";
+    }
+
+    if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " BLayout =  RowMajor";
+    }
+    else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " BLayout =  ColumnMajor";
+    }
+
+    std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
+              << " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_avg_time
+              << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    return pass ? 0 : 1;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_reduce_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_reduce_impl.hpp
new file mode 100644
index 0000000..370121a
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_reduce_impl.hpp
@@ -0,0 +1,353 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32                 = float;
+using F16                 = ck::half_t;
+using ReducePtrsGlobal    = ck::Tuple<F32*, F32*>;
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using DeviceGemmReduceNoOpPtr =
+    ck::tensor_operation::device::DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>;
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ReduceDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_gemm_reduce_impl(int do_verification,
+                              int init_method,
+                              bool do_log,
+                              bool time_kernel,
+                              int M,
+                              int N,
+                              int K,
+                              int StrideA,
+                              int StrideB,
+                              int StrideC)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> reduce0_m_host_result({M});
+    Tensor<ReduceDataType> reduce1_m_host_result({M});
+
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> reduce0_m_device_result({M});
+    Tensor<ReduceDataType> reduce1_m_device_result({M});
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+    std::cout << "reduce0_m: " << reduce0_m_host_result.mDesc << std::endl;
+    std::cout << "reduce1_m: " << reduce1_m_host_result.mDesc << std::endl;
+
+    std::size_t num_thread = 1;
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        std::srand(0);
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        break;
+    default:
+        std::srand(0);
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+    }
+
+    using AElementOp            = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp            = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp            = ck::tensor_operation::element_wise::PassThrough;
+    using ReduceOp0             = ck::reduce::Add;
+    using ReduceOp1             = ck::reduce::Add;
+    using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
+    using UnaryDivElementOp     = ck::tensor_operation::element_wise::UnaryDivide;
+
+    auto a_element_op                     = AElementOp{};
+    auto b_element_op                     = BElementOp{};
+    auto c_element_op                     = CElementOp{};
+    std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
+
+    const auto reduce0_op = ReduceOp0{};
+    const auto reduce1_op = ReduceOp1{};
+
+    auto passthrough                            = UnaryIdenticElementOp{};
+    auto square                                 = UnarySquareElementOp{};
+    auto div                                    = UnaryDivElementOp{N};
+    std::array<void*, 2> reduce_in_element_ops  = {&passthrough, &square};
+    std::array<void*, 2> reduce_out_element_ops = {&div, &div};
+
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                ReduceDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        using ReduceAccDataType = ReduceDataType;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+            auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType d0_val =
+                    ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
+                ReduceAccDataType d1_val;
+
+                square(d1_val, d0_val);
+                reduce0_op(reduce0_acc, d0_val);
+                reduce1_op(reduce1_acc, d1_val);
+            }
+
+            div(reduce0_acc, reduce0_acc);
+            div(reduce1_acc, reduce1_acc);
+            reduce0_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce0_acc);
+            reduce1_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce1_acc);
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
+                                 reduce0_m_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
+                                 reduce1_m_device_result.mDesc.GetElementSpaceSize());
+
+    std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
+                                      reduce1_device_buf.GetDeviceBuffer()};
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    // add device GEMM instances
+    std::vector<ck::tensor_operation::device::instance::DeviceGemmReduceNoOpPtr> gemm_ptrs;
+
+    if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
+                 is_same<CDataType, half_t>::value)
+    {
+        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances(
+                    gemm_ptrs);
+        }
+    }
+
+    if(gemm_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device GEMM instance found");
+    }
+
+    std::string best_gemm_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device GEMM instances
+    for(auto& gemm_ptr : gemm_ptrs)
+    {
+        auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                          b_device_buf.GetDeviceBuffer(),
+                                                          nullptr,
+                                                          {},
+                                                          c_device_buf.GetDeviceBuffer(),
+                                                          p_reduces,
+                                                          M,
+                                                          N,
+                                                          K,
+                                                          StrideA,
+                                                          StrideB,
+                                                          StrideC,
+                                                          {},
+                                                          gemm_element_ops,
+                                                          {},
+                                                          reduce_in_element_ops,
+                                                          reduce_out_element_ops);
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // init DO, D1 to 0
+            reduce0_device_buf.SetZero();
+            reduce1_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::string gemm_name = gemm_ptr->GetTypeString();
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                    sizeof(CDataType) * M * N + sizeof(CDataType) * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << gemm_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_gemm_name  = gemm_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+                reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
+                reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
+
+                ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+                ck::utils::check_err(reduce0_m_device_result, reduce0_m_host_result);
+                ck::utils::check_err(reduce1_m_device_result, reduce1_m_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host: ", c_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d0_host: ", reduce0_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d0_device: ", reduce0_m_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d1_host: ", reduce1_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d1_device: ", reduce1_m_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << "does not support this GEMM problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_splitk_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_splitk_impl.hpp
new file mode 100644
index 0000000..e5d5f87
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_gemm_splitk_impl.hpp
@@ -0,0 +1,257 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_gemm_splitk_impl(int do_verification,
+                              int init_method,
+                              bool do_log,
+                              bool time_kernel,
+                              int M,
+                              int N,
+                              int K,
+                              int StrideA,
+                              int StrideB,
+                              int StrideC,
+                              int KBatch)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    c_device_buf.ToDevice(c_m_n_device_result.mData.data());
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmSplitK<ALayout,
+                                                                    BLayout,
+                                                                    CLayout,
+                                                                    ADataType,
+                                                                    BDataType,
+                                                                    CDataType,
+                                                                    AElementOp,
+                                                                    BElementOp,
+                                                                    CElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // Run reference GEMM
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device GEMM instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                        static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                        M,
+                                        N,
+                                        K,
+                                        StrideA,
+                                        StrideB,
+                                        StrideC,
+                                        a_element_op,
+                                        b_element_op,
+                                        c_element_op,
+                                        KBatch);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init C to zero before profiling next kernel
+            c_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    if constexpr(is_same<CDataType, float>::value)
+    {
+        std::cout << "Best Perf for datatype = f32";
+    }
+    else if constexpr(is_same<CDataType, half_t>::value)
+    {
+        std::cout << "Best Perf for datatype = f16";
+    }
+    else if constexpr(is_same<CDataType, bhalf_t>::value)
+    {
+        std::cout << "Best Perf for datatype = bf16";
+    }
+    else if constexpr(is_same<CDataType, int8_t>::value)
+    {
+        std::cout << "Best Perf for datatype = int8";
+    }
+
+    if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " ALayout =  RowMajor";
+    }
+    else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " ALayout =  ColumnMajor";
+    }
+
+    if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " BLayout =  RowMajor";
+    }
+    else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " BLayout =  ColumnMajor";
+    }
+
+    std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
+              << " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_ave_time
+              << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
new file mode 100644
index 0000000..4f9aa98
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
@@ -0,0 +1,252 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+bool profile_grouped_conv_bwd_weight_impl(int do_verification,
+                                          int init_method,
+                                          bool do_log,
+                                          bool time_kernel,
+                                          const ck::utils::conv::ConvParam& conv_param,
+                                          ck::index_t split_k)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    Tensor<InDataType> input(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weight_host_result(wei_g_k_c_xs_desc);
+    Tensor<WeiDataType> weight_device_result(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> output(out_g_n_k_wos_desc);
+
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "weight: " << weight_host_result.mDesc << std::endl;
+    std::cout << "output: " << output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        output.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        break;
+    default:
+        input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        output.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) *
+                             weight_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * output.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(input.mData.data());
+    out_device_buf.ToDevice(output.mData.data());
+
+    if(do_verification)
+    {
+        auto ref_conv     = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+                                                                           InDataType,
+                                                                           WeiDataType,
+                                                                           OutDataType,
+                                                                           InElementOp,
+                                                                           WeiElementOp,
+                                                                           OutElementOp>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(input,
+                                                  weight_host_result,
+                                                  output,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NDimSpatial,
+                                                                              InLayout,
+                                                                              WeiLayout,
+                                                                              OutLayout,
+                                                                              InDataType,
+                                                                              WeiDataType,
+                                                                              OutDataType,
+                                                                              InElementOp,
+                                                                              WeiElementOp,
+                                                                              OutElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device Conv instances
+    bool all_pass = true;
+
+    std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto range_copy = [](const auto& from, auto to) { std::copy(begin(from), end(from), to); };
+
+    range_copy(conv_param.input_spatial_lengths_, begin(input_spatial_lengths));
+    range_copy(conv_param.filter_spatial_lengths_, begin(filter_spatial_lengths));
+    range_copy(conv_param.output_spatial_lengths_, begin(output_spatial_lengths));
+    range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
+    range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
+    range_copy(conv_param.input_left_pads_, begin(input_left_pads));
+    range_copy(conv_param.input_right_pads_, begin(input_right_pads));
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                        conv_param.G_,
+                                        conv_param.N_,
+                                        conv_param.K_,
+                                        conv_param.C_,
+                                        input_spatial_lengths,
+                                        filter_spatial_lengths,
+                                        output_spatial_lengths,
+                                        conv_filter_strides,
+                                        conv_filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        in_element_op,
+                                        wei_element_op,
+                                        out_element_op,
+                                        split_k);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // using atomic add, so need to reset input
+            wei_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                wei_device_buf.FromDevice(weight_device_result.mData.data());
+
+                bool pass = ck::utils::check_err(weight_device_result, weight_host_result);
+
+                if(!pass)
+                {
+                    std::cout << "Fail info: " << op_ptr->GetTypeString() << std::endl;
+                }
+
+                all_pass &= pass;
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "output : ", output.mData, ",") << std::endl;
+                    ;
+                    LogRangeAsType<float>(
+                        std::cout << "weight (device): ", weight_device_result.mData, ",")
+                        << std::endl;
+                    ;
+                    LogRangeAsType<float>(
+                        std::cout << "weight (host): ", weight_host_result.mData, ",")
+                        << std::endl;
+                    ;
+                    LogRangeAsType<float>(std::cout << "input: ", input.mData, ",") << std::endl;
+                    ;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return all_pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
new file mode 100644
index 0000000..b201a2e
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
@@ -0,0 +1,253 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+bool profile_grouped_conv_fwd_impl(int do_verification,
+                                   int init_method,
+                                   bool do_log,
+                                   bool time_kernel,
+                                   const ck::utils::conv::ConvParam& conv_param)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    Tensor<InDataType> input(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> host_output(out_g_n_k_wos_desc);
+    Tensor<OutDataType> device_output(out_g_n_k_wos_desc);
+
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "weight: " << weight.mDesc << std::endl;
+    std::cout << "output: " << host_output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        weight.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * device_output.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(input.mData.data());
+    wei_device_buf.ToDevice(weight.mData.data());
+
+    // run reference op
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>{};
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(input,
+                                                  weight,
+                                                  host_output,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        // init host output to zero
+        host_output.SetZero();
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    bool pass = true;
+
+    auto run_impl = [&](auto& op_ptr, auto& argument_ptr) {
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init output to zero before profiling next kernel
+            out_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+            float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                out_device_buf.FromDevice(device_output.mData.data());
+
+                pass = pass & ck::utils::check_err(device_output, host_output);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "input : ", input.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "weight: ", weight.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "host_output  : ", host_output.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "device_output: ", device_output.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    };
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                                                 InLayout,
+                                                                                 WeiLayout,
+                                                                                 ck::Tuple<>,
+                                                                                 OutLayout,
+                                                                                 InDataType,
+                                                                                 WeiDataType,
+                                                                                 ck::Tuple<>,
+                                                                                 OutDataType,
+                                                                                 InElementOp,
+                                                                                 WeiElementOp,
+                                                                                 OutElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "xdl found " << op_ptrs.size() << " instances" << std::endl;
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(),
+                                                        wei_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        out_device_buf.GetDeviceBuffer(),
+                                                        a_g_n_c_wis_lengths,
+                                                        a_g_n_c_wis_strides,
+                                                        b_g_k_c_xs_lengths,
+                                                        b_g_k_c_xs_strides,
+                                                        {},
+                                                        {},
+                                                        e_g_n_k_wos_lengths,
+                                                        e_g_n_k_wos_strides,
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        in_element_op,
+                                                        wei_element_op,
+                                                        out_element_op);
+
+        run_impl(op_ptr, argument_ptr);
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_grouped_gemm_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_grouped_gemm_impl.hpp
new file mode 100644
index 0000000..04f94a0
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_grouped_gemm_impl.hpp
@@ -0,0 +1,291 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_grouped_gemm_impl(int do_verification,
+                               int init_method,
+                               bool do_log,
+                               bool time_kernel,
+                               const std::vector<int>& Ms,
+                               const std::vector<int>& Ns,
+                               const std::vector<int>& Ks,
+                               const std::vector<int>& StrideAs,
+                               const std::vector<int>& StrideBs,
+                               const std::vector<int>& StrideCs)
+{
+
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    std::size_t group_count = Ms.size();
+
+    if(!(group_count == Ns.size() && group_count == Ks.size() && group_count == StrideAs.size() &&
+         group_count == StrideBs.size() && group_count == StrideCs.size()))
+    {
+        throw std::runtime_error("wrong! inconsistent M/N/Ks, StrideA/B/Cs size\n");
+    }
+
+    std::vector<Tensor<ADataType>> a_m_k;
+    std::vector<Tensor<BDataType>> b_k_n;
+    std::vector<Tensor<CDataType>> c_m_n_device_results;
+
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        a_m_k.push_back(
+            Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
+        b_k_n.push_back(
+            Tensor<BDataType>(f_host_tensor_descriptor(Ks[i], Ns[i], StrideBs[i], BLayout{})));
+
+        c_m_n_device_results.push_back(
+            Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
+
+        std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n[" << i
+                  << "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
+                  << "]:" << c_m_n_device_results[i].mDesc << std::endl;
+
+        std::size_t num_thread = 1;
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_m_k[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
+            b_k_n[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            a_m_k[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
+            b_k_n[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+        }
+
+        c_m_n_device_results[i].GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    // if(do_verification)
+    // {
+
+    // }
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+    std::vector<DeviceMemPtr> a_device_buf, b_device_buf, c_device_buf;
+
+    a_device_buf.reserve(group_count);
+    b_device_buf.reserve(group_count);
+    c_device_buf.reserve(group_count);
+
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_c;
+
+    p_a.reserve(group_count);
+    p_b.reserve(group_count);
+    p_c.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        a_device_buf.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpaceSize()));
+        b_device_buf.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpaceSize()));
+
+        c_device_buf.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(CDataType) * c_m_n_device_results[i].mDesc.GetElementSpaceSize()));
+
+        a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
+        b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
+        c_device_buf[i]->ToDevice(c_m_n_device_results[i].mData.data());
+
+        gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
+
+        p_a.push_back(a_device_buf[i]->GetDeviceBuffer());
+        p_b.push_back(b_device_buf[i]->GetDeviceBuffer());
+        p_c.push_back(c_device_buf[i]->GetDeviceBuffer());
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
+                                                                     BLayout,
+                                                                     ck::Tuple<>,
+                                                                     CLayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     ck::Tuple<>,
+                                                                     CDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     CElementOp>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    if(op_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device GEMM instance found");
+    }
+
+    std::string best_gemm_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    auto p_ds = std::vector<std::array<const void*, 0>>{};
+
+    // profile device GEMM instances
+    for(auto& gemm_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            gemm_ptr->MakeArgumentPointer(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_c,
+                                          gemm_descs,
+                                          ck::tensor_operation::element_wise::PassThrough{},
+                                          ck::tensor_operation::element_wise::PassThrough{},
+                                          ck::tensor_operation::element_wise::PassThrough{});
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+        DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
+
+        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string gemm_name = gemm_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = 0, num_btype = 0;
+            for(std::size_t i = 0; i < gemm_descs.size(); i++)
+            {
+                flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
+
+                num_btype += sizeof(ADataType) * Ms[i] * Ks[i] + sizeof(BDataType) * Ks[i] * Ns[i] +
+                             sizeof(CDataType) * Ms[i] * Ns[i];
+            }
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << gemm_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_gemm_name  = gemm_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                {
+
+                    c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
+
+                    Tensor<CDataType> c_m_n_host_result(
+                        f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{}));
+
+                    using ReferenceGemmInstance =
+                        ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                  BDataType,
+                                                                  CDataType,
+                                                                  AccDataType,
+                                                                  AElementOp,
+                                                                  BElementOp,
+                                                                  CElementOp>;
+
+                    auto ref_gemm    = ReferenceGemmInstance{};
+                    auto ref_invoker = ref_gemm.MakeInvoker();
+
+                    auto ref_argument = ref_gemm.MakeArgument(a_m_k[i],
+                                                              b_k_n[i],
+                                                              c_m_n_host_result,
+                                                              a_element_op,
+                                                              b_element_op,
+                                                              c_element_op);
+
+                    ref_invoker.Run(ref_argument);
+                    pass = pass && ck::utils::check_err(c_m_n_device_results[i], c_m_n_host_result);
+
+                    if(do_log)
+                    {
+                        LogRangeAsType<float>(std::cout << "a : ", a_m_k[i].mData, ",")
+                            << std::endl;
+                        LogRangeAsType<float>(std::cout << "b: ", b_k_n[i].mData, ",") << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_device: ", c_m_n_device_results[i].mData, ",")
+                            << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
+                            << std::endl;
+                    }
+                }
+            }
+        }
+        else
+        {
+            std::cout << "does not support this GEMM problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
+
+    return pass;
+} // namespace profiler
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_groupnorm_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_groupnorm_impl.hpp
new file mode 100644
index 0000000..81fec55
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_groupnorm_impl.hpp
@@ -0,0 +1,208 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/normalization.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType>
+bool profile_groupnorm_impl(int do_verification,
+                            int init_method,
+                            bool do_log,
+                            bool time_kernel,
+                            std::vector<index_t> length)
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    if(length.size() != 5)
+        return false;
+
+    index_t G = length[3];
+    index_t C = length[4];
+
+    std::vector<index_t> reduce_dim      = {1, 2, 4};
+    std::vector<index_t> gammaBetaLength = {G, C};
+    std::vector<index_t> gammaBetaStride = {0, 0, 0, C, 1};
+
+    Tensor<XDataType> x(length);
+    Tensor<GammaDataType> gamma(gammaBetaLength);
+    Tensor<BetaDataType> beta(gammaBetaLength);
+    Tensor<YDataType> y(length);
+    Tensor<YDataType> host_y(length);
+
+    switch(init_method)
+    {
+    case 0:
+        x.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        gamma.GenerateTensorValue(GeneratorTensor_1<GammaDataType>{});
+        beta.GenerateTensorValue(GeneratorTensor_1<BetaDataType>{});
+        break;
+    case 1:
+        x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5});
+        gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
+        beta.GenerateTensorValue(GeneratorTensor_2<BetaDataType>{-5, 5});
+        break;
+    default:
+        x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1});
+        gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-0.5, 0.5});
+        beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    // add device normalization instances
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
+                                                                       GammaDataType,
+                                                                       BetaDataType,
+                                                                       AccDataType,
+                                                                       YDataType,
+                                                                       PassThrough,
+                                                                       5,
+                                                                       3>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferenceInstance = ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
+                                                                                 GammaDataType,
+                                                                                 BetaDataType,
+                                                                                 YDataType,
+                                                                                 AccDataType,
+                                                                                 PassThrough>;
+
+        ReferenceInstance ref;
+        auto ref_argument = ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, length, 1e-6);
+        auto ref_invoker  = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            length,
+            std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
+            gammaBetaStride,
+            gammaBetaStride,
+            std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+            reduce_dim,
+            1e-6,
+            x_dev.GetDeviceBuffer(),
+            gamma_dev.GetDeviceBuffer(),
+            beta_dev.GetDeviceBuffer(),
+            y_dev.GetDeviceBuffer(),
+            nullptr,
+            nullptr,
+            PassThrough{});
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            continue;
+        }
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes = x.mDesc.GetElementSize() * sizeof(XDataType) +
+                                gamma.mDesc.GetElementSize() * sizeof(GammaDataType) +
+                                beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
+                                y.mDesc.GetElementSize() * sizeof(YDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            y_dev.FromDevice(y.mData.data());
+
+            bool pass = ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(std::cout << "x  : ", x.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "host_y  : ", host_y.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "y  : ", y.mData, ",") << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", length, ",") << ", ";
+        std::cout << "num_kernel = " << num_kernel << ", best perf = " << best_avg_time << " ms, "
+                  << best_gb_per_sec << " GB/s, " << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_layernorm_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_layernorm_impl.hpp
new file mode 100644
index 0000000..eb21d4a
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_layernorm_impl.hpp
@@ -0,0 +1,227 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/normalization.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          index_t Rank>
+bool profile_layernorm_impl(int do_verification,
+                            int init_method,
+                            bool do_log,
+                            bool time_kernel,
+                            std::vector<index_t> length)
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    if(length.size() < 2)
+        return false;
+
+    // Assume normalize dimension except for batch (first) dimension
+    std::vector<index_t> reduce_length{length.begin() + 1, length.end()};
+    std::vector<index_t> reduce_dim;
+    for(int i = 1; i < Rank; ++i)
+        reduce_dim.push_back(i);
+
+    Tensor<XDataType> x(length);
+    Tensor<GammaDataType> gamma(reduce_length);
+    Tensor<BetaDataType> beta(reduce_length);
+    Tensor<YDataType> y(length);
+    Tensor<YDataType> host_y(length);
+
+    std::vector<index_t> strideXY =
+        std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()};
+    std::vector<index_t> strideGammaBeta = strideXY;
+    strideGammaBeta[0]                   = 0;
+
+    switch(init_method)
+    {
+    case 0:
+        x.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        gamma.GenerateTensorValue(GeneratorTensor_1<GammaDataType>{});
+        beta.GenerateTensorValue(GeneratorTensor_1<BetaDataType>{});
+        y.GenerateTensorValue(GeneratorTensor_1<YDataType>{});
+        break;
+    case 1:
+        x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5});
+        gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
+        beta.GenerateTensorValue(GeneratorTensor_2<BetaDataType>{-5, 5});
+        y.GenerateTensorValue(GeneratorTensor_2<YDataType>{-5, 5});
+        break;
+    default:
+        x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1});
+        gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-0.5, 0.5});
+        beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-0.5, 0.5});
+        y.GenerateTensorValue(GeneratorTensor_3<YDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    constexpr int NumReduceDim = Rank - 1;
+
+    // add device normalization instances
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
+                                                                       GammaDataType,
+                                                                       BetaDataType,
+                                                                       AccDataType,
+                                                                       YDataType,
+                                                                       PassThrough,
+                                                                       Rank,
+                                                                       NumReduceDim>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferenceInstance = ck::tensor_operation::host::ReferenceLayernorm<XDataType,
+                                                                                 GammaDataType,
+                                                                                 BetaDataType,
+                                                                                 YDataType,
+                                                                                 AccDataType,
+                                                                                 PassThrough,
+                                                                                 Rank,
+                                                                                 NumReduceDim>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, length, reduce_dim, 1e-4);
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(length,
+                                                          strideXY,
+                                                          strideGammaBeta,
+                                                          strideGammaBeta,
+                                                          strideXY,
+                                                          reduce_dim,
+                                                          1e-4,
+                                                          x_dev.GetDeviceBuffer(),
+                                                          gamma_dev.GetDeviceBuffer(),
+                                                          beta_dev.GetDeviceBuffer(),
+                                                          y_dev.GetDeviceBuffer(),
+                                                          nullptr,
+                                                          nullptr,
+                                                          PassThrough{});
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
+                LogRange(std::cout << "input lengths = ", length, ", ") << std::endl;
+            }
+
+            continue;
+        }
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes = x.mDesc.GetElementSize() * sizeof(XDataType) +
+                                gamma.mDesc.GetElementSize() * sizeof(GammaDataType) +
+                                beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
+                                y.mDesc.GetElementSize() * sizeof(YDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            y_dev.FromDevice(y.mData.data());
+
+            bool pass = ck::utils::check_err(
+                y.mData, host_y.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(std::cout << "x  : ", x.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "host_y  : ", host_y.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "y  : ", y.mData, ",") << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", length, ",") << ", ";
+        LogRange(std::cout << "stride = ", strideXY, ",") << ", ";
+        LogRange(std::cout << "reduce dims ", reduce_dim, ",") << std::endl;
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_reduce_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_reduce_impl.hpp
new file mode 100644
index 0000000..ccb9939
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_reduce_impl.hpp
@@ -0,0 +1,520 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_reduction.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <index_t Rank,
+          index_t NumReduceDim,
+          ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool UseIndex>
+struct ReduceDescription
+{
+    static constexpr index_t Rank_              = Rank;
+    static constexpr index_t NumReduceDim_      = NumReduceDim;
+    static constexpr ReduceTensorOp ReduceOpId_ = ReduceOpId;
+    static constexpr bool PropagateNan_         = PropagateNan;
+    static constexpr bool UseIndex_             = UseIndex;
+};
+
+using reduce_description_instances =
+    std::tuple<ReduceDescription<4, 3, ReduceTensorOp::ADD, false, false>, // for ADD
+               ReduceDescription<4, 4, ReduceTensorOp::ADD, false, false>,
+               ReduceDescription<4, 1, ReduceTensorOp::ADD, false, false>,
+               ReduceDescription<2, 1, ReduceTensorOp::ADD, false, false>,
+
+               ReduceDescription<4, 3, ReduceTensorOp::AVG, false, false>, // for AVG
+               ReduceDescription<4, 4, ReduceTensorOp::AVG, false, false>,
+               ReduceDescription<4, 1, ReduceTensorOp::AVG, false, false>,
+               ReduceDescription<2, 1, ReduceTensorOp::AVG, false, false>,
+
+               ReduceDescription<4, 3, ReduceTensorOp::NORM2, false, false>, // for NORM2
+               ReduceDescription<4, 4, ReduceTensorOp::NORM2, false, false>,
+               ReduceDescription<4, 1, ReduceTensorOp::NORM2, false, false>,
+               ReduceDescription<2, 1, ReduceTensorOp::NORM2, false, false>,
+
+               ReduceDescription<4, 3, ReduceTensorOp::MIN, false, false>, // for MIN
+               ReduceDescription<4, 4, ReduceTensorOp::MIN, false, false>,
+               ReduceDescription<4, 1, ReduceTensorOp::MIN, false, false>,
+               ReduceDescription<2, 1, ReduceTensorOp::MIN, false, false>,
+               ReduceDescription<4, 3, ReduceTensorOp::MAX, false, false>, // for MAX
+               ReduceDescription<4, 4, ReduceTensorOp::MAX, false, false>,
+               ReduceDescription<4, 1, ReduceTensorOp::MAX, false, false>,
+               ReduceDescription<2, 1, ReduceTensorOp::MAX, false, false>,
+               ReduceDescription<4, 3, ReduceTensorOp::AMAX, false, false>, // for AMAX
+               ReduceDescription<4, 4, ReduceTensorOp::AMAX, false, false>,
+               ReduceDescription<4, 1, ReduceTensorOp::AMAX, false, false>,
+               ReduceDescription<2, 1, ReduceTensorOp::AMAX, false, false>,
+
+               ReduceDescription<4, 3, ReduceTensorOp::MIN, false, true>, // for MIN
+               ReduceDescription<4, 4, ReduceTensorOp::MIN, false, true>,
+               ReduceDescription<4, 1, ReduceTensorOp::MIN, false, true>,
+               ReduceDescription<2, 1, ReduceTensorOp::MIN, false, true>,
+               ReduceDescription<4, 3, ReduceTensorOp::MAX, false, true>, // for MAX
+               ReduceDescription<4, 4, ReduceTensorOp::MAX, false, true>,
+               ReduceDescription<4, 1, ReduceTensorOp::MAX, false, true>,
+               ReduceDescription<2, 1, ReduceTensorOp::MAX, false, true>,
+               ReduceDescription<4, 3, ReduceTensorOp::AMAX, false, true>, // for AMAX
+               ReduceDescription<4, 4, ReduceTensorOp::AMAX, false, true>,
+               ReduceDescription<4, 1, ReduceTensorOp::AMAX, false, true>,
+               ReduceDescription<2, 1, ReduceTensorOp::AMAX, false, true>>;
+
+template <typename DescriptionType>
+bool description_match(const DescriptionType& description,
+                       int Rank,
+                       const std::vector<int>& reduceDims,
+                       ReduceTensorOp ReduceOpId,
+                       bool PropagateNan,
+                       bool UseIndex)
+{
+    if(description.Rank_ != Rank || description.ReduceOpId_ != ReduceOpId ||
+       description.PropagateNan_ != PropagateNan || description.UseIndex_ != UseIndex)
+        return (false);
+
+    if(DescriptionType::NumReduceDim_ != reduceDims.size())
+        return (false);
+
+    bool result = true;
+
+    return (result);
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <int Rank, int NumReduceDim>
+static inline std::array<int, Rank - NumReduceDim>
+get_invariant_dims(const std::array<int, NumReduceDim>& reduceDims)
+{
+    int reduceFlag = 0;
+
+    // flag the bits for the reduceDims
+    for(int i = 0; i < NumReduceDim; i++)
+    {
+        reduceFlag |= 1 << reduceDims[i];
+    };
+
+    std::array<int, Rank - NumReduceDim> invariantDims;
+
+    // collect invariant dimensions
+    int dim = 0;
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) == 0)
+        {
+            invariantDims[dim] = i;
+            dim++;
+        };
+
+    return invariantDims;
+};
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          int Rank,
+          int NumReduceDim,
+          ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool UseIndex>
+bool profile_reduce_impl_impl(bool do_verification,
+                              int init_method,
+                              bool do_dumpout,
+                              bool time_kernel,
+                              const std::vector<size_t>& inLengths,
+                              const std::array<int, NumReduceDim>& reduceDims,
+                              float alpha,
+                              float beta)
+{
+    using namespace ck::tensor_operation::device;
+    using namespace ck::tensor_operation::device::instance;
+    using ck::host_common::dumpBufferToFile;
+
+    constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
+    constexpr bool op_support_indices =
+        (ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
+         ReduceOpId == ReduceTensorOp::AMAX);
+
+    constexpr bool OutputIndex = (op_support_indices && UseIndex);
+
+    constexpr bool out_support_atomic_add = std::is_same<OutDataType, float>::value;
+    constexpr bool op_support_atomic_add =
+        !op_support_indices && ReduceOpId != ReduceTensorOp::NORM2;
+    constexpr bool use_atomic_add = (out_support_atomic_add && op_support_atomic_add);
+
+    // 1) If InDataType is half_t, must use half_t as AccDataType for indexable reduction operations
+    // 2) If InDataType is half_t, must use float as AccDataType for non-indexable reduction
+    // operations
+    constexpr bool invalid_reduce_1 =
+        std::is_same<InDataType, half_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, float>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, half_t>::value));
+
+    // 1) If InDataType is float, must use float as AccDataType for indexable reduction operations
+    constexpr bool invalid_reduce_2 =
+        std::is_same<InDataType, float>::value &&
+        (op_support_indices && !std::is_same<AccDataType, float>::value);
+
+    // 1) The indices can only be used when the reduction operation is indexable
+    constexpr bool invalid_reduce_3 = (!op_support_indices && UseIndex);
+
+    // 1) If InDataType is int8_t, must use int8_t as AccDataType for indexable reduction operations
+    // 2) If InDataType is int8_t, must use int32_t as AccDataType for non-indexable reduction
+    // operations
+    constexpr bool invalid_reduce_4 =
+        std::is_same<InDataType, int8_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, int8_t>::value));
+
+    // 1) If InDataType is int8_t, the supported operation must be either indexable operations or
+    // ADD/AVG
+    constexpr bool invalid_reduce_5 = std::is_same<InDataType, int8_t>::value &&
+                                      (!op_support_indices && ReduceOpId != ReduceTensorOp::ADD &&
+                                       ReduceOpId != ReduceTensorOp::AVG);
+
+    // 1) If InDataType is bhalf_t, must use float as AccDataType for all reduction operations
+    constexpr bool invalid_reduce_6 =
+        std::is_same<InDataType, bhalf_t>::value && !std::is_same<AccDataType, float>::value;
+
+    constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
+                                     invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
+
+    bool pass = true;
+
+    if constexpr(!invalid_reduce)
+    {
+        Tensor<InDataType> in(inLengths);
+
+        std::vector<size_t> outLengths;
+
+        const auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
+
+        if(reduceDims.size() == Rank)
+            outLengths.push_back(1);
+        else
+            for(auto dim : invariantDims)
+                outLengths.push_back(inLengths[dim]);
+
+        Tensor<OutDataType> out_ref(outLengths);
+        Tensor<OutDataType> out(outLengths);
+        Tensor<int32_t> out_indices_ref(outLengths);
+        Tensor<int32_t> out_indices(outLengths);
+
+        auto inStrides  = in.mDesc.GetStrides();
+        auto outStrides = out.mDesc.GetStrides();
+
+        size_t invariant_total_length = out.mDesc.GetElementSize();
+        size_t reduce_total_length    = in.mDesc.GetElementSize() / invariant_total_length;
+
+        std::size_t num_thread = 1;
+
+        if(do_verification)
+        {
+            switch(init_method)
+            {
+            case 0: break;
+            case 1:
+                in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
+                if(beta != 0.0f)
+                    out_ref.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
+                break;
+            case 2:
+                in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
+                if(beta != 0.0f)
+                    out_ref.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
+                break;
+            default:
+                in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
+                if(beta != 0.0f)
+                    out_ref.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0},
+                                                num_thread);
+            }
+
+            if(beta != 0.0f)
+                for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                    out.mData[i] = out_ref.mData[i];
+        };
+
+        // these buffers are usually provided by the user application
+        DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+        DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+
+        in_dev.ToDevice(in.mData.data());
+
+        if(beta != 0.0f)
+            out_dev.ToDevice(out.mData.data());
+
+        size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int) : 0;
+
+        DeviceMem out_indices_dev(indicesSizeInBytes);
+
+        float best_avg_time   = 0;
+        float best_gb_per_sec = 0;
+
+        using InElementwiseOperation =
+            typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+        using AccElementwiseOperation =
+            typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+        using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+
+        InElementwiseOperation in_elementwise_op;
+        AccElementwiseOperation acc_elementwise_op;
+
+        std::tie(in_elementwise_op, acc_elementwise_op) =
+            reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+                static_cast<int32_t>(reduce_total_length));
+
+        using DeviceReduceInstPtr =
+            DeviceReducePtr<Rank, NumReduceDim, InElementwiseOperation, AccElementwiseOperation>;
+
+        std::vector<DeviceReduceInstPtr> reduce_ptrs;
+
+        add_device_reduce_instance_threadwise<InDataType,
+                                              AccDataType,
+                                              OutDataType,
+                                              Rank,
+                                              NumReduceDim,
+                                              ReduceOperation,
+                                              InElementwiseOperation,
+                                              AccElementwiseOperation,
+                                              PropagateNan,
+                                              UseIndex>(reduce_ptrs);
+
+        add_device_reduce_instance_blockwise<InDataType,
+                                             AccDataType,
+                                             OutDataType,
+                                             Rank,
+                                             NumReduceDim,
+                                             ReduceOperation,
+                                             InElementwiseOperation,
+                                             AccElementwiseOperation,
+                                             PropagateNan,
+                                             UseIndex>(reduce_ptrs);
+
+        if constexpr(use_atomic_add)
+        {
+            add_device_reduce_instance_multiblock_atomic_add<InDataType,
+                                                             AccDataType,
+                                                             OutDataType,
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             ReduceOperation,
+                                                             InElementwiseOperation,
+                                                             AccElementwiseOperation,
+                                                             PropagateNan,
+                                                             UseIndex>(reduce_ptrs);
+        }
+
+        if(reduce_ptrs.empty())
+        {
+            throw std::runtime_error("Wrong! No device REDUCE instance found");
+        };
+
+        if(do_verification)
+        {
+            ReductionHost<InDataType,
+                          AccDataType,
+                          OutDataType,
+                          ReduceOperation,
+                          InElementwiseOperation,
+                          AccElementwiseOperation,
+                          Rank,
+                          NumReduceDim,
+                          PropagateNan,
+                          OutputIndex>
+                hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
+
+            hostReduce.Run(alpha,
+                           in.mData.data(),
+                           beta,
+                           out_ref.mData.data(),
+                           out_indices_ref.mData.data(),
+                           in_elementwise_op,
+                           acc_elementwise_op);
+        };
+
+        std::array<index_t, Rank> arrInLengths;
+        std::array<index_t, Rank> arrInStrides;
+        std::array<index_t, NumOutDim> arrOutLengths;
+        std::array<index_t, NumOutDim> arrOutStrides;
+
+        ck::ranges::copy(inLengths, arrInLengths.begin());
+        ck::ranges::copy(inStrides, arrInStrides.begin());
+        ck::ranges::copy(outLengths, arrOutLengths.begin());
+        ck::ranges::copy(outStrides, arrOutStrides.begin());
+
+        for(auto& reduce_ptr : reduce_ptrs)
+        {
+            auto argument_ptr = reduce_ptr->MakeArgumentPointer(arrInLengths,
+                                                                arrInStrides,
+                                                                arrOutLengths,
+                                                                arrOutStrides,
+                                                                reduceDims,
+                                                                alpha,
+                                                                beta,
+                                                                in_dev.GetDeviceBuffer(),
+                                                                nullptr,
+                                                                out_dev.GetDeviceBuffer(),
+                                                                out_indices_dev.GetDeviceBuffer(),
+                                                                in_elementwise_op,
+                                                                acc_elementwise_op);
+
+            if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
+                continue;
+
+            std::string reduce_name = reduce_ptr->GetTypeString();
+
+            auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t num_bytes =
+                invariant_total_length * reduce_total_length * sizeof(InDataType) +
+                invariant_total_length * sizeof(OutDataType);
+
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            if(time_kernel)
+                std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                          << reduce_name << std::endl;
+
+            if(gb_per_sec > best_gb_per_sec)
+            {
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                bool single_pass;
+
+                out_dev.FromDevice(out.mData.data());
+                single_pass = ck::utils::check_err(out, out_ref);
+
+                if(OutputIndex)
+                {
+                    out_indices_dev.FromDevice(out_indices.mData.data());
+                    single_pass = single_pass && ck::utils::check_err(out_indices, out_indices_ref);
+                };
+
+                if(!single_pass)
+                {
+                    std::cout << "Fail Info: " << reduce_ptr->GetTypeString() << std::endl;
+                }
+
+                pass = pass && single_pass;
+            };
+
+            if(do_dumpout)
+            {
+                dumpBufferToFile("dump_in.bin", in.mData.data(), in.mDesc.GetElementSize());
+                dumpBufferToFile("dump_out.bin", out.mData.data(), out.mDesc.GetElementSize());
+                dumpBufferToFile(
+                    "dump_out_host.bin", out_ref.mData.data(), out_ref.mDesc.GetElementSize());
+                if(OutputIndex)
+                {
+                    dumpBufferToFile("dump_indices.bin",
+                                     out_indices.mData.data(),
+                                     out_indices.mDesc.GetElementSize());
+                    dumpBufferToFile("dump_indices_host.bin",
+                                     out_indices_ref.mData.data(),
+                                     out_indices_ref.mDesc.GetElementSize());
+                };
+            };
+        };
+
+        if(time_kernel)
+            std::cout << "Best Perf: " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s"
+                      << std::endl;
+    }
+    else
+    {
+        std::cout << "The requested reduction operation is not supported, please check !!!"
+                  << std::endl;
+    };
+
+    return pass;
+};
+
+template <typename InDataType, typename AccDataType, typename OutDataType>
+bool profile_reduce_impl(bool do_verification,
+                         int init_method,
+                         bool do_dumpout,
+                         bool time_kernel,
+                         const std::vector<size_t>& inLengths,
+                         const std::vector<int>& reduceDims,
+                         ReduceTensorOp ReduceOpId,
+                         bool PropagateNan,
+                         bool UseIndex,
+                         float alpha,
+                         float beta)
+{
+    bool matched = false;
+    bool pass    = true;
+
+    using tuple_of_description_instances =
+        tensor_operation::device::instance::reduce_description_instances;
+
+    const auto tuple_object = tuple_of_description_instances{};
+
+    static_for<0, std::tuple_size<tuple_of_description_instances>::value, 1>{}([&](auto i) {
+        if(matched)
+            return;
+
+        using descType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
+
+        if(!description_match(
+               descType{}, inLengths.size(), reduceDims, ReduceOpId, PropagateNan, UseIndex))
+            return;
+
+        std::array<ck::index_t, descType::NumReduceDim_> arrReduceDims;
+
+        ck::ranges::copy(reduceDims, arrReduceDims.begin());
+
+        pass = pass && profile_reduce_impl_impl<InDataType,
+                                                AccDataType,
+                                                OutDataType,
+                                                descType::Rank_,
+                                                descType::NumReduceDim_,
+                                                static_cast<ReduceTensorOp>(descType::ReduceOpId_),
+                                                descType::PropagateNan_,
+                                                descType::UseIndex_>(do_verification,
+                                                                     init_method,
+                                                                     do_dumpout,
+                                                                     time_kernel,
+                                                                     inLengths,
+                                                                     arrReduceDims,
+                                                                     alpha,
+                                                                     beta);
+
+        matched = true;
+    });
+
+    return pass;
+};
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/include/profiler/profile_softmax_impl.hpp b/3rdparty/composable_kernel/profiler/include/profiler/profile_softmax_impl.hpp
new file mode 100644
index 0000000..090cdaa
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/include/profiler/profile_softmax_impl.hpp
@@ -0,0 +1,219 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <iomanip>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/data_type.hpp"
+
+namespace ck {
+namespace profiler {
+
+enum struct SoftmaxDataType
+{
+    F32_F32, // in, out
+    F16_F16,
+    BF16_BF16,
+    INT8_INT8,
+};
+
+// clang-format off
+template <typename SoftmaxDataType> std::string type_to_string();
+template <> std::string type_to_string<float>()   { return "f32"; }
+template <> std::string type_to_string<half_t>()  { return "f16"; }
+template <> std::string type_to_string<bhalf_t>() { return "bf16"; }
+template <> std::string type_to_string<int8_t>()  { return "int8"; }
+template <> std::string type_to_string<int32_t>() { return "int32"; }
+// clang-format on
+
+template <typename InDataType, typename AccDataType, typename OutDataType, index_t Rank>
+bool profile_softmax_impl(int do_verification,
+                          int init_method,
+                          bool do_log,
+                          bool time_kernel,
+                          std::vector<index_t> in_length,
+                          std::vector<index_t> in_strides,
+                          std::vector<index_t> reduce_dims,
+                          AccDataType alpha,
+                          AccDataType beta)
+{
+    if(Rank != in_length.size())
+    {
+        throw std::runtime_error("Input tensor rank is different from template argument Rank!");
+    }
+
+    Tensor<InDataType> in = in_strides.empty() ? Tensor<InDataType>(in_length)
+                                               : Tensor<InDataType>(in_length, in_strides);
+    Tensor<OutDataType> out(in.mDesc);
+    Tensor<OutDataType> prior_out(in.mDesc);
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<InDataType>{-5.f, 5.f}(in.begin(), in.end());
+        ck::utils::FillUniformDistributionIntegerValue<OutDataType>{-5.f, 5.f}(prior_out.begin(),
+                                                                               prior_out.end());
+        break;
+    default:
+        ck::utils::FillUniformDistribution<InDataType>{0.0f, 1.0f}(in);
+        ck::utils::FillUniformDistribution<OutDataType>{-0.5f, 0.5f}(prior_out);
+    }
+
+    Tensor<OutDataType> out_ref(prior_out);
+
+    if(do_verification)
+    {
+        using ReferenceSoftmax =
+            tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>;
+        ReferenceSoftmax{}.MakeInvoker().Run({in, out_ref, alpha, beta, reduce_dims});
+    }
+
+    DeviceMem in_dev(in.GetElementSpaceSizeInBytes());
+    DeviceMem out_dev(out.GetElementSpaceSizeInBytes());
+    in_dev.ToDevice(in.data());
+
+    std::vector<index_t> in_tensor_lengths(in.GetLengths().begin(), in.GetLengths().end());
+    std::vector<index_t> in_tensor_strides(in.GetStrides().begin(), in.GetStrides().end());
+
+    // add device softmax instances
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using DeviceOp    = tensor_operation::device::
+        DeviceSoftmax<InDataType, AccDataType, OutDataType, PassThrough, PassThrough, Rank>;
+
+    // get device op instances
+    const auto instances = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+    std::cout << "found " << instances.size() << " instances" << std::endl;
+
+    if(instances.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device normalization instance found");
+    }
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    std::vector<bool> instance_pass;
+
+    for(auto& inst_ptr : instances)
+    {
+        // Is this user's responsibility to check if problem mismatches kernel instance (ie. rank 3
+        // problem to rank 4 kernel) other than invoking IsSupportedArgument()?
+        if(!(inst_ptr->GetNumReduceDim() == static_cast<index_t>(reduce_dims.size())))
+        {
+            continue;
+        }
+
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(in_tensor_lengths,
+                                                          in_tensor_strides,
+                                                          reduce_dims,
+                                                          &alpha,
+                                                          &beta,
+                                                          in_dev.GetDeviceBuffer(),
+                                                          out_dev.GetDeviceBuffer(),
+                                                          PassThrough{},
+                                                          PassThrough{});
+
+        if(!inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
+            LogRange(std::cout << "input lengths = [", in_length, ", ")
+                << "], "
+                << "scaler = [" << alpha << ", " << beta << "]";
+            LogRange(std::cout << ", reduce dims = [", reduce_dims, ", ") << "]." << std::endl;
+            instance_pass.push_back(true);
+            continue;
+        }
+
+        out_dev.ToDevice(prior_out.data());
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+        float avg_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+        {
+            std::size_t num_bytes =
+                in.GetElementSize() * sizeof(InDataType) +
+                (beta == 0.0f ? 1 : 2) * out.GetElementSize() * sizeof(OutDataType);
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+            if(avg_time < best_avg_time)
+            {
+                best_instance_name = inst_ptr->GetTypeString();
+                best_avg_time      = avg_time;
+                best_gb_per_sec    = gb_per_sec;
+            }
+        }
+
+        if(do_verification)
+        {
+            out_dev.FromDevice(out.data());
+            bool pass = true;
+            if(std::is_same<InDataType, int8_t>::value)
+            {
+                pass = pass && ck::utils::check_err(
+                                   out.mData, out_ref.mData, "Error: Incorrect results!", 0, 1);
+                if(do_log)
+                {
+                    LogRangeAsType<int>(std::cout << "in  : ", in.mData, ",") << std::endl;
+                    LogRangeAsType<int>(std::cout << "out_ref  : ", out_ref.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<int>(std::cout << "out  : ", out.mData, ",") << std::endl;
+                }
+            }
+            else
+            {
+                pass = pass && ck::utils::check_err(out.mData, out_ref.mData);
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "in  : ", in.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "out_ref  : ", out_ref.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "out  : ", out.mData, ",") << std::endl;
+                }
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "input lengths = [", in_length, ", ")
+                    << "], "
+                    << "scaler = [" << alpha << ", " << beta << "]." << std::endl;
+            }
+            instance_pass.push_back(pass);
+        }
+    }
+    if(time_kernel)
+    {
+        std::cout << "Best Perf for datatype = " << type_to_string<InDataType>() << "_"
+                  << type_to_string<OutDataType>() << ", ";
+        LogRange(std::cout << "length = ", in_tensor_lengths, ",") << ", ";
+        LogRange(std::cout << "stride = ", in_tensor_strides, ",") << ", ";
+        LogRange(std::cout << "reduce dims ", reduce_dims, ",") << ", ";
+        std::cout << "alpha = " << alpha << ", "
+                  << "beta = " << beta << ", " << best_avg_time << " ms, " << best_gb_per_sec
+                  << " GB/s, " << best_instance_name << std::endl;
+    }
+    return std::all_of(
+        std::begin(instance_pass), std::end(instance_pass), [](bool p) { return p; });
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/3rdparty/composable_kernel/profiler/src/CMakeLists.txt b/3rdparty/composable_kernel/profiler/src/CMakeLists.txt
new file mode 100644
index 0000000..bc87554
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/CMakeLists.txt
@@ -0,0 +1,67 @@
+# ckProfiler
+set(PROFILER_SOURCES
+    profiler.cpp
+    profile_gemm.cpp
+    profile_gemm_splitk.cpp
+    profile_gemm_bilinear.cpp
+    profile_gemm_bias_add_reduce.cpp
+    profile_gemm_add_add_fastgelu.cpp
+    profile_gemm_add_fastgelu.cpp
+    profile_gemm_fastgelu.cpp
+    profile_gemm_reduce.cpp
+    profile_batched_gemm.cpp
+    profile_batched_gemm_gemm.cpp
+    profile_batched_gemm_add_relu_gemm_add.cpp
+    profile_batched_gemm_reduce.cpp
+    profile_grouped_gemm.cpp
+    profile_conv_fwd.cpp
+    profile_conv_fwd_bias_relu.cpp
+    profile_conv_fwd_bias_relu_add.cpp
+    profile_conv_bwd_data.cpp
+    profile_grouped_conv_fwd.cpp
+    profile_grouped_conv_bwd_weight.cpp
+    profile_reduce.cpp
+    profile_groupnorm.cpp
+    profile_layernorm.cpp
+    profile_softmax.cpp
+    profile_batchnorm_fwd.cpp
+    profile_batchnorm_bwd.cpp
+)
+
+set(PROFILER_EXECUTABLE ckProfiler)
+
+add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES})
+target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors)
+
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
+
+rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)
diff --git a/3rdparty/composable_kernel/profiler/src/profile_batched_gemm.cpp b/3rdparty/composable_kernel/profiler/src/profile_batched_gemm.cpp
new file mode 100644
index 0000000..907a373
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_batched_gemm.cpp
@@ -0,0 +1,203 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdint>
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_batched_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "batched_gemm"
+#define OP_DESC "Batched GEMM"
+
+int profile_batched_gemm(int argc, char* argv[])
+{
+    if(argc != 18)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16, 2: bf16, 3: int8)\n");
+        printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
+        printf("                     1: A[g, m, k] * B[g, n, k] = C[g, m, n];\n");
+        printf("                     2: A[g, k, m] * B[g, k, n] = C[g, m, n];\n");
+        printf("                     3: A[g, k, m] * B[g, n, k] = C[g, m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+
+    const int BatchStrideA = std::stoi(argv[14]);
+    const int BatchStrideB = std::stoi(argv[15]);
+    const int BatchStrideC = std::stoi(argv[16]);
+
+    const int BatchCount = std::stoi(argv[17]);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using INT8 = int8_t;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto c_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto c_layout) {
+        using ADataType = decltype(a_type);
+        using BDataType = decltype(b_type);
+        using CDataType = decltype(c_type);
+
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using CLayout = decltype(c_layout);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
+
+        const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA;
+        const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
+        const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+        const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
+        const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
+        const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
+
+        const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
+        const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
+        const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
+
+        bool pass = ck::profiler::
+            profile_batched_gemm_impl<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>(
+                do_verification,
+                init_method,
+                do_log,
+                time_kernel,
+                M,
+                N,
+                K,
+                BatchStrideA_,
+                BatchStrideB_,
+                BatchStrideC_,
+                StrideA_,
+                StrideB_,
+                StrideC_,
+                BatchCount);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F16{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F16{}, F16{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F16{}, F16{}, F16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F16{}, F16{}, F16{}, Col{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(BF16{}, BF16{}, BF16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(BF16{}, BF16{}, BF16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(BF16{}, BF16{}, BF16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(BF16{}, BF16{}, BF16{}, Col{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(INT8{}, INT8{}, INT8{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(INT8{}, INT8{}, INT8{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(INT8{}, INT8{}, INT8{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(INT8{}, INT8{}, INT8{}, Col{}, Col{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp b/3rdparty/composable_kernel/profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp
new file mode 100644
index 0000000..f440a30
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp
@@ -0,0 +1,214 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+#define OP_NAME "batched_gemm_add_relu_gemm_add"
+#define OP_DESC "Batched GEMM+Add+Relu+GEMM+Add"
+
+int profile_batched_gemm_add_relu_gemm_add(int argc, char* argv[])
+{
+    enum struct GemmMatrixLayout
+    {
+        MK_NK_MN_NO_MO_MO, // 0
+        MK_NK_MN_ON_MO_MO, // 1
+    };
+
+    enum struct GemmDataType
+    {
+        F32_F32_F32_F32_F32_F32, // 0
+        F16_F16_F16_F16_F16_F16, // 1
+    };
+
+    GemmDataType data_type  = GemmDataType::F16_F16_F16_F16_F16_F16;
+    GemmMatrixLayout layout = GemmMatrixLayout::MK_NK_MN_NO_MO_MO;
+    bool do_verification    = true;
+    int init_method         = 1;
+    bool do_log             = 0;
+    bool time_kernel        = false;
+
+    // GEMM shape
+    ck::index_t M             = 1024;
+    ck::index_t N             = 1024;
+    ck::index_t K             = 64;
+    ck::index_t O             = 128;
+    ck::index_t BatchCount    = 4;
+    ck::index_t StrideA0      = -1;
+    ck::index_t StrideB0      = -1;
+    ck::index_t StrideD0      = -1;
+    ck::index_t StrideB1      = -1;
+    ck::index_t StrideD1      = -1;
+    ck::index_t StrideE1      = -1;
+    ck::index_t BatchStrideA0 = -1;
+    ck::index_t BatchStrideB0 = -1;
+    ck::index_t BatchStrideD0 = -1;
+    ck::index_t BatchStrideB1 = -1;
+    ck::index_t BatchStrideD1 = -1;
+    ck::index_t BatchStrideE1 = -1;
+
+    if(argc == 8)
+    {
+        data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+        layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+        do_verification = std::stoi(argv[4]);
+        init_method     = std::stoi(argv[5]);
+        do_log          = std::stoi(argv[6]);
+        time_kernel     = std::stoi(argv[7]);
+    }
+    else if(argc == 13)
+    {
+        data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+        layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+        do_verification = std::stoi(argv[4]);
+        init_method     = std::stoi(argv[5]);
+        do_log          = std::stoi(argv[6]);
+        time_kernel     = std::stoi(argv[7]);
+
+        M          = std::stoi(argv[8]);
+        N          = std::stoi(argv[9]);
+        K          = std::stoi(argv[10]);
+        O          = std::stoi(argv[11]);
+        BatchCount = std::stoi(argv[12]);
+    }
+    else if(argc == 25)
+    {
+        data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+        layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+        do_verification = std::stoi(argv[4]);
+        init_method     = std::stoi(argv[5]);
+        do_log          = std::stoi(argv[6]);
+        time_kernel     = std::stoi(argv[7]);
+
+        M          = std::stoi(argv[8]);
+        N          = std::stoi(argv[9]);
+        K          = std::stoi(argv[10]);
+        O          = std::stoi(argv[11]);
+        BatchCount = std::stoi(argv[12]);
+
+        StrideA0 = std::stoi(argv[13]);
+        StrideB0 = std::stoi(argv[14]);
+        StrideD0 = std::stoi(argv[15]);
+        StrideB1 = std::stoi(argv[16]);
+        StrideD1 = std::stoi(argv[17]);
+        StrideE1 = std::stoi(argv[18]);
+
+        BatchStrideA0 = std::stoi(argv[19]);
+        BatchStrideB0 = std::stoi(argv[20]);
+        BatchStrideD0 = std::stoi(argv[21]);
+        BatchStrideB1 = std::stoi(argv[22]);
+        BatchStrideD1 = std::stoi(argv[23]);
+        BatchStrideE1 = std::stoi(argv[24]);
+    }
+    else
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (1: fp16)\n");
+        printf("arg3: matrix layout (0: Relu(A0[m, k] * B0[n, k] + D0[m, n]) * B1[n, o] + D1[m, o] "
+               "= E1[m, o]; 1: Relu(A0[m, k] * B0[n, k] + D0[m, n]) * B1[o, n] + D1[m, o] = "
+               "E1[m, o];)\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 12: M, N, K, O, Batch\n");
+        printf("arg13 to 18: StrideA0, StrideB0, StrideD0, StrideB1, StrideD1, StrideE1\n");
+        printf("arg19 to 24: BatchStrideA0, BatchStrideB0, BatchStrideD0, BatchStrideB1, "
+               "BatchStrideD1, BatchStrideE1 \n");
+        exit(1);
+    }
+
+    if(data_type == GemmDataType::F16_F16_F16_F16_F16_F16 &&
+       layout == GemmMatrixLayout::MK_NK_MN_NO_MO_MO)
+    {
+        ck::profiler::profile_batched_gemm_add_relu_gemm_add_impl<Row,            // A0Layout,
+                                                                  Col,            // B0Layout,
+                                                                  ck::Tuple<Row>, // D0sLayout,
+                                                                  Row,            // B1Layout,
+                                                                  ck::Tuple<Row>, // D1sLayout,
+                                                                  Row,            // E1Layout,
+                                                                  F16,            // A0DataType,
+                                                                  F16,            // B0DataType,
+                                                                  ck::Tuple<F16>, // D0DataType,
+                                                                  F16,            // B1DataType,
+                                                                  ck::Tuple<F16>, // D1sDataType
+                                                                  F16>            // E1DataType,
+            (do_verification,
+             init_method,
+             do_log,
+             time_kernel,
+             M,
+             N,
+             K,
+             O,
+             BatchCount,
+             StrideA0,
+             StrideB0,
+             StrideD0,
+             StrideB1,
+             StrideD1,
+             StrideE1,
+             BatchStrideA0,
+             BatchStrideB0,
+             BatchStrideD0,
+             BatchStrideB1,
+             BatchStrideD1,
+             BatchStrideE1);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16_F16_F16_F16 &&
+            layout == GemmMatrixLayout::MK_NK_MN_ON_MO_MO)
+    {
+        ck::profiler::profile_batched_gemm_add_relu_gemm_add_impl<Row,            // A0Layout,
+                                                                  Col,            // B0Layout,
+                                                                  ck::Tuple<Row>, // D0sLayout,
+                                                                  Col,            // B1Layout,
+                                                                  ck::Tuple<Row>, // D1sLayout,
+                                                                  Row,            // E1Layout,
+                                                                  F16,            // A0DataType,
+                                                                  F16,            // B0DataType,
+                                                                  ck::Tuple<F16>, // D0DataType,
+                                                                  F16,            // B1DataType,
+                                                                  ck::Tuple<F16>, // D1sDataType
+                                                                  F16>            // E1DataType,
+            (do_verification,
+             init_method,
+             do_log,
+             time_kernel,
+             M,
+             N,
+             K,
+             O,
+             BatchCount,
+             StrideA0,
+             StrideB0,
+             StrideD0,
+             StrideB1,
+             StrideD1,
+             StrideE1,
+             BatchStrideA0,
+             BatchStrideB0,
+             BatchStrideD0,
+             BatchStrideB1,
+             BatchStrideD1,
+             BatchStrideE1);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_add_relu_gemm_add);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_batched_gemm_gemm.cpp b/3rdparty/composable_kernel/profiler/src/profile_batched_gemm_gemm.cpp
new file mode 100644
index 0000000..6015c93
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_batched_gemm_gemm.cpp
@@ -0,0 +1,187 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_batched_gemm_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+#define OP_NAME "batched_gemm_gemm"
+#define OP_DESC "Batched GEMM+GEMM"
+
+int profile_batched_gemm_gemm(int argc, char* argv[])
+{
+    enum struct GemmMatrixLayout
+    {
+        MK_NK_NO_MO, // 0
+        MK_NK_ON_MO, // 0
+    };
+
+    enum struct GemmDataType
+    {
+        F32_F32_F32_F32, // 0
+        F16_F16_F16_F16, // 1
+    };
+
+    GemmDataType data_type  = GemmDataType::F16_F16_F16_F16;
+    GemmMatrixLayout layout = GemmMatrixLayout::MK_NK_NO_MO;
+    bool do_verification    = true;
+    int init_method         = 1;
+    bool do_log             = 0;
+    bool time_kernel        = false;
+
+    // GEMM shape
+    ck::index_t M             = 1024;
+    ck::index_t N             = 1024;
+    ck::index_t K             = 64;
+    ck::index_t O             = 128;
+    ck::index_t BatchCount    = 4;
+    ck::index_t StrideA0      = -1;
+    ck::index_t StrideB0      = -1;
+    ck::index_t StrideB1      = -1;
+    ck::index_t StrideE1      = -1;
+    ck::index_t BatchStrideA0 = -1;
+    ck::index_t BatchStrideB0 = -1;
+    ck::index_t BatchStrideB1 = -1;
+    ck::index_t BatchStrideE1 = -1;
+
+    if(argc == 8)
+    {
+        data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+        layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+        do_verification = std::stoi(argv[4]);
+        init_method     = std::stoi(argv[5]);
+        do_log          = std::stoi(argv[6]);
+        time_kernel     = std::stoi(argv[7]);
+    }
+    else if(argc == 13)
+    {
+        data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+        layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+        do_verification = std::stoi(argv[4]);
+        init_method     = std::stoi(argv[5]);
+        do_log          = std::stoi(argv[6]);
+        time_kernel     = std::stoi(argv[7]);
+
+        M          = std::stoi(argv[8]);
+        N          = std::stoi(argv[9]);
+        K          = std::stoi(argv[10]);
+        O          = std::stoi(argv[11]);
+        BatchCount = std::stoi(argv[12]);
+    }
+    else if(argc == 21)
+    {
+        data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+        layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+        do_verification = std::stoi(argv[4]);
+        init_method     = std::stoi(argv[5]);
+        do_log          = std::stoi(argv[6]);
+        time_kernel     = std::stoi(argv[7]);
+
+        M          = std::stoi(argv[8]);
+        N          = std::stoi(argv[9]);
+        K          = std::stoi(argv[10]);
+        O          = std::stoi(argv[11]);
+        BatchCount = std::stoi(argv[12]);
+
+        StrideA0 = std::stoi(argv[13]);
+        StrideB0 = std::stoi(argv[14]);
+        StrideB1 = std::stoi(argv[15]);
+        StrideE1 = std::stoi(argv[16]);
+
+        BatchStrideA0 = std::stoi(argv[17]);
+        BatchStrideB0 = std::stoi(argv[18]);
+        BatchStrideB1 = std::stoi(argv[19]);
+        BatchStrideE1 = std::stoi(argv[20]);
+    }
+    else
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (1: fp16)\n");
+        printf("arg3: matrix layout (0: Relu(A0[m, k] * B0[n, k] + D0[m, n]) * B1[n, o] + D1[m, o] "
+               "= E1[m, o];  1: Relu(A0[m, k] * B0[n, k] + D0[m, n]) * B1[o, n] + D1[m, o] = E1[m, "
+               "o];)\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 12: M, N, K, O, Batch\n");
+        printf("arg13 to 16: StrideA0, StrideB0, StrideB1, StrideE1\n");
+        printf("arg17 to 20: BatchStrideA0, BatchStrideB0, BatchStrideB1, BatchStrideE1 \n");
+        exit(1);
+    }
+
+    if(data_type == GemmDataType::F16_F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_NO_MO)
+    {
+        ck::profiler::profile_batched_gemm_gemm_impl<F16, // A0DataType,
+                                                     F16, // B0DataType,
+                                                     F16, // B1DataType,
+                                                     F16, // E1DataType,
+                                                     Row, // A0Layout,
+                                                     Col, // B0Layout,
+                                                     Row, // B1Layout,
+                                                     Row> // E1Layout,
+            (do_verification,
+             init_method,
+             do_log,
+             time_kernel,
+             M,
+             N,
+             K,
+             O,
+             BatchCount,
+             StrideA0,
+             StrideB0,
+             StrideB1,
+             StrideE1,
+             BatchStrideA0,
+             BatchStrideB0,
+             BatchStrideB1,
+             BatchStrideE1);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_ON_MO)
+    {
+        ck::profiler::profile_batched_gemm_gemm_impl<F16, // A0DataType,
+                                                     F16, // B0DataType,
+                                                     F16, // B1DataType,
+                                                     F16, // E1DataType,
+                                                     Row, // A0Layout,
+                                                     Col, // B0Layout,
+                                                     Col, // B1Layout,
+                                                     Row> // E1Layout,
+            (do_verification,
+             init_method,
+             do_log,
+             time_kernel,
+             M,
+             N,
+             K,
+             O,
+             BatchCount,
+             StrideA0,
+             StrideB0,
+             StrideB1,
+             StrideE1,
+             BatchStrideA0,
+             BatchStrideB0,
+             BatchStrideB1,
+             BatchStrideE1);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_gemm);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_batched_gemm_reduce.cpp b/3rdparty/composable_kernel/profiler/src/profile_batched_gemm_reduce.cpp
new file mode 100644
index 0000000..6b1dfc0
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_batched_gemm_reduce.cpp
@@ -0,0 +1,159 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_batched_gemm_reduce_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "batched_gemm_reduce"
+#define OP_DESC "Batched GEMM+Reduce"
+
+int profile_batched_gemm_reduce(int argc, char* argv[])
+{
+    enum struct GemmMatrixLayout
+    {
+        MK_KN_MN, // 0
+        MK_NK_MN, // 1
+        KM_KN_MN, // 2
+        KM_NK_MN, // 3
+    };
+
+    enum struct GemmReduceDataType
+    {
+        F32_F32_F32_F32_F32, // 0
+        F16_F16_F16_F32_F32, // 1
+    };
+
+    if(argc != 15)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmReduceDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+
+    const int BatchCount = std::stoi(argv[14]);
+
+    if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                       ck::half_t,
+                                                       ck::half_t,
+                                                       float,
+                                                       ck::tensor_layout::gemm::RowMajor,
+                                                       ck::tensor_layout::gemm::RowMajor,
+                                                       ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? K : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            BatchCount);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                       ck::half_t,
+                                                       ck::half_t,
+                                                       float,
+                                                       ck::tensor_layout::gemm::RowMajor,
+                                                       ck::tensor_layout::gemm::ColumnMajor,
+                                                       ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? K : StrideA,
+            (StrideB < 0) ? K : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            BatchCount);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                       ck::half_t,
+                                                       ck::half_t,
+                                                       float,
+                                                       ck::tensor_layout::gemm::ColumnMajor,
+                                                       ck::tensor_layout::gemm::RowMajor,
+                                                       ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            BatchCount);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                       ck::half_t,
+                                                       ck::half_t,
+                                                       float,
+                                                       ck::tensor_layout::gemm::ColumnMajor,
+                                                       ck::tensor_layout::gemm::ColumnMajor,
+                                                       ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? K : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            BatchCount);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_reduce);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_batchnorm_bwd.cpp b/3rdparty/composable_kernel/profiler/src/profile_batchnorm_bwd.cpp
new file mode 100644
index 0000000..44ce735
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_batchnorm_bwd.cpp
@@ -0,0 +1,207 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <getopt.h>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "profiler/profile_batchnorm_backward_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+using namespace std;
+
+static const struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                             {"reduceDims", required_argument, nullptr, 'R'},
+                                             {"dumpout", required_argument, nullptr, 'o'},
+                                             {"verify", required_argument, nullptr, 'v'},
+                                             {"help", no_argument, nullptr, '?'},
+                                             {nullptr, 0, nullptr, 0}};
+
+class BatchnormBwdArgParser
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths;
+    std::vector<int> reduceDims;
+
+    bool do_verification = false;
+    bool do_dumpout      = false;
+
+    bool haveSavedMeanInvVar;
+
+    int data_type    = 0;
+    int init_method  = 2;
+    bool time_kernel = false;
+
+    BatchnormBwdArgParser()  = default;
+    ~BatchnormBwdArgParser() = default;
+
+    void show_usage(const char* cmd)
+    {
+        // clang-format off
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of dimensions to reduce on" << std::endl;  
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2 -- 1/0 to indicate whether to use saved mean and invVariance" << std::endl;
+        std::cout << "Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
+        std::cout << "Arg4 -- time kernel (0=no, 1=yes)" << std::endl;
+        // clang-format on
+    };
+
+    int operator()(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        optind++; // to skip the module name
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:o:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case 'o':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_dumpout = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return -1;
+                };
+                break;
+
+            default:
+                show_usage(argv[0]);
+                std::cerr << "Invalid cmd-line options!" << std::endl;
+                return -1;
+            };
+        };
+
+        if(optind + 4 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type           = std::atoi(argv[optind++]);
+        haveSavedMeanInvVar = std::atoi(argv[optind++]);
+        init_method         = std::atoi(argv[optind++]);
+        time_kernel         = static_cast<bool>(std::atoi(argv[optind++]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return -1;
+
+        return 0;
+    };
+}; // end of class AppArgs
+
+static const double epsilon = std::numeric_limits<float>::epsilon();
+
+int profile_batchnorm_backward(int argc, char* argv[])
+{
+    using ck::profiler::profile_batchnorm_backward_impl;
+
+    BatchnormBwdArgParser arg_parser;
+
+    if(arg_parser(argc, argv) != 0)
+        return -1;
+
+    using F16  = ck::half_t;
+    using F32  = float;
+    using BF16 = ck::bhalf_t;
+    using F64  = double;
+
+    if(arg_parser.data_type == 0)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<F16, F32, F32, F32, F16, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 1)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<F32, F32, F32, F32, F32, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 5)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<BF16, F32, F32, F32, BF16, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 6)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<F64, F64, F64, F64, F64, F64, F64, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("bnorm_bwd", "Batchnorm backward", profile_batchnorm_backward);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_batchnorm_fwd.cpp b/3rdparty/composable_kernel/profiler/src/profile_batchnorm_fwd.cpp
new file mode 100644
index 0000000..902a1fc
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_batchnorm_fwd.cpp
@@ -0,0 +1,219 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <getopt.h>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "profiler/profile_batchnorm_forward_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+using namespace std;
+
+static const struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                             {"reduceDims", required_argument, nullptr, 'R'},
+                                             {"dumpout", required_argument, nullptr, 'o'},
+                                             {"verify", required_argument, nullptr, 'v'},
+                                             {"help", no_argument, nullptr, '?'},
+                                             {nullptr, 0, nullptr, 0}};
+
+class BatchnormFwdArgParser
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths;
+    std::vector<int> reduceDims;
+
+    bool do_verification = false;
+    bool do_dumpout      = false;
+
+    bool updateMovingAverage;
+    bool saveMeanAndInvVariance;
+
+    int data_type    = 0;
+    int init_method  = 2;
+    bool time_kernel = false;
+
+    BatchnormFwdArgParser()  = default;
+    ~BatchnormFwdArgParser() = default;
+
+    void show_usage(const char* cmd)
+    {
+        // clang-format off
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of dimensions to reduce on" << std::endl;  
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2: 1/0 to indicate whether to update the moving average and variance (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg3: 1/0 to indicate whether to save the calculated mean and invVariance (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg4: init method used for bnScale and bnBias (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
+        std::cout << "Arg5: time kernel (0=no, 1=yes)" << std::endl;
+        // clang-format on
+    };
+
+    int operator()(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        optind++; // to skip the module name
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:o:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case 'o':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_dumpout = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return -1;
+                };
+                break;
+
+            default:
+                show_usage(argv[0]);
+                std::cerr << "Invalid cmd-line options!" << std::endl;
+                return -1;
+            };
+        };
+
+        if(optind + 5 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type              = std::atoi(argv[optind++]);
+        updateMovingAverage    = std::atoi(argv[optind++]);
+        saveMeanAndInvVariance = std::atoi(argv[optind++]);
+        init_method            = std::atoi(argv[optind++]);
+        time_kernel            = static_cast<bool>(std::atoi(argv[optind++]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return -1;
+
+        return 0;
+    };
+}; // end of class AppArgs
+
+static const double epsilon       = std::numeric_limits<float>::epsilon();
+static const double averageFactor = 0.1;
+
+int profile_batchnorm_forward(int argc, char* argv[])
+{
+    using ck::profiler::profile_batchnorm_forward_impl;
+
+    BatchnormFwdArgParser arg_parser;
+
+    if(arg_parser(argc, argv) != 0)
+        return -1;
+
+    using F16  = ck::half_t;
+    using F32  = float;
+    using BF16 = ck::bhalf_t;
+    using F64  = double;
+
+    if(arg_parser.data_type == 0)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_forward_impl<F16, F16, F32, F16, F16, F16, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.updateMovingAverage,
+                arg_parser.saveMeanAndInvVariance,
+                epsilon,
+                averageFactor);
+        };
+    }
+    else if(arg_parser.data_type == 1)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_forward_impl<F32, F32, F32, F32, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.updateMovingAverage,
+                arg_parser.saveMeanAndInvVariance,
+                epsilon,
+                averageFactor);
+        };
+    }
+    else if(arg_parser.data_type == 5)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_forward_impl<BF16, BF16, F32, BF16, BF16, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.updateMovingAverage,
+                arg_parser.saveMeanAndInvVariance,
+                epsilon,
+                averageFactor);
+        };
+    }
+    else if(arg_parser.data_type == 6)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_forward_impl<F64, F64, F64, F64, F64, F64, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.updateMovingAverage,
+                arg_parser.saveMeanAndInvVariance,
+                epsilon,
+                averageFactor);
+        };
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("bnorm_fwd", "Batchnorm forward", profile_batchnorm_forward);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_conv_bwd_data.cpp b/3rdparty/composable_kernel/profiler/src/profile_conv_bwd_data.cpp
new file mode 100644
index 0000000..9241ead
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_conv_bwd_data.cpp
@@ -0,0 +1,190 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_conv_bwd_data_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    NCHW_KCYX_NKHW, // 0
+    NHWC_KYXC_NHWK, // 1
+};
+
+enum struct ConvDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "conv_bwd_data"
+#define OP_DESC "Convolution Backward Data"
+
+static void print_helper_msg()
+{
+    std::cout
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+        << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
+        << "                 1: Input fp16, Weight fp16, Output fp16\n"
+        << "                 2: Input bf16, Weight bf16, Output bf16\n"
+        << "                 3: Input int8, Weight int8, Output int8)\n"
+        << "arg3: tensor layout (0: Input[N, C, Hi, Wi], Weight[K, C, Y, X], Output[N, K, Ho, Wo]\n"
+        << "                     1: Input[N, Hi, Wi, C], Weight[K, Y, X, C], Output[N, Ho, Wo, "
+           "K])\n"
+        << "arg4: verification (0: no, 1: yes)\n"
+        << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+        << "arg6: print tensor value (0: no; 1: yes)\n"
+        << "arg7: time kernel (0: no, 1: yes)\n"
+        << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+} // namespace
+
+int profile_conv_bwd_data(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using INT8 = int8_t;
+
+    using NWC   = ck::tensor_layout::convolution::NWC;
+    using NHWC  = ck::tensor_layout::convolution::NHWC;
+    using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+    using KXC   = ck::tensor_layout::convolution::KXC;
+    using KYXC  = ck::tensor_layout::convolution::KYXC;
+    using KZYXC = ck::tensor_layout::convolution::KZYXC;
+
+    using NWK   = ck::tensor_layout::convolution::NWK;
+    using NHWK  = ck::tensor_layout::convolution::NHWK;
+    using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+    constexpr auto I1 = ck::Number<1>{};
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto in_layout,
+                       auto wei_layout,
+                       auto out_layout,
+                       auto in_type,
+                       auto wei_type,
+                       auto out_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        using InDataType  = decltype(in_type);
+        using WeiDataType = decltype(wei_type);
+        using OutDataType = decltype(out_type);
+
+        bool pass = ck::profiler::profile_conv_bwd_data_impl<NDimSpatial,
+                                                             InLayout,
+                                                             WeiLayout,
+                                                             OutLayout,
+                                                             InDataType,
+                                                             WeiDataType,
+                                                             OutDataType>(
+            do_verification, init_method, do_log, time_kernel, params);
+
+        return pass ? 0 : 1;
+    };
+
+    if(num_dim_spatial == 1 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_bwd_data);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_conv_fwd.cpp b/3rdparty/composable_kernel/profiler/src/profile_conv_fwd.cpp
new file mode 100644
index 0000000..b57ee7f
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_conv_fwd.cpp
@@ -0,0 +1,192 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_conv_fwd_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    NCHW_KCYX_NKHW, // 0
+    NHWC_KYXC_NHWK, // 1
+};
+
+enum struct ConvDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "conv_fwd"
+#define OP_DESC "Convolution Forward"
+
+static void print_helper_msg()
+{
+    std::cout
+        // clang-format-off
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+        << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
+        << "                 1: Input fp16, Weight fp16, Output fp16\n"
+        << "                 2: Input bf16, Weight bf16, Output bf16\n"
+        << "                 3: Input int8, Weight int8, Output int8)\n"
+        << "arg3: tensor layout (0: Input[N, C, Hi, Wi], Weight[K, C, Y, X], Output[N, K, Ho, Wo]\n"
+        << "                     1: Input[N, Hi, Wi, C], Weight[K, Y, X, C], Output[N, Ho, Wo, "
+           "K])\n"
+        << "arg4: verification (0: no, 1: yes)\n"
+        << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+        << "arg6: print tensor value (0: no; 1: yes)\n"
+        << "arg7: time kernel (0: no, 1: yes)\n"
+        << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+    // clang-format-on
+}
+
+} // namespace
+
+int profile_conv_fwd(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using INT8 = int8_t;
+
+    using NWC   = ck::tensor_layout::convolution::NWC;
+    using NHWC  = ck::tensor_layout::convolution::NHWC;
+    using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+    using KXC   = ck::tensor_layout::convolution::KXC;
+    using KYXC  = ck::tensor_layout::convolution::KYXC;
+    using KZYXC = ck::tensor_layout::convolution::KZYXC;
+
+    using NWK   = ck::tensor_layout::convolution::NWK;
+    using NHWK  = ck::tensor_layout::convolution::NHWK;
+    using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+    constexpr auto I1 = ck::Number<1>{};
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto in_layout,
+                       auto wei_layout,
+                       auto out_layout,
+                       auto in_type,
+                       auto wei_type,
+                       auto out_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        using InDataType  = decltype(in_type);
+        using WeiDataType = decltype(wei_type);
+        using OutDataType = decltype(out_type);
+
+        bool pass = ck::profiler::profile_conv_fwd_impl<NDimSpatial,
+                                                        InLayout,
+                                                        WeiLayout,
+                                                        OutLayout,
+                                                        InDataType,
+                                                        WeiDataType,
+                                                        OutDataType>(
+            do_verification, init_method, do_log, time_kernel, params);
+
+        return pass ? 0 : 1;
+    };
+
+    if(num_dim_spatial == 1 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_conv_fwd_bias_relu.cpp b/3rdparty/composable_kernel/profiler/src/profile_conv_fwd_bias_relu.cpp
new file mode 100644
index 0000000..b44007c
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_conv_fwd_bias_relu.cpp
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_conv_fwd_bias_relu_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct ConvDataType
+{
+    F32_F32_F32, // 0
+    F16_F16_F16, // 1
+};
+
+enum struct ConvInputLayout
+{
+    NCHW, // 0
+    NHWC, // 1
+};
+
+enum struct ConvWeightLayout
+{
+    KCYX, // 0
+    KYXC, // 1
+};
+
+enum struct ConvOutputLayout
+{
+    NKHW, // 0
+    NHWK, // 1
+};
+
+#define OP_NAME "conv_fwd_bias_relu"
+#define OP_DESC "Convolution Forward+Bias+ReLU"
+
+int profile_conv_fwd_bias_relu(int argc, char* argv[])
+{
+    if(argc != 25)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
+        printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
+        printf("arg5: output tensor layout (0: NKHW; 1: NHWK)\n");
+        printf("arg6: verification (0: no; 1: yes)\n");
+        printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg8: print tensor value (0: no; 1: yes)\n");
+        printf("arg9: time kernel (0=n0, 1=yes)\n");
+        printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto in_layout       = static_cast<ConvInputLayout>(std::stoi(argv[3]));
+    const auto wei_layout      = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
+    const auto out_layout      = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
+    const bool do_verification = std::stoi(argv[6]);
+    const int init_method      = std::stoi(argv[7]);
+    const bool do_log          = std::stoi(argv[8]);
+    const bool time_kernel     = std::stoi(argv[9]);
+
+    const ck::index_t N  = std::stoi(argv[10]);
+    const ck::index_t K  = std::stoi(argv[11]);
+    const ck::index_t C  = std::stoi(argv[12]);
+    const ck::index_t Y  = std::stoi(argv[13]);
+    const ck::index_t X  = std::stoi(argv[14]);
+    const ck::index_t Hi = std::stoi(argv[15]);
+    const ck::index_t Wi = std::stoi(argv[16]);
+
+    const ck::index_t conv_stride_h   = std::stoi(argv[17]);
+    const ck::index_t conv_stride_w   = std::stoi(argv[18]);
+    const ck::index_t conv_dilation_h = std::stoi(argv[19]);
+    const ck::index_t conv_dilation_w = std::stoi(argv[20]);
+    const ck::index_t in_left_pad_h   = std::stoi(argv[21]);
+    const ck::index_t in_left_pad_w   = std::stoi(argv[22]);
+    const ck::index_t in_right_pad_h  = std::stoi(argv[23]);
+    const ck::index_t in_right_pad_w  = std::stoi(argv[24]);
+
+    const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
+    const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
+
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+
+    if(data_type == ConvDataType::F16_F16_F16 && in_layout == ConvInputLayout::NHWC &&
+       wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
+    {
+        ck::profiler::profile_conv_fwd_bias_relu_impl<2,
+                                                      ck::half_t,
+                                                      ck::half_t,
+                                                      ck::half_t,
+                                                      ck::tensor_layout::convolution::NHWC,
+                                                      ck::tensor_layout::convolution::KYXC,
+                                                      ck::tensor_layout::convolution::NHWK>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            N,
+            K,
+            C,
+            std::vector<ck::index_t>{Hi, Wi},
+            std::vector<ck::index_t>{Y, X},
+            std::vector<ck::index_t>{Ho, Wo},
+            std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
+            std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
+            std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
+            std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
+    }
+    else
+    {
+        throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd_bias_relu);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_conv_fwd_bias_relu_add.cpp b/3rdparty/composable_kernel/profiler/src/profile_conv_fwd_bias_relu_add.cpp
new file mode 100644
index 0000000..408dd02
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_conv_fwd_bias_relu_add.cpp
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_conv_fwd_bias_relu_add_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct ConvDataType
+{
+    F32_F32_F32, // 0
+    F16_F16_F16, // 1
+};
+
+enum struct ConvInputLayout
+{
+    NCHW, // 0
+    NHWC, // 1
+};
+
+enum struct ConvWeightLayout
+{
+    KCYX, // 0
+    KYXC, // 1
+};
+
+enum struct ConvOutputLayout
+{
+    NKHW, // 0
+    NHWK, // 1
+};
+
+#define OP_NAME "conv_fwd_bias_relu_add"
+#define OP_DESC "Convolution Forward+Bias+ReLU+Add"
+
+int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
+{
+    if(argc != 25)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
+        printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
+        printf("arg5: output tensor layout (0: NKHW; 1: NHWK)\n");
+        printf("arg6: verification (0: no; 1: yes)\n");
+        printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg8: print tensor value (0: no; 1: yes)\n");
+        printf("arg9: time kernel (0=n0, 1=yes)\n");
+        printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto in_layout       = static_cast<ConvInputLayout>(std::stoi(argv[3]));
+    const auto wei_layout      = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
+    const auto out_layout      = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
+    const bool do_verification = std::stoi(argv[6]);
+    const int init_method      = std::stoi(argv[7]);
+    const bool do_log          = std::stoi(argv[8]);
+    const bool time_kernel     = std::stoi(argv[9]);
+
+    const ck::index_t N  = std::stoi(argv[10]);
+    const ck::index_t K  = std::stoi(argv[11]);
+    const ck::index_t C  = std::stoi(argv[12]);
+    const ck::index_t Y  = std::stoi(argv[13]);
+    const ck::index_t X  = std::stoi(argv[14]);
+    const ck::index_t Hi = std::stoi(argv[15]);
+    const ck::index_t Wi = std::stoi(argv[16]);
+
+    const ck::index_t conv_stride_h   = std::stoi(argv[17]);
+    const ck::index_t conv_stride_w   = std::stoi(argv[18]);
+    const ck::index_t conv_dilation_h = std::stoi(argv[19]);
+    const ck::index_t conv_dilation_w = std::stoi(argv[20]);
+    const ck::index_t in_left_pad_h   = std::stoi(argv[21]);
+    const ck::index_t in_left_pad_w   = std::stoi(argv[22]);
+    const ck::index_t in_right_pad_h  = std::stoi(argv[23]);
+    const ck::index_t in_right_pad_w  = std::stoi(argv[24]);
+
+    const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
+    const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
+
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+
+    if(data_type == ConvDataType::F16_F16_F16 && in_layout == ConvInputLayout::NHWC &&
+       wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
+    {
+        ck::profiler::profile_conv_fwd_bias_relu_add_impl<2,
+                                                          ck::half_t,
+                                                          ck::half_t,
+                                                          ck::half_t,
+                                                          ck::tensor_layout::convolution::NHWC,
+                                                          ck::tensor_layout::convolution::KYXC,
+                                                          ck::tensor_layout::convolution::NHWK>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            N,
+            K,
+            C,
+            std::vector<ck::index_t>{Hi, Wi},
+            std::vector<ck::index_t>{Y, X},
+            std::vector<ck::index_t>{Ho, Wo},
+            std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
+            std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
+            std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
+            std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
+    }
+    else
+    {
+        throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd_bias_relu_add);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_gemm.cpp b/3rdparty/composable_kernel/profiler/src/profile_gemm.cpp
new file mode 100644
index 0000000..61bae6a
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_gemm.cpp
@@ -0,0 +1,192 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "gemm"
+#define OP_DESC "GEMM"
+
+static void print_helper_msg()
+{
+    std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+              << "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
+              << "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"
+              << "                     1: A[m, k] * B[n, k] = C[m, n];\n"
+              << "                     2: A[k, m] * B[k, n] = C[m, n];\n"
+              << "                     3: A[k, m] * B[n, k] = C[m, n])\n"
+              << "arg4: verification (0: no; 1: yes)\n"
+              << "arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg6: print tensor value (0: no; 1: yes)\n"
+              << "arg7: time kernel (0: no, 1: yes)\n"
+              << "arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n"
+              << std::endl;
+}
+
+int profile_gemm(int argc, char* argv[])
+{
+    if(argc != 14)
+    {
+        print_helper_msg();
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+
+    using F32   = float;
+    using F16   = ck::half_t;
+    using BF16  = ck::bhalf_t;
+    using INT8  = int8_t;
+    using INT32 = int32_t;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_layout,
+                       auto b_layout,
+                       auto c_layout,
+                       auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto c_type) {
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using CLayout = decltype(c_layout);
+
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using CDataType   = decltype(c_type);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
+
+        bool pass =
+            ck::profiler::profile_gemm_impl<ALayout,
+                                            BLayout,
+                                            CLayout,
+                                            ADataType,
+                                            BDataType,
+                                            AccDataType,
+                                            CDataType>(do_verification,
+                                                       init_method,
+                                                       do_log,
+                                                       time_kernel,
+                                                       M,
+                                                       N,
+                                                       K,
+                                                       (StrideA < 0) ? DefaultStrideA : StrideA,
+                                                       (StrideB < 0) ? DefaultStrideB : StrideB,
+                                                       (StrideC < 0) ? DefaultStrideC : StrideC);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(Row{}, Row{}, Row{}, F32{}, F32{}, F32{}, F32{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(Row{}, Col{}, Row{}, F32{}, F32{}, F32{}, F32{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(Col{}, Row{}, Row{}, F32{}, F32{}, F32{}, F32{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(Col{}, Col{}, Row{}, F32{}, F32{}, F32{}, F32{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(Row{}, Row{}, Row{}, F16{}, F16{}, F32{}, F16{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(Row{}, Col{}, Row{}, F16{}, F16{}, F32{}, F16{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(Col{}, Row{}, Row{}, F16{}, F16{}, F32{}, F16{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(Col{}, Col{}, Row{}, F16{}, F16{}, F32{}, F16{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(Row{}, Row{}, Row{}, BF16{}, BF16{}, F32{}, BF16{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(Row{}, Col{}, Row{}, BF16{}, BF16{}, F32{}, BF16{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(Col{}, Row{}, Row{}, BF16{}, BF16{}, F32{}, BF16{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(Col{}, Col{}, Row{}, BF16{}, BF16{}, F32{}, BF16{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(Row{}, Row{}, Row{}, INT8{}, INT8{}, INT32{}, INT8{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(Row{}, Col{}, Row{}, INT8{}, INT8{}, INT32{}, INT8{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(Col{}, Row{}, Row{}, INT8{}, INT8{}, INT32{}, INT8{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(Col{}, Col{}, Row{}, INT8{}, INT8{}, INT32{}, INT8{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_gemm_add_add_fastgelu.cpp b/3rdparty/composable_kernel/profiler/src/profile_gemm_add_add_fastgelu.cpp
new file mode 100644
index 0000000..c3c0fb7
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_gemm_add_add_fastgelu.cpp
@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_add_add_fastgelu_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_add_add_fastgelu"
+#define OP_DESC "GEMM+Add+Add+FastGeLU"
+
+int profile_gemm_add_add_fastgelu(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN_MN_MN, // 0
+        MK_NK_MN_MN_MN, // 1
+        KM_KN_MN_MN_MN, // 2
+        KM_NK_MN_MN_MN, // 3
+    };
+
+    enum struct MatrixDataType
+    {
+        F32_F32_F32_F32_F32,      // 0
+        F16_F16_F16_F16_F16,      // 1
+        BF16_BF16_BF16_BF16_BF16, // 2
+        INT8_INT8_INT8_INT8_INT8, // 3
+    };
+
+    if(argc != 16)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\n");
+        printf("                     1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
+        printf("                     2: E[m, n] = FastGeLU(A[k, m] * B[k, n] + D0[m, n] + D1[m, n]);\n");
+        printf("                     3: E[m, n] = FastGeLU(A[k, m] * B[n, k] + D0[m, n] + D1[m, n]))\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA  = std::stoi(argv[11]);
+    const int StrideB  = std::stoi(argv[12]);
+    const int StrideD0 = std::stoi(argv[13]);
+    const int StrideD1 = std::stoi(argv[14]);
+    const int StrideE  = std::stoi(argv[15]);
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto d0_type,
+                       auto d1_type,
+                       auto e_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto d0_layout,
+                       auto d1_layout,
+                       auto e_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using D0DataType  = decltype(d0_type);
+        using D1DataType  = decltype(d1_type);
+        using EDataType   = decltype(e_type);
+
+        using ALayout  = decltype(a_layout);
+        using BLayout  = decltype(b_layout);
+        using D0Layout = decltype(d0_layout);
+        using D1Layout = decltype(d1_layout);
+        using ELayout  = decltype(e_layout);
+
+        const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB  = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
+        const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
+        const int DefaultStrideE  = ck::is_same_v<ELayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_add_add_fastgelu_impl<ADataType,
+                                                                     BDataType,
+                                                                     AccDataType,
+                                                                     D0DataType,
+                                                                     D1DataType,
+                                                                     EDataType,
+                                                                     ALayout,
+                                                                     BLayout,
+                                                                     D0Layout,
+                                                                     D1Layout,
+                                                                     ELayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
+            (StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
+            (StrideE < 0) ? DefaultStrideE : StrideE);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
+            layout == MatrixLayout::MK_NK_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
+            layout == MatrixLayout::KM_KN_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
+            layout == MatrixLayout::KM_NK_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_add_fastgelu);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_gemm_add_fastgelu.cpp b/3rdparty/composable_kernel/profiler/src/profile_gemm_add_fastgelu.cpp
new file mode 100644
index 0000000..380b25a
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_gemm_add_fastgelu.cpp
@@ -0,0 +1,146 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_add_fastgelu_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_add_fastgelu"
+#define OP_DESC "GEMM+Add+FastGeLU"
+
+int profile_gemm_add_fastgelu(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN_MN, // 0
+        MK_NK_MN_MN, // 1
+        KM_KN_MN_MN, // 2
+        KM_NK_MN_MN, // 3
+    };
+
+    enum struct MatrixDataType
+    {
+        F32_F32_F32_F32,     // 0
+        F16_F16_F16_F16,     // 1
+        BF16_BF16_BF16_BF16, // 2
+        INT8_INT8_INT8_INT8, // 3
+    };
+
+    if(argc != 15)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n]);\n");
+        printf("                     1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n]);\n");
+        printf("                     2: E[m, n] = FastGeLU(A[k, m] * B[k, n] + D0[m, n]);\n");
+        printf("                     3: E[m, n] = FastGeLU(A[k, m] * B[n, k] + D0[m, n]))\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideD0, StrideE\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA  = std::stoi(argv[11]);
+    const int StrideB  = std::stoi(argv[12]);
+    const int StrideD0 = std::stoi(argv[13]);
+    const int StrideE  = std::stoi(argv[14]);
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto d0_type,
+                       auto e_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto d0_layout,
+                       auto e_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using D0DataType  = decltype(d0_type);
+        using EDataType   = decltype(e_type);
+
+        using ALayout  = decltype(a_layout);
+        using BLayout  = decltype(b_layout);
+        using D0Layout = decltype(d0_layout);
+        using ELayout  = decltype(e_layout);
+
+        const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB  = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
+        const int DefaultStrideE  = ck::is_same_v<ELayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_add_fastgelu_impl<ADataType,
+                                                                 BDataType,
+                                                                 AccDataType,
+                                                                 D0DataType,
+                                                                 EDataType,
+                                                                 ALayout,
+                                                                 BLayout,
+                                                                 D0Layout,
+                                                                 ELayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
+            (StrideE < 0) ? DefaultStrideE : StrideE);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_NK_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_KN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_NK_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_fastgelu);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_gemm_bias_add_reduce.cpp b/3rdparty/composable_kernel/profiler/src/profile_gemm_bias_add_reduce.cpp
new file mode 100644
index 0000000..6d86db0
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_gemm_bias_add_reduce.cpp
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_bias_add_reduce_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_bias_add_reduce"
+#define OP_DESC "GEMM+Bias+Add+Reduce"
+
+int profile_gemm_bias_add_reduce(int argc, char* argv[])
+{
+    enum struct GemmMatrixLayout
+    {
+        MK_KN_MN, // 0
+        MK_NK_MN, // 1
+        KM_KN_MN, // 2
+        KM_NK_MN, // 3
+    };
+
+    enum struct GemmReduceDataType
+    {
+        F32_F32_F32_F32_F32_F32_F32, // 0
+        F16_F16_F16_F16_F16_F32_F32, // 1
+    };
+
+    if(!(argc == 14 || argc == 15))
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, StrideC1\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmReduceDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA  = std::stoi(argv[11]);
+    const int StrideB  = std::stoi(argv[12]);
+    const int StrideC  = std::stoi(argv[13]);
+    const int StrideC1 = std::stoi(argv[14]);
+
+    if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
+       layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        float,
+                                                        ck::tensor_layout::gemm::RowMajor,
+                                                        ck::tensor_layout::gemm::RowMajor,
+                                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            (StrideC1 < 0) ? N : StrideC1);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        float,
+                                                        ck::tensor_layout::gemm::RowMajor,
+                                                        ck::tensor_layout::gemm::ColumnMajor,
+                                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            (StrideC1 < 0) ? N : StrideC1);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        float,
+                                                        ck::tensor_layout::gemm::ColumnMajor,
+                                                        ck::tensor_layout::gemm::RowMajor,
+                                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            (StrideC1 < 0) ? N : StrideC1);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        float,
+                                                        ck::tensor_layout::gemm::ColumnMajor,
+                                                        ck::tensor_layout::gemm::ColumnMajor,
+                                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            (StrideC1 < 0) ? N : StrideC1);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_bias_add_reduce);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_gemm_bilinear.cpp b/3rdparty/composable_kernel/profiler/src/profile_gemm_bilinear.cpp
new file mode 100644
index 0000000..3480014
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_gemm_bilinear.cpp
@@ -0,0 +1,152 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_bilinear_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_bilinear"
+#define OP_DESC "GEMM+Bilinear"
+
+int profile_gemm_bilinear(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN_MN, // 0
+        MK_NK_MN_MN, // 1
+        KM_KN_MN_MN, // 2
+        KM_NK_MN_MN, // 3
+    };
+
+    enum struct MatrixDataType
+    {
+        F32_F32_F32_F32,     // 0
+        F16_F16_F16_F16,     // 1
+        BF16_BF16_BF16_BF16, // 2
+        INT8_INT8_INT8_INT8, // 3
+    };
+
+    if(argc != 17)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: E[m, n] = alpha * A[m, k] * B[k, n] + beta * D[m, n];\n");
+        printf("                     1: E[m, n] = alpha * A[m, k] * B[n, k] + beta * D[m, n];\n");
+        printf("                     2: E[m, n] = alpha * A[k, m] * B[k, n] + beta * D[m, n];\n");
+        printf("                     3: E[m, n] = alpha * A[k, m] * B[n, k] + beta * D[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideD, StrideE\n");
+        printf("arg15 to 16: alhpa, beta\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideD = std::stoi(argv[13]);
+    const int StrideE = std::stoi(argv[14]);
+
+    const float alpha = std::stof(argv[15]);
+    const float beta  = std::stof(argv[16]);
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto d_type,
+                       auto e_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto d_layout,
+                       auto e_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using DDataType   = decltype(d_type);
+        using EDataType   = decltype(e_type);
+
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using DLayout = decltype(d_layout);
+        using ELayout = decltype(e_layout);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideD = ck::is_same_v<DLayout, Row> ? N : M;
+        const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_bilinear_impl<ADataType,
+                                                             BDataType,
+                                                             AccDataType,
+                                                             DDataType,
+                                                             EDataType,
+                                                             ALayout,
+                                                             BLayout,
+                                                             DLayout,
+                                                             ELayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideD < 0) ? DefaultStrideD : StrideD,
+            (StrideE < 0) ? DefaultStrideE : StrideE,
+            alpha,
+            beta);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_NK_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_KN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_NK_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_bilinear);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_gemm_fastgelu.cpp b/3rdparty/composable_kernel/profiler/src/profile_gemm_fastgelu.cpp
new file mode 100644
index 0000000..2a13722
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_gemm_fastgelu.cpp
@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_fastgelu_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_fastgelu"
+#define OP_DESC "GEMM+FastGeLU"
+
+int profile_gemm_fastgelu(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN, // 0
+        MK_NK_MN, // 1
+        KM_KN_MN, // 2
+        KM_NK_MN, // 3
+    };
+
+    enum struct MatrixDataType
+    {
+        F32_F32_F32,    // 0
+        F16_F16_F16,    // 1
+        BF16_BF16_BF16, // 2
+        INT8_INT8_INT8, // 3
+    };
+
+    if(argc != 14)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n]);\n");
+        printf("                     1: E[m, n] = FastGeLU(A[m, k] * B[n, k]);\n");
+        printf("                     2: E[m, n] = FastGeLU(A[k, m] * B[k, n]);\n");
+        printf("                     3: E[m, n] = FastGeLU(A[k, m] * B[n, k]))\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideE\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideE = std::stoi(argv[13]);
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto e_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto e_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using EDataType   = decltype(e_type);
+
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using ELayout = decltype(e_layout);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_fastgelu_impl<ADataType,
+                                                             BDataType,
+                                                             AccDataType,
+                                                             EDataType,
+                                                             ALayout,
+                                                             BLayout,
+                                                             ELayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideE < 0) ? DefaultStrideE : StrideE);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16_F16_F16 && layout == MatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16 && layout == MatrixLayout::MK_NK_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16 && layout == MatrixLayout::KM_KN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16 && layout == MatrixLayout::KM_NK_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_fastgelu);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_gemm_reduce.cpp b/3rdparty/composable_kernel/profiler/src/profile_gemm_reduce.cpp
new file mode 100644
index 0000000..395bf06
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_gemm_reduce.cpp
@@ -0,0 +1,154 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_reduce_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_reduce"
+#define OP_DESC "GEMM+Reduce"
+
+int profile_gemm_reduce(int argc, char* argv[])
+{
+    enum struct GemmMatrixLayout
+    {
+        MK_KN_MN, // 0
+        MK_NK_MN, // 1
+        KM_KN_MN, // 2
+        KM_NK_MN, // 3
+    };
+
+    enum struct GemmReduceDataType
+    {
+        F32_F32_F32_F32_F32, // 0
+        F16_F16_F16_F32_F32, // 1
+    };
+
+    if(!(argc == 14 || argc == 15))
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
+        printf("arg14: split k into  mulitiple batch\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmReduceDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+
+    if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_gemm_reduce_impl<ck::half_t,
+                                               ck::half_t,
+                                               ck::half_t,
+                                               float,
+                                               ck::tensor_layout::gemm::RowMajor,
+                                               ck::tensor_layout::gemm::RowMajor,
+                                               ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? K : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_gemm_reduce_impl<ck::half_t,
+                                               ck::half_t,
+                                               ck::half_t,
+                                               float,
+                                               ck::tensor_layout::gemm::RowMajor,
+                                               ck::tensor_layout::gemm::ColumnMajor,
+                                               ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? K : StrideA,
+            (StrideB < 0) ? K : StrideB,
+            (StrideC < 0) ? N : StrideC);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        ck::profiler::profile_gemm_reduce_impl<ck::half_t,
+                                               ck::half_t,
+                                               ck::half_t,
+                                               float,
+                                               ck::tensor_layout::gemm::ColumnMajor,
+                                               ck::tensor_layout::gemm::RowMajor,
+                                               ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        ck::profiler::profile_gemm_reduce_impl<ck::half_t,
+                                               ck::half_t,
+                                               ck::half_t,
+                                               float,
+                                               ck::tensor_layout::gemm::ColumnMajor,
+                                               ck::tensor_layout::gemm::ColumnMajor,
+                                               ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? K : StrideB,
+            (StrideC < 0) ? N : StrideC);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_reduce);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_gemm_splitk.cpp b/3rdparty/composable_kernel/profiler/src/profile_gemm_splitk.cpp
new file mode 100644
index 0000000..f636ce7
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_gemm_splitk.cpp
@@ -0,0 +1,154 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_splitk_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "gemm_splitk"
+#define OP_DESC "Split-K GEMM"
+
+int profile_gemm_splitk(int argc, char* argv[])
+{
+    if(argc != 15)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
+        printf("arg14: split k into  mulitiple batch\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+    const int KBatch  = std::stoi(argv[14]);
+
+    using F32 = float;
+    using F16 = ck::half_t;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto c_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto c_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using CDataType   = decltype(c_type);
+
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using CLayout = decltype(c_layout);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_splitk_impl<ADataType,
+                                                           BDataType,
+                                                           AccDataType,
+                                                           CDataType,
+                                                           ALayout,
+                                                           BLayout,
+                                                           CLayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideC < 0) ? DefaultStrideC : StrideC,
+            KBatch);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_splitk);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_grouped_conv_bwd_weight.cpp b/3rdparty/composable_kernel/profiler/src/profile_grouped_conv_bwd_weight.cpp
new file mode 100644
index 0000000..dfd8a09
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_grouped_conv_bwd_weight.cpp
@@ -0,0 +1,182 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <initializer_list>
+#include <iostream>
+#include <numeric>
+
+#include "profiler/profile_grouped_conv_bwd_weight_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    GNCHW_GKCYX_GNKHW, // 0
+    GNHWC_GKYXC_GNHWK, // 1
+};
+
+enum struct ConvDataType
+{
+    F32_F32_F32,   // 0
+    F16_F16_F16,   // 1
+    BF16_F32_BF16, // 2
+};
+
+#define OP_NAME "grouped_conv_bwd_weight"
+#define OP_DESC "Grouped Convolution Backward Weight"
+
+static void print_helper_msg()
+{
+    std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+              << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
+              << "                 1: Input fp16, Weight fp16, Output fp16\n"
+              << "                 2: Input bf16, Weight fp32, Output bf16)\n"
+              << "arg3: tensor layout (0: Input[G, N, C, Hi, Wi], Weight[G, K, C, Y, X], Output[G, "
+                 "N, K, Ho, Wo]\n"
+              << "                     1: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, "
+                 "N, Ho, Wo, K]\n"
+              << "arg4: verification (0: no, 1: yes)\n"
+              << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+              << "arg6: print tensor value (0: no; 1: yes)\n"
+              << "arg7: time kernel (0: no, 1: yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << " SplitK\n"
+              << std::endl;
+}
+
+} // namespace
+
+int profile_grouped_conv_bwd_weight(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial, 1 for split-K
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial + 1)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    ck::index_t split_k = std::stoi(argv[8 + 1 + 4 + 6 * num_dim_spatial]);
+    split_k             = std::max(1, split_k);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+
+    using GNWC   = ck::tensor_layout::convolution::GNWC;
+    using GNHWC  = ck::tensor_layout::convolution::GNHWC;
+    using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
+
+    using GKXC   = ck::tensor_layout::convolution::GKXC;
+    using GKYXC  = ck::tensor_layout::convolution::GKYXC;
+    using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
+
+    using GNWK   = ck::tensor_layout::convolution::GNWK;
+    using GNHWK  = ck::tensor_layout::convolution::GNHWK;
+    using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
+
+    constexpr auto I1 = ck::Number<1>{};
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto in_layout,
+                       auto wei_layout,
+                       auto out_layout,
+                       auto in_type,
+                       auto wei_type,
+                       auto out_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        using InDataType  = decltype(in_type);
+        using WeiDataType = decltype(wei_type);
+        using OutDataType = decltype(out_type);
+
+        bool pass = ck::profiler::profile_grouped_conv_bwd_weight_impl<NDimSpatial,
+                                                                       InLayout,
+                                                                       WeiLayout,
+                                                                       OutLayout,
+                                                                       InDataType,
+                                                                       WeiDataType,
+                                                                       OutDataType>(
+            do_verification, init_method, do_log, time_kernel, params, split_k);
+
+        return pass ? 0 : 1;
+    };
+
+    if(num_dim_spatial == 1 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_F32_BF16)
+        {
+            // fp32 atomic add is used for weight tensor in bf16 kernel
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, BF16{}, F32{}, BF16{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_F32_BF16)
+        {
+            // fp32 atomic add is used for weight tensor in bf16 kernel
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, BF16{}, F32{}, BF16{});
+        }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_F32_BF16)
+        {
+            // fp32 atomic add is used for weight tensor in bf16 kernel
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, BF16{}, F32{}, BF16{});
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_conv_bwd_weight);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_grouped_conv_fwd.cpp b/3rdparty/composable_kernel/profiler/src/profile_grouped_conv_fwd.cpp
new file mode 100644
index 0000000..9ff3c15
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_grouped_conv_fwd.cpp
@@ -0,0 +1,260 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_grouped_conv_fwd_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    GNHWC_GKYXC_GNHWK, // 0
+    NHWGC_GKYXC_NHWGK, // 1
+};
+
+enum struct ConvDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "grouped_conv_fwd"
+#define OP_DESC "Grouped Convolution Forward"
+
+static void print_helper_msg()
+{
+    std::cout
+        // clang-format off
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+        << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
+        << "                 1: Input fp16, Weight fp16, Output fp16\n"
+        << "                 2: Input bf16, Weight bf16, Output bf16\n"
+        << "                 3: Input int8, Weight int8, Output int8)\n"
+        << "arg3: tensor layout (0: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]\n"
+        << "                     1: Input[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Output[N, Ho, Wo, G, K])\n"
+        << "arg4: verification (0: no, 1: yes)\n"
+        << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+        << "arg6: print tensor value (0: no; 1: yes)\n"
+        << "arg7: time kernel (0: no, 1: yes)\n"
+        << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+    // clang-format on
+}
+
+} // namespace
+
+int profile_grouped_conv_fwd(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using INT8 = int8_t;
+
+    //
+    using GNWC   = ck::tensor_layout::convolution::GNWC;
+    using GNHWC  = ck::tensor_layout::convolution::GNHWC;
+    using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
+
+    using GKXC   = ck::tensor_layout::convolution::GKXC;
+    using GKYXC  = ck::tensor_layout::convolution::GKYXC;
+    using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
+
+    using GNWK   = ck::tensor_layout::convolution::GNWK;
+    using GNHWK  = ck::tensor_layout::convolution::GNHWK;
+    using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
+
+    //
+    using NWGC   = ck::tensor_layout::convolution::NWGC;
+    using NHWGC  = ck::tensor_layout::convolution::NHWGC;
+    using NDHWGC = ck::tensor_layout::convolution::NDHWGC;
+
+    using NWGK   = ck::tensor_layout::convolution::NWGK;
+    using NHWGK  = ck::tensor_layout::convolution::NHWGK;
+    using NDHWGK = ck::tensor_layout::convolution::NDHWGK;
+
+    constexpr auto I1 = ck::Number<1>{};
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto in_layout,
+                       auto wei_layout,
+                       auto out_layout,
+                       auto in_type,
+                       auto wei_type,
+                       auto out_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        using InDataType  = decltype(in_type);
+        using WeiDataType = decltype(wei_type);
+        using OutDataType = decltype(out_type);
+
+        bool pass = ck::profiler::profile_grouped_conv_fwd_impl<NDimSpatial,
+                                                                InLayout,
+                                                                WeiLayout,
+                                                                OutLayout,
+                                                                InDataType,
+                                                                WeiDataType,
+                                                                OutDataType>(
+            do_verification, init_method, do_log, time_kernel, params);
+
+        return pass ? 0 : 1;
+    };
+
+    // GNHWC_GKYXC_GNHWK
+    if(num_dim_spatial == 1 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    // NHWGC_GKYXC_NHWGK
+    else if(num_dim_spatial == 1 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, NWGC{}, GKXC{}, NWGK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, NWGC{}, GKXC{}, NWGK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I1, NWGC{}, GKXC{}, NWGK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I1, NWGC{}, GKXC{}, NWGK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_conv_fwd);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_grouped_gemm.cpp b/3rdparty/composable_kernel/profiler/src/profile_grouped_gemm.cpp
new file mode 100644
index 0000000..65e24bd
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_grouped_gemm.cpp
@@ -0,0 +1,169 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_grouped_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+    MK_KN_NM, // 4
+    MK_NK_NM, // 5
+    KM_KN_NM, // 6
+    KM_NK_NM, // 7
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+std::vector<int> argToIntArray(char* input)
+{
+    std::vector<int> out;
+
+    std::istringstream in(input);
+
+    std::string item;
+
+    while(std::getline(in, item, ','))
+    {
+        out.push_back(std::stoi(item));
+    }
+
+    return out;
+}
+
+#define OP_NAME "grouped_gemm"
+#define OP_DESC "Grouped GEMM"
+
+int profile_grouped_gemm(int argc, char* argv[])
+{
+    if(!(argc == 14))
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 13: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
+               "64,64 64,64 128,128)\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const auto Ms = argToIntArray(argv[8]);
+    const auto Ns = argToIntArray(argv[9]);
+    const auto Ks = argToIntArray(argv[10]);
+
+    const auto StrideAs = argToIntArray(argv[11]);
+    const auto StrideBs = argToIntArray(argv[12]);
+    const auto StrideCs = argToIntArray(argv[13]);
+
+    if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_grouped_gemm_impl<ck::half_t,
+                                                ck::half_t,
+                                                ck::half_t,
+                                                float,
+                                                ck::tensor_layout::gemm::RowMajor,
+                                                ck::tensor_layout::gemm::RowMajor,
+                                                ck::tensor_layout::gemm::RowMajor>(do_verification,
+                                                                                   init_method,
+                                                                                   do_log,
+                                                                                   time_kernel,
+                                                                                   Ms,
+                                                                                   Ns,
+                                                                                   Ks,
+                                                                                   StrideAs,
+                                                                                   StrideBs,
+                                                                                   StrideCs);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_grouped_gemm_impl<ck::half_t,
+                                                ck::half_t,
+                                                ck::half_t,
+                                                float,
+                                                ck::tensor_layout::gemm::RowMajor,
+                                                ck::tensor_layout::gemm::ColumnMajor,
+                                                ck::tensor_layout::gemm::RowMajor>(do_verification,
+                                                                                   init_method,
+                                                                                   do_log,
+                                                                                   time_kernel,
+                                                                                   Ms,
+                                                                                   Ns,
+                                                                                   Ks,
+                                                                                   StrideAs,
+                                                                                   StrideBs,
+                                                                                   StrideCs);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        ck::profiler::profile_grouped_gemm_impl<ck::half_t,
+                                                ck::half_t,
+                                                ck::half_t,
+                                                float,
+                                                ck::tensor_layout::gemm::ColumnMajor,
+                                                ck::tensor_layout::gemm::RowMajor,
+                                                ck::tensor_layout::gemm::RowMajor>(do_verification,
+                                                                                   init_method,
+                                                                                   do_log,
+                                                                                   time_kernel,
+                                                                                   Ms,
+                                                                                   Ns,
+                                                                                   Ks,
+                                                                                   StrideAs,
+                                                                                   StrideBs,
+                                                                                   StrideCs);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        ck::profiler::profile_grouped_gemm_impl<ck::half_t,
+                                                ck::half_t,
+                                                ck::half_t,
+                                                float,
+                                                ck::tensor_layout::gemm::ColumnMajor,
+                                                ck::tensor_layout::gemm::ColumnMajor,
+                                                ck::tensor_layout::gemm::RowMajor>(do_verification,
+                                                                                   init_method,
+                                                                                   do_log,
+                                                                                   time_kernel,
+                                                                                   Ms,
+                                                                                   Ns,
+                                                                                   Ks,
+                                                                                   StrideAs,
+                                                                                   StrideBs,
+                                                                                   StrideCs);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_gemm);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_groupnorm.cpp b/3rdparty/composable_kernel/profiler/src/profile_groupnorm.cpp
new file mode 100644
index 0000000..2741f52
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_groupnorm.cpp
@@ -0,0 +1,112 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_groupnorm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+struct GroupnormArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+#define OP_NAME "groupnorm"
+#define OP_DESC "Group Normalization"
+
+void print_help_groupnorm()
+{
+    std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+              << "arg2: data type (0: fp16; 1: fp32)\n"
+              << "arg3: verification (0: no; 1: yes)\n"
+              << "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg5: print tensor value (0: no; 1: yes)\n"
+              << "arg6: time kernel (0=no, 1=yes)\n"
+              << "--length: tensor extents (e.g, --length 1 16 16 32 40) \n"
+              << std::endl;
+}
+
+int profile_groupnorm(int argc, char* argv[])
+{
+    ck::DataTypeEnum data_type  = ck::DataTypeEnum::Half;
+    bool do_verification        = false;
+    int init_method             = 0;
+    bool do_log                 = 0;
+    bool time_kernel            = 1;
+    std::vector<index_t> length = {64, 16, 16, 32, 40};
+
+    if(argc != 1 && argc != 13)
+    {
+        print_help_groupnorm();
+        return 0;
+    }
+
+    if(argc == 13)
+    {
+        data_type       = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
+        do_verification = std::stoi(argv[3]);
+        init_method     = std::stoi(argv[4]);
+        do_log          = std::stoi(argv[5]);
+        time_kernel     = std::stoi(argv[6]);
+
+        // parse the long options
+        GroupnormArgParser arg_parser;
+        arg_parser(argc, argv);
+        length = arg_parser.long_opts["length"];
+    }
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    if(data_type == ck::DataTypeEnum::Float)
+    {
+        ck::profiler::profile_groupnorm_impl<F32, F32, F32, F32, F32>(
+            do_verification, init_method, do_log, time_kernel, length);
+    }
+    else if(data_type == ck::DataTypeEnum::Half)
+    {
+        ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16>(
+            do_verification, init_method, do_log, time_kernel, length);
+    }
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_groupnorm);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_layernorm.cpp b/3rdparty/composable_kernel/profiler/src/profile_layernorm.cpp
new file mode 100644
index 0000000..e93fc2d
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_layernorm.cpp
@@ -0,0 +1,101 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_layernorm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+struct LayernormArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+void print_help_layernorm()
+{
+    std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
+              << "arg2: verification (0: no; 1: yes)\n"
+              << "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg4: print tensor value (0: no; 1: yes)\n"
+              << "arg5: time kernel (0=no, 1=yes)\n"
+              << "--length: tensor extents (e.g, --length 1024 1024) \n"
+              << std::endl;
+}
+
+int profile_layernorm(int argc, char* argv[])
+{
+    if(argc <= 2)
+    {
+        print_help_layernorm();
+        return 0;
+    }
+
+    LayernormArgParser arg_parser;
+
+    // short unnamed options
+    const ck::DataTypeEnum data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
+    const bool do_verification       = std::stoi(argv[3]);
+    const int init_method            = std::stoi(argv[4]);
+    const bool do_log                = std::stoi(argv[5]);
+    const bool time_kernel           = std::stoi(argv[6]);
+
+    // parse the long options
+    arg_parser(argc, argv);
+    const std::vector<index_t> length = arg_parser.long_opts["length"];
+
+    using F16          = ck::half_t;
+    using F32          = float;
+    constexpr int rank = 2;
+
+    if(data_type == ck::DataTypeEnum::Half)
+    {
+        ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, rank>(
+            do_verification, init_method, do_log, time_kernel, length);
+    }
+    else if(data_type == ck::DataTypeEnum::Float)
+    {
+        ck::profiler::profile_layernorm_impl<F32, F32, F32, F32, F32, rank>(
+            do_verification, init_method, do_log, time_kernel, length);
+    }
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("layernorm", "Layer Normalization", profile_layernorm);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_reduce.cpp b/3rdparty/composable_kernel/profiler/src/profile_reduce.cpp
new file mode 100644
index 0000000..6925371
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_reduce.cpp
@@ -0,0 +1,434 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <fstream>
+#include <cstdlib>
+#include <vector>
+#include <stdexcept>
+#include <sstream>
+#include <getopt.h>
+
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/utility/host_common_util.hpp"
+
+#include "profiler/profile_reduce_impl.hpp"
+#include "profiler/data_type_enum.hpp"
+#include "profiler_operation_registry.hpp"
+
+using namespace std;
+
+using ck::ReduceTensorOp;
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"reduceDims", required_argument, nullptr, 'R'},
+                                       {"reduceOp", required_argument, nullptr, 'O'},
+                                       {"compType", required_argument, nullptr, 'C'},
+                                       {"outType", required_argument, nullptr, 'W'},
+                                       {"nanOpt", required_argument, nullptr, 'N'},
+                                       {"indicesOpt", required_argument, nullptr, 'I'},
+                                       {"scales", required_argument, nullptr, 'S'},
+                                       {"half", no_argument, nullptr, '?'},
+                                       {"double", no_argument, nullptr, '?'},
+                                       {"int8", no_argument, nullptr, '?'},
+                                       {"bf16", no_argument, nullptr, '?'},
+                                       {"dumpout", required_argument, nullptr, 'o'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+static void check_reduce_dims(const int rank, const std::vector<int>& reduceDims)
+{
+    for(auto dim : reduceDims)
+    {
+        if(dim < 0 || dim >= rank)
+            throw std::runtime_error("Invalid dimension index specified for Reducing");
+    };
+
+    unsigned int flag = 0;
+
+    for(auto dim : reduceDims)
+    {
+        if(flag & (0x1 << dim))
+            throw std::runtime_error("All toReduce dimensions should be different!");
+        flag = flag | (0x1 << dim);
+    };
+};
+
+class ReduceProfilerArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    bool use_half   = false;
+    bool use_double = false;
+    bool use_int8   = false;
+    bool use_bf16   = false;
+
+    std::vector<size_t> inLengths;
+    std::vector<size_t> outLengths;
+    std::vector<int> reduceDims;
+
+    std::vector<float> scales;
+
+    ReduceTensorOp reduceOp     = ReduceTensorOp::ADD;
+    ck::DataTypeEnum compTypeId = ck::DataTypeEnum::Float;
+    ck::DataTypeEnum outTypeId  = ck::DataTypeEnum::Float;
+
+    bool compType_assigned = false;
+    bool outType_assigned  = false;
+
+    int nanOpt           = 0;
+    int indicesOpt       = 0;
+    bool do_verification = false;
+    bool do_dumpout      = false;
+
+    int init_method;
+    bool time_kernel;
+
+    ReduceProfilerArgs()  = default;
+    ~ReduceProfilerArgs() = default;
+
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of to-reduce dimensions"
+                  << std::endl;
+        std::cout << "--reduceOp or -O, enum value indicating the reduction operations"
+                  << std::endl;
+        std::cout << "--compType or -C, enum value indicating the type of accumulated values used "
+                     "during the reduction"
+                  << std::endl;
+        std::cout << "--outType or -W, optional enum value indicating the type of the reduced "
+                     "output, which could be float when the input data is half"
+                  << std::endl;
+        std::cout
+            << "--nanOpt or -N, 1/0 value indicates the selection to use or not use Nan-Propagation"
+            << std::endl;
+        std::cout << "--indicesOpt or -I, 1/0 value indicates the selection to use or not use "
+                     "index in reduction"
+                  << std::endl;
+        std::cout << "--scales or -S, comma separated two float values for alpha and beta"
+                  << std::endl;
+        std::cout << "--half, use fp16 for the input and output tensor data types" << std::endl;
+        std::cout << "--double, use fp64 for the input and output tensor data types" << std::endl;
+        std::cout << "--int8, use int8 for the input and output tensor data types" << std::endl;
+        std::cout << "--bf16, use bfloat16 for the input and output tensor data types" << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "--dumpout or -o, 1/0 to indicate where to save the reduction result to files "
+                     "for further analysis"
+                  << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        optind++; // to skip the "reduce" module name
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:O:C:W:N:I:S:v:o:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'O':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceOp = static_cast<ReduceTensorOp>(std::atoi(optarg));
+                break;
+            case 'C':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                compTypeId        = static_cast<ck::DataTypeEnum>(std::atoi(optarg));
+                compType_assigned = true;
+                break;
+            case 'W':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                outTypeId        = static_cast<ck::DataTypeEnum>(std::atoi(optarg));
+                outType_assigned = true;
+                break;
+            case 'N':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                nanOpt = std::atoi(optarg);
+                break;
+            case 'I':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                indicesOpt = std::atoi(optarg);
+                break;
+            case 'S':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                scales = getTypeValuesFromString<float>(optarg);
+
+                if(scales.size() != 2)
+                    throw std::runtime_error("Invalid option format!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case 'o':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_dumpout = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "half")
+                    use_half = true;
+                else if(std::string(long_options[option_index].name) == "double")
+                    use_double = true;
+                else if(std::string(long_options[option_index].name) == "int8")
+                    use_int8 = true;
+                else if(std::string(long_options[option_index].name) == "bf16")
+                    use_bf16 = true;
+                else if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+
+            default:
+                show_usage(argv[0]);
+                std::cerr << "Invalid cmd-line options!" << std::endl;
+                return (-1);
+            };
+        };
+
+        if(optind + 2 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        if(reduceOp == ReduceTensorOp::MIN || reduceOp == ReduceTensorOp::MAX ||
+           reduceOp == ReduceTensorOp::AMAX)
+        {
+            // for indexable operations, no need to assign compType and outType, just let them be
+            // same as inType
+            compType_assigned = false;
+            outType_assigned  = false;
+        };
+
+        return (0);
+    };
+
+}; // end of class AppArgs
+
+int profile_reduce(int argc, char* argv[])
+{
+    using ck::DataTypeEnum;
+    using ck::profiler::profile_reduce_impl;
+
+    ReduceProfilerArgs args;
+
+    if(args.processArgs(argc, argv) < 0)
+        return (-1);
+
+    int rank = args.inLengths.size();
+
+    check_reduce_dims(rank, args.reduceDims);
+
+    if(args.reduceOp == ReduceTensorOp::MUL || args.reduceOp == ReduceTensorOp::NORM1)
+        throw std::runtime_error("MUL and NORM1 are not supported by composable kernel!");
+
+    if(args.use_half)
+    {
+        if(!args.compType_assigned)
+            args.compTypeId = DataTypeEnum::Half;
+
+        if(args.outType_assigned &&
+           (args.outTypeId != DataTypeEnum::Half && args.outTypeId != DataTypeEnum::Float))
+            args.outTypeId = DataTypeEnum::Float;
+
+        if(!args.outType_assigned)
+            args.outTypeId = DataTypeEnum::Half;
+
+        if(args.compTypeId == DataTypeEnum::Half)
+        {
+            profile_reduce_impl<ck::half_t, ck::half_t, ck::half_t>(
+                args.do_verification,
+                args.init_method,
+                args.do_dumpout,
+                args.time_kernel,
+                args.inLengths,
+                args.reduceDims,
+                args.reduceOp,
+                static_cast<bool>(args.nanOpt),
+                static_cast<bool>(args.indicesOpt),
+                args.scales[0],
+                args.scales[1]);
+        }
+        else if(args.compTypeId == DataTypeEnum::Float)
+        {
+            profile_reduce_impl<ck::half_t, float, ck::half_t>(args.do_verification,
+                                                               args.init_method,
+                                                               args.do_dumpout,
+                                                               args.time_kernel,
+                                                               args.inLengths,
+                                                               args.reduceDims,
+                                                               args.reduceOp,
+                                                               static_cast<bool>(args.nanOpt),
+                                                               static_cast<bool>(args.indicesOpt),
+                                                               args.scales[0],
+                                                               args.scales[1]);
+        }
+        else
+            throw std::runtime_error("Invalid compType assignment!");
+    }
+    else if(args.use_double)
+    {
+        profile_reduce_impl<double, double, double>(args.do_verification,
+                                                    args.init_method,
+                                                    args.do_dumpout,
+                                                    args.time_kernel,
+                                                    args.inLengths,
+                                                    args.reduceDims,
+                                                    args.reduceOp,
+                                                    static_cast<bool>(args.nanOpt),
+                                                    static_cast<bool>(args.indicesOpt),
+                                                    args.scales[0],
+                                                    args.scales[1]);
+    }
+    else if(args.use_int8)
+    {
+        if(!args.compType_assigned)
+            args.compTypeId = DataTypeEnum::Int8;
+
+        if(args.outType_assigned &&
+           (args.outTypeId != DataTypeEnum::Int8 && args.outTypeId != DataTypeEnum::Int32))
+            args.outTypeId = DataTypeEnum::Int32;
+
+        if(!args.outType_assigned)
+            args.outTypeId = DataTypeEnum::Int8;
+
+        if(args.compTypeId == DataTypeEnum::Int8)
+        {
+            profile_reduce_impl<int8_t, int8_t, int8_t>(args.do_verification,
+                                                        args.init_method,
+                                                        args.do_dumpout,
+                                                        args.time_kernel,
+                                                        args.inLengths,
+                                                        args.reduceDims,
+                                                        args.reduceOp,
+                                                        static_cast<bool>(args.nanOpt),
+                                                        static_cast<bool>(args.indicesOpt),
+                                                        args.scales[0],
+                                                        args.scales[1]);
+        }
+        else if(args.compTypeId == DataTypeEnum::Int32)
+        {
+            profile_reduce_impl<int8_t, int32_t, int8_t>(args.do_verification,
+                                                         args.init_method,
+                                                         args.do_dumpout,
+                                                         args.time_kernel,
+                                                         args.inLengths,
+                                                         args.reduceDims,
+                                                         args.reduceOp,
+                                                         static_cast<bool>(args.nanOpt),
+                                                         static_cast<bool>(args.indicesOpt),
+                                                         args.scales[0],
+                                                         args.scales[1]);
+        }
+        else
+            throw std::runtime_error("Invalid compType assignment!");
+    }
+    else if(args.use_bf16)
+    {
+        if(args.outType_assigned &&
+           (args.outTypeId != DataTypeEnum::BFloat16 && args.outTypeId != DataTypeEnum::Float))
+            args.outTypeId = DataTypeEnum::Float;
+
+        if(!args.outType_assigned)
+            args.outTypeId = DataTypeEnum::BFloat16;
+
+        profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(args.do_verification,
+                                                             args.init_method,
+                                                             args.do_dumpout,
+                                                             args.time_kernel,
+                                                             args.inLengths,
+                                                             args.reduceDims,
+                                                             args.reduceOp,
+                                                             static_cast<bool>(args.nanOpt),
+                                                             static_cast<bool>(args.indicesOpt),
+                                                             args.scales[0],
+                                                             args.scales[1]);
+    }
+    else
+    {
+        if(args.compTypeId == DataTypeEnum::Float)
+        {
+            profile_reduce_impl<float, float, float>(args.do_verification,
+                                                     args.init_method,
+                                                     args.do_dumpout,
+                                                     args.time_kernel,
+                                                     args.inLengths,
+                                                     args.reduceDims,
+                                                     args.reduceOp,
+                                                     static_cast<bool>(args.nanOpt),
+                                                     static_cast<bool>(args.indicesOpt),
+                                                     args.scales[0],
+                                                     args.scales[1]);
+        }
+        else if(args.compTypeId == DataTypeEnum::Double)
+        {
+            profile_reduce_impl<float, double, float>(args.do_verification,
+                                                      args.init_method,
+                                                      args.do_dumpout,
+                                                      args.time_kernel,
+                                                      args.inLengths,
+                                                      args.reduceDims,
+                                                      args.reduceOp,
+                                                      static_cast<bool>(args.nanOpt),
+                                                      static_cast<bool>(args.indicesOpt),
+                                                      args.scales[0],
+                                                      args.scales[1]);
+        }
+        else
+            throw std::runtime_error("Invalid compType assignment!");
+    };
+
+    return (0);
+};
+
+REGISTER_PROFILER_OPERATION("reduce", "Reduce", profile_reduce);
diff --git a/3rdparty/composable_kernel/profiler/src/profile_softmax.cpp b/3rdparty/composable_kernel/profiler/src/profile_softmax.cpp
new file mode 100644
index 0000000..30f627d
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profile_softmax.cpp
@@ -0,0 +1,169 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/profile_softmax_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+using ck::profiler::SoftmaxDataType;
+
+struct ArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {
+        {"length", {}}, {"stride", {}}, {"reduce", {}}, {"alpha", {}}, {"beta", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+void print_help()
+{
+    std::cout << "arg1: tensor operation (softmax)\n"
+              << "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
+              << "arg3: verification (0: no; 1: yes)\n"
+              << "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg5: print tensor value (0: no; 1: yes)\n"
+              << "arg6: time kernel (0=n0, 1=yes)\n"
+              << "--length: tensor extents (e.g, --length 8 4 256) \n"
+              << "--stride: tensor strides (e.g, --stride 1024 256 1)\n"
+              << "--reduce: to-reduce dimensions (e.g, --reduce 2)\n"
+              << "--alpha: alpha scaling value\n"
+              << "--beta: beta scaling value\n"
+              << std::endl;
+}
+
+int profile_softmax(int argc, char* argv[])
+{
+    if(argc <= 2)
+    {
+        print_help();
+        return 0;
+    }
+
+    ArgParser arg_parser;
+
+    // short unnamed options
+    const SoftmaxDataType data_type = static_cast<SoftmaxDataType>(std::stoi(argv[2]));
+    const bool do_verification      = std::stoi(argv[3]);
+    const int init_method           = std::stoi(argv[4]);
+    const bool do_log               = std::stoi(argv[5]);
+    const bool time_kernel          = std::stoi(argv[6]);
+
+    // parse the long options
+    arg_parser(argc, argv);
+    const std::vector<index_t> length = arg_parser.long_opts["length"];
+    const std::vector<index_t> stride = arg_parser.long_opts["stride"];
+    const std::vector<index_t> reduce = arg_parser.long_opts["reduce"];
+    const index_t alpha =
+        arg_parser.long_opts["alpha"].empty() ? 1 : arg_parser.long_opts["alpha"][0];
+    const index_t beta = arg_parser.long_opts["beta"].empty() ? 0 : arg_parser.long_opts["beta"][0];
+
+    // Rank 3
+    if(length.size() == 3)
+    {
+        if(data_type == SoftmaxDataType::F16_F16)
+        {
+            ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 3>(do_verification,
+                                                                                 init_method,
+                                                                                 do_log,
+                                                                                 time_kernel,
+                                                                                 length,
+                                                                                 stride,
+                                                                                 reduce,
+                                                                                 float(alpha),
+                                                                                 float(beta));
+        }
+        else if(data_type == SoftmaxDataType::F32_F32)
+        {
+            ck::profiler::profile_softmax_impl<float, float, float, 3>(do_verification,
+                                                                       init_method,
+                                                                       do_log,
+                                                                       time_kernel,
+                                                                       length,
+                                                                       stride,
+                                                                       reduce,
+                                                                       float(alpha),
+                                                                       float(beta));
+        }
+        else
+        {
+            throw std::runtime_error("not implemented yet");
+        }
+    }
+    // Rank 4
+    else if(length.size() == 4)
+    {
+        if(data_type == SoftmaxDataType::F16_F16)
+        {
+            ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 4>(do_verification,
+                                                                                 init_method,
+                                                                                 do_log,
+                                                                                 time_kernel,
+                                                                                 length,
+                                                                                 stride,
+                                                                                 reduce,
+                                                                                 float(alpha),
+                                                                                 float(beta));
+        }
+        else if(data_type == SoftmaxDataType::F32_F32)
+        {
+            ck::profiler::profile_softmax_impl<float, float, float, 4>(do_verification,
+                                                                       init_method,
+                                                                       do_log,
+                                                                       time_kernel,
+                                                                       length,
+                                                                       stride,
+                                                                       reduce,
+                                                                       float(alpha),
+                                                                       float(beta));
+        }
+        else
+        {
+            throw std::runtime_error("not implemented yet");
+        }
+    }
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+// hijack main() for quick debugging
+// int main(int argc, char* argv[])
+// {
+//     profile_normalization(argc, argv);
+//     return 0;
+// }
+
+REGISTER_PROFILER_OPERATION("softmax", "Softmax", profile_softmax);
diff --git a/3rdparty/composable_kernel/profiler/src/profiler.cpp b/3rdparty/composable_kernel/profiler/src/profiler.cpp
new file mode 100644
index 0000000..080117e
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profiler.cpp
@@ -0,0 +1,30 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+
+#include "profiler_operation_registry.hpp"
+
+static void print_helper_message()
+{
+    std::cout << "arg1: tensor operation " << ProfilerOperationRegistry::GetInstance() << std::endl;
+}
+
+int main(int argc, char* argv[])
+{
+    if(argc == 1)
+    {
+        print_helper_message();
+    }
+    else if(const auto operation = ProfilerOperationRegistry::GetInstance().Get(argv[1]);
+            operation.has_value())
+    {
+        return (*operation)(argc, argv);
+    }
+    else
+    {
+        std::cerr << "cannot find operation: " << argv[1] << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/3rdparty/composable_kernel/profiler/src/profiler_operation_registry.hpp b/3rdparty/composable_kernel/profiler/src/profiler_operation_registry.hpp
new file mode 100644
index 0000000..91ff291
--- /dev/null
+++ b/3rdparty/composable_kernel/profiler/src/profiler_operation_registry.hpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <map>
+#include <optional>
+#include <string_view>
+#include <utility>
+
+class ProfilerOperationRegistry final
+{
+    ProfilerOperationRegistry()  = default;
+    ~ProfilerOperationRegistry() = default;
+
+    public:
+    using Operation = std::function<int(int, char*[])>;
+
+    private:
+    struct Entry final
+    {
+        explicit Entry(std::string_view description, Operation operation) noexcept
+            : description_(description), operation_(std::move(operation))
+        {
+        }
+
+        std::string_view description_;
+        Operation operation_;
+    };
+
+    std::map<std::string_view, Entry> entries_;
+
+    friend std::ostream& operator<<(std::ostream& stream, const ProfilerOperationRegistry& registry)
+    {
+        stream << "{\n";
+        for(auto& [name, entry] : registry.entries_)
+        {
+            stream << "\t" << name << ": " << entry.description_ << "\n";
+        }
+        stream << "}";
+
+        return stream;
+    }
+
+    public:
+    static ProfilerOperationRegistry& GetInstance()
+    {
+        static ProfilerOperationRegistry registry;
+        return registry;
+    }
+
+    std::optional<Operation> Get(std::string_view name) const
+    {
+        const auto found = entries_.find(name);
+        if(found == end(entries_))
+        {
+            return std::nullopt;
+        }
+
+        return (found->second).operation_;
+    }
+
+    bool Add(std::string_view name, std::string_view description, Operation operation)
+    {
+        return entries_
+            .emplace(std::piecewise_construct,
+                     std::forward_as_tuple(name),
+                     std::forward_as_tuple(description, std::move(operation)))
+            .second;
+    }
+};
+
+#define PP_CONCAT(x, y) PP_CONCAT_IMPL(x, y)
+#define PP_CONCAT_IMPL(x, y) x##y
+
+#define REGISTER_PROFILER_OPERATION(name, description, operation)              \
+    static const bool PP_CONCAT(operation_registration_result_, __COUNTER__) = \
+        ::ProfilerOperationRegistry::GetInstance().Add(name, description, operation)
diff --git a/3rdparty/composable_kernel/rbuild.ini b/3rdparty/composable_kernel/rbuild.ini
new file mode 100644
index 0000000..3649ced
--- /dev/null
+++ b/3rdparty/composable_kernel/rbuild.ini
@@ -0,0 +1,8 @@
+[develop]
+cxx = ${rocm_path}/bin/hipcc
+cc = ${rocm_path}/llvm/bin/clang
+ignore = pcre
+deps =
+    -f dev-requirements.txt
+define =
+    BUILD_DEV=On
diff --git a/3rdparty/composable_kernel/requirements.txt b/3rdparty/composable_kernel/requirements.txt
new file mode 100644
index 0000000..8b13789
--- /dev/null
+++ b/3rdparty/composable_kernel/requirements.txt
@@ -0,0 +1 @@
+
diff --git a/3rdparty/composable_kernel/script/clang-format-overwrite.sh b/3rdparty/composable_kernel/script/clang-format-overwrite.sh
new file mode 100644
index 0000000..f9d11fc
--- /dev/null
+++ b/3rdparty/composable_kernel/script/clang-format-overwrite.sh
@@ -0,0 +1,2 @@
+#find . -name deps -prune -o -name build -prune -o -iname '*.h' -o -iname '*.hpp' -o -iname '*.cpp' -o -iname '*.h.in' -o -iname '*.hpp.in' -o -iname '*.cpp.in' -o -iname '*.cl' -o -iname '*.cuh' -o -iname '*.cu' -o -iname '*.inc' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-10 -i -style=file {}'
+git status --porcelain | awk '$1 != "D" && (match($2, "\\.cpp|hpp|inc")) {print $2}' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-10 -i -style=file {}'
diff --git a/3rdparty/composable_kernel/script/cmake-ck-dev.sh b/3rdparty/composable_kernel/script/cmake-ck-dev.sh
new file mode 100644
index 0000000..2e605ce
--- /dev/null
+++ b/3rdparty/composable_kernel/script/cmake-ck-dev.sh
@@ -0,0 +1,19 @@
+#!/bin/bash
+rm -f CMakeCache.txt
+rm -f *.cmake
+rm -rf CMakeFiles
+
+MY_PROJECT_SOURCE=$1
+
+cmake                                                                                             \
+-D CMAKE_PREFIX_PATH=/opt/rocm                                                                    \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
+-D CMAKE_CXX_FLAGS="-O3 -ftemplate-backtrace-limit=0 -gline-tables-only -save-temps=$PWD"         \
+-D CMAKE_BUILD_TYPE=Release                                                                       \
+-D BUILD_DEV=ON                                                                                   \
+-D GPU_TARGETS="gfx908;gfx90a"                                                                    \
+-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON                                                                 \
+-D USE_BITINT_EXTENSION_INT4=OFF                                                                  \
+${MY_PROJECT_SOURCE}
+
+#-D AMDGPU_TARGETS=gfx90a;gfx908
diff --git a/3rdparty/composable_kernel/script/cmake-ck-release.sh b/3rdparty/composable_kernel/script/cmake-ck-release.sh
new file mode 100644
index 0000000..268b1eb
--- /dev/null
+++ b/3rdparty/composable_kernel/script/cmake-ck-release.sh
@@ -0,0 +1,19 @@
+#!/bin/bash
+rm -f CMakeCache.txt
+rm -f *.cmake
+rm -rf CMakeFiles
+
+MY_PROJECT_SOURCE=$1
+
+cmake                                                                                             \
+-D CMAKE_PREFIX_PATH=/opt/rocm                                                                    \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
+-D CMAKE_CXX_FLAGS="-O3"                                                                          \
+-D CMAKE_BUILD_TYPE=Release                                                                       \
+-D BUILD_DEV=OFF                                                                                  \
+-D GPU_TARGETS="gfx908;gfx90a"                                                                      \
+-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON                                                                 \
+-D USE_BITINT_EXTENSION_INT4=OFF                                                                  \
+${MY_PROJECT_SOURCE}
+
+#-D AMDGPU_TARGETS=gfx90a;gfx908
diff --git a/3rdparty/composable_kernel/script/count_vgpr.sh b/3rdparty/composable_kernel/script/count_vgpr.sh
new file mode 100644
index 0000000..07debc5
--- /dev/null
+++ b/3rdparty/composable_kernel/script/count_vgpr.sh
@@ -0,0 +1,20 @@
+#!/bin/bash
+FILE=$1
+
+for num in {0..255}
+do
+    base_pattern="(\[?${num}\b|\[\d*:${num}\])"
+    spattern="s${base_pattern}"
+    vpattern="v${base_pattern}"
+    apattern="a${base_pattern}"
+    scount=$(grep -P $spattern $FILE | wc -l)
+    vcount=$(grep -P $vpattern $FILE | wc -l)
+    acount=$(grep -P $apattern $FILE | wc -l)
+    bash -c "echo -n v${num}   $vcount && \
+             echo -n , s${num} $scount && \
+             echo -n , a${num} $acount"
+    if [[ $scount -ne 0 || $vcount -ne 0 || $acount -ne 0 ]]; then
+        echo -n "  *"
+    fi
+    echo ""
+done
diff --git a/3rdparty/composable_kernel/script/hipclang_opt.sh b/3rdparty/composable_kernel/script/hipclang_opt.sh
new file mode 100644
index 0000000..c51bd51
--- /dev/null
+++ b/3rdparty/composable_kernel/script/hipclang_opt.sh
@@ -0,0 +1,25 @@
+rm *.ll *.s
+
+BC_FILE=$1
+
+/opt/rocm/llvm/bin/llvm-dis $BC_FILE -o original.ll
+/opt/rocm/llvm/bin/opt -S -inline -inline-threshold=104857 original.ll > inline.ll
+/opt/rocm/llvm/bin/opt -S -sroa inline.ll > sroa.ll
+/opt/rocm/llvm/bin/opt -S -O3 sroa.ll > o3.ll
+
+/opt/rocm/llvm/bin/llc -mcpu=gfx906 original.ll
+/opt/rocm/llvm/bin/llc -mcpu=gfx906 inline.ll
+/opt/rocm/llvm/bin/llc -mcpu=gfx906 sroa.ll
+/opt/rocm/llvm/bin/llc -mcpu=gfx906 o3.ll
+
+#/opt/rocm/llvm/bin/opt -S -O3 -sroa inline.ll > o3.ll
+#/opt/rocm/llvm/bin/opt -S -O3 -sroa o3.ll > o3_2.ll
+#/opt/rocm/llvm/bin/opt -S -O3 -sroa o3_2.ll > o3_3.ll
+#/opt/rocm/llvm/bin/opt -S -O3 -sroa o3_3.ll > o3_4.ll
+
+#/opt/rocm/llvm/bin/llc -mcpu=gfx908 opt.ll
+#/opt/rocm/llvm/bin/llc -mcpu=gfx908 inline.ll
+#/opt/rocm/llvm/bin/llc -mcpu=gfx908 o3.ll
+#/opt/rocm/llvm/bin/llc -mcpu=gfx908 o3_2.ll
+#/opt/rocm/llvm/bin/llc -mcpu=gfx908 o3_3.ll
+#/opt/rocm/llvm/bin/llc -mcpu=gfx908 o3_4.ll
diff --git a/3rdparty/composable_kernel/script/parse_perf_data.py b/3rdparty/composable_kernel/script/parse_perf_data.py
new file mode 100644
index 0000000..4cb13e6
--- /dev/null
+++ b/3rdparty/composable_kernel/script/parse_perf_data.py
@@ -0,0 +1,290 @@
+#!/usr/bin/env python3
+import os, io, argparse, datetime, re
+import numpy as np
+import sqlalchemy
+from sqlalchemy.types import NVARCHAR, Float, Integer
+import pymysql
+import pandas as pd
+from sshtunnel import SSHTunnelForwarder
+
+def print_to_string(*args, **kwargs):
+    output = io.StringIO()
+    print(*args, file=output, **kwargs)
+    contents = output.getvalue()
+    output.close()
+    return contents
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Parse results from tf benchmark runs')
+    parser.add_argument('filename', type=str, help='Log file to prase or directory containing log files')
+    args = parser.parse_args()
+    files = []
+    if os.path.isdir(args.filename):
+        all_files = os.listdir(args.filename)
+        for name in all_files:
+            if not 'log' in name:
+                continue
+            files.append(os.path.join(args.filename, name))
+    else:
+        files = [args.filename]
+    args.files = files
+    return args
+
+def main():
+    args = parse_args()
+    tests = []
+    kernels=[]
+    tflops=[]
+    dtype=[]
+    alayout=[]
+    blayout=[]
+    M=[]
+    N=[]
+    K=[]
+    StrideA=[]
+    StrideB=[]
+    StrideC=[]
+    #parse results, get the Tflops value for "Best Perf" kernels
+
+    glue=""
+    for filename in args.files:
+        for line in open(filename):
+            if 'Branch name' in line:
+                lst=line.split()
+                branch_name=lst[2]
+            if 'On branch' in line:
+                lst=line.split()
+                branch_name=lst[2]
+            if 'Node name' in line:
+                lst=line.split()
+                node_id=lst[2]
+            if 'GPU_arch' in line:
+                lst=line.split()
+                gpu_arch=lst[2]
+            if 'HIP version' in line:
+                lst=line.split()
+                hip_vers=lst[2]
+            if 'Compute Unit' in line:
+                lst=line.split()
+                compute_units=lst[2]
+            if 'InstalledDir' in line:
+                lst=line.split()
+                rocm_vers=lst[1][lst[1].find('/opt/rocm-')+len('/opt/rocm-'):lst[1].rfind('/llvm/bin')]
+    print("Branch name:",branch_name)
+    print("Node name:",node_id)
+    print("GPU_arch:",gpu_arch)
+    print("Compute units:",compute_units)
+    print("ROCM_version:",rocm_vers)
+    print("HIP_version:",hip_vers)
+
+
+    #parse gemm performance tests:
+    if 'gemm' in filename:
+        for filename in args.files:
+            for line in open(filename):
+                if 'Best Perf' in line:
+                    lst=line.split()
+                    if len(lst)>=37: #the line is complete
+                        tests.append(glue.join(lst[5:30]))
+                        kernels.append(glue.join(lst[37:]))
+                        tflops.append(lst[33])
+                        dtype.append(lst[5])
+                        alayout.append(lst[8])
+                        blayout.append(lst[11])
+                        M.append(lst[14])
+                        N.append(lst[17])
+                        K.append(lst[20])
+                        StrideA.append(lst[23])
+                        StrideB.append(lst[26])
+                        StrideC.append(lst[29])
+                    elif len(lst)<37 and len(lst)>=33: #the tflops are available
+                        tests.append(glue.join(lst[5:30]))
+                        kernels.append("N/A")
+                        tflops.append(lst[33])
+                        dtype.append(lst[5])
+                        alayout.append(lst[8])
+                        blayout.append(lst[11])
+                        M.append(lst[14])
+                        N.append(lst[17])
+                        K.append(lst[20])
+                        StrideA.append(lst[23])
+                        StrideB.append(lst[26])
+                        StrideC.append(lst[29])
+                        print("warning: incomplete line:",lst)
+                    elif len(lst)<33: #even the tflops are not available
+                        print("Error in ckProfiler output!")
+                        print("warning: incomplete line=",lst)
+        #sort results
+        #sorted_tests = sorted(tests)
+        #print("sorted tests:",sorted_tests)
+        sorted_tflops = [x for _,x in sorted(zip(tests,tflops))]
+        #sorted_kernels = [x for _,x in sorted(zip(tests,kernels))]
+        test_list=list(range(1,len(tests)+1))
+
+    #parse resnet50 performance tests:
+    if 'resnet50' in filename:
+        for filename in args.files:
+            for line in open(filename):
+                if 'Best Perf' in line:
+                    lst=line.split()
+                    tflops.append(lst[4])
+
+    print("Number of tests:",len(tflops))
+    sql_hostname = '127.0.0.1'
+    sql_username = os.environ["dbuser"]
+    sql_password = os.environ["dbpassword"]
+    sql_main_database = 'miopen_perf'
+    sql_port = 3306
+    ssh_host = os.environ["dbsship"]
+    ssh_user = os.environ["dbsshuser"]
+    ssh_port = int(os.environ["dbsshport"])
+    ssh_pass = os.environ["dbsshpassword"]
+
+    with SSHTunnelForwarder(
+            (ssh_host, ssh_port),
+            ssh_username=ssh_user,
+            ssh_password=ssh_pass,
+            remote_bind_address=(sql_hostname, sql_port)) as tunnel:
+
+        sqlEngine = sqlalchemy.create_engine('mysql+pymysql://{0}:{1}@{2}:{3}/{4}'.
+            format(sql_username, sql_password, sql_hostname, tunnel.local_bind_port, sql_main_database))
+        conn = sqlEngine.connect()
+
+        #save gemm performance tests:
+        if 'gemm' in filename:
+
+            #write the ck_gemm_test_params table
+            #only needed once the test set changes
+            '''
+            sorted_dtypes = [x for _,x in sorted(zip(tests,dtype))]
+            sorted_alayout = [x for _,x in sorted(zip(tests,alayout))]
+            sorted_blayout = [x for _,x in sorted(zip(tests,blayout))]
+            sorted_M = [x for _,x in sorted(zip(tests,M))]
+            sorted_N = [x for _,x in sorted(zip(tests,N))]
+            sorted_K = [x for _,x in sorted(zip(tests,K))]
+            sorted_StrideA = [x for _,x in sorted(zip(tests,StrideA))]
+            sorted_StrideB = [x for _,x in sorted(zip(tests,StrideB))]
+            sorted_StrideC = [x for _,x in sorted(zip(tests,StrideC))]
+            ck_gemm_params=[test_list,sorted_dtypes,sorted_alayout,sorted_blayout,
+                        sorted_M,sorted_N,sorted_K,sorted_StrideA,sorted_StrideB,
+                        sorted_StrideC]
+            df=pd.DataFrame(np.transpose(ck_gemm_params),columns=['Test_number','Data_type',
+                'Alayout','BLayout','M','N','K', 'StrideA','StrideB','StrideC'])
+            print(df)
+
+            dtypes = {
+                'Test_number': Integer(),
+                'Data_type': NVARCHAR(length=5),
+                'Alayout': NVARCHAR(length=12),
+                'Blayout': NVARCHAR(length=12),
+                'M': Integer(),
+                'N': Integer(),
+                'K': Integer(),
+                'StrideA': Integer(),
+                'StrideB': Integer(),
+                'StrideC': Integer()
+                }
+            df.to_sql("ck_gemm_test_params",conn,if_exists='replace',index=False, dtype=dtypes)
+            '''
+
+            #read baseline results for the latest develop branch
+            query = '''SELECT * from ck_gemm_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_gemm_tflops where Branch_ID='develop' );'''
+            tflops_base = pd.read_sql_query(query, conn)
+
+            #write new results to the db
+            testlist=[]
+            for i in range(1,len(tests)+1):
+                testlist.append("Test%i"%i)
+            ck_gemm_tflops=[str(branch_name),str(node_id),str(gpu_arch),compute_units,str(rocm_vers),str(hip_vers),str(datetime.datetime.now())]
+            flops=pd.DataFrame(data=[ck_gemm_tflops],columns=['Branch_ID','Node_ID','GPU_arch','Compute Units','ROCM_version','HIP_version','Datetime'])
+            df_add=pd.DataFrame(data=[sorted_tflops],columns=testlist)
+            flops=pd.concat([flops,df_add],axis=1)
+            print("new tflops for gemm tests:",flops)
+            flops.to_sql("ck_gemm_tflops",conn,if_exists='append',index=False)
+
+        #save resnet50 performance tests:
+        if 'resnet50' in filename:
+            #read baseline results for the latest develop branch
+            query = '''SELECT * from ck_resnet50_N256_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_resnet50_N256_tflops where Branch_ID='develop' );'''
+            tflops_base_N256 = pd.read_sql_query(query, conn)
+            query = '''SELECT * from ck_resnet50_N4_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_resnet50_N4_tflops where Branch_ID='develop' );'''
+            tflops_base_N4 = pd.read_sql_query(query, conn)
+
+            #write new results to the db
+            testlist=[]
+            for i in range(1,50):
+                testlist.append("Layer%i"%i)
+            ck_resnet_tflops=[str(branch_name),str(node_id),str(gpu_arch),compute_units,str(rocm_vers),str(hip_vers),str(datetime.datetime.now())]
+            flops0=pd.DataFrame(data=[ck_resnet_tflops],columns=['Branch_ID','Node_ID','GPU_arch','Compute Units','ROCM_version','HIP_version','Datetime'])
+            df_add=pd.DataFrame(data=[tflops[0:49]],columns=testlist)
+            flops=pd.concat([flops0,df_add],axis=1)
+            print("new tflops for N=256 resnet50 test:",flops)
+            flops.to_sql("ck_resnet50_N256_tflops",conn,if_exists='append',index=False)
+            df_add=pd.DataFrame(data=[tflops[49:98]],columns=testlist)
+            flops=pd.concat([flops0,df_add],axis=1)
+            print("new tflops for N=4 resnet50 test:",flops)
+            flops.to_sql("ck_resnet50_N4_tflops",conn,if_exists='append',index=False)
+
+        conn.close()
+
+    #compare the results to the baseline if baseline exists
+    regression=0
+    if 'gemm' in filename:
+        if not tflops_base.empty:
+            base=tflops_base[testlist].to_numpy(dtype='float')
+            base_list=base[0]
+            ave_perf=0
+            for i in range(len(base_list)):
+                # success criterion:
+                if base_list[i]>1.01*float(sorted_tflops[i]):
+                    print("test # ",i,"shows regression by {:.3f}%".format(
+                        (float(sorted_tflops[i])-base_list[i])/base_list[i]*100))
+                    regression=1
+                ave_perf=ave_perf+float(sorted_tflops[i])/base_list[i]
+            if regression==0:
+                print("no regressions found")
+            ave_perf=ave_perf/len(base_list)
+            print("average performance relative to baseline:",ave_perf)
+        else:
+            print("could not find a baseline")
+    if 'resnet50' in filename:
+        if not tflops_base_N256.empty:
+            base=tflops_base_N256[testlist].to_numpy(dtype='float')
+            base_list=base[0]
+            ave_perf=0
+            for i in range(len(base_list)):
+                # success criterion:
+                if base_list[i]>1.01*float(tflops[i]):
+                    print("layer # ",i,"shows regression by {:.3f}%".format(
+                        (float(tflops[i])-base_list[i])/base_list[i]*100))
+                    regression=1
+                ave_perf=ave_perf+float(tflops[i])/base_list[i]
+            if regression==0:
+                print("no regressions found")
+            ave_perf=ave_perf/len(base_list)
+            print("average performance relative to baseline:",ave_perf)
+        else:
+            print("could not find a baseline for N=256")
+        if not tflops_base_N4.empty:
+            base=tflops_base_N4[testlist].to_numpy(dtype='float')
+            base_list=base[0]
+            ave_perf=0
+            for i in range(len(base_list)):
+                # success criterion:
+                if base_list[i]>1.01*float(tflops[i+49]):
+                    print("layer # ",i,"shows regression by {:.3f}%".format(
+                        (float(tflops[i+49])-base_list[i])/base_list[i]*100))
+                    regression=1
+                ave_perf=ave_perf+float(tflops[i+49])/base_list[i]
+            if regression==0:
+                print("no regressions found")
+            ave_perf=ave_perf/len(base_list)
+            print("average performance relative to baseline:",ave_perf)
+        else:
+            print("could not find a baseline for N=4")
+
+    #return 0 if performance criteria met, otherwise return 1
+    return regression
+
+if __name__ == '__main__':
+    main()
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/script/process_perf_data.py b/3rdparty/composable_kernel/script/process_perf_data.py
new file mode 100644
index 0000000..638e4ef
--- /dev/null
+++ b/3rdparty/composable_kernel/script/process_perf_data.py
@@ -0,0 +1,308 @@
+#!/usr/bin/env python3
+import os, io, argparse, datetime
+#import numpy as np
+import sqlalchemy
+from sqlalchemy.types import NVARCHAR, Float, Integer
+import pymysql
+import pandas as pd
+from sshtunnel import SSHTunnelForwarder
+
+def print_to_string(*args, **kwargs):
+    output = io.StringIO()
+    print(*args, file=output, **kwargs)
+    contents = output.getvalue()
+    output.close()
+    return contents
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Parse results from tf benchmark runs')
+    parser.add_argument('filename', type=str, help='Log file to prase or directory containing log files')
+    args = parser.parse_args()
+    files = []
+    if os.path.isdir(args.filename):
+        all_files = os.listdir(args.filename)
+        for name in all_files:
+            if not 'log' in name:
+                continue
+            files.append(os.path.join(args.filename, name))
+    else:
+        files = [args.filename]
+    args.files = files
+    return args
+
+def get_log_params(logfile):
+    print("logfile=",logfile)
+    branch_name=' '
+    node_id=' '
+    gpu_arch=' '
+    hip_vers=' '
+    compute_units=0
+    environment=' '
+    rocm_vers=' '
+    for line in open(logfile):
+        if 'Branch name' in line:
+            lst=line.split()
+            branch_name=lst[2]
+        if 'On branch' in line:
+            lst=line.split()
+            branch_name=lst[2]
+        if 'Node name' in line:
+            lst=line.split()
+            node_id=lst[2]
+        if 'GPU_arch' in line:
+            lst=line.split()
+            gpu_arch=lst[2]
+        if 'HIP version' in line:
+            lst=line.split()
+            hip_vers=lst[2]
+        if 'Compute Unit' in line:
+            lst=line.split()
+            compute_units=lst[2]
+        if 'Environment type' in line:
+            lst=line.split()
+            environment=lst[2]
+        if 'InstalledDir' in line:
+            lst=line.split()
+            rocm_vers=lst[1][lst[1].find('/opt/rocm-')+len('/opt/rocm-'):lst[1].rfind('/llvm/bin')]
+    return branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment
+
+def parse_logfile(logfile):
+    glue=''
+    res=[]
+    tests=[]
+    kernels=[]
+    tflops=[]
+    dtype=[]
+    alayout=[]
+    blayout=[]
+    M=[]
+    N=[]
+    K=[]
+    StrideA=[]
+    StrideB=[]
+    StrideC=[]
+    if 'perf_gemm.log' in logfile:
+        for line in open(logfile):
+            if 'Best Perf' in line:
+                lst=line.split()
+                if len(lst)>=37: #the line is complete
+                    tests.append(glue.join(lst[5:30]))
+                    kernels.append(glue.join(lst[37:]))
+                    tflops.append(lst[33])
+                    dtype.append(lst[5])
+                    alayout.append(lst[8])
+                    blayout.append(lst[11])
+                    M.append(lst[14])
+                    N.append(lst[17])
+                    K.append(lst[20])
+                    StrideA.append(lst[23])
+                    StrideB.append(lst[26])
+                    StrideC.append(lst[29])
+                elif len(lst)<37 and len(lst)>=33: #the tflops are available
+                    tests.append(glue.join(lst[5:30]))
+                    kernels.append("N/A")
+                    tflops.append(lst[33])
+                    dtype.append(lst[5])
+                    alayout.append(lst[8])
+                    blayout.append(lst[11])
+                    M.append(lst[14])
+                    N.append(lst[17])
+                    K.append(lst[20])
+                    StrideA.append(lst[23])
+                    StrideB.append(lst[26])
+                    StrideC.append(lst[29])
+                    print("warning: incomplete line:",lst)
+                elif len(lst)<33: #even the tflops are not available
+                    print("Error in ckProfiler output!")
+                    print("warning: incomplete line=",lst)
+        #sort results
+        #sorted_tests = sorted(tests)
+        res = [x for _,x in sorted(zip(tests,tflops))]
+        #sorted_kernels = [x for _,x in sorted(zip(tests,kernels))]
+        test_list=list(range(1,len(tests)+1))
+    #parse conv_fwd and conv_bwd performance tests:
+    elif 'conv_fwd' in logfile or 'conv_bwd_data' in logfile:
+        for line in open(logfile):
+            if 'tflops:' in line:
+                lst=line.split()
+                res.append(lst[1])
+    #parse all other performance tests:
+    elif 'resnet50' in logfile or 'batched_gemm' in logfile or 'grouped_gemm' in logfile  or 'gemm_bilinear' in logfile or 'reduction' in logfile:
+        for line in open(logfile):
+            if 'Best Perf' in line:
+                lst=line.split()
+                res.append(lst[4])
+    elif 'onnx_gemm' in logfile or 'splitK_gemm' in logfile:
+        for line in open(logfile):
+            if 'Best Perf' in line:
+                lst=line.split()
+                res.append(lst[33])
+    return res
+
+
+def get_baseline(table, connection):
+    query = '''SELECT * from '''+table+''' WHERE Datetime = (SELECT MAX(Datetime) FROM '''+table+''' where Branch_ID='develop' );'''
+    return pd.read_sql_query(query, connection)
+
+def store_new_test_result(table_name, test_results, testlist, branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment, connection):
+    params=[str(branch_name),str(node_id),str(gpu_arch),compute_units,str(rocm_vers),str(hip_vers),str(environment),str(datetime.datetime.now())]
+    df=pd.DataFrame(data=[params],columns=['Branch_ID','Node_ID','GPU_arch','Compute Units','ROCM_version','HIP_version','Environment','Datetime'])
+    df_add=pd.DataFrame(data=[test_results],columns=testlist)
+    df=pd.concat([df,df_add],axis=1)
+    #print("new test results dataframe:",df)
+    df.to_sql(table_name,connection,if_exists='append',index=False)
+    return 0
+
+def compare_test_to_baseline(baseline,test,testlist):
+    regression=0
+    if not baseline.empty:
+        base=baseline[testlist].to_numpy(dtype='float')
+        base_list=base[0]
+        ave_perf=0
+        for i in range(len(base_list)):
+            # success criterion:
+            if base_list[i]>1.01*float(test[i]):
+                print("test # ",i,"shows regression by {:.3f}%".format(
+                    (float(test[i])-base_list[i])/base_list[i]*100))
+                regression=1
+            if base_list[i]>0: ave_perf=ave_perf+float(test[i])/base_list[i]
+        if regression==0:
+            print("no regressions found")
+        ave_perf=ave_perf/len(base_list)
+        print("average performance relative to baseline:",ave_perf)
+    else:
+        print("could not find a baseline")
+    return regression
+
+'''
+def post_test_params(tlist,connection):
+    sorted_dtypes = [x for _,x in sorted(zip(tests,dtype))]
+    sorted_alayout = [x for _,x in sorted(zip(tests,alayout))]
+    sorted_blayout = [x for _,x in sorted(zip(tests,blayout))]
+    sorted_M = [x for _,x in sorted(zip(tests,M))]
+    sorted_N = [x for _,x in sorted(zip(tests,N))]
+    sorted_K = [x for _,x in sorted(zip(tests,K))]
+    sorted_StrideA = [x for _,x in sorted(zip(tests,StrideA))]
+    sorted_StrideB = [x for _,x in sorted(zip(tests,StrideB))]
+    sorted_StrideC = [x for _,x in sorted(zip(tests,StrideC))]
+    ck_gemm_params=[tlist,sorted_dtypes,sorted_alayout,sorted_blayout,
+                sorted_M,sorted_N,sorted_K,sorted_StrideA,sorted_StrideB,
+                sorted_StrideC]
+    df=pd.DataFrame(np.transpose(ck_gemm_params),columns=['Test_number','Data_type',
+        'Alayout','BLayout','M','N','K', 'StrideA','StrideB','StrideC'])
+    print(df)
+
+    dtypes = {
+        'Test_number': Integer(),
+        'Data_type': NVARCHAR(length=5),
+        'Alayout': NVARCHAR(length=12),
+        'Blayout': NVARCHAR(length=12),
+        'M': Integer(),
+        'N': Integer(),
+        'K': Integer(),
+        'StrideA': Integer(),
+        'StrideB': Integer(),
+        'StrideC': Integer()
+        }
+    df.to_sql("ck_gemm_test_params",connection,if_exists='replace',index=False, dtype=dtypes)
+'''
+
+def main():
+    args = parse_args()
+    results=[]
+    tflops_base=[]
+    testlist=[]
+    #parse the test parameters from the logfile
+    for filename in args.files:
+        branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment = get_log_params(filename)
+
+    print("Branch name:",branch_name)
+    print("Node name:",node_id)
+    print("GPU_arch:",gpu_arch)
+    print("Compute units:",compute_units)
+    print("ROCM_version:",rocm_vers)
+    print("HIP_version:",hip_vers)
+    print("Environment:",environment)
+    #parse results, get the Tflops value for "Best Perf" kernels
+    results=parse_logfile(filename)
+
+    print("Number of tests:",len(results))
+    sql_hostname = '127.0.0.1'
+    sql_username = os.environ["dbuser"]
+    sql_password = os.environ["dbpassword"]
+    sql_main_database = 'miopen_perf'
+    sql_port = 3306
+    ssh_host = os.environ["dbsship"]
+    ssh_user = os.environ["dbsshuser"]
+    ssh_port = int(os.environ["dbsshport"])
+    ssh_pass = os.environ["dbsshpassword"]
+
+    with SSHTunnelForwarder(
+            (ssh_host, ssh_port),
+            ssh_username=ssh_user,
+            ssh_password=ssh_pass,
+            remote_bind_address=(sql_hostname, sql_port)) as tunnel:
+
+        sqlEngine = sqlalchemy.create_engine('mysql+pymysql://{0}:{1}@{2}:{3}/{4}'.
+            format(sql_username, sql_password, sql_hostname, tunnel.local_bind_port, sql_main_database))
+        conn = sqlEngine.connect()
+
+        #save gemm performance tests:
+        if 'perf_gemm.log' in filename:
+            #write the ck_gemm_test_params table only needed once the test set changes
+            #post_test_params(test_list,conn)
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_gemm_tflops"
+        if 'batched_gemm' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_batched_gemm_tflops"
+        if 'grouped_gemm' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_grouped_gemm_tflops"
+        if 'conv_fwd' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_conv_fwd_tflops"
+        if 'conv_bwd_data' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_conv_bwd_data_tflops"
+        if 'gemm_bilinear' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_gemm_bilinear_tflops"
+        if 'reduction' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_reduction_GBps"
+        if 'resnet50_N4' in filename:
+            for i in range(1,50):
+                testlist.append("Layer%i"%i)
+            table_name="ck_resnet50_N4_tflops"
+        if 'resnet50_N256' in filename:
+            for i in range(1,50):
+                testlist.append("Layer%i"%i)
+            table_name="ck_resnet50_N256_tflops"
+        if 'onnx_gemm' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_onnx_gemm_tflops"
+        if 'splitK_gemm' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_splitK_gemm_tflops"
+
+        tflops_base = get_baseline(table_name,conn)
+        store_new_test_result(table_name, results, testlist, branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment, conn)
+        conn.close()
+
+    #compare the results to the baseline if baseline exists
+    regression=0
+    regression=compare_test_to_baseline(tflops_base,results,testlist)
+    return regression
+
+if __name__ == '__main__':
+    main()
diff --git a/3rdparty/composable_kernel/script/process_perf_data.sh b/3rdparty/composable_kernel/script/process_perf_data.sh
new file mode 100644
index 0000000..15fc5cb
--- /dev/null
+++ b/3rdparty/composable_kernel/script/process_perf_data.sh
@@ -0,0 +1,15 @@
+#!/bin/bash 
+#
+# in order to run this script you'd need the following python packages:
+
+#pip3 install --upgrade pip
+#pip3 install sqlalchemy pymysql pandas sshtunnel
+
+# you would also need to set up some environment variables in order to 
+# post your new test results to the database and compare them to the baseline
+# please contact Illia.Silin@amd.com for more details
+
+#process results
+python3 process_perf_data.py perf_gemm.log
+python3 process_perf_data.py perf_resnet50_N256.log
+python3 process_perf_data.py perf_resnet50_N4.log
diff --git a/3rdparty/composable_kernel/script/process_qa_data.sh b/3rdparty/composable_kernel/script/process_qa_data.sh
new file mode 100644
index 0000000..abf1e62
--- /dev/null
+++ b/3rdparty/composable_kernel/script/process_qa_data.sh
@@ -0,0 +1,23 @@
+#!/bin/bash 
+#
+# in order to run this script you'd need the following python packages:
+
+#pip3 install --upgrade pip
+#pip3 install sqlalchemy pymysql pandas sshtunnel
+
+# you would also need to set up some environment variables in order to 
+# post your new test results to the database and compare them to the baseline
+# please contact Illia.Silin@amd.com for more details
+
+#process results
+python3 process_perf_data.py perf_gemm.log
+python3 process_perf_data.py perf_resnet50_N256.log
+python3 process_perf_data.py perf_resnet50_N4.log
+python3 process_perf_data.py perf_batched_gemm.log
+python3 process_perf_data.py perf_grouped_gemm.log
+python3 process_perf_data.py perf_conv_fwd.log
+python3 process_perf_data.py perf_conv_bwd_data.log
+python3 process_perf_data.py perf_gemm_bilinear.log
+python3 process_perf_data.py perf_reduction.log
+python3 process_perf_data.py perf_splitK_gemm.log
+python3 process_perf_data.py perf_onnx_gemm.log
diff --git a/3rdparty/composable_kernel/script/profile_batched_gemm.sh b/3rdparty/composable_kernel/script/profile_batched_gemm.sh
new file mode 100644
index 0000000..d19ddd0
--- /dev/null
+++ b/3rdparty/composable_kernel/script/profile_batched_gemm.sh
@@ -0,0 +1,44 @@
+#!/bin/bash
+ 
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+ 
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+ 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC BatchStrideA BatchStrideB BatchStrideC BatchCount
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  1024 1024       -1     -1      -1            -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1920  2048 2048       -1     -1      -1            -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 3840  4096 4096       -1     -1      -1            -1           -1           -1          4
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 7680  8192 8192       -1     -1      -1            -1           -1           -1          2
+ 
+ #######  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC BatchStrideA BatchStrideB BatchStrideC BatchCount
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024     1024    1024    1024           -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048     2048    2048    2048           -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096     4096    4096    4096           -1           -1           -1          4
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192     8192    8192    8192           -1           -1           -1          2
+ 
+ #######  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC BatchStrideA BatchStrideB BatchStrideC BatchCount
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024     1056    1056    1056           -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048     2080    2080    2080           -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096     4128    4128    4128           -1           -1           -1          4
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192     8224    8224    8224           -1           -1           -1          2
+ 
+ #######  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC BatchStrideA BatchStrideB BatchStrideC BatchCount
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024     1088    1088    1088           -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048     2112    2112    2112           -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096     4160    4160    4160           -1           -1           -1          4
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192     8256    8256    8256           -1           -1           -1          2
diff --git a/3rdparty/composable_kernel/script/profile_conv_bwd_data.sh b/3rdparty/composable_kernel/script/profile_conv_bwd_data.sh
new file mode 100644
index 0000000..a1d2f45
--- /dev/null
+++ b/3rdparty/composable_kernel/script/profile_conv_bwd_data.sh
@@ -0,0 +1,38 @@
+#!/bin/bash
+ 
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+
+ N=$8
+
+# Resnet50
+########  op  datatype  layout  verify  init  log  time conv_dim G__ N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256 1024 1 1   14   14     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512 1024 1 1   14   14     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  128 3 3   28   28     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512  128 1 1   28   28     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  128 3 3   56   56     2 2       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512 2048 1 1    7    7     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N 1024  256 1 1   14   14     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256  256 3 3   14   14     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256  256 3 3   28   28     2 2       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  256 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64  256 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512  512 3 3   14   14     2 2       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  512 1 1   28   28     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256  512 1 1   28   28     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N 2048  512 1 1    7    7     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512  512 3 3    7    7     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256   64 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64   64 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64   64 3 3   56   56     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64    3 7 7  224  224     2 2       1 1      3 3       3 3
diff --git a/3rdparty/composable_kernel/script/profile_conv_fwd.sh b/3rdparty/composable_kernel/script/profile_conv_fwd.sh
new file mode 100644
index 0000000..a1d2f45
--- /dev/null
+++ b/3rdparty/composable_kernel/script/profile_conv_fwd.sh
@@ -0,0 +1,38 @@
+#!/bin/bash
+ 
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+
+ N=$8
+
+# Resnet50
+########  op  datatype  layout  verify  init  log  time conv_dim G__ N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256 1024 1 1   14   14     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512 1024 1 1   14   14     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  128 3 3   28   28     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512  128 1 1   28   28     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  128 3 3   56   56     2 2       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512 2048 1 1    7    7     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N 1024  256 1 1   14   14     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256  256 3 3   14   14     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256  256 3 3   28   28     2 2       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  256 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64  256 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512  512 3 3   14   14     2 2       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  512 1 1   28   28     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256  512 1 1   28   28     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N 2048  512 1 1    7    7     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512  512 3 3    7    7     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256   64 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64   64 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64   64 3 3   56   56     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64    3 7 7  224  224     2 2       1 1      3 3       3 3
diff --git a/3rdparty/composable_kernel/script/profile_gemm.sh b/3rdparty/composable_kernel/script/profile_gemm.sh
new file mode 100644
index 0000000..b88159e
--- /dev/null
+++ b/3rdparty/composable_kernel/script/profile_gemm.sh
@@ -0,0 +1,58 @@
+#!/bin/bash
+
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+echo $DRIVER
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+
+
+# 120 CU
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  1024 1024       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  2048 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1920  1024 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1920  2048 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 3840  4096 4096       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 7680  8192 8192       -1     -1      -1
+ 
+# 104 CU
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  832  1024 1024       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  832  2048 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1664  1024 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1664  2048 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 3328  4096 4096       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 6656  8192 8192       -1     -1      -1
+ 
+# 110 CU
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1280  1408 1024       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1280  2816 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2560  1408 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2560  2816 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 5120  5632 4096       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 7040  8192 8192       -1     -1      -1
+
+# testing different strides
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024	    1024   1024    1024
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048	    2048   2048    2048
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096	    4096   4096    4096
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192	    8192   8192    8192
+ 
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024	    1056   1056    1056
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048	    2080   2080    2080
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096	    4128   4128    4128
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192	    8224   8224    8224
+ 
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024	    1088   1088    1088
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048	    2112   2112    2112
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096	    4160   4160    4160
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192	    8256   8256    8256
diff --git a/3rdparty/composable_kernel/script/profile_gemm_bilinear.sh b/3rdparty/composable_kernel/script/profile_gemm_bilinear.sh
new file mode 100644
index 0000000..e6edefa
--- /dev/null
+++ b/3rdparty/composable_kernel/script/profile_gemm_bilinear.sh
@@ -0,0 +1,41 @@
+#!/bin/bash
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+ 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideD StrideE Alpha Beta
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  1024 1024       -1      -1      -1      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1920  2048 2048       -1      -1      -1      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 3840  4096 4096       -1      -1      -1      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 7680  8192 8192       -1      -1      -1      -1     1    1
+ 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideD StrideE Alpha Beta
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  1024 1024       -1      -1       0      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1920  2048 2048       -1      -1       0      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 3840  4096 4096       -1      -1       0      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 7680  8192 8192       -1      -1       0      -1     1    1
+ 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideD StrideE Alpha Beta
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1000  1000 1000       -1      -1       0      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2000  2000 2000       -1      -1       0      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4000  4000 4000       -1      -1       0      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8000  8000 8000       -1      -1       0      -1     1    1
+ 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideD StrideE Alpha Beta
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024     1056    1056    1056    1056     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048     2080    2080    2080    2080     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096     4128    4128    4128    4128     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192     8224    8224    8224    8224     1    1
+ 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideD StrideE Alpha Beta
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024     1088    1088    1088    1088     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048     2112    2112    2112    2112     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096     4160    4160    4160    4160     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192     8256    8256    8256    8256     1    1
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/script/profile_grouped_gemm.sh b/3rdparty/composable_kernel/script/profile_grouped_gemm.sh
new file mode 100644
index 0000000..8adb7c8
--- /dev/null
+++ b/3rdparty/composable_kernel/script/profile_grouped_gemm.sh
@@ -0,0 +1,18 @@
+#!/bin/bash
+ 
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+ 
+########  op  datatype  layout  verify  init  log  time  Ms______________ Ns______________ Ks_____________ StrideAs___________ StrideBs__________  StrideCs___________
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  256,512,1024,768 128,256,384,1024 128,192,256,512 1024,1025,1044,1026 1024,1024,1024,1024 1025,1024,1028,1024
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  512,768,2048,128 128,256,384,1024 128,192,256,512 1024,1025,2053,1026 1024,1024,1024,1024 1025,1024,2054,1024
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  256,512,1024,768 512,256,768,1024 128,192,256,512 1024,1045,1034,1026 1024,1024,1024,1024 1025,1063,1028,1024
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  512,768,4096,768 128,768,512,2048 128,192,256,512 1024,1027,4096,2050 1024,1024,1024,2048 1025,1024,4099,2049
diff --git a/3rdparty/composable_kernel/script/profile_onnx_gemm.sh b/3rdparty/composable_kernel/script/profile_onnx_gemm.sh
new file mode 100644
index 0000000..c2721e7
--- /dev/null
+++ b/3rdparty/composable_kernel/script/profile_onnx_gemm.sh
@@ -0,0 +1,31 @@
+#!/bin/bash
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+echo $DRIVER
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+# GEMM kernel benchmarks used by ONNX 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  768  768        -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  768  2304       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  768  3072       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  3072 768        -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  1024 1024       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  1024 3072       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  1024 4096       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  4096 1024       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 768 768        -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 768 2304       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 768 3072       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 3072 768       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 1024 1024      -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 1024 3072      -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 1024 4096      -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 4096 1024      -1     -1      -1
+ 
diff --git a/3rdparty/composable_kernel/script/profile_reduce_no_index.sh b/3rdparty/composable_kernel/script/profile_reduce_no_index.sh
new file mode 100644
index 0000000..66bfe1d
--- /dev/null
+++ b/3rdparty/composable_kernel/script/profile_reduce_no_index.sh
@@ -0,0 +1,78 @@
+#!/bin/bash
+DRIVER="../build/bin/ckProfiler"
+VERIFY="-v $1"
+INIT=$2
+NREPEAT=$3
+PRECISION=$4
+##PRECISION=--half
+##PRECISION=--double
+##PRECISION=--int8
+##PRECISION=--bf16
+
+if [ -n $PRECISION ] && [ "$PRECISION" = "--half" -o "$PRECISION" = "--bf16" ]; then
+   ACCTYPE="-C 1"
+elif [ -n $PRECISION ] && [ "$PRECISION" = "--int8" ]; then
+   ACCTYPE="-C 2"
+fi
+
+#### 0 - ADD,  5 - AVG,  7 - NORM2
+Operations="0 5"
+
+#### 0 - ADD,  5 - AVG,    for int8, no NORM2 supported
+if [ -n $PRECISION ] && [ "$PRECISION" = "--int8" -o "$PRECISION" = "--half" ]; then
+   Operations=5
+fi
+
+## for generic validation
+for op in $Operations; do
+    set -x
+    #######        datatype   layout          reduce dims  op     acctype   verify  init  repeats
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,1,2,3   -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 1         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 2         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 3         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 1,2,3     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,2,3     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,1,3     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,22960    -R 0         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,22960    -R 1         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 4,1469440    -R 0         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 4,1469440    -R 1         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    set +x
+done
+
+#### 0 - ADD,  5 - AVG,  7 - NORM2
+Operations=5
+
+## for performance evaluation (resnet50 NHWC => C)
+for op in $Operations; do
+    set -x
+    #######        datatype   layout             reduce dims  op     acctype   verify  init  repeats
+    $DRIVER reduce $PRECISION -D 256,14,14,1024  -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT 
+    $DRIVER reduce $PRECISION -D 256,28,28,128   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,58,58,128   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,7,7,2048    -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,14,14,256   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,30,30,256   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,56,56,256   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,16,16,512   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,28,28,512   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,7,7,512     -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,56,56,64    -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,230,230,3   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,14,14,1024  -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,28,28,128   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,58,58,128   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,7,7,2048    -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,14,14,256   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,30,30,256   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,56,56,256   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,16,16,512   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,28,28,512   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,7,7,512     -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,56,56,64    -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    set +x
+done 
+
diff --git a/3rdparty/composable_kernel/script/profile_reduce_with_index.sh b/3rdparty/composable_kernel/script/profile_reduce_with_index.sh
new file mode 100644
index 0000000..43543f4
--- /dev/null
+++ b/3rdparty/composable_kernel/script/profile_reduce_with_index.sh
@@ -0,0 +1,70 @@
+#!/bin/bash
+DRIVER="../build/bin/ckProfiler"
+VERIFY="-v $1"
+INIT=$2
+NREPEAT=$3
+PRECISION=$4
+##PRECISION=--half
+##PRECISION=--double
+##PRECISION=--int8
+##PRECISION=--bf16
+
+#### 2 - MIN,  3 - MAX,  4 - AMAX
+Operations="2 4"
+
+## for generic validation
+for op in $Operations; do
+    for use_idx in 0 1; do
+        set -x
+        #######        datatype   layout          reduce dims  op     use index    verify  init  repeats
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,1,2,3   -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 1         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 2         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 3         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 1,2,3     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,2,3     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,1,3     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,22960    -R 0         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,22960    -R 1         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 4,1469440    -R 0         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 4,1469440    -R 1         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        set +x
+    done
+done
+
+Operations=2
+
+## for performance evaluation (resnet50 NHWC => C)
+for op in $Operations; do
+    for use_idx in 0 1; do
+        set -x
+        #######        datatype   layout             reduce dims  op     use index    verify  init  repeats
+        $DRIVER reduce $PRECISION -D 256,14,14,1024  -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,28,28,128   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,58,58,128   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,7,7,2048    -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,14,14,256   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,30,30,256   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,56,56,256   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,16,16,512   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,28,28,512   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,7,7,512     -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,56,56,64    -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,230,230,3   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,14,14,1024  -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,28,28,128   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,58,58,128   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,7,7,2048    -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,14,14,256   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,30,30,256   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,56,56,256   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,16,16,512   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,28,28,512   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,7,7,512     -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,56,56,64    -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        set +x
+    done
+done 
+
diff --git a/3rdparty/composable_kernel/script/profile_resnet50.sh b/3rdparty/composable_kernel/script/profile_resnet50.sh
new file mode 100644
index 0000000..b55cb2c
--- /dev/null
+++ b/3rdparty/composable_kernel/script/profile_resnet50.sh
@@ -0,0 +1,69 @@
+#!/bin/bash
+
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+
+OP=$1
+DATATYPE=$2
+IN_LAYOUT=$3
+WEI_LAYOUT=$4
+OUT_LAYOUT=$5
+VERIFY=$6
+INIT=$7
+LOG=$8
+TIME=$9
+
+ N=${10}
+
+# Resnet50
+######## op____________________  datatype  in_layout   wei_layout  out_layout  verify  init  log  time  N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64    3 7 7  224 224    2   2     1   1    3   3     3   3
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64   64 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64   64 3 3   56  56    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256   64 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64  256 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64   64 3 3   56  56    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256   64 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64  256 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64   64 3 3   56  56    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256   64 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  256 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  128 3 3   56  56    2   2     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  128 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  512 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  128 3 3   28  28    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  128 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  512 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  128 3 3   28  28    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  128 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  512 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  128 3 3   28  28    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  128 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  512 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  256 3 3   28  28    2   2     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 1024  256 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256 1024 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  256 3 3   14  14    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 1024  256 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256 1024 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  256 3 3   14  14    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 1024  256 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256 1024 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  256 3 3   14  14    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 1024  256 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256 1024 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  256 3 3   14  14    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 1024  256 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256 1024 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  256 3 3   14  14    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 1024  256 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512 1024 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  512 3 3   14  14    2   2     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 2048  512 1 1    7   7    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512 2048 1 1    7   7    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  512 3 3    7   7    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 2048  512 1 1    7   7    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512 2048 1 1    7   7    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  512 3 3    7   7    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 2048  512 1 1    7   7    1   1     1   1    0   0     0   0
diff --git a/3rdparty/composable_kernel/script/profile_splitK_gemm.sh b/3rdparty/composable_kernel/script/profile_splitK_gemm.sh
new file mode 100644
index 0000000..d62f0e4
--- /dev/null
+++ b/3rdparty/composable_kernel/script/profile_splitK_gemm.sh
@@ -0,0 +1,41 @@
+#!/bin/bash
+
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+echo $DRIVER
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+KBatch=$8
+
+
+# 120 CU
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC  KBatch_
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  1024 1024       -1     -1      -1   $KBatch
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  2048 2048       -1     -1      -1   $KBatch
+ 
+# 104 CU
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC KBatch_
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  832  1024 1024       -1     -1      -1  $KBatch
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  832  2048 2048       -1     -1      -1  $KBatch
+ 
+# 110 CU
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC KBatch_
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1280  1408 1024       -1     -1      -1  $KBatch
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1280  2816 2048       -1     -1      -1  $KBatch
+
+# testing different strides
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC KBatch_
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024	    1024   1024    1024  $KBatch
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048	    2048   2048    2048  $KBatch
+ 
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024	    1056   1056    1056  $KBatch
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048	    2080   2080    2080  $KBatch
+ 
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024	    1088   1088    1088  $KBatch
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048	    2112   2112    2112  $KBatch
diff --git a/3rdparty/composable_kernel/script/run_full_performance_tests.sh b/3rdparty/composable_kernel/script/run_full_performance_tests.sh
new file mode 100644
index 0000000..eae334a
--- /dev/null
+++ b/3rdparty/composable_kernel/script/run_full_performance_tests.sh
@@ -0,0 +1,149 @@
+#!/bin/bash 
+#
+# in order to run this script you'd first need to build the ckProfiler executable in ../build/bin/
+# you would also need to set up some environment variables in order to 
+# post your new test results to the database and compare them to the baseline
+# please contact Illia.Silin@amd.com for more details
+#
+# run the script as "./run_full_performance_tests.sh <verification> <tag for your test environment> <branch name> < node name>
+# input arguments: 
+# verification = 0 : do not verify result correctness on CPU
+#              = 1 : verifuy correctness on CPU (may take a long time)
+# environment tag  : a string describing the specifics of your test environment
+# branch name      : name of the branch in git repo (git status | grep -e 'On branch')
+# node name        : $hostname
+
+#get the command line arguments:
+export verify=$1
+echo 'Verification: ' $verify
+export env_type=$2
+echo 'Environment type: ' $env_type
+export branch=$3
+echo 'Branch name: ' $branch
+export host_name=$4
+echo 'Host name: ' $host_name
+function print_log_header(){
+	rm -f $1;
+	echo 'On branch ' $3 &> $1;
+	echo 'Node name: ' $4 >> $1;
+	#get GPU_arch and number of compute units from rocminfo
+	echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
+	rocminfo | grep "Compute Unit:" >> $1;
+	hipcc --version | grep -e 'HIP version'  >> $1;
+	echo 'Environment type: ' $2 >> $1;
+	/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
+}
+
+#run gemm tests
+export gemm_log="perf_gemm.log"
+print_log_header $gemm_log $env_type $branch $host_name
+./profile_gemm.sh gemm 0 0 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 0 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 0 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 0 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 0 1 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 1 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 1 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 1 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 0 2 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 2 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 2 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 2 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 0 3 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 3 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 3 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 3 $verify 1 0 1 2>&1 | tee -a $gemm_log
+
+#run batched_gemm tests
+export batched_gemm_log="perf_batched_gemm.log"
+print_log_header $batched_gemm_log $env_type $branch $host_name
+./profile_batched_gemm.sh batched_gemm 0 0 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 0 1 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 0 2 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 0 3 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 1 0 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 1 1 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 1 2 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 1 3 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 2 0 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 2 1 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 2 2 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 2 3 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 3 0 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 3 1 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 3 2 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 3 3 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+
+#run grouped_gemm tests
+export grouped_gemm_log="perf_grouped_gemm.log"
+print_log_header $grouped_gemm_log $env_type $branch $host_name
+./profile_grouped_gemm.sh grouped_gemm 1 0 $verify 1 0 1 2>&1 | tee -a $grouped_gemm_log
+./profile_grouped_gemm.sh grouped_gemm 1 1 $verify 1 0 1 2>&1 | tee -a $grouped_gemm_log
+./profile_grouped_gemm.sh grouped_gemm 1 2 $verify 1 0 1 2>&1 | tee -a $grouped_gemm_log
+./profile_grouped_gemm.sh grouped_gemm 1 3 $verify 1 0 1 2>&1 | tee -a $grouped_gemm_log
+
+#run GEMM+Bilinear tests
+export gemm_bilinear_log="perf_gemm_bilinear.log"
+print_log_header $gemm_bilinear_log $env_type $branch $host_name
+./profile_gemm_bilinear.sh gemm_bilinear 1 0 $verify 1 0 1 2>&1 | tee -a $gemm_bilinear_log
+./profile_gemm_bilinear.sh gemm_bilinear 1 1 $verify 1 0 1 2>&1 | tee -a $gemm_bilinear_log
+./profile_gemm_bilinear.sh gemm_bilinear 1 2 $verify 1 0 1 2>&1 | tee -a $gemm_bilinear_log
+./profile_gemm_bilinear.sh gemm_bilinear 1 3 $verify 1 0 1 2>&1 | tee -a $gemm_bilinear_log
+
+#run conv_fwd tests
+export conv_fwd_log="perf_conv_fwd.log"
+print_log_header $conv_fwd_log $env_type $branch $host_name
+./profile_conv_fwd.sh conv_fwd 0 1 $verify 1 0 1 256 2>&1 | tee -a $conv_fwd_log
+./profile_conv_fwd.sh conv_fwd 1 1 $verify 1 0 1 256 2>&1 | tee -a $conv_fwd_log
+./profile_conv_fwd.sh conv_fwd 2 1 $verify 1 0 1 256 2>&1 | tee -a $conv_fwd_log
+./profile_conv_fwd.sh conv_fwd 3 1 $verify 1 0 1 256 2>&1 | tee -a $conv_fwd_log
+
+#run conv_bwd_data tests
+export conv_bwd_data_log="perf_conv_bwd_data.log"
+print_log_header $conv_bwd_data_log $env_type $branch $host_name
+./profile_conv_bwd_data.sh conv_bwd_data 0 1 $verify 1 0 1 256 2>&1 | tee -a $conv_bwd_data_log
+./profile_conv_bwd_data.sh conv_bwd_data 1 1 $verify 1 0 1 256 2>&1 | tee -a $conv_bwd_data_log
+./profile_conv_bwd_data.sh conv_bwd_data 2 1 $verify 1 0 1 256 2>&1 | tee -a $conv_bwd_data_log
+./profile_conv_bwd_data.sh conv_bwd_data 3 1 $verify 1 0 1 256 2>&1 | tee -a $conv_bwd_data_log
+
+#run resnet50 tests
+export resnet256_log="perf_resnet50_N256.log"
+print_log_header $resnet256_log $env_type $branch $host_name
+./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 256 2>&1 | tee -a $resnet256_log
+export resnet4_log="perf_resnet50_N4.log"
+print_log_header $resnet4_log $env_type $branch $host_name
+./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 4 2>&1 | tee -a $resnet4_log
+
+#run reduction tests
+export reduction_log="perf_reduction.log"
+print_log_header $reduction_log $env_type $branch $host_name
+./profile_reduce_with_index.sh $verify 2 10 --half 2>&1 | tee -a $reduction_log
+./profile_reduce_no_index.sh $verify 2 10 --half 2>&1 | tee -a $reduction_log
+
+#run splitK_gemm tests, first correctness verification, then performance
+export splitK_gemm_ver_log="perf_splitK_gemm_verify.log"
+print_log_header $splitK_gemm_ver_log $env_type $branch $host_name
+./profile_splitK_gemm.sh gemm_splitk 0 0 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 0 1 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 0 2 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 0 3 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 1 0 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 1 1 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 1 2 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 1 3 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+export splitK_gemm_log="perf_splitK_gemm.log"
+print_log_header $splitK_gemm_log $env_type $branch $host_name
+./profile_splitK_gemm.sh gemm_splitk 0 0 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 0 1 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 0 2 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 0 3 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 1 0 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 1 1 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 1 2 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 1 3 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+
+#run ONNX gemm tests
+export onnx_log="perf_onnx_gemm.log"
+print_log_header $onnx_log $env_type $branch $host_name
+./profile_onnx_gemm.sh gemm 0 0 $verify 1 0 1 2>&1 | tee -a $onnx_log
+./profile_onnx_gemm.sh gemm 1 0 $verify 1 0 1 2>&1 | tee -a $onnx_log
diff --git a/3rdparty/composable_kernel/script/run_performance_tests.sh b/3rdparty/composable_kernel/script/run_performance_tests.sh
new file mode 100644
index 0000000..4e3a6fc
--- /dev/null
+++ b/3rdparty/composable_kernel/script/run_performance_tests.sh
@@ -0,0 +1,60 @@
+#!/bin/bash 
+#
+# in order to run this script you'd first need to build the ckProfiler executable in ../build/bin/
+# run the script as "./run_performance_tests.sh <verification> <tag for your test environment> <branch name> < node name>
+# input arguments: 
+# verification = 0 : do not verify result correctness on CPU
+#              = 1 : verify correctness on CPU (may take a long time)
+# environment tag  : a string describing the specifics of your test environment
+# branch name      : name of the branch in git repo (git status | grep -e 'On branch')
+# node name        : $hostname
+
+#get the command line arguments:
+export verify=$1
+echo 'Verification: ' $verify
+export env_type=$2
+echo 'Environment type: ' $env_type
+export branch=$3
+echo 'Branch name: ' $branch
+export host_name=$4
+echo 'Host name: ' $host_name
+
+function print_log_header(){
+	rm -f $1;
+	echo 'On branch ' $3 &> $1;
+	echo 'Node name: ' $4 >> $1;
+	#get GPU_arch and number of compute units from rocminfo
+	echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
+	rocminfo | grep "Compute Unit:" >> $1;
+	hipcc --version | grep -e 'HIP version'  >> $1;
+	echo 'Environment type: ' $2 >> $1;
+	/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
+}
+
+#run gemm tests
+export gemm_log="perf_gemm.log"
+print_log_header $gemm_log $env_type $branch $host_name
+./profile_gemm.sh gemm 0 0 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 0 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 0 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 0 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 0 1 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 1 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 1 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 1 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 0 2 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 2 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 2 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 2 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 0 3 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 3 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 3 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 3 $verify 1 0 1 | tee -a $gemm_log
+
+#run resnet50 tests
+export resnet256_log="perf_resnet50_N256.log"
+print_log_header $resnet256_log $env_type $branch $host_name
+./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 256 | tee -a $resnet256_log
+export resnet4_log="perf_resnet50_N4.log"
+print_log_header $resnet4_log $env_type $branch $host_name
+./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 4 | tee -a $resnet4_log
diff --git a/3rdparty/composable_kernel/script/test_convnd_fwd.sh b/3rdparty/composable_kernel/script/test_convnd_fwd.sh
new file mode 100644
index 0000000..1bd7a6b
--- /dev/null
+++ b/3rdparty/composable_kernel/script/test_convnd_fwd.sh
@@ -0,0 +1,110 @@
+#!/usr/bin/env bash
+
+# set -e
+
+DIM1=False
+DIM2=True
+DIM3=False
+DATE=220317
+GIT_HASH=4e6dfda
+LOG_DIR=${DATE}_${GIT_HASH}
+SUFFIX=${GIT_HASH}
+
+
+#--------------------------------------------------------------------------
+#   Commandline arguments parsing
+#   like: cmd -key[--key] value
+#--------------------------------------------------------------------------
+
+POSITIONAL=()
+while [[ $# -gt 0 ]]
+do
+key="$1"
+
+case $key in
+    -d1|--d1)
+    DIM1=True
+    echo DIM1: "${DIM1}"
+    shift # past argument
+    ;;
+    -d2|--d2)
+    DIM2=True
+    echo DIM2: "${DIM2}"
+    shift # past argument
+    ;;
+    -d3|--d3)
+    DIM3=True
+    echo DIM3: "${DIM3}"
+    shift # past argument
+    ;;
+    -all|--all)
+    DIM1=True
+    DIM2=True
+    DIM3=True
+    echo DIM1: "${DIM1}"
+    echo DIM2: "${DIM2}"
+    echo DIM3: "${DIM3}"
+    shift # past argument
+    ;;
+    -s|--suffix)
+    SUFFIX=${SUFFIX}_"$2"
+    echo SUFFIX: "${SUFFIX}"
+    shift # past argument
+    shift # past value
+    ;;
+    *)    # unknown option
+    POSITIONAL+=("$1") # save it in an array for later
+    shift # past argument
+    ;;
+esac
+done
+set -- "${POSITIONAL[@]}" # restore positional parameters
+
+#--------------------------------------------------------------------------
+
+# NUMACTL="numactl --cpunodebind=1 --membind=1"
+NUMACTL=
+# ENV_CONF=
+GPU=mi100
+PROF_ITER_COUNT=10000
+LOG_DIR_PATH=../log/${LOG_DIR}
+set -x
+
+#-------------------------------------------------------------------------------
+#               1D
+#-------------------------------------------------------------------------------
+
+if [[ "${DIM1}" == "True" ]]; then
+    mkdir -p ${LOG_DIR_PATH}
+    echo ">>>>>>>> RUN test conv1d nwc <<<<<<<<<<"
+    CMD="./../build/bin/test_conv1d_fwd"
+    ${NUMACTL} ${CMD} 2>&1 \
+        | tee ${LOG_DIR_PATH}/test_conv1d_fwd_nwc_${SUFFIX}_${GPU}.log
+
+fi
+
+#-------------------------------------------------------------------------------
+#               2D
+#-------------------------------------------------------------------------------
+
+if [[ "${DIM2}" == "True" ]]; then
+    mkdir -p ${LOG_DIR_PATH}
+    echo ">>>>>>>> RUN test conv2d nhwc <<<<<<<<<<"
+    CMD="./../build/bin/test_conv2d_fwd"
+    ${NUMACTL} ${CMD} 2>&1 \
+        | tee ${LOG_DIR_PATH}/test_conv2d_fwd_nhwc_${SUFFIX}_${GPU}.log
+
+fi
+
+#-------------------------------------------------------------------------------
+#               3D
+#-------------------------------------------------------------------------------
+
+if [[ "${DIM3}" == "True" ]]; then
+    mkdir -p ${LOG_DIR_PATH}
+    echo ">>>>>>>> RUN test conv3d ndhwc <<<<<<<<<<"
+    CMD="./../build/bin/test_conv3d_fwd"
+    ${NUMACTL} ${CMD} 2>&1 \
+        | tee ${LOG_DIR_PATH}/test_conv3d_fwd_ndhwc_${SUFFIX}_${GPU}.log
+
+fi
diff --git a/3rdparty/composable_kernel/script/test_reduce_no_index.sh b/3rdparty/composable_kernel/script/test_reduce_no_index.sh
new file mode 100644
index 0000000..b956303
--- /dev/null
+++ b/3rdparty/composable_kernel/script/test_reduce_no_index.sh
@@ -0,0 +1,63 @@
+#!/bin/bash
+
+## The following will be used for CI
+
+set -x
+
+## for float
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2,3  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,3  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,2,3  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1,2,3  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 2  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 3  0 2
+
+## for float64
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2,3  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,3  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,2,3  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1,2,3  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 2  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 3  6 2
+
+## for float16
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2,3  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,3  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,2,3  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1,2,3  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 2  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 3  1 2
+
+## for int8_t
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2,3  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,3  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,2,3  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1,2,3  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 2  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 3  3 2
+
+## for bfloat16
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2,3  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,3  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,2,3  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1,2,3  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 2  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 3  5 2
+
+set +x
+
diff --git a/3rdparty/composable_kernel/script/test_reduce_with_index.sh b/3rdparty/composable_kernel/script/test_reduce_with_index.sh
new file mode 100644
index 0000000..b0843ba
--- /dev/null
+++ b/3rdparty/composable_kernel/script/test_reduce_with_index.sh
@@ -0,0 +1,63 @@
+#!/bin/bash
+
+## The following will be used for CI
+
+set -x
+
+## for float
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2,3  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,3  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,2,3  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1,2,3  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 2  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 3  0 2
+
+## for float64
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2,3  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,3  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,2,3  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1,2,3  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 2  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 3  6 2
+
+## for float16
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2,3  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,3  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,2,3  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1,2,3  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 2  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 3  1 2
+
+## for int8_t
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2,3  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,3  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,2,3  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1,2,3  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 2  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 3  3 2
+
+## for bfloat16
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2,3  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,3  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,2,3  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1,2,3  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 2  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 3  5 2
+
+set +x
+
diff --git a/3rdparty/composable_kernel/test/CMakeLists.txt b/3rdparty/composable_kernel/test/CMakeLists.txt
new file mode 100644
index 0000000..b2e25e4
--- /dev/null
+++ b/3rdparty/composable_kernel/test/CMakeLists.txt
@@ -0,0 +1,60 @@
+include_directories(BEFORE
+    ${PROJECT_SOURCE_DIR}/
+    ${PROJECT_SOURCE_DIR}/profiler/include
+)
+
+include(googletest)
+
+add_custom_target(tests)
+
+function(add_test_executable TEST_NAME)
+    message("adding test ${TEST_NAME}")
+    add_executable(${TEST_NAME} ${ARGN})
+    add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}>)
+    add_dependencies(tests ${TEST_NAME})
+    add_dependencies(check ${TEST_NAME})
+    rocm_install(TARGETS ${TEST_NAME} COMPONENT tests)
+endfunction(add_test_executable TEST_NAME)
+
+include(GoogleTest)
+
+function(add_gtest_executable TEST_NAME)
+    message("adding gtest ${TEST_NAME}")
+    add_executable(${TEST_NAME} ${ARGN})
+    add_dependencies(tests ${TEST_NAME})
+    add_dependencies(check ${TEST_NAME})
+
+    # suppress gtest warnings
+    target_compile_options(${TEST_NAME} PRIVATE -Wno-global-constructors -Wno-undef)
+    target_link_libraries(${TEST_NAME} PRIVATE gtest_main)
+    add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}> )
+    rocm_install(TARGETS ${TEST_NAME} COMPONENT tests)
+endfunction(add_gtest_executable TEST_NAME)
+
+add_subdirectory(magic_number_division)
+add_subdirectory(space_filling_curve)
+add_subdirectory(conv_util)
+add_subdirectory(reference_conv_fwd)
+add_subdirectory(gemm)
+add_subdirectory(gemm_split_k)
+add_subdirectory(gemm_reduce)
+add_subdirectory(batched_gemm)
+add_subdirectory(batched_gemm_reduce)
+add_subdirectory(batched_gemm_gemm)
+add_subdirectory(batched_gemm_softmax_gemm)
+add_subdirectory(batched_gemm_softmax_gemm_permute)
+add_subdirectory(grouped_gemm)
+add_subdirectory(reduce)
+add_subdirectory(convnd_fwd)
+add_subdirectory(convnd_bwd_data)
+add_subdirectory(grouped_convnd_fwd)
+add_subdirectory(grouped_convnd_bwd_weight)
+add_subdirectory(block_to_ctile_map)
+add_subdirectory(softmax)
+add_subdirectory(normalization)
+add_subdirectory(data_type)
+add_subdirectory(elementwise_normalization)
+add_subdirectory(batchnorm)
+if(GPU_TARGETS MATCHES "gfx1100")
+    add_subdirectory(wmma_op)
+endif()
diff --git a/3rdparty/composable_kernel/test/batched_gemm/CMakeLists.txt b/3rdparty/composable_kernel/test/batched_gemm/CMakeLists.txt
new file mode 100644
index 0000000..0574f98
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm/CMakeLists.txt
@@ -0,0 +1,15 @@
+add_test_executable(test_batched_gemm_fp16 batched_gemm_fp16.cpp)
+target_link_libraries(test_batched_gemm_fp16 PRIVATE utility)
+target_link_libraries(test_batched_gemm_fp16 PRIVATE device_batched_gemm_instance)
+
+add_test_executable(test_batched_gemm_fp32 batched_gemm_fp32.cpp)
+target_link_libraries(test_batched_gemm_fp32 PRIVATE utility)
+target_link_libraries(test_batched_gemm_fp32 PRIVATE device_batched_gemm_instance)
+
+add_test_executable(test_batched_gemm_bf16 batched_gemm_bf16.cpp)
+target_link_libraries(test_batched_gemm_bf16 PRIVATE utility)
+target_link_libraries(test_batched_gemm_bf16 PRIVATE device_batched_gemm_instance)
+
+add_test_executable(test_batched_gemm_int8 batched_gemm_int8.cpp)
+target_link_libraries(test_batched_gemm_int8 PRIVATE utility)
+target_link_libraries(test_batched_gemm_int8 PRIVATE device_batched_gemm_instance)
diff --git a/3rdparty/composable_kernel/test/batched_gemm/batched_gemm_bf16.cpp b/3rdparty/composable_kernel/test/batched_gemm/batched_gemm_bf16.cpp
new file mode 100644
index 0000000..78be540
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm/batched_gemm_bf16.cpp
@@ -0,0 +1,44 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_batched_gemm_impl.hpp"
+
+namespace {
+using ADataType = ck::bhalf_t;
+using BDataType = ck::bhalf_t;
+using CDataType = ck::bhalf_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+} // namespace
+
+int main()
+{
+    int M          = 256;
+    int N          = 256;
+    int K          = 128;
+    int BatchCount = 3;
+
+    bool pass = true;
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Row, Row, Row>(
+               true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Row, Col, Row>(
+               true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Col, Row, Row>(
+               true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Col, Col, Row>(
+               true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
+
+    std::cout << "test BatchedGEMM bf16: " << (pass ? "Pass" : "Fail") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/test/batched_gemm/batched_gemm_fp16.cpp b/3rdparty/composable_kernel/test/batched_gemm/batched_gemm_fp16.cpp
new file mode 100644
index 0000000..6cbbedf
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm/batched_gemm_fp16.cpp
@@ -0,0 +1,44 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_batched_gemm_impl.hpp"
+
+namespace {
+using ADataType = ck::half_t;
+using BDataType = ck::half_t;
+using CDataType = ck::half_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+} // namespace
+
+int main()
+{
+    int M          = 512;
+    int N          = 256;
+    int K          = 128;
+    int BatchCount = 3;
+
+    bool pass = true;
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Row, Row, Row>(
+               true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Row, Col, Row>(
+               true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Col, Row, Row>(
+               true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Col, Col, Row>(
+               true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
+
+    std::cout << "test BatchedGEMM fp16: " << (pass ? "Pass" : "Fail") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/test/batched_gemm/batched_gemm_fp32.cpp b/3rdparty/composable_kernel/test/batched_gemm/batched_gemm_fp32.cpp
new file mode 100644
index 0000000..c9e565e
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm/batched_gemm_fp32.cpp
@@ -0,0 +1,44 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_batched_gemm_impl.hpp"
+
+namespace {
+using ADataType = float;
+using BDataType = float;
+using CDataType = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+} // namespace
+
+int main()
+{
+    int M          = 256;
+    int N          = 256;
+    int K          = 128;
+    int BatchCount = 3;
+
+    bool pass = true;
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Row, Row, Row>(
+               true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Row, Col, Row>(
+               true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Col, Row, Row>(
+               true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Col, Col, Row>(
+               true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
+
+    std::cout << "test BatchedGEMM fp32: " << (pass ? "Pass" : "Fail") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/test/batched_gemm/batched_gemm_int8.cpp b/3rdparty/composable_kernel/test/batched_gemm/batched_gemm_int8.cpp
new file mode 100644
index 0000000..4da941a
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm/batched_gemm_int8.cpp
@@ -0,0 +1,44 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_batched_gemm_impl.hpp"
+
+namespace {
+using ADataType = int8_t;
+using BDataType = int8_t;
+using CDataType = int8_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+} // namespace
+
+int main()
+{
+    int M          = 256;
+    int N          = 256;
+    int K          = 128;
+    int BatchCount = 3;
+
+    bool pass = true;
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Row, Row, Row>(
+               true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Row, Col, Row>(
+               true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Col, Row, Row>(
+               true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass &&
+           ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Col, Col, Row>(
+               true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
+
+    std::cout << "test BatchedGEMM int8: " << (pass ? "Pass" : "Fail") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/test/batched_gemm_gemm/CMakeLists.txt b/3rdparty/composable_kernel/test/batched_gemm_gemm/CMakeLists.txt
new file mode 100644
index 0000000..3868097
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm_gemm/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_custom_target(test_batched_gemm_gemm)
+
+add_gtest_executable(test_batched_gemm_gemm_fp16 test_batched_gemm_gemm_fp16.cpp)
+target_link_libraries(test_batched_gemm_gemm_fp16 PRIVATE utility device_batched_gemm_gemm_instance)
+add_dependencies(test_batched_gemm_gemm test_batched_gemm_gemm_fp16)
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp b/3rdparty/composable_kernel/test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp
new file mode 100644
index 0000000..aa113de
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp
@@ -0,0 +1,149 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_gemm_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmGemmFP16 : public TestBatchedGemmGemm<Tuple>
+{
+};
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<F16, F16, F16, F16, Row, Col, Row, Row>,
+    std::tuple<F16, F16, F16, F16, Row, Col, Col, Row>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmGemmFP16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_PadM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {136, 128, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_PadN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 136, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_PadK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 40, 128, 1},
+        {128, 128, 136, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_PadO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 136, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_OddM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {129, 128, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_OddN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 129, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_OddK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 33, 128, 1},
+        {128, 128, 129, 128, 1},
+    };
+    this->Run();
+}
+
+// If kernel B1Layout is RowMajor, expect not to support odd O size
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_OddO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 129, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, DISABLED_Bench_FP16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 768},
+        {256, 256, 128, 128, 768},
+        {512, 512, 64, 64, 768},
+        {512, 512, 128, 128, 768},
+        {1024, 1024, 64, 64, 768},
+        {1024, 1024, 128, 128, 768},
+        {2048, 2048, 64, 64, 768},
+        {2048, 2048, 128, 128, 768},
+        {4096, 4096, 64, 64, 768},
+        {4096, 4096, 128, 128, 768},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+TEST(TestBatchedGemmGemmInterface, GemmSpecializationSizeMatch)
+{
+    int P = 120; // requires padding
+    int Q = 128; // do not require padding
+
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // clang-format on
+}
+
+TEST(TestBatchedGemmGemmInterface, GemmSpecializationSizeMismatch)
+{
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // clang-format on
+}
diff --git a/3rdparty/composable_kernel/test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp b/3rdparty/composable_kernel/test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp
new file mode 100644
index 0000000..53c4d37
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp
@@ -0,0 +1,189 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include <vector>
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "profiler/profile_batched_gemm_gemm_impl.hpp"
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+
+using F16 = ck::half_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <typename Tuple>
+struct TestBatchedGemmGemm : public ::testing::Test
+{
+    using ADataType  = std::tuple_element_t<0, Tuple>;
+    using B0DataType = std::tuple_element_t<1, Tuple>;
+    using B1DataType = std::tuple_element_t<2, Tuple>;
+    using CDataType  = std::tuple_element_t<3, Tuple>;
+    using ALayout    = std::tuple_element_t<4, Tuple>;
+    using B0Layout   = std::tuple_element_t<5, Tuple>;
+    using B1Layout   = std::tuple_element_t<6, Tuple>;
+    using CLayout    = std::tuple_element_t<7, Tuple>;
+
+    std::vector<std::vector<int>> lengths_ = {
+        {256, 256, 64, 64, 4},
+        {256, 256, 128, 128, 4},
+        {512, 512, 64, 64, 2},
+        {512, 512, 128, 128, 2},
+        {1024, 1024, 64, 64, 1},
+        {1024, 1024, 128, 128, 1},
+    };
+    bool bench_  = false;
+    bool verify_ = true;
+
+    void RunSingle(int M, int N, int K, int O, int BatchCount)
+    {
+        bool pass = ck::profiler::profile_batched_gemm_gemm_impl<ADataType,
+                                                                 B0DataType,
+                                                                 B1DataType,
+                                                                 CDataType,
+                                                                 ALayout,
+                                                                 B0Layout,
+                                                                 B1Layout,
+                                                                 CLayout>(
+            verify_, 1, false, bench_, M, N, K, O, BatchCount);
+
+        EXPECT_TRUE(pass);
+    }
+
+    void Run()
+    {
+        for(auto lengths : this->lengths_)
+        {
+            int M          = lengths[0];
+            int N          = lengths[1];
+            int K          = lengths[2];
+            int O          = lengths[3];
+            int BatchCount = lengths[4];
+
+            this->RunSingle(M, N, K, O, BatchCount);
+        }
+    }
+};
+
+template <GemmSpecialization GemmSpec>
+struct DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    using ALayout  = Row;
+    using B0Layout = Col;
+    using B1Layout = Row;
+    using CLayout  = Row;
+
+    using ADataType        = F16;
+    using B0DataType       = F16;
+    using B1DataType       = F16;
+    using AccDataType      = float;
+    using CShuffleDataType = float;
+    using CDataType        = F16;
+
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = PassThrough;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    // static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
+
+    using DeviceGemmGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+        ALayout,
+        B0Layout,
+        B1Layout,
+        CLayout,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<8, 32, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+    bool IsSupported(int M, int N, int K, int O)
+    {
+        auto gemm     = DeviceGemmGemmInstance{};
+        auto invoker  = gemm.MakeInvoker();
+        auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
+                                          static_cast<B0DataType*>(nullptr),
+                                          static_cast<B1DataType*>(nullptr),
+                                          static_cast<CDataType*>(nullptr),
+                                          M,
+                                          N,
+                                          K,
+                                          O,
+                                          0,              // BatchCount
+                                          0,              // StrideA
+                                          0,              // StrideB0
+                                          0,              // StrideB1
+                                          0,              // StrideC
+                                          0,              // BatchStrideA
+                                          0,              // BatchStrideB0
+                                          0,              // BatchStrideB1
+                                          0,              // BatchStrideC
+                                          PassThrough{},  // a_element_op
+                                          PassThrough{},  // b0_element_op
+                                          PassThrough{},  // acc0_element_op
+                                          PassThrough{},  // b1_element_op
+                                          PassThrough{}); // c_element_op
+
+        return gemm.IsSupportedArgument(argument);
+    }
+};
diff --git a/3rdparty/composable_kernel/test/batched_gemm_reduce/CMakeLists.txt b/3rdparty/composable_kernel/test/batched_gemm_reduce/CMakeLists.txt
new file mode 100644
index 0000000..4dc0b08
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm_reduce/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_test_executable(test_batched_gemm_reduce_fp16 batched_gemm_reduce_fp16.cpp)
+target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE utility)
+target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE device_batched_gemm_reduce_instance)
diff --git a/3rdparty/composable_kernel/test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp b/3rdparty/composable_kernel/test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp
new file mode 100644
index 0000000..b150ce5
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_batched_gemm_reduce_impl.hpp"
+
+int main()
+{
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    int M = 512;
+    int N = 256;
+    int K = 128;
+
+    int BatchCount = 3;
+
+    bool pass = true;
+
+    pass = pass && ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                                  ck::half_t,
+                                                                  ck::half_t,
+                                                                  float,
+                                                                  Row,
+                                                                  Row,
+                                                                  Row>(
+                       true, 1, false, false, M, N, K, K, N, N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                                  ck::half_t,
+                                                                  ck::half_t,
+                                                                  float,
+                                                                  Row,
+                                                                  Col,
+                                                                  Row>(
+                       true, 1, false, false, M, N, K, K, K, N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                                  ck::half_t,
+                                                                  ck::half_t,
+                                                                  float,
+                                                                  Col,
+                                                                  Row,
+                                                                  Row>(
+                       true, 1, false, false, M, N, K, M, N, N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                                  ck::half_t,
+                                                                  ck::half_t,
+                                                                  float,
+                                                                  Col,
+                                                                  Col,
+                                                                  Row>(
+                       true, 1, false, false, M, N, K, M, K, N, BatchCount);
+
+    if(pass)
+    {
+        std::cout << "test BatchedGEMM+Reduce fp16: Pass" << std::endl;
+        return 0;
+    }
+    else
+    {
+        std::cout << "test BatchedGEMM+Reduce fp16: Fail" << std::endl;
+        return -1;
+    }
+}
diff --git a/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm/CMakeLists.txt b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm/CMakeLists.txt
new file mode 100644
index 0000000..1ceecef
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_custom_target(test_batched_gemm_softmax_gemm)
+
+add_gtest_executable(test_batched_gemm_softmax_gemm_fp16 test_batched_gemm_softmax_gemm_fp16.cpp)
+target_link_libraries(test_batched_gemm_softmax_gemm_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_instance)
+add_dependencies(test_batched_gemm_softmax_gemm test_batched_gemm_softmax_gemm_fp16)
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp
new file mode 100644
index 0000000..5df7769
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp
@@ -0,0 +1,176 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_softmax_gemm_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmSoftmaxGemmFP16 : public TestBatchedGemmSoftmaxGemm<Tuple>
+{
+};
+
+using Masked = std::true_type;
+using NoMask = std::false_type;
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<F16, F16, F16, F16, Row, Col, Row, Row, NoMask>,
+    std::tuple<F16, F16, F16, F16, Row, Col, Row, Row, Masked>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmSoftmaxGemmFP16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_PadM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {136, 128, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_PadN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 136, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_PadK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 40, 128, 1},
+        {128, 128, 136, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_PadO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 136, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_OddM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {129, 128, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_OddN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 129, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_OddK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 33, 128, 1},
+        {128, 128, 129, 128, 1},
+    };
+    this->Run();
+}
+
+// If kernel B1Layout is RowMajor, expect not to support odd O size
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_OddO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 129, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, DISABLED_Bench_FP16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 768},
+        {256, 256, 128, 128, 768},
+        {512, 512, 64, 64, 768},
+        {512, 512, 128, 128, 768},
+        {1024, 1024, 64, 64, 768},
+        {1024, 1024, 128, 128, 768},
+        {2048, 2048, 64, 64, 768},
+        {2048, 2048, 128, 128, 768},
+        {4096, 4096, 64, 64, 768},
+        {4096, 4096, 128, 128, 768},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, DISABLED_Bench_FP16_IrregularK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{{256, 256, 160, 160, 16},
+                                                   {256, 64, 160, 64, 16},
+                                                   {1024, 1024, 80, 80, 16},
+                                                   {1024, 64, 80, 64, 16},
+                                                   {4096, 4096, 40, 40, 16},
+                                                   {4096, 64, 40, 64, 16}};
+    this->bench_   = true;
+    this->verify_  = false;
+    this->Run();
+}
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+TEST(TestBatchedGemmSoftmaxGemmInterface, GemmSpecializationSizeMatch)
+{
+    int P = 120; // requires padding
+    int Q = 128; // do not require padding
+
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // clang-format on
+}
+
+TEST(TestBatchedGemmSoftmaxGemmInterface, GemmSpecializationSizeMismatch)
+{
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // clang-format on
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, AdhocTest)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {49, 49, 64, 64, 24},
+        {64, 49, 64, 64, 24},
+        {1020, 1020, 64, 128, 24},
+        {576, 576, 64, 64, 24},
+    };
+    this->Run();
+}
diff --git a/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp
new file mode 100644
index 0000000..98debe1
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp
@@ -0,0 +1,197 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include <vector>
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
+#include "profiler/profile_batched_gemm_softmax_gemm_impl.hpp"
+using ck::tensor_operation::device::GemmSpecialization;
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+
+using F16 = ck::half_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <typename Tuple>
+struct TestBatchedGemmSoftmaxGemm : public ::testing::Test
+{
+    using ADataType   = std::tuple_element_t<0, Tuple>;
+    using B0DataType  = std::tuple_element_t<1, Tuple>;
+    using B1DataType  = std::tuple_element_t<2, Tuple>;
+    using CDataType   = std::tuple_element_t<3, Tuple>;
+    using ALayout     = std::tuple_element_t<4, Tuple>;
+    using B0Layout    = std::tuple_element_t<5, Tuple>;
+    using B1Layout    = std::tuple_element_t<6, Tuple>;
+    using CLayout     = std::tuple_element_t<7, Tuple>;
+    using MaskingType = std::tuple_element_t<8, Tuple>;
+
+    std::vector<std::vector<int>> lengths_ = {{256, 256, 64, 64, 4},
+                                              {256, 256, 128, 128, 4},
+                                              {512, 512, 64, 64, 2},
+                                              {512, 512, 128, 128, 2},
+                                              {1024, 1024, 64, 64, 1},
+                                              {1024, 1024, 128, 128, 1},
+                                              {256, 256, 160, 160, 4},
+                                              {256, 64, 160, 64, 4},
+                                              {1024, 1024, 80, 80, 2},
+                                              {1024, 64, 80, 64, 2},
+                                              {4096, 4096, 40, 40, 1},
+                                              {4096, 64, 40, 64, 1}};
+
+    bool bench_  = false;
+    bool verify_ = true;
+
+    void RunSingle(int M, int N, int K, int O, int BatchCount)
+    {
+        bool pass = ck::profiler::profile_batched_gemm_softmax_gemm_impl<ADataType,
+                                                                         B0DataType,
+                                                                         B1DataType,
+                                                                         CDataType,
+                                                                         ALayout,
+                                                                         B0Layout,
+                                                                         B1Layout,
+                                                                         CLayout,
+                                                                         MaskingType::value>(
+            verify_, 1, false, bench_, M, N, K, O, BatchCount);
+
+        EXPECT_TRUE(pass);
+    }
+
+    void Run()
+    {
+        for(auto lengths : this->lengths_)
+        {
+            int M          = lengths[0];
+            int N          = lengths[1];
+            int K          = lengths[2];
+            int O          = lengths[3];
+            int BatchCount = lengths[4];
+
+            this->RunSingle(M, N, K, O, BatchCount);
+        }
+    }
+};
+
+template <GemmSpecialization GemmSpec>
+struct DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    using ALayout  = Row;
+    using B0Layout = Col;
+    using B1Layout = Row;
+    using CLayout  = Row;
+
+    using ADataType        = F16;
+    using B0DataType       = F16;
+    using B1DataType       = F16;
+    using AccDataType      = float;
+    using CShuffleDataType = float;
+    using CDataType        = F16;
+
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = PassThrough;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    // static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
+
+    using DeviceGemmGemmInstance =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+            ALayout,
+            B0Layout,
+            B1Layout,
+            CLayout,
+            ADataType,
+            B0DataType,
+            B1DataType,
+            CDataType,
+            AccDataType,
+            CShuffleDataType,
+            AElementOp,
+            B0ElementOp,
+            Acc0ElementOp,
+            B1ElementOp,
+            CElementOp,
+            GemmSpec,
+            1,
+            256,
+            128,         // MPerBlock
+            128,         // NPerBlock
+            32,          // KPerBlock
+            128,         // Gemm1NPerBlock
+            32,          // Gemm1KPerBlock
+            8,           // AK1
+            8,           // BK1
+            2,           // B1K1
+            32,          // MPerXDL
+            32,          // NPerXDL
+            1,           // MXdlPerWave
+            4,           // NXdlPerWave
+            4,           // Gemm1NXdlPerWave
+            S<4, 64, 1>, // ABlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<4, 64, 1>, // BBlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<8, 32, 1>, // B1BlockTransfer
+            S<0, 2, 1>,
+            S<0, 2, 1>,
+            1,
+            4,
+            2,
+            false,
+            1,              // CShuffleMXdlPerWavePerShuffle
+            2,              // CShuffleNXdlPerWavePerShuffle
+            S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+            8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+            false>;
+
+    bool IsSupported(int M, int N, int K, int O)
+    {
+        auto gemm     = DeviceGemmGemmInstance{};
+        auto invoker  = gemm.MakeInvoker();
+        auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
+                                          static_cast<B0DataType*>(nullptr),
+                                          static_cast<B1DataType*>(nullptr),
+                                          static_cast<CDataType*>(nullptr),
+                                          M,
+                                          N,
+                                          K,
+                                          O,
+                                          0,              // BatchCount
+                                          0,              // StrideA
+                                          0,              // StrideB0
+                                          0,              // StrideB1
+                                          0,              // StrideC
+                                          0,              // BatchStrideA
+                                          0,              // BatchStrideB0
+                                          0,              // BatchStrideB1
+                                          0,              // BatchStrideC
+                                          PassThrough{},  // a_element_op
+                                          PassThrough{},  // b0_element_op
+                                          PassThrough{},  // acc0_element_op
+                                          PassThrough{},  // b1_element_op
+                                          PassThrough{}); // c_element_op
+
+        return gemm.IsSupportedArgument(argument);
+    }
+};
diff --git a/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
new file mode 100644
index 0000000..f858d9f
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
@@ -0,0 +1,8 @@
+add_custom_target(test_batched_gemm_softmax_gemm_permute)
+
+add_gtest_executable(test_batched_gemm_softmax_gemm_permute_fp16 test_batched_gemm_softmax_gemm_permute_fp16.cpp)
+add_gtest_executable(test_batched_gemm_softmax_gemm_permute_bf16 test_batched_gemm_softmax_gemm_permute_bf16.cpp)
+target_link_libraries(test_batched_gemm_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
+target_link_libraries(test_batched_gemm_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
+add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_fp16)
+add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_bf16)
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp
new file mode 100644
index 0000000..4a775e6
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp
@@ -0,0 +1,182 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_softmax_gemm_permute_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16
+    : public TestBatchedGemmMaskingScaleSoftmaxGemmPermute<Tuple>
+{
+};
+
+using I1_t = ck::Number<1>;
+using I2_t = ck::Number<2>;
+
+using MaskDisabled_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskDisabled>;
+using MaskOutUpperTriangle_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskOutUpperTriangle>;
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, MaskDisabled_t>,
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, MaskOutUpperTriangle_t>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, DISABLED_Test_BF16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {136, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 136, 32, 128, 3, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 40, 128, 2, 4},
+        {128, 128, 136, 128, 4, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 136, 1, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {129, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 129, 32, 128, 4, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 33, 128, 2, 3},
+        {128, 128, 129, 128, 2, 3},
+    };
+    this->Run();
+}
+
+// If kernel B1Layout is RowMajor, expect not to support odd O size
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 129, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, DISABLED_Bench_BF16_IrregularK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{{256, 256, 160, 160, 1, 16},
+                                                   {256, 64, 160, 64, 1, 16},
+                                                   {1024, 1024, 80, 80, 1, 16},
+                                                   {1024, 64, 80, 64, 1, 16},
+                                                   {4096, 4096, 40, 40, 1, 16},
+                                                   {4096, 64, 40, 64, 1, 16}};
+    this->bench_   = true;
+    this->verify_  = false;
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, DISABLED_Bench_BF16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 48, 16},
+        {256, 256, 128, 128, 48, 16},
+        {512, 512, 64, 64, 48, 16},
+        {512, 512, 128, 128, 48, 16},
+        {1024, 1024, 64, 64, 48, 16},
+        {1024, 1024, 128, 128, 48, 16},
+        {2048, 2048, 64, 64, 48, 16},
+        {2048, 2048, 128, 128, 48, 16},
+        {4096, 4096, 64, 64, 48, 16},
+        {4096, 4096, 128, 128, 48, 16},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMatch)
+{
+    int P = 120; // requires padding
+    int Q = 128; // do not require padding
+
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // clang-format on
+}
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMismatch)
+{
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // clang-format on
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, AdhocTest)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {49, 49, 64, 64, 4, 6},
+        {64, 49, 64, 64, 4, 6},
+        {1020, 1020, 64, 128, 4, 6},
+        {576, 576, 64, 64, 4, 6},
+    };
+    this->Run();
+}
diff --git a/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp
new file mode 100644
index 0000000..293acd6
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp
@@ -0,0 +1,182 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_softmax_gemm_permute_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16
+    : public TestBatchedGemmMaskingScaleSoftmaxGemmPermute<Tuple>
+{
+};
+
+using I1_t = ck::Number<1>;
+using I2_t = ck::Number<2>;
+
+using MaskDisabled_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskDisabled>;
+using MaskOutUpperTriangle_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskOutUpperTriangle>;
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, MaskDisabled_t>,
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, MaskOutUpperTriangle_t>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {136, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 136, 32, 128, 3, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 40, 128, 2, 4},
+        {128, 128, 136, 128, 4, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 136, 1, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {129, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 129, 32, 128, 4, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 33, 128, 2, 3},
+        {128, 128, 129, 128, 2, 3},
+    };
+    this->Run();
+}
+
+// If kernel B1Layout is RowMajor, expect not to support odd O size
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 129, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, DISABLED_Bench_FP16_IrregularK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{{256, 256, 160, 160, 1, 16},
+                                                   {256, 64, 160, 64, 1, 16},
+                                                   {1024, 1024, 80, 80, 1, 16},
+                                                   {1024, 64, 80, 64, 1, 16},
+                                                   {4096, 4096, 40, 40, 1, 16},
+                                                   {4096, 64, 40, 64, 1, 16}};
+    this->bench_   = true;
+    this->verify_  = false;
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, DISABLED_Bench_FP16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 48, 16},
+        {256, 256, 128, 128, 48, 16},
+        {512, 512, 64, 64, 48, 16},
+        {512, 512, 128, 128, 48, 16},
+        {1024, 1024, 64, 64, 48, 16},
+        {1024, 1024, 128, 128, 48, 16},
+        {2048, 2048, 64, 64, 48, 16},
+        {2048, 2048, 128, 128, 48, 16},
+        {4096, 4096, 64, 64, 48, 16},
+        {4096, 4096, 128, 128, 48, 16},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMatch)
+{
+    int P = 120; // requires padding
+    int Q = 128; // do not require padding
+
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // clang-format on
+}
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMismatch)
+{
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // clang-format on
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, AdhocTest)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {49, 49, 64, 64, 4, 6},
+        {64, 49, 64, 64, 4, 6},
+        {1020, 1020, 64, 128, 4, 6},
+        {576, 576, 64, 64, 4, 6},
+    };
+    this->Run();
+}
diff --git a/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp
new file mode 100644
index 0000000..912bbc9
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp
@@ -0,0 +1,368 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include <vector>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp"
+
+using ck::tensor_operation::device::GemmSpecialization;
+using ck::tensor_operation::device::MaskingSpecialization;
+using ck::tensor_operation::device::TensorSpecialization;
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <typename Tuple>
+struct TestBatchedGemmMaskingScaleSoftmaxGemmPermute : public ::testing::Test
+{
+    using NumDimGType      = std::tuple_element_t<0, Tuple>;
+    using NumDimMType      = std::tuple_element_t<1, Tuple>;
+    using NumDimNType      = std::tuple_element_t<2, Tuple>;
+    using NumDimKType      = std::tuple_element_t<3, Tuple>;
+    using NumDimOType      = std::tuple_element_t<4, Tuple>;
+    using ADataType        = std::tuple_element_t<5, Tuple>;
+    using B0DataType       = std::tuple_element_t<6, Tuple>;
+    using B1DataType       = std::tuple_element_t<7, Tuple>;
+    using CDataType        = std::tuple_element_t<8, Tuple>;
+    using Acc0BiasDataType = std::tuple_element_t<9, Tuple>;
+    using Acc1BiasDataType = std::tuple_element_t<10, Tuple>;
+    using MaskingType      = std::tuple_element_t<11, Tuple>;
+
+    std::vector<std::vector<int>> lengths_ = {
+        {256, 256, 64, 64, 6, 4},
+        {256, 256, 128, 128, 4, 6},
+        {512, 512, 64, 64, 3, 2},
+        {512, 512, 128, 128, 2, 3},
+        {1024, 1024, 64, 64, 3, 1},
+        {1024, 1024, 128, 128, 1, 1},
+    };
+    bool bench_  = false;
+    bool verify_ = true;
+
+    void RunSingle(int M, int N, int K, int O, int G0, int G1)
+    {
+        bool pass =
+            ck::profiler::profile_batched_gemm_softmax_gemm_permute_impl<NumDimGType::value,
+                                                                         NumDimMType::value,
+                                                                         NumDimNType::value,
+                                                                         NumDimKType::value,
+                                                                         NumDimOType::value,
+                                                                         ADataType,
+                                                                         B0DataType,
+                                                                         B1DataType,
+                                                                         CDataType,
+                                                                         ck::Tuple<>,
+                                                                         ck::Tuple<>,
+                                                                         MaskingType::value>(
+                verify_, 2, false, bench_, M, N, K, O, G0, G1);
+
+        EXPECT_TRUE(pass);
+    }
+
+    void Run()
+    {
+        for(auto lengths : this->lengths_)
+        {
+            int M  = lengths[0];
+            int N  = lengths[1];
+            int K  = lengths[2];
+            int O  = lengths[3];
+            int G0 = lengths[4];
+            int G1 = lengths[5];
+
+            this->RunSingle(M, N, K, O, G0, G1);
+        }
+    }
+};
+
+template <GemmSpecialization GemmSpec>
+struct DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using Scale       = ck::tensor_operation::element_wise::Scale;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    using ADataType        = F16;
+    using B0DataType       = F16;
+    using B1DataType       = F16;
+    using AccDataType      = float;
+    using CShuffleDataType = F16;
+    using CDataType        = F16;
+
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = Scale;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+
+    // static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
+
+    using DeviceGemmGemmInstance =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+            2,
+            1,
+            1,
+            1,
+            1,
+            ADataType,
+            B0DataType,
+            B1DataType,
+            CDataType,
+            ck::Tuple<>,
+            ck::Tuple<>,
+            AccDataType,
+            CShuffleDataType,
+            AElementOp,
+            B0ElementOp,
+            Acc0ElementOp,
+            B1ElementOp,
+            CElementOp,
+            GemmSpec,
+            TensorSpecialization::Default, // ATensorSpec
+            TensorSpecialization::Default, // B0TensorSpec
+            TensorSpecialization::Default, // B1TensorSpec
+            TensorSpecialization::Default, // CTensorSpec
+            1,
+            256,
+            128,         // MPerBlock
+            128,         // NPerBlock
+            32,          // KPerBlock
+            128,         // Gemm1NPerBlock
+            32,          // Gemm1KPerBlock
+            8,           // AK1
+            8,           // BK1
+            2,           // B1K1
+            32,          // MPerXDL
+            32,          // NPerXDL
+            1,           // MXdlPerWave
+            4,           // NXdlPerWave
+            4,           // Gemm1NXdlPerWave
+            S<4, 64, 1>, // ABlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<4, 64, 1>, // BBlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<8, 32, 1>, // B1BlockTransfer
+            S<0, 2, 1>,
+            S<0, 2, 1>,
+            1,
+            4,
+            2,
+            false,
+            1,              // CShuffleMXdlPerWavePerShuffle
+            2,              // CShuffleNXdlPerWavePerShuffle
+            S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+            8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+            MaskingSpecialization::MaskOutUpperTriangle>; // MaskOutUpperTriangle
+
+    bool IsSupported(int M, int N, int K, int O)
+    {
+        const int G0 = 1, G1 = 1;
+
+        // A layout [G0, M, G1, K]
+        std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+        std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+        // B0 layout [G0, N, G1, K]
+        std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+        std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+        // B1 layout [G0, N, G1, O]
+        std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+        std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+        // C layout [G0, M, G1, O]
+        std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+        std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+        auto gemm     = DeviceGemmGemmInstance{};
+        auto invoker  = gemm.MakeInvoker();
+        auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
+                                          static_cast<B0DataType*>(nullptr),
+                                          static_cast<B1DataType*>(nullptr),
+                                          static_cast<CDataType*>(nullptr),
+                                          {}, // p_acc0_biases
+                                          {}, // p_acc1_biases
+                                          a_gs_ms_ks_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b0_gs_ns_ks_lengths,
+                                          b0_gs_ns_ks_strides,
+                                          b1_gs_os_ns_lengths,
+                                          b1_gs_os_ns_strides,
+                                          c_gs_ms_os_lengths,
+                                          c_gs_ms_os_strides,
+                                          {},             // acc0_biases_gs_ms_ns_lengths
+                                          {},             // acc0_biases_gs_ms_ns_strides
+                                          {},             // acc1_biases_gs_ms_os_lengths
+                                          {},             // acc1_biases_gs_ms_os_strides
+                                          PassThrough{},  // a_element_op
+                                          PassThrough{},  // b0_element_op
+                                          Scale{1.f},     // acc0_element_op
+                                          PassThrough{},  // b1_element_op
+                                          PassThrough{}); // c_element_op
+
+        return gemm.IsSupportedArgument(argument);
+    }
+};
+
+template <GemmSpecialization GemmSpec>
+struct DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using Scale       = ck::tensor_operation::element_wise::Scale;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    using ADataType        = BF16;
+    using B0DataType       = BF16;
+    using B1DataType       = BF16;
+    using AccDataType      = float;
+    using CShuffleDataType = BF16;
+    using CDataType        = BF16;
+
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = Scale;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+
+    // static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
+
+    using DeviceGemmGemmInstance =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+            2,
+            1,
+            1,
+            1,
+            1,
+            ADataType,
+            B0DataType,
+            B1DataType,
+            CDataType,
+            ck::Tuple<>,
+            ck::Tuple<>,
+            AccDataType,
+            CShuffleDataType,
+            AElementOp,
+            B0ElementOp,
+            Acc0ElementOp,
+            B1ElementOp,
+            CElementOp,
+            GemmSpec,
+            TensorSpecialization::Default, // ATensorSpec
+            TensorSpecialization::Default, // B0TensorSpec
+            TensorSpecialization::Default, // B1TensorSpec
+            TensorSpecialization::Default, // CTensorSpec
+            1,
+            256,
+            128,         // MPerBlock
+            128,         // NPerBlock
+            32,          // KPerBlock
+            128,         // Gemm1NPerBlock
+            32,          // Gemm1KPerBlock
+            8,           // AK1
+            8,           // BK1
+            2,           // B1K1
+            32,          // MPerXDL
+            32,          // NPerXDL
+            1,           // MXdlPerWave
+            4,           // NXdlPerWave
+            4,           // Gemm1NXdlPerWave
+            S<4, 64, 1>, // ABlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<4, 64, 1>, // BBlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<8, 32, 1>, // B1BlockTransfer
+            S<0, 2, 1>,
+            S<0, 2, 1>,
+            1,
+            4,
+            2,
+            false,
+            1,              // CShuffleMXdlPerWavePerShuffle
+            2,              // CShuffleNXdlPerWavePerShuffle
+            S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+            8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+            MaskingSpecialization::MaskOutUpperTriangle>; // MaskOutUpperTriangle
+
+    bool IsSupported(int M, int N, int K, int O)
+    {
+        const int G0 = 1, G1 = 1;
+
+        // A layout [G0, M, G1, K]
+        std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+        std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+        // B0 layout [G0, N, G1, K]
+        std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+        std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+        // B1 layout [G0, N, G1, O]
+        std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+        std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+        // C layout [G0, M, G1, O]
+        std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+        std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+        auto gemm     = DeviceGemmGemmInstance{};
+        auto invoker  = gemm.MakeInvoker();
+        auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
+                                          static_cast<B0DataType*>(nullptr),
+                                          static_cast<B1DataType*>(nullptr),
+                                          static_cast<CDataType*>(nullptr),
+                                          {}, // p_acc0_biases
+                                          {}, // p_acc1_biases
+                                          a_gs_ms_ks_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b0_gs_ns_ks_lengths,
+                                          b0_gs_ns_ks_strides,
+                                          b1_gs_os_ns_lengths,
+                                          b1_gs_os_ns_strides,
+                                          c_gs_ms_os_lengths,
+                                          c_gs_ms_os_strides,
+                                          {},             // acc0_biases_gs_ms_ns_lengths
+                                          {},             // acc0_biases_gs_ms_ns_strides
+                                          {},             // acc1_biases_gs_ms_os_lengths
+                                          {},             // acc1_biases_gs_ms_os_strides
+                                          PassThrough{},  // a_element_op
+                                          PassThrough{},  // b0_element_op
+                                          Scale{1.f},     // acc0_element_op
+                                          PassThrough{},  // b1_element_op
+                                          PassThrough{}); // c_element_op
+
+        return gemm.IsSupportedArgument(argument);
+    }
+};
diff --git a/3rdparty/composable_kernel/test/batchnorm/CMakeLists.txt b/3rdparty/composable_kernel/test/batchnorm/CMakeLists.txt
new file mode 100644
index 0000000..52f1508
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batchnorm/CMakeLists.txt
@@ -0,0 +1,4 @@
+add_gtest_executable(test_batchnorm_fwd_rank_4 batchnorm_fwd_rank_4.cpp)
+add_gtest_executable(test_batchnorm_bwd_rank_4 batchnorm_bwd_rank_4.cpp)
+target_link_libraries(test_batchnorm_fwd_rank_4 PRIVATE utility device_batchnorm_instance)
+target_link_libraries(test_batchnorm_bwd_rank_4 PRIVATE utility device_batchnorm_instance)
diff --git a/3rdparty/composable_kernel/test/batchnorm/batchnorm_bwd_rank_4.cpp b/3rdparty/composable_kernel/test/batchnorm/batchnorm_bwd_rank_4.cpp
new file mode 100644
index 0000000..caa7331
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batchnorm/batchnorm_bwd_rank_4.cpp
@@ -0,0 +1,92 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <tuple>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_batchnorm_backward_impl.hpp"
+
+using F16  = ck::half_t;
+using F32  = float;
+using BF16 = ck::bhalf_t;
+using F64  = double;
+
+template <typename Tuple>
+class TestBatchNormBwdRank4 : public ::testing::Test
+{
+    private:
+    const double epsilon = std::numeric_limits<float>::epsilon();
+
+    protected:
+    using XDataType       = std::tuple_element_t<0, Tuple>;
+    using DxDataType      = std::tuple_element_t<1, Tuple>;
+    using DyDataType      = std::tuple_element_t<2, Tuple>;
+    using AccDataType     = std::tuple_element_t<3, Tuple>;
+    using ScaleDataType   = std::tuple_element_t<4, Tuple>;
+    using BiasDataType    = std::tuple_element_t<5, Tuple>;
+    using MeanVarDataType = std::tuple_element_t<6, Tuple>;
+
+    std::vector<std::vector<size_t>> list_of_lengths = {
+        {128, 16, 3, 1024}, {128, 16, 6, 512}, {1, 1, 1, 1}, {4, 4, 4, 4}, {32, 32, 32, 32}};
+    std::vector<int> reduceDims;
+
+    template <int NumReduceDim>
+    void Run()
+    {
+        for(auto& inOutLengths : list_of_lengths)
+        {
+            bool pass = true;
+
+            EXPECT_FALSE(reduceDims.size() != NumReduceDim);
+
+            pass = pass && ck::profiler::profile_batchnorm_backward_impl<XDataType,
+                                                                         DxDataType,
+                                                                         DyDataType,
+                                                                         AccDataType,
+                                                                         ScaleDataType,
+                                                                         BiasDataType,
+                                                                         MeanVarDataType,
+                                                                         4,
+                                                                         NumReduceDim>(
+                               true, 3, false, false, inOutLengths, reduceDims, true, epsilon);
+
+            pass = pass && ck::profiler::profile_batchnorm_backward_impl<XDataType,
+                                                                         DxDataType,
+                                                                         DyDataType,
+                                                                         AccDataType,
+                                                                         ScaleDataType,
+                                                                         BiasDataType,
+                                                                         MeanVarDataType,
+                                                                         4,
+                                                                         NumReduceDim>(
+                               true, 3, false, false, inOutLengths, reduceDims, false, epsilon);
+
+            EXPECT_TRUE(pass);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<std::tuple<F16, F32, F32, F32, F16, F32, F32>,
+                                     std::tuple<F32, F32, F32, F32, F32, F32, F32>,
+                                     std::tuple<BF16, F32, F32, F32, BF16, F32, F32>,
+                                     std::tuple<F64, F64, F64, F64, F64, F64, F64>>;
+
+TYPED_TEST_SUITE(TestBatchNormBwdRank4, KernelTypes);
+
+// nhwc
+TYPED_TEST(TestBatchNormBwdRank4, nhwc)
+{
+    this->reduceDims = {0, 1, 2};
+    this->template Run<3>();
+}
+
+// nchw
+TYPED_TEST(TestBatchNormBwdRank4, nchw)
+{
+    this->reduceDims = {0, 2, 3};
+    this->template Run<3>();
+}
diff --git a/3rdparty/composable_kernel/test/batchnorm/batchnorm_fwd_rank_4.cpp b/3rdparty/composable_kernel/test/batchnorm/batchnorm_fwd_rank_4.cpp
new file mode 100644
index 0000000..13aef7d
--- /dev/null
+++ b/3rdparty/composable_kernel/test/batchnorm/batchnorm_fwd_rank_4.cpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <tuple>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_batchnorm_forward_impl.hpp"
+
+using F16  = ck::half_t;
+using F32  = float;
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F64  = double;
+
+template <typename Tuple>
+class TestBatchNormFwdRank4 : public ::testing::Test
+{
+    private:
+    const double epsilon       = std::numeric_limits<float>::epsilon();
+    const double averageFactor = 0.1;
+
+    protected:
+    using XDataType       = std::tuple_element_t<0, Tuple>;
+    using YDataType       = std::tuple_element_t<1, Tuple>;
+    using AccDataType     = std::tuple_element_t<2, Tuple>;
+    using ScaleDataType   = std::tuple_element_t<3, Tuple>;
+    using BiasDataType    = std::tuple_element_t<4, Tuple>;
+    using MeanVarDataType = std::tuple_element_t<5, Tuple>;
+
+    std::vector<std::vector<size_t>> list_of_lengths = {
+        {128, 16, 3, 1024}, {128, 16, 6, 512}, {1, 1, 1, 1}, {4, 4, 4, 4}, {32, 32, 32, 32}};
+    std::vector<int> reduceDims;
+
+    template <int NumReduceDim>
+    void Run()
+    {
+        for(auto& inOutLengths : list_of_lengths)
+        {
+            bool pass = true;
+
+            EXPECT_FALSE(reduceDims.size() != NumReduceDim);
+
+            pass =
+                pass && ck::profiler::profile_batchnorm_forward_impl<XDataType,
+                                                                     YDataType,
+                                                                     AccDataType,
+                                                                     ScaleDataType,
+                                                                     BiasDataType,
+                                                                     MeanVarDataType,
+                                                                     4,
+                                                                     NumReduceDim>(true,
+                                                                                   3,
+                                                                                   false,
+                                                                                   false,
+                                                                                   inOutLengths,
+                                                                                   reduceDims,
+                                                                                   true,
+                                                                                   true,
+                                                                                   epsilon,
+                                                                                   averageFactor);
+
+            pass =
+                pass && ck::profiler::profile_batchnorm_forward_impl<XDataType,
+                                                                     YDataType,
+                                                                     AccDataType,
+                                                                     ScaleDataType,
+                                                                     BiasDataType,
+                                                                     MeanVarDataType,
+                                                                     4,
+                                                                     NumReduceDim>(true,
+                                                                                   3,
+                                                                                   false,
+                                                                                   false,
+                                                                                   inOutLengths,
+                                                                                   reduceDims,
+                                                                                   false,
+                                                                                   false,
+                                                                                   epsilon,
+                                                                                   averageFactor);
+
+            EXPECT_TRUE(pass);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<std::tuple<F16, F16, F32, F16, F16, F32>,
+                                     std::tuple<F32, F32, F32, F32, F32, F32>,
+                                     std::tuple<BF16, BF16, F32, BF16, BF16, F32>,
+                                     std::tuple<F64, F64, F64, F64, F64, F64>>;
+
+TYPED_TEST_SUITE(TestBatchNormFwdRank4, KernelTypes);
+
+// nhwc
+TYPED_TEST(TestBatchNormFwdRank4, nhwc)
+{
+    this->reduceDims = {0, 1, 2};
+    this->template Run<3>();
+}
+
+// nchw
+TYPED_TEST(TestBatchNormFwdRank4, nchw)
+{
+    this->reduceDims = {0, 2, 3};
+    this->template Run<3>();
+}
diff --git a/3rdparty/composable_kernel/test/block_to_ctile_map/CMakeLists.txt b/3rdparty/composable_kernel/test/block_to_ctile_map/CMakeLists.txt
new file mode 100644
index 0000000..97dfbb2
--- /dev/null
+++ b/3rdparty/composable_kernel/test/block_to_ctile_map/CMakeLists.txt
@@ -0,0 +1 @@
+add_gtest_executable(test_block_to_ctile_map test_block_to_ctile_map.cpp)
\ No newline at end of file
diff --git a/3rdparty/composable_kernel/test/block_to_ctile_map/test_block_to_ctile_map.cpp b/3rdparty/composable_kernel/test/block_to_ctile_map/test_block_to_ctile_map.cpp
new file mode 100644
index 0000000..55d9b59
--- /dev/null
+++ b/3rdparty/composable_kernel/test/block_to_ctile_map/test_block_to_ctile_map.cpp
@@ -0,0 +1,322 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+
+using namespace ck;
+
+static auto I0 = Number<0>{};
+static auto I1 = Number<1>{};
+static auto I2 = Number<2>{};
+
+TEST(BlockToCTileMap, TestBlockToCTileMap_M00_N00_M01_N01_DeviceCTileIndexCheck1)
+{
+    const index_t M         = 384;
+    const index_t N         = 384;
+    const index_t MPerBlock = 128;
+    const index_t NPerBlock = 128;
+    const index_t MBlock    = M / MPerBlock;
+    const index_t NBlock    = N / NPerBlock;
+    const index_t M01       = 4;
+    const index_t N01       = 4;
+
+    auto c_grid_desc_m_n = make_naive_tensor_descriptor_packed(make_tuple(M, N));
+
+    printf("(M, N, MPerBlock, NPerBlock, M01, N01) = (%d, %d, %d, %d, %d, %d)\n",
+           M,
+           N,
+           MPerBlock,
+           NPerBlock,
+           M01,
+           N01);
+
+    BlockToCTileMap_M00_N00_M01_N01<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n), true> tile_map(
+        c_grid_desc_m_n, M01, N01);
+
+    EXPECT_TRUE(tile_map.CheckValidity(c_grid_desc_m_n) == true);
+    EXPECT_TRUE(tile_map.CalculateGridSize(c_grid_desc_m_n) == 16);
+
+    // clang-format off
+    std::vector<std::vector<int>> expected_m0idx_n0idx_valid = {
+        {0, 0, 1},
+        {0, 1, 1},
+        {0, 2, 1},
+        {0, 3, 0},
+        {1, 0, 1},
+        {1, 1, 1},
+        {1, 2, 1},
+        {1, 3, 0},
+        {2, 0, 1},
+        {2, 1, 1},
+        {2, 2, 1},
+        {2, 3, 0},
+        {3, 0, 0},
+        {3, 1, 0},
+        {3, 2, 0},
+        {3, 3, 0}
+    };
+    // clang-format on
+
+    for(index_t i = 0; i < tile_map.CalculateGridSize(c_grid_desc_m_n); i++)
+    {
+        auto m0n0_idx = tile_map.CalculateBottomIndex(make_multi_index(i));
+        std::cout << "block_1d_id = " << i << ", m0, n0 = " << m0n0_idx[I0] << ", " << m0n0_idx[I1];
+        std::cout << ", valid = " << tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))
+                  << std::endl;
+        bool equal =
+            expected_m0idx_n0idx_valid[i] ==
+            std::vector<int>{m0n0_idx[I0],
+                             m0n0_idx[I1],
+                             tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))};
+        EXPECT_TRUE(equal);
+    }
+}
+
+TEST(BlockToCTileMap, TestBlockToCTileMap_M00_N00_M01_N01_DeviceCTileIndexCheck0)
+{
+    const index_t M         = 384;
+    const index_t N         = 384;
+    const index_t MPerBlock = 128;
+    const index_t NPerBlock = 128;
+
+    const index_t M01 = 4;
+    const index_t N01 = 4;
+
+    auto c_grid_desc_m_n = make_naive_tensor_descriptor_packed(make_tuple(M, N));
+
+    printf("(M, N, MPerBlock, NPerBlock, M01, N01) = (%d, %d, %d, %d, %d, %d)\n",
+           M,
+           N,
+           MPerBlock,
+           NPerBlock,
+           M01,
+           N01);
+
+    BlockToCTileMap_M00_N00_M01_N01<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n), false>
+        tile_map(c_grid_desc_m_n, M01, N01);
+
+    EXPECT_TRUE(tile_map.CheckValidity(c_grid_desc_m_n) == false);
+}
+
+TEST(BlockToCTileMap, TestBlockToCTileMap_M00_N0_M01_DeviceCTileIndexCheck1)
+{
+    const index_t M         = 384;
+    const index_t N         = 512;
+    const index_t MPerBlock = 128;
+    const index_t NPerBlock = 128;
+    const index_t MBlock    = M / MPerBlock;
+    const index_t NBlock    = N / NPerBlock;
+    const index_t M01       = 4;
+
+    auto c_grid_desc_m_n = make_naive_tensor_descriptor_packed(make_tuple(M, N));
+
+    printf("(M, N, MPerBlock, NPerBlock, M01) = (%d, %d, %d, %d, %d)\n",
+           M,
+           N,
+           MPerBlock,
+           NPerBlock,
+           M01);
+
+    BlockToCTileMap_M00_N0_M01<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n), true> tile_map(
+        c_grid_desc_m_n, M01);
+
+    EXPECT_TRUE(tile_map.CheckValidity(c_grid_desc_m_n) == true);
+    EXPECT_TRUE(tile_map.CalculateGridSize(c_grid_desc_m_n) == 16);
+
+    // clang-format off
+    std::vector<std::vector<int>> expected_m0idx_n0idx_valid = {
+        {0, 0, 1},
+        {1, 0, 1},
+        {2, 0, 1},
+        {3, 0, 0},
+        {0, 1, 1},
+        {1, 1, 1},
+        {2, 1, 1},
+        {3, 1, 0},
+        {0, 2, 1},
+        {1, 2, 1},
+        {2, 2, 1},
+        {3, 2, 0},
+        {0, 3, 1},
+        {1, 3, 1},
+        {2, 3, 1},
+        {3, 3, 0}
+    };
+    // clang-format on
+
+    for(index_t i = 0; i < tile_map.CalculateGridSize(c_grid_desc_m_n); i++)
+    {
+        auto m0n0_idx = tile_map.CalculateBottomIndex(make_multi_index(i));
+        std::cout << "block_1d_id = " << i << ", m0, n0 = " << m0n0_idx[I0] << ", " << m0n0_idx[I1];
+        std::cout << ", valid = " << tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))
+                  << std::endl;
+        bool equal =
+            expected_m0idx_n0idx_valid[i] ==
+            std::vector<int>{m0n0_idx[I0],
+                             m0n0_idx[I1],
+                             tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))};
+        EXPECT_TRUE(equal);
+    }
+}
+
+TEST(BlockToCTileMap, TestBlockToCTileMap_M00_N0_M01_DeviceCTileIndexCheck0)
+{
+    const index_t M         = 512;
+    const index_t N         = 384;
+    const index_t MPerBlock = 128;
+    const index_t NPerBlock = 128;
+
+    auto c_grid_desc_m_n = make_naive_tensor_descriptor_packed(make_tuple(M, N));
+
+    // clang-format off
+    std::vector<std::tuple<int, int, bool>> expected_m0_gridsize_validity = {
+        {5, 15, false},
+        {4, 12, true},
+        {3, 18, false},
+        {2, 12, true},
+        {1, 12, true}
+    };
+    // clang-format on
+
+    for(auto e : expected_m0_gridsize_validity)
+    {
+        const index_t M01 = std::get<0>(e);
+
+        printf("(M, N, MPerBlock, NPerBlock, M01) = (%d, %d, %d, %d, %d)\n",
+               M,
+               N,
+               MPerBlock,
+               NPerBlock,
+               M01);
+
+        BlockToCTileMap_M00_N0_M01<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n), false> tile_map(
+            c_grid_desc_m_n, M01);
+
+        EXPECT_EQ(tile_map.CalculateGridSize(c_grid_desc_m_n), std::get<1>(e));
+        EXPECT_EQ(tile_map.CheckValidity(c_grid_desc_m_n), std::get<2>(e));
+    }
+}
+
+TEST(BlockToCTileMap, TestBlockToCTileMap_M00_N0_M01Adapt)
+{
+    const index_t M         = 768;
+    const index_t N         = 384;
+    const index_t MPerBlock = 128;
+    const index_t NPerBlock = 128;
+    const index_t MBlock    = M / MPerBlock;
+    const index_t NBlock    = N / NPerBlock;
+    constexpr index_t M01   = 4;
+
+    auto c_grid_desc_m_n = make_naive_tensor_descriptor_packed(make_tuple(M, N));
+
+    printf("(M, N, MPerBlock, NPerBlock, M01) = (%d, %d, %d, %d, %d)\n",
+           M,
+           N,
+           MPerBlock,
+           NPerBlock,
+           M01);
+
+    BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n)> tile_map(
+        c_grid_desc_m_n, M01);
+
+    EXPECT_TRUE(tile_map.CheckValidity(c_grid_desc_m_n) == true);
+    EXPECT_TRUE(tile_map.CalculateGridSize(c_grid_desc_m_n) == 18);
+
+    // clang-format off
+    std::vector<std::vector<int>> expected_m0idx_n0idx_valid = {
+        {0, 0, 1},
+        {1, 0, 1},
+        {2, 0, 1},
+        {3, 0, 1},
+        {0, 1, 1},
+        {1, 1, 1},
+        {2, 1, 1},
+        {3, 1, 1},
+        {0, 2, 1},
+        {1, 2, 1},
+        {2, 2, 1},
+        {3, 2, 1},
+        {4, 0, 1},
+        {5, 0, 1},
+        {4, 1, 1},
+        {5, 1, 1},
+        {4, 2, 1},
+        {5, 2, 1},
+    };
+    // clang-format on
+
+    for(index_t i = 0; i < tile_map.CalculateGridSize(c_grid_desc_m_n); i++)
+    {
+        auto m0n0_idx = tile_map.CalculateBottomIndex(make_multi_index(i));
+        std::cout << "block_1d_id = " << i << ", m0, n0 = " << m0n0_idx[I0] << ", " << m0n0_idx[I1];
+        std::cout << ", valid = " << tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))
+                  << std::endl;
+        bool equal =
+            expected_m0idx_n0idx_valid[i] ==
+            std::vector<int>{m0n0_idx[I0],
+                             m0n0_idx[I1],
+                             tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))};
+        EXPECT_TRUE(equal);
+    }
+}
+
+TEST(BlockToCTileMap, TestBlockToCTileMap_KSplit_M00_N0_M01Adapt)
+{
+    const index_t M         = 768;
+    const index_t N         = 384;
+    const index_t MPerBlock = 128;
+    const index_t NPerBlock = 128;
+    const index_t MBlock    = M / MPerBlock;
+    const index_t NBlock    = N / NPerBlock;
+    constexpr index_t M01   = 4;
+    const index_t KSplit    = 3;
+
+    auto c_grid_desc_m_n = make_naive_tensor_descriptor_packed(make_tuple(M, N));
+
+    printf("(M, N, MPerBlock, NPerBlock, M01) = (%d, %d, %d, %d, %d)\n",
+           M,
+           N,
+           MPerBlock,
+           NPerBlock,
+           M01);
+
+    BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n)>
+        tile_map(c_grid_desc_m_n, M01, KSplit);
+
+    EXPECT_TRUE(tile_map.CheckValidity(c_grid_desc_m_n) == true);
+    EXPECT_TRUE(tile_map.CalculateGridSize(c_grid_desc_m_n) == 18 * KSplit);
+
+    std::vector<std::vector<int>> expected_ksplitidx_m0idx_n0idx_valid = {
+        {0, 0, 0, 1}, {0, 1, 0, 1}, {0, 2, 0, 1}, {0, 3, 0, 1}, {0, 0, 1, 1}, {0, 1, 1, 1},
+        {0, 2, 1, 1}, {0, 3, 1, 1}, {0, 0, 2, 1}, {0, 1, 2, 1}, {0, 2, 2, 1}, {0, 3, 2, 1},
+        {0, 4, 0, 1}, {0, 5, 0, 1}, {0, 4, 1, 1}, {0, 5, 1, 1}, {0, 4, 2, 1}, {0, 5, 2, 1},
+        {1, 0, 0, 1}, {1, 1, 0, 1}, {1, 2, 0, 1}, {1, 3, 0, 1}, {1, 0, 1, 1}, {1, 1, 1, 1},
+        {1, 2, 1, 1}, {1, 3, 1, 1}, {1, 0, 2, 1}, {1, 1, 2, 1}, {1, 2, 2, 1}, {1, 3, 2, 1},
+        {1, 4, 0, 1}, {1, 5, 0, 1}, {1, 4, 1, 1}, {1, 5, 1, 1}, {1, 4, 2, 1}, {1, 5, 2, 1},
+        {2, 0, 0, 1}, {2, 1, 0, 1}, {2, 2, 0, 1}, {2, 3, 0, 1}, {2, 0, 1, 1}, {2, 1, 1, 1},
+        {2, 2, 1, 1}, {2, 3, 1, 1}, {2, 0, 2, 1}, {2, 1, 2, 1}, {2, 2, 2, 1}, {2, 3, 2, 1},
+        {2, 4, 0, 1}, {2, 5, 0, 1}, {2, 4, 1, 1}, {2, 5, 1, 1}, {2, 4, 2, 1}, {2, 5, 2, 1},
+    };
+
+    for(index_t i = 0; i < tile_map.CalculateGridSize(c_grid_desc_m_n); i++)
+    {
+        auto ksplitm0n0_idx = tile_map.CalculateBottomIndex(make_multi_index(i));
+        std::cout << "block_1d_id = " << i << ", ksplit, m0, n0 = " << ksplitm0n0_idx[I0] << ", "
+                  << ksplitm0n0_idx[I1] << ", " << ksplitm0n0_idx[I2];
+        std::cout << ", valid = "
+                  << tile_map.ValidCTileIndex(ksplitm0n0_idx, make_tuple(MBlock, NBlock))
+                  << std::endl;
+        bool equal =
+            expected_ksplitidx_m0idx_n0idx_valid[i] ==
+            std::vector<int>{ksplitm0n0_idx[I0],
+                             ksplitm0n0_idx[I1],
+                             ksplitm0n0_idx[I2],
+                             tile_map.ValidCTileIndex(ksplitm0n0_idx, make_tuple(MBlock, NBlock))};
+        EXPECT_TRUE(equal);
+    }
+}
diff --git a/3rdparty/composable_kernel/test/conv_util/CMakeLists.txt b/3rdparty/composable_kernel/test/conv_util/CMakeLists.txt
new file mode 100644
index 0000000..7a46039
--- /dev/null
+++ b/3rdparty/composable_kernel/test/conv_util/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_gtest_executable(test_conv_util conv_util.cpp)
+target_link_libraries(test_conv_util PRIVATE utility)
diff --git a/3rdparty/composable_kernel/test/conv_util/conv_util.cpp b/3rdparty/composable_kernel/test/conv_util/conv_util.cpp
new file mode 100644
index 0000000..73797a7
--- /dev/null
+++ b/3rdparty/composable_kernel/test/conv_util/conv_util.cpp
@@ -0,0 +1,156 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <string>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+
+namespace {
+
+class TestConvUtil : public ::testing::Test
+{
+    public:
+    void SetNDParams(std::size_t ndims, std::size_t s, std::size_t d, std::size_t p)
+    {
+        conv_params = ck::utils::conv::ConvParam(ndims,
+                                                 2,
+                                                 128,
+                                                 192,
+                                                 256,
+                                                 std::vector<ck::index_t>(ndims, 3),
+                                                 std::vector<ck::index_t>(ndims, 71),
+                                                 std::vector<ck::index_t>(ndims, s),
+                                                 std::vector<ck::index_t>(ndims, d),
+                                                 std::vector<ck::index_t>(ndims, p),
+                                                 std::vector<ck::index_t>(ndims, p));
+    }
+
+    protected:
+    // -------  default 2D -------
+    // input GNCHW {2, 128, 192, 71, 71},
+    // weights GKCYX {2, 256, 192, 3, 3},
+    // stride {s, s},
+    // dilations {d, d},
+    // padding {{p, p}, {p, p}
+    ck::utils::conv::ConvParam conv_params;
+};
+
+} // namespace
+
+TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths1D)
+{
+    // stride 2, dilation 1, pad 1
+    SetNDParams(1, 2, 1, 1);
+    std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D."));
+
+    // stride 1, dilation 1, pad 1
+    SetNDParams(1, 1, 1, 1);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len, std::vector<ck::index_t>{71}, "Error: ConvParams 1D stride {1}."));
+
+    // stride 2, dilation 1, pad 2
+    SetNDParams(1, 2, 1, 2);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
+                                     std::vector<ck::index_t>{37},
+                                     "Error: ConvParams 1D padding left/right {2}."));
+
+    // stride 2, dilation 2, pad 2
+    SetNDParams(1, 2, 2, 2);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D dilation {2}."));
+
+    // stride 3, dilation 2, pad 1
+    SetNDParams(1, 3, 2, 1);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(
+        ck::utils::check_err(out_spatial_len,
+                             std::vector<ck::index_t>{23},
+                             "Error: ConvParams 1D strides{3}, padding {1}, dilations {2}."));
+}
+
+TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths2D)
+{
+    // stride 2, dilation 1, pad 1
+    SetNDParams(2, 2, 1, 1);
+    std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
+                                     std::vector<ck::index_t>{36, 36},
+                                     "Error: ConvParams 2D default constructor."));
+
+    // stride 1, dilation 1, pad 1
+    SetNDParams(2, 1, 1, 1);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len, std::vector<ck::index_t>{71, 71}, "Error: ConvParams 2D stride {1,1}."));
+
+    // stride 2, dilation 1, pad 2
+    SetNDParams(2, 2, 1, 2);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
+                                     std::vector<ck::index_t>{37, 37},
+                                     "Error: ConvParams 2D padding left/right {2,2}."));
+
+    // stride 2, dilation 2, pad 2
+    SetNDParams(2, 2, 2, 2);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len, std::vector<ck::index_t>{36, 36}, "Error: ConvParams 2D dilation {2,2}."));
+
+    // stride 3, dilation 2, pad 1
+    SetNDParams(2, 3, 2, 1);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(
+        ck::utils::check_err(out_spatial_len,
+                             std::vector<ck::index_t>{23, 23},
+                             "Error: ConvParams 2D strides{3,3}, padding {1,1}, dilations {2,2}."));
+}
+
+TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths3D)
+{
+    // stride 2, dilation 1, pad 1
+    SetNDParams(3, 2, 1, 1);
+    std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len, std::vector<ck::index_t>{36, 36, 36}, "Error: ConvParams 3D."));
+
+    // stride 1, dilation 1, pad 1
+    SetNDParams(3, 1, 1, 1);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
+                                     std::vector<ck::index_t>{71, 71, 71},
+                                     "Error: ConvParams 3D stride {1, 1, 1}."));
+
+    // stride 2, dilation 1, pad 2
+    SetNDParams(3, 2, 1, 2);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
+                                     std::vector<ck::index_t>{37, 37, 37},
+                                     "Error: ConvParams 3D padding left/right {2, 2, 2}."));
+
+    // stride 2, dilation 2, pad 2
+    SetNDParams(3, 2, 2, 2);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
+                                     std::vector<ck::index_t>{36, 36, 36},
+                                     "Error: ConvParams 3D dilation {2, 2, 2}."));
+
+    // stride 3, dilation 2, pad 1
+    SetNDParams(3, 3, 2, 1);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len,
+        std::vector<ck::index_t>{23, 23, 23},
+        "Error: ConvParams 3D strides{3, 3, 3}, padding {1, 1, 1}, dilations {2, 2, 2}."));
+}
diff --git a/3rdparty/composable_kernel/test/convnd_bwd_data/CMakeLists.txt b/3rdparty/composable_kernel/test/convnd_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000..16ca4de
--- /dev/null
+++ b/3rdparty/composable_kernel/test/convnd_bwd_data/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_gtest_executable(test_convnd_bwd_data convnd_bwd_data.cpp)
+target_link_libraries(test_convnd_bwd_data PRIVATE utility device_conv1d_bwd_data_instance device_conv2d_bwd_data_instance device_conv3d_bwd_data_instance)
diff --git a/3rdparty/composable_kernel/test/convnd_bwd_data/convnd_bwd_data.cpp b/3rdparty/composable_kernel/test/convnd_bwd_data/convnd_bwd_data.cpp
new file mode 100644
index 0000000..70231d4
--- /dev/null
+++ b/3rdparty/composable_kernel/test/convnd_bwd_data/convnd_bwd_data.cpp
@@ -0,0 +1,93 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <tuple>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_conv_bwd_data_impl.hpp"
+
+template <typename Tuple>
+class TestConvndBwdData : public ::testing::Test
+{
+    protected:
+    using DataType = std::tuple_element_t<0, Tuple>;
+    std::vector<ck::utils::conv::ConvParam> conv_params;
+
+    template <ck::index_t NDimSpatial>
+    void Run()
+    {
+        for(auto& param : conv_params)
+        {
+            bool pass;
+            EXPECT_FALSE(conv_params.empty());
+            pass = ck::profiler::profile_conv_bwd_data_impl<
+                NDimSpatial,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::NWC,
+                                              ck::tensor_layout::convolution::NHWC,
+                                              ck::tensor_layout::convolution::NDHWC>>,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::KXC,
+                                              ck::tensor_layout::convolution::KYXC,
+                                              ck::tensor_layout::convolution::KZYXC>>,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::NWK,
+                                              ck::tensor_layout::convolution::NHWK,
+                                              ck::tensor_layout::convolution::NDHWK>>,
+                DataType,
+                DataType,
+                DataType>(true,  // do_verification
+                          1,     // init_method integer value
+                          false, // do_log
+                          false, // time_kernel
+                          param);
+            EXPECT_TRUE(pass);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<std::tuple<float>,
+                                     std::tuple<ck::half_t>,
+                                     std::tuple<ck::bhalf_t>,
+                                     std::tuple<std::int8_t>>;
+TYPED_TEST_SUITE(TestConvndBwdData, KernelTypes);
+
+// 1d
+TYPED_TEST(TestConvndBwdData, Conv1dBwdData)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back({1, 1, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
+    this->conv_params.push_back({1, 1, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
+    this->conv_params.push_back({1, 1, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
+    this->template Run<1>();
+}
+
+// 2d
+TYPED_TEST(TestConvndBwdData, Conv2dBwdData)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {2, 1, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    this->conv_params.push_back(
+        {2, 1, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back(
+        {2, 1, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+    this->template Run<2>();
+}
+
+// 3d
+TYPED_TEST(TestConvndBwdData, Conv3dBwdData)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {3, 1, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->conv_params.push_back(
+        {3, 1, 128, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 1, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->template Run<3>();
+}
diff --git a/3rdparty/composable_kernel/test/convnd_fwd/CMakeLists.txt b/3rdparty/composable_kernel/test/convnd_fwd/CMakeLists.txt
new file mode 100644
index 0000000..97e170d
--- /dev/null
+++ b/3rdparty/composable_kernel/test/convnd_fwd/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_gtest_executable(test_convnd_fwd convnd_fwd.cpp)
+target_link_libraries(test_convnd_fwd PRIVATE utility device_conv2d_fwd_instance)
diff --git a/3rdparty/composable_kernel/test/convnd_fwd/convnd_fwd.cpp b/3rdparty/composable_kernel/test/convnd_fwd/convnd_fwd.cpp
new file mode 100644
index 0000000..a1921a9
--- /dev/null
+++ b/3rdparty/composable_kernel/test/convnd_fwd/convnd_fwd.cpp
@@ -0,0 +1,92 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <tuple>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_conv_fwd_impl.hpp"
+
+template <typename Tuple>
+class TestConvndFwd : public ::testing::Test
+{
+    protected:
+    using DataType = std::tuple_element_t<0, Tuple>;
+    std::vector<ck::utils::conv::ConvParam> conv_params;
+
+    template <ck::index_t NDimSpatial>
+    void Run()
+    {
+        for(auto& param : conv_params)
+        {
+            bool pass;
+            EXPECT_FALSE(conv_params.empty());
+            pass = ck::profiler::profile_conv_fwd_impl<
+                NDimSpatial,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::NWC,
+                                              ck::tensor_layout::convolution::NHWC,
+                                              ck::tensor_layout::convolution::NDHWC>>,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::KXC,
+                                              ck::tensor_layout::convolution::KYXC,
+                                              ck::tensor_layout::convolution::KZYXC>>,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::NWK,
+                                              ck::tensor_layout::convolution::NHWK,
+                                              ck::tensor_layout::convolution::NDHWK>>,
+                DataType,
+                DataType,
+                DataType>(true,  // do_verification
+                          1,     // init_method integer value
+                          false, // do_log
+                          false, // time_kernel
+                          param);
+            EXPECT_TRUE(pass);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<std::tuple<float>,
+                                     std::tuple<ck::half_t>,
+                                     std::tuple<ck::bhalf_t>,
+                                     std::tuple<std::int8_t>>;
+TYPED_TEST_SUITE(TestConvndFwd, KernelTypes);
+
+// 1d
+TYPED_TEST(TestConvndFwd, Conv1dFwd)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back({1, 1, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
+    this->conv_params.push_back({1, 1, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
+    this->conv_params.push_back({1, 1, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
+    this->template Run<1>();
+}
+
+// 2d
+TYPED_TEST(TestConvndFwd, Conv2dFwd)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {2, 1, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    this->conv_params.push_back(
+        {2, 1, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back(
+        {2, 1, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+    this->template Run<2>();
+}
+// 3d
+TYPED_TEST(TestConvndFwd, Conv3dFwd)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {3, 1, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->conv_params.push_back(
+        {3, 1, 128, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 1, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->template Run<3>();
+}
diff --git a/3rdparty/composable_kernel/test/data_type/CMakeLists.txt b/3rdparty/composable_kernel/test/data_type/CMakeLists.txt
new file mode 100644
index 0000000..088fbfe
--- /dev/null
+++ b/3rdparty/composable_kernel/test/data_type/CMakeLists.txt
@@ -0,0 +1,4 @@
+if (USE_BITINT_EXTENSION_INT4)
+  add_gtest_executable(test_int4 int4.cpp)
+  target_link_libraries(test_int4 PRIVATE utility)
+endif()
diff --git a/3rdparty/composable_kernel/test/data_type/int4.cpp b/3rdparty/composable_kernel/test/data_type/int4.cpp
new file mode 100644
index 0000000..252a450
--- /dev/null
+++ b/3rdparty/composable_kernel/test/data_type/int4.cpp
@@ -0,0 +1,211 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <bitset>
+#include <cinttypes>
+#include <cstdint>
+#include <iomanip>
+#include "gtest/gtest.h"
+#include <hip/hip_runtime.h>
+
+#include "ck/host_utility/hip_check_error.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/get_id.hpp"
+#include "ck/library/utility/device_memory.hpp"
+
+using ck::int4_t;
+
+TEST(Int4, BaseArithmetic)
+{
+    int4_t a{1};
+    int4_t b{-2};
+    EXPECT_EQ(a + a, int4_t{2});
+    EXPECT_EQ(a - a, int4_t{0});
+    EXPECT_EQ(a + b, int4_t{-1});
+    EXPECT_EQ(a - b, int4_t{3});
+    EXPECT_EQ(a * a, int4_t{1});
+    EXPECT_EQ(a * b, int4_t{-2});
+    EXPECT_EQ(b * b, int4_t{4});
+    EXPECT_EQ(a / b, int4_t{0});
+    a = int4_t{4};
+    EXPECT_EQ(a / b, int4_t{-2});
+    b = int4_t{2};
+    EXPECT_EQ(a % b, int4_t{0});
+}
+
+TEST(Int4, NumericLimits)
+{
+    EXPECT_EQ(ck::NumericLimits<int4_t>::Min(), int4_t{-8});
+    EXPECT_EQ(ck::NumericLimits<int4_t>::Max(), int4_t{7});
+    EXPECT_EQ(ck::NumericLimits<int4_t>::Lowest(), int4_t{-8});
+}
+
+TEST(Int4, MathOpsV2)
+{
+    int4_t a{4};
+    int4_t b{-5};
+
+    EXPECT_EQ(ck::math::abs(a), int4_t{4});
+    EXPECT_EQ(ck::math::abs(b), int4_t{5});
+    EXPECT_FALSE(ck::math::isnan(b));
+}
+
+namespace {
+
+__global__ void copy(const int4_t* src, std::int8_t* dst, ck::index_t N)
+{
+    ck::index_t tid = ck::get_thread_global_1d_id();
+
+    const int8_t* src_i8 = reinterpret_cast<const int8_t*>(src);
+
+    if(tid < N)
+    {
+        for(ck::index_t i = tid; i < N; i += ck::get_grid_size())
+        {
+            dst[i] = src_i8[i];
+        }
+    }
+}
+
+__global__ void copy_with_static_cast(const int4_t* src, std::int8_t* dst, ck::index_t N)
+{
+    ck::index_t tid = ck::get_thread_global_1d_id();
+
+    if(tid < N)
+    {
+        for(ck::index_t i = tid; i < N; i += ck::get_grid_size())
+        {
+            dst[i] = static_cast<std::int8_t>(src[i]);
+        }
+    }
+}
+
+} // anonymous namespace
+
+TEST(Int4, CopyAsI8PositiveValue)
+{
+    constexpr std::size_t SIZE = 100;
+    std::vector<int4_t> h_src_i4(SIZE, 7);
+    std::vector<std::int8_t> h_src_i8(SIZE, 7);
+    std::vector<std::int8_t> h_dst_i8(SIZE, 0);
+
+    DeviceMem d_src_i4(h_src_i4.size() * sizeof(int4_t));
+    DeviceMem d_dst_i8(h_dst_i8.size() * sizeof(std::int8_t));
+
+    d_src_i4.SetZero();
+    d_dst_i8.SetZero();
+
+    d_src_i4.ToDevice(h_src_i4.data());
+
+    copy<<<1, 64>>>(reinterpret_cast<const int4_t*>(d_src_i4.GetDeviceBuffer()),
+                    reinterpret_cast<std::int8_t*>(d_dst_i8.GetDeviceBuffer()),
+                    SIZE);
+    hip_check_error(hipDeviceSynchronize());
+    d_dst_i8.FromDevice(h_dst_i8.data());
+
+    for(std::size_t i = 0; i < SIZE; ++i)
+    {
+        EXPECT_EQ(h_src_i8[i], h_dst_i8[i]);
+    }
+}
+
+TEST(Int4, DISABLED_CopyAsI8NegativeValue)
+{
+    constexpr std::size_t SIZE = 32;
+    std::vector<int4_t> h_src_i4(SIZE, -8);
+    std::vector<std::int8_t> h_src_i8(SIZE, -8);
+    std::vector<std::int8_t> h_dst_i8(SIZE, 0);
+
+    DeviceMem d_src_i4(h_src_i4.size() * sizeof(int4_t));
+    DeviceMem d_dst_i8(h_dst_i8.size() * sizeof(std::int8_t));
+
+    d_src_i4.SetZero();
+    d_dst_i8.SetZero();
+
+    d_src_i4.ToDevice(h_src_i4.data());
+
+    copy<<<1, 64>>>(reinterpret_cast<const int4_t*>(d_src_i4.GetDeviceBuffer()),
+                    reinterpret_cast<std::int8_t*>(d_dst_i8.GetDeviceBuffer()),
+                    SIZE);
+    hip_check_error(hipDeviceSynchronize());
+    d_dst_i8.FromDevice(h_dst_i8.data());
+
+    for(std::size_t i = 0; i < SIZE; ++i)
+    {
+        EXPECT_EQ(h_src_i8[i], h_dst_i8[i]);
+    }
+}
+
+TEST(Int4, CopyAsI8NegativeValueStaticCast)
+{
+    constexpr std::size_t SIZE = 32;
+    std::vector<int4_t> h_src_i4(SIZE, -8);
+    std::vector<std::int8_t> h_src_i8(SIZE, -8);
+    std::vector<std::int8_t> h_dst_i8(SIZE, 0);
+
+    DeviceMem d_src_i4(h_src_i4.size() * sizeof(int4_t));
+    DeviceMem d_dst_i8(h_dst_i8.size() * sizeof(std::int8_t));
+
+    d_src_i4.SetZero();
+    d_dst_i8.SetZero();
+
+    d_src_i4.ToDevice(h_src_i4.data());
+
+    copy_with_static_cast<<<1, 64>>>(reinterpret_cast<const int4_t*>(d_src_i4.GetDeviceBuffer()),
+                                     reinterpret_cast<std::int8_t*>(d_dst_i8.GetDeviceBuffer()),
+                                     SIZE);
+    hip_check_error(hipDeviceSynchronize());
+    d_dst_i8.FromDevice(h_dst_i8.data());
+
+    for(std::size_t i = 0; i < SIZE; ++i)
+    {
+        EXPECT_EQ(h_src_i8[i], h_dst_i8[i]);
+    }
+}
+
+TEST(Int4, DISABLED_BitwiseRepresentation)
+{
+    using bit8_t = std::bitset<8>;
+
+    int4_t a_i4{3};
+    std::int8_t a_i8 = *reinterpret_cast<std::int8_t*>(&a_i4);
+    std::int8_t b_i8{3};
+#if 0
+    std::cout << std::hex << std::showbase << static_cast<int32_t>(a_i8)
+              << ", " << static_cast<int32_t>(b_i8) << std::endl;
+#endif
+    EXPECT_EQ(bit8_t{static_cast<std::uint64_t>(a_i8)}, bit8_t{static_cast<std::uint64_t>(b_i8)});
+
+    a_i4 = int4_t{-3};
+    a_i8 = *reinterpret_cast<std::int8_t*>(&a_i4);
+    b_i8 = std::int8_t{-3};
+#if 0
+    std::cout << std::hex << std::showbase << static_cast<int32_t>(a_i8)
+              << ", " << static_cast<int32_t>(b_i8) << std::endl;
+#endif
+    EXPECT_EQ(bit8_t{static_cast<std::uint64_t>(a_i8)}, bit8_t{static_cast<std::uint64_t>(b_i8)});
+}
+
+TEST(Int4, BitwiseRepresentationStaticCast)
+{
+    using bit8_t = std::bitset<8>;
+
+    int4_t a_i4{3};
+    std::int8_t a_i8 = static_cast<std::int8_t>(a_i4);
+    std::int8_t b_i8{3};
+#if 0
+    std::cout << std::hex << std::showbase << static_cast<int32_t>(a_i8)
+              << ", " << static_cast<int32_t>(b_i8) << std::endl;
+#endif
+    EXPECT_EQ(bit8_t{static_cast<std::uint64_t>(a_i8)}, bit8_t{static_cast<std::uint64_t>(b_i8)});
+
+    a_i4 = int4_t{-3};
+    a_i8 = static_cast<std::int8_t>(a_i4);
+    b_i8 = std::int8_t{-3};
+#if 0
+    std::cout << std::hex << std::showbase << static_cast<int32_t>(a_i8)
+              << ", " << static_cast<int32_t>(b_i8) << std::endl;
+#endif
+    EXPECT_EQ(bit8_t{static_cast<std::uint64_t>(a_i8)}, bit8_t{static_cast<std::uint64_t>(b_i8)});
+}
diff --git a/3rdparty/composable_kernel/test/elementwise_normalization/CMakeLists.txt b/3rdparty/composable_kernel/test/elementwise_normalization/CMakeLists.txt
new file mode 100644
index 0000000..a20eb26
--- /dev/null
+++ b/3rdparty/composable_kernel/test/elementwise_normalization/CMakeLists.txt
@@ -0,0 +1,7 @@
+add_custom_target(test_elementwise_normalization)
+
+add_gtest_executable(test_elementwise_layernorm_fp16 test_elementwise_layernorm_fp16.cpp)
+
+target_link_libraries(test_elementwise_layernorm_fp16 PRIVATE utility device_elementwise_normalization_instance)
+
+add_dependencies(test_elementwise_normalization test_elementwise_layernorm_fp16)
diff --git a/3rdparty/composable_kernel/test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp b/3rdparty/composable_kernel/test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp
new file mode 100644
index 0000000..403881b
--- /dev/null
+++ b/3rdparty/composable_kernel/test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_elementwise_layernorm_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestElementwiseLayernorm : public ::testing::Test
+{
+    protected:
+    using ADataType     = std::tuple_element_t<0, Tuple>;
+    using BDataType     = std::tuple_element_t<1, Tuple>;
+    using GammaDataType = std::tuple_element_t<2, Tuple>;
+    using BetaDataType  = std::tuple_element_t<3, Tuple>;
+    using AccDataType   = std::tuple_element_t<4, Tuple>;
+    using YDataType     = std::tuple_element_t<5, Tuple>;
+
+    void Run()
+    {
+        // M, N
+        std::vector<std::vector<ck::index_t>> lengths = {
+            {1, 1}, {25, 16}, {39, 777}, {100, 200}, {1024, 1024}, {48 * 256, 2048}};
+
+        for(auto length : lengths)
+        {
+            bool success = ck::profiler::profile_elementwise_layernorm_impl<ADataType,
+                                                                            BDataType,
+                                                                            GammaDataType,
+                                                                            BetaDataType,
+                                                                            AccDataType,
+                                                                            YDataType>(
+                true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // ADataType, BDataType, GammaDataType, BetaDataType, AccDataType, YDataType>
+    std::tuple<F16, F16, F16, F16, F32, F16>>;
+
+TYPED_TEST_SUITE(TestElementwiseLayernorm, KernelTypes);
+TYPED_TEST(TestElementwiseLayernorm, Test_FP16) { this->Run(); }
diff --git a/3rdparty/composable_kernel/test/gemm/CMakeLists.txt b/3rdparty/composable_kernel/test/gemm/CMakeLists.txt
new file mode 100644
index 0000000..c427586
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/CMakeLists.txt
@@ -0,0 +1,25 @@
+add_test_executable(test_gemm_fp32 gemm_fp32.cpp)
+target_link_libraries(test_gemm_fp32 PRIVATE utility)
+target_link_libraries(test_gemm_fp32 PRIVATE device_gemm_instance)
+
+add_test_executable(test_gemm_fp16 gemm_fp16.cpp)
+target_link_libraries(test_gemm_fp16 PRIVATE utility)
+target_link_libraries(test_gemm_fp16 PRIVATE device_gemm_instance)
+
+add_test_executable(test_gemm_bf16 gemm_bf16.cpp)
+target_link_libraries(test_gemm_bf16 PRIVATE utility)
+target_link_libraries(test_gemm_bf16 PRIVATE device_gemm_instance)
+
+add_test_executable(test_gemm_int8 gemm_int8.cpp)
+target_link_libraries(test_gemm_int8 PRIVATE utility)
+target_link_libraries(test_gemm_int8 PRIVATE device_gemm_instance)
+
+add_library(gemm_standalone_xdl_fp16_instances STATIC
+    instance/gemm_f16_nn_instance.cpp
+    instance/gemm_f16_nt_instance.cpp
+    instance/gemm_f16_tn_instance.cpp
+    instance/gemm_f16_tt_instance.cpp
+)
+add_test_executable(test_gemm_standalone_xdl_fp16 gemm_standalone_xdl_fp16.cpp)
+target_link_libraries(test_gemm_standalone_xdl_fp16 PRIVATE gemm_standalone_xdl_fp16_instances utility)
+target_include_directories(test_gemm_standalone_xdl_fp16 PRIVATE instance/)
diff --git a/3rdparty/composable_kernel/test/gemm/gemm_bf16.cpp b/3rdparty/composable_kernel/test/gemm/gemm_bf16.cpp
new file mode 100644
index 0000000..5290d46
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/gemm_bf16.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+#include "test/gemm/gemm_util.hpp"
+
+using ADataType   = ck::bhalf_t;
+using BDataType   = ck::bhalf_t;
+using CDataType   = ck::bhalf_t;
+using AccDataType = float;
+
+#include "run_gemm_test.inc"
+
+int main() { return run_gemm_test(); }
diff --git a/3rdparty/composable_kernel/test/gemm/gemm_fp16.cpp b/3rdparty/composable_kernel/test/gemm/gemm_fp16.cpp
new file mode 100644
index 0000000..92e225d
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/gemm_fp16.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+#include "test/gemm/gemm_util.hpp"
+
+using ADataType   = ck::half_t;
+using BDataType   = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+
+#include "run_gemm_test.inc"
+
+int main() { return run_gemm_test(); }
diff --git a/3rdparty/composable_kernel/test/gemm/gemm_fp32.cpp b/3rdparty/composable_kernel/test/gemm/gemm_fp32.cpp
new file mode 100644
index 0000000..5d8c488
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/gemm_fp32.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+#include "test/gemm/gemm_util.hpp"
+
+using ADataType   = float;
+using BDataType   = float;
+using CDataType   = float;
+using AccDataType = float;
+
+#include "run_gemm_test.inc"
+
+int main() { return run_gemm_test(); }
diff --git a/3rdparty/composable_kernel/test/gemm/gemm_fp64.cpp b/3rdparty/composable_kernel/test/gemm/gemm_fp64.cpp
new file mode 100644
index 0000000..85d7f95
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/gemm_fp64.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+#include "test/gemm/gemm_util.hpp"
+
+using ADataType   = double;
+using BDataType   = double;
+using CDataType   = double;
+using AccDataType = double;
+
+#include "run_gemm_test.inc"
+
+int main() { return run_gemm_test(); }
diff --git a/3rdparty/composable_kernel/test/gemm/gemm_int8.cpp b/3rdparty/composable_kernel/test/gemm/gemm_int8.cpp
new file mode 100644
index 0000000..e73b22c
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/gemm_int8.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+#include "test/gemm/gemm_util.hpp"
+
+using ADataType   = int8_t;
+using BDataType   = int8_t;
+using CDataType   = int8_t;
+using AccDataType = int32_t;
+
+#include "run_gemm_test.inc"
+
+int main() { return run_gemm_test(); }
diff --git a/3rdparty/composable_kernel/test/gemm/gemm_standalone_xdl_fp16.cpp b/3rdparty/composable_kernel/test/gemm/gemm_standalone_xdl_fp16.cpp
new file mode 100644
index 0000000..8f5a5c5
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/gemm_standalone_xdl_fp16.cpp
@@ -0,0 +1,162 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_util.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+#include "gemm_f16_nn_instance.hpp"
+#include "gemm_f16_nt_instance.hpp"
+#include "gemm_f16_tn_instance.hpp"
+#include "gemm_f16_tt_instance.hpp"
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using F16         = ck::half_t;
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = float;
+using CDataType   = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+using ck::gemm_util::GemmParams;
+using ck::tensor_operation::device::BaseOperator;
+using ck::tensor_operation::device::DeviceGemm;
+using namespace ck::tensor_operation::device::instance;
+
+using DeviceGemmNN =
+    DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>;
+using DeviceGemmNT =
+    DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>;
+using DeviceGemmTN =
+    DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>;
+using DeviceGemmTT =
+    DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>;
+
+struct LayoutConfig
+{
+    bool ARowMajor;
+    bool BRowMajor;
+    bool CRowMajor;
+};
+
+int main(int argc, char* argv[])
+{
+    // Class DeviceGemm is templated by layout and precision types so it is not an option to contain
+    // them in a single vector. Instead we use abstract BaseOperator class and dynamic_cast() it
+    // upon invocation.
+    // And since DeviceGemm does not expose template arg information, an extra book keeping class
+    // LayoutConfig is used for determining which type a BaseOperator instance should be cast to.
+    using OpFactoryFn = void (*)(std::vector<std::unique_ptr<BaseOperator>>&);
+
+    std::vector<std::tuple<GemmParams, LayoutConfig, OpFactoryFn>> problems = {
+        // clang-format off
+    // 104 tiles
+    {GemmParams{2048, 3328, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_256x256},
+    {GemmParams{2048, 1664, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_256x128},
+    {GemmParams{1024, 1664, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_128x128},
+    {GemmParams{1024,  832, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_128x64},
+    {GemmParams{2048, 3328, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_256x256},
+    {GemmParams{2048, 1664, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_256x128},
+    {GemmParams{1024, 1664, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_128x128},
+    {GemmParams{1024,  832, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_128x64},
+    {GemmParams{2048, 3328, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_256x256},
+    {GemmParams{2048, 1664, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_256x128},
+    {GemmParams{1024, 1664, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_128x128},
+    {GemmParams{1024,  832, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_128x64},
+    {GemmParams{2048, 3328, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_256x256},
+    {GemmParams{2048, 1664, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_256x128},
+    {GemmParams{1024, 1664, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_128x128},
+    {GemmParams{1024,  832, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_128x64},
+    // 110 tiles
+    {GemmParams{2560, 2816, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_256x256},
+    {GemmParams{2560, 1408, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_256x128},
+    {GemmParams{1280, 1408, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_128x128},
+    {GemmParams{1280,  704, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_128x64},
+    {GemmParams{2560, 2816, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_256x256},
+    {GemmParams{2560, 1408, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_256x128},
+    {GemmParams{1280, 1408, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_128x128},
+    {GemmParams{1280,  704, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_128x64},
+    {GemmParams{2560, 2816, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_256x256},
+    {GemmParams{2560, 1408, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_256x128},
+    {GemmParams{1280, 1408, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_128x128},
+    {GemmParams{1280,  704, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_128x64},
+    {GemmParams{2560, 2816, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_256x256},
+    {GemmParams{2560, 1408, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_256x128},
+    {GemmParams{1280, 1408, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_128x128},
+    {GemmParams{1280,  704, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_128x64},
+        // clang-format on
+    };
+
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 3)
+    {
+        do_verification = std::stoi(argv[1]);
+        time_kernel     = std::stoi(argv[2]);
+    }
+    else
+    {
+        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+                  << "arg2: time kernel (0=no, 1=yes)" << std::endl;
+        return 0;
+    }
+
+    bool pass = true;
+    for(auto& p : problems)
+    {
+        GemmParams& problem_size          = std::get<0>(p);
+        const LayoutConfig& layout_config = std::get<1>(p);
+        const auto& factory               = std::get<2>(p);
+        std::vector<std::unique_ptr<BaseOperator>> ops;
+        factory(ops);
+
+        // overwrite strides
+        problem_size.StrideA = layout_config.ARowMajor ? problem_size.K : problem_size.M;
+        problem_size.StrideB = layout_config.BRowMajor ? problem_size.N : problem_size.K;
+        problem_size.StrideC = layout_config.CRowMajor ? problem_size.N : problem_size.M;
+
+        if(!layout_config.ARowMajor && !layout_config.BRowMajor)
+        {
+            auto op_ptr = dynamic_cast<DeviceGemmNN*>(ops[0].get());
+            pass &= ck::gemm_util::TestGemm<AccDataType>{}(
+                op_ptr, problem_size, do_verification, time_kernel);
+        }
+        else if(!layout_config.ARowMajor && layout_config.BRowMajor)
+        {
+            auto op_ptr = dynamic_cast<DeviceGemmNT*>(ops[0].get());
+            pass &= ck::gemm_util::TestGemm<AccDataType>{}(
+                op_ptr, problem_size, do_verification, time_kernel);
+        }
+        else if(layout_config.ARowMajor && !layout_config.BRowMajor)
+        {
+            auto op_ptr = dynamic_cast<DeviceGemmTN*>(ops[0].get());
+            pass &= ck::gemm_util::TestGemm<AccDataType>{}(
+                op_ptr, problem_size, do_verification, time_kernel);
+        }
+        else if(layout_config.ARowMajor && layout_config.BRowMajor)
+        {
+            auto op_ptr = dynamic_cast<DeviceGemmTT*>(ops[0].get());
+            pass &= ck::gemm_util::TestGemm<AccDataType>{}(
+                op_ptr, problem_size, do_verification, time_kernel);
+        }
+    }
+
+    std::cout << (pass ? "ALL TESTS PASSED" : "SOME TESTS FAILED") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/test/gemm/gemm_util.hpp b/3rdparty/composable_kernel/test/gemm/gemm_util.hpp
new file mode 100644
index 0000000..9057c0a
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/gemm_util.hpp
@@ -0,0 +1,267 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace gemm_util {
+
+struct GemmParams
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideC = 1024;
+};
+
+template <typename GemmInstance,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+void RunHostGEMM(const Tensor<ADataType>& A,
+                 const Tensor<BDataType>& B,
+                 Tensor<CDataType>& C,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+{
+    auto ref_gemm    = GemmInstance{};
+    auto ref_invoker = ref_gemm.MakeInvoker();
+
+    auto ref_argument = ref_gemm.MakeArgument(A, B, C, a_element_op, b_element_op, c_element_op);
+
+    ref_invoker.Run(ref_argument);
+}
+
+template <typename DeviceGemmPtr_,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+bool RunDeviceGEMM(DeviceGemmPtr_& gemmPtr,
+                   const ck::gemm_util::GemmParams& params,
+                   const Tensor<ADataType>& A,
+                   const Tensor<BDataType>& B,
+                   Tensor<CDataType>& C,
+                   AElementwiseOperation a_element_op,
+                   BElementwiseOperation b_element_op,
+                   CElementwiseOperation c_element_op,
+                   bool time_kernel)
+{
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpaceSize());
+
+    auto invoker_ptr = gemmPtr->MakeInvokerPointer();
+    auto argument_ptr =
+        gemmPtr->MakeArgumentPointer(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                                     static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+                                     static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+                                     params.M,
+                                     params.N,
+                                     params.K,
+                                     params.StrideA,
+                                     params.StrideB,
+                                     params.StrideC,
+                                     a_element_op,
+                                     b_element_op,
+                                     c_element_op);
+
+    if(gemmPtr->IsSupportedArgument(argument_ptr.get()))
+    {
+        a_m_k_device_buf.ToDevice(A.mData.data());
+        b_k_n_device_buf.ToDevice(B.mData.data());
+        float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t flop      = std::size_t(2) * params.M * params.N * params.K;
+        std::size_t num_btype = sizeof(ADataType) * params.M * params.K +
+                                sizeof(BDataType) * params.K * params.N +
+                                sizeof(CDataType) * params.M * params.N;
+
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << std::endl;
+
+        c_m_n_device_buf.FromDevice(C.mData.data());
+
+        return true;
+    }
+    else
+    {
+        std::cout << "device_gemm with the specified compilation parameters does "
+                     "not support this GEMM problem"
+                  << std::endl;
+
+        return false;
+    }
+}
+
+template <typename AccDataType>
+struct TestGemm
+{
+    template <typename ADataType,
+              typename BDataType,
+              typename CDataType,
+              typename ALayout,
+              typename BLayout,
+              typename CLayout>
+    auto PrepareGemmTensor(const ck::gemm_util::GemmParams& params)
+    {
+        auto f_host_tensor_descriptor =
+            [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+                using namespace ck::literals;
+
+                if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+                {
+                    return HostTensorDescriptor({row, col}, {stride, 1_uz});
+                }
+                else
+                {
+                    return HostTensorDescriptor({row, col}, {1_uz, stride});
+                }
+            };
+
+        Tensor<ADataType> a_m_k(
+            f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
+        Tensor<BDataType> b_k_n(
+            f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
+        Tensor<CDataType> c_m_n_host_result(
+            f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+        Tensor<CDataType> c_m_n_device_result(
+            f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+
+        auto f_generate_tensor_value = [](auto& tensor, auto type) {
+            using dataType = decltype(type);
+
+            tensor.GenerateTensorValue(GeneratorTensor_2<dataType>{-5, 5});
+        };
+
+        f_generate_tensor_value(a_m_k, ADataType{});
+        f_generate_tensor_value(b_k_n, BDataType{});
+
+        std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+        std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+        std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+        return std::make_tuple(a_m_k, b_k_n, c_m_n_host_result, c_m_n_device_result);
+    }
+
+    template <template <class...> class DeviceGemmPtr_,
+              typename ALayout,
+              typename BLayout,
+              typename CLayout,
+              typename ADataType,
+              typename BDataType,
+              typename CDataType,
+              typename AElementwiseOperation,
+              typename BElementwiseOperation,
+              typename CElementwiseOperation>
+    auto operator()(DeviceGemmPtr_<ALayout,
+                                   BLayout,
+                                   CLayout,
+                                   ADataType,
+                                   BDataType,
+                                   CDataType,
+                                   AElementwiseOperation,
+                                   BElementwiseOperation,
+                                   CElementwiseOperation>* gemmPtr,
+                    const GemmParams& params = GemmParams{},
+                    bool do_verification     = true,
+                    bool time_kernel         = false)
+    {
+        std::cout << "ALayout = " << ALayout{}.name << ", BLayout = " << BLayout{}.name
+                  << ", CLayout = " << CLayout{}.name << std::endl;
+        std::cout << gemmPtr->GetTypeString() << std::endl;
+
+        auto host_tensors =
+            PrepareGemmTensor<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>(params);
+
+        const Tensor<ADataType>& a  = std::get<0>(host_tensors);
+        const Tensor<BDataType>& b  = std::get<1>(host_tensors);
+        Tensor<CDataType>& c_host   = std::get<2>(host_tensors);
+        Tensor<CDataType>& c_device = std::get<3>(host_tensors);
+
+        auto a_element_op = AElementwiseOperation{};
+        auto b_element_op = BElementwiseOperation{};
+        auto c_element_op = CElementwiseOperation{};
+
+        using ReferenceGemmInstance =
+            ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                      BDataType,
+                                                      CDataType,
+                                                      AccDataType,
+                                                      AElementwiseOperation,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation>;
+
+        if(do_verification)
+        {
+            ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
+                a, b, c_host, a_element_op, b_element_op, c_element_op);
+        }
+
+        // Act
+        bool is_supported = ck::gemm_util::RunDeviceGEMM(
+            gemmPtr, params, a, b, c_device, a_element_op, b_element_op, c_element_op, time_kernel);
+
+        if(is_supported && do_verification)
+        {
+            // Assert
+            bool res = false;
+            if(std::is_same<CDataType, float>::value)
+            {
+                res = ck::utils::check_err(c_device, c_host);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, ck::half_t>::value)
+            {
+                res = ck::utils::check_err(c_device, c_host);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, ck::bhalf_t>::value)
+            {
+                res = ck::utils::check_err(c_device, c_host);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, int8_t>::value)
+            {
+                res = ck::utils::check_err(c_device, c_host);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, double>::value)
+            {
+                res = ck::utils::check_err(c_device, c_host);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+
+            return res;
+        }
+        else
+        {
+            return true;
+        }
+    }
+};
+
+} // namespace gemm_util
+} // namespace ck
diff --git a/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nn_instance.cpp b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nn_instance.cpp
new file mode 100644
index 0000000..4d65c58
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nn_instance.cpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "gemm_f16_nn_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using gemm_f16_nn_256x256 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   256,    32,   2,   8,   32,   32,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_nn_256x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_nn_128x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_nn_128x64 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_gemm_f16_nn_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nn_256x256{});
+}
+
+void add_gemm_f16_nn_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nn_256x128{});
+}
+
+void add_gemm_f16_nn_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nn_128x128{});
+}
+
+void add_gemm_f16_nn_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nn_128x64{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nn_instance.hpp b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nn_instance.hpp
new file mode 100644
index 0000000..5ae3928
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nn_instance.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+void add_gemm_f16_nn_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_nn_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_nn_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_nn_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nt_instance.cpp b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nt_instance.cpp
new file mode 100644
index 0000000..431ff1e
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nt_instance.cpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "gemm_f16_nt_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using gemm_f16_nt_256x256 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   256,    32,   2,   2,   32,   32,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_nt_256x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_nt_128x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_nt_128x64 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_gemm_f16_nt_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nt_256x256{});
+}
+
+void add_gemm_f16_nt_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nt_256x128{});
+}
+
+void add_gemm_f16_nt_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nt_128x128{});
+}
+
+void add_gemm_f16_nt_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nt_128x64{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nt_instance.hpp b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nt_instance.hpp
new file mode 100644
index 0000000..99f9ffb
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_nt_instance.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+void add_gemm_f16_nt_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_nt_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_nt_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_nt_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tn_instance.cpp b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tn_instance.cpp
new file mode 100644
index 0000000..6f5dbc3
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tn_instance.cpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "gemm_f16_tn_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using gemm_f16_tn_256x256 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   256,    32,   8,   8,   32,   32,    4,   4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tn_256x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tn_128x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tn_128x64 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_gemm_f16_tn_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tn_256x256{});
+}
+
+void add_gemm_f16_tn_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tn_256x128{});
+}
+
+void add_gemm_f16_tn_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tn_128x128{});
+}
+
+void add_gemm_f16_tn_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tn_128x64{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tn_instance.hpp b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tn_instance.hpp
new file mode 100644
index 0000000..62388ae
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tn_instance.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+void add_gemm_f16_tn_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_tn_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_tn_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_tn_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tt_instance.cpp b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tt_instance.cpp
new file mode 100644
index 0000000..b6ef5b1
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tt_instance.cpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "gemm_f16_tt_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using gemm_f16_tt_256x256 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   256,    32,   8,   2,   32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tt_256x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tt_128x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tt_128x64 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_gemm_f16_tt_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tt_256x256{});
+}
+
+void add_gemm_f16_tt_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tt_256x128{});
+}
+
+void add_gemm_f16_tt_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tt_128x128{});
+}
+
+void add_gemm_f16_tt_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tt_128x64{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tt_instance.hpp b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tt_instance.hpp
new file mode 100644
index 0000000..9d75b4e
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/instance/gemm_f16_tt_instance.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+void add_gemm_f16_tt_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_tt_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_tt_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_tt_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/3rdparty/composable_kernel/test/gemm/run_gemm_test.inc b/3rdparty/composable_kernel/test/gemm/run_gemm_test.inc
new file mode 100644
index 0000000..ec27729
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm/run_gemm_test.inc
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+int run_gemm_test()
+{
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    auto test = [&](auto a_layout, auto b_layout, auto c_layout) {
+        bool pass = true;
+
+        using DeviceOp = ck::tensor_operation::device::DeviceGemm<decltype(a_layout),
+                                                                  decltype(b_layout),
+                                                                  decltype(c_layout),
+                                                                  ADataType,
+                                                                  BDataType,
+                                                                  CDataType,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>;
+
+        const auto gemmPtrs =
+            ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+                DeviceOp>::GetInstances();
+
+        for(auto& gemmPtr : gemmPtrs)
+        {
+            pass &= ck::gemm_util::TestGemm<AccDataType>{}(gemmPtr.get());
+        }
+
+        return pass;
+    };
+
+    bool pass = test(Row{}, Row{}, Row{}) && test(Row{}, Col{}, Row{}) &&
+                test(Col{}, Row{}, Row{}) && test(Col{}, Col{}, Row{});
+
+    std::cout << "TestGemm ..... " << (pass ? "SUCCESS" : "FAILURE") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/test/gemm_reduce/CMakeLists.txt b/3rdparty/composable_kernel/test/gemm_reduce/CMakeLists.txt
new file mode 100644
index 0000000..349f892
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm_reduce/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_test_executable(test_gemm_reduce_fp16 gemm_reduce_fp16.cpp)
+target_link_libraries(test_gemm_reduce_fp16 PRIVATE utility)
+target_link_libraries(test_gemm_reduce_fp16 PRIVATE device_gemm_reduce_instance)
diff --git a/3rdparty/composable_kernel/test/gemm_reduce/gemm_reduce_fp16.cpp b/3rdparty/composable_kernel/test/gemm_reduce/gemm_reduce_fp16.cpp
new file mode 100644
index 0000000..029165e
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm_reduce/gemm_reduce_fp16.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_gemm_reduce_impl.hpp"
+
+int main()
+{
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    int M = 512;
+    int N = 256;
+    int K = 128;
+
+    bool pass = true;
+
+    pass = pass &&
+           ck::profiler::
+               profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Row, Row, Row>(
+                   true, 1, false, false, M, N, K, K, N, N);
+
+    pass = pass &&
+           ck::profiler::
+               profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Row, Col, Row>(
+                   true, 1, false, false, M, N, K, K, K, N);
+
+    pass = pass &&
+           ck::profiler::
+               profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Col, Row, Row>(
+                   true, 1, false, false, M, N, K, M, N, N);
+
+    pass = pass &&
+           ck::profiler::
+               profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Col, Col, Row>(
+                   true, 1, false, false, M, N, K, M, K, N);
+
+    if(pass)
+    {
+        std::cout << "test GEMM+Reduce fp16: Pass" << std::endl;
+        return 0;
+    }
+    else
+    {
+        std::cout << "test GEMM+Reduce fp16: Fail" << std::endl;
+        return -1;
+    }
+}
diff --git a/3rdparty/composable_kernel/test/gemm_split_k/CMakeLists.txt b/3rdparty/composable_kernel/test/gemm_split_k/CMakeLists.txt
new file mode 100644
index 0000000..793091e
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm_split_k/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_test_executable(test_gemm_split_k gemm_split_k.cpp)
+target_link_libraries(test_gemm_split_k PRIVATE utility)
+target_link_libraries(test_gemm_split_k PRIVATE device_gemm_splitk_instance)
diff --git a/3rdparty/composable_kernel/test/gemm_split_k/gemm_split_k.cpp b/3rdparty/composable_kernel/test/gemm_split_k/gemm_split_k.cpp
new file mode 100644
index 0000000..1edb576
--- /dev/null
+++ b/3rdparty/composable_kernel/test/gemm_split_k/gemm_split_k.cpp
@@ -0,0 +1,261 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+#include "ck/library/utility/host_gemm.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+template <typename T>
+static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
+{
+    float max_diff = 1e-6;
+
+    for(std::size_t i = 0; i < ref.mData.size(); ++i)
+    {
+        float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
+        if(max_diff < diff)
+        {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+struct gemmArgs
+{
+    GemmMatrixLayout layout;
+    int M;
+    int N;
+    int K;
+    int StrideA;
+    int StrideB;
+    int StrideC;
+    int KBatch;
+};
+
+int test_gemm(const gemmArgs& args)
+{
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    bool a_row_major, b_row_major, c_row_major;
+
+    switch(args.layout)
+    {
+    case GemmMatrixLayout::MK_KN_MN:
+        a_row_major = true;
+        b_row_major = true;
+        c_row_major = true;
+        break;
+    case GemmMatrixLayout::MK_NK_MN:
+        a_row_major = true;
+        b_row_major = false;
+        c_row_major = true;
+        break;
+    case GemmMatrixLayout::KM_KN_MN:
+        a_row_major = false;
+        b_row_major = true;
+        c_row_major = true;
+        break;
+    case GemmMatrixLayout::KM_NK_MN:
+        a_row_major = false;
+        b_row_major = false;
+        c_row_major = true;
+        break;
+    default: printf("not supported layout"); return 1;
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, bool row_major) {
+            using namespace ck::literals;
+
+            if(row_major)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<float> a_m_k(f_host_tensor_descriptor(args.M, args.K, args.StrideA, a_row_major));
+    Tensor<float> b_k_n(f_host_tensor_descriptor(args.K, args.N, args.StrideB, b_row_major));
+    Tensor<float> c_m_n_host_result(
+        f_host_tensor_descriptor(args.M, args.N, args.StrideC, c_row_major));
+    Tensor<float> c_m_n_device_result(
+        f_host_tensor_descriptor(args.M, args.N, args.StrideC, c_row_major));
+
+    // init data
+    std::size_t num_thread = 1;
+    a_m_k.GenerateTensorValue(GeneratorTensor_2<float>{-5, 5}, num_thread);
+    b_k_n.GenerateTensorValue(GeneratorTensor_2<float>{-5, 5}, num_thread);
+    // set zero to c_device_buf
+    c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<float>{}, num_thread);
+
+    host_gemm_mk_kn_mn(a_m_k,
+                       b_k_n,
+                       c_m_n_host_result,
+                       ck::tensor_operation::element_wise::PassThrough{},
+                       ck::tensor_operation::element_wise::PassThrough{},
+                       ck::tensor_operation::element_wise::PassThrough{});
+
+    DeviceMem a_device_buf(sizeof(float) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(float) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(float) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    c_device_buf.ToDevice(c_m_n_device_result.mData.data());
+
+    auto test = [&](auto a_layout, auto b_layout, auto c_layout) {
+        bool success = false;
+
+        using DeviceOp = ck::tensor_operation::device::DeviceGemmSplitK<decltype(a_layout),
+                                                                        decltype(b_layout),
+                                                                        decltype(c_layout),
+                                                                        float,
+                                                                        float,
+                                                                        float,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        PassThrough>;
+
+        const auto gemm_ptrs =
+            ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+                DeviceOp>::GetInstances();
+
+        for(auto& gemm_ptr : gemm_ptrs)
+        {
+            auto argument_ptr =
+                gemm_ptr->MakeArgumentPointer(static_cast<float*>(a_device_buf.GetDeviceBuffer()),
+                                              static_cast<float*>(b_device_buf.GetDeviceBuffer()),
+                                              static_cast<float*>(c_device_buf.GetDeviceBuffer()),
+                                              args.M,
+                                              args.N,
+                                              args.K,
+                                              args.StrideA,
+                                              args.StrideB,
+                                              args.StrideC,
+                                              ck::tensor_operation::element_wise::PassThrough{},
+                                              ck::tensor_operation::element_wise::PassThrough{},
+                                              ck::tensor_operation::element_wise::PassThrough{},
+                                              args.KBatch);
+
+            auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+            if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                invoker_ptr->Run(argument_ptr.get());
+
+                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+                if(!check_out(c_m_n_host_result, c_m_n_device_result))
+                {
+                    success = false;
+                    break;
+                }
+                success = true;
+            }
+        }
+
+        return success;
+    };
+
+    bool success = false;
+
+    if(args.layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        success = test(Row{}, Row{}, Row{});
+    }
+    else if(args.layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        success = test(Row{}, Col{}, Row{});
+    }
+    else if(args.layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        success = test(Col{}, Row{}, Row{});
+    }
+    else
+    {
+        success = test(Col{}, Col{}, Row{});
+    }
+
+    auto error_code = 0;
+    if(success)
+    {
+        std::cout << "test split k : Pass" << std::endl;
+    }
+    else
+    {
+        std::cout << "test split k: Fail " << std::endl;
+        error_code = -1; // test needs to report failure
+    }
+    return error_code;
+}
+
+int main(int argc, char* argv[])
+{
+    std::vector<gemmArgs> test_cases;
+    if(argc == 1)
+    {
+        test_cases = {{GemmMatrixLayout::MK_KN_MN, 1024, 1024, 1024, 1024, 1024, 1024, 2},
+                      {GemmMatrixLayout::MK_KN_MN, 1024, 1024, 1024, 1024, 1024, 1024, 8}};
+    }
+    else if(argc == 9)
+    {
+        const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[1]));
+
+        const int M = std::stoi(argv[2]);
+        const int N = std::stoi(argv[3]);
+        const int K = std::stoi(argv[4]);
+
+        const int StrideA = std::stoi(argv[5]);
+        const int StrideB = std::stoi(argv[6]);
+        const int StrideC = std::stoi(argv[7]);
+        const int KBatch  = std::stoi(argv[8]);
+        test_cases        = {{layout, M, N, K, StrideA, StrideB, StrideC, KBatch}};
+    }
+    else
+    {
+        printf("arg1: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg2 to 7: M, N, K, StrideA, StrideB, StrideC KBatch\n");
+        return -1;
+    }
+    bool error = false;
+    for(const auto& kinder : test_cases)
+    {
+        error |= test_gemm(kinder);
+    }
+    return error ? 1 : 0;
+}
diff --git a/3rdparty/composable_kernel/test/grouped_convnd_bwd_weight/CMakeLists.txt b/3rdparty/composable_kernel/test/grouped_convnd_bwd_weight/CMakeLists.txt
new file mode 100644
index 0000000..e2f0790
--- /dev/null
+++ b/3rdparty/composable_kernel/test/grouped_convnd_bwd_weight/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_gtest_executable(test_grouped_convnd_bwd_weight grouped_convnd_bwd_weight.cpp) 
+target_link_libraries(test_grouped_convnd_bwd_weight PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance)
diff --git a/3rdparty/composable_kernel/test/grouped_convnd_bwd_weight/grouped_convnd_bwd_weight.cpp b/3rdparty/composable_kernel/test/grouped_convnd_bwd_weight/grouped_convnd_bwd_weight.cpp
new file mode 100644
index 0000000..6545b6e
--- /dev/null
+++ b/3rdparty/composable_kernel/test/grouped_convnd_bwd_weight/grouped_convnd_bwd_weight.cpp
@@ -0,0 +1,91 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <tuple>
+#include <vector>
+
+#include <gtest/gtest.h>
+
+#include "profiler/profile_grouped_conv_bwd_weight_impl.hpp"
+
+template <typename Tuple>
+class TestGroupedConvndBwdWeight : public ::testing::Test
+{
+    protected:
+    using DataType = std::tuple_element_t<0, Tuple>;
+    std::vector<ck::utils::conv::ConvParam> conv_params;
+    ck::index_t split_k{2};
+
+    template <ck::index_t NDimSpatial>
+    void Run()
+    {
+        for(auto& param : conv_params)
+        {
+            bool pass;
+            EXPECT_FALSE(conv_params.empty());
+            pass = ck::profiler::profile_grouped_conv_bwd_weight_impl<
+                NDimSpatial,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                              ck::tensor_layout::convolution::GNHWC,
+                                              ck::tensor_layout::convolution::GNDHWC>>,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                              ck::tensor_layout::convolution::GKYXC,
+                                              ck::tensor_layout::convolution::GKZYXC>>,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                              ck::tensor_layout::convolution::GNHWK,
+                                              ck::tensor_layout::convolution::GNDHWK>>,
+                DataType,
+                DataType,
+                DataType>(true,  // do_verification
+                          1,     // init_method integer value
+                          false, // do_log
+                          false, // time_kernel
+                          param,
+                          split_k);
+            EXPECT_TRUE(pass);
+        }
+    }
+};
+
+using KernelTypes =
+    ::testing::Types<std::tuple<float>, std::tuple<ck::half_t>, std::tuple<ck::bhalf_t>>;
+TYPED_TEST_SUITE(TestGroupedConvndBwdWeight, KernelTypes);
+
+TYPED_TEST(TestGroupedConvndBwdWeight, Test1D)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back({1, 4, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
+    this->conv_params.push_back({1, 4, 64, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
+    this->conv_params.push_back({1, 4, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
+    this->template Run<1>();
+}
+
+TYPED_TEST(TestGroupedConvndBwdWeight, Test2D)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {2, 4, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    this->conv_params.push_back(
+        {2, 4, 8, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back(
+        {2, 4, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+    this->template Run<2>();
+}
+
+TYPED_TEST(TestGroupedConvndBwdWeight, Test3D)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {3, 4, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->conv_params.push_back(
+        {3, 4, 8, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 4, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->template Run<3>();
+}
diff --git a/3rdparty/composable_kernel/test/grouped_convnd_fwd/CMakeLists.txt b/3rdparty/composable_kernel/test/grouped_convnd_fwd/CMakeLists.txt
new file mode 100644
index 0000000..38da884
--- /dev/null
+++ b/3rdparty/composable_kernel/test/grouped_convnd_fwd/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_gtest_executable(test_grouped_convnd_fwd grouped_convnd_fwd.cpp)
+target_link_libraries(test_grouped_convnd_fwd PRIVATE utility device_grouped_conv1d_fwd_instance device_grouped_conv2d_fwd_instance device_grouped_conv3d_fwd_instance)
+
diff --git a/3rdparty/composable_kernel/test/grouped_convnd_fwd/grouped_convnd_fwd.cpp b/3rdparty/composable_kernel/test/grouped_convnd_fwd/grouped_convnd_fwd.cpp
new file mode 100644
index 0000000..6df7f99
--- /dev/null
+++ b/3rdparty/composable_kernel/test/grouped_convnd_fwd/grouped_convnd_fwd.cpp
@@ -0,0 +1,270 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_grouped_conv_fwd_impl.hpp"
+
+class TestGroupedConvNdFwd : public ::testing::Test
+{
+    protected:
+    std::vector<ck::utils::conv::ConvParam> conv_params;
+};
+
+// 1d GNWC/GKXC/GNWK
+TEST_F(TestGroupedConvNdFwd, GroupedConv1dFwdGNWC)
+{
+    conv_params.clear();
+    conv_params.push_back({1, 2, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
+    conv_params.push_back({1, 2, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
+    conv_params.push_back({1, 2, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
+
+    for(auto& param : conv_params)
+    {
+        bool pass;
+
+        // fp32
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<1,
+                                                           ck::tensor_layout::convolution::GNWC,
+                                                           ck::tensor_layout::convolution::GKXC,
+                                                           ck::tensor_layout::convolution::GNWK,
+                                                           float,
+                                                           float,
+                                                           float>(true,  // do_verification
+                                                                  1,     // init_method
+                                                                  false, // do_log
+                                                                  false, // time_kernel
+                                                                  param);
+
+        EXPECT_TRUE(pass);
+
+        // fp16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<1,
+                                                           ck::tensor_layout::convolution::GNWC,
+                                                           ck::tensor_layout::convolution::GKXC,
+                                                           ck::tensor_layout::convolution::GNWK,
+                                                           ck::half_t,
+                                                           ck::half_t,
+                                                           ck::half_t>(true,  // do_verification
+                                                                       1,     // init_method
+                                                                       false, // do_log
+                                                                       false, // time_kernel
+                                                                       param);
+
+        EXPECT_TRUE(pass);
+
+        // bf16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<1,
+                                                           ck::tensor_layout::convolution::GNWC,
+                                                           ck::tensor_layout::convolution::GKXC,
+                                                           ck::tensor_layout::convolution::GNWK,
+                                                           ck::bhalf_t,
+                                                           ck::bhalf_t,
+                                                           ck::bhalf_t>(true,  // do_verification
+                                                                        1,     // init_method
+                                                                        false, // do_log
+                                                                        false, // time_kernel
+                                                                        param);
+
+        EXPECT_TRUE(pass);
+
+        // int8
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<1,
+                                                           ck::tensor_layout::convolution::GNWC,
+                                                           ck::tensor_layout::convolution::GKXC,
+                                                           ck::tensor_layout::convolution::GNWK,
+                                                           int8_t,
+                                                           int8_t,
+                                                           int8_t>(true,  // do_verification
+                                                                   1,     // init_method
+                                                                   false, // do_log
+                                                                   false, // time_kernel
+                                                                   param);
+
+        EXPECT_TRUE(pass);
+    }
+}
+
+// 2d GNHWC/GKYXC/GNHWK
+TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdGNHWC)
+{
+    conv_params.clear();
+    conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    conv_params.push_back({2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+
+    for(auto& param : conv_params)
+    {
+        bool pass;
+
+        // fp32
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
+                                                           ck::tensor_layout::convolution::GNHWC,
+                                                           ck::tensor_layout::convolution::GKYXC,
+                                                           ck::tensor_layout::convolution::GNHWK,
+                                                           float,
+                                                           float,
+                                                           float>(true,  // do_verification
+                                                                  1,     // init_method
+                                                                  false, // do_log
+                                                                  false, // time_kernel
+                                                                  param);
+
+        EXPECT_TRUE(pass);
+
+        // fp16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
+                                                           ck::tensor_layout::convolution::GNHWC,
+                                                           ck::tensor_layout::convolution::GKYXC,
+                                                           ck::tensor_layout::convolution::GNHWK,
+                                                           ck::half_t,
+                                                           ck::half_t,
+                                                           ck::half_t>(true,  // do_verification
+                                                                       1,     // init_method
+                                                                       false, // do_log
+                                                                       false, // time_kernel
+                                                                       param);
+
+        EXPECT_TRUE(pass);
+
+        // bf16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
+                                                           ck::tensor_layout::convolution::GNHWC,
+                                                           ck::tensor_layout::convolution::GKYXC,
+                                                           ck::tensor_layout::convolution::GNHWK,
+                                                           ck::bhalf_t,
+                                                           ck::bhalf_t,
+                                                           ck::bhalf_t>(true,  // do_verification
+                                                                        1,     // init_method
+                                                                        false, // do_log
+                                                                        false, // time_kernel
+                                                                        param);
+
+        EXPECT_TRUE(pass);
+
+        // int8
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
+                                                           ck::tensor_layout::convolution::GNHWC,
+                                                           ck::tensor_layout::convolution::GKYXC,
+                                                           ck::tensor_layout::convolution::GNHWK,
+                                                           int8_t,
+                                                           int8_t,
+                                                           int8_t>(true,  // do_verification
+                                                                   1,     // init_method
+                                                                   false, // do_log
+                                                                   false, // time_kernel
+                                                                   param);
+
+        EXPECT_TRUE(pass);
+    }
+}
+
+// 3d GNDHWC/GKZYXC/GNDHWK
+TEST_F(TestGroupedConvNdFwd, GroupedConv3dFwdGNDHWC)
+{
+    conv_params.clear();
+    conv_params.push_back(
+        {3, 2, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    conv_params.push_back(
+        {3, 2, 128, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    conv_params.push_back(
+        {3, 2, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+
+    for(auto& param : conv_params)
+    {
+        bool pass;
+
+        // fp32
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<3,
+                                                           ck::tensor_layout::convolution::GNDHWC,
+                                                           ck::tensor_layout::convolution::GKZYXC,
+                                                           ck::tensor_layout::convolution::GNDHWK,
+                                                           float,
+                                                           float,
+                                                           float>(true,  // do_verification
+                                                                  1,     // init_method
+                                                                  false, // do_log
+                                                                  false, // time_kernel
+                                                                  param);
+
+        EXPECT_TRUE(pass);
+
+        // fp16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<3,
+                                                           ck::tensor_layout::convolution::GNDHWC,
+                                                           ck::tensor_layout::convolution::GKZYXC,
+                                                           ck::tensor_layout::convolution::GNDHWK,
+                                                           ck::half_t,
+                                                           ck::half_t,
+                                                           ck::half_t>(true,  // do_verification
+                                                                       1,     // init_method
+                                                                       false, // do_log
+                                                                       false, // time_kernel
+                                                                       param);
+
+        EXPECT_TRUE(pass);
+
+        // bf16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<3,
+                                                           ck::tensor_layout::convolution::GNDHWC,
+                                                           ck::tensor_layout::convolution::GKZYXC,
+                                                           ck::tensor_layout::convolution::GNDHWK,
+                                                           ck::bhalf_t,
+                                                           ck::bhalf_t,
+                                                           ck::bhalf_t>(true,  // do_verification
+                                                                        1,     // init_method
+                                                                        false, // do_log
+                                                                        false, // time_kernel
+                                                                        param);
+
+        EXPECT_TRUE(pass);
+
+        // int8
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<3,
+                                                           ck::tensor_layout::convolution::GNDHWC,
+                                                           ck::tensor_layout::convolution::GKZYXC,
+                                                           ck::tensor_layout::convolution::GNDHWK,
+                                                           int8_t,
+                                                           int8_t,
+                                                           int8_t>(true,  // do_verification
+                                                                   1,     // init_method
+                                                                   false, // do_log
+                                                                   false, // time_kernel
+                                                                   param);
+
+        EXPECT_TRUE(pass);
+    }
+}
+
+// 2d NHWGC/KYXGC/NHWGK
+TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdNHWGC)
+{
+    conv_params.clear();
+    conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    conv_params.push_back({2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+
+    for(auto& param : conv_params)
+    {
+        bool pass;
+
+        // fp16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
+                                                           ck::tensor_layout::convolution::NHWGC,
+                                                           ck::tensor_layout::convolution::KYXGC,
+                                                           ck::tensor_layout::convolution::NHWGK,
+                                                           ck::half_t,
+                                                           ck::half_t,
+                                                           ck::half_t>(true,  // do_verification
+                                                                       1,     // init_method
+                                                                       false, // do_log
+                                                                       false, // time_kernel
+                                                                       param);
+
+        EXPECT_TRUE(pass);
+    }
+}
diff --git a/3rdparty/composable_kernel/test/grouped_gemm/CMakeLists.txt b/3rdparty/composable_kernel/test/grouped_gemm/CMakeLists.txt
new file mode 100644
index 0000000..31a7873
--- /dev/null
+++ b/3rdparty/composable_kernel/test/grouped_gemm/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_test_executable(test_grouped_gemm_fp16 grouped_gemm_fp16.cpp)
+target_link_libraries(test_grouped_gemm_fp16 PRIVATE utility)
+target_link_libraries(test_grouped_gemm_fp16 PRIVATE device_grouped_gemm_instance)
diff --git a/3rdparty/composable_kernel/test/grouped_gemm/grouped_gemm_fp16.cpp b/3rdparty/composable_kernel/test/grouped_gemm/grouped_gemm_fp16.cpp
new file mode 100644
index 0000000..b3f7cca
--- /dev/null
+++ b/3rdparty/composable_kernel/test/grouped_gemm/grouped_gemm_fp16.cpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_grouped_gemm_impl.hpp"
+
+namespace {
+
+using ADataType   = ck::half_t;
+using BDataType   = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <typename ALayout, typename BLayout, typename CLayout>
+bool TestGroupedGemm()
+{
+    int group_count = rand() % 10 + 1;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_c;
+
+    std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideCs;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        Ms.push_back(256 + 256 * (rand() % 10));
+        Ns.push_back(256 + 256 * (rand() % 10));
+        Ks.push_back(128 + 128 * (rand() % 10));
+
+        StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
+        StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
+        StrideCs.push_back(std::is_same<Row, CLayout>::value ? Ns[i] : Ms[i]);
+    }
+
+    return ck::profiler::profile_grouped_gemm_impl<ADataType,
+                                                   BDataType,
+                                                   CDataType,
+                                                   AccDataType,
+                                                   ALayout,
+                                                   BLayout,
+                                                   CLayout>(
+        true, 1, false, 1, Ms, Ns, Ks, StrideAs, StrideBs, StrideCs);
+}
+
+} // anonymous namespace
+
+int main()
+{
+    bool res = true;
+
+    res = res && TestGroupedGemm<Row, Row, Row>();
+    res = res && TestGroupedGemm<Row, Col, Row>();
+    res = res && TestGroupedGemm<Col, Row, Row>();
+    res = res && TestGroupedGemm<Col, Col, Row>();
+
+    std::cout << "TestGroupedGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
+
+    return res ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/test/magic_number_division/CMakeLists.txt b/3rdparty/composable_kernel/test/magic_number_division/CMakeLists.txt
new file mode 100644
index 0000000..e7fc6ee
--- /dev/null
+++ b/3rdparty/composable_kernel/test/magic_number_division/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_test_executable(test_magic_number_division magic_number_division.cpp)
+target_link_libraries(test_magic_number_division PRIVATE utility)
diff --git a/3rdparty/composable_kernel/test/magic_number_division/magic_number_division.cpp b/3rdparty/composable_kernel/test/magic_number_division/magic_number_division.cpp
new file mode 100644
index 0000000..680fddf
--- /dev/null
+++ b/3rdparty/composable_kernel/test/magic_number_division/magic_number_division.cpp
@@ -0,0 +1,150 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/magic_division.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+__global__ void gpu_magic_number_division(uint32_t magic_multiplier,
+                                          uint32_t magic_shift,
+                                          const int32_t* p_dividend,
+                                          int32_t* p_result,
+                                          uint64_t num)
+{
+    uint64_t global_thread_num = blockDim.x * gridDim.x;
+
+    uint64_t global_thread_id = blockIdx.x * blockDim.x + threadIdx.x;
+
+    for(uint64_t data_id = global_thread_id; data_id < num; data_id += global_thread_num)
+    {
+        p_result[data_id] =
+            ck::MagicDivision::DoMagicDivision(p_dividend[data_id], magic_multiplier, magic_shift);
+    }
+}
+
+__global__ void
+gpu_naive_division(int32_t divisor, const int32_t* p_dividend, int32_t* p_result, uint64_t num)
+{
+    uint64_t global_thread_num = blockDim.x * gridDim.x;
+
+    uint64_t global_thread_id = blockIdx.x * blockDim.x + threadIdx.x;
+
+    for(uint64_t data_id = global_thread_id; data_id < num; data_id += global_thread_num)
+    {
+        p_result[data_id] = p_dividend[data_id] / divisor;
+    }
+}
+
+__host__ void cpu_magic_number_division(uint32_t magic_multiplier,
+                                        uint32_t magic_shift,
+                                        const int32_t* p_dividend,
+                                        int32_t* p_result,
+                                        uint64_t num)
+{
+    for(uint64_t data_id = 0; data_id < num; ++data_id)
+    {
+        p_result[data_id] =
+            ck::MagicDivision::DoMagicDivision(p_dividend[data_id], magic_multiplier, magic_shift);
+    }
+}
+
+int main(int, char*[])
+{
+    uint64_t num_divisor  = 4096;
+    uint64_t num_dividend = 1L << 16;
+
+    std::vector<int32_t> divisors_host(num_divisor);
+    std::vector<int32_t> dividends_host(num_dividend);
+
+    // generate divisor
+    for(uint64_t i = 0; i < num_divisor; ++i)
+    {
+        divisors_host[i] = i + 1;
+    }
+
+    // generate dividend
+    for(uint64_t i = 0; i < num_divisor; ++i)
+    {
+        dividends_host[i] = i;
+    }
+
+    DeviceMem dividends_dev_buf(sizeof(int32_t) * num_dividend);
+    DeviceMem naive_result_dev_buf(sizeof(int32_t) * num_dividend);
+    DeviceMem magic_result_dev_buf(sizeof(int32_t) * num_dividend);
+
+    std::vector<int32_t> naive_result_host(num_dividend);
+    std::vector<int32_t> magic_result_host(num_dividend);
+    std::vector<int32_t> magic_result_host2(num_dividend);
+
+    dividends_dev_buf.ToDevice(dividends_host.data());
+
+    bool pass = true;
+
+    for(std::size_t i = 0; i < num_divisor; ++i)
+    {
+        // run naive division on GPU
+        gpu_naive_division<<<1024, 256>>>(
+            divisors_host[i],
+            static_cast<const int32_t*>(dividends_dev_buf.GetDeviceBuffer()),
+            static_cast<int32_t*>(naive_result_dev_buf.GetDeviceBuffer()),
+            num_dividend);
+
+        // calculate magic number
+        uint32_t magic_multiplier, magic_shift;
+
+        ck::tie(magic_multiplier, magic_shift) =
+            ck::MagicDivision::CalculateMagicNumbers(divisors_host[i]);
+
+        // run magic division on GPU
+        gpu_magic_number_division<<<1024, 256>>>(
+            magic_multiplier,
+            magic_shift,
+            static_cast<const int32_t*>(dividends_dev_buf.GetDeviceBuffer()),
+            static_cast<int32_t*>(magic_result_dev_buf.GetDeviceBuffer()),
+            num_dividend);
+
+        naive_result_dev_buf.FromDevice(naive_result_host.data());
+        magic_result_dev_buf.FromDevice(magic_result_host.data());
+
+        bool res = ck::utils::check_err(magic_result_host, naive_result_host);
+
+        if(!res)
+        {
+            pass = false;
+            continue;
+        }
+
+        cpu_magic_number_division(magic_multiplier,
+                                  magic_shift,
+                                  dividends_host.data(),
+                                  magic_result_host2.data(),
+                                  num_dividend);
+
+        res = ck::utils::check_err(magic_result_host2, naive_result_host);
+
+        if(!res)
+        {
+            pass = false;
+            continue;
+        }
+    }
+
+    if(pass)
+    {
+        std::cout << "test magic number division: Pass" << std::endl;
+        return 0;
+    }
+    else
+    {
+        std::cout << "test magic number division: Fail" << std::endl;
+        return -1;
+    }
+}
diff --git a/3rdparty/composable_kernel/test/normalization/CMakeLists.txt b/3rdparty/composable_kernel/test/normalization/CMakeLists.txt
new file mode 100644
index 0000000..456423f
--- /dev/null
+++ b/3rdparty/composable_kernel/test/normalization/CMakeLists.txt
@@ -0,0 +1,17 @@
+add_custom_target(test_layernorm)
+
+add_gtest_executable(test_layernorm2d_fp32 test_layernorm2d_fp32.cpp)
+add_gtest_executable(test_layernorm2d_fp16 test_layernorm2d_fp16.cpp)
+add_gtest_executable(test_groupnorm_fp16 test_groupnorm_fp16.cpp)
+add_gtest_executable(test_groupnorm_fp32 test_groupnorm_fp32.cpp) 
+
+
+target_link_libraries(test_layernorm2d_fp32 PRIVATE utility device_normalization_instance)
+target_link_libraries(test_layernorm2d_fp16 PRIVATE utility device_normalization_instance)
+target_link_libraries(test_groupnorm_fp16 PRIVATE utility device_normalization_instance)
+target_link_libraries(test_groupnorm_fp32 PRIVATE utility device_normalization_instance)
+
+add_dependencies(test_layernorm test_layernorm2d_fp32)
+add_dependencies(test_layernorm test_layernorm2d_fp16)
+add_dependencies(test_layernorm test_groupnorm_fp16)
+add_dependencies(test_layernorm test_groupnorm_fp32)
diff --git a/3rdparty/composable_kernel/test/normalization/test_groupnorm_fp16.cpp b/3rdparty/composable_kernel/test/normalization/test_groupnorm_fp16.cpp
new file mode 100644
index 0000000..636e522
--- /dev/null
+++ b/3rdparty/composable_kernel/test/normalization/test_groupnorm_fp16.cpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_groupnorm_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestGroupnorm : public ::testing::Test
+{
+    protected:
+    using XDataType     = std::tuple_element_t<0, Tuple>;
+    using GammaDataType = std::tuple_element_t<1, Tuple>;
+    using BetaDataType  = std::tuple_element_t<2, Tuple>;
+    using AccDataType   = std::tuple_element_t<3, Tuple>;
+    using YDataType     = std::tuple_element_t<4, Tuple>;
+
+    void Run()
+    {
+        // [N, H, W, G, C], reduce H, W, C
+        std::vector<std::vector<ck::index_t>> lengths = {{1, 1, 1, 1, 1},
+                                                         {1, 2, 3, 4, 5},
+                                                         {256, 9, 9, 9, 9},
+                                                         {1, 64, 64, 32, 10},
+                                                         {1, 32, 32, 32, 20},
+                                                         {2, 32, 32, 32, 30},
+                                                         {2, 32, 32, 32, 40},
+                                                         {1, 16, 16, 32, 40}};
+
+        for(auto length : lengths)
+        {
+            bool success =
+                ck::profiler::profile_groupnorm_impl<XDataType,
+                                                     GammaDataType,
+                                                     BetaDataType,
+                                                     AccDataType,
+                                                     YDataType>(true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType>
+    std::tuple<F16, F16, F16, F32, F16>>;
+
+TYPED_TEST_SUITE(TestGroupnorm, KernelTypes);
+TYPED_TEST(TestGroupnorm, Test_FP16) { this->Run(); }
diff --git a/3rdparty/composable_kernel/test/normalization/test_groupnorm_fp32.cpp b/3rdparty/composable_kernel/test/normalization/test_groupnorm_fp32.cpp
new file mode 100644
index 0000000..ef49266
--- /dev/null
+++ b/3rdparty/composable_kernel/test/normalization/test_groupnorm_fp32.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_groupnorm_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestGroupnorm : public ::testing::Test
+{
+    protected:
+    using XDataType     = std::tuple_element_t<0, Tuple>;
+    using GammaDataType = std::tuple_element_t<1, Tuple>;
+    using BetaDataType  = std::tuple_element_t<2, Tuple>;
+    using AccDataType   = std::tuple_element_t<3, Tuple>;
+    using YDataType     = std::tuple_element_t<4, Tuple>;
+
+    void Run()
+    {
+        // [N, H, W, G, C], reduce H, W, C
+        std::vector<std::vector<ck::index_t>> lengths = {{1, 1, 1, 1, 1},
+                                                         {1, 2, 3, 4, 5},
+                                                         {256, 9, 9, 9, 9},
+                                                         {1, 64, 64, 32, 10},
+                                                         {1, 32, 32, 32, 20},
+                                                         {1, 16, 16, 32, 40}};
+
+        for(auto length : lengths)
+        {
+            bool success =
+                ck::profiler::profile_groupnorm_impl<XDataType,
+                                                     GammaDataType,
+                                                     BetaDataType,
+                                                     AccDataType,
+                                                     YDataType>(true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType>
+    std::tuple<F32, F32, F32, F32, F32>>;
+
+TYPED_TEST_SUITE(TestGroupnorm, KernelTypes);
+TYPED_TEST(TestGroupnorm, Test_FP32) { this->Run(); }
diff --git a/3rdparty/composable_kernel/test/normalization/test_layernorm2d_fp16.cpp b/3rdparty/composable_kernel/test/normalization/test_layernorm2d_fp16.cpp
new file mode 100644
index 0000000..eeb8ec1
--- /dev/null
+++ b/3rdparty/composable_kernel/test/normalization/test_layernorm2d_fp16.cpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_layernorm_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestLayernorm2d : public ::testing::Test
+{
+    protected:
+    using XDataType     = std::tuple_element_t<0, Tuple>;
+    using GammaDataType = std::tuple_element_t<1, Tuple>;
+    using BetaDataType  = std::tuple_element_t<2, Tuple>;
+    using AccDataType   = std::tuple_element_t<3, Tuple>;
+    using YDataType     = std::tuple_element_t<4, Tuple>;
+
+    void Run()
+    {
+        // [N, D], reduce D
+        std::vector<std::vector<ck::index_t>> lengths = {
+            {4, 256}, {8, 511}, {9, 1032}, {4, 2048}, {1, 8192}, {4000, 2000}};
+
+        for(auto length : lengths)
+        {
+            bool success = ck::profiler::profile_layernorm_impl<XDataType,
+                                                                GammaDataType,
+                                                                BetaDataType,
+                                                                AccDataType,
+                                                                YDataType,
+                                                                2>(true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType>
+    std::tuple<F16, F16, F16, F32, F16>>;
+
+TYPED_TEST_SUITE(TestLayernorm2d, KernelTypes);
+TYPED_TEST(TestLayernorm2d, Test_FP16) { this->Run(); }
diff --git a/3rdparty/composable_kernel/test/normalization/test_layernorm2d_fp32.cpp b/3rdparty/composable_kernel/test/normalization/test_layernorm2d_fp32.cpp
new file mode 100644
index 0000000..f555b42
--- /dev/null
+++ b/3rdparty/composable_kernel/test/normalization/test_layernorm2d_fp32.cpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_layernorm_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestLayernorm2d : public ::testing::Test
+{
+    protected:
+    using XDataType     = std::tuple_element_t<0, Tuple>;
+    using GammaDataType = std::tuple_element_t<1, Tuple>;
+    using BetaDataType  = std::tuple_element_t<2, Tuple>;
+    using AccDataType   = std::tuple_element_t<3, Tuple>;
+    using YDataType     = std::tuple_element_t<4, Tuple>;
+
+    void Run()
+    {
+        // [N, D], reduce D
+        std::vector<std::vector<ck::index_t>> lengths = {
+            {4, 256}, {8, 511}, {9, 1032}, {4, 2048}, {1, 8192}, {4000, 2000}};
+
+        for(auto length : lengths)
+        {
+            bool success = ck::profiler::profile_layernorm_impl<XDataType,
+                                                                GammaDataType,
+                                                                BetaDataType,
+                                                                AccDataType,
+                                                                YDataType,
+                                                                2>(true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType>
+    std::tuple<F32, F32, F32, F32, F32>>;
+
+TYPED_TEST_SUITE(TestLayernorm2d, KernelTypes);
+TYPED_TEST(TestLayernorm2d, Test_FP32) { this->Run(); }
diff --git a/3rdparty/composable_kernel/test/reduce/CMakeLists.txt b/3rdparty/composable_kernel/test/reduce/CMakeLists.txt
new file mode 100644
index 0000000..fb43616
--- /dev/null
+++ b/3rdparty/composable_kernel/test/reduce/CMakeLists.txt
@@ -0,0 +1,7 @@
+add_test_executable(test_reduce_no_index reduce_no_index.cpp)
+add_test_executable(test_reduce_with_index reduce_with_index.cpp)
+target_link_libraries(test_reduce_no_index PRIVATE utility)
+target_link_libraries(test_reduce_no_index PRIVATE device_reduce_instance)
+target_link_libraries(test_reduce_with_index PRIVATE utility)
+target_link_libraries(test_reduce_with_index PRIVATE device_reduce_instance)
+
diff --git a/3rdparty/composable_kernel/test/reduce/reduce_no_index.cpp b/3rdparty/composable_kernel/test/reduce/reduce_no_index.cpp
new file mode 100644
index 0000000..3f4d067
--- /dev/null
+++ b/3rdparty/composable_kernel/test/reduce/reduce_no_index.cpp
@@ -0,0 +1,248 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <getopt.h>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "profiler/profile_reduce_impl.hpp"
+
+using namespace ck;
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"reduceDimensions", required_argument, nullptr, 'R'},
+                                       {"scales", required_argument, nullptr, 'S'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths;
+    std::vector<int> reduceDims;
+    std::vector<float> scales;
+
+    int data_type;
+    int init_method = 1;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths "
+                     "(only 4-d tensor supported)"
+                  << std::endl;
+        std::cout << "--reduceDimensions or -R comma seperated list of dimension indexes to reduce "
+                     "(only 1 or 3 or 4 dimensions supported)"
+                  << std::endl;
+        std::cout << "--scales or -S, comma separated two float values for alpha and beta"
+                  << std::endl;
+        std::cout << "Arg1 -- data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2 -- init method(0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:S:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'S':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                scales = getTypeValuesFromString<float>(optarg);
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 2 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind]);
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        if(inLengths.size() != 4 ||
+           (reduceDims.size() != 1 && reduceDims.size() != 3 && reduceDims.size() != 4))
+            return (-1);
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return (-1);
+
+        return (0);
+    };
+};
+
+bool test_reduce_no_index(int data_type,
+                          int init_method,
+                          std::vector<int> reduceDims,
+                          std::vector<size_t> inLengths,
+                          ReduceTensorOp reduceOpId,
+                          bool propagateNan,
+                          float alpha,
+                          float beta)
+{
+    using ck::profiler::profile_reduce_impl;
+
+    bool result = true;
+
+    if(data_type == 0)
+    {
+        result = profile_reduce_impl<float, float, float>(true,
+                                                          init_method,
+                                                          false,
+                                                          false,
+                                                          inLengths,
+                                                          reduceDims,
+                                                          reduceOpId,
+                                                          propagateNan,
+                                                          false,
+                                                          alpha,
+                                                          beta);
+    }
+    else if(data_type == 1)
+    {
+        result = profile_reduce_impl<ck::half_t, float, ck::half_t>(true,
+                                                                    init_method,
+                                                                    false,
+                                                                    false,
+                                                                    inLengths,
+                                                                    reduceDims,
+                                                                    reduceOpId,
+                                                                    propagateNan,
+                                                                    false,
+                                                                    alpha,
+                                                                    beta);
+    }
+    else if(data_type == 3)
+    {
+        result = profile_reduce_impl<int8_t, int32_t, int8_t>(true,
+                                                              init_method,
+                                                              false,
+                                                              false,
+                                                              inLengths,
+                                                              reduceDims,
+                                                              reduceOpId,
+                                                              propagateNan,
+                                                              false,
+                                                              alpha,
+                                                              beta);
+    }
+    else if(data_type == 5)
+    {
+        result = profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(true,
+                                                                      init_method,
+                                                                      false,
+                                                                      false,
+                                                                      inLengths,
+                                                                      reduceDims,
+                                                                      reduceOpId,
+                                                                      propagateNan,
+                                                                      false,
+                                                                      alpha,
+                                                                      beta);
+    }
+    else if(data_type == 6)
+    {
+        result = profile_reduce_impl<double, double, double>(true,
+                                                             init_method,
+                                                             false,
+                                                             false,
+                                                             inLengths,
+                                                             reduceDims,
+                                                             reduceOpId,
+                                                             propagateNan,
+                                                             false,
+                                                             alpha,
+                                                             beta);
+    }
+
+    return (result);
+};
+
+constexpr ReduceTensorOp reduceOpId = ReduceTensorOp::AVG;
+constexpr bool propagateNan         = false;
+
+int main(int argc, char* argv[])
+{
+    SimpleAppArgs args;
+
+    bool result = true;
+
+    if(argc == 1)
+    {
+        int data_type   = 1;
+        int init_method = 2;
+        std::vector<size_t> inLengths{64, 4, 280, 80};
+        std::vector<std::vector<int>> v_reduceDims{
+            {0, 1, 2, 3}, {0, 1, 2}, {1, 2, 3}, {0, 1, 3}, {0, 2, 3}, {0}, {1}, {2}, {3}};
+
+        for(auto& reduceDims : v_reduceDims)
+            result = result && test_reduce_no_index(data_type,
+                                                    init_method,
+                                                    reduceDims,
+                                                    inLengths,
+                                                    reduceOpId,
+                                                    propagateNan,
+                                                    1.0f,
+                                                    0.0f);
+    }
+    else
+    {
+        if(args.processArgs(argc, argv) < 0)
+        {
+            throw std::runtime_error(
+                "Invalid input arguments, test_reduce_no_index could not be executed!");
+        };
+
+        result = test_reduce_no_index(args.data_type,
+                                      args.init_method,
+                                      args.reduceDims,
+                                      args.inLengths,
+                                      reduceOpId,
+                                      propagateNan,
+                                      args.scales[0],
+                                      args.scales[1]);
+    }
+
+    std::cout << "test_reduce_no_index ..... " << (result ? "SUCCESS" : "FAILURE") << std::endl;
+
+    return (result ? 0 : -1);
+}
diff --git a/3rdparty/composable_kernel/test/reduce/reduce_with_index.cpp b/3rdparty/composable_kernel/test/reduce/reduce_with_index.cpp
new file mode 100644
index 0000000..c616a68
--- /dev/null
+++ b/3rdparty/composable_kernel/test/reduce/reduce_with_index.cpp
@@ -0,0 +1,248 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <getopt.h>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "profiler/profile_reduce_impl.hpp"
+
+using namespace ck;
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"reduceDimensions", required_argument, nullptr, 'R'},
+                                       {"scales", required_argument, nullptr, 'S'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths;
+    std::vector<int> reduceDims;
+    std::vector<float> scales;
+
+    int data_type;
+    int init_method = 1;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths "
+                     "(only 4-d tensor supported)"
+                  << std::endl;
+        std::cout << "--reduceDimensions or -R comma seperated list of dimension indexes to reduce "
+                     "(only 1 or 3 or 4 dimensions supported)"
+                  << std::endl;
+        std::cout << "--scales or -S, comma separated two float values for alpha and beta"
+                  << std::endl;
+        std::cout << "Arg1 -- data type (1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2 -- init method(0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:S:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'S':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                scales = getTypeValuesFromString<float>(optarg);
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 2 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind]);
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        if(inLengths.size() != 4 ||
+           (reduceDims.size() != 1 && reduceDims.size() != 3 && reduceDims.size() != 4))
+            return (-1);
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return (-1);
+
+        return (0);
+    };
+};
+
+bool test_reduce_with_index(int data_type,
+                            int init_method,
+                            std::vector<int> reduceDims,
+                            std::vector<size_t> inLengths,
+                            ReduceTensorOp reduceOpId,
+                            bool propagateNan,
+                            float alpha,
+                            float beta)
+{
+    using ck::profiler::profile_reduce_impl;
+
+    bool result = true;
+
+    if(data_type == 0)
+    {
+        result = profile_reduce_impl<float, float, float>(true,
+                                                          init_method,
+                                                          false,
+                                                          false,
+                                                          inLengths,
+                                                          reduceDims,
+                                                          reduceOpId,
+                                                          propagateNan,
+                                                          true,
+                                                          alpha,
+                                                          beta);
+    }
+    else if(data_type == 1)
+    {
+        result = profile_reduce_impl<ck::half_t, ck::half_t, ck::half_t>(true,
+                                                                         init_method,
+                                                                         false,
+                                                                         false,
+                                                                         inLengths,
+                                                                         reduceDims,
+                                                                         reduceOpId,
+                                                                         propagateNan,
+                                                                         true,
+                                                                         alpha,
+                                                                         beta);
+    }
+    else if(data_type == 3)
+    {
+        result = profile_reduce_impl<int8_t, int8_t, int8_t>(true,
+                                                             init_method,
+                                                             false,
+                                                             false,
+                                                             inLengths,
+                                                             reduceDims,
+                                                             reduceOpId,
+                                                             propagateNan,
+                                                             true,
+                                                             alpha,
+                                                             beta);
+    }
+    else if(data_type == 5)
+    {
+        result = profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(true,
+                                                                      init_method,
+                                                                      false,
+                                                                      false,
+                                                                      inLengths,
+                                                                      reduceDims,
+                                                                      reduceOpId,
+                                                                      propagateNan,
+                                                                      true,
+                                                                      alpha,
+                                                                      beta);
+    }
+    else if(data_type == 6)
+    {
+        result = profile_reduce_impl<double, double, double>(true,
+                                                             init_method,
+                                                             false,
+                                                             false,
+                                                             inLengths,
+                                                             reduceDims,
+                                                             reduceOpId,
+                                                             propagateNan,
+                                                             true,
+                                                             alpha,
+                                                             beta);
+    }
+
+    return (result);
+};
+
+constexpr ReduceTensorOp reduceOpId = ReduceTensorOp::AMAX;
+constexpr bool propagateNan         = false;
+
+int main(int argc, char* argv[])
+{
+    SimpleAppArgs args;
+
+    bool result = true;
+
+    if(argc == 1)
+    {
+        int data_type   = 1;
+        int init_method = 2;
+        std::vector<size_t> inLengths{64, 4, 280, 80};
+        std::vector<std::vector<int>> v_reduceDims{
+            {0, 1, 2, 3}, {0, 1, 2}, {1, 2, 3}, {0, 1, 3}, {0, 2, 3}, {0}, {1}, {2}, {3}};
+
+        for(auto& reduceDims : v_reduceDims)
+            result = result && test_reduce_with_index(data_type,
+                                                      init_method,
+                                                      reduceDims,
+                                                      inLengths,
+                                                      reduceOpId,
+                                                      propagateNan,
+                                                      1.0f,
+                                                      0.0f);
+    }
+    else
+    {
+        if(args.processArgs(argc, argv) < 0)
+        {
+            throw std::runtime_error(
+                "Invalid input arguments, test_reduce_with_index could not be executed!");
+        };
+
+        result = test_reduce_with_index(args.data_type,
+                                        args.init_method,
+                                        args.reduceDims,
+                                        args.inLengths,
+                                        reduceOpId,
+                                        propagateNan,
+                                        args.scales[0],
+                                        args.scales[1]);
+    }
+
+    std::cout << "test_reduce_with_index ..... " << (result ? "SUCCESS" : "FAILURE") << std::endl;
+
+    return (result ? 0 : -1);
+}
diff --git a/3rdparty/composable_kernel/test/reference_conv_fwd/CMakeLists.txt b/3rdparty/composable_kernel/test/reference_conv_fwd/CMakeLists.txt
new file mode 100644
index 0000000..b40b9a1
--- /dev/null
+++ b/3rdparty/composable_kernel/test/reference_conv_fwd/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_gtest_executable(test_reference_conv_fwd reference_conv_fwd.cpp)
+target_link_libraries(test_reference_conv_fwd PRIVATE utility)
diff --git a/3rdparty/composable_kernel/test/reference_conv_fwd/reference_conv_fwd.cpp b/3rdparty/composable_kernel/test/reference_conv_fwd/reference_conv_fwd.cpp
new file mode 100644
index 0000000..1f9ba00
--- /dev/null
+++ b/3rdparty/composable_kernel/test/reference_conv_fwd/reference_conv_fwd.cpp
@@ -0,0 +1,392 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cmath>
+#include <cstdlib>
+#include <numeric>
+#include <type_traits>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+namespace {
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+template <ck::index_t NDimSpatial,
+          typename InDataType    = float,
+          typename WeiDataType   = float,
+          typename OutDataType   = float,
+          typename InLayout      = ck::tensor_layout::convolution::GNHWC,
+          typename WeiLayout     = ck::tensor_layout::convolution::GKYXC,
+          typename OutLayout     = ck::tensor_layout::convolution::GNHWK,
+          typename FillInputOp   = ck::utils::FillMonotonicSeq<InDataType>,
+          typename FillWeightsOp = ck::utils::FillConstant<WeiDataType>>
+Tensor<OutDataType>
+run_reference_convolution_forward(const ck::utils::conv::ConvParam& conv_param,
+                                  const FillInputOp& fill_input_op     = FillInputOp{},
+                                  const FillWeightsOp& fill_weights_op = FillWeightsOp{0.5f})
+{
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    Tensor<InDataType> input(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weights(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> host_output(out_g_n_k_wos_desc);
+
+    fill_input_op(input.begin(), input.end());
+    fill_weights_op(weights.begin(), weights.end());
+    ck::ranges::fill<OutDataType>(host_output, 0.f);
+
+    auto ref_conv     = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                 InDataType,
+                                                                 WeiDataType,
+                                                                 OutDataType,
+                                                                 InElementOp,
+                                                                 WeiElementOp,
+                                                                 OutElementOp>();
+    auto ref_invoker  = ref_conv.MakeInvoker();
+    auto ref_argument = ref_conv.MakeArgument(input,
+                                              weights,
+                                              host_output,
+                                              conv_param.conv_filter_strides_,
+                                              conv_param.conv_filter_dilations_,
+                                              conv_param.input_left_pads_,
+                                              conv_param.input_right_pads_,
+                                              InElementOp{},
+                                              WeiElementOp{},
+                                              OutElementOp{});
+
+    ref_invoker.Run(ref_argument);
+    return host_output;
+}
+
+} // anonymous namespace
+
+// Eeference convolution assume dimensions of tensor descriptors are in GNCDHW/GKCZYX/GNKDHW order,
+// regardless of physical tensor layouts in  memory.
+// Some tests below assume dimensions of tensor descriptors can be in other order, and therefore
+// are disabled
+// TODO: add more tests, which comply with assumption about dimension order of reference convolution
+// and add tests for more physical layout
+#if 0
+TEST(ReferenceConvolutionFWD, Conv2DGNHWC)
+{
+    ck::utils::conv::ConvParam conv_param(2,
+                                          1,
+                                          1,
+                                          1,
+                                          2,
+                                          std::vector<ck::index_t>{3, 3},
+                                          std::vector<ck::index_t>{6, 6},
+                                          std::vector<ck::index_t>{1, 1},
+                                          std::vector<ck::index_t>{1, 1},
+                                          std::vector<ck::index_t>{0, 0},
+                                          std::vector<ck::index_t>{0, 0});
+
+    auto out_tensor = run_reference_convolution_forward<2>(conv_param);
+    std::vector<std::size_t> ref_dims{1, 1, 4, 4, 1};
+    std::vector<float> ref_data{130.5,
+                                148.5,
+                                166.5,
+                                184.5,
+                                238.5,
+                                256.5,
+                                274.5,
+                                292.5,
+                                346.5,
+                                364.5,
+                                382.5,
+                                400.5,
+                                454.5,
+                                472.5,
+                                490.5,
+                                508.5};
+    EXPECT_TRUE(ck::utils::check_err(
+        out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(out_tensor, ref_data, "Error: incorrect results!"));
+}
+
+TEST(ReferenceConvolutionFWD, Conv2DGNHWCStridesDilationsPadding)
+{
+    ck::utils::conv::ConvParam conv_param(2,
+                                          1,
+                                          1,
+                                          2,
+                                          2,
+                                          std::vector<ck::index_t>{3, 3},
+                                          std::vector<ck::index_t>{12, 12},
+                                          std::vector<ck::index_t>{2, 2},
+                                          std::vector<ck::index_t>{2, 2},
+                                          std::vector<ck::index_t>{1, 1},
+                                          std::vector<ck::index_t>{1, 1});
+
+    auto out_tensor                   = run_reference_convolution_forward<2>(conv_param);
+    std::vector<std::size_t> ref_dims = std::vector<std::size_t>{1, 5, 5, 2};
+    std::vector<float> ref_data{
+        210.,  210.,  327.,   327.,   351.,   351.,   375.,   375.,   399.,   399.,
+        459.,  459.,  706.5,  706.5,  742.5,  742.5,  778.5,  778.5,  814.5,  814.5,
+        747.,  747.,  1138.5, 1138.5, 1174.5, 1174.5, 1210.5, 1210.5, 1246.5, 1246.5,
+        1035., 1035., 1570.5, 1570.5, 1606.5, 1606.5, 1642.5, 1642.5, 1678.5, 1678.5,
+        1323., 1323., 2002.5, 2002.5, 2038.5, 2038.5, 2074.5, 2074.5, 2110.5, 2110.5};
+    EXPECT_TRUE(ck::utils::check_err(
+        out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(out_tensor, ref_data, "Error: incorrect results!"));
+}
+
+TEST(ReferenceConvolutionFWD, Conv1DGNWC)
+{
+    ck::utils::conv::ConvParam conv_param(1,
+                                          1,
+                                          1,
+                                          1,
+                                          2,
+                                          std::vector<ck::index_t>{3},
+                                          std::vector<ck::index_t>{6},
+                                          std::vector<ck::index_t>{1},
+                                          std::vector<ck::index_t>{1},
+                                          std::vector<ck::index_t>{0},
+                                          std::vector<ck::index_t>{0});
+
+    auto out_tensor =
+        run_reference_convolution_forward<1,
+                                          float,
+                                          float,
+                                          float,
+                                          ck::tensor_layout::convolution::GNWC,
+                                          ck::tensor_layout::convolution::GKXC,
+                                          ck::tensor_layout::convolution::GNWK>(conv_param);
+    std::vector<std::size_t> ref_dims{1, 1, 4, 1};
+    std::vector<float> ref_data{7.5, 13.5, 19.5, 25.5};
+    EXPECT_TRUE(ck::utils::check_err(
+        out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(out_tensor, ref_data, "Error: incorrect results!"));
+}
+
+TEST(ReferenceConvolutionFWD, Conv1DGNWCStridesDilationsPadding)
+{
+    ck::utils::conv::ConvParam conv_param(1,
+                                          1,
+                                          1,
+                                          2,
+                                          2,
+                                          std::vector<ck::index_t>{3},
+                                          std::vector<ck::index_t>{12},
+                                          std::vector<ck::index_t>{2},
+                                          std::vector<ck::index_t>{2},
+                                          std::vector<ck::index_t>{1},
+                                          std::vector<ck::index_t>{1});
+
+    auto out_tensor =
+        run_reference_convolution_forward<1,
+                                          float,
+                                          float,
+                                          float,
+                                          ck::tensor_layout::convolution::GNWC,
+                                          ck::tensor_layout::convolution::GKXC,
+                                          ck::tensor_layout::convolution::GNWK>(conv_param);
+    std::vector<std::size_t> ref_dims{1, 1, 5, 2};
+    std::vector<float> ref_data{9., 9., 19.5, 19.5, 31.5, 31.5, 43.5, 43.5, 55.5, 55.5};
+    EXPECT_TRUE(ck::utils::check_err(
+        out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(out_tensor, ref_data, "Error: incorrect results!"));
+}
+
+TEST(ReferenceConvolutionFWD, Conv1DGNWCSameOutputSize)
+{
+    ck::utils::conv::ConvParam conv_param(1,
+                                          1,
+                                          2,
+                                          16,
+                                          4,
+                                          std::vector<ck::index_t>{3},
+                                          std::vector<ck::index_t>{16},
+                                          std::vector<ck::index_t>{1},
+                                          std::vector<ck::index_t>{1},
+                                          std::vector<ck::index_t>{1},
+                                          std::vector<ck::index_t>{1});
+
+    auto out_tensor2 = run_reference_convolution_forward<1,
+                                                         float,
+                                                         float,
+                                                         float,
+                                                         ck::tensor_layout::convolution::GNWC,
+                                                         ck::tensor_layout::convolution::GKXC,
+                                                         ck::tensor_layout::convolution::GNWK>(
+        conv_param, ck::utils::FillMonotonicSeq<float>{0.f, 0.1f});
+
+    std::vector<std::size_t> ref_dims{1, 2, 16, 16};
+    std::vector<float> ref_data{
+        1.4,       1.4,       1.4,       1.4,       1.4,       1.4,       1.4,       1.4,
+        1.4,       1.4,       1.4,       1.4,       1.4,       1.4,       1.4,       1.4,
+        3.3,       3.3,       3.3,       3.3,       3.3,       3.3,       3.3,       3.3,
+        3.3,       3.3,       3.3,       3.3,       3.3,       3.3,       3.3,       3.3,
+        5.7,       5.7,       5.7,       5.7,       5.7,       5.7,       5.7,       5.7,
+        5.7,       5.7,       5.7,       5.7,       5.7,       5.7,       5.7,       5.7,
+        8.1,       8.1,       8.1,       8.1,       8.1,       8.1,       8.1,       8.1,
+        8.1,       8.1,       8.1,       8.1,       8.1,       8.1,       8.1,       8.1,
+        10.5,      10.5,      10.5,      10.5,      10.5,      10.5,      10.5,      10.5,
+        10.5,      10.5,      10.5,      10.5,      10.5,      10.5,      10.5,      10.5,
+        12.900001, 12.900001, 12.900001, 12.900001, 12.900001, 12.900001, 12.900001, 12.900001,
+        12.900001, 12.900001, 12.900001, 12.900001, 12.900001, 12.900001, 12.900001, 12.900001,
+        15.3,      15.3,      15.3,      15.3,      15.3,      15.3,      15.3,      15.3,
+        15.3,      15.3,      15.3,      15.3,      15.3,      15.3,      15.3,      15.3,
+        17.7,      17.7,      17.7,      17.7,      17.7,      17.7,      17.7,      17.7,
+        17.7,      17.7,      17.7,      17.7,      17.7,      17.7,      17.7,      17.7,
+        20.1,      20.1,      20.1,      20.1,      20.1,      20.1,      20.1,      20.1,
+        20.1,      20.1,      20.1,      20.1,      20.1,      20.1,      20.1,      20.1,
+        22.5,      22.5,      22.5,      22.5,      22.5,      22.5,      22.5,      22.5,
+        22.5,      22.5,      22.5,      22.5,      22.5,      22.5,      22.5,      22.5,
+        24.900002, 24.900002, 24.900002, 24.900002, 24.900002, 24.900002, 24.900002, 24.900002,
+        24.900002, 24.900002, 24.900002, 24.900002, 24.900002, 24.900002, 24.900002, 24.900002,
+        27.300001, 27.300001, 27.300001, 27.300001, 27.300001, 27.300001, 27.300001, 27.300001,
+        27.300001, 27.300001, 27.300001, 27.300001, 27.300001, 27.300001, 27.300001, 27.300001,
+        29.7,      29.7,      29.7,      29.7,      29.7,      29.7,      29.7,      29.7,
+        29.7,      29.7,      29.7,      29.7,      29.7,      29.7,      29.7,      29.7,
+        32.100002, 32.100002, 32.100002, 32.100002, 32.100002, 32.100002, 32.100002, 32.100002,
+        32.100002, 32.100002, 32.100002, 32.100002, 32.100002, 32.100002, 32.100002, 32.100002,
+        34.5,      34.5,      34.5,      34.5,      34.5,      34.5,      34.5,      34.5,
+        34.5,      34.5,      34.5,      34.5,      34.5,      34.5,      34.5,      34.5,
+        23.8,      23.8,      23.8,      23.8,      23.8,      23.8,      23.8,      23.8,
+        23.8,      23.8,      23.8,      23.8,      23.8,      23.8,      23.8,      23.8,
+        27.,       27.,       27.,       27.,       27.,       27.,       27.,       27.,
+        27.,       27.,       27.,       27.,       27.,       27.,       27.,       27.,
+        41.7,      41.7,      41.7,      41.7,      41.7,      41.7,      41.7,      41.7,
+        41.7,      41.7,      41.7,      41.7,      41.7,      41.7,      41.7,      41.7,
+        44.100002, 44.100002, 44.100002, 44.100002, 44.100002, 44.100002, 44.100002, 44.100002,
+        44.100002, 44.100002, 44.100002, 44.100002, 44.100002, 44.100002, 44.100002, 44.100002,
+        46.5,      46.5,      46.5,      46.5,      46.5,      46.5,      46.5,      46.5,
+        46.5,      46.5,      46.5,      46.5,      46.5,      46.5,      46.5,      46.5,
+        48.899998, 48.899998, 48.899998, 48.899998, 48.899998, 48.899998, 48.899998, 48.899998,
+        48.899998, 48.899998, 48.899998, 48.899998, 48.899998, 48.899998, 48.899998, 48.899998,
+        51.3,      51.3,      51.3,      51.3,      51.3,      51.3,      51.3,      51.3,
+        51.3,      51.3,      51.3,      51.3,      51.3,      51.3,      51.3,      51.3,
+        53.7,      53.7,      53.7,      53.7,      53.7,      53.7,      53.7,      53.7,
+        53.7,      53.7,      53.7,      53.7,      53.7,      53.7,      53.7,      53.7,
+        56.100002, 56.100002, 56.100002, 56.100002, 56.100002, 56.100002, 56.100002, 56.100002,
+        56.100002, 56.100002, 56.100002, 56.100002, 56.100002, 56.100002, 56.100002, 56.100002,
+        58.5,      58.5,      58.5,      58.5,      58.5,      58.5,      58.5,      58.5,
+        58.5,      58.5,      58.5,      58.5,      58.5,      58.5,      58.5,      58.5,
+        60.899998, 60.899998, 60.899998, 60.899998, 60.899998, 60.899998, 60.899998, 60.899998,
+        60.899998, 60.899998, 60.899998, 60.899998, 60.899998, 60.899998, 60.899998, 60.899998,
+        63.3,      63.3,      63.3,      63.3,      63.3,      63.3,      63.3,      63.3,
+        63.3,      63.3,      63.3,      63.3,      63.3,      63.3,      63.3,      63.3,
+        65.7,      65.7,      65.7,      65.7,      65.7,      65.7,      65.7,      65.7,
+        65.7,      65.7,      65.7,      65.7,      65.7,      65.7,      65.7,      65.7,
+        68.1,      68.1,      68.1,      68.1,      68.1,      68.1,      68.1,      68.1,
+        68.1,      68.1,      68.1,      68.1,      68.1,      68.1,      68.1,      68.1,
+        70.5,      70.5,      70.5,      70.5,      70.5,      70.5,      70.5,      70.5,
+        70.5,      70.5,      70.5,      70.5,      70.5,      70.5,      70.5,      70.5,
+        72.9,      72.9,      72.9,      72.9,      72.9,      72.9,      72.9,      72.9,
+        72.9,      72.9,      72.9,      72.9,      72.9,      72.9,      72.9,      72.9,
+        49.4,      49.4,      49.4,      49.4,      49.4,      49.4,      49.4,      49.4,
+        49.4,      49.4,      49.4,      49.4,      49.4,      49.4,      49.4,      49.4};
+    EXPECT_TRUE(ck::utils::check_err(
+        out_tensor2.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(out_tensor2, ref_data, "Error: incorrect results!"));
+}
+#endif
+
+TEST(ReferenceConvolutionFWD, Conv3DGNCDHW)
+{
+    ck::utils::conv::ConvParam conv_param(3,
+                                          1,
+                                          1,
+                                          1,
+                                          2,
+                                          std::vector<ck::index_t>{3, 3, 3},
+                                          std::vector<ck::index_t>{6, 6, 6},
+                                          std::vector<ck::index_t>{1, 1, 1},
+                                          std::vector<ck::index_t>{1, 1, 1},
+                                          std::vector<ck::index_t>{0, 0, 0},
+                                          std::vector<ck::index_t>{0, 0, 0});
+
+    auto out_tensor = run_reference_convolution_forward<3,
+                                                        float,
+                                                        float,
+                                                        float,
+                                                        ck::tensor_layout::convolution::GNCDHW,
+                                                        ck::tensor_layout::convolution::GKCZYX,
+                                                        ck::tensor_layout::convolution::GNKDHW>(
+        conv_param, ck::utils::FillMonotonicSeq<float>{0.f, 0.1f});
+    std::vector<std::size_t> ref_dims{1, 1, 1, 4, 4, 4};
+    std::vector<float> ref_data{
+        407.7,     410.40002, 413.09998, 415.80002, 423.90002, 426.6,     429.30002, 432.,
+        440.1,     442.80002, 445.5,     448.2,     456.30002, 459.,      461.7,     464.40002,
+        504.90002, 507.6,     510.30002, 513.,      521.1,     523.8,     526.5,     529.2001,
+        537.3,     540.,      542.7001,  545.4,     553.5,     556.2001,  558.9,     561.6,
+        602.10004, 604.8,     607.5,     610.2,     618.3,     621.,      623.7,     626.4,
+        634.5,     637.2,     639.9,     642.60004, 650.7,     653.4,     656.10004, 658.8,
+        699.3,     702.,      704.7,     707.4,     715.5,     718.2,     720.9,     723.60004,
+        731.7,     734.4001,  737.10004, 739.8,     747.9001,  750.60004, 753.3,     756.};
+    EXPECT_TRUE(ck::utils::check_err(out_tensor.mDesc.GetLengths(),
+                                     ref_dims,
+                                     "Error [case 1]: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(out_tensor, ref_data, "Error [case 1]: incorrect results!"));
+}
+
+TEST(ReferenceConvolutionFWD, Conv3DGNCDHWStridesDilations)
+{
+    ck::utils::conv::ConvParam conv_param(3,
+                                          1,
+                                          1,
+                                          2,
+                                          2,
+                                          std::vector<ck::index_t>{3, 3, 3},
+                                          std::vector<ck::index_t>{12, 12, 12},
+                                          std::vector<ck::index_t>{3, 3, 3},
+                                          std::vector<ck::index_t>{1, 1, 1},
+                                          std::vector<ck::index_t>{0, 0, 0},
+                                          std::vector<ck::index_t>{0, 0, 0});
+
+    auto out_tensor = run_reference_convolution_forward<3,
+                                                        float,
+                                                        float,
+                                                        float,
+                                                        ck::tensor_layout::convolution::GNCDHW,
+                                                        ck::tensor_layout::convolution::GKCZYX,
+                                                        ck::tensor_layout::convolution::GNKDHW>(
+        conv_param, ck::utils::FillMonotonicSeq<float>{0.f, 0.1f});
+    std::vector<std::size_t> ref_dims{1, 1, 2, 4, 4, 4};
+    std::vector<float> ref_data{
+        2756.7002, 2764.7998, 2772.9001, 2781.,     2853.9001, 2862.,     2870.1,    2878.2002,
+        2951.1,    2959.2002, 2967.2998, 2975.4001, 3048.2998, 3056.4001, 3064.5,    3072.6,
+        3923.1,    3931.2,    3939.2998, 3947.4,    4020.2998, 4028.4001, 4036.5002, 4044.5999,
+        4117.5,    4125.6,    4133.7,    4141.8,    4214.7,    4222.8,    4230.9004, 4239.,
+        5089.5,    5097.5996, 5105.7,    5113.8,    5186.7,    5194.8,    5202.9,    5211.,
+        5283.9004, 5292.,     5300.0996, 5308.2,    5381.0996, 5389.2,    5397.3,    5405.4004,
+        6255.9004, 6264.0005, 6272.1,    6280.2,    6353.1,    6361.2,    6369.301,  6377.4,
+        6450.301,  6458.4,    6466.5,    6474.6,    6547.5,    6555.6,    6563.699,  6571.801,
+        2756.7002, 2764.7998, 2772.9001, 2781.,     2853.9001, 2862.,     2870.1,    2878.2002,
+        2951.1,    2959.2002, 2967.2998, 2975.4001, 3048.2998, 3056.4001, 3064.5,    3072.6,
+        3923.1,    3931.2,    3939.2998, 3947.4,    4020.2998, 4028.4001, 4036.5002, 4044.5999,
+        4117.5,    4125.6,    4133.7,    4141.8,    4214.7,    4222.8,    4230.9004, 4239.,
+        5089.5,    5097.5996, 5105.7,    5113.8,    5186.7,    5194.8,    5202.9,    5211.,
+        5283.9004, 5292.,     5300.0996, 5308.2,    5381.0996, 5389.2,    5397.3,    5405.4004,
+        6255.9004, 6264.0005, 6272.1,    6280.2,    6353.1,    6361.2,    6369.301,  6377.4,
+        6450.301,  6458.4,    6466.5,    6474.6,    6547.5,    6555.6,    6563.699,  6571.801};
+    EXPECT_TRUE(ck::utils::check_err(out_tensor.mDesc.GetLengths(),
+                                     ref_dims,
+                                     "Error [case 2]: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(
+        out_tensor, ref_data, "Error [case 2]: incorrect results!", 1e-4f, 1e-6f));
+}
diff --git a/3rdparty/composable_kernel/test/softmax/CMakeLists.txt b/3rdparty/composable_kernel/test/softmax/CMakeLists.txt
new file mode 100644
index 0000000..4ba4012
--- /dev/null
+++ b/3rdparty/composable_kernel/test/softmax/CMakeLists.txt
@@ -0,0 +1,11 @@
+add_custom_target(test_softmax)
+
+add_gtest_executable(test_softmax_rank3 test_softmax_rank3.cpp)
+add_gtest_executable(test_softmax_rank4 test_softmax_rank4.cpp)
+add_gtest_executable(test_softmax_interface test_softmax_interface.cpp)
+target_link_libraries(test_softmax_rank3 PRIVATE utility device_softmax_instance)
+target_link_libraries(test_softmax_rank4 PRIVATE utility device_softmax_instance)
+target_link_libraries(test_softmax_interface PRIVATE utility device_softmax_instance)
+add_dependencies(test_softmax test_softmax_rank3)
+add_dependencies(test_softmax test_softmax_rank4)
+add_dependencies(test_softmax test_softmax_interface)
diff --git a/3rdparty/composable_kernel/test/softmax/test_softmax_interface.cpp b/3rdparty/composable_kernel/test/softmax/test_softmax_interface.cpp
new file mode 100644
index 0000000..8cac0ba
--- /dev/null
+++ b/3rdparty/composable_kernel/test/softmax/test_softmax_interface.cpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <stdexcept>
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "test_softmax_util.hpp"
+
+class TestSoftmaxInterface : public ::testing::Test
+{
+    protected:
+    template <ck::index_t Rank, ck::index_t NumReduceDims>
+    using SoftmaxInstance =
+        ck::DeviceSoftmaxInstanceWrapper<Rank, NumReduceDims, 256, 1, 256, 1, 8, 1, 8, 8>;
+};
+
+TEST_F(TestSoftmaxInterface, IncorrectReduceDims)
+{
+    std::vector<ck::index_t> lengths{2, 128, 1536};
+    std::vector<ck::index_t> strides{128 * 1536, 1536, 1};
+
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, strides, {-1})), std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, strides, {3})), std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, strides, {0, 1})),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, strides, {})), std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 2>{}.IsSupported(lengths, strides, {2, -1})),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 2>{}.IsSupported(lengths, strides, {2, 4})),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 2>{}.IsSupported(lengths, strides, {2})), std::runtime_error);
+}
+
+TEST_F(TestSoftmaxInterface, IncorrectLengthsSize)
+{
+    std::vector<ck::index_t> lengths{128, 1536};
+    std::vector<ck::index_t> strides{128 * 1536, 1536, 1};
+    std::vector<ck::index_t> reduce_dims{2};
+
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported({128, 1536}, strides, reduce_dims)),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported({}, strides, reduce_dims)),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported({1, 8, 128, 1536}, strides, reduce_dims)),
+                 std::runtime_error);
+}
+
+TEST_F(TestSoftmaxInterface, IncorrectStridesSize)
+{
+    std::vector<ck::index_t> lengths{2, 128, 1536};
+    std::vector<ck::index_t> reduce_dims{2};
+
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, {1536, 1}, reduce_dims)),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, {}, reduce_dims)),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, {1, 2, 3, 4}, reduce_dims)),
+                 std::runtime_error);
+}
+
+TEST_F(TestSoftmaxInterface, UnsupportedLengths)
+{
+    using SoftmaxInstance1 = ck::DeviceSoftmaxInstanceWrapper<3, 1, 256, 1, 256, 1, 8, 1, 8, 4>;
+    EXPECT_FALSE(SoftmaxInstance1{}.IsSupported({2, 128, 1500}, {128 * 1500, 1500, 1}, {2}));
+    EXPECT_FALSE(SoftmaxInstance1{}.IsSupported({2, 127, 1536}, {127 * 1536, 1536, 1}, {2}));
+    EXPECT_FALSE(SoftmaxInstance1{}.IsSupported({2, 128, 1537}, {128 * 1537, 1537, 1}, {2}));
+
+    // Reduction of middle dimensions
+    using SoftmaxInstance2 = ck::DeviceSoftmaxInstanceWrapper<3, 3, 256, 8, 32, 8, 8, 0, 8, 4>;
+    EXPECT_FALSE(SoftmaxInstance2{}.IsSupported({2, 128, 1536}, {128 * 1536, 1536, 1}, {0, 1, 2}));
+
+    // Reduction of middle dimensions
+    using SoftmaxInstance3 = ck::DeviceSoftmaxInstanceWrapper<3, 1, 256, 8, 32, 8, 8, 0, 4, 8>;
+    EXPECT_FALSE(SoftmaxInstance3{}.IsSupported({2, 128, 1536}, {128 * 1536, 1536, 1}, {2}));
+    EXPECT_FALSE(SoftmaxInstance3{}.IsSupported({2, 128, 1537}, {128 * 1537, 1537, 1}, {1}));
+    EXPECT_FALSE(SoftmaxInstance3{}.IsSupported({2, 128, 1540}, {128 * 1540, 1540, 1}, {1}));
+    EXPECT_FALSE(SoftmaxInstance3{}.IsSupported({2, 127, 1536}, {127 * 1536, 1536, 1}, {1}));
+}
+
+TEST_F(TestSoftmaxInterface, UnsupportedInstance)
+{
+    // Instance with InSrcVectorDim = 1, can't reduce middle dims if in/out vec size != 1
+    using SoftmaxInstance1 = ck::DeviceSoftmaxInstanceWrapper<3, 1, 256, 8, 32, 1, 8, 1, 8, 8>;
+    EXPECT_FALSE(SoftmaxInstance1{}.IsSupported({2, 128, 1024}, {128 * 1024, 1024, 1}, {0}));
+}
diff --git a/3rdparty/composable_kernel/test/softmax/test_softmax_rank3.cpp b/3rdparty/composable_kernel/test/softmax/test_softmax_rank3.cpp
new file mode 100644
index 0000000..5691ee3
--- /dev/null
+++ b/3rdparty/composable_kernel/test/softmax/test_softmax_rank3.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <stdexcept>
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "test_softmax_util.hpp"
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+
+using F16 = ck::half_t;
+using F32 = float;
+using I8  = int8_t;
+
+template <typename Tuple>
+class TestSoftmax : public ck::TestSoftmax<Tuple>
+{
+};
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    //         InDataType, AccDataType, OutDataType, Rank
+    std::tuple<       F16,         F32,         F16,    I<3>>,
+    std::tuple<       F32,         F32,         F32,    I<3>>,
+    std::tuple<        I8,         F32,          I8,    I<3>>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestSoftmax, KernelTypes);
+
+#include "test_softmax_ut_cases.inc"
diff --git a/3rdparty/composable_kernel/test/softmax/test_softmax_rank4.cpp b/3rdparty/composable_kernel/test/softmax/test_softmax_rank4.cpp
new file mode 100644
index 0000000..f0b22df
--- /dev/null
+++ b/3rdparty/composable_kernel/test/softmax/test_softmax_rank4.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <stdexcept>
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "test_softmax_util.hpp"
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+
+using F16 = ck::half_t;
+using F32 = float;
+using I8  = int8_t;
+
+template <typename Tuple>
+class TestSoftmax : public ck::TestSoftmax<Tuple>
+{
+};
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    //         InDataType, AccDataType, OutDataType, Rank
+    std::tuple<       F16,         F32,         F16,    I<4>>,
+    std::tuple<       F32,         F32,         F32,    I<4>>,
+    std::tuple<        I8,         F32,          I8,    I<4>>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestSoftmax, KernelTypes);
+
+#include "test_softmax_ut_cases.inc"
diff --git a/3rdparty/composable_kernel/test/softmax/test_softmax_ut_cases.inc b/3rdparty/composable_kernel/test/softmax/test_softmax_ut_cases.inc
new file mode 100644
index 0000000..cf5e4d2
--- /dev/null
+++ b/3rdparty/composable_kernel/test/softmax/test_softmax_ut_cases.inc
@@ -0,0 +1,60 @@
+#pragma once
+
+TYPED_TEST(TestSoftmax, ReduceOutermostDim)
+{
+    std::vector<ck::index_t> reduce_dims{this->Rank - 1};
+    this->Run(reduce_dims);
+}
+
+TYPED_TEST(TestSoftmax, ReduceMiddleDim)
+{
+    for(int dim = 0; dim < this->Rank - 1; ++dim)
+    {
+        std::vector<ck::index_t> reduce_dims{dim};
+        this->Run(reduce_dims);
+    }
+}
+
+TYPED_TEST(TestSoftmax, ReduceMultipleDimsWithOutermost)
+{
+    for(int dim = 0; dim < this->Rank - 1; ++dim)
+    {
+        std::vector<ck::index_t> reduce_dims{dim, this->Rank - 1};
+        this->Run(reduce_dims);
+    }
+}
+
+TYPED_TEST(TestSoftmax, ReduceMultipleMiddleDims)
+{
+    std::vector<ck::index_t> reduce_dims{0, 1};
+    if(this->Rank >= 3)
+    {
+        this->Run(reduce_dims);
+    }
+
+    if(this->Rank >= 4)
+    {
+        reduce_dims = std::vector<ck::index_t>{0, 2};
+        this->Run(reduce_dims);
+        reduce_dims = std::vector<ck::index_t>{0, 1, 2};
+        this->Run(reduce_dims);
+    }
+}
+
+TYPED_TEST(TestSoftmax, ReduceAllDims)
+{
+    std::vector<ck::index_t> reduce_dims(this->Rank);
+    std::iota(std::begin(reduce_dims), std::end(reduce_dims), 0);
+    this->Run(reduce_dims);
+}
+
+TYPED_TEST(TestSoftmax, ReduceOddLengths)
+{
+    this->in_lengths_ = {{3, 63, 1032}};
+    if(this->Rank >= 4)
+    {
+        this->in_lengths_ = {{1, 3, 63, 1032}};
+    }
+    this->Run({this->Rank - 1});
+    this->Run({this->Rank - 2});
+}
diff --git a/3rdparty/composable_kernel/test/softmax/test_softmax_util.hpp b/3rdparty/composable_kernel/test/softmax/test_softmax_util.hpp
new file mode 100644
index 0000000..40b300c
--- /dev/null
+++ b/3rdparty/composable_kernel/test/softmax/test_softmax_util.hpp
@@ -0,0 +1,142 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <string>
+#include <sstream>
+#include <tuple>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "include/ck/utility/data_type.hpp"
+#include "profiler/profile_softmax_impl.hpp"
+
+namespace ck {
+
+template <typename Range>
+std::string serialize_range(const Range& range)
+{
+    std::stringstream ss;
+    for(auto& r : range)
+    {
+        ss << r << ", ";
+    }
+    std::string str = ss.str();
+    return std::string(str.begin(), str.end() - 2);
+}
+
+template <typename Tuple>
+class TestSoftmax : public ::testing::Test
+{
+    protected:
+    using InDataType              = std::tuple_element_t<0, Tuple>;
+    using AccDataType             = std::tuple_element_t<1, Tuple>;
+    using OutDataType             = std::tuple_element_t<2, Tuple>;
+    static constexpr index_t Rank = std::tuple_element_t<3, Tuple>{}.value;
+
+    public:
+    std::vector<std::vector<index_t>> in_lengths_ = {{2, 128, 1024}, {4, 16, 8448}, {128, 128, 64}};
+    std::vector<std::vector<AccDataType>> scales_ = {{2, 0}, {0, 2}, {2, 2}};
+    bool bench_                                   = false; // measure kernel performance
+    bool verify_                                  = true;
+
+    void SetUp() override
+    {
+        if constexpr(Rank == 4)
+        {
+            in_lengths_ = std::vector<std::vector<index_t>>{
+                {1, 2, 128, 1024}, {2, 4, 16, 8448}, {1, 128, 128, 64}};
+        }
+    }
+
+    void RunSingle(std::vector<index_t> in_length,
+                   std::vector<index_t> reduce_dims,
+                   AccDataType alpha,
+                   AccDataType beta)
+    {
+        int init_method = 1; // integer value initialization
+        bool log        = false;
+        std::vector<ck::index_t> strides; // intenionally empty, to get packed layout.
+        bool pass = ck::profiler::profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank>(
+            verify_, init_method, log, bench_, in_length, strides, reduce_dims, alpha, beta);
+        EXPECT_TRUE(pass);
+    }
+
+    void Run(std::vector<index_t> reduce_dims = {})
+    {
+        if(reduce_dims.empty())
+        {
+            reduce_dims.push_back(Rank - 1);
+        }
+
+        for(auto in_length : this->in_lengths_)
+        {
+            for(auto scale : this->scales_)
+            {
+                this->RunSingle(in_length, reduce_dims, scale[0], scale[1]);
+            }
+        }
+    }
+};
+
+template <index_t Rank,
+          index_t NumReduceDim,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize>
+struct DeviceSoftmaxInstanceWrapper
+{
+    using F16  = half_t;
+    using F32  = float;
+    using Pass = tensor_operation::element_wise::PassThrough;
+
+    using InDataType   = F16;
+    using AccDataType  = F32;
+    using OutDataType  = F16;
+    using InElementOp  = Pass;
+    using AccElementOp = Pass;
+
+    using DeviceSoftmaxInstance = tensor_operation::device::DeviceSoftmaxImpl<InDataType,
+                                                                              AccDataType,
+                                                                              OutDataType,
+                                                                              InElementOp,
+                                                                              AccElementOp,
+                                                                              Rank,
+                                                                              NumReduceDim,
+                                                                              BlockSize,
+                                                                              MThreadClusterSize,
+                                                                              KThreadClusterSize,
+                                                                              MThreadSliceSize,
+                                                                              KThreadSliceSize,
+                                                                              InSrcVectorDim,
+                                                                              InSrcVectorSize,
+                                                                              OutDstVectorSize>;
+
+    bool IsSupported(const std::vector<index_t> in_lengths,
+                     const std::vector<index_t> in_strides,
+                     const std::vector<index_t> reduce_dims) const
+    {
+        auto softmax  = DeviceSoftmaxInstance{};
+        auto argument = softmax.MakeArgument(in_lengths,
+                                             in_strides,
+                                             reduce_dims,
+                                             1,       // alpha
+                                             1,       // beta
+                                             nullptr, // in_dev
+                                             nullptr, // in_out
+                                             Pass{},  // in elementwise op
+                                             Pass{}); // acc elementwise op
+        return softmax.IsSupportedArgument(argument);
+    }
+};
+
+} // namespace ck
diff --git a/3rdparty/composable_kernel/test/space_filling_curve/CMakeLists.txt b/3rdparty/composable_kernel/test/space_filling_curve/CMakeLists.txt
new file mode 100644
index 0000000..a527268
--- /dev/null
+++ b/3rdparty/composable_kernel/test/space_filling_curve/CMakeLists.txt
@@ -0,0 +1 @@
+add_test_executable(test_space_filling_curve space_filling_curve.cpp)
diff --git a/3rdparty/composable_kernel/test/space_filling_curve/space_filling_curve.cpp b/3rdparty/composable_kernel/test/space_filling_curve/space_filling_curve.cpp
new file mode 100644
index 0000000..c7f6759
--- /dev/null
+++ b/3rdparty/composable_kernel/test/space_filling_curve/space_filling_curve.cpp
@@ -0,0 +1,195 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+#include <iostream>
+#include <numeric>
+#include <cassert>
+
+#include "ck/ck.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+using namespace ck;
+
+void traverse_using_space_filling_curve_linear();
+void traverse_using_space_filling_curve_snakecurved();
+
+int main(int argc, char** argv)
+{
+    (void)argc;
+    (void)argv;
+
+    traverse_using_space_filling_curve_linear();
+    traverse_using_space_filling_curve_snakecurved();
+
+    return 0;
+}
+
+void traverse_using_space_filling_curve_linear()
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+
+    using TensorLengths    = Sequence<3, 2, 2>;
+    using DimAccessOrder   = Sequence<2, 0, 1>;
+    using ScalarsPerAccess = Sequence<1, 1, 1>;
+    using SpaceFillingCurve =
+        SpaceFillingCurve<TensorLengths, DimAccessOrder, ScalarsPerAccess, false>;
+
+    constexpr auto expected = make_tuple(make_tuple(0, 0, 0),
+                                         make_tuple(0, 1, 0),
+                                         make_tuple(1, 0, 0),
+                                         make_tuple(1, 1, 0),
+                                         make_tuple(2, 0, 0),
+                                         make_tuple(2, 1, 0),
+                                         make_tuple(0, 0, 1),
+                                         make_tuple(0, 1, 1),
+                                         make_tuple(1, 0, 1),
+                                         make_tuple(1, 1, 1),
+                                         make_tuple(2, 0, 1),
+                                         make_tuple(2, 1, 1));
+
+    constexpr index_t num_access = SpaceFillingCurve::GetNumOfAccess();
+
+    static_assert(num_access == reduce_on_sequence(TensorLengths{} / ScalarsPerAccess{},
+                                                   math::multiplies{},
+                                                   Number<1>{}));
+
+    static_for<1, num_access, 1>{}([&](auto i) {
+        constexpr auto idx_curr = SpaceFillingCurve::GetIndex(i);
+
+        static_assert(idx_curr[I0] == expected[i][I0]);
+        static_assert(idx_curr[I1] == expected[i][I1]);
+        static_assert(idx_curr[I2] == expected[i][I2]);
+
+        constexpr auto backward_step = SpaceFillingCurve::GetBackwardStep(i);
+        constexpr auto expected_step = expected[i - I1] - expected[i];
+        static_assert(backward_step[I0] == expected_step[I0]);
+        static_assert(backward_step[I1] == expected_step[I1]);
+        static_assert(backward_step[I2] == expected_step[I2]);
+    });
+
+    static_for<0, num_access - 1, 1>{}([&](auto i) {
+        constexpr auto idx_curr = SpaceFillingCurve::GetIndex(i);
+
+        static_assert(idx_curr[I0] == expected[i][I0]);
+        static_assert(idx_curr[I1] == expected[i][I1]);
+        static_assert(idx_curr[I2] == expected[i][I2]);
+
+        constexpr auto forward_step  = SpaceFillingCurve::GetForwardStep(i);
+        constexpr auto expected_step = expected[i + I1] - expected[i];
+        static_assert(forward_step[I0] == expected_step[I0]);
+        static_assert(forward_step[I1] == expected_step[I1]);
+        static_assert(forward_step[I2] == expected_step[I2]);
+    });
+}
+
+void traverse_using_space_filling_curve_snakecurved()
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+
+    using TensorLengths    = Sequence<16, 10, 9>;
+    using DimAccessOrder   = Sequence<2, 0, 1>;
+    using ScalarsPerAccess = Sequence<4, 2, 3>;
+    using SpaceFillingCurve =
+        SpaceFillingCurve<TensorLengths, DimAccessOrder, ScalarsPerAccess, true>;
+
+    constexpr auto expected = make_tuple(make_tuple(0, 0, 0),
+                                         make_tuple(0, 2, 0),
+                                         make_tuple(0, 4, 0),
+                                         make_tuple(0, 6, 0),
+                                         make_tuple(0, 8, 0),
+                                         make_tuple(4, 8, 0),
+                                         make_tuple(4, 6, 0),
+                                         make_tuple(4, 4, 0),
+                                         make_tuple(4, 2, 0),
+                                         make_tuple(4, 0, 0),
+                                         make_tuple(8, 0, 0),
+                                         make_tuple(8, 2, 0),
+                                         make_tuple(8, 4, 0),
+                                         make_tuple(8, 6, 0),
+                                         make_tuple(8, 8, 0),
+                                         make_tuple(12, 8, 0),
+                                         make_tuple(12, 6, 0),
+                                         make_tuple(12, 4, 0),
+                                         make_tuple(12, 2, 0),
+                                         make_tuple(12, 0, 0),
+                                         make_tuple(12, 0, 3),
+                                         make_tuple(12, 2, 3),
+                                         make_tuple(12, 4, 3),
+                                         make_tuple(12, 6, 3),
+                                         make_tuple(12, 8, 3),
+                                         make_tuple(8, 8, 3),
+                                         make_tuple(8, 6, 3),
+                                         make_tuple(8, 4, 3),
+                                         make_tuple(8, 2, 3),
+                                         make_tuple(8, 0, 3),
+                                         make_tuple(4, 0, 3),
+                                         make_tuple(4, 2, 3),
+                                         make_tuple(4, 4, 3),
+                                         make_tuple(4, 6, 3),
+                                         make_tuple(4, 8, 3),
+                                         make_tuple(0, 8, 3),
+                                         make_tuple(0, 6, 3),
+                                         make_tuple(0, 4, 3),
+                                         make_tuple(0, 2, 3),
+                                         make_tuple(0, 0, 3),
+                                         make_tuple(0, 0, 6),
+                                         make_tuple(0, 2, 6),
+                                         make_tuple(0, 4, 6),
+                                         make_tuple(0, 6, 6),
+                                         make_tuple(0, 8, 6),
+                                         make_tuple(4, 8, 6),
+                                         make_tuple(4, 6, 6),
+                                         make_tuple(4, 4, 6),
+                                         make_tuple(4, 2, 6),
+                                         make_tuple(4, 0, 6),
+                                         make_tuple(8, 0, 6),
+                                         make_tuple(8, 2, 6),
+                                         make_tuple(8, 4, 6),
+                                         make_tuple(8, 6, 6),
+                                         make_tuple(8, 8, 6),
+                                         make_tuple(12, 8, 6),
+                                         make_tuple(12, 6, 6),
+                                         make_tuple(12, 4, 6),
+                                         make_tuple(12, 2, 6),
+                                         make_tuple(12, 0, 6));
+
+    constexpr index_t num_access = SpaceFillingCurve::GetNumOfAccess();
+
+    static_assert(num_access == reduce_on_sequence(TensorLengths{} / ScalarsPerAccess{},
+                                                   math::multiplies{},
+                                                   Number<1>{}));
+
+    static_for<1, num_access, 1>{}([&](auto i) {
+        constexpr auto idx_curr = SpaceFillingCurve::GetIndex(i);
+
+        static_assert(idx_curr[I0] == expected[i][I0]);
+        static_assert(idx_curr[I1] == expected[i][I1]);
+        static_assert(idx_curr[I2] == expected[i][I2]);
+
+        constexpr auto backward_step = SpaceFillingCurve::GetBackwardStep(i);
+        constexpr auto expected_step = expected[i - I1] - expected[i];
+        static_assert(backward_step[I0] == expected_step[I0]);
+        static_assert(backward_step[I1] == expected_step[I1]);
+        static_assert(backward_step[I2] == expected_step[I2]);
+    });
+
+    static_for<0, num_access - 1, 1>{}([&](auto i) {
+        constexpr auto idx_curr = SpaceFillingCurve::GetIndex(i);
+
+        static_assert(idx_curr[I0] == expected[i][I0]);
+        static_assert(idx_curr[I1] == expected[i][I1]);
+        static_assert(idx_curr[I2] == expected[i][I2]);
+
+        constexpr auto forward_step  = SpaceFillingCurve::GetForwardStep(i);
+        constexpr auto expected_step = expected[i + I1] - expected[i];
+        static_assert(forward_step[I0] == expected_step[I0]);
+        static_assert(forward_step[I1] == expected_step[I1]);
+        static_assert(forward_step[I2] == expected_step[I2]);
+    });
+}
diff --git a/3rdparty/composable_kernel/test/wmma_op/CMakeLists.txt b/3rdparty/composable_kernel/test/wmma_op/CMakeLists.txt
new file mode 100644
index 0000000..e553253
--- /dev/null
+++ b/3rdparty/composable_kernel/test/wmma_op/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_test_executable(test_wmma_op wmma_op.cpp)
+target_link_libraries(test_wmma_op PRIVATE utility)
diff --git a/3rdparty/composable_kernel/test/wmma_op/wmma_op.cpp b/3rdparty/composable_kernel/test/wmma_op/wmma_op.cpp
new file mode 100644
index 0000000..761c15f
--- /dev/null
+++ b/3rdparty/composable_kernel/test/wmma_op/wmma_op.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "test/wmma_op/wmma_op_util.hpp"
+
+template <typename SrcType,
+          typename DstType,
+          typename GPUAccType,
+          typename CPUAccType,
+          ck::index_t AccNum>
+bool run_test()
+{
+    using Row         = ck::tensor_layout::gemm::RowMajor;
+    using Col         = ck::tensor_layout::gemm::ColumnMajor;
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    bool pass         = true;
+
+    const auto matmul_default = ck::wmma_op_util::matmul<SrcType, DstType, GPUAccType, AccNum>;
+    const auto matmul_swizzle_a =
+        ck::wmma_op_util::matmul_swizzle_a<SrcType, DstType, GPUAccType, AccNum>;
+
+    const auto wmma_kernel_container = std::make_tuple(matmul_default, matmul_swizzle_a);
+
+    ck::static_for<0, 2, 1>{}([&](auto i) {
+        pass &=
+            ck::wmma_op_util::TestWmma<decltype(std::get<ck::Number<i>{}>(wmma_kernel_container)),
+                                       SrcType,
+                                       SrcType,
+                                       DstType,
+                                       GPUAccType,
+                                       CPUAccType,
+                                       decltype(Row{}),
+                                       decltype(Col{}),
+                                       decltype(Row{}),
+                                       PassThrough,
+                                       PassThrough,
+                                       PassThrough,
+                                       AccNum>{}(std::get<ck::Number<i>{}>(wmma_kernel_container));
+    });
+
+    return pass ? 1 : 0;
+}
+int main(int, char*[])
+{
+    bool pass = true;
+    // clang-format off
+    //              |SrcType     |DstType     |GPUAccType  |CPUAccType |AccNum
+    pass &= run_test<ck::half_t,  ck::half_t,  float,       float,      8     >();
+    pass &= run_test<ck::bhalf_t, ck::bhalf_t, float,       float,      8     >();
+    pass &= run_test<ck::half_t,  ck::half_t,  ck::half_t,  ck::half_t, 16    >();
+    pass &= run_test<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t, float,      16    >();
+    pass &= run_test<int8_t,      int8_t,      int32_t,     int32_t,    8     >();
+    // clang-format on
+
+    std::cout << "TestGemm ..... " << (pass ? "SUCCESS" : "FAILURE") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/3rdparty/composable_kernel/test/wmma_op/wmma_op_util.hpp b/3rdparty/composable_kernel/test/wmma_op/wmma_op_util.hpp
new file mode 100644
index 0000000..ef3f831
--- /dev/null
+++ b/3rdparty/composable_kernel/test/wmma_op/wmma_op_util.hpp
@@ -0,0 +1,369 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/utility/amd_wmma.hpp"
+
+namespace ck {
+namespace wmma_op_util {
+
+template <typename src_vec, typename acc_vec>
+__device__ void builtin_wmma_naive_selector(const src_vec&, const src_vec&, acc_vec&)
+{
+}
+
+template <>
+__device__ void
+builtin_wmma_naive_selector<half16_t,
+                            StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, float, 1, 8, true>>(
+    const half16_t& reg_a,
+    const half16_t& reg_b,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, float, 1, 8, true>& reg_c)
+{
+    intrin_wmma_f32_16x16x16_f16_w32<16, 16>::Run(
+        reg_a, reg_b, reg_c.GetVectorTypeReference(Number<0>{}));
+}
+
+template <>
+__device__ void
+builtin_wmma_naive_selector<bhalf16_t,
+                            StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, float, 1, 8, true>>(
+    const bhalf16_t& reg_a,
+    const bhalf16_t& reg_b,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, float, 1, 8, true>& reg_c)
+{
+    intrin_wmma_f32_16x16x16_bf16_w32<16, 16>::Run(
+        reg_a, reg_b, reg_c.GetVectorTypeReference(Number<0>{}));
+}
+
+template <>
+__device__ void
+builtin_wmma_naive_selector<half16_t,
+                            StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, half_t, 1, 16, true>>(
+    const half16_t& reg_a,
+    const half16_t& reg_b,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, half_t, 1, 16, true>& reg_c)
+{
+    intrin_wmma_f16_16x16x16_f16_w32<16, 16, 0>::Run(
+        reg_a, reg_b, reg_c.GetVectorTypeReference(Number<0>{}));
+}
+
+template <>
+__device__ void builtin_wmma_naive_selector<
+    bhalf16_t,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, bhalf_t, 1, 16, true>>(
+    const bhalf16_t& reg_a,
+    const bhalf16_t& reg_b,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, bhalf_t, 1, 16, true>& reg_c)
+{
+    intrin_wmma_bf16_16x16x16_bf16_w32<16, 16, 0>::Run(
+        reg_a, reg_b, reg_c.GetVectorTypeReference(Number<0>{}));
+}
+
+template <>
+__device__ void
+builtin_wmma_naive_selector<int8x16_t,
+                            StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, int32_t, 1, 8, true>>(
+    const int8x16_t& reg_a,
+    const int8x16_t& reg_b,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, int32_t, 1, 8, true>& reg_c)
+{
+    intrin_wmma_i32_16x16x16_iu8_w32<16, 16, true, true, false>::Run(
+        reg_a, reg_b, reg_c.GetVectorTypeReference(Number<0>{}));
+}
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+template <>
+__device__ void
+builtin_wmma_naive_selector<int4x16_t,
+                            StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, int32_t, 1, 8, true>>(
+    const int4x16_t& reg_a,
+    const int4x16_t& reg_b,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, int32_t, 1, 8, true>& reg_c)
+{
+    intrin_wmma_i32_16x16x16_iu4_w32<16, 16, true, true, false>::Run(
+        reg_a, reg_b, reg_c.GetVectorTypeReference(Number<0>{}));
+}
+#endif
+
+template <typename src_t, typename dst_t, typename acc_t, index_t acc_num>
+__global__ void matmul(const src_t* a, const src_t* b, dst_t* c)
+{
+    const int lIdx = threadIdx.x;
+    // a and b fragments are stored in 8 VGPRs each, in packed format, so 16 elements each for a and
+    // b a_frag will store one column of the 16x16 matrix tile b_frag will store one row of the
+    // 16x16 matrix tile
+    using src_vec  = typename vector_type<src_t, 16>::type;
+    src_vec a_frag = {};
+    src_vec b_frag = {};
+    // initialize c fragment to 0
+    using acc_vec = StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, acc_t, 1, acc_num, true>;
+    acc_vec c_thread_buf_;
+
+    // lane is (0-31) mod 16 instead of 0-31 due to matrix replication in gfx11
+    // see https://atlvsp3.amd.com/sp3_gfx11_5_instructions.pdf page 482
+    // TODO: remove this dependency in gfx12 https://ontrack-internal.amd.com/browse/DEGFXSP3-101
+    const int lane = lIdx % 16;
+
+    for(int ele = 0; ele < 16; ++ele)
+    {
+        b_frag[ele] = b[16 * lane + ele];
+    }
+    // follow origin design
+    for(int ele = 0; ele < 16; ++ele)
+    {
+        a_frag[ele] = a[16 * lane + ele];
+    }
+
+    // sync threads, similar to mma_sync
+    __syncthreads();
+    builtin_wmma_naive_selector<src_vec, acc_vec>(a_frag, b_frag, c_thread_buf_);
+    __syncthreads();
+    // wait for results, similar to mma_sync
+    static_for<0, 8, 1>{}([&](auto ele) {
+        const int r = ele * 2 + (lIdx / 16);
+        // store results from unpacked c_thread_buf_ output
+        c[16 * r + lane] = ck::type_convert<dst_t>(c_thread_buf_[Number<ele * acc_num / 8>{}]);
+    });
+}
+
+template <typename src_t, typename dst_t, typename acc_t, index_t acc_num>
+__global__ void matmul_swizzle_a(const src_t* a, const src_t* b, dst_t* c)
+{
+    const int lIdx = threadIdx.x;
+
+    using src_vec  = typename vector_type<src_t, 16>::type;
+    src_vec a_frag = {};
+    src_vec b_frag = {};
+    using acc_vec  = StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, acc_t, 1, acc_num, true>;
+    acc_vec c_thread_buf_;
+
+    const int lane = lIdx % 16;
+
+    for(int ele = 0; ele < 16; ++ele)
+    {
+        b_frag[ele] = b[16 * lane + ele];
+    }
+
+    const int offset_m = (((lane & 1) << 3) | (lane >> 1));
+    for(int ele = 0; ele < 16; ++ele)
+    {
+        a_frag[ele] = a[16 * offset_m + ele];
+    }
+
+    __syncthreads();
+    builtin_wmma_naive_selector<src_vec, acc_vec>(a_frag, b_frag, c_thread_buf_);
+    __syncthreads();
+
+    static_for<0, 8, 1>{}([&](auto ele) {
+        const int blk = lIdx / 16;
+        const int r   = ele;
+        c[16 * 8 * blk + 16 * r + lane] =
+            ck::type_convert<dst_t>(c_thread_buf_[Number<ele * acc_num / 8>{}]);
+    });
+}
+
+struct GemmParams
+{
+    GemmParams() : M(16), N(16), K(16), StrideA(16), StrideB(16), StrideC(16), alpha(1), beta(0) {}
+
+    ck::index_t M;
+    ck::index_t N;
+    ck::index_t K;
+
+    ck::index_t StrideA;
+    ck::index_t StrideB;
+    ck::index_t StrideC;
+
+    float alpha;
+    float beta;
+};
+
+template <typename GemmInstance,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+void RunHostGEMM(const Tensor<ADataType>& A,
+                 const Tensor<BDataType>& B,
+                 Tensor<CDataType>& C,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+{
+    auto ref_gemm     = GemmInstance{};
+    auto ref_invoker  = ref_gemm.MakeInvoker();
+    auto ref_argument = ref_gemm.MakeArgument(A, B, C, a_element_op, b_element_op, c_element_op);
+
+    ref_invoker.Run(ref_argument);
+}
+
+template <typename KernelType, typename ADataType, typename BDataType, typename CDataType>
+bool RunDeviceGEMM(KernelType kernel,
+                   const Tensor<ADataType>& A,
+                   const Tensor<BDataType>& B,
+                   Tensor<CDataType>& C)
+{
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpaceSize());
+    DeviceMem b_n_k_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpaceSize());
+
+    a_m_k_device_buf.ToDevice(A.mData.data());
+    b_n_k_device_buf.ToDevice(B.mData.data());
+    kernel<<<1, 32>>>(static_cast<const ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                      static_cast<const BDataType*>(b_n_k_device_buf.GetDeviceBuffer()),
+                      static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()));
+    c_m_n_device_buf.FromDevice(C.mData.data());
+
+    return true;
+}
+
+template <typename DeviceWmma,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename GPUAccDataType,
+          typename CPUAccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t CAccNum>
+struct TestWmma
+{
+    auto PrepareGemmTensor(const ck::wmma_op_util::GemmParams& params)
+    {
+        auto f_host_tensor_descriptor =
+            [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+                if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+                {
+                    return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                                std::vector<std::size_t>({stride, 1}));
+                }
+                else
+                {
+                    return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                                std::vector<std::size_t>({1, stride}));
+                }
+            };
+
+        Tensor<ADataType> a_m_k(
+            f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
+        Tensor<BDataType> b_n_k(
+            f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
+        Tensor<CDataType> c_m_n_host_result(
+            f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+        Tensor<CDataType> c_m_n_device_result(
+            f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+
+        auto f_generate_tensor_value = [](auto& tensor, auto type) {
+            using dataType = decltype(type);
+
+            tensor.GenerateTensorValue(GeneratorTensor_2<dataType>{-5, 5});
+        };
+
+        f_generate_tensor_value(a_m_k, ADataType{});
+        f_generate_tensor_value(b_n_k, BDataType{});
+
+        return std::make_tuple(a_m_k, b_n_k, c_m_n_host_result, c_m_n_device_result);
+    }
+
+    auto operator()(const DeviceWmma& wmma_kernel)
+    {
+        std::cout << "ALayout = " << ALayout{}.name << ", BLayout = " << BLayout{}.name
+                  << ", CLayout = " << CLayout{}.name << std::endl;
+
+        // Arrange
+        ck::wmma_op_util::GemmParams params;
+        params.M       = 16;
+        params.N       = 16;
+        params.K       = 16;
+        params.StrideA = 16;
+        params.StrideB = 16;
+        params.StrideC = 16;
+
+        auto host_tensors = PrepareGemmTensor(params);
+
+        const Tensor<ADataType>& a  = std::get<0>(host_tensors);
+        const Tensor<BDataType>& b  = std::get<1>(host_tensors);
+        Tensor<CDataType>& c_host   = std::get<2>(host_tensors);
+        Tensor<CDataType>& c_device = std::get<3>(host_tensors);
+
+        auto a_element_op = AElementwiseOperation{};
+        auto b_element_op = BElementwiseOperation{};
+        auto c_element_op = CElementwiseOperation{};
+
+        using ReferenceGemmInstance =
+            ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                      BDataType,
+                                                      CDataType,
+                                                      CPUAccDataType,
+                                                      AElementwiseOperation,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation>;
+        ck::wmma_op_util::RunHostGEMM<ReferenceGemmInstance>(
+            a, b, c_host, a_element_op, b_element_op, c_element_op);
+
+        // Act
+        bool is_supported = ck::wmma_op_util::RunDeviceGEMM(wmma_kernel, a, b, c_device);
+
+        if(is_supported)
+        {
+            // Assert
+            bool res = false;
+            if(std::is_same<CDataType, float>::value)
+            {
+                res = ck::utils::check_err(c_device.mData, c_host.mData);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, ck::half_t>::value)
+            {
+                res = ck::utils::check_err(c_device.mData, c_host.mData);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, ck::bhalf_t>::value)
+            {
+                // 0.5 Pixel Error Tolerance is introduced by Accumulator difference.
+                // BF16 WMMA Accumulator is in BF16 Type while On Host-side Accumulator is Float.
+                res = ck::utils::check_err(
+                    c_device.mData, c_host.mData, "Error: Incorrect results!", 0, 1.0);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, int8_t>::value)
+            {
+                res = ck::utils::check_err(c_device.mData, c_host.mData);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, double>::value)
+            {
+                res = ck::utils::check_err(c_device.mData, c_host.mData);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else
+            {
+                std::cout << "UNSUPPORTED CDataType" << std::endl;
+            }
+
+            return res;
+        }
+        else
+        {
+            return true;
+        }
+    }
+};
+
+} // namespace wmma_op_util
+} // namespace ck
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9ceb30a..a536a24 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -69,8 +69,12 @@ if(GIT_FOUND AND EXISTS "${PROJECT_SOURCE_DIR}/.git")
   endif()
 endif()
 
-set(CUTLASS_HEADER_DIR ${PROJECT_SOURCE_DIR}/3rdparty/cutlass/include)
-set(CUTLASS_EXTENSIONS_DIR ${PROJECT_SOURCE_DIR}/src/fastertransformer/cutlass_extensions/include)
+set(CK_HEADER_DIR ${PROJECT_SOURCE_DIR}/3rdparty/composable_kernel/include)
+set(CK_LIBRARY_HEADER_DIR ${PROJECT_SOURCE_DIR}/3rdparty/composable_kernel/library/include/)
+set(CK_EXTENSIONS_DIR ${PROJECT_SOURCE_DIR}/src/fastertransformer/ck_extensions/include)
+
+# set(CUTLASS_HEADER_DIR ${PROJECT_SOURCE_DIR}/3rdparty/cutlass/include)
+# set(CUTLASS_EXTENSIONS_DIR ${PROJECT_SOURCE_DIR}/src/fastertransformer/cutlass_extensions/include)
 
 option(SPARSITY_SUPPORT "Build project with Ampere sparsity feature support" OFF)
 
@@ -121,6 +125,8 @@ if(USE_NVTX)
   add_definitions("-DUSE_NVTX")
 endif()
 
+add_definitions("-D__HIP_PLATFORM_AMD__")
+
 # setting compiler flags
 set(CMAKE_C_FLAGS    "${CMAKE_C_FLAGS}")
 set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} -std=c++17")
@@ -210,8 +216,11 @@ set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 set(COMMON_HEADER_DIRS
   ${PROJECT_SOURCE_DIR}
   ${CUDA_PATH}/include
-  ${CUTLASS_HEADER_DIR}
-  ${CUTLASS_EXTENSIONS_DIR}
+  ${CK_HEADER_DIR}
+  ${CK_EXTENSIONS_DIR}
+  ${CK_LIBRARY_HEADER_DIR}
+  # ${CUTLASS_HEADER_DIR}
+  # ${CUTLASS_EXTENSIONS_DIR}
   ${PROJECT_SOURCE_DIR}/3rdparty/trt_fp8_fmha/src
   ${PROJECT_SOURCE_DIR}/3rdparty/trt_fp8_fmha/generated
 )
@@ -391,10 +400,10 @@ add_library(transformer-shared SHARED
   # $<TARGET_OBJECTS:cutlass_preprocessors>
   $<TARGET_OBJECTS:decoder_masked_multihead_attention>
   $<TARGET_OBJECTS:decoding_kernels>
-  # $<TARGET_OBJECTS:fpA_intB_gemm>
+  $<TARGET_OBJECTS:fpA_intB_gemm>
   $<TARGET_OBJECTS:gen_relative_pos_bias>
   $<TARGET_OBJECTS:gpt_kernels>
-  # $<TARGET_OBJECTS:int8_gemm>
+  $<TARGET_OBJECTS:int8_gemm>
   $<TARGET_OBJECTS:layernorm_int8_kernels>
   $<TARGET_OBJECTS:layernorm_kernels>
   $<TARGET_OBJECTS:layout_transformer_int8_kernels>
@@ -404,7 +413,7 @@ add_library(transformer-shared SHARED
   $<TARGET_OBJECTS:matrix_transpose_kernels>
   $<TARGET_OBJECTS:matrix_vector_multiplication>
   $<TARGET_OBJECTS:memory_utils>
-  # $<TARGET_OBJECTS:moe_gemm_kernels>
+  $<TARGET_OBJECTS:moe_gemm_kernels>
   # $<TARGET_OBJECTS:moe_kernels>
   $<TARGET_OBJECTS:mpi_utils>
   $<TARGET_OBJECTS:nccl_utils>
diff --git a/src/fastertransformer/ck_extensions/include/ck_extensions/ck_utils.h b/src/fastertransformer/ck_extensions/include/ck_extensions/ck_utils.h
new file mode 100644
index 0000000..e55ac63
--- /dev/null
+++ b/src/fastertransformer/ck_extensions/include/ck_extensions/ck_utils.h
@@ -0,0 +1,37 @@
+#pragma once
+
+#include "ck/utility/sequence.hpp"
+#include "hip/hip_runtime.h"
+
+namespace fastertransformer {
+
+static const char* _hipGetErrorEnum(hipError_t error)
+{
+    return hipGetErrorString(error);
+}
+
+template<typename T>
+void check(T result, char const* const func, const char* const file, int const line)
+{
+    if (result) {
+        throw std::runtime_error(std::string("[FT][ERROR] CUDA runtime error: ") + (_hipGetErrorEnum(result)) + " "
+                                 + file + ":" + std::to_string(line) + " \n");
+    }
+}
+
+#define check_hip_error(val) check((val), #val, __FILE__, __LINE__)
+#define check_hip_error_2(val, file, line) check((val), #val, file, line)
+
+inline int getSMVersion()
+{
+    int device{-1};
+    check_hip_error(hipGetDevice(&device));
+    int sm_major = 0;
+    int sm_minor = 0;
+    check_hip_error(hipDeviceGetAttribute(&sm_major, hipDeviceAttributeComputeCapabilityMinor, device));
+    check_hip_error(hipDeviceGetAttribute(&sm_minor, hipDeviceAttributeComputeCapabilityMinor, device));
+    return sm_major * 10 + sm_minor;
+}
+
+
+}
diff --git a/src/fastertransformer/ck_extensions/include/ck_extensions/epilogue/epilogue_quant_helper.h b/src/fastertransformer/ck_extensions/include/ck_extensions/epilogue/epilogue_quant_helper.h
new file mode 100644
index 0000000..3697b67
--- /dev/null
+++ b/src/fastertransformer/ck_extensions/include/ck_extensions/epilogue/epilogue_quant_helper.h
@@ -0,0 +1,48 @@
+/***************************************************************************************************
+ * Copyright (c) 2017 - 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice, this
+ * list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ * this list of conditions and the following disclaimer in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ * 3. Neither the name of the copyright holder nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+ * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+ * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+ * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+
+#pragma once
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+
+namespace cutlass {
+namespace epilogue {
+
+// define scaling mode
+enum class QuantMode {
+    PerTensorQuant,
+    PerTokenQuant,
+    PerChannelQuant,
+    PerTokenChannelQuant
+};
+
+}  // namespace epilogue
+}  // namespace cutlass
diff --git a/src/fastertransformer/ck_extensions/include/ck_extensions/ft_gemm_configs.h b/src/fastertransformer/ck_extensions/include/ck_extensions/ft_gemm_configs.h
new file mode 100644
index 0000000..fbc01b0
--- /dev/null
+++ b/src/fastertransformer/ck_extensions/include/ck_extensions/ft_gemm_configs.h
@@ -0,0 +1,58 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+namespace fastertransformer {
+// Note: The shapes are in the format MxNxK. The K shape of the runtime config MUST match the K shape
+//       in the kernel layout details when doing weight only quantization.
+enum class CutlassTileConfig {
+    // Signals that we should run heuristics do choose a config
+    Undefined,
+
+    // Signals that we should run heuristics do choose a config
+    ChooseWithHeuristic,
+
+    // SiMT config
+    CtaShape128x128x8_WarpShape64x64x8,
+
+    // TensorCore configs CTA_N = 128, CTA_K = 64
+    // Warp configs for M=32
+    CtaShape32x128x64_WarpShape32x32x64,
+
+    // Warp configs for M=64
+    CtaShape64x128x64_WarpShape32x64x64,
+    CtaShape64x128x64_WarpShape64x32x64,
+
+    // Warp configs for M=128
+    CtaShape128x128x64_WarpShape64x32x64,
+    CtaShape128x128x64_WarpShape128x32x64
+};
+
+enum class SplitKStyle {
+    NO_SPLIT_K,
+    SPLIT_K_SERIAL,
+    // SPLIT_K_PARALLEL // Not supported yet
+};
+
+struct CutlassGemmConfig {
+    CutlassTileConfig tile_config    = CutlassTileConfig::ChooseWithHeuristic;
+    SplitKStyle       split_k_style  = SplitKStyle::NO_SPLIT_K;
+    int               split_k_factor = -1;
+    int               stages         = -1;
+};
+
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/ck_extensions/include/ck_extensions/gemm/device/DeviceGemmDequant.hpp b/src/fastertransformer/ck_extensions/include/ck_extensions/gemm/device/DeviceGemmDequant.hpp
new file mode 100644
index 0000000..b2fd59e
--- /dev/null
+++ b/src/fastertransformer/ck_extensions/include/ck_extensions/gemm/device/DeviceGemmDequant.hpp
@@ -0,0 +1,137 @@
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+
+namespace ck::tensor_operation::device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct DeviceGemmDequant {
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // desc for problem definition
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+
+};
+
+
+} // namespace
\ No newline at end of file
diff --git a/src/fastertransformer/ck_extensions/include/ck_extensions/gemm/kernel/default_fpA_intB_traits.h b/src/fastertransformer/ck_extensions/include/ck_extensions/gemm/kernel/default_fpA_intB_traits.h
new file mode 100644
index 0000000..12539aa
--- /dev/null
+++ b/src/fastertransformer/ck_extensions/include/ck_extensions/gemm/kernel/default_fpA_intB_traits.h
@@ -0,0 +1,125 @@
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+// #include "ck/arch/mma.h"
+// #include "ck/bfloat16.h"
+// #include "ck/cutlass.h"
+// #include "ck/gemm/gemm.h"
+// #include "ck/layout/matrix.h"
+
+// #include "cutlass_extensions/arch/mma.h"
+// #include "cutlass_extensions/gemm/kernel/mixed_gemm_B_layout.h"
+
+// namespace cutlass {
+// namespace gemm {
+// namespace kernel {
+
+// template<typename TypeA, typename TypeB, typename arch, typename Enable = void>
+// struct MixedGemmArchTraits {
+// };
+
+// template<typename arch>
+// struct MixedGemmArchTraits<float, float, arch> {
+//     static constexpr int Stages = 2;
+//     using OperatorClass         = cutlass::arch::OpClassSimt;
+//     using AccType               = float;
+//     using LayoutB               = cutlass::layout::RowMajor;
+
+//     static constexpr int ElementsPerAccessA = 1;
+//     static constexpr int ElementsPerAccessB = 1;
+//     static constexpr int ElementsPerAccessC = 1;
+//     static constexpr int ThreadblockK       = 8;
+//     using InstructionShape                  = cutlass::gemm::GemmShape<1, 1, 1>;
+
+//     using Operator = cutlass::arch::OpMultiplyAdd;
+// };
+
+// // ========================= Volta Traits ===========================
+// // Volta will always dequantize after the global memory load.
+// // This will instantiate any HMMA tensorcore kernels for Volta.
+// // Note that volta does not have native bfloat support so weights and activations will be casted to fp16
+// // and compute will happen in fp16 then will be converted for bf16 output.
+// template<typename TypeA, typename TypeB>
+// struct MixedGemmArchTraits<
+//     TypeA,
+//     TypeB,
+//     cutlass::arch::Sm70,
+//     typename cutlass::platform::enable_if<cutlass::platform::is_same<TypeA, cutlass::half_t>::value
+//                                           || cutlass::platform::is_same<TypeA, cutlass::bfloat16_t>::value>::type> {
+// private:
+//     using LayoutDetails = LayoutDetailsB<TypeB, cutlass::arch::Sm70>;
+
+// public:
+//     static constexpr int ThreadblockK = LayoutDetails::ThreadblockK;
+
+//     using OperatorClass = cutlass::arch::OpClassTensorOp;
+//     using AccType       = float;
+//     using LayoutB       = typename LayoutDetails::Layout;
+
+//     static constexpr int ElementsPerAccessA = 128 / cutlass::sizeof_bits<TypeA>::value;
+//     static constexpr int ElementsPerAccessB = LayoutDetails::ElementsPerAccess;
+//     static constexpr int ElementsPerAccessC = 128 / cutlass::sizeof_bits<TypeA>::value;
+//     using InstructionShape                  = cutlass::gemm::GemmShape<8, 8, 4>;
+
+//     using Operator = typename LayoutDetails::Operator;
+// };
+
+// // ======================= Turing Traits ==============================
+// // Note that turing does not have native bfloat support so weights and activations will be casted to fp16
+// // and compute will happen in fp16 then will be converted for bf16 output.
+// template<typename TypeA, typename TypeB>
+// struct MixedGemmArchTraits<
+//     TypeA,
+//     TypeB,
+//     cutlass::arch::Sm75,
+//     typename cutlass::platform::enable_if<cutlass::platform::is_same<TypeA, cutlass::half_t>::value
+//                                           || cutlass::platform::is_same<TypeA, cutlass::bfloat16_t>::value>::type> {
+// private:
+//     using LayoutDetails = LayoutDetailsB<TypeB, cutlass::arch::Sm75>;
+
+// public:
+//     static constexpr int ThreadblockK = LayoutDetails::ThreadblockK;
+
+//     using OperatorClass = cutlass::arch::OpClassTensorOp;
+//     using AccType       = float;
+//     using LayoutB       = typename LayoutDetails::Layout;
+
+//     static constexpr int ElementsPerAccessA = 128 / cutlass::sizeof_bits<TypeA>::value;
+//     static constexpr int ElementsPerAccessB = LayoutDetails::ElementsPerAccess;
+//     static constexpr int ElementsPerAccessC = 128 / cutlass::sizeof_bits<TypeA>::value;
+//     using InstructionShape                  = cutlass::gemm::GemmShape<16, 8, 8>;
+
+//     using Operator = typename LayoutDetails::Operator;
+// };
+
+// // ======================= Ampere Traits ==============================
+// template<typename TypeA, typename TypeB>
+// struct MixedGemmArchTraits<
+//     TypeA,
+//     TypeB,
+//     cutlass::arch::Sm80,
+//     typename cutlass::platform::enable_if<cutlass::platform::is_same<TypeA, cutlass::half_t>::value
+//                                           || cutlass::platform::is_same<TypeA, cutlass::bfloat16_t>::value>::type> {
+// private:
+//     using LayoutDetails = LayoutDetailsB<TypeB, cutlass::arch::Sm80>;
+
+// public:
+//     static constexpr int ThreadblockK = LayoutDetails::ThreadblockK;
+
+//     using OperatorClass = cutlass::arch::OpClassTensorOp;
+//     using AccType       = float;
+//     using LayoutB       = typename LayoutDetails::Layout;
+
+//     static constexpr int ElementsPerAccessA = 128 / cutlass::sizeof_bits<TypeA>::value;
+//     static constexpr int ElementsPerAccessB = LayoutDetails::ElementsPerAccess;
+//     static constexpr int ElementsPerAccessC = 128 / cutlass::sizeof_bits<TypeA>::value;
+//     using InstructionShape                  = cutlass::gemm::GemmShape<16, 8, 16>;
+
+//     using Operator = typename LayoutDetails::Operator;
+// };
+
+// }  // namespace kernel
+// }  // namespace gemm
+// }  // namespace cutlass
\ No newline at end of file
diff --git a/src/fastertransformer/ck_extensions/include/ck_extensions/interleaved_numeric_conversion.h b/src/fastertransformer/ck_extensions/include/ck_extensions/interleaved_numeric_conversion.h
new file mode 100644
index 0000000..a71f982
--- /dev/null
+++ b/src/fastertransformer/ck_extensions/include/ck_extensions/interleaved_numeric_conversion.h
@@ -0,0 +1,120 @@
+
+
+
+
+
+
+#pragma once
+#include <ck/ck.hpp>
+
+namespace ck {
+
+template<typename T, typename S, int N>
+struct FastInterleavedAndBiasedNumericArrayConverter {
+};
+
+
+template<>
+struct FastInterleavedAndBiasedNumericArrayConverter<half_t, uint8_t, 4> {
+    using result_type = Array<half_t, 4>;
+    using source_type = Array<uint8_t, 4>;
+
+
+    CUTLASS_DEVICE
+    static result_type convert(source_type const& source)
+    {
+        result_type result;
+        uint32_t*      h   = reinterpret_cast<uint32_t*>(&result);
+        uint32_t const i8s = reinterpret_cast<uint32_t const&>(source);
+
+        static constexpr uint32_t mask_for_elt_01     = 0x05020500;
+        static constexpr uint32_t mask_for_elt_23     = 0x05030501;
+        static constexpr uint32_t start_byte_for_fp16 = 0x64646464;
+
+        asm volatile("v_perm_b32 %0,%1,%2,%3;\n" : "=v"(h[0]) : "v"(start_byte_for_fp16), "v"(i8s), "v"(mask_for_elt_01));
+        asm volatile("v_perm_b32 %0,%1,%2,%3;\n" : "=v"(h[1]) : "v"(start_byte_for_fp16), "v"(i8s), "v"(mask_for_elt_23));
+
+        static constexpr uint32_t I8s_TO_F16s_MAGIC_NUM = 0x6480;
+        asm volatile("v_sub_f16x2 %0, %1, %2;\n" : "=v"(h[0]) : "v"(h[0]), "v"(I8s_TO_F16s_MAGIC_NUM));
+        asm volatile("v_sub_f16x2 %0, %1, %2;\n" : "=v"(h[1]) : "v"(h[1]), "v"(I8s_TO_F16s_MAGIC_NUM));
+
+        return result;
+    }
+
+    CUTLASS_DEVICE
+    result_type operator()(source_type const& s)
+    {
+        return convert(s);
+    }
+};
+
+template<int N>
+struct FastInterleavedAndBiasedNumericArrayConverter<half_t, uint8_t, N> {
+    static constexpr int VEC_WIDTH = 4;
+    static_assert(!(N % VEC_WIDTH), "N must be multiple of 4.");
+
+    using result_type = Array<half_t, N>;
+    using source_type = Array<uint8_t, N>;
+
+    CUTLASS_DEVICE
+    static result_type convert(source_type const& source)
+    {
+        using scalar_result_type = typename result_type::Element;
+        using scalar_source_type = typename source_type::Element;
+        FastInterleavedAndBiasedNumericArrayConverter<scalar_result_type, scalar_source_type, VEC_WIDTH>
+            convert_vector_;
+
+        result_type result;
+        using vec_result = Array<scalar_result_type, VEC_WIDTH>;
+        using vec_source = Array<scalar_source_type, VEC_WIDTH>;
+
+        vec_result*       result_ptr = reinterpret_cast<vec_result*>(&result);
+        vec_source const* source_ptr = reinterpret_cast<vec_source const*>(&source);
+
+        CUTLASS_PRAGMA_UNROLL
+        for (int i = 0; i < N / VEC_WIDTH; ++i) {
+            result_ptr[i] = convert_vector_(source_ptr[i]);
+        }
+
+        return result;
+    }
+
+    CUTLASS_DEVICE
+    result_type operator()(source_type const& s)
+    {
+        return convert(s);
+    }
+};
+
+} // namespace ck
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
diff --git a/src/fastertransformer/kernels/CMakeLists.txt b/src/fastertransformer/kernels/CMakeLists.txt
index 79bc570..43659aa 100644
--- a/src/fastertransformer/kernels/CMakeLists.txt
+++ b/src/fastertransformer/kernels/CMakeLists.txt
@@ -14,7 +14,8 @@
 
 cmake_minimum_required(VERSION 3.8)
 
-add_subdirectory(cutlass_kernels)
+# add_subdirectory(cutlass_kernels)
+add_subdirectory(ck_kernels)
 
 add_library(image_shift_partition_kernels image_shift_partition_kernels.cu)
 set_property(TARGET image_shift_partition_kernels PROPERTY POSITION_INDEPENDENT_CODE  ON)
diff --git a/src/fastertransformer/kernels/activation_fp8_kernels.cu b/src/fastertransformer/kernels/activation_fp8_kernels.cu
index c44581b..965b109 100644
--- a/src/fastertransformer/kernels/activation_fp8_kernels.cu
+++ b/src/fastertransformer/kernels/activation_fp8_kernels.cu
@@ -18,6 +18,7 @@
 #include "src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h"
 #include "src/fastertransformer/utils/cuda_bf16_fallbacks.cuh"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 
diff --git a/src/fastertransformer/kernels/activation_kernels.cu b/src/fastertransformer/kernels/activation_kernels.cu
index 3532b5f..7bf3733 100644
--- a/src/fastertransformer/kernels/activation_kernels.cu
+++ b/src/fastertransformer/kernels/activation_kernels.cu
@@ -17,6 +17,7 @@
 #include "src/fastertransformer/kernels/activation_kernels.h"
 #include "src/fastertransformer/utils/cuda_type_utils.cuh"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include "src/fastertransformer/utils/memory_utils.h"
 
 #ifndef CUDART_VERSION
diff --git a/src/fastertransformer/kernels/add_residual_kernels.h b/src/fastertransformer/kernels/add_residual_kernels.h
index 34951d8..dfd351c 100644
--- a/src/fastertransformer/kernels/add_residual_kernels.h
+++ b/src/fastertransformer/kernels/add_residual_kernels.h
@@ -17,6 +17,7 @@
 #pragma once
 
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <assert.h>
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
diff --git a/src/fastertransformer/kernels/ban_bad_words.cu b/src/fastertransformer/kernels/ban_bad_words.cu
index e5fb77f..408aa80 100644
--- a/src/fastertransformer/kernels/ban_bad_words.cu
+++ b/src/fastertransformer/kernels/ban_bad_words.cu
@@ -16,6 +16,7 @@
 
 #include "src/fastertransformer/kernels/ban_bad_words.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 
diff --git a/src/fastertransformer/kernels/beam_search_penalty_kernels.h b/src/fastertransformer/kernels/beam_search_penalty_kernels.h
index ae67f96..80b1dda 100644
--- a/src/fastertransformer/kernels/beam_search_penalty_kernels.h
+++ b/src/fastertransformer/kernels/beam_search_penalty_kernels.h
@@ -19,6 +19,7 @@
 
 #include "src/fastertransformer/kernels/penalty_types.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 
diff --git a/src/fastertransformer/kernels/beam_search_topk_kernels.cu b/src/fastertransformer/kernels/beam_search_topk_kernels.cu
index fcaf644..ed54a19 100644
--- a/src/fastertransformer/kernels/beam_search_topk_kernels.cu
+++ b/src/fastertransformer/kernels/beam_search_topk_kernels.cu
@@ -26,6 +26,7 @@
 #include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
 #include "src/fastertransformer/utils/cuda_type_utils.cuh"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include "src/fastertransformer/utils/logger.h"
 
 namespace fastertransformer {
diff --git a/src/fastertransformer/kernels/bert_preprocess_kernels.h b/src/fastertransformer/kernels/bert_preprocess_kernels.h
index 2de4865..696d783 100644
--- a/src/fastertransformer/kernels/bert_preprocess_kernels.h
+++ b/src/fastertransformer/kernels/bert_preprocess_kernels.h
@@ -17,6 +17,7 @@
 #pragma once
 #include "src/fastertransformer/kernels/gen_relative_pos_bias.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
 #ifdef ENABLE_FP8
diff --git a/src/fastertransformer/kernels/ck_kernels/CMakeLists.txt b/src/fastertransformer/kernels/ck_kernels/CMakeLists.txt
new file mode 100644
index 0000000..6071ebd
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/CMakeLists.txt
@@ -0,0 +1,47 @@
+# 测试用
+# cmake_minimum_required(VERSION 3.16)
+# project(ck)
+# include_directories(/code/llama_ft/)
+# include_directories(/code/llama_ft/src/fastertransformer/ck_extensions/include)
+# include_directories(/code/llama_ft/3rdparty/composable_kernel/include/)
+# include_directories(/code/llama_ft/3rdparty/composable_kernel/library/include/)
+# set(CMAKE_SYSTEM_NAME Linux)
+# set(TOOLCHAIN_PATH ${ROCM_PATH}/hip/cmake)
+# set(CMAKE_C_COMPILER ${ROCM_PATH}/bin/hipcc)
+# list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
+
+
+enable_language(CXX)
+find_package(HIP REQUIRED)
+
+add_definitions("-D__HIP_PLATFORM_AMD__")
+set(CMAKE_PREFIX_PATH "/opt/dtk-23.04" CACHE PATH "Path to which RoCm has been installed")
+set(CMAKE_CXX_COMPILER ${CMAKE_PREFIX_PATH}/bin/hipcc)
+set(CMAKE_CXX_FLAGS "-O3")
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_BUILD_TYPE Release)
+set(GPU_TARGETS "gfx906;gfx926")
+
+
+# add_library(cutlass_preprocessors STATIC cutlass_preprocessors.cc)
+# set_property(TARGET cutlass_preprocessors PROPERTY POSITION_INDEPENDENT_CODE ON)
+
+set(moe_gemm_files "")
+file(GLOB moe_gemm_files ${moe_gemm_files} ./moe_gemm/*.cpp)
+add_library(moe_gemm_kernels STATIC ${moe_gemm_files})
+set_property(TARGET moe_gemm_kernels PROPERTY POSITION_INDEPENDENT_CODE ON)
+set_property(TARGET moe_gemm_kernels PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
+
+
+set(fpA_intB_files "")
+file(GLOB fpA_intB_files ${fpA_intB_files} ./fpA_intB_gemm/*.cpp)
+add_library(fpA_intB_gemm STATIC ${fpA_intB_files})
+set_property(TARGET fpA_intB_gemm PROPERTY POSITION_INDEPENDENT_CODE ON)
+set_property(TARGET fpA_intB_gemm PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
+
+
+set(int8_files "")
+file(GLOB int8_files ${int8_files} ./int8_gemm/*.cpp)
+add_library(int8_gemm STATIC ${int8_files} )
+set_property(TARGET int8_gemm PROPERTY POSITION_INDEPENDENT_CODE ON)
+set_property(TARGET int8_gemm PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/cmake/Analyzers.cmake b/src/fastertransformer/kernels/ck_kernels/cmake/Analyzers.cmake
new file mode 100644
index 0000000..1bf1a52
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/cmake/Analyzers.cmake
@@ -0,0 +1,34 @@
+################################################################################
+# 
+# MIT License
+# 
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+# 
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+# 
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+# 
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+# 
+################################################################################
+
+if(NOT TARGET analyze)
+    add_custom_target(analyze)
+endif()
+
+function(mark_as_analyzer)
+    add_dependencies(analyze ${ARGN})
+endfunction()
+
diff --git a/src/fastertransformer/kernels/ck_kernels/cmake/ClangTidy.cmake b/src/fastertransformer/kernels/ck_kernels/cmake/ClangTidy.cmake
new file mode 100644
index 0000000..654d4b6
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/cmake/ClangTidy.cmake
@@ -0,0 +1,162 @@
+################################################################################
+#
+# MIT License
+#
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+#
+################################################################################
+include(CMakeParseArguments)
+include(Analyzers)
+
+get_filename_component(CLANG_TIDY_EXE_HINT "${CMAKE_CXX_COMPILER}" PATH)
+set(ROCM_PATH $ENV{ROCM_PATH})
+find_program(CLANG_TIDY_EXE
+    NAMES
+        clang-tidy
+        clang-tidy-5.0
+        clang-tidy-4.0
+        clang-tidy-3.9
+        clang-tidy-3.8
+        clang-tidy-3.7
+        clang-tidy-3.6
+        clang-tidy-3.5
+    HINTS
+        ${CLANG_TIDY_EXE_HINT}
+    PATH_SUFFIXES
+        compiler/bin
+    PATHS
+        ${ROCM_PATH}/llvm/bin
+        ${ROCM_PATH}/hcc
+        /usr/local/opt/llvm/bin
+)
+
+function(find_clang_tidy_version VAR)
+    execute_process(COMMAND ${CLANG_TIDY_EXE} -version OUTPUT_VARIABLE VERSION_OUTPUT)
+    separate_arguments(VERSION_OUTPUT_LIST UNIX_COMMAND "${VERSION_OUTPUT}")
+    list(FIND VERSION_OUTPUT_LIST "version" VERSION_INDEX)
+    if(VERSION_INDEX GREATER 0)
+        math(EXPR VERSION_INDEX "${VERSION_INDEX} + 1")
+        list(GET VERSION_OUTPUT_LIST ${VERSION_INDEX} VERSION)
+        set(${VAR} ${VERSION} PARENT_SCOPE)
+    else()
+        set(${VAR} "0.0" PARENT_SCOPE)
+    endif()
+
+endfunction()
+
+if( NOT CLANG_TIDY_EXE )
+    message( STATUS "Clang tidy not found" )
+    set(CLANG_TIDY_VERSION "0.0")
+else()
+    find_clang_tidy_version(CLANG_TIDY_VERSION)
+    message( STATUS "Clang tidy found: ${CLANG_TIDY_VERSION}")
+endif()
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+set(CLANG_TIDY_FIXIT_DIR ${CMAKE_BINARY_DIR}/fixits)
+file(MAKE_DIRECTORY ${CLANG_TIDY_FIXIT_DIR})
+set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES ${CLANG_TIDY_FIXIT_DIR})
+
+macro(enable_clang_tidy)
+    set(options ANALYZE_TEMPORARY_DTORS ALL)
+    set(oneValueArgs HEADER_FILTER)
+    set(multiValueArgs CHECKS ERRORS EXTRA_ARGS)
+
+    cmake_parse_arguments(PARSE "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    string(REPLACE ";" "," CLANG_TIDY_CHECKS "${PARSE_CHECKS}")
+    string(REPLACE ";" "," CLANG_TIDY_ERRORS "${PARSE_ERRORS}")
+    set(CLANG_TIDY_EXTRA_ARGS)
+    foreach(ARG ${PARSE_EXTRA_ARGS})
+        list(APPEND CLANG_TIDY_EXTRA_ARGS "-extra-arg=${ARG}")
+    endforeach()
+
+    set(CLANG_TIDY_ALL)
+    if(PARSE_ALL)
+        set(CLANG_TIDY_ALL ALL)
+    endif()
+
+    message(STATUS "Clang tidy checks: ${CLANG_TIDY_CHECKS}")
+
+    if (${PARSE_ANALYZE_TEMPORARY_DTORS})
+        set(CLANG_TIDY_ANALYZE_TEMPORARY_DTORS "-analyze-temporary-dtors")
+    endif()
+
+    if (${CLANG_TIDY_VERSION} VERSION_LESS "3.9.0")
+        set(CLANG_TIDY_ERRORS_ARG "")
+    else()
+        set(CLANG_TIDY_ERRORS_ARG "-warnings-as-errors='${CLANG_TIDY_ERRORS}'")
+    endif()
+
+    if (${CLANG_TIDY_VERSION} VERSION_LESS "3.9.0")
+        set(CLANG_TIDY_QUIET_ARG "")
+    else()
+        set(CLANG_TIDY_QUIET_ARG "-quiet")
+    endif()
+
+    if(PARSE_HEADER_FILTER)
+        string(REPLACE "$" "$$" CLANG_TIDY_HEADER_FILTER "${PARSE_HEADER_FILTER}")
+    else()
+        set(CLANG_TIDY_HEADER_FILTER ".*")
+    endif()
+
+    set(CLANG_TIDY_COMMAND
+        ${CLANG_TIDY_EXE}
+        ${CLANG_TIDY_QUIET_ARG}
+        -p ${CMAKE_BINARY_DIR}
+        -checks='${CLANG_TIDY_CHECKS}'
+        ${CLANG_TIDY_ERRORS_ARG}
+        ${CLANG_TIDY_EXTRA_ARGS}
+        ${CLANG_TIDY_ANALYZE_TEMPORARY_DTORS}
+        -header-filter='${CLANG_TIDY_HEADER_FILTER}'
+    )
+    add_custom_target(tidy ${CLANG_TIDY_ALL})
+    mark_as_analyzer(tidy)
+    add_custom_target(tidy-base)
+    add_custom_target(tidy-make-fixit-dir COMMAND ${CMAKE_COMMAND} -E make_directory ${CLANG_TIDY_FIXIT_DIR})
+    add_custom_target(tidy-rm-fixit-dir COMMAND ${CMAKE_COMMAND} -E remove_directory ${CLANG_TIDY_FIXIT_DIR})
+    add_dependencies(tidy-make-fixit-dir tidy-rm-fixit-dir)
+    add_dependencies(tidy-base tidy-make-fixit-dir)
+endmacro()
+
+function(clang_tidy_check TARGET)
+    get_target_property(SOURCES ${TARGET} SOURCES)
+    # TODO: Use generator expressions instead
+    # COMMAND ${CLANG_TIDY_COMMAND} $<TARGET_PROPERTY:${TARGET},SOURCES>
+    # COMMAND ${CLANG_TIDY_COMMAND} $<JOIN:$<TARGET_PROPERTY:${TARGET},SOURCES>, >
+    foreach(SOURCE ${SOURCES})
+        if((NOT "${SOURCE}" MATCHES "(h|hpp|hxx)$") AND (NOT "${SOURCE}" MATCHES "TARGET_OBJECTS"))
+            string(MAKE_C_IDENTIFIER "${SOURCE}" tidy_file)
+            set(tidy_target tidy-target-${TARGET}-${tidy_file})
+            add_custom_target(${tidy_target}
+                # for some targets clang-tidy not able to get information from .clang-tidy
+                DEPENDS ${SOURCE}
+                COMMAND ${CLANG_TIDY_COMMAND} "-config=\{CheckOptions: \[\{key: bugprone-reserved-identifier.AllowedIdentifiers,value: __HIP_PLATFORM_HCC__\; __HIP_ROCclr__\}\]\}" ${SOURCE} "-export-fixes=${CLANG_TIDY_FIXIT_DIR}/${TARGET}-${tidy_file}.yaml"
+                WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+                COMMENT "clang-tidy: Running clang-tidy on target ${SOURCE}..."
+            )
+            add_dependencies(${tidy_target} ${TARGET})
+            add_dependencies(${tidy_target} tidy-base)
+            add_dependencies(tidy ${tidy_target})
+        endif()
+    endforeach()
+endfunction()
+
diff --git a/src/fastertransformer/kernels/ck_kernels/cmake/CppCheck.cmake b/src/fastertransformer/kernels/ck_kernels/cmake/CppCheck.cmake
new file mode 100644
index 0000000..63e1441
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/cmake/CppCheck.cmake
@@ -0,0 +1,130 @@
+################################################################################
+# 
+# MIT License
+# 
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+# 
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+# 
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+# 
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+# 
+################################################################################
+
+include(CMakeParseArguments)
+include(ProcessorCount)
+include(Analyzers)
+
+find_program(CPPCHECK_EXE 
+    NAMES 
+        cppcheck
+    PATHS
+        $ENV{ROCM_PATH}/bin
+)
+
+ProcessorCount(CPPCHECK_JOBS)
+
+set(CPPCHECK_BUILD_DIR ${CMAKE_BINARY_DIR}/cppcheck-build)
+file(MAKE_DIRECTORY ${CPPCHECK_BUILD_DIR})
+set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES ${CPPCHECK_BUILD_DIR})
+
+macro(enable_cppcheck)
+    set(options FORCE)
+    set(oneValueArgs)
+    set(multiValueArgs CHECKS SUPPRESS DEFINE UNDEFINE INCLUDE SOURCES)
+
+    cmake_parse_arguments(PARSE "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    string(REPLACE ";" "," CPPCHECK_CHECKS "${PARSE_CHECKS}")
+    string(REPLACE ";" "\n" CPPCHECK_SUPPRESS "${PARSE_SUPPRESS};*:/usr/*")
+    file(WRITE ${CMAKE_BINARY_DIR}/cppcheck-supressions "${CPPCHECK_SUPPRESS}")
+    set(CPPCHECK_DEFINES)
+    foreach(DEF ${PARSE_DEFINE})
+        set(CPPCHECK_DEFINES "${CPPCHECK_DEFINES} -D${DEF}")
+    endforeach()
+
+    set(CPPCHECK_UNDEFINES)
+    foreach(DEF ${PARSE_UNDEFINE})
+        set(CPPCHECK_UNDEFINES "${CPPCHECK_UNDEFINES} -U${DEF}")
+    endforeach()
+
+    set(CPPCHECK_INCLUDES)
+    foreach(INC ${PARSE_INCLUDE})
+        set(CPPCHECK_INCLUDES "${CPPCHECK_INCLUDES} -I${INC}")
+    endforeach()
+
+    # set(CPPCHECK_FORCE)
+    set(CPPCHECK_FORCE "--project=${CMAKE_BINARY_DIR}/compile_commands.json")
+    if(PARSE_FORCE)
+        set(CPPCHECK_FORCE --force)
+    endif()
+
+    set(SOURCES)
+    set(GLOBS)
+    foreach(SOURCE ${PARSE_SOURCES})
+        get_filename_component(ABS_SOURCE ${SOURCE} ABSOLUTE)
+        if(EXISTS ${ABS_SOURCE})
+            if(IS_DIRECTORY ${ABS_SOURCE})
+                set(GLOBS "${GLOBS} ${ABS_SOURCE}/*.cpp ${ABS_SOURCE}/*.hpp ${ABS_SOURCE}/*.cxx ${ABS_SOURCE}/*.c ${ABS_SOURCE}/*.h")
+            else()
+                set(SOURCES "${SOURCES} ${ABS_SOURCE}")
+            endif()
+        else()
+            set(GLOBS "${GLOBS} ${ABS_SOURCE}")
+        endif()
+    endforeach()
+
+    file(WRITE ${CMAKE_BINARY_DIR}/cppcheck.cmake "
+        file(GLOB_RECURSE GSRCS ${GLOBS})
+        set(CPPCHECK_COMMAND
+            ${CPPCHECK_EXE}
+            -q
+            # -v
+            # --report-progress
+            ${CPPCHECK_FORCE}
+            --cppcheck-build-dir=${CPPCHECK_BUILD_DIR}
+            --platform=native
+            --template=gcc
+            --error-exitcode=1
+            -j ${CPPCHECK_JOBS}
+            ${CPPCHECK_DEFINES}
+            ${CPPCHECK_UNDEFINES}
+            ${CPPCHECK_INCLUDES}
+            --enable=${CPPCHECK_CHECKS}
+            --inline-suppr
+            --suppressions-list=${CMAKE_BINARY_DIR}/cppcheck-supressions
+            ${SOURCES} \${GSRCS}
+        )
+        string(REPLACE \";\" \" \" CPPCHECK_SHOW_COMMAND \"\${CPPCHECK_COMMAND}\")
+        message(\"\${CPPCHECK_SHOW_COMMAND}\")
+        execute_process(
+            COMMAND \${CPPCHECK_COMMAND}
+            WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+            RESULT_VARIABLE RESULT
+        )
+        if(NOT RESULT EQUAL 0)
+            message(FATAL_ERROR \"Cppcheck failed\")
+        endif()
+")
+
+    add_custom_target(cppcheck
+        COMMAND ${CMAKE_COMMAND} -P ${CMAKE_BINARY_DIR}/cppcheck.cmake
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        COMMENT "cppcheck: Running cppcheck..."
+    )
+    mark_as_analyzer(cppcheck)
+endmacro()
+
+
diff --git a/src/fastertransformer/kernels/ck_kernels/cmake/DoxygenDoc.cmake b/src/fastertransformer/kernels/ck_kernels/cmake/DoxygenDoc.cmake
new file mode 100644
index 0000000..2e3669f
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/cmake/DoxygenDoc.cmake
@@ -0,0 +1,355 @@
+################################################################################
+# 
+# MIT License
+# 
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+# 
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+# 
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+# 
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+# 
+################################################################################
+include(CMakeParseArguments)
+include(MainDoc)
+
+find_program(DOXYGEN_EXECUTABLE NAMES doxygen
+    PATH_SUFFIXES bin
+    DOC "Doxygen documentation generator"
+)
+mark_as_advanced(DOXYGEN_EXECUTABLE)
+
+find_path(DOT_EXECUTABLE NAMES dot
+    PATH_SUFFIXES bin
+    DOC "Graphviz"
+)
+mark_as_advanced(DOT_EXECUTABLE)
+
+set(DOXYGEN_ARGS
+ABBREVIATE_BRIEF
+ALIASES
+ALLEXTERNALS
+ALLOW_UNICODE_NAMES
+ALPHABETICAL_INDEX
+ALWAYS_DETAILED_SEC
+AUTOLINK_SUPPORT
+BINARY_TOC
+BRIEF_MEMBER_DESC
+BUILTIN_STL_SUPPORT
+CALLER_GRAPH
+CALL_GRAPH
+CASE_SENSE_NAMES
+CHM_FILE
+CHM_INDEX_ENCODING
+CITE_BIB_FILES
+CLANG_ASSISTED_PARSING
+CLANG_OPTIONS
+CLASS_DIAGRAMS
+CLASS_GRAPH
+COLLABORATION_GRAPH
+COLS_IN_ALPHA_INDEX
+COMPACT_LATEX
+COMPACT_RTF
+CPP_CLI_SUPPORT
+CREATE_SUBDIRS
+DIAFILE_DIRS
+DIA_PATH
+DIRECTORY_GRAPH
+DISABLE_INDEX
+DISTRIBUTE_GROUP_DOC
+DOCBOOK_OUTPUT
+DOCBOOK_PROGRAMLISTING
+DOCSET_BUNDLE_ID
+DOCSET_FEEDNAME
+DOCSET_PUBLISHER_ID
+DOCSET_PUBLISHER_NAME
+DOTFILE_DIRS
+DOT_CLEANUP
+DOT_FONTNAME
+DOT_FONTPATH
+DOT_FONTSIZE
+DOT_GRAPH_MAX_NODES
+DOT_IMAGE_FORMAT
+DOT_MULTI_TARGETS
+DOT_NUM_THREADS
+# DOT_PATH
+DOT_TRANSPARENT
+DOXYFILE_ENCODING
+ECLIPSE_DOC_ID
+ENABLED_SECTIONS
+ENABLE_PREPROCESSING
+ENUM_VALUES_PER_LINE
+EXAMPLE_PATH
+EXAMPLE_PATTERNS
+EXAMPLE_RECURSIVE
+EXCLUDE
+EXCLUDE_PATTERNS
+EXCLUDE_SYMBOLS
+EXCLUDE_SYMLINKS
+EXPAND_AS_DEFINED
+EXPAND_ONLY_PREDEF
+EXTENSION_MAPPING
+EXTERNAL_GROUPS
+EXTERNAL_PAGES
+EXTERNAL_SEARCH
+EXTERNAL_SEARCH_ID
+EXTRACT_ALL
+EXTRACT_ANON_NSPACES
+EXTRACT_LOCAL_CLASSES
+EXTRACT_LOCAL_METHODS
+EXTRACT_PACKAGE
+EXTRACT_PRIVATE
+EXTRACT_STATIC
+EXTRA_PACKAGES
+EXTRA_SEARCH_MAPPINGS
+EXT_LINKS_IN_WINDOW
+FILE_PATTERNS
+FILE_VERSION_FILTER
+FILTER_PATTERNS
+FILTER_SOURCE_FILES
+FILTER_SOURCE_PATTERNS
+FORCE_LOCAL_INCLUDES
+FORMULA_FONTSIZE
+FORMULA_TRANSPARENT
+FULL_PATH_NAMES
+GENERATE_AUTOGEN_DEF
+GENERATE_BUGLIST
+GENERATE_CHI
+GENERATE_DEPRECATEDLIST
+GENERATE_DOCBOOK
+GENERATE_DOCSET
+GENERATE_ECLIPSEHELP
+GENERATE_HTML
+GENERATE_HTMLHELP
+GENERATE_LATEX
+GENERATE_LEGEND
+GENERATE_MAN
+GENERATE_PERLMOD
+GENERATE_QHP
+GENERATE_RTF
+GENERATE_TAGFILE
+GENERATE_TESTLIST
+GENERATE_TODOLIST
+GENERATE_TREEVIEW
+GENERATE_XML
+GRAPHICAL_HIERARCHY
+GROUP_GRAPHS
+GROUP_NESTED_COMPOUNDS
+# HAVE_DOT
+HHC_LOCATION
+HIDE_COMPOUND_REFERENCE
+HIDE_FRIEND_COMPOUNDS
+HIDE_IN_BODY_DOCS
+HIDE_SCOPE_NAMES
+HIDE_UNDOC_CLASSES
+HIDE_UNDOC_MEMBERS
+HIDE_UNDOC_RELATIONS
+HTML_COLORSTYLE_GAMMA
+HTML_COLORSTYLE_HUE
+HTML_COLORSTYLE_SAT
+HTML_DYNAMIC_SECTIONS
+HTML_EXTRA_FILES
+HTML_EXTRA_STYLESHEET
+HTML_FILE_EXTENSION
+HTML_FOOTER
+HTML_HEADER
+HTML_INDEX_NUM_ENTRIES
+HTML_OUTPUT
+HTML_STYLESHEET
+HTML_TIMESTAMP
+IDL_PROPERTY_SUPPORT
+IGNORE_PREFIX
+IMAGE_PATH
+INCLUDED_BY_GRAPH
+INCLUDE_FILE_PATTERNS
+INCLUDE_GRAPH
+INCLUDE_PATH
+INHERIT_DOCS
+INLINE_GROUPED_CLASSES
+INLINE_INFO
+INLINE_INHERITED_MEMB
+INLINE_SIMPLE_STRUCTS
+INLINE_SOURCES
+INPUT
+INPUT_ENCODING
+INPUT_FILTER
+INTERACTIVE_SVG
+INTERNAL_DOCS
+JAVADOC_AUTOBRIEF
+LATEX_BATCHMODE
+LATEX_BIB_STYLE
+LATEX_CMD_NAME
+LATEX_EXTRA_FILES
+LATEX_EXTRA_STYLESHEET
+LATEX_FOOTER
+LATEX_HEADER
+LATEX_HIDE_INDICES
+LATEX_OUTPUT
+LATEX_SOURCE_CODE
+LATEX_TIMESTAMP
+LAYOUT_FILE
+LOOKUP_CACHE_SIZE
+MACRO_EXPANSION
+MAKEINDEX_CMD_NAME
+MAN_EXTENSION
+MAN_LINKS
+MAN_OUTPUT
+MAN_SUBDIR
+MARKDOWN_SUPPORT
+MATHJAX_CODEFILE
+MATHJAX_EXTENSIONS
+MATHJAX_FORMAT
+MATHJAX_RELPATH
+MAX_DOT_GRAPH_DEPTH
+MAX_INITIALIZER_LINES
+MSCFILE_DIRS
+MSCGEN_PATH
+MULTILINE_CPP_IS_BRIEF
+OPTIMIZE_FOR_FORTRAN
+OPTIMIZE_OUTPUT_FOR_C
+OPTIMIZE_OUTPUT_JAVA
+OPTIMIZE_OUTPUT_VHDL
+OUTPUT_DIRECTORY
+OUTPUT_LANGUAGE
+PAPER_TYPE
+PDF_HYPERLINKS
+PERLMOD_LATEX
+PERLMOD_MAKEVAR_PREFIX
+PERLMOD_PRETTY
+PERL_PATH
+PLANTUML_CFG_FILE
+PLANTUML_INCLUDE_PATH
+PLANTUML_JAR_PATH
+PREDEFINED
+PROJECT_BRIEF
+PROJECT_LOGO
+PROJECT_NAME
+PROJECT_NUMBER
+QCH_FILE
+QHG_LOCATION
+QHP_CUST_FILTER_ATTRS
+QHP_CUST_FILTER_NAME
+QHP_NAMESPACE
+QHP_SECT_FILTER_ATTRS
+QHP_VIRTUAL_FOLDER
+QT_AUTOBRIEF
+QUIET
+RECURSIVE
+REFERENCED_BY_RELATION
+REFERENCES_LINK_SOURCE
+REFERENCES_RELATION
+REPEAT_BRIEF
+RTF_EXTENSIONS_FILE
+RTF_HYPERLINKS
+RTF_OUTPUT
+RTF_SOURCE_CODE
+RTF_STYLESHEET_FILE
+SEARCHDATA_FILE
+SEARCHENGINE
+SEARCHENGINE_URL
+SEARCH_INCLUDES
+SEPARATE_MEMBER_PAGES
+SERVER_BASED_SEARCH
+SHORT_NAMES
+SHOW_FILES
+SHOW_GROUPED_MEMB_INC
+SHOW_INCLUDE_FILES
+SHOW_NAMESPACES
+SHOW_USED_FILES
+SIP_SUPPORT
+SKIP_FUNCTION_MACROS
+SORT_BRIEF_DOCS
+SORT_BY_SCOPE_NAME
+SORT_GROUP_NAMES
+SORT_MEMBERS_CTORS_1ST
+SORT_MEMBER_DOCS
+SOURCE_BROWSER
+SOURCE_TOOLTIPS
+STRICT_PROTO_MATCHING
+STRIP_CODE_COMMENTS
+STRIP_FROM_INC_PATH
+STRIP_FROM_PATH
+SUBGROUPING
+TAB_SIZE
+TAGFILES
+TCL_SUBST
+TEMPLATE_RELATIONS
+TOC_EXPAND
+TOC_INCLUDE_HEADINGS
+TREEVIEW_WIDTH
+TYPEDEF_HIDES_STRUCT
+UML_LIMIT_NUM_FIELDS
+UML_LOOK
+USE_HTAGS
+USE_MATHJAX
+USE_MDFILE_AS_MAINPAGE
+USE_PDFLATEX
+VERBATIM_HEADERS
+WARNINGS
+WARN_AS_ERROR
+WARN_FORMAT
+WARN_IF_DOC_ERROR
+WARN_IF_UNDOCUMENTED
+WARN_LOGFILE
+WARN_NO_PARAMDOC
+XML_OUTPUT
+XML_PROGRAMLISTING
+)
+
+set(DOXYGEN_CONFIG_FILE "${CMAKE_CURRENT_BINARY_DIR}/doxygen/doxygen.conf" CACHE PATH "Path to generated doxygen configuration file")
+
+function(add_doxygen_doc)
+    set(options)
+    set(oneValueArgs)
+    set(multiValueArgs DEPENDS ${DOXYGEN_ARGS})
+
+    cmake_parse_arguments(PARSE "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+
+    file(WRITE ${DOXYGEN_CONFIG_FILE} "# Auto-generated doxygen configuration file\n")
+
+    foreach(ARG ${DOXYGEN_ARGS})
+        if(PARSE_${ARG})
+            string(REPLACE ";" " " ARG_VALUE ${PARSE_${ARG}})
+            file(APPEND ${DOXYGEN_CONFIG_FILE} "\n${ARG} = ${ARG_VALUE}\n")
+        endif()
+    endforeach()
+
+    if(PARSE_OUTPUT_DIRECTORY)
+        if(NOT EXISTS ${PARSE_OUTPUT_DIRECTORY})
+            file(MAKE_DIRECTORY ${PARSE_OUTPUT_DIRECTORY})
+        endif()
+    endif()
+
+    if(DOT_EXECUTABLE)
+        file(APPEND ${DOXYGEN_CONFIG_FILE} "\nDOT_PATH = \"${DOT_EXECUTABLE}\"\n")
+        file(APPEND ${DOXYGEN_CONFIG_FILE} "\nHAVE_DOT = YES\n")
+    else()
+        file(APPEND ${DOXYGEN_CONFIG_FILE} "\nHAVE_DOT = NO\n")
+    endif()
+
+    add_custom_target(doxygen
+        ${DOXYGEN_EXECUTABLE} ${DOXYGEN_CONFIG_FILE}
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        COMMENT "Building documentation with doxygen"
+    )
+    if(PARSE_OUTPUT_DIRECTORY)
+        clean_doc_output(${PARSE_OUTPUT_DIRECTORY})
+    endif()
+    mark_as_doc(doxygen)
+    if(PARSE_DEPENDS)
+        add_dependencies(doxygen ${PARSE_DEPENDS})
+    endif()
+endfunction()
diff --git a/src/fastertransformer/kernels/ck_kernels/cmake/EnableCompilerWarnings.cmake b/src/fastertransformer/kernels/ck_kernels/cmake/EnableCompilerWarnings.cmake
new file mode 100644
index 0000000..78133af
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/cmake/EnableCompilerWarnings.cmake
@@ -0,0 +1,110 @@
+################################################################################
+#
+# MIT License
+#
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+#
+################################################################################
+# - Enable warning all for gcc/clang or use /W4 for visual studio
+
+## Strict warning level
+if (MSVC)
+    # Use the highest warning level for visual studio.
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /w")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /w")
+    # set(CMAKE_CXX_WARNING_LEVEL 4)
+    # if (CMAKE_CXX_FLAGS MATCHES "/W[0-4]")
+    #     string(REGEX REPLACE "/W[0-4]" "/W4" CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
+    # else ()
+    #     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /W4")
+    # endif ()
+
+    # set(CMAKE_C_WARNING_LEVEL 4)
+    # if (CMAKE_C_FLAGS MATCHES "/W[0-4]")
+    #     string(REGEX REPLACE "/W[0-4]" "/W4" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
+    # else ()
+    #     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /W4")
+    # endif ()
+
+else()
+    foreach(COMPILER C CXX)
+        set(CMAKE_COMPILER_WARNINGS)
+        # use -Wall for gcc and clang
+        list(APPEND CMAKE_COMPILER_WARNINGS
+            -Wall
+            -Wextra
+            -Wcomment
+            -Wendif-labels
+            -Wformat
+            -Winit-self
+            -Wreturn-type
+            -Wsequence-point
+            # Shadow is broken on gcc when using lambdas
+            # -Wshadow
+            -Wswitch
+            -Wtrigraphs
+            -Wundef
+            -Wuninitialized
+            -Wunreachable-code
+            -Wunused
+
+            -Wsign-compare
+            -Wno-extra-semi-stmt
+        )
+        if (CMAKE_${COMPILER}_COMPILER_ID MATCHES "Clang")
+            list(APPEND CMAKE_COMPILER_WARNINGS
+                -Weverything
+                -Wno-c++98-compat
+                -Wno-c++98-compat-pedantic
+                -Wno-conversion
+                -Wno-double-promotion
+                -Wno-exit-time-destructors
+                -Wno-extra-semi
+                -Wno-float-conversion
+                -Wno-gnu-anonymous-struct
+                -Wno-gnu-zero-variadic-macro-arguments
+                -Wno-missing-prototypes
+                -Wno-nested-anon-types
+                -Wno-padded
+                -Wno-return-std-move-in-c++11
+                -Wno-shorten-64-to-32
+                -Wno-sign-conversion
+                -Wno-unknown-warning-option
+                -Wno-unused-command-line-argument
+                -Wno-weak-vtables
+                -Wno-covered-switch-default
+            )
+        else()
+            if (CMAKE_${COMPILER}_COMPILER_ID MATCHES "GNU" AND ${COMPILER} MATCHES "CXX")
+                # cmake 3.5.2 does not support >=.
+                if(NOT CMAKE_CXX_COMPILER_VERSION VERSION_LESS "6.1")
+                    list(APPEND CMAKE_COMPILER_WARNINGS
+                        -Wno-ignored-attributes)
+                endif()
+            endif()
+            list(APPEND CMAKE_COMPILER_WARNINGS
+                -Wno-missing-field-initializers
+                -Wno-deprecated-declarations
+            )
+        endif()
+        add_definitions(${CMAKE_COMPILER_WARNINGS})
+    endforeach()
+endif ()
diff --git a/src/fastertransformer/kernels/ck_kernels/cmake/TargetFlags.cmake b/src/fastertransformer/kernels/ck_kernels/cmake/TargetFlags.cmake
new file mode 100644
index 0000000..4f83fb5
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/cmake/TargetFlags.cmake
@@ -0,0 +1,50 @@
+
+function(get_target_property2 VAR TARGET PROPERTY)
+    get_target_property(_pflags ${TARGET} ${PROPERTY})
+    if(_pflags)
+        set(${VAR} ${_pflags} PARENT_SCOPE)
+    else()
+        set(${VAR} "" PARENT_SCOPE)
+    endif()
+endfunction()
+
+
+macro(append_flags FLAGS TARGET PROPERTY PREFIX)
+    get_target_property2(_pflags ${TARGET} ${PROPERTY})
+    foreach(FLAG ${_pflags})
+        if(TARGET ${FLAG})
+            target_flags(_pflags2 ${FLAG})
+            string(APPEND ${FLAGS} " ${_pflags2}")
+        else()
+            string(APPEND ${FLAGS} " ${PREFIX}${FLAG}")
+        endif()
+    endforeach()
+endmacro()
+
+macro(append_link_flags FLAGS TARGET PROPERTY)
+    get_target_property2(_pflags ${TARGET} ${PROPERTY})
+    foreach(FLAG ${_pflags})
+        if(TARGET ${FLAG})
+            target_flags(_pflags2 ${FLAG})
+            string(APPEND ${FLAGS} " ${_pflags2}")
+        elseif(FLAG MATCHES "^-.*")
+            string(APPEND ${FLAGS} " ${FLAG}")
+        elseif(EXISTS ${FLAG})
+            string(APPEND ${FLAGS} " ${FLAG}")
+        else()
+            string(APPEND ${FLAGS} " -l${FLAG}")
+        endif()
+    endforeach()
+endmacro()
+
+function(target_flags FLAGS TARGET)
+    set(_flags)
+    append_flags(_flags ${TARGET} "INTERFACE_COMPILE_OPTIONS" "")
+    append_flags(_flags ${TARGET} "INTERFACE_COMPILE_DEFINITIONS" "-D")
+    append_flags(_flags ${TARGET} "INTERFACE_INCLUDE_DIRECTORIES" "-isystem ")
+    append_flags(_flags ${TARGET} "INTERFACE_LINK_DIRECTORIES" "-L ")
+    append_flags(_flags ${TARGET} "INTERFACE_LINK_OPTIONS" "")
+    append_link_flags(_flags ${TARGET} "INTERFACE_LINK_LIBRARIES" "")
+    # message("_flags: ${_flags}")
+    set(${FLAGS} ${_flags} PARENT_SCOPE)
+endfunction()
diff --git a/src/fastertransformer/kernels/ck_kernels/cmake/googletest.cmake b/src/fastertransformer/kernels/ck_kernels/cmake/googletest.cmake
new file mode 100644
index 0000000..3da4877
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/cmake/googletest.cmake
@@ -0,0 +1,50 @@
+include(FetchContent)
+
+set(GOOGLETEST_DIR "" CACHE STRING "Location of local GoogleTest repo to build against")
+
+if(GOOGLETEST_DIR)
+  set(FETCHCONTENT_SOURCE_DIR_GOOGLETEST ${GOOGLETEST_DIR} CACHE STRING "GoogleTest source directory override")
+endif()
+
+message(STATUS "Fetching GoogleTest")
+
+list(APPEND GTEST_CMAKE_CXX_FLAGS
+     -Wno-undef
+     -Wno-reserved-identifier
+     -Wno-global-constructors
+     -Wno-missing-noreturn
+     -Wno-disabled-macro-expansion
+     -Wno-used-but-marked-unused
+     -Wno-switch-enum
+     -Wno-zero-as-null-pointer-constant
+     -Wno-unused-member-function
+     -Wno-comma
+     -Wno-old-style-cast
+     -Wno-deprecated
+)
+message(STATUS "Suppressing googltest warnings with flags: ${GTEST_CMAKE_CXX_FLAGS}")
+
+FetchContent_Declare(
+  googletest
+  GIT_REPOSITORY http://10.0.50.24/Mirrors/googletest.git
+# GIT_REPOSITORY /work/home/zhangshao/installer/googletest
+  GIT_TAG        b85864c64758dec007208e56af933fc3f52044ee
+)
+
+# Will be necessary for windows build
+# set(gtest_force_shared_crt ON CACHE BOOL "" FORCE)
+FetchContent_GetProperties(googletest)
+if(NOT googletest_POPULATED)
+  FetchContent_Populate(googletest)
+  add_subdirectory(${googletest_SOURCE_DIR} ${googletest_BINARY_DIR} EXCLUDE_FROM_ALL)
+endif()
+
+target_compile_options(gtest PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+target_compile_options(gtest_main PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+target_compile_options(gmock PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+target_compile_options(gmock_main PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+
+set_target_properties(gtest PROPERTIES POSITION_INDEPENDENT_CODE ON)
+set_target_properties(gtest_main PROPERTIES POSITION_INDEPENDENT_CODE ON)
+set_target_properties(gmock PROPERTIES POSITION_INDEPENDENT_CODE ON)
+set_target_properties(gmock_main PROPERTIES POSITION_INDEPENDENT_CODE ON)
diff --git a/src/fastertransformer/kernels/ck_kernels/cutlass_preprocessors.cc b/src/fastertransformer/kernels/ck_kernels/cutlass_preprocessors.cc
new file mode 100644
index 0000000..7f2ec59
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/cutlass_preprocessors.cc
@@ -0,0 +1,719 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/cutlass_preprocessors.h"
+#include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
+
+#include "cutlass_extensions/gemm/kernel/mixed_gemm_B_layout.h"
+
+namespace fastertransformer {
+
+int get_bits_in_quant_type(QuantType quant_type)
+{
+    switch (quant_type) {
+        case QuantType::INT8_WEIGHT_ONLY:
+            return 8;
+        case QuantType::PACKED_INT4_WEIGHT_ONLY:
+            return 4;
+        default:
+            FT_CHECK_WITH_INFO(false, "Invalid quant_type");
+            return -1;
+    }
+}
+
+struct LayoutDetails {
+    enum class Layout {
+        UNKNOWN,
+        ROW_MAJOR,
+        COLUMN_MAJOR
+    };
+
+    Layout layoutB              = Layout::UNKNOWN;
+    int    rows_per_column_tile = 1;
+    int    columns_interleaved  = 1;
+
+    bool uses_imma_ldsm = false;
+};
+
+template<typename Layout>
+struct getLayoutDetails {
+};
+
+template<>
+struct getLayoutDetails<cutlass::layout::RowMajor> {
+    LayoutDetails operator()()
+    {
+        LayoutDetails layout_details;
+        layout_details.layoutB = LayoutDetails::Layout::ROW_MAJOR;
+        return layout_details;
+    }
+};
+
+template<>
+struct getLayoutDetails<cutlass::layout::ColumnMajor> {
+    LayoutDetails operator()()
+    {
+        LayoutDetails layout_details;
+        layout_details.layoutB = LayoutDetails::Layout::COLUMN_MAJOR;
+        return layout_details;
+    }
+};
+
+template<int RowsPerTile, int ColumnsInterleaved>
+struct getLayoutDetails<cutlass::layout::ColumnMajorTileInterleave<RowsPerTile, ColumnsInterleaved>> {
+    LayoutDetails operator()()
+    {
+        LayoutDetails layout_details;
+        layout_details.layoutB              = LayoutDetails::Layout::COLUMN_MAJOR;
+        layout_details.rows_per_column_tile = RowsPerTile;
+        layout_details.columns_interleaved  = ColumnsInterleaved;
+        return layout_details;
+    }
+};
+
+template<typename cutlassArch, typename TypeB>
+LayoutDetails getLayoutDetailsForArchAndQuantType()
+{
+
+    using CompileTraits    = cutlass::gemm::kernel::LayoutDetailsB<TypeB, cutlassArch>;
+    using LayoutB          = typename CompileTraits::Layout;
+    using MmaOperator      = typename CompileTraits::Operator;
+    LayoutDetails details  = getLayoutDetails<LayoutB>()();
+    details.uses_imma_ldsm = std::is_same<MmaOperator, cutlass::arch::OpMultiplyAddDequantizeInterleavedBToA>::value;
+    return details;
+}
+
+template<typename cutlassArch>
+LayoutDetails getLayoutDetailsForArch(QuantType quant_type)
+{
+    LayoutDetails details;
+    if (quant_type == QuantType::INT8_WEIGHT_ONLY) {
+        details = getLayoutDetailsForArchAndQuantType<cutlassArch, uint8_t>();
+    }
+    else if (quant_type == QuantType::PACKED_INT4_WEIGHT_ONLY) {
+        details = getLayoutDetailsForArchAndQuantType<cutlassArch, cutlass::uint4b_t>();
+    }
+    else {
+        FT_CHECK_WITH_INFO(false, "Unsupported quantization type");
+    }
+    return details;
+}
+
+LayoutDetails getLayoutDetailsForTransform(QuantType quant_type)
+{
+    const int arch = fastertransformer::getSMVersion();
+    if (arch >= 70 && arch < 75) {
+        return getLayoutDetailsForArch<cutlass::arch::Sm70>(quant_type);
+    }
+    else if (arch >= 75 && arch < 80) {
+        return getLayoutDetailsForArch<cutlass::arch::Sm75>(quant_type);
+    }
+    else if (arch >= 80 && arch < 90) {
+        return getLayoutDetailsForArch<cutlass::arch::Sm80>(quant_type);
+    }
+    else {
+        FT_CHECK_WITH_INFO(false, "Unsupported Arch");
+        return LayoutDetails();
+    }
+}
+
+// Permutes the rows of B for Turing and Ampere. Throws an error for other architectures.
+// The data is permuted such that:
+// For int8, each group of 16 rows is permuted using the map below:
+//  0 1 8 9 2 3 10 11 4 5 12 13 6 7 14 15
+// For int4, each group of 32 rows is permuted using the map below:
+//  0 1 8 9 16 17 24 25 2 3 10 11 18 19 26 27 4 5 12 13 20 21 28 29 6 7 14 15 22 23 30 31
+void permute_B_rows_for_mixed_gemm(int8_t*                    permuted_quantized_tensor,
+                                   const int8_t*              quantized_tensor,
+                                   const std::vector<size_t>& shape,
+                                   QuantType                  quant_type,
+                                   const int64_t              arch_version)
+{
+
+    // We only want to run this step for weight only quant.
+    FT_CHECK(quant_type == QuantType::PACKED_INT4_WEIGHT_ONLY || quant_type == QuantType::INT8_WEIGHT_ONLY);
+
+    FT_CHECK_WITH_INFO(shape.size() == 2 || shape.size() == 3, "Shape must be 2-D or 3-D");
+    const size_t num_experts = shape.size() == 2 ? 1 : shape[0];
+    const size_t num_rows    = shape.size() == 2 ? shape[0] : shape[1];
+    const size_t num_cols    = shape.size() == 2 ? shape[1] : shape[2];
+
+    const int BITS_PER_ELT  = get_bits_in_quant_type(quant_type);
+    const int K             = 16 / BITS_PER_ELT;
+    const int ELTS_PER_BYTE = 8 / BITS_PER_ELT;
+    const int ELTS_PER_REG  = 32 / BITS_PER_ELT;
+
+    const uint32_t* input_byte_ptr  = reinterpret_cast<const uint32_t*>(quantized_tensor);
+    uint32_t*       output_byte_ptr = reinterpret_cast<uint32_t*>(permuted_quantized_tensor);
+
+    int       MMA_SHAPE_N    = 8;
+    int       B_ROWS_PER_MMA = 8 * K;
+    const int elts_in_int32  = 32 / BITS_PER_ELT;
+
+    const int num_vec_cols = num_cols / elts_in_int32;
+
+    FT_CHECK_WITH_INFO(arch_version >= 75,
+                       "Unsupported Arch. Pre-volta not supported. Column interleave not needed on Volta.");
+
+    FT_CHECK_WITH_INFO(
+        num_rows % B_ROWS_PER_MMA == 0,
+        fmtstr("Invalid shape for quantized tensor. Number of rows of quantized matrix must be a multiple of %d",
+               B_ROWS_PER_MMA));
+    FT_CHECK_WITH_INFO(
+        num_cols % MMA_SHAPE_N == 0,
+        fmtstr("Invalid shape for quantized tensor. On turing/Ampere, the number of cols must be a multiple of %d.",
+               MMA_SHAPE_N));
+
+    // The code is written as below so it works for both int8 and packed int4.
+    for (int expert = 0; expert < num_experts; ++expert) {
+        const int64_t matrix_offset = expert * int64_t(num_rows) * int64_t(num_vec_cols);
+        for (int base_row = 0; base_row < num_rows; base_row += B_ROWS_PER_MMA) {
+            for (int tile_row = 0; tile_row < B_ROWS_PER_MMA; ++tile_row) {
+
+                for (int write_col = 0; write_col < num_vec_cols; ++write_col) {
+                    const int write_row = base_row + tile_row;
+                    const int tile_read_row =
+                        8 * (((tile_row % ELTS_PER_REG) / 2)) + tile_row % 2 + 2 * (tile_row / ELTS_PER_REG);
+                    const int read_row = base_row + tile_read_row;
+                    const int read_col = write_col;
+
+                    const int64_t read_offset  = matrix_offset + int64_t(read_row) * num_vec_cols + read_col;
+                    const int64_t write_offset = matrix_offset + int64_t(write_row) * num_vec_cols + write_col;
+
+                    output_byte_ptr[write_offset] = input_byte_ptr[read_offset];
+                }
+            }
+        }
+    }
+}
+
+// We need to use this transpose to correctly handle packed int4 and int8 data
+// The reason this code is relatively complex is that the "trivial" loops took a substantial
+// amount of time to transpose leading to long preprocessing times. This seemed to be a big
+// issue for relatively large models.
+template<QuantType quant_type>
+void subbyte_transpose_impl(int8_t*                    transposed_quantized_tensor,
+                            const int8_t*              quantized_tensor,
+                            const std::vector<size_t>& shape)
+{
+    const int bits_per_elt = get_bits_in_quant_type(quant_type);
+
+    FT_CHECK_WITH_INFO(shape.size() == 2 || shape.size() == 3, "Shape must be 2-D or 3-D");
+    const size_t num_experts = shape.size() == 2 ? 1 : shape[0];
+    const size_t num_rows    = shape.size() == 2 ? shape[0] : shape[1];
+    const size_t num_cols    = shape.size() == 2 ? shape[1] : shape[2];
+
+    const size_t col_bytes       = num_cols * bits_per_elt / 8;
+    const size_t col_bytes_trans = num_rows * bits_per_elt / 8;
+    const size_t num_bytes       = size_t(num_experts) * num_rows * col_bytes;
+
+    const uint8_t* input_byte_ptr  = reinterpret_cast<const uint8_t*>(quantized_tensor);
+    uint8_t*       output_byte_ptr = reinterpret_cast<uint8_t*>(transposed_quantized_tensor);
+
+    static_assert(quant_type == QuantType::INT8_WEIGHT_ONLY || quant_type == QuantType::PACKED_INT4_WEIGHT_ONLY, "");
+    static constexpr int ELTS_PER_BYTE = quant_type == QuantType::INT8_WEIGHT_ONLY ? 1 : 2;
+
+    static constexpr int M_TILE_L1 = 64;
+    static constexpr int N_TILE_L1 = M_TILE_L1 / ELTS_PER_BYTE;
+    uint8_t              cache_buf[M_TILE_L1][N_TILE_L1];
+
+    static constexpr int VECTOR_WIDTH = std::min(32, N_TILE_L1);
+
+    // We assume the dims are a multiple of vector width. Our kernels only handle dims which are multiples
+    // of 64 for weight-only quantization. As a result, this seemed like a reasonable tradeoff because it
+    // allows GCC to emit vector instructions.
+    FT_CHECK_WITH_INFO(
+        !(col_bytes_trans % VECTOR_WIDTH) && !(col_bytes % VECTOR_WIDTH),
+        fmtstr(
+            "Number of bytes for rows and cols must be a multiple of %d. However, num_rows_bytes = %ld and num_col_bytes = %d.",
+            VECTOR_WIDTH,
+            col_bytes_trans,
+            col_bytes));
+
+    const int num_m_tiles = (num_rows + M_TILE_L1 - 1) / M_TILE_L1;
+    const int num_n_tiles = (col_bytes + N_TILE_L1 - 1) / N_TILE_L1;
+
+    for (size_t expert = 0; expert < num_experts; ++expert) {
+        const size_t matrix_offset = expert * num_rows * col_bytes;
+        for (size_t row_tile_start = 0; row_tile_start < num_rows; row_tile_start += M_TILE_L1) {
+            for (size_t col_tile_start_byte = 0; col_tile_start_byte < col_bytes; col_tile_start_byte += N_TILE_L1) {
+
+                const int row_limit = std::min(row_tile_start + M_TILE_L1, num_rows);
+                const int col_limit = std::min(col_tile_start_byte + N_TILE_L1, col_bytes);
+
+                for (int ii = 0; ii < M_TILE_L1; ++ii) {
+                    const int row = row_tile_start + ii;
+
+                    for (int jj = 0; jj < N_TILE_L1; jj += VECTOR_WIDTH) {
+                        const int col = col_tile_start_byte + jj;
+
+                        const size_t logical_src_offset = matrix_offset + row * col_bytes + col;
+
+                        if (row < row_limit && col < col_limit) {
+                            for (int v = 0; v < VECTOR_WIDTH; ++v) {
+                                cache_buf[ii][jj + v] = input_byte_ptr[logical_src_offset + v];
+                            }
+                        }
+                    }
+                }
+
+                if (quant_type == QuantType::INT8_WEIGHT_ONLY) {
+                    for (int ii = 0; ii < M_TILE_L1; ++ii) {
+                        for (int jj = ii + 1; jj < N_TILE_L1; ++jj) {
+                            std::swap(cache_buf[ii][jj], cache_buf[jj][ii]);
+                        }
+                    }
+                }
+                else if (quant_type == QuantType::PACKED_INT4_WEIGHT_ONLY) {
+
+                    for (int ii = 0; ii < M_TILE_L1; ++ii) {
+                        // Using M_TILE_L1 here is deliberate since we assume that the cache tile
+                        // is square in the number of elements (not necessarily the number of bytes).
+                        for (int jj = ii + 1; jj < M_TILE_L1; ++jj) {
+                            const int ii_byte       = ii / ELTS_PER_BYTE;
+                            const int ii_bit_offset = ii % ELTS_PER_BYTE;
+
+                            const int jj_byte       = jj / ELTS_PER_BYTE;
+                            const int jj_bit_offset = jj % ELTS_PER_BYTE;
+
+                            uint8_t src_elt = 0xF & (cache_buf[ii][jj_byte] >> (4 * jj_bit_offset));
+                            uint8_t tgt_elt = 0xF & (cache_buf[jj][ii_byte] >> (4 * ii_bit_offset));
+
+                            cache_buf[ii][jj_byte] &= (0xF0 >> (4 * jj_bit_offset));
+                            cache_buf[jj][ii_byte] &= (0xF0 >> (4 * ii_bit_offset));
+
+                            cache_buf[ii][jj_byte] |= (tgt_elt << (4 * jj_bit_offset));
+                            cache_buf[jj][ii_byte] |= (src_elt << (4 * ii_bit_offset));
+                        }
+                    }
+                }
+                else {
+                    FT_CHECK_WITH_INFO(false, "Unsupported quantization type.");
+                }
+
+                const size_t row_tile_start_trans      = col_tile_start_byte * ELTS_PER_BYTE;
+                const size_t col_tile_start_byte_trans = row_tile_start / ELTS_PER_BYTE;
+
+                const int row_limit_trans = std::min(row_tile_start_trans + M_TILE_L1, num_cols);
+                const int col_limit_trans = std::min(col_tile_start_byte_trans + N_TILE_L1, col_bytes_trans);
+
+                for (int ii = 0; ii < M_TILE_L1; ++ii) {
+                    const int row = row_tile_start_trans + ii;
+                    for (int jj = 0; jj < N_TILE_L1; jj += VECTOR_WIDTH) {
+                        const int col = col_tile_start_byte_trans + jj;
+
+                        const size_t logical_tgt_offset = matrix_offset + row * col_bytes_trans + col;
+
+                        if (row < row_limit_trans && col < col_limit_trans) {
+                            for (int v = 0; v < VECTOR_WIDTH; ++v) {
+                                output_byte_ptr[logical_tgt_offset + v] = cache_buf[ii][jj + v];
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+}
+
+void subbyte_transpose(int8_t*                    transposed_quantized_tensor,
+                       const int8_t*              quantized_tensor,
+                       const std::vector<size_t>& shape,
+                       QuantType                  quant_type)
+{
+
+    if (quant_type == QuantType::INT8_WEIGHT_ONLY) {
+        subbyte_transpose_impl<QuantType::INT8_WEIGHT_ONLY>(transposed_quantized_tensor, quantized_tensor, shape);
+    }
+    else if (quant_type == QuantType::PACKED_INT4_WEIGHT_ONLY) {
+        subbyte_transpose_impl<QuantType::PACKED_INT4_WEIGHT_ONLY>(
+            transposed_quantized_tensor, quantized_tensor, shape);
+    }
+    else {
+        FT_CHECK_WITH_INFO(false, "Invalid quant_tye");
+    }
+}
+
+void add_bias_and_interleave_int8s_inplace(int8_t* int8_tensor, const size_t num_elts)
+{
+    for (int ii = 0; ii < num_elts; ++ii) {
+        int8_tensor[ii] = int8_t(int(int8_tensor[ii]) + 128);
+    }
+
+    // Step 2 will transform the layout of a 32-bit register in CUDA in order to match the int4 layout. This has no
+    // performance benefit and is purely so that int4 and int8 have the same layout.
+    // Pictorially, this does the following:
+    // bit 32                                                      0
+    //      [elt_3  elt_2  elt_1  elt_0] (each elt occupies 8 bits)
+    //
+    // And it will rearrange the output 32 bit register to be the following:
+    // bit 32                                                      0
+    //      [elt_3  elt_1  elt_2  elt_0] (each elt occupies 8 bits)
+
+    FT_CHECK_WITH_INFO(num_elts % 4 == 0, "Dimensions of int8 tensor must be a multiple of 4 for register relayout");
+    for (size_t base = 0; base < num_elts; base += 4) {
+        std::swap(int8_tensor[base + 1], int8_tensor[base + 2]);
+    }
+}
+
+void add_bias_and_interleave_int4s_inplace(int8_t* packed_int4_tensor, const size_t num_elts)
+{
+    const int num_bytes = num_elts / 2;
+
+    // Step 1 will be to transform all the int4s to unsigned in order to make the dequantize take as little
+    // instructions as possible in the CUDA code.
+    for (size_t ii = 0; ii < num_bytes; ++ii) {
+        int8_t transformed_packed_int4s = 0;
+        int8_t transformed_first_elt =
+            (int8_t(packed_int4_tensor[ii] << 4) >> 4) + 8;  // The double shift here is to ensure sign extension
+        int8_t transformed_second_elt = (packed_int4_tensor[ii] >> 4) + 8;
+
+        FT_CHECK_WITH_INFO(transformed_first_elt >= 0 && transformed_first_elt <= 15,
+                           "Illegal result for int4 transform (first elt)");
+        FT_CHECK_WITH_INFO(transformed_second_elt >= 0 && transformed_second_elt <= 15,
+                           "Illegal result for int4 transform (second elt)");
+
+        // We don't need to mask in these ops since everything should be in the range 0-15
+        transformed_packed_int4s |= transformed_first_elt;
+        transformed_packed_int4s |= (transformed_second_elt << 4);
+        packed_int4_tensor[ii] = transformed_packed_int4s;
+    }
+
+    // Step 2 will transform the layout of a 32-bit register in CUDA in order to minimize the number of shift & logical
+    // instructions That are needed to extract the int4s in the GEMM main loop. Pictorially, the loop below will do the
+    // following: Take as input a 32 bit register with layout: bit 32 0
+    //      [elt_7  elt_6  elt_5  elt_4  elt_3  elt_2  elt_1  elt_0] (each elt occupies 4 bits)
+    //
+    // And it will rearrange the output 32 bit register to be the following:
+    // bit 32                                                      0
+    //      [elt_7  elt_5  elt_3  elt_1  elt_6  elt_4  elt_2  elt_0] (each elt occupies 4 bits)
+
+    FT_CHECK_WITH_INFO(num_bytes % 4 == 0, "Dimensions of int4 tensor must be a multiple of 8 for register relayout");
+    const size_t num_registers = num_bytes / 4;
+
+    uint32_t* register_ptr = reinterpret_cast<uint32_t*>(packed_int4_tensor);
+    for (size_t ii = 0; ii < num_registers; ++ii) {
+        const uint32_t current_register     = register_ptr[ii];
+        uint32_t       transformed_register = 0;
+
+        for (int dest_idx = 0; dest_idx < 8; ++dest_idx) {
+            const int src_idx    = dest_idx < 4 ? 2 * dest_idx : 2 * (dest_idx - 4) + 1;
+            const int src_shift  = 4 * src_idx;
+            const int dest_shift = 4 * dest_idx;
+
+            const uint32_t src_bits = (current_register >> src_shift) & 0xF;
+            transformed_register |= (src_bits << dest_shift);
+        }
+        register_ptr[ii] = transformed_register;
+    }
+}
+
+void add_bias_and_interleave_quantized_tensor_inplace(int8_t* tensor, const size_t num_elts, QuantType quant_type)
+{
+    if (quant_type == QuantType::INT8_WEIGHT_ONLY) {
+        add_bias_and_interleave_int8s_inplace(tensor, num_elts);
+    }
+    else if (quant_type == QuantType::PACKED_INT4_WEIGHT_ONLY) {
+        add_bias_and_interleave_int4s_inplace(tensor, num_elts);
+    }
+    else {
+        FT_CHECK_WITH_INFO(false, "Invalid quantization type for interleaving.");
+    }
+}
+
+void interleave_column_major_tensor(int8_t*                    interleaved_quantized_tensor,
+                                    const int8_t*              quantized_tensor,
+                                    const std::vector<size_t>& shape,
+                                    QuantType                  quant_type,
+                                    LayoutDetails              details)
+{
+
+    // We only want to run this step for weight only quant.
+    FT_CHECK(quant_type == QuantType::PACKED_INT4_WEIGHT_ONLY || quant_type == QuantType::INT8_WEIGHT_ONLY);
+
+    FT_CHECK_WITH_INFO(shape.size() == 2 || shape.size() == 3, "Shape must be 2-D or 3-D");
+    const size_t num_experts = shape.size() == 2 ? 1 : shape[0];
+    const size_t num_rows    = shape.size() == 2 ? shape[0] : shape[1];
+    const size_t num_cols    = shape.size() == 2 ? shape[1] : shape[2];
+
+    const int BITS_PER_ELT  = get_bits_in_quant_type(quant_type);
+    const int elts_in_int32 = 32 / BITS_PER_ELT;
+
+    const int rows_per_tile = details.rows_per_column_tile;
+
+    FT_CHECK_WITH_INFO(
+        !(num_rows % elts_in_int32),
+        fmtstr("The number of rows must be a multiple of %d but the number of rows is %d.", elts_in_int32, num_rows));
+
+    FT_CHECK_WITH_INFO(!(num_cols % rows_per_tile),
+                       fmtstr("The number of columns must be a multiple of %d but the number of columns is %ld",
+                              rows_per_tile,
+                              num_cols));
+
+    const uint32_t* input_byte_ptr  = reinterpret_cast<const uint32_t*>(quantized_tensor);
+    uint32_t*       output_byte_ptr = reinterpret_cast<uint32_t*>(interleaved_quantized_tensor);
+
+    FT_CHECK_WITH_INFO(!(num_cols % rows_per_tile),
+                       fmtstr("The number of columns must be a multiple of %d but the number of columns is %d.",
+                              rows_per_tile,
+                              num_cols));
+
+    const int num_vec_rows      = num_rows / elts_in_int32;
+    const int vec_rows_per_tile = rows_per_tile / elts_in_int32;
+    const int interleave        = details.columns_interleaved;
+
+    for (int expert = 0; expert < num_experts; ++expert) {
+        const int64_t matrix_offset = expert * int64_t(num_vec_rows) * int64_t(num_cols);
+        for (int read_col = 0; read_col < num_cols; ++read_col) {
+            const int64_t write_col = read_col / interleave;
+            for (int base_vec_row = 0; base_vec_row < num_vec_rows; base_vec_row += vec_rows_per_tile) {
+                for (int vec_read_row = base_vec_row;
+                     vec_read_row < std::min(num_vec_rows, base_vec_row + vec_rows_per_tile);
+                     ++vec_read_row) {
+                    const int64_t vec_write_row = interleave * base_vec_row
+                                                  + vec_rows_per_tile * (read_col % interleave)
+                                                  + vec_read_row % vec_rows_per_tile;
+
+                    const int64_t read_offset = matrix_offset + int64_t(read_col) * num_vec_rows + vec_read_row;
+                    const int64_t write_offset =
+                        matrix_offset + int64_t(write_col) * num_vec_rows * interleave + vec_write_row;
+                    output_byte_ptr[write_offset] = input_byte_ptr[read_offset];
+                }
+            }
+        }
+    }
+}
+
+void preprocess_weights_for_mixed_gemm(int8_t*                    preprocessed_quantized_weight,
+                                       const int8_t*              row_major_quantized_weight,
+                                       const std::vector<size_t>& shape,
+                                       QuantType                  quant_type)
+{
+    LayoutDetails details = getLayoutDetailsForTransform(quant_type);
+
+    FT_CHECK_WITH_INFO(shape.size() == 2 || shape.size() == 3, "Shape must be 2-D or 3-D");
+
+    size_t num_elts = 1;
+    for (const auto& dim : shape) {
+        num_elts *= dim;
+    }
+
+    const size_t num_bytes = num_elts * get_bits_in_quant_type(quant_type) / 8;
+
+    std::vector<int8_t> src_buf(num_bytes);
+    std::vector<int8_t> dst_buf(num_bytes);
+    std::copy(row_major_quantized_weight, row_major_quantized_weight + num_bytes, src_buf.begin());
+
+    // Works on row major data, so issue this permutation first.
+    if (details.uses_imma_ldsm) {
+        const int arch = fastertransformer::getSMVersion();
+        permute_B_rows_for_mixed_gemm(dst_buf.data(), src_buf.data(), shape, quant_type, arch);
+        src_buf.swap(dst_buf);
+    }
+
+    if (details.layoutB == LayoutDetails::Layout::COLUMN_MAJOR) {
+        subbyte_transpose(dst_buf.data(), src_buf.data(), shape, quant_type);
+        src_buf.swap(dst_buf);
+    }
+
+    if (details.columns_interleaved > 1) {
+        interleave_column_major_tensor(dst_buf.data(), src_buf.data(), shape, quant_type, details);
+        src_buf.swap(dst_buf);
+    }
+
+    add_bias_and_interleave_quantized_tensor_inplace(src_buf.data(), num_elts, quant_type);
+    std::copy(src_buf.begin(), src_buf.end(), preprocessed_quantized_weight);
+}
+
+/*
+    Arguments:
+      input_weight_ptr - the weight tensor to be quantized. Must be 2-D or 3-D and of type FP16.
+
+      quant_type - the type of the output quantization weight.
+
+    This function does symmetric quantization on 2-D or 3-D tensors. It uses the full int range and assumes the
+    zero-point is zero and will automatically construct the scales.
+
+    It always quantizes the last axis of the tensor. For 3-D tensors, it operates in "batched" mode where the tensor is
+    viewed as a stack of matrices and a scale is produced for each column of every matrix.
+
+Outputs
+    processed_quantized_weight - quantized AND processed weight for GEMM. This MUST be used with the CUTLASS GEMM
+    unprocessed_quantized_weight - quantized but unprocessed weights. Useful for reference checking.
+    scale_ptr - scales for the quantized weight.
+
+    Note that the returned quantized_weights will be preprocessed in a way to accelerate the mixed type GEMM. The data
+    layout may not make sense if printed.
+
+    Shapes:
+      quant_type == int8:
+        If weight is a [m,n] matrix, quantized_weights will have shape [m,n] and scales of shape [n]
+        If weight is a [b,m,n] tensor, unprocessed_quantized_weight will have shape [b,m,n] and scales of shape [b,n]
+      quant_type == int4:
+        If weight is a [m,n] matrix, quantized_weights will have shape [m, ceil(n/2)] and scales of shape [n]
+        If weight is a [b,m,n] tensor, unprocessed_quantized_weight will have shape [b,m, ceil(n/2)] and scales of shape
+          [b,n]
+
+      The quantized_weight will be of type torch.int8 and have two int4 values packed in a single byte. This is the
+      reason for halving the shape. At the time of writing this code, there was not an elegant way to handle this kind
+      of batched quantization using torch's quantized tensors (to the best of the author's knowledge). Scale tensors
+      must have a dimension of 1, which breaks the semantics we need for batched weights.
+  */
+
+template<typename ComputeType, typename WeightType>
+void symmetric_quantize(int8_t*                    processed_quantized_weight,
+                        int8_t*                    unprocessed_quantized_weight,
+                        ComputeType*               scale_ptr,
+                        const WeightType*          input_weight_ptr,
+                        const std::vector<size_t>& shape,
+                        QuantType                  quant_type)
+{
+
+    FT_CHECK_WITH_INFO(processed_quantized_weight, "Processed quantized tensor is NULL");
+    FT_CHECK_WITH_INFO(scale_ptr, "Scale output pointer is NULL");
+    FT_CHECK_WITH_INFO(input_weight_ptr, "Input weight pointer is NULL");
+
+    FT_CHECK_WITH_INFO(shape.size() == 2 || shape.size() == 3, "Shape must be 2-D or 3-D");
+    const size_t num_experts = shape.size() == 2 ? 1 : shape[0];
+    const size_t num_rows    = shape.size() == 2 ? shape[0] : shape[1];
+    const size_t num_cols    = shape.size() == 2 ? shape[1] : shape[2];
+
+    const int bits_in_type      = get_bits_in_quant_type(quant_type);
+    const int bytes_per_out_col = num_cols * bits_in_type / 8;
+
+    std::vector<int8_t> weight_buf;
+    if (unprocessed_quantized_weight == nullptr) {
+        weight_buf.resize(num_experts * num_rows * num_cols);
+        unprocessed_quantized_weight = weight_buf.data();
+    }
+
+    const int   input_mat_size     = num_rows * num_cols;
+    const int   quantized_mat_size = num_rows * bytes_per_out_col;
+    const float quant_range_scale  = 1.f / float(1 << (bits_in_type - 1));
+
+    std::vector<float> per_col_max(num_cols);
+
+    for (int expert = 0; expert < num_experts; ++expert) {
+        const WeightType* current_weight           = input_weight_ptr + expert * input_mat_size;
+        int8_t*           current_quantized_weight = unprocessed_quantized_weight + expert * quantized_mat_size;
+
+        // First we find the per column max for this expert weight.
+        for (int jj = 0; jj < num_cols; ++jj) {
+            per_col_max[jj] = 0.f;
+        }
+
+        for (int ii = 0; ii < num_rows; ++ii) {
+            const WeightType* current_weight_row = current_weight + ii * num_cols;
+            for (int jj = 0; jj < num_cols; ++jj) {
+                per_col_max[jj] = std::max(per_col_max[jj], std::abs(float(current_weight_row[jj])));
+            }
+        }
+
+        // Then, we construct the scales
+        ComputeType* current_scales = scale_ptr + expert * num_cols;
+        for (int jj = 0; jj < num_cols; ++jj) {
+            per_col_max[jj] *= quant_range_scale;
+            current_scales[jj] = ComputeType(per_col_max[jj]);
+        }
+
+        // Finally, construct the weights.
+        for (int ii = 0; ii < num_rows; ++ii) {
+            int8_t*           current_quantized_weight_row = current_quantized_weight + ii * bytes_per_out_col;
+            const WeightType* current_weight_row           = current_weight + ii * num_cols;
+            for (int jj = 0; jj < bytes_per_out_col; ++jj) {
+
+                if (quant_type == QuantType::INT8_WEIGHT_ONLY) {
+                    const float  col_scale           = per_col_max[jj];
+                    const float  weight_elt          = float(current_weight_row[jj]);
+                    const float  scaled_weight       = round(weight_elt / col_scale);
+                    const int8_t clipped_weight      = int8_t(std::max(-128.f, std::min(127.f, scaled_weight)));
+                    current_quantized_weight_row[jj] = clipped_weight;
+                }
+                else if (quant_type == QuantType::PACKED_INT4_WEIGHT_ONLY) {
+
+                    // We will pack two int4 elements per iteration of the inner loop.
+                    int8_t packed_int4s = 0;
+                    for (int packed_idx = 0; packed_idx < 2; ++packed_idx) {
+                        const int input_idx = 2 * jj + packed_idx;
+                        if (input_idx < num_cols) {
+                            const float  col_scale      = per_col_max[input_idx];
+                            const float  weight_elt     = float(current_weight_row[input_idx]);
+                            const float  scaled_weight  = round(weight_elt / col_scale);
+                            int          int_weight     = int(scaled_weight);
+                            const int8_t clipped_weight = std::max(-8, std::min(7, int_weight));
+
+                            // Kill the sign extension bits (hence 0x0F mask) then shift to upper bits
+                            // if packing the second int4 and or the bits into the final result.
+                            packed_int4s |= ((clipped_weight & 0x0F) << (4 * packed_idx));
+                        }
+                    }
+                    current_quantized_weight_row[jj] = packed_int4s;
+                }
+                else {
+                    FT_CHECK_WITH_INFO(false, "Unsupported quantization type");
+                }
+            }
+        }
+    }
+
+    preprocess_weights_for_mixed_gemm(processed_quantized_weight, unprocessed_quantized_weight, shape, quant_type);
+}
+
+template void
+symmetric_quantize<half, float>(int8_t*, int8_t*, half*, const float*, const std::vector<size_t>&, QuantType);
+
+template void
+symmetric_quantize<half, half>(int8_t*, int8_t*, half*, const half*, const std::vector<size_t>&, QuantType);
+
+#ifdef ENABLE_BF16
+template void symmetric_quantize<__nv_bfloat16, __nv_bfloat16>(
+    int8_t*, int8_t*, __nv_bfloat16*, const __nv_bfloat16*, const std::vector<size_t>&, QuantType);
+
+template void symmetric_quantize<__nv_bfloat16, float>(
+    int8_t*, int8_t*, __nv_bfloat16*, const float*, const std::vector<size_t>&, QuantType);
+#endif
+
+template<typename ComputeType, typename WeightType>
+void symmetric_quantize(int8_t*                    processed_quantized_weight,
+                        ComputeType*               scale_ptr,
+                        const WeightType*          input_weight_ptr,
+                        const std::vector<size_t>& shape,
+                        QuantType                  quant_type)
+{
+    symmetric_quantize(processed_quantized_weight, nullptr, scale_ptr, input_weight_ptr, shape, quant_type);
+}
+
+template void symmetric_quantize<float, float>(int8_t*, float*, const float*, const std::vector<size_t>&, QuantType);
+
+template void symmetric_quantize<half, float>(int8_t*, half*, const float*, const std::vector<size_t>&, QuantType);
+
+template void symmetric_quantize<half, half>(int8_t*, half*, const half*, const std::vector<size_t>&, QuantType);
+
+#ifdef ENABLE_BF16
+template void symmetric_quantize<__nv_bfloat16, __nv_bfloat16>(
+    int8_t*, __nv_bfloat16*, const __nv_bfloat16*, const std::vector<size_t>&, QuantType);
+
+template void
+symmetric_quantize<__nv_bfloat16, half>(int8_t*, __nv_bfloat16*, const half*, const std::vector<size_t>&, QuantType);
+
+template void
+symmetric_quantize<half, __nv_bfloat16>(int8_t*, half*, const __nv_bfloat16*, const std::vector<size_t>&, QuantType);
+
+template void
+symmetric_quantize<__nv_bfloat16, float>(int8_t*, __nv_bfloat16*, const float*, const std::vector<size_t>&, QuantType);
+#endif
+
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/cutlass_preprocessors.h b/src/fastertransformer/kernels/ck_kernels/cutlass_preprocessors.h
new file mode 100644
index 0000000..ad7cfa0
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/cutlass_preprocessors.h
@@ -0,0 +1,70 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include <cstddef>
+#include <stdint.h>
+#include <vector>
+
+#include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
+
+namespace fastertransformer {
+enum class QuantType {
+    INT8_WEIGHT_ONLY,
+    PACKED_INT4_WEIGHT_ONLY
+};
+int get_bits_in_quant_type(QuantType quant_type);
+
+// Shapes here can be 2 or 3D. 2-D shapes are [num_rows, num_cols]
+// 3-D shapes are [num_experts, num_rows, num_cols]
+void permute_B_rows_for_mixed_gemm(int8_t*                    permuted_quantized_tensor,
+                                   const int8_t*              quantized_tensor,
+                                   const std::vector<size_t>& shape,
+                                   QuantType                  quant_type,
+                                   const int64_t              arch_version);
+
+void subbyte_transpose(int8_t*                    transposed_quantized_tensor,
+                       const int8_t*              quantized_tensor,
+                       const std::vector<size_t>& shape,
+                       QuantType                  quant_type);
+
+void add_bias_and_interleave_quantized_tensor_inplace(int8_t* tensor, const size_t num_elts, QuantType quant_type);
+
+void preprocess_weights_for_mixed_gemm(int8_t*                    preprocessed_quantized_weight,
+                                       const int8_t*              row_major_quantized_weight,
+                                       const std::vector<size_t>& shape,
+                                       QuantType                  quant_type);
+
+template<typename ComputeType, typename WeightType>
+void symmetric_quantize(int8_t*                    processed_quantized_weight,
+                        ComputeType*               scale_ptr,
+                        const WeightType*          input_weight_ptr,
+                        const std::vector<size_t>& shape,
+                        QuantType                  quant_type);
+
+// This is exposed so that we can write tests that use the processed weights for CUTLASS but the unprocessed weight
+// to implement a simple reference implementation.
+template<typename ComputeType, typename WeightType>
+void symmetric_quantize(int8_t*                    processed_quantized_weight,
+                        int8_t*                    unprocessed_quantized_weight,
+                        ComputeType*               scale_ptr,
+                        const WeightType*          input_weight_ptr,
+                        const std::vector<size_t>& shape,
+                        QuantType                  quant_type);
+
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm.h b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm.h
new file mode 100644
index 0000000..c792e7d
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm.h
@@ -0,0 +1,131 @@
+#pragma once
+
+#include "src/fastertransformer/ck_extensions/include/ck_extensions/ft_gemm_configs.h"
+#include "src/fastertransformer/utils/activation_types.h"
+// #include "hip/hip_runtime.h"
+
+namespace fastertransformer {
+
+/*
+  This runner only supports:
+  T in {half, __nv_bfloat} WeightType in {int8_t, cutlass::uint4b_t}
+
+  Activations, biases, scales and outputs are all assumed to be row-major.
+
+  However, it is assumed that B is in a special format governed by cutlass_extensions/gemm/kernel/mixed_gemm_B_layout.
+  In this case, B must be preprocessed using the cutlass weight only quant preprocessors. The weight preprocessor
+  will instantiate the layout and preprocess based on the instantiation, so layout changes should only require
+  modifications to mix_gemm_B_layout.h.
+*/
+
+template<typename T, typename WeightType>
+class CutlassFpAIntBGemmRunner {
+public:
+    CutlassFpAIntBGemmRunner();
+    ~CutlassFpAIntBGemmRunner();
+
+    void gemm(const T*          A,
+              const WeightType* B,
+              const T*          weight_scales,
+              T*                C,
+              int               m,
+              int               n,
+              int               k,
+              char*             workspace_ptr,
+              const size_t      workspace_bytes
+              );
+            //   hipStream_t      stream);
+
+    void gemm_bias_act(const T*          A,
+                       const WeightType* B,
+                       const T*          weight_scales,
+                       const T*          biases,
+                       T*                C,
+                       int               m,
+                       int               n,
+                       int               k,
+                       ActivationType    activation_type,
+                       char*             workspace_ptr,
+                       const size_t      workspace_bytes
+                       );
+                    //    hipStream_t      stream);
+
+    // Returns desired workspace size in bytes.
+    int getWorkspaceSize(const int m, const int n, const int k);
+
+private:
+    template<typename EpilogueTag>
+    void dispatch_to_arch(const T*          A,
+                          const WeightType* B,
+                          const T*          weight_scales,
+                          const T*          biases,
+                          T*                C,
+                          int               m,
+                          int               n,
+                          int               k,
+                          CutlassGemmConfig gemm_config,
+                          char*             workspace_ptr,
+                          const size_t      workspace_bytes,
+                        //   hipStream_t      stream,
+                          int*              occupancy = nullptr);
+
+    template<typename EpilogueTag>
+    void run_gemm(const T*          A,
+                  const WeightType* B,
+                  const T*          weight_scales,
+                  const T*          biases,
+                  T*                C,
+                  int               m,
+                  int               n,
+                  int               k,
+                  char*             workspace_ptr,
+                  const size_t      workspace_bytes
+                  );
+                //   hipStream_t      stream);
+
+private:
+    static constexpr int split_k_limit = 7;
+
+    int sm_;
+    int multi_processor_count_;
+};
+
+// This allocation is present to help with compiling with other structures in FT. It will throw an error
+// in all functions because this runner assumes the weight type and the activation type are different.
+// We allow empty classes to be created, but any calls to gemm or gemm_bias_act will throw an error.
+template<typename WeightType>
+class CutlassFpAIntBGemmRunner<float, WeightType> {
+public:
+    CutlassFpAIntBGemmRunner()  = default;
+    ~CutlassFpAIntBGemmRunner() = default;
+
+    void gemm(const float*      A,
+              const WeightType* B,
+              const float*      weight_scales,
+              float*            C,
+              int               m,
+              int               n,
+              int               k,
+              char*             workspace_ptr,
+              const size_t      workspace_bytes
+              );
+            //   hipStream_t      stream);
+
+    void gemm_bias_act(const float*      A,
+                       const WeightType* B,
+                       const float*      weight_scales,
+                       const float*      biases,
+                       float*            C,
+                       int               m,
+                       int               n,
+                       int               k,
+                       ActivationType    activation_type,
+                       char*             workspace_ptr,
+                       const size_t      workspace_bytes
+                       );
+                    //    hipStream_t      stream);
+
+    int getWorkspaceSize(const int m, const int n, const int k);
+};
+
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_bf16_uint4.cpp b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_bf16_uint4.cpp
new file mode 100644
index 0000000..6e7ab95
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_bf16_uint4.cpp
@@ -0,0 +1,23 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h"
+
+namespace fastertransformer {
+#ifdef ENABLE_BF16
+template class CutlassFpAIntBGemmRunner<__nv_bfloat16, cutlass::uint4b_t>;
+#endif
+}  // namespace fastertransformer
diff --git a/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_bf16_uint8.cpp b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_bf16_uint8.cpp
new file mode 100644
index 0000000..87aa4e8
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_bf16_uint8.cpp
@@ -0,0 +1,23 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h"
+
+namespace fastertransformer {
+#ifdef ENABLE_BF16
+template class CutlassFpAIntBGemmRunner<__nv_bfloat16, uint8_t>;
+#endif
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_dummy_stubs.cpp b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_dummy_stubs.cpp
new file mode 100644
index 0000000..c664f34
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_dummy_stubs.cpp
@@ -0,0 +1,22 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h"
+
+namespace fastertransformer {
+template class CutlassFpAIntBGemmRunner<float, uint8_t>;
+// template class CutlassFpAIntBGemmRunner<float, cutlass::uint4b_t>;
+}  // namespace fastertransformer
diff --git a/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_fp16_int4.cpp b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_fp16_int4.cpp
new file mode 100644
index 0000000..6df08a2
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_fp16_int4.cpp
@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h"
+
+namespace fastertransformer {
+// template class CutlassFpAIntBGemmRunner<half, cutlass::uint4b_t>;
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_fp16_int8.cpp b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_fp16_int8.cpp
new file mode 100644
index 0000000..44bed0c
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_fp16_int8.cpp
@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h"
+
+namespace fastertransformer {
+template class CutlassFpAIntBGemmRunner<half, uint8_t>;
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h
new file mode 100644
index 0000000..0987812
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h
@@ -0,0 +1,120 @@
+#include "src/fastertransformer/kernels/ck_kernels/kernels.h"
+#include "src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
+#include "ck_extensions/ck_utils.h"
+#include "src/fastertransformer/utils/logger.h"
+#include "hip/hip_runtime.h"
+
+namespace fastertransformer {
+
+template<typename T, typename WeightType>
+CutlassFpAIntBGemmRunner<T, WeightType>::CutlassFpAIntBGemmRunner()
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+    int device{-1};
+    check_hip_error(hipGetDevice(&device));
+    sm_ = getSMVersion();
+    check_hip_error(hipDeviceGetAttribute(&multi_processor_count_, hipDeviceAttributeMultiprocessorCount, device));
+}
+
+
+template<typename T, typename WeightType>
+CutlassFpAIntBGemmRunner<T, WeightType>::~CutlassFpAIntBGemmRunner()
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+}
+
+
+template<typename T, typename WeightType>
+void CutlassFpAIntBGemmRunner<T, WeightType>::gemm_bias_act(const T*          A,
+                                                            const WeightType* B,
+                                                            const T*          weight_scales,
+                                                            const T*          biases,
+                                                            T*                C,
+                                                            int               m,
+                                                            int               n,
+                                                            int               k,
+                                                            ActivationType    activation_type,
+                                                            char*             workspace_ptr,
+                                                            const size_t      workspace_bytes
+                                                            )
+                                                            // hipStream_t      stream)
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+
+
+}
+
+template<typename T, typename WeightType>
+void CutlassFpAIntBGemmRunner<T, WeightType>::gemm(const T*          A,
+                                                   const WeightType* B,
+                                                   const T*          weight_scales,
+                                                   T*                C,
+                                                   int               m,
+                                                   int               n,
+                                                   int               k,
+                                                   char*             workspace_ptr,
+                                                   const size_t      workspace_bytes
+                                                   )
+                                                //    hipStream_t      stream)
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+}
+
+
+template<typename T, typename WeightType>
+int CutlassFpAIntBGemmRunner<T, WeightType>::getWorkspaceSize(const int m, const int n, const int k)
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+    // These are the min tile sizes for each config, which would launch the maximum number of blocks
+    const int max_grid_m = (m + 31) / 32;
+    const int max_grid_n = (n + 127) / 128;
+    // We need 4 bytes per block in the worst case. We launch split_k_limit in z dim.
+    return max_grid_m * max_grid_n * split_k_limit * 4;
+}
+
+// =============================== Specialization T == WeightType =======================================
+template<typename WeightType>
+void CutlassFpAIntBGemmRunner<float, WeightType>::gemm_bias_act(const float*      A,
+                                                                const WeightType* B,
+                                                                const float*      weight_scales,
+                                                                const float*      biases,
+                                                                float*            C,
+                                                                int               m,
+                                                                int               n,
+                                                                int               k,
+                                                                ActivationType    activation_type,
+                                                                char*             workspace_ptr,
+                                                                const size_t      workspace_bytes
+                                                                )
+                                                                // hipStream_t      stream)
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+    FT_CHECK_WITH_INFO(false, "Attempting to run mixed gemm bias act when the types are the same is an error.");
+}
+
+template<typename WeightType>
+void CutlassFpAIntBGemmRunner<float, WeightType>::gemm(const float*      A,
+                                                       const WeightType* B,
+                                                       const float*      weight_scales,
+                                                       float*            C,
+                                                       int               m,
+                                                       int               n,
+                                                       int               k,
+                                                       char*             workspace_ptr,
+                                                       const size_t      workspace_bytes
+                                                       )
+                                                    //    hipStream_t      stream)
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+    FT_CHECK_WITH_INFO(false, "Attempting to run mixed gemm when the types are the same is an error.");
+}
+
+template<typename WeightType>
+int CutlassFpAIntBGemmRunner<float, WeightType>::getWorkspaceSize(const int m, const int n, const int k)
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+    return 0;
+}
+
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm.h b/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm.h
new file mode 100644
index 0000000..d7f7dfd
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm.h
@@ -0,0 +1,49 @@
+#pragma once
+
+#include "src/fastertransformer/ck_extensions/include/ck_extensions/epilogue/epilogue_quant_helper.h"
+#include "src/fastertransformer/ck_extensions/include/ck_extensions/ft_gemm_configs.h"
+// #include "hip/hip_runtime.h"
+
+using cutlass::epilogue::QuantMode;
+namespace fastertransformer {
+
+/*
+  This runner supports:
+  int8_t inputs (A and B)
+  float alpha scalings (either per-col, or per-col x per-row)
+  T output (D) where T = {float, half, __nv_bfloat16} // TODO(mseznec)
+
+  Activations, biases, scales and outputs are all assumed to be row-major.
+  Weights are assumed to be column-major.
+*/
+
+template<typename T>
+class CutlassInt8GemmRunner {
+public:
+    CutlassInt8GemmRunner();
+    ~CutlassInt8GemmRunner();
+
+    void gemm(const int8_t* A,
+              const int8_t* B,
+              QuantMode     quant_mode,
+              const float*  alpha_col,
+              const float*  alpha_row,
+              T*            C,
+              int           m,
+              int           n,
+              int           k,
+              char*         workspace_ptr,
+              const size_t  workspace_bytes
+              );
+              // hipStream_t  stream);
+
+    // Returns desired workspace size in bytes.
+    int getWorkspaceSize(const int m, const int n, const int k);
+
+private:
+    static constexpr int split_k_limit = 7;
+
+    int multi_processor_count_;
+};
+
+}  // namespace fastertransformer
diff --git a/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_bf16.cpp b/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_bf16.cpp
new file mode 100644
index 0000000..c8182c6
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_bf16.cpp
@@ -0,0 +1,23 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_template.h"
+
+namespace fastertransformer {
+#ifdef ENABLE_BF16
+template class CutlassInt8GemmRunner<__nv_bfloat16>;
+#endif
+}  // namespace fastertransformer
diff --git a/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_fp16.cpp b/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_fp16.cpp
new file mode 100644
index 0000000..3db4935
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_fp16.cpp
@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_template.h"
+
+namespace fastertransformer {
+template class CutlassInt8GemmRunner<half>;
+}  // namespace fastertransformer
diff --git a/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_fp32.cpp b/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_fp32.cpp
new file mode 100644
index 0000000..71b1c6c
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_fp32.cpp
@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_template.h"
+
+namespace fastertransformer {
+template class CutlassInt8GemmRunner<float>;  // for compilation only
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_int32.cpp b/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_int32.cpp
new file mode 100644
index 0000000..35c0704
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_int32.cpp
@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_template.h"
+
+namespace fastertransformer {
+template class CutlassInt8GemmRunner<int32_t>;
+}  // namespace fastertransformer
diff --git a/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_template.h b/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_template.h
new file mode 100644
index 0000000..1ec1001
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm_template.h
@@ -0,0 +1,71 @@
+#include "src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm.h"
+#include "src/fastertransformer/kernels/ck_kernels/kernels.h"
+#include "src/fastertransformer/utils/logger.h"
+
+#include "ck_extensions/gemm/kernel/default_fpA_intB_traits.h"
+#include "ck_extensions/ck_utils.h"
+#include "ck/utility/sequence.hpp"
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "hip/hip_runtime.h"
+
+namespace fastertransformer {
+template<typename T>
+CutlassInt8GemmRunner<T>::CutlassInt8GemmRunner()
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+
+}
+
+template<typename T>
+CutlassInt8GemmRunner<T>::~CutlassInt8GemmRunner()
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+}
+
+template<ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+template<typename T>
+void CutlassInt8GemmRunner<T>::gemm(const int8_t* A,
+                                    const int8_t* B,
+                                    QuantMode     quant_mode,
+                                    const float*  alpha_row,
+                                    const float*  alpha_col,
+                                    T*            C,
+                                    int           m,
+                                    int           n,
+                                    int           k,
+                                    char*         workspace_ptr,
+                                    const size_t  workspace_bytes
+                                    )
+                                    // hipStream_t  stream)
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+    
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+    using Gemm = ck::tensor_operation::device::DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>;
+
+}
+
+template<typename T>
+int CutlassInt8GemmRunner<T>::getWorkspaceSize(const int m, const int n, const int k)
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+    // These are the min tile sizes for each config, which would launch the maximum number of blocks
+    const int max_grid_m = (m + 31) / 32;
+    const int max_grid_n = (n + 127) / 128;
+    // We need 4 bytes per block in the worst case. We launch split_k_limit in z dim.
+    return max_grid_m * max_grid_n * split_k_limit * 4;
+}
+
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/kernels.h b/src/fastertransformer/kernels/ck_kernels/kernels.h
new file mode 100644
index 0000000..5fa55ba
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/kernels.h
@@ -0,0 +1,19 @@
+// #include "src/fastertransformer/utils/cuda_utils.h"
+#include "hip/hip_runtime.h"
+
+namespace fastertransformer{
+template<typename T, typename U>
+void ck_int8_gemm(const T* A,
+                                        const T* B,
+                                        const U*  alpha_col,
+                                        const U*  alpha_row,
+                                        T*            C,
+                                        int           m,
+                                        int           n,
+                                        int           k,
+                                        char*         workspace_ptr,
+                                        const size_t  workspace_bytes,
+                                        hipStream_t  stream);
+
+
+} // namespace
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels.h b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels.h
new file mode 100644
index 0000000..41f0e50
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels.h
@@ -0,0 +1,75 @@
+#pragma once
+#include "src/fastertransformer/ck_extensions/include/ck_extensions/ft_gemm_configs.h"
+#include "src/fastertransformer/utils/activation_types.h"
+// #include "hip/hip_runtime.h"
+
+namespace fastertransformer {
+
+template<typename T, /*The type used for activations/scales/compute*/
+         typename WeightType /* The type for the MoE weights */>
+class MoeGemmRunner {
+public:
+    MoeGemmRunner();
+
+    void moe_gemm_bias_act(const T*          A,
+                           const WeightType* B,
+                           const T*          weight_scales,
+                           const T*          biases,
+                           T*                C,
+                           int64_t*          total_rows_before_expert,
+                           int64_t           total_rows,
+                           int64_t           gemm_n,
+                           int64_t           gemm_k,
+                           int               num_experts,
+                           ActivationType    activation_type
+                           );
+                        //    hipStream_t      stream);
+
+    void moe_gemm(const T*          A,
+                  const WeightType* B,
+                  const T*          weight_scales,
+                  T*                C,
+                  int64_t*          total_rows_before_expert,
+                  int64_t           total_rows,
+                  int64_t           gemm_n,
+                  int64_t           gemm_k,
+                  int               num_experts
+                  );
+                //   hipStream_t      stream);
+
+private:
+    template<typename EpilogueTag>
+    void dispatch_to_arch(const T*          A,
+                          const WeightType* B,
+                          const T*          weight_scales,
+                          const T*          biases,
+                          T*                C,
+                          int64_t*          total_rows_before_expert,
+                          int64_t           total_rows,
+                          int64_t           gemm_n,
+                          int64_t           gemm_k,
+                          int               num_experts,
+                          CutlassGemmConfig gemm_config,
+                        //   hipStream_t      stream,
+                          int*              occupancy = nullptr);
+
+    template<typename EpilogueTag>
+    void run_gemm(const T*          A,
+                  const WeightType* B,
+                  const T*          weight_scales,
+                  const T*          biases,
+                  T*                C,
+                  int64_t*          total_rows_before_expert,
+                  int64_t           total_rows,
+                  int64_t           gemm_n,
+                  int64_t           gemm_k,
+                  int               num_experts
+                  );
+                //   hipStream_t      stream);
+
+private:
+    int sm_;
+    int multi_processor_count_;
+};
+
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_bf16_bf16.cpp b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_bf16_bf16.cpp
new file mode 100644
index 0000000..32d461d
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_bf16_bf16.cpp
@@ -0,0 +1,23 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_template.h"
+
+namespace fastertransformer {
+#ifdef ENABLE_BF16
+template class MoeGemmRunner<__nv_bfloat16, __nv_bfloat16>;
+#endif
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_bf16_uint4.cpp b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_bf16_uint4.cpp
new file mode 100644
index 0000000..12ec924
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_bf16_uint4.cpp
@@ -0,0 +1,23 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_template.h"
+
+namespace fastertransformer {
+#ifdef ENABLE_BF16
+template class MoeGemmRunner<__nv_bfloat16, cutlass::uint4b_t>;
+#endif
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_bf16_uint8.cpp b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_bf16_uint8.cpp
new file mode 100644
index 0000000..6ae094c
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_bf16_uint8.cpp
@@ -0,0 +1,23 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_template.h"
+
+namespace fastertransformer {
+#ifdef ENABLE_BF16
+template class MoeGemmRunner<__nv_bfloat16, uint8_t>;
+#endif
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp16_fp16.cpp b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp16_fp16.cpp
new file mode 100644
index 0000000..7b4b910
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp16_fp16.cpp
@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_template.h"
+
+namespace fastertransformer {
+template class MoeGemmRunner<half, half>;
+}
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp16_uint4.cpp b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp16_uint4.cpp
new file mode 100644
index 0000000..8a21782
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp16_uint4.cpp
@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_template.h"
+
+namespace fastertransformer {
+// template class MoeGemmRunner<half, cutlass::uint4b_t>;
+}
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp16_uint8.cpp b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp16_uint8.cpp
new file mode 100644
index 0000000..fa7eea1
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp16_uint8.cpp
@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_template.h"
+
+namespace fastertransformer {
+template class MoeGemmRunner<half, uint8_t>;
+}
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp32_fp32.cpp b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp32_fp32.cpp
new file mode 100644
index 0000000..7aeb338
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_fp32_fp32.cpp
@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_template.h"
+
+namespace fastertransformer {
+template class MoeGemmRunner<float, float>;
+}
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_template.h b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_template.h
new file mode 100644
index 0000000..40ea69a
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels_template.h
@@ -0,0 +1,56 @@
+// Ignore CUTLASS warnings about type punning
+#include "src/fastertransformer/kernels/ck_kernels/kernels.h"
+#include "src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels.h"
+#include "src/fastertransformer/ck_extensions/include/ck_extensions/ck_utils.h"
+#include "src/fastertransformer/utils/logger.h"
+#include <math.h>
+#include <sstream>
+#include "hip/hip_runtime.h"
+
+namespace fastertransformer {
+template<typename T, typename WeightType>
+MoeGemmRunner<T, WeightType>::MoeGemmRunner()
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+    int device{-1};
+    check_hip_error(hipGetDevice(&device));
+    sm_ = getSMVersion();
+    check_hip_error(hipDeviceGetAttribute(&multi_processor_count_, hipDeviceAttributeMultiprocessorCount, device));
+}
+
+template<typename T, typename WeightType>
+void MoeGemmRunner<T, WeightType>::moe_gemm_bias_act(const T*          A,
+                                                     const WeightType* B,
+                                                     const T*          weight_scales,
+                                                     const T*          biases,
+                                                     T*                C,
+                                                     int64_t*          total_rows_before_expert,
+                                                     int64_t           total_rows,
+                                                     int64_t           gemm_n,
+                                                     int64_t           gemm_k,
+                                                     int               num_experts,
+                                                     ActivationType    activation_type
+                                                     )
+                                                    //  hipStream_t      stream)
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+}
+
+template<typename T, typename WeightType>
+void MoeGemmRunner<T, WeightType>::moe_gemm(const T*          A,
+                                            const WeightType* B,
+                                            const T*          weight_scales,
+                                            T*                C,
+                                            int64_t*          total_rows_before_expert,
+                                            int64_t           total_rows,
+                                            int64_t           gemm_n,
+                                            int64_t           gemm_k,
+                                            int               num_experts
+                                            )
+                                            // hipStream_t      stream)
+{
+    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+
+}
+
+}  // namespace fastertransformer
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/ck_kernels/utils.h b/src/fastertransformer/kernels/ck_kernels/utils.h
new file mode 100644
index 0000000..b0efdec
--- /dev/null
+++ b/src/fastertransformer/kernels/ck_kernels/utils.h
@@ -0,0 +1,30 @@
+#pragma once
+
+namespace fastertransformer {
+
+[[noreturn]] inline void throwRuntimeError(const char* const file, int const line, std::string const& info = "")
+{
+    throw std::runtime_error(std::string("[FT][ERROR] ") + info + " Assertion fail: " + file + ":"
+                             + std::to_string(line) + " \n");
+}
+
+
+inline void myAssert(bool result, const char* const file, int const line, std::string const& info = "")
+{
+    if (!result) {
+        throwRuntimeError(file, line, info);
+    }
+}
+
+#define FT_CHECK(val) myAssert(val, __FILE__, __LINE__)
+#define FT_CHECK_WITH_INFO(val, info)                                                                                  \
+    do {                                                                                                               \
+        bool is_valid_val = (val);                                                                                     \
+        if (!is_valid_val) {                                                                                           \
+            fastertransformer::myAssert(is_valid_val, __FILE__, __LINE__, (info));                                     \
+        }                                                                                                              \
+    } while (0)
+
+#define FT_THROW(info) throwRuntimeError(__FILE__, __LINE__, info)
+
+}
\ No newline at end of file
diff --git a/src/fastertransformer/kernels/custom_ar_kernels.h b/src/fastertransformer/kernels/custom_ar_kernels.h
index 2466f27..d6b7ec1 100644
--- a/src/fastertransformer/kernels/custom_ar_kernels.h
+++ b/src/fastertransformer/kernels/custom_ar_kernels.h
@@ -22,6 +22,7 @@
 #include <iostream>
 
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 #define CUSTOM_AR_SIZE_THRESHOLD 50331648
 #define MAX_ALL_REDUCE_BLOCKS 24
diff --git a/src/fastertransformer/kernels/cutlass_kernels/CMakeLists.txt b/src/fastertransformer/kernels/cutlass_kernels/CMakeLists.txt
index 77374d6..519a50a 100644
--- a/src/fastertransformer/kernels/cutlass_kernels/CMakeLists.txt
+++ b/src/fastertransformer/kernels/cutlass_kernels/CMakeLists.txt
@@ -27,12 +27,11 @@
 # target_link_libraries(moe_gemm_kernels PRIVATE cutlass_heuristic)
 
 # set(fpA_intB_files "")
-# file(GLOB fpA_intB_files ${fpA_intB_files} ./fpA_intB_gemm/*.cu)
-# add_library(fpA_intB_gemm STATIC ${fpA_intB_files})
-# # target_compile_options(fpA_intB_gemm PRIVATE "-Xptxas=-v")
-# set_property(TARGET fpA_intB_gemm PROPERTY POSITION_INDEPENDENT_CODE ON)
-# set_property(TARGET fpA_intB_gemm PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
-# target_link_libraries(fpA_intB_gemm PRIVATE cutlass_heuristic)
+file(GLOB fpA_intB_files ${fpA_intB_files} ./fpA_intB_gemm/*.cu)
+add_library(fpA_intB_gemm STATIC ${fpA_intB_files})
+set_property(TARGET fpA_intB_gemm PROPERTY POSITION_INDEPENDENT_CODE ON)
+set_property(TARGET fpA_intB_gemm PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
+target_link_libraries(fpA_intB_gemm PRIVATE cutlass_heuristic)
 
 # set(int8_files "")
 # file(GLOB int8_files ${int8_files} ./int8_gemm/*.cu)
diff --git a/src/fastertransformer/kernels/cutlass_kernels/cutlass_heuristic.h b/src/fastertransformer/kernels/cutlass_kernels/cutlass_heuristic.h
index cd955d2..819e9b5 100644
--- a/src/fastertransformer/kernels/cutlass_kernels/cutlass_heuristic.h
+++ b/src/fastertransformer/kernels/cutlass_kernels/cutlass_heuristic.h
@@ -18,6 +18,7 @@
 
 #include "cutlass_extensions/ft_gemm_configs.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 
diff --git a/src/fastertransformer/kernels/cutlass_kernels/cutlass_preprocessors.h b/src/fastertransformer/kernels/cutlass_kernels/cutlass_preprocessors.h
index fc67a9f..2cb8d78 100644
--- a/src/fastertransformer/kernels/cutlass_kernels/cutlass_preprocessors.h
+++ b/src/fastertransformer/kernels/cutlass_kernels/cutlass_preprocessors.h
@@ -21,6 +21,7 @@
 #include <vector>
 
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 enum class QuantType {
diff --git a/src/fastertransformer/kernels/cutlass_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h b/src/fastertransformer/kernels/cutlass_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h
index 6255363..917e569 100644
--- a/src/fastertransformer/kernels/cutlass_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h
+++ b/src/fastertransformer/kernels/cutlass_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h
@@ -32,6 +32,7 @@
 #include "src/fastertransformer/kernels/cutlass_kernels/cutlass_heuristic.h"
 #include "src/fastertransformer/kernels/cutlass_kernels/fpA_intB_gemm/fpA_intB_gemm.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 
diff --git a/src/fastertransformer/kernels/cutlass_kernels/int8_gemm/int8_gemm_template.h b/src/fastertransformer/kernels/cutlass_kernels/int8_gemm/int8_gemm_template.h
index ab9d5e5..5a59b6a 100644
--- a/src/fastertransformer/kernels/cutlass_kernels/int8_gemm/int8_gemm_template.h
+++ b/src/fastertransformer/kernels/cutlass_kernels/int8_gemm/int8_gemm_template.h
@@ -38,6 +38,7 @@
 #include "src/fastertransformer/kernels/cutlass_kernels/cutlass_heuristic.h"
 #include "src/fastertransformer/kernels/cutlass_kernels/int8_gemm/int8_gemm.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 
diff --git a/src/fastertransformer/kernels/cutlass_kernels/moe_gemm/moe_gemm_kernels_template.h b/src/fastertransformer/kernels/cutlass_kernels/moe_gemm/moe_gemm_kernels_template.h
index b198925..fdfc2e7 100644
--- a/src/fastertransformer/kernels/cutlass_kernels/moe_gemm/moe_gemm_kernels_template.h
+++ b/src/fastertransformer/kernels/cutlass_kernels/moe_gemm/moe_gemm_kernels_template.h
@@ -35,6 +35,7 @@
 #include "src/fastertransformer/kernels/cutlass_kernels/cutlass_heuristic.h"
 #include "src/fastertransformer/kernels/cutlass_kernels/moe_gemm/moe_gemm_kernels.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <cuda.h>
 #include <cuda_fp16.h>
 #include <math.h>
diff --git a/src/fastertransformer/kernels/decoding_kernels.cu b/src/fastertransformer/kernels/decoding_kernels.cu
index 89f0d50..8746c39 100644
--- a/src/fastertransformer/kernels/decoding_kernels.cu
+++ b/src/fastertransformer/kernels/decoding_kernels.cu
@@ -18,6 +18,7 @@
 #include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
 #include "src/fastertransformer/utils/cuda_type_utils.cuh"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 
diff --git a/src/fastertransformer/kernels/dequantize_kernels.h b/src/fastertransformer/kernels/dequantize_kernels.h
index 91b634a..9a95deb 100644
--- a/src/fastertransformer/kernels/dequantize_kernels.h
+++ b/src/fastertransformer/kernels/dequantize_kernels.h
@@ -16,6 +16,7 @@
 
 #pragma once
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <assert.h>
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
diff --git a/src/fastertransformer/kernels/disentangled_attention_kernels.cu b/src/fastertransformer/kernels/disentangled_attention_kernels.cu
index b766081..48231f1 100644
--- a/src/fastertransformer/kernels/disentangled_attention_kernels.cu
+++ b/src/fastertransformer/kernels/disentangled_attention_kernels.cu
@@ -17,6 +17,7 @@
 #include "src/fastertransformer/kernels/disentangled_attention_kernels.h"
 #include "src/fastertransformer/utils/cuda_type_utils.cuh"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <assert.h>
 
 namespace fastertransformer {
diff --git a/src/fastertransformer/kernels/gen_relative_pos_bias.cu b/src/fastertransformer/kernels/gen_relative_pos_bias.cu
index f70a8a9..6d468bb 100644
--- a/src/fastertransformer/kernels/gen_relative_pos_bias.cu
+++ b/src/fastertransformer/kernels/gen_relative_pos_bias.cu
@@ -19,6 +19,7 @@
 #include "reduce_kernel_utils.cuh"
 #include "src/fastertransformer/kernels/activation_kernels.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <cstdio>
 
 namespace fastertransformer {
diff --git a/src/fastertransformer/kernels/image_merge_kernels.h b/src/fastertransformer/kernels/image_merge_kernels.h
index d2032ec..9b4a821 100644
--- a/src/fastertransformer/kernels/image_merge_kernels.h
+++ b/src/fastertransformer/kernels/image_merge_kernels.h
@@ -19,6 +19,7 @@
 #include "int8_utils.cuh"
 #include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <assert.h>
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
diff --git a/src/fastertransformer/kernels/image_shift_partition_kernels.h b/src/fastertransformer/kernels/image_shift_partition_kernels.h
index e067d20..7e27f24 100644
--- a/src/fastertransformer/kernels/image_shift_partition_kernels.h
+++ b/src/fastertransformer/kernels/image_shift_partition_kernels.h
@@ -19,6 +19,7 @@
 #include "int8_utils.cuh"
 #include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <assert.h>
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
diff --git a/src/fastertransformer/kernels/layernorm_fp8_kernels.cu b/src/fastertransformer/kernels/layernorm_fp8_kernels.cu
index a350c98..1155bfe 100644
--- a/src/fastertransformer/kernels/layernorm_fp8_kernels.cu
+++ b/src/fastertransformer/kernels/layernorm_fp8_kernels.cu
@@ -18,6 +18,7 @@
 #include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
 #include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <cuda_fp16.h>
 
 namespace fastertransformer {
diff --git a/src/fastertransformer/kernels/layernorm_int8_kernels.h b/src/fastertransformer/kernels/layernorm_int8_kernels.h
index a6d1704..f43de0a 100644
--- a/src/fastertransformer/kernels/layernorm_int8_kernels.h
+++ b/src/fastertransformer/kernels/layernorm_int8_kernels.h
@@ -19,6 +19,7 @@
 #include "int8_utils.cuh"
 #include "src/fastertransformer/kernels/layernorm_kernels.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <assert.h>
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
diff --git a/src/fastertransformer/kernels/layernorm_kernels.h b/src/fastertransformer/kernels/layernorm_kernels.h
index d8ac092..52bd685 100644
--- a/src/fastertransformer/kernels/layernorm_kernels.h
+++ b/src/fastertransformer/kernels/layernorm_kernels.h
@@ -18,6 +18,7 @@
 
 #include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <assert.h>
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
diff --git a/src/fastertransformer/kernels/longformer_kernels.cu b/src/fastertransformer/kernels/longformer_kernels.cu
index 62c5703..6b99156 100644
--- a/src/fastertransformer/kernels/longformer_kernels.cu
+++ b/src/fastertransformer/kernels/longformer_kernels.cu
@@ -26,6 +26,7 @@
 #include "longformer_kernels.h"
 #include "src/fastertransformer/utils/cuda_type_utils.cuh"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 #include <stdio.h>
 #include <float.h>
diff --git a/src/fastertransformer/kernels/moe_kernels.cu b/src/fastertransformer/kernels/moe_kernels.cu
index d884725..1444a18 100644
--- a/src/fastertransformer/kernels/moe_kernels.cu
+++ b/src/fastertransformer/kernels/moe_kernels.cu
@@ -31,6 +31,7 @@
 
 #include "src/fastertransformer/kernels/moe_kernels.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 #ifndef CUDART_VERSION
 #error CUDART_VERSION Undefined!
diff --git a/src/fastertransformer/kernels/moe_kernels.h b/src/fastertransformer/kernels/moe_kernels.h
index c1fa9ec..2cb5b97 100644
--- a/src/fastertransformer/kernels/moe_kernels.h
+++ b/src/fastertransformer/kernels/moe_kernels.h
@@ -16,7 +16,8 @@
 
 #pragma once
 #include "cutlass/gemm/gemm.h"
-#include "src/fastertransformer/kernels/cutlass_kernels/moe_gemm/moe_gemm_kernels.h"
+#include "src/fastertransformer/kernels/ck_kernels/moe_gemm/moe_gemm_kernels.h"
+// #include "src/fastertransformer/kernels/cutlass_kernels/moe_gemm/moe_gemm_kernels.h"
 #include <cuda_runtime_api.h>
 
 namespace fastertransformer {
diff --git a/src/fastertransformer/kernels/normalize_kernels.h b/src/fastertransformer/kernels/normalize_kernels.h
index d821f45..da1cd2d 100644
--- a/src/fastertransformer/kernels/normalize_kernels.h
+++ b/src/fastertransformer/kernels/normalize_kernels.h
@@ -18,6 +18,7 @@
 
 #include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <assert.h>
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
diff --git a/src/fastertransformer/kernels/online_softmax_beamsearch_kernels.cu b/src/fastertransformer/kernels/online_softmax_beamsearch_kernels.cu
index 67d382b..5f3e2ba 100644
--- a/src/fastertransformer/kernels/online_softmax_beamsearch_kernels.cu
+++ b/src/fastertransformer/kernels/online_softmax_beamsearch_kernels.cu
@@ -25,6 +25,7 @@
 #include "src/fastertransformer/kernels/online_softmax_beamsearch_kernels.h"
 #include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 
diff --git a/src/fastertransformer/kernels/sampling_penalty_kernels.h b/src/fastertransformer/kernels/sampling_penalty_kernels.h
index 1ceb9c8..e964922 100644
--- a/src/fastertransformer/kernels/sampling_penalty_kernels.h
+++ b/src/fastertransformer/kernels/sampling_penalty_kernels.h
@@ -19,6 +19,7 @@
 
 #include "src/fastertransformer/kernels/penalty_types.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 
diff --git a/src/fastertransformer/kernels/sampling_topp_kernels.cu b/src/fastertransformer/kernels/sampling_topp_kernels.cu
index 78122dc..61b9c3c 100644
--- a/src/fastertransformer/kernels/sampling_topp_kernels.cu
+++ b/src/fastertransformer/kernels/sampling_topp_kernels.cu
@@ -25,6 +25,7 @@
 #include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
 #include "src/fastertransformer/kernels/sampling_topp_kernels.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 constexpr int   ENABLE_SINGLE_PASS_TOP_P = 0;
 constexpr float SINGLE_PASS_THRESHOLD    = 0.9;
diff --git a/src/fastertransformer/kernels/softmax_int8_kernels.cu b/src/fastertransformer/kernels/softmax_int8_kernels.cu
index 2dbf49b..d9ec64a 100644
--- a/src/fastertransformer/kernels/softmax_int8_kernels.cu
+++ b/src/fastertransformer/kernels/softmax_int8_kernels.cu
@@ -18,6 +18,7 @@
 #include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
 #include "src/fastertransformer/kernels/softmax_int8_kernels.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 
diff --git a/src/fastertransformer/kernels/stop_criteria_kernels.cu b/src/fastertransformer/kernels/stop_criteria_kernels.cu
index 6579ec6..29a985c 100644
--- a/src/fastertransformer/kernels/stop_criteria_kernels.cu
+++ b/src/fastertransformer/kernels/stop_criteria_kernels.cu
@@ -17,6 +17,7 @@
 #include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
 #include "src/fastertransformer/kernels/stop_criteria_kernels.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include "src/fastertransformer/utils/memory_utils.h"
 
 namespace fastertransformer {
diff --git a/src/fastertransformer/kernels/unfused_attention_fp8_kernels.cu b/src/fastertransformer/kernels/unfused_attention_fp8_kernels.cu
index ab46490..d141b93 100644
--- a/src/fastertransformer/kernels/unfused_attention_fp8_kernels.cu
+++ b/src/fastertransformer/kernels/unfused_attention_fp8_kernels.cu
@@ -19,6 +19,7 @@
 #include "src/fastertransformer/kernels/unfused_attention_fp8_kernels.h"
 #include "src/fastertransformer/utils/cuda_bf16_fallbacks.cuh"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include "src/fastertransformer/utils/memory_utils.h"
 
 namespace fastertransformer {
diff --git a/src/fastertransformer/kernels/unfused_attention_kernels.cu b/src/fastertransformer/kernels/unfused_attention_kernels.cu
index d0fb0a1..5169d56 100644
--- a/src/fastertransformer/kernels/unfused_attention_kernels.cu
+++ b/src/fastertransformer/kernels/unfused_attention_kernels.cu
@@ -20,6 +20,7 @@
 #include "src/fastertransformer/kernels/unfused_attention_kernels.h"
 #include "src/fastertransformer/utils/cuda_type_utils.cuh"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 
diff --git a/src/fastertransformer/kernels/vit_kernels.cu b/src/fastertransformer/kernels/vit_kernels.cu
index 0d67c5c..f7b5cdb 100644
--- a/src/fastertransformer/kernels/vit_kernels.cu
+++ b/src/fastertransformer/kernels/vit_kernels.cu
@@ -16,6 +16,7 @@
 
 #include "src/fastertransformer/kernels/vit_kernels.h"
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 namespace fastertransformer {
 
 template<typename T>
diff --git a/src/fastertransformer/kernels/xlnet_preprocess_kernels.h b/src/fastertransformer/kernels/xlnet_preprocess_kernels.h
index 9ab49a6..9f6e422 100644
--- a/src/fastertransformer/kernels/xlnet_preprocess_kernels.h
+++ b/src/fastertransformer/kernels/xlnet_preprocess_kernels.h
@@ -16,6 +16,7 @@
 
 #pragma once
 #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
 
diff --git a/src/fastertransformer/layers/CMakeLists.txt b/src/fastertransformer/layers/CMakeLists.txt
index aa7c31c..1bc7bf8 100644
--- a/src/fastertransformer/layers/CMakeLists.txt
+++ b/src/fastertransformer/layers/CMakeLists.txt
@@ -27,7 +27,7 @@ add_library(FfnLayer STATIC FfnLayer.cc)
 set_property(TARGET FfnLayer PROPERTY POSITION_INDEPENDENT_CODE  ON)
 set_property(TARGET FfnLayer PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS  ON)
 # target_link_libraries(FfnLayer PUBLIC -lcublas -lcudart cublasMMWrapper activation_kernels transpose_int8_kernels memory_utils matrix_vector_multiplication tensor moe_kernels fpA_intB_gemm int8_gemm nvtx_utils)
-target_link_libraries(FfnLayer PUBLIC -lcublas -lcudart cublasMMWrapper activation_kernels transpose_int8_kernels memory_utils matrix_vector_multiplication tensor)
+target_link_libraries(FfnLayer PUBLIC -lcublas -lcudart cublasMMWrapper activation_kernels transpose_int8_kernels memory_utils matrix_vector_multiplication tensor fpA_intB_gemm int8_gemm)
 
 # add_library(FfnLayerINT8 STATIC FfnLayerINT8.cc)
 # set_property(TARGET FfnLayerINT8 PROPERTY POSITION_INDEPENDENT_CODE  ON)
diff --git a/src/fastertransformer/layers/FfnLayer.cc b/src/fastertransformer/layers/FfnLayer.cc
index dddab73..a058401 100644
--- a/src/fastertransformer/layers/FfnLayer.cc
+++ b/src/fastertransformer/layers/FfnLayer.cc
@@ -195,74 +195,76 @@ void FfnLayer<T>::forward(TensorMap* output_tensors, TensorMap* input_tensors, c
 #endif
     }
     else {
-        // if (int8_mode_ == 1) {
-        //     FT_CHECK_WITH_INFO(weight_only_int8_fc_runner_.get() != NULL, "weight only runner was not initialized.");
-        //     FT_CHECK(ffn_weights->intermediate_weight.int8_kernel != NULL
-        //              && ffn_weights->intermediate_weight.weight_only_quant_scale != NULL);
-
-        //     if (ia3_tasks == nullptr && !use_gated_activation) {
-        //         // launch fused GEMM + activation
-        //         weight_only_int8_fc_runner_->gemm_bias_act(
-        //             input_tensor,
-        //             reinterpret_cast<const uint8_t*>(ffn_weights->intermediate_weight.int8_kernel),
-        //             ffn_weights->intermediate_weight.weight_only_quant_scale,
-        //             ffn_weights->intermediate_weight.bias,
-        //             inter_buf_,
-        //             m,
-        //             inter_size_,
-        //             hidden_units_,
-        //             activation_type,
-        //             mixed_gemm_workspace_,
-        //             mixed_gemm_ws_bytes_,
-        //             stream_);
-        //     }
-        //     else {
-        //         // Otherwise, let FT handle activation
-        //         weight_only_int8_fc_runner_->gemm(
-        //             input_tensor,
-        //             reinterpret_cast<const uint8_t*>(ffn_weights->intermediate_weight.int8_kernel),
-        //             ffn_weights->intermediate_weight.weight_only_quant_scale,
-        //             inter_buf_,
-        //             m,
-        //             inter_size_,
-        //             hidden_units_,
-        //             mixed_gemm_workspace_,
-        //             mixed_gemm_ws_bytes_,
-        //             stream_);
-
-        //         if (use_gated_activation) {
-        //             FT_CHECK(ffn_weights->intermediate_weight2.int8_kernel != NULL
-        //                      && ffn_weights->intermediate_weight2.weight_only_quant_scale != NULL);
-
-        //             weight_only_int8_fc_runner_->gemm(
-        //                 input_tensor,
-        //                 reinterpret_cast<const uint8_t*>(ffn_weights->intermediate_weight2.int8_kernel),
-        //                 ffn_weights->intermediate_weight2.weight_only_quant_scale,
-        //                 inter_buf_2_,
-        //                 m,
-        //                 inter_size_,
-        //                 hidden_units_,
-        //                 mixed_gemm_workspace_,
-        //                 mixed_gemm_ws_bytes_,
-        //                 stream_);
-        //         }
-        //     }
-        // }
-        // else if (int8_mode_ == 2) {
-        //     FT_CHECK(!use_gated_activation);
-        //     cublas_wrapper_->Int8Gemm(inter_size_,
-        //                               m,
-        //                               hidden_units_,
-        //                               ffn_weights->intermediate_weight.int8_kernel,
-        //                               hidden_units_,
-        //                               input_tensors->getPtr<int8_t>("ffn_input"),
-        //                               hidden_units_,
-        //                               reinterpret_cast<int8_t*>(inter_buf_),
-        //                               inter_size_,
-        //                               ffn_weights->intermediate_weight.scale_inter);
-        // }
-        // else {
-        {
+        if (int8_mode_ == 1) {
+            FT_CHECK_WITH_INFO(weight_only_int8_fc_runner_.get() != NULL, "weight only runner was not initialized.");
+            FT_CHECK(ffn_weights->intermediate_weight.int8_kernel != NULL
+                     && ffn_weights->intermediate_weight.weight_only_quant_scale != NULL);
+
+            if (ia3_tasks == nullptr && !use_gated_activation) {
+                // launch fused GEMM + activation
+                weight_only_int8_fc_runner_->gemm_bias_act(
+                    input_tensor,
+                    reinterpret_cast<const uint8_t*>(ffn_weights->intermediate_weight.int8_kernel),
+                    ffn_weights->intermediate_weight.weight_only_quant_scale,
+                    ffn_weights->intermediate_weight.bias,
+                    inter_buf_,
+                    m,
+                    inter_size_,
+                    hidden_units_,
+                    activation_type,
+                    mixed_gemm_workspace_,
+                    mixed_gemm_ws_bytes_
+                    );
+                    // stream_);
+            }
+            else {
+                // Otherwise, let FT handle activation
+                weight_only_int8_fc_runner_->gemm(
+                    input_tensor,
+                    reinterpret_cast<const uint8_t*>(ffn_weights->intermediate_weight.int8_kernel),
+                    ffn_weights->intermediate_weight.weight_only_quant_scale,
+                    inter_buf_,
+                    m,
+                    inter_size_,
+                    hidden_units_,
+                    mixed_gemm_workspace_,
+                    mixed_gemm_ws_bytes_
+                    );
+                    // stream_);
+
+                if (use_gated_activation) {
+                    FT_CHECK(ffn_weights->intermediate_weight2.int8_kernel != NULL
+                             && ffn_weights->intermediate_weight2.weight_only_quant_scale != NULL);
+
+                    weight_only_int8_fc_runner_->gemm(
+                        input_tensor,
+                        reinterpret_cast<const uint8_t*>(ffn_weights->intermediate_weight2.int8_kernel),
+                        ffn_weights->intermediate_weight2.weight_only_quant_scale,
+                        inter_buf_2_,
+                        m,
+                        inter_size_,
+                        hidden_units_,
+                        mixed_gemm_workspace_,
+                        mixed_gemm_ws_bytes_
+                        );
+                        // stream_);
+                }
+            }
+        }
+        else if (int8_mode_ == 2) {
+            FT_CHECK(!use_gated_activation);
+            cublas_wrapper_->Int8Gemm(inter_size_,
+                                      m,
+                                      hidden_units_,
+                                      ffn_weights->intermediate_weight.int8_kernel,
+                                      hidden_units_,
+                                      input_tensors->getPtr<int8_t>("ffn_input"),
+                                      hidden_units_,
+                                      reinterpret_cast<int8_t*>(inter_buf_),
+                                      inter_size_,
+                                      ffn_weights->intermediate_weight.scale_inter);
+        }
+        else {
             cublas_wrapper_->Gemm(CUBLAS_OP_N,
                                   CUBLAS_OP_N,
                                   inter_size_,
@@ -328,37 +330,38 @@ void FfnLayer<T>::forward(TensorMap* output_tensors, TensorMap* input_tensors, c
 #endif
     }
     else {
-        // if (int8_mode_ == 1) {
-        //     FT_CHECK_WITH_INFO(weight_only_int8_fc_runner_.get() != NULL, "weight only runner was not initialized.");
-        //     FT_CHECK(ffn_weights->output_weight.int8_kernel != NULL
-        //              && ffn_weights->output_weight.weight_only_quant_scale != NULL);
-        //     weight_only_int8_fc_runner_->gemm(inter_buf_,
-        //                                       reinterpret_cast<const uint8_t*>(ffn_weights->output_weight.int8_kernel),
-        //                                       ffn_weights->output_weight.weight_only_quant_scale,
-        //                                       output_tensor,
-        //                                       m,
-        //                                       hidden_units_,
-        //                                       inter_size_,
-        //                                       mixed_gemm_workspace_,
-        //                                       mixed_gemm_ws_bytes_,
-        //                                       stream_);
-        // }
-        // else if (int8_mode_ == 2) {
-        //     int8_fc_runner_->gemm(reinterpret_cast<int8_t*>(inter_buf_),
-        //                           ffn_weights->output_weight.int8_kernel,
-        //                           QuantMode::PerTensorQuant,
-        //                           ffn_weights->output_weight.scale_inter,
-        //                           ffn_weights->output_weight.scale_out,
-        //                           output_tensors->getPtr<T>("ffn_output"),
-        //                           m,
-        //                           hidden_units_,
-        //                           inter_size_,
-        //                           nullptr,
-        //                           0,
-        //                           stream_);
-        // }
-        // else {
-        {
+        if (int8_mode_ == 1) {
+            FT_CHECK_WITH_INFO(weight_only_int8_fc_runner_.get() != NULL, "weight only runner was not initialized.");
+            FT_CHECK(ffn_weights->output_weight.int8_kernel != NULL
+                     && ffn_weights->output_weight.weight_only_quant_scale != NULL);
+            weight_only_int8_fc_runner_->gemm(inter_buf_,
+                                              reinterpret_cast<const uint8_t*>(ffn_weights->output_weight.int8_kernel),
+                                              ffn_weights->output_weight.weight_only_quant_scale,
+                                              output_tensor,
+                                              m,
+                                              hidden_units_,
+                                              inter_size_,
+                                              mixed_gemm_workspace_,
+                                              mixed_gemm_ws_bytes_
+                                              );
+                                            //   stream_);
+        }
+        else if (int8_mode_ == 2) {
+            int8_fc_runner_->gemm(reinterpret_cast<int8_t*>(inter_buf_),
+                                  ffn_weights->output_weight.int8_kernel,
+                                  QuantMode::PerTensorQuant,
+                                  ffn_weights->output_weight.scale_inter,
+                                  ffn_weights->output_weight.scale_out,
+                                  output_tensors->getPtr<T>("ffn_output"),
+                                  m,
+                                  hidden_units_,
+                                  inter_size_,
+                                  nullptr,
+                                  0
+                                  );
+                                //   stream_);
+        }
+        else {
             cublas_wrapper_->Gemm(CUBLAS_OP_N,
                                   CUBLAS_OP_N,
                                   hidden_units_,
@@ -404,8 +407,8 @@ FfnLayer<T>::FfnLayer(size_t           max_batch_size,
     max_inter_size_(inter_size),
     inter_size_(inter_size),
     int8_mode_(int8_mode),
-    use_gated_activation_(use_gated_activation)
-    // int8_fc_runner_(int8_mode == 2 ? std::make_shared<CutlassInt8GemmRunner<T>>() : nullptr)
+    use_gated_activation_(use_gated_activation),
+    int8_fc_runner_(int8_mode == 2 ? std::make_shared<CutlassInt8GemmRunner<T>>() : nullptr)
 {
     FT_LOG_DEBUG(__PRETTY_FUNCTION__);
     // if (int8_mode_ == 0) {
@@ -434,11 +437,11 @@ FfnLayer<T>::FfnLayer(FfnLayer<T> const& ffn_layer):
     max_inter_size_(ffn_layer.max_inter_size_),
     inter_size_(ffn_layer.inter_size_),
     int8_mode_(ffn_layer.int8_mode_),
-    use_gated_activation_(ffn_layer.use_gated_activation_)
+    use_gated_activation_(ffn_layer.use_gated_activation_),
     // moe_fc_runner_(ffn_layer.moe_fc_runner_),
     // moe_int8_weight_only_fc_runner_(ffn_layer.moe_int8_weight_only_fc_runner_),
-    // weight_only_int8_fc_runner_(ffn_layer.weight_only_int8_fc_runner_),
-    // int8_fc_runner_(ffn_layer.int8_fc_runner_)
+    weight_only_int8_fc_runner_(ffn_layer.weight_only_int8_fc_runner_),
+    int8_fc_runner_(ffn_layer.int8_fc_runner_)
 {
     FT_LOG_DEBUG(__PRETTY_FUNCTION__);
 }
@@ -478,8 +481,7 @@ void FfnLayer<T>::allocateBuffer(size_t token_num, int moe_k, bool use_moe)
     //     }
 
     //     moe_fc_workspace_ = (char*)allocator_->reMalloc(moe_fc_workspace_, sizeof(char) * ws_size_moe, false);
-    // }
-    // else {
+    // }else {
     //     const auto type_size = int8_mode_ == 2 ? sizeof(int8_t) : sizeof(T);
     //     inter_buf_           = (T*)allocator_->reMalloc(inter_buf_, type_size * token_num * max_inter_size_, false);
     //     if (use_gated_activation_) {
@@ -500,13 +502,26 @@ void FfnLayer<T>::allocateBuffer(size_t token_num, int moe_k, bool use_moe)
     //         int8_gemm_workspace_ = (char*)allocator_->reMalloc(int8_gemm_workspace_, int8_gemm_ws_bytes_, false);
     //     }
     // }
-    {
+    if (use_moe == false) {
         const auto type_size = int8_mode_ == 2 ? sizeof(int8_t) : sizeof(T);
         inter_buf_           = (T*)allocator_->reMalloc(inter_buf_, type_size * token_num * max_inter_size_, false);
         if (use_gated_activation_) {
             inter_buf_2_ = (T*)allocator_->reMalloc(inter_buf_2_, sizeof(T) * token_num * max_inter_size_, false);
         }
 
+        if (int8_mode_ == 1) {
+            FT_CHECK_WITH_INFO(weight_only_int8_fc_runner_.get() != NULL, "weight only runner was not initialized.");
+            // We use max_size for n and k since we reuse buffers for both FCs and want to allocate the max
+            // possible memory that would be required by any of the individual gemms.
+            const int max_size    = std::max(hidden_units_, inter_size_);
+            mixed_gemm_ws_bytes_  = weight_only_int8_fc_runner_->getWorkspaceSize(token_num, max_size, max_size);
+            mixed_gemm_workspace_ = (char*)allocator_->reMalloc(mixed_gemm_workspace_, mixed_gemm_ws_bytes_, false);
+        }
+        else if (int8_mode_ == 2) {
+            const int max_size   = std::max(hidden_units_, inter_size_);
+            int8_gemm_ws_bytes_  = int8_fc_runner_->getWorkspaceSize(token_num, max_size, max_size);
+            int8_gemm_workspace_ = (char*)allocator_->reMalloc(int8_gemm_workspace_, int8_gemm_ws_bytes_, false);
+        }
     }
 
     is_allocate_buffer_ = true;
diff --git a/src/fastertransformer/layers/FfnLayer.h b/src/fastertransformer/layers/FfnLayer.h
index 60c2e21..9ed7555 100644
--- a/src/fastertransformer/layers/FfnLayer.h
+++ b/src/fastertransformer/layers/FfnLayer.h
@@ -19,8 +19,10 @@
 #include "src/fastertransformer/kernels/activation_kernels.h"
 // #include "src/fastertransformer/kernels/cutlass_kernels/fpA_intB_gemm/fpA_intB_gemm.h"
 // #include "src/fastertransformer/kernels/cutlass_kernels/int8_gemm/int8_gemm.h"
-#include "src/fastertransformer/kernels/matrix_vector_multiplication.h"
 // #include "src/fastertransformer/kernels/moe_kernels.h"
+#include "src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm.h"
+#include "src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm.h"
+#include "src/fastertransformer/kernels/matrix_vector_multiplication.h"
 #include "src/fastertransformer/layers/BaseLayer.h"
 #include "src/fastertransformer/layers/FfnWeight.h"
 #include "src/fastertransformer/utils/activation_types.h"
@@ -51,8 +53,8 @@ private:
     // std::shared_ptr<CutlassMoeFCRunner<T, T>>       moe_fc_runner_;
     // std::shared_ptr<CutlassMoeFCRunner<T, uint8_t>> moe_int8_weight_only_fc_runner_;
 
-    // std::shared_ptr<CutlassFpAIntBGemmRunner<T, uint8_t>> weight_only_int8_fc_runner_;
-    // std::shared_ptr<CutlassInt8GemmRunner<T>>             int8_fc_runner_;
+    std::shared_ptr<CutlassFpAIntBGemmRunner<T, uint8_t>> weight_only_int8_fc_runner_;
+    std::shared_ptr<CutlassInt8GemmRunner<T>>             int8_fc_runner_;
 
     void allocateBuffer() override;
     void freeBuffer() override;
diff --git a/src/fastertransformer/layers/attention_layers/CMakeLists.txt b/src/fastertransformer/layers/attention_layers/CMakeLists.txt
index 2fca973..b31841e 100644
--- a/src/fastertransformer/layers/attention_layers/CMakeLists.txt
+++ b/src/fastertransformer/layers/attention_layers/CMakeLists.txt
@@ -38,13 +38,13 @@ add_library(DecoderSelfAttentionLayer STATIC DecoderSelfAttentionLayer.cc)
 set_property(TARGET DecoderSelfAttentionLayer PROPERTY POSITION_INDEPENDENT_CODE  ON)
 set_property(TARGET DecoderSelfAttentionLayer PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS  ON)
 # target_link_libraries(DecoderSelfAttentionLayer PUBLIC -lcublas -lcudart cublasMMWrapper memory_utils decoder_masked_multihead_attention fpA_intB_gemm int8_gemm tensor nvtx_utils)
-target_link_libraries(DecoderSelfAttentionLayer PUBLIC -lcublas -lcudart cublasMMWrapper memory_utils decoder_masked_multihead_attention tensor)
+target_link_libraries(DecoderSelfAttentionLayer PUBLIC -lcublas -lcudart cublasMMWrapper memory_utils decoder_masked_multihead_attention tensor fpA_intB_gemm int8_gemm)
 
 add_library(GptContextAttentionLayer STATIC GptContextAttentionLayer.cc)
 set_property(TARGET GptContextAttentionLayer PROPERTY POSITION_INDEPENDENT_CODE  ON)
 set_property(TARGET GptContextAttentionLayer PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS  ON)
 # target_link_libraries(GptContextAttentionLayer PUBLIC -lcublas -lcudart cublasMMWrapper memory_utils unfused_attention_kernels trt_fused_multi_head_attention fpA_intB_gemm int8_gemm nvtx_utils)
-target_link_libraries(GptContextAttentionLayer PUBLIC -lcublas -lcudart cublasMMWrapper memory_utils unfused_attention_kernels trt_fused_multi_head_attention)
+target_link_libraries(GptContextAttentionLayer PUBLIC -lcublas -lcudart cublasMMWrapper memory_utils unfused_attention_kernels trt_fused_multi_head_attention fpA_intB_gemm int8_gemm)
 
 add_library(DisentangledAttentionLayer STATIC DisentangledAttentionLayer.cc)
 set_property(TARGET DisentangledAttentionLayer PROPERTY POSITION_INDEPENDENT_CODE  ON)
diff --git a/src/fastertransformer/layers/attention_layers/DecoderSelfAttentionLayer.cc b/src/fastertransformer/layers/attention_layers/DecoderSelfAttentionLayer.cc
index 0d70aa2..7e6e2df 100644
--- a/src/fastertransformer/layers/attention_layers/DecoderSelfAttentionLayer.cc
+++ b/src/fastertransformer/layers/attention_layers/DecoderSelfAttentionLayer.cc
@@ -208,12 +208,12 @@ void DecoderSelfAttentionLayer<T>::allocateBuffer(size_t batch_size)
         // We use max_size for n and k since we reuse buffers for both FCs and want to allocate the max
         // possible memory that would be required by any of the individual gemms.
         const int max_size    = std::max(d_model_, 3 * local_hidden_units_);
-        // mixed_gemm_ws_bytes_  = weight_only_int8_fc_runner_->getWorkspaceSize(batch_size, max_size, max_size);
+        mixed_gemm_ws_bytes_  = weight_only_int8_fc_runner_->getWorkspaceSize(batch_size, max_size, max_size);
         mixed_gemm_workspace_ = (char*)allocator_->reMalloc(mixed_gemm_workspace_, mixed_gemm_ws_bytes_, false);
     }
     else if (int8_mode_ == 2) {
         const int max_size   = std::max(d_model_, 3 * local_hidden_units_);
-        // int8_gemm_ws_bytes_  = int8_fc_runner_->getWorkspaceSize(batch_size, max_size, max_size);
+        int8_gemm_ws_bytes_  = int8_fc_runner_->getWorkspaceSize(batch_size, max_size, max_size);
         int8_gemm_workspace_ = (char*)allocator_->reMalloc(int8_gemm_workspace_, int8_gemm_ws_bytes_, false);
     }
 
@@ -274,16 +274,16 @@ DecoderSelfAttentionLayer<T>::DecoderSelfAttentionLayer(size_t           max_bat
     neox_rotary_style_(neox_rotary_style),
     d_model_(d_model),
     q_scaling_(q_scaling),
-    // int8_fc_runner_(int8_mode == 2 ? std::make_shared<CutlassInt8GemmRunner<T>>() : nullptr),
+    int8_fc_runner_(int8_mode == 2 ? std::make_shared<CutlassInt8GemmRunner<T>>() : nullptr),
     int8_mode_(int8_mode)
 {
     FT_CHECK(size_per_head_ == 32 || size_per_head_ == 48 || size_per_head_ == 64 || size_per_head_ == 80
              || size_per_head_ == 96 || size_per_head_ == 112 || size_per_head_ == 128 || size_per_head_ == 144
              || size_per_head_ == 160 || size_per_head_ == 192 || size_per_head_ == 224 || size_per_head_ == 256);
-    // if (int8_mode_ == 1) {
-    //     FT_CHECK_WITH_INFO(!(std::is_same<T, float>::value), "Weight only quant not supported for fp32.");
-    //     weight_only_int8_fc_runner_ = std::make_shared<CutlassFpAIntBGemmRunner<T, uint8_t>>();
-    // }
+    if (int8_mode_ == 1) {
+        FT_CHECK_WITH_INFO(!(std::is_same<T, float>::value), "Weight only quant not supported for fp32.");
+        weight_only_int8_fc_runner_ = std::make_shared<CutlassFpAIntBGemmRunner<T, uint8_t>>();
+    }
 }
 
 template<typename T>
@@ -529,52 +529,53 @@ void DecoderSelfAttentionLayer<T>::forward(TensorMap*                output_tens
 #endif
     }
     else {
-        // if (int8_mode_ == 1) {
-        //     FT_CHECK(weight_only_int8_fc_runner_.get() != NULL && attention_weights->query_weight.int8_kernel != NULL
-        //              && attention_weights->query_weight.weight_only_quant_scale != NULL);
-
-        //     weight_only_int8_fc_runner_->gemm(
-        //         attention_input,
-        //         reinterpret_cast<const uint8_t*>(attention_weights->query_weight.int8_kernel),
-        //         attention_weights->query_weight.weight_only_quant_scale,
-        //         qkv_buf_,
-        //         batch_size,
-        //         3 * local_hidden_units_,
-        //         d_model_,
-        //         mixed_gemm_workspace_,
-        //         mixed_gemm_ws_bytes_,
-        //         stream_);
-        // }
-        // else if (int8_mode_ == 2) {
-        //     // Here, we set per_column_scaling to be true because q, k, v may
-        //     // use different scales. So, we pass a pointer with shape [3, local_hidden_units_] like
-        //     // [s_q, s_q, ..., s_q, s_k, s_k, ..., s_k, s_v, s_v, ..., s_v],
-        //     // where s_q are scales of q, s_k are scales of k and s_v are scales of v.
-        //     cublas_wrapper_->Int8Gemm(3 * local_hidden_units_,
-        //                               batch_size,
-        //                               d_model_,
-        //                               attention_weights->query_weight.int8_kernel,
-        //                               d_model_,
-        //                               input_tensors->getPtr<int8_t>("input_query"),
-        //                               d_model_,
-        //                               reinterpret_cast<int8_t*>(qkv_buf_),
-        //                               3 * local_hidden_units_,
-        //                               attention_weights->query_weight.scale_inter,
-        //                               true);
-        // }
-        // else {
-        //     cublas_wrapper_->Gemm(CUBLAS_OP_N,
-        //                           CUBLAS_OP_N,
-        //                           3 * local_hidden_units_,  // n
-        //                           batch_size,
-        //                           d_model_,  // k
-        //                           attention_weights->query_weight.kernel,
-        //                           3 * local_hidden_units_,  // n
-        //                           attention_input,
-        //                           d_model_,  // k
-        //                           qkv_buf_,
-        //                           3 * local_hidden_units_ /* n */);
-        // }
+        if (int8_mode_ == 1) {
+            FT_CHECK(weight_only_int8_fc_runner_.get() != NULL && attention_weights->query_weight.int8_kernel != NULL
+                     && attention_weights->query_weight.weight_only_quant_scale != NULL);
+
+            weight_only_int8_fc_runner_->gemm(
+                attention_input,
+                reinterpret_cast<const uint8_t*>(attention_weights->query_weight.int8_kernel),
+                attention_weights->query_weight.weight_only_quant_scale,
+                qkv_buf_,
+                batch_size,
+                3 * local_hidden_units_,
+                d_model_,
+                mixed_gemm_workspace_,
+                mixed_gemm_ws_bytes_
+                );
+                // stream_);
+        }
+        else if (int8_mode_ == 2) {
+            // Here, we set per_column_scaling to be true because q, k, v may
+            // use different scales. So, we pass a pointer with shape [3, local_hidden_units_] like
+            // [s_q, s_q, ..., s_q, s_k, s_k, ..., s_k, s_v, s_v, ..., s_v],
+            // where s_q are scales of q, s_k are scales of k and s_v are scales of v.
+            cublas_wrapper_->Int8Gemm(3 * local_hidden_units_,
+                                      batch_size,
+                                      d_model_,
+                                      attention_weights->query_weight.int8_kernel,
+                                      d_model_,
+                                      input_tensors->getPtr<int8_t>("input_query"),
+                                      d_model_,
+                                      reinterpret_cast<int8_t*>(qkv_buf_),
+                                      3 * local_hidden_units_,
+                                      attention_weights->query_weight.scale_inter,
+                                      true);
+        }
+        else {
+            cublas_wrapper_->Gemm(CUBLAS_OP_N,
+                                  CUBLAS_OP_N,
+                                  3 * local_hidden_units_,  // n
+                                  batch_size,
+                                  d_model_,  // k
+                                  attention_weights->query_weight.kernel,
+                                  3 * local_hidden_units_,  // n
+                                  attention_input,
+                                  d_model_,  // k
+                                  qkv_buf_,
+                                  3 * local_hidden_units_ /* n */);
+        }
         {
             cublas_wrapper_->Gemm(CUBLAS_OP_N,
                                   CUBLAS_OP_N,
@@ -645,50 +646,52 @@ void DecoderSelfAttentionLayer<T>::forward(TensorMap*                output_tens
 #endif
     }
     else {
-        // if (int8_mode_ == 1) {
-        //     FT_CHECK(weight_only_int8_fc_runner_.get() != NULL
-        //              && attention_weights->attention_output_weight.int8_kernel != NULL
-        //              && attention_weights->attention_output_weight.weight_only_quant_scale != NULL);
-
-        //     weight_only_int8_fc_runner_->gemm(
-        //         context_buf_,
-        //         reinterpret_cast<const uint8_t*>(attention_weights->attention_output_weight.int8_kernel),
-        //         attention_weights->attention_output_weight.weight_only_quant_scale,
-        //         attention_out,
-        //         batch_size,
-        //         d_model_,
-        //         local_hidden_units_,
-        //         mixed_gemm_workspace_,
-        //         mixed_gemm_ws_bytes_,
-        //         stream_);
-        // }
-        // else if (int8_mode_ == 2) {
-        //     int8_fc_runner_->gemm(reinterpret_cast<int8_t*>(context_buf_),
-        //                           attention_weights->attention_output_weight.int8_kernel,
-        //                           QuantMode::PerTensorQuant,
-        //                           attention_weights->attention_output_weight.scale_inter,
-        //                           attention_weights->attention_output_weight.scale_out,
-        //                           output_tensors->getPtr<T>("hidden_features"),
-        //                           batch_size,
-        //                           d_model_,
-        //                           local_hidden_units_,
-        //                           nullptr,
-        //                           0,
-        //                           stream_);
-        // }
-        // else {
-        //     cublas_wrapper_->Gemm(CUBLAS_OP_N,
-        //                           CUBLAS_OP_N,
-        //                           d_model_,  // n
-        //                           batch_size,
-        //                           local_hidden_units_,  // k
-        //                           attention_weights->attention_output_weight.kernel,
-        //                           d_model_,  // n
-        //                           context_buf_,
-        //                           local_hidden_units_,  // k
-        //                           attention_out,
-        //                           d_model_ /* n */);
-        // }
+        if (int8_mode_ == 1) {
+            FT_CHECK(weight_only_int8_fc_runner_.get() != NULL
+                     && attention_weights->attention_output_weight.int8_kernel != NULL
+                     && attention_weights->attention_output_weight.weight_only_quant_scale != NULL);
+
+            weight_only_int8_fc_runner_->gemm(
+                context_buf_,
+                reinterpret_cast<const uint8_t*>(attention_weights->attention_output_weight.int8_kernel),
+                attention_weights->attention_output_weight.weight_only_quant_scale,
+                attention_out,
+                batch_size,
+                d_model_,
+                local_hidden_units_,
+                mixed_gemm_workspace_,
+                mixed_gemm_ws_bytes_
+                );
+                // stream_);
+        }
+        else if (int8_mode_ == 2) {
+            int8_fc_runner_->gemm(reinterpret_cast<int8_t*>(context_buf_),
+                                  attention_weights->attention_output_weight.int8_kernel,
+                                  QuantMode::PerTensorQuant,
+                                  attention_weights->attention_output_weight.scale_inter,
+                                  attention_weights->attention_output_weight.scale_out,
+                                  output_tensors->getPtr<T>("hidden_features"),
+                                  batch_size,
+                                  d_model_,
+                                  local_hidden_units_,
+                                  nullptr,
+                                  0
+                                  );
+                                //   stream_);
+        }
+        else {
+            cublas_wrapper_->Gemm(CUBLAS_OP_N,
+                                  CUBLAS_OP_N,
+                                  d_model_,  // n
+                                  batch_size,
+                                  local_hidden_units_,  // k
+                                  attention_weights->attention_output_weight.kernel,
+                                  d_model_,  // n
+                                  context_buf_,
+                                  local_hidden_units_,  // k
+                                  attention_out,
+                                  d_model_ /* n */);
+        }
         {
             cublas_wrapper_->Gemm(CUBLAS_OP_N,
                                   CUBLAS_OP_N,
diff --git a/src/fastertransformer/layers/attention_layers/DecoderSelfAttentionLayer.h b/src/fastertransformer/layers/attention_layers/DecoderSelfAttentionLayer.h
index 11211f1..43f53ae 100644
--- a/src/fastertransformer/layers/attention_layers/DecoderSelfAttentionLayer.h
+++ b/src/fastertransformer/layers/attention_layers/DecoderSelfAttentionLayer.h
@@ -18,6 +18,8 @@
 
 // #include "src/fastertransformer/kernels/cutlass_kernels/fpA_intB_gemm/fpA_intB_gemm.h"
 // #include "src/fastertransformer/kernels/cutlass_kernels/int8_gemm/int8_gemm.h"
+#include "src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm.h"
+#include "src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm.h"
 #include "src/fastertransformer/kernels/matrix_vector_multiplication.h"
 #include "src/fastertransformer/layers/attention_layers/BaseAttentionLayer.h"
 
@@ -40,8 +42,8 @@ private:
     const size_t rotary_embedding_dim_;
     const bool   neox_rotary_style_;
 
-    // std::shared_ptr<CutlassFpAIntBGemmRunner<T, uint8_t>> weight_only_int8_fc_runner_;
-    // std::shared_ptr<CutlassInt8GemmRunner<T>>             int8_fc_runner_;
+    std::shared_ptr<CutlassFpAIntBGemmRunner<T, uint8_t>> weight_only_int8_fc_runner_;
+    std::shared_ptr<CutlassInt8GemmRunner<T>>             int8_fc_runner_;
 
     void allocateBuffer() override;
     void freeBuffer() override;
diff --git a/src/fastertransformer/layers/attention_layers/GptContextAttentionLayer.cc b/src/fastertransformer/layers/attention_layers/GptContextAttentionLayer.cc
index 0505659..2de7b90 100644
--- a/src/fastertransformer/layers/attention_layers/GptContextAttentionLayer.cc
+++ b/src/fastertransformer/layers/attention_layers/GptContextAttentionLayer.cc
@@ -98,47 +98,48 @@ void GptContextAttentionLayer<T>::forward(TensorMap*                output_tenso
                                 qkv_buf_);
 #endif
     }
-    // else if (int8_mode_ == 1) {
-    //     FT_CHECK(weight_only_int8_fc_runner_.get() != NULL && attention_weights->query_weight.int8_kernel != NULL
-    //              && attention_weights->query_weight.weight_only_quant_scale != NULL);
-
-    //     weight_only_int8_fc_runner_->gemm(attention_input,
-    //                                       reinterpret_cast<const uint8_t*>(attention_weights->query_weight.int8_kernel),
-    //                                       attention_weights->query_weight.weight_only_quant_scale,
-    //                                       qkv_buf_,
-    //                                       m,
-    //                                       3 * local_hidden_units_,
-    //                                       hidden_units_,
-    //                                       mixed_gemm_workspace_,
-    //                                       mixed_gemm_ws_bytes_,
-    //                                       stream_);
-    // }
-    // else if (int8_mode_ == 2) {
-    //     cublas_wrapper_->Int8Gemm(3 * local_hidden_units_,
-    //                               m,
-    //                               hidden_units_,
-    //                               attention_weights->query_weight.int8_kernel,
-    //                               hidden_units_,
-    //                               input_tensors->at("input_query").getPtr<int8_t>(),
-    //                               hidden_units_,
-    //                               reinterpret_cast<int8_t*>(qkv_buf_),
-    //                               3 * local_hidden_units_,
-    //                               attention_weights->query_weight.scale_inter,
-    //                               true);
-    // }
-    // else {
-    //     cublas_wrapper_->Gemm(CUBLAS_OP_N,
-    //                           CUBLAS_OP_N,
-    //                           3 * local_hidden_units_,  // n
-    //                           m,
-    //                           hidden_units_,  // k
-    //                           attention_weights->query_weight.kernel,
-    //                           3 * local_hidden_units_,  // n
-    //                           attention_input,
-    //                           hidden_units_,  // k
-    //                           qkv_buf_,
-    //                           3 * local_hidden_units_ /* n */);
-    // }
+    else if (int8_mode_ == 1) {
+        FT_CHECK(weight_only_int8_fc_runner_.get() != NULL && attention_weights->query_weight.int8_kernel != NULL
+                 && attention_weights->query_weight.weight_only_quant_scale != NULL);
+
+        weight_only_int8_fc_runner_->gemm(attention_input,
+                                          reinterpret_cast<const uint8_t*>(attention_weights->query_weight.int8_kernel),
+                                          attention_weights->query_weight.weight_only_quant_scale,
+                                          qkv_buf_,
+                                          m,
+                                          3 * local_hidden_units_,
+                                          hidden_units_,
+                                          mixed_gemm_workspace_,
+                                          mixed_gemm_ws_bytes_
+                                          );
+                                        //   stream_);
+    }
+    else if (int8_mode_ == 2) {
+        cublas_wrapper_->Int8Gemm(3 * local_hidden_units_,
+                                  m,
+                                  hidden_units_,
+                                  attention_weights->query_weight.int8_kernel,
+                                  hidden_units_,
+                                  input_tensors->at("input_query").getPtr<int8_t>(),
+                                  hidden_units_,
+                                  reinterpret_cast<int8_t*>(qkv_buf_),
+                                  3 * local_hidden_units_,
+                                  attention_weights->query_weight.scale_inter,
+                                  true);
+    }
+    else {
+        cublas_wrapper_->Gemm(CUBLAS_OP_N,
+                              CUBLAS_OP_N,
+                              3 * local_hidden_units_,  // n
+                              m,
+                              hidden_units_,  // k
+                              attention_weights->query_weight.kernel,
+                              3 * local_hidden_units_,  // n
+                              attention_input,
+                              hidden_units_,  // k
+                              qkv_buf_,
+                              3 * local_hidden_units_ /* n */);
+    }
     {
         cublas_wrapper_->Gemm(CUBLAS_OP_N,
                               CUBLAS_OP_N,
@@ -360,50 +361,52 @@ void GptContextAttentionLayer<T>::forward(TensorMap*                output_tenso
 #endif
         }
         else {
-            // if (int8_mode_ == 1) {
-            //     FT_CHECK(weight_only_int8_fc_runner_.get() != NULL
-            //              && attention_weights->attention_output_weight.int8_kernel != NULL
-            //              && attention_weights->attention_output_weight.weight_only_quant_scale != NULL);
-
-            //     weight_only_int8_fc_runner_->gemm(
-            //         qkv_buf_3_,
-            //         reinterpret_cast<const uint8_t*>(attention_weights->attention_output_weight.int8_kernel),
-            //         attention_weights->attention_output_weight.weight_only_quant_scale,
-            //         attention_out,
-            //         m,
-            //         hidden_units_,
-            //         local_hidden_units_,
-            //         mixed_gemm_workspace_,
-            //         mixed_gemm_ws_bytes_,
-            //         stream_);
-            // }
-            // else if (int8_mode_ == 2) {
-            //     int8_fc_runner_->gemm(reinterpret_cast<int8_t*>(qkv_buf_3_),
-            //                           attention_weights->attention_output_weight.int8_kernel,
-            //                           QuantMode::PerTensorQuant,
-            //                           attention_weights->attention_output_weight.scale_inter,
-            //                           attention_weights->attention_output_weight.scale_out,
-            //                           output_tensors->at("hidden_features").getPtr<T>(),
-            //                           m,
-            //                           hidden_units_,
-            //                           local_hidden_units_,
-            //                           nullptr,
-            //                           0,
-            //                           stream_);
-            // }
-            // else {
-            //     cublas_wrapper_->Gemm(CUBLAS_OP_N,
-            //                           CUBLAS_OP_N,
-            //                           hidden_units_,
-            //                           m,
-            //                           local_hidden_units_,
-            //                           attention_weights->attention_output_weight.kernel,
-            //                           hidden_units_,
-            //                           qkv_buf_3_,
-            //                           local_hidden_units_,
-            //                           attention_out,
-            //                           hidden_units_);
-            // }
+            if (int8_mode_ == 1) {
+                FT_CHECK(weight_only_int8_fc_runner_.get() != NULL
+                         && attention_weights->attention_output_weight.int8_kernel != NULL
+                         && attention_weights->attention_output_weight.weight_only_quant_scale != NULL);
+
+                weight_only_int8_fc_runner_->gemm(
+                    qkv_buf_3_,
+                    reinterpret_cast<const uint8_t*>(attention_weights->attention_output_weight.int8_kernel),
+                    attention_weights->attention_output_weight.weight_only_quant_scale,
+                    attention_out,
+                    m,
+                    hidden_units_,
+                    local_hidden_units_,
+                    mixed_gemm_workspace_,
+                    mixed_gemm_ws_bytes_
+                    );
+                    // stream_);
+            }
+            else if (int8_mode_ == 2) {
+                int8_fc_runner_->gemm(reinterpret_cast<int8_t*>(qkv_buf_3_),
+                                      attention_weights->attention_output_weight.int8_kernel,
+                                      QuantMode::PerTensorQuant,
+                                      attention_weights->attention_output_weight.scale_inter,
+                                      attention_weights->attention_output_weight.scale_out,
+                                      output_tensors->at("hidden_features").getPtr<T>(),
+                                      m,
+                                      hidden_units_,
+                                      local_hidden_units_,
+                                      nullptr,
+                                      0
+                                      );
+                                    //   stream_);
+            }
+            else {
+                cublas_wrapper_->Gemm(CUBLAS_OP_N,
+                                      CUBLAS_OP_N,
+                                      hidden_units_,
+                                      m,
+                                      local_hidden_units_,
+                                      attention_weights->attention_output_weight.kernel,
+                                      hidden_units_,
+                                      qkv_buf_3_,
+                                      local_hidden_units_,
+                                      attention_out,
+                                      hidden_units_);
+            }
             {
                 cublas_wrapper_->Gemm(CUBLAS_OP_N,
                                       CUBLAS_OP_N,
@@ -451,8 +454,8 @@ GptContextAttentionLayer<T>::GptContextAttentionLayer(size_t           max_batch
     rotary_embedding_dim_(0),
     neox_rotary_style_(false),
     is_qk_buf_float_(is_qk_buf_float || int8_mode == 2),
-    // weight_only_int8_fc_runner_(int8_mode == 1 ? std::make_shared<CutlassFpAIntBGemmRunner<T, uint8_t>>() : nullptr),
-    // int8_fc_runner_(int8_mode == 2 ? std::make_shared<CutlassInt8GemmRunner<T>>() : nullptr),
+    weight_only_int8_fc_runner_(int8_mode == 1 ? std::make_shared<CutlassFpAIntBGemmRunner<T, uint8_t>>() : nullptr),
+    int8_fc_runner_(int8_mode == 2 ? std::make_shared<CutlassInt8GemmRunner<T>>() : nullptr),
     int8_mode_(int8_mode)
 {
 }
@@ -481,8 +484,8 @@ GptContextAttentionLayer<T>::GptContextAttentionLayer(size_t           max_batch
     rotary_embedding_dim_(0),
     neox_rotary_style_(false),
     is_qk_buf_float_(is_qk_buf_float || int8_mode == 2),
-    // weight_only_int8_fc_runner_(int8_mode == 1 ? std::make_shared<CutlassFpAIntBGemmRunner<T, uint8_t>>() : nullptr),
-    // int8_fc_runner_(int8_mode == 2 ? std::make_shared<CutlassInt8GemmRunner<T>>() : nullptr),
+    weight_only_int8_fc_runner_(int8_mode == 1 ? std::make_shared<CutlassFpAIntBGemmRunner<T, uint8_t>>() : nullptr),
+    int8_fc_runner_(int8_mode == 2 ? std::make_shared<CutlassInt8GemmRunner<T>>() : nullptr),
     int8_mode_(int8_mode)
 {
     FT_LOG_DEBUG(__PRETTY_FUNCTION__);
@@ -515,8 +518,8 @@ GptContextAttentionLayer<T>::GptContextAttentionLayer(size_t           max_batch
     rotary_embedding_dim_(rotary_embedding_dim),
     neox_rotary_style_(neox_rotary_style),
     is_qk_buf_float_(is_qk_buf_float),
-    // weight_only_int8_fc_runner_(int8_mode == 1 ? std::make_shared<CutlassFpAIntBGemmRunner<T, uint8_t>>() : nullptr),
-    // int8_fc_runner_(int8_mode == 2 ? std::make_shared<CutlassInt8GemmRunner<T>>() : nullptr),
+    weight_only_int8_fc_runner_(int8_mode == 1 ? std::make_shared<CutlassFpAIntBGemmRunner<T, uint8_t>>() : nullptr),
+    int8_fc_runner_(int8_mode == 2 ? std::make_shared<CutlassInt8GemmRunner<T>>() : nullptr),
     int8_mode_(int8_mode)
 {
     FT_LOG_DEBUG(__PRETTY_FUNCTION__);
@@ -540,8 +543,8 @@ GptContextAttentionLayer<T>::GptContextAttentionLayer(GptContextAttentionLayer<T
     rotary_embedding_dim_(attention_layer.rotary_embedding_dim_),
     neox_rotary_style_(attention_layer.neox_rotary_style_),
     is_qk_buf_float_(attention_layer.is_qk_buf_float_),
-    // weight_only_int8_fc_runner_(attention_layer.weight_only_int8_fc_runner_),
-    // int8_fc_runner_(attention_layer.int8_fc_runner_),
+    weight_only_int8_fc_runner_(attention_layer.weight_only_int8_fc_runner_),
+    int8_fc_runner_(attention_layer.int8_fc_runner_),
     int8_mode_(attention_layer.int8_mode_)
 {
 }
@@ -591,26 +594,26 @@ void GptContextAttentionLayer<T>::allocateBuffer(size_t batch_size, size_t seq_l
         }
     }
 
-    // if (int8_mode_ == 1) {
-    //     // We use max_size for n and k since we reuse buffers for both FCs and want to allocate the max
-    //     // possible memory that would be required by any of the individual gemms.
-    //     const int max_size    = std::max(hidden_units_, 3 * local_hidden_units_);
-    //     mixed_gemm_ws_bytes_  = weight_only_int8_fc_runner_->getWorkspaceSize(batch_size * seq_len, max_size, max_size);
-    //     mixed_gemm_workspace_ = (char*)allocator_->reMalloc(mixed_gemm_workspace_, mixed_gemm_ws_bytes_, false);
-    // }
-
-    // if (int8_mode_ == 1) {
-    //     // We use max_size for n and k since we reuse buffers for both FCs and want to allocate the max
-    //     // possible memory that would be required by any of the individual gemms.
-    //     const int max_size    = std::max(hidden_units_, 3 * local_hidden_units_);
-    //     mixed_gemm_ws_bytes_  = weight_only_int8_fc_runner_->getWorkspaceSize(batch_size * seq_len, max_size, max_size);
-    //     mixed_gemm_workspace_ = (char*)allocator_->reMalloc(mixed_gemm_workspace_, mixed_gemm_ws_bytes_, false);
-    // }
-    // else if (int8_mode_ == 2) {
-    //     const int max_size   = std::max(hidden_units_, 3 * local_hidden_units_);
-    //     int8_gemm_ws_bytes_  = int8_fc_runner_->getWorkspaceSize(batch_size * seq_len, max_size, max_size);
-    //     int8_gemm_workspace_ = (char*)allocator_->reMalloc(int8_gemm_workspace_, int8_gemm_ws_bytes_, false);
-    // }
+    if (int8_mode_ == 1) {
+        // We use max_size for n and k since we reuse buffers for both FCs and want to allocate the max
+        // possible memory that would be required by any of the individual gemms.
+        const int max_size    = std::max(hidden_units_, 3 * local_hidden_units_);
+        mixed_gemm_ws_bytes_  = weight_only_int8_fc_runner_->getWorkspaceSize(batch_size * seq_len, max_size, max_size);
+        mixed_gemm_workspace_ = (char*)allocator_->reMalloc(mixed_gemm_workspace_, mixed_gemm_ws_bytes_, false);
+    }
+
+    if (int8_mode_ == 1) {
+        // We use max_size for n and k since we reuse buffers for both FCs and want to allocate the max
+        // possible memory that would be required by any of the individual gemms.
+        const int max_size    = std::max(hidden_units_, 3 * local_hidden_units_);
+        mixed_gemm_ws_bytes_  = weight_only_int8_fc_runner_->getWorkspaceSize(batch_size * seq_len, max_size, max_size);
+        mixed_gemm_workspace_ = (char*)allocator_->reMalloc(mixed_gemm_workspace_, mixed_gemm_ws_bytes_, false);
+    }
+    else if (int8_mode_ == 2) {
+        const int max_size   = std::max(hidden_units_, 3 * local_hidden_units_);
+        int8_gemm_ws_bytes_  = int8_fc_runner_->getWorkspaceSize(batch_size * seq_len, max_size, max_size);
+        int8_gemm_workspace_ = (char*)allocator_->reMalloc(int8_gemm_workspace_, int8_gemm_ws_bytes_, false);
+    }
     is_allocate_buffer_ = true;
 }
 
diff --git a/src/fastertransformer/layers/attention_layers/GptContextAttentionLayer.h b/src/fastertransformer/layers/attention_layers/GptContextAttentionLayer.h
index 4432d78..a2b2f40 100644
--- a/src/fastertransformer/layers/attention_layers/GptContextAttentionLayer.h
+++ b/src/fastertransformer/layers/attention_layers/GptContextAttentionLayer.h
@@ -20,6 +20,8 @@
 #include "3rdparty/trt_fused_multihead_attention/qkvToContext.h"
 // #include "src/fastertransformer/kernels/cutlass_kernels/fpA_intB_gemm/fpA_intB_gemm.h"
 // #include "src/fastertransformer/kernels/cutlass_kernels/int8_gemm/int8_gemm.h"
+#include "src/fastertransformer/kernels/ck_kernels/fpA_intB_gemm/fpA_intB_gemm.h"
+#include "src/fastertransformer/kernels/ck_kernels/int8_gemm/int8_gemm.h"
 #include "src/fastertransformer/layers/attention_layers/BaseAttentionLayer.h"
 
 namespace fastertransformer {
@@ -54,8 +56,8 @@ private:
 
     bool is_qk_buf_float_;
 
-    // std::shared_ptr<CutlassFpAIntBGemmRunner<T, uint8_t>> weight_only_int8_fc_runner_;
-    // std::shared_ptr<CutlassInt8GemmRunner<T>>             int8_fc_runner_;
+    std::shared_ptr<CutlassFpAIntBGemmRunner<T, uint8_t>> weight_only_int8_fc_runner_;
+    std::shared_ptr<CutlassInt8GemmRunner<T>>             int8_fc_runner_;
 
 protected:
     using BaseAttentionLayer<T>::allocator_;
diff --git a/src/fastertransformer/utils/activation_types.h b/src/fastertransformer/utils/activation_types.h
index c608d3d..3dd1c43 100644
--- a/src/fastertransformer/utils/activation_types.h
+++ b/src/fastertransformer/utils/activation_types.h
@@ -16,7 +16,8 @@
 
 #pragma once
 
-#include "src/fastertransformer/utils/cuda_utils.h"
+// #include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 
 namespace fastertransformer {
 
diff --git a/src/fastertransformer/utils/cuda_utils.h b/src/fastertransformer/utils/cuda_utils.h
index 331b6f2..c83694e 100644
--- a/src/fastertransformer/utils/cuda_utils.h
+++ b/src/fastertransformer/utils/cuda_utils.h
@@ -19,7 +19,7 @@
 #include "3rdparty/INIReader.h"
 #include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
 #include "src/fastertransformer/utils/logger.h"
-
+#include "src/fastertransformer/kernels/ck_kernels/utils.h"
 #include <cublasLt.h>
 #include <cublas_v2.h>
 #include <cuda_runtime.h>
@@ -194,29 +194,29 @@ void check_abs_mean_val(const T* result, const int size);
         std::cout << "[FT][CALL] " << __FUNCTION__ << " " << std::endl;                                                \
     } while (0)
 
-[[noreturn]] inline void throwRuntimeError(const char* const file, int const line, std::string const& info = "")
-{
-    throw std::runtime_error(std::string("[FT][ERROR] ") + info + " Assertion fail: " + file + ":"
-                             + std::to_string(line) + " \n");
-}
-
-inline void myAssert(bool result, const char* const file, int const line, std::string const& info = "")
-{
-    if (!result) {
-        throwRuntimeError(file, line, info);
-    }
-}
-
-#define FT_CHECK(val) myAssert(val, __FILE__, __LINE__)
-#define FT_CHECK_WITH_INFO(val, info)                                                                                  \
-    do {                                                                                                               \
-        bool is_valid_val = (val);                                                                                     \
-        if (!is_valid_val) {                                                                                           \
-            fastertransformer::myAssert(is_valid_val, __FILE__, __LINE__, (info));                                     \
-        }                                                                                                              \
-    } while (0)
-
-#define FT_THROW(info) throwRuntimeError(__FILE__, __LINE__, info)
+// [[noreturn]] inline void throwRuntimeError(const char* const file, int const line, std::string const& info = "")
+// {
+//     throw std::runtime_error(std::string("[FT][ERROR] ") + info + " Assertion fail: " + file + ":"
+//                              + std::to_string(line) + " \n");
+// }
+
+// inline void myAssert(bool result, const char* const file, int const line, std::string const& info = "")
+// {
+//     if (!result) {
+//         throwRuntimeError(file, line, info);
+//     }
+// }
+
+// #define FT_CHECK(val) myAssert(val, __FILE__, __LINE__)
+// #define FT_CHECK_WITH_INFO(val, info)                                                                                  \
+//     do {                                                                                                               \
+//         bool is_valid_val = (val);                                                                                     \
+//         if (!is_valid_val) {                                                                                           \
+//             fastertransformer::myAssert(is_valid_val, __FILE__, __LINE__, (info));                                     \
+//         }                                                                                                              \
+//     } while (0)
+
+// #define FT_THROW(info) throwRuntimeError(__FILE__, __LINE__, info)
 
 #ifdef SPARSITY_ENABLED
 #define CHECK_CUSPARSE(func)                                                                                           \
-- 
GitLab